diff --git a/.circleci/test.yml b/.circleci/test.yml
index 12ea1636d72..a0fcb0fe05d 100644
--- a/.circleci/test.yml
+++ b/.circleci/test.yml
@@ -207,9 +207,9 @@ workflows:
- lint
- build_cpu_with_3rdparty:
name: maximum_version_cpu
- torch: 1.12.1
- torchvision: 0.13.1
- python: 3.9.0
+ torch: 1.13.0
+ torchvision: 0.14.0
+ python: 3.10.0
requires:
- minimum_version_cpu
- hold:
diff --git a/.dev_scripts/benchmark_regression/1-benchmark_valid.py b/.dev_scripts/benchmark_regression/1-benchmark_valid.py
index f2335a4fdd5..c8c63c2daa1 100644
--- a/.dev_scripts/benchmark_regression/1-benchmark_valid.py
+++ b/.dev_scripts/benchmark_regression/1-benchmark_valid.py
@@ -12,11 +12,12 @@
from mmengine import Config, DictAction, MMLogger
from mmengine.dataset import Compose, default_collate
from mmengine.fileio import FileClient
-from mmengine.runner import Runner
+from mmengine.runner import Runner, load_checkpoint
from modelindex.load_model_index import load
from rich.console import Console
from rich.table import Table
+from mmcls.apis import init_model
from mmcls.datasets import CIFAR10, CIFAR100, ImageNet
from mmcls.utils import register_all_modules
from mmcls.visualization import ClsVisualizer
@@ -82,20 +83,27 @@ def inference(config_file, checkpoint, work_dir, args, exp_name):
if args.cfg_options is not None:
cfg.merge_from_dict(args.cfg_options)
- # build the data pipeline
- test_dataset = cfg.test_dataloader.dataset
- if test_dataset.pipeline[0]['type'] != 'LoadImageFromFile':
- test_dataset.pipeline.insert(0, dict(type='LoadImageFromFile'))
- if test_dataset.type in ['CIFAR10', 'CIFAR100']:
- # The image shape of CIFAR is (32, 32, 3)
- test_dataset.pipeline.insert(1, dict(type='Resize', scale=32))
-
- data = Compose(test_dataset.pipeline)({'img_path': args.img})
- data = default_collate([data] * args.batch_size)
- resolution = tuple(data['inputs'].shape[-2:])
-
- runner: Runner = Runner.from_cfg(cfg)
- model = runner.model
+ if 'test_dataloader' in cfg:
+ # build the data pipeline
+ test_dataset = cfg.test_dataloader.dataset
+ if test_dataset.pipeline[0]['type'] != 'LoadImageFromFile':
+ test_dataset.pipeline.insert(0, dict(type='LoadImageFromFile'))
+ if test_dataset.type in ['CIFAR10', 'CIFAR100']:
+ # The image shape of CIFAR is (32, 32, 3)
+ test_dataset.pipeline.insert(1, dict(type='Resize', scale=32))
+
+ data = Compose(test_dataset.pipeline)({'img_path': args.img})
+ data = default_collate([data] * args.batch_size)
+ resolution = tuple(data['inputs'].shape[-2:])
+ model = Runner.from_cfg(cfg).model
+ forward = model.val_step
+ else:
+ # For configs only for get model.
+ model = init_model(cfg)
+ load_checkpoint(model, checkpoint, map_location='cpu')
+ data = torch.empty(1, 3, 224, 224).to(model.data_preprocessor.device)
+ resolution = (224, 224)
+ forward = model.extract_feat
# forward the model
result = {'resolution': resolution}
@@ -103,16 +111,16 @@ def inference(config_file, checkpoint, work_dir, args, exp_name):
if args.inference_time:
time_record = []
for _ in range(10):
- model.val_step(data) # warmup before profiling
+ forward(data) # warmup before profiling
torch.cuda.synchronize()
start = time()
- model.val_step(data)
+ forward(data)
torch.cuda.synchronize()
time_record.append((time() - start) / args.batch_size * 1000)
result['time_mean'] = np.mean(time_record[1:-1])
result['time_std'] = np.std(time_record[1:-1])
else:
- model.val_step(data)
+ forward(data)
result['model'] = config_file.stem
@@ -144,8 +152,8 @@ def show_summary(summary_data, args):
if args.inference_time:
table.add_column('Inference Time (std) (ms/im)')
if args.flops:
- table.add_column('Flops', justify='right', width=11)
- table.add_column('Params', justify='right')
+ table.add_column('Flops', justify='right', width=13)
+ table.add_column('Params', justify='right', width=11)
for model_name, summary in summary_data.items():
row = [model_name]
diff --git a/.dev_scripts/ckpt_tree.py b/.dev_scripts/ckpt_tree.py
new file mode 100644
index 00000000000..787020e1be0
--- /dev/null
+++ b/.dev_scripts/ckpt_tree.py
@@ -0,0 +1,186 @@
+import argparse
+import math
+from pathlib import Path
+
+import torch
+from rich.console import Console
+
+console = Console()
+
+prog_description = """\
+Draw the state dict tree.
+"""
+
+
+def parse_args():
+ parser = argparse.ArgumentParser(description=prog_description)
+ parser.add_argument(
+ 'path',
+ type=Path,
+ help='The path of the checkpoint or model config to draw.')
+ parser.add_argument('--depth', type=int, help='The max depth to draw.')
+ parser.add_argument(
+ '--full-name',
+ action='store_true',
+ help='Whether to print the full name of the key.')
+ parser.add_argument(
+ '--shape',
+ action='store_true',
+ help='Whether to print the shape of the parameter.')
+ parser.add_argument(
+ '--state-key',
+ type=str,
+ help='The key of the state dict in the checkpoint.')
+ parser.add_argument(
+ '--number',
+ action='store_true',
+ help='Mark all parameters and their index number.')
+ parser.add_argument(
+ '--node',
+ type=str,
+ help='Show the sub-tree of a node, like "backbone.layers".')
+ args = parser.parse_args()
+ return args
+
+
+def ckpt_to_state_dict(checkpoint, key=None):
+ if key is not None:
+ state_dict = checkpoint[key]
+ elif 'state_dict' in checkpoint:
+ # try mmcls style
+ state_dict = checkpoint['state_dict']
+ elif 'model' in checkpoint:
+ state_dict = checkpoint['model']
+ elif isinstance(next(iter(checkpoint.values())), torch.Tensor):
+ # try native style
+ state_dict = checkpoint
+ else:
+ raise KeyError('Please specify the key of state '
+ f'dict from {list(checkpoint.keys())}.')
+ return state_dict
+
+
+class StateDictTree:
+
+ def __init__(self, key='', value=None):
+ self.children = {}
+ self.key: str = key
+ self.value = value
+
+ def add_parameter(self, key, value):
+ keys = key.split('.', 1)
+ if len(keys) == 1:
+ self.children[key] = StateDictTree(key, value)
+ elif keys[0] in self.children:
+ self.children[keys[0]].add_parameter(keys[1], value)
+ else:
+ node = StateDictTree(keys[0])
+ node.add_parameter(keys[1], value)
+ self.children[keys[0]] = node
+
+ def __getitem__(self, key: str):
+ return self.children[key]
+
+ def __repr__(self) -> str:
+ with console.capture() as capture:
+ for line in self.iter_tree():
+ console.print(line)
+ return capture.get()
+
+ def __len__(self):
+ return len(self.children)
+
+ def draw_tree(self,
+ max_depth=None,
+ full_name=False,
+ with_shape=False,
+ with_value=False):
+ for line in self.iter_tree(
+ max_depth=max_depth,
+ full_name=full_name,
+ with_shape=with_shape,
+ with_value=with_value,
+ ):
+ console.print(line, highlight=False)
+
+ def iter_tree(
+ self,
+ lead='',
+ prefix='',
+ max_depth=None,
+ full_name=False,
+ with_shape=False,
+ with_value=False,
+ ):
+ if self.value is None:
+ key_str = f'[blue]{self.key}[/]'
+ elif with_shape:
+ key_str = f'[green]{self.key}[/] {tuple(self.value.shape)}'
+ elif with_value:
+ key_str = f'[green]{self.key}[/] {self.value}'
+ else:
+ key_str = f'[green]{self.key}[/]'
+
+ yield lead + prefix + key_str
+
+ lead = lead.replace('├─', '│ ')
+ lead = lead.replace('└─', ' ')
+ if self.key and full_name:
+ prefix = f'{prefix}{self.key}.'
+
+ if max_depth == 0:
+ return
+ elif max_depth is not None:
+ max_depth -= 1
+
+ for i, child in enumerate(self.children.values()):
+ level_lead = '├─' if i < len(self.children) - 1 else '└─'
+ yield from child.iter_tree(
+ lead=f'{lead}{level_lead} ',
+ prefix=prefix,
+ max_depth=max_depth,
+ full_name=full_name,
+ with_shape=with_shape,
+ with_value=with_value)
+
+
+def main():
+ args = parse_args()
+ if args.path.suffix in ['.json', '.py', '.yml']:
+ from mmengine.runner import get_state_dict
+
+ from mmcls.apis import init_model
+ model = init_model(args.path, device='cpu')
+ state_dict = get_state_dict(model)
+ else:
+ ckpt = torch.load(args.path, map_location='cpu')
+ state_dict = ckpt_to_state_dict(ckpt, args.state_key)
+
+ root = StateDictTree()
+ for k, v in state_dict.items():
+ root.add_parameter(k, v)
+
+ para_index = 0
+ mark_width = math.floor(math.log(len(state_dict), 10) + 1)
+ if args.node is not None:
+ for key in args.node.split('.'):
+ root = root[key]
+
+ for line in root.iter_tree(
+ max_depth=args.depth,
+ full_name=args.full_name,
+ with_shape=args.shape,
+ ):
+ if not args.number:
+ mark = ''
+ # A hack method to determine whether a line is parameter.
+ elif '[green]' in line:
+ mark = f'[red]({str(para_index).ljust(mark_width)})[/]'
+ para_index += 1
+ else:
+ mark = ' ' * (mark_width + 2)
+ console.print(mark + line, highlight=False)
+
+
+if __name__ == '__main__':
+ main()
diff --git a/.dev_scripts/compare_init.py b/.dev_scripts/compare_init.py
new file mode 100644
index 00000000000..71030f67d46
--- /dev/null
+++ b/.dev_scripts/compare_init.py
@@ -0,0 +1,121 @@
+#!/usr/bin/env python
+import argparse
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import torch
+from ckpt_tree import StateDictTree, ckpt_to_state_dict
+from rich.progress import track
+from scipy import stats
+
+prog_description = """\
+Compare the initialization distribution between state dicts by Kolmogorov-Smirnov test.
+""" # noqa: E501
+
+
+def parse_args():
+ parser = argparse.ArgumentParser(
+ formatter_class=argparse.RawDescriptionHelpFormatter,
+ description=prog_description)
+ parser.add_argument(
+ 'model_a',
+ type=Path,
+ help='The path of the first checkpoint or model config.')
+ parser.add_argument(
+ 'model_b',
+ type=Path,
+ help='The path of the second checkpoint or model config.')
+ parser.add_argument(
+ '--show',
+ action='store_true',
+ help='Whether to draw the KDE of variables')
+ parser.add_argument(
+ '-p',
+ default=0.01,
+ type=float,
+ help='The threshold of p-value. '
+ 'Higher threshold means more strict test.')
+ args = parser.parse_args()
+ return args
+
+
+def compare_distribution(state_dict_a, state_dict_b, p_thres):
+ assert len(state_dict_a) == len(state_dict_b)
+ for k, v1 in state_dict_a.items():
+ assert k in state_dict_b
+ v2 = state_dict_b[k]
+ v1 = v1.cpu().flatten()
+ v2 = v2.cpu().flatten()
+ pvalue = stats.kstest(v1, v2).pvalue
+ if pvalue < p_thres:
+ yield k, pvalue, v1, v2
+
+
+def state_dict_from_cfg_or_ckpt(path, state_key=None):
+ if path.suffix in ['.json', '.py', '.yml']:
+ from mmengine.runner import get_state_dict
+
+ from mmcls.apis import init_model
+ model = init_model(path, device='cpu')
+ model.init_weights()
+ return get_state_dict(model)
+ else:
+ ckpt = torch.load(path, map_location='cpu')
+ return ckpt_to_state_dict(ckpt, state_key)
+
+
+def main():
+ args = parse_args()
+
+ state_dict_a = state_dict_from_cfg_or_ckpt(args.model_a)
+ state_dict_b = state_dict_from_cfg_or_ckpt(args.model_b)
+ compare_keys = state_dict_a.keys() & state_dict_b.keys()
+ if len(compare_keys) == 0:
+ raise ValueError("The state dicts don't match, please convert "
+ 'to the same keys before comparison.')
+
+ root = StateDictTree()
+ for key in track(compare_keys):
+ if state_dict_a[key].shape != state_dict_b[key].shape:
+ raise ValueError(f'The shapes of "{key}" are different. '
+ 'Please check models in the same architecture.')
+
+ # Sample at most 30000 items to prevent long-time calcuation.
+ perm_ids = torch.randperm(state_dict_a[key].numel())[:30000]
+ value_a = state_dict_a[key].flatten()[perm_ids]
+ value_b = state_dict_b[key].flatten()[perm_ids]
+ pvalue = stats.kstest(value_a, value_b).pvalue
+ if pvalue < args.p:
+ root.add_parameter(key, round(pvalue, 4))
+ if args.show:
+ try:
+ import seaborn as sns
+ except ImportError:
+ raise ImportError('Please install `seaborn` by '
+ '`pip install seaborn` to show KDE.')
+ sample_a = str([round(v.item(), 2) for v in value_a[:10]])
+ sample_b = str([round(v.item(), 2) for v in value_b[:10]])
+ if value_a.std() > 0:
+ sns.kdeplot(value_a, fill=True)
+ else:
+ sns.scatterplot(x=[value_a[0].item()], y=[1])
+ if value_b.std() > 0:
+ sns.kdeplot(value_b, fill=True)
+ else:
+ sns.scatterplot(x=[value_b[0].item()], y=[1])
+ plt.legend([
+ f'{args.model_a.stem}: {sample_a}',
+ f'{args.model_b.stem}: {sample_b}'
+ ])
+ plt.title(key)
+ plt.show()
+ if len(root) > 0:
+ root.draw_tree(with_value=True)
+ print("Above parameters didn't pass the test, "
+ 'and the values are their similarity score.')
+ else:
+ print('The distributions of all weights are the same.')
+
+
+if __name__ == '__main__':
+ main()
diff --git a/.dev_scripts/generate_readme.py b/.dev_scripts/generate_readme.py
new file mode 100644
index 00000000000..bb7482b94f6
--- /dev/null
+++ b/.dev_scripts/generate_readme.py
@@ -0,0 +1,161 @@
+# flake8: noqa
+import argparse
+import warnings
+from collections import defaultdict
+from pathlib import Path
+
+from modelindex.load_model_index import load
+from modelindex.models.ModelIndex import ModelIndex
+
+prog_description = """\
+Use metafile to generate a README.md.
+
+Notice that the tool may fail in some corner cases, and you still need to check and fill some contents manually in the generated README.
+"""
+
+
+def parse_args():
+ parser = argparse.ArgumentParser(description=prog_description)
+ parser.add_argument('metafile', type=Path, help='The path of metafile')
+ parser.add_argument(
+ '--table', action='store_true', help='Only generate summary tables')
+ args = parser.parse_args()
+ return args
+
+
+def add_title(metafile: ModelIndex, readme: list):
+ paper = metafile.collections[0].paper
+ title = paper['Title']
+ url = paper['URL']
+ abbr = metafile.collections[0].name
+ papertype = metafile.collections[0].data.get('type', 'Algorithm')
+
+ readme.append(f'# {abbr}\n')
+ readme.append(f'> [{title}]({url})')
+ readme.append(f'')
+ readme.append('')
+
+
+def add_abstract(metafile, readme):
+ paper = metafile.collections[0].paper
+ url = paper['URL']
+ if 'arxiv' in url:
+ try:
+ import arxiv
+ search = arxiv.Search(id_list=[url.split('/')[-1]])
+ info = next(search.results())
+ abstract = info.summary
+ except ImportError:
+ warnings.warn('Install arxiv parser by `pip install arxiv` '
+ 'to automatically generate abstract.')
+ abstract = None
+ else:
+ abstract = None
+
+ readme.append('## Abstract\n')
+ if abstract is not None:
+ readme.append(abstract.replace('\n', ' '))
+
+ readme.append('')
+ readme.append('\n'
+ '
![]()
\n'
+ '
+
+
+
+
+
+
+
+
-## Results and models
+## Abstract
-### ImageNet-1k
+
-| Model | Epochs | Params(M) | Flops(G) | Top-1 (%) | Top-5 (%) | Config | Download |
-| :-----------: | :----: | :-------------------------------: | :-----------------------------: | :-------: | :-------: | :----------------------------------------------: | :-------------------------------------------------: |
-| RepVGG-A0\* | 120 | 9.11（train) \| 8.31 (deploy) | 1.52 (train) \| 1.36 (deploy) | 72.41 | 90.50 | [config (train)](./repvgg-A0_4xb64-coslr-120e_in1k.py) \| [config (deploy)](./deploy/repvgg-A0_deploy_4xb64-coslr-120e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_3rdparty_4xb64-coslr-120e_in1k_20210909-883ab98c.pth) |
-| RepVGG-A1\* | 120 | 14.09 (train) \| 12.79 (deploy) | 2.64 (train) \| 2.37 (deploy) | 74.47 | 91.85 | [config (train)](./repvgg-A1_4xb64-coslr-120e_in1k.py) \| [config (deploy)](./deploy/repvgg-A1_deploy_4xb64-coslr-120e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A1_3rdparty_4xb64-coslr-120e_in1k_20210909-24003a24.pth) |
-| RepVGG-A2\* | 120 | 28.21 (train) \| 25.5 (deploy) | 5.7 (train) \| 5.12 (deploy) | 76.48 | 93.01 | [config (train)](./repvgg-A2_4xb64-coslr-120e_in1k.py) \|[config (deploy)](./deploy/repvgg-A2_deploy_4xb64-coslr-120e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A2_3rdparty_4xb64-coslr-120e_in1k_20210909-97d7695a.pth) |
-| RepVGG-B0\* | 120 | 15.82 (train) \| 14.34 (deploy) | 3.42 (train) \| 3.06 (deploy) | 75.14 | 92.42 | [config (train)](./repvgg-B0_4xb64-coslr-120e_in1k.py) \|[config (deploy)](./deploy/repvgg-B0_deploy_4xb64-coslr-120e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B0_3rdparty_4xb64-coslr-120e_in1k_20210909-446375f4.pth) |
-| RepVGG-B1\* | 120 | 57.42 (train) \| 51.83 (deploy) | 13.16 (train) \| 11.82 (deploy) | 78.37 | 94.11 | [config (train)](./repvgg-B1_4xb64-coslr-120e_in1k.py) \|[config (deploy)](./deploy/repvgg-B1_deploy_4xb64-coslr-120e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1_3rdparty_4xb64-coslr-120e_in1k_20210909-750cdf67.pth) |
-| RepVGG-B1g2\* | 120 | 45.78 (train) \| 41.36 (deploy) | 9.82 (train) \| 8.82 (deploy) | 77.79 | 93.88 | [config (train)](./repvgg-B1g2_4xb64-coslr-120e_in1k.py) \|[config (deploy)](./deploy/repvgg-B1g2_deploy_4xb64-coslr-120e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g2_3rdparty_4xb64-coslr-120e_in1k_20210909-344f6422.pth) |
-| RepVGG-B1g4\* | 120 | 39.97 (train) \| 36.13 (deploy) | 8.15 (train) \| 7.32 (deploy) | 77.58 | 93.84 | [config (train)](./repvgg-B1g4_4xb64-coslr-120e_in1k.py) \|[config (deploy)](./deploy/repvgg-B1g4_deploy_4xb64-coslr-120e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g4_3rdparty_4xb64-coslr-120e_in1k_20210909-d4c1a642.pth) |
-| RepVGG-B2\* | 120 | 89.02 (train) \| 80.32 (deploy) | 20.46 (train) \| 18.39 (deploy) | 78.78 | 94.42 | [config (train)](./repvgg-B2_4xb64-coslr-120e_in1k.py) \|[config (deploy)](./deploy/repvgg-B2_deploy_4xb64-coslr-120e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2_3rdparty_4xb64-coslr-120e_in1k_20210909-bd6b937c.pth) |
-| RepVGG-B2g4\* | 200 | 61.76 (train) \| 55.78 (deploy) | 12.63 (train) \| 11.34 (deploy) | 79.38 | 94.68 | [config (train)](./repvgg-B2g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py) \|[config (deploy)](./deploy/repvgg-B2g4_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2g4_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-7b7955f0.pth) |
-| RepVGG-B3\* | 200 | 123.09 (train) \| 110.96 (deploy) | 29.17 (train) \| 26.22 (deploy) | 80.52 | 95.26 | [config (train)](./repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py) \|[config (deploy)](./deploy/repvgg-B3_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-dda968bf.pth) |
-| RepVGG-B3g4\* | 200 | 83.83 (train) \| 75.63 (deploy) | 17.9 (train) \| 16.08 (deploy) | 80.22 | 95.10 | [config (train)](./repvgg-B3g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py) \|[config (deploy)](./deploy/repvgg-B3g4_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3g4_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-4e54846a.pth) |
-| RepVGG-D2se\* | 200 | 133.33 (train) \| 120.39 (deploy) | 36.56 (train) \| 32.85 (deploy) | 81.81 | 95.94 | [config (train)](./repvgg-D2se_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py) \|[config (deploy)](./deploy/repvgg-D2se_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-D2se_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-cf3139b7.pth) |
-
-*Models with * are converted from the [official repo](https://github.com/DingXiaoH/RepVGG). The config files of these models are only for validation. We don't ensure these config files' training accuracy and welcome you to contribute your reproduction results.*
+Show the paper's abstract
+
+
+We present a simple but powerful architecture of convolutional neural network, which has a VGG-like inference-time body composed of nothing but a stack of 3x3 convolution and ReLU, while the training-time model has a multi-branch topology. Such decoupling of the training-time and inference-time architecture is realized by a structural re-parameterization technique so that the model is named RepVGG. On ImageNet, RepVGG reaches over 80% top-1 accuracy, which is the first time for a plain model, to the best of our knowledge. On NVIDIA 1080Ti GPU, RepVGG models run 83% faster than ResNet-50 or 101% faster than ResNet-101 with higher accuracy and show favorable accuracy-speed trade-off compared to the state-of-the-art models like EfficientNet and RegNet.
+
+
+
## How to use
-The checkpoints provided are all `training-time` models. Use the reparameterize tool to switch them to more efficient `inference-time` architecture, which not only has fewer parameters but also less calculations.
+The checkpoints provided are all `training-time` models. Use the reparameterize tool or `switch_to_deploy` interface to switch them to more efficient `inference-time` architecture, which not only has fewer parameters but also less calculations.
+
+
+
+**Predict image**
+
+Use `classifier.backbone.switch_to_deploy()` interface to switch the RepVGG models into inference mode.
+
+```python
+>>> import torch
+>>> from mmcls.apis import init_model, inference_model
+>>>
+>>> model = init_model('configs/repvgg/repvgg-A0_8xb32_in1k.py', 'https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth')
+>>> results = inference_model(model, 'demo/demo.JPEG')
+>>> print( (results['pred_class'], results['pred_score']) )
+('sea snake' 0.8338906168937683)
+>>>
+>>> # switch to deploy mode
+>>> model.backbone.switch_to_deploy()
+>>> results = inference_model(model, 'demo/demo.JPEG')
+>>> print( (results['pred_class'], results['pred_score']) )
+('sea snake', 0.7883061170578003)
+```
+
+**Use the model**
+
+```python
+>>> import torch
+>>> from mmcls.apis import get_model
+>>>
+>>> model = get_model("repvgg-a0_8xb32_in1k", pretrained=True)
+>>> model.eval()
+>>> inputs = torch.rand(1, 3, 224, 224).to(model.data_preprocessor.device)
+>>> # To get classification scores.
+>>> out = model(inputs)
+>>> print(out.shape)
+torch.Size([1, 1000])
+>>> # To extract features.
+>>> outs = model.extract_feat(inputs)
+>>> print(outs[0].shape)
+torch.Size([1, 1280])
+>>>
+>>> # switch to deploy mode
+>>> model.backbone.switch_to_deploy()
+>>> out_deploy = model(inputs)
+>>> print(out.shape)
+torch.Size([1, 1000])
+>>> assert torch.allclose(out, out_deploy, rtol=1e-4, atol=1e-5) # pass without error
+```
+
+**Train/Test Command**
+
+Place the ImageNet dataset to the `data/imagenet/` directory, or prepare datasets according to the [docs](https://mmclassification.readthedocs.io/en/1.x/user_guides/dataset_prepare.html#prepare-dataset).
+
+Train:
+
+```shell
+python tools/train.py configs/repvgg/repvgg-a0_8xb32_in1k.py
+```
+
+Download Checkpoint:
+
+```shell
+wget https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth
+```
+
+Test use unfused model:
+
+```shell
+python tools/test.py configs/repvgg/repvgg-a0_8xb32_in1k.py repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth
+```
-### Use tool
+Reparameterize checkpoint:
+
+```shell
+python ./tools/convert_models/reparameterize_model.py configs/repvgg/repvgg-a0_8xb32_in1k.py repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth repvgg_A0_deploy.pth
+```
+
+Test use fused model:
+
+```shell
+python tools/test.py configs/repvgg/repvgg-A0_8xb32_in1k.py repvgg_A0_deploy.pth --cfg-options model.backbone.deploy=True
+```
+
+or
+
+```shell
+python tools/test.py configs/repvgg/repvgg-A0_deploy_in1k.py repvgg_A0_deploy.pth
+```
+
+
+
+For more configurable parameters, please refer to the [API](https://mmclassification.readthedocs.io/en/1.x/api/generated/mmcls.models.backbones.RepVGG.html#mmcls.models.backbones.RepVGG).
+
+
+
+How to use the reparameterisation tool(click to show)
+
+
Use provided tool to reparameterize the given model and save the checkpoint:
@@ -45,52 +136,68 @@ Use provided tool to reparameterize the given model and save the checkpoint:
python tools/convert_models/reparameterize_model.py ${CFG_PATH} ${SRC_CKPT_PATH} ${TARGET_CKPT_PATH}
```
-`${CFG_PATH}` is the config file, `${SRC_CKPT_PATH}` is the source chenpoint file, `${TARGET_CKPT_PATH}` is the target deploy weight file path.
+`${CFG_PATH}` is the config file path, `${SRC_CKPT_PATH}` is the source chenpoint file path, `${TARGET_CKPT_PATH}` is the target deploy weight file path.
-To use reparameterized weights, the config file must switch to the deploy config files.
+For example:
-```bash
-python tools/test.py ${Deploy_CFG} ${Deploy_Checkpoint} --metrics accuracy
+```shell
+# download the weight
+wget https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth
+# reparameterize unfused weight to fused weight
+python ./tools/convert_models/reparameterize_model.py configs/repvgg/repvgg-a0_8xb32_in1k.py repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth repvgg-A0_deploy.pth
```
-### In the code
+To use reparameterized weights, the config file must switch to **the deploy config files** as [the deploy_A0 example](./repvgg-A0_deploy_in1k.py) or add `--cfg-options model.backbone.deploy=True` in command.
-Use `backbone.switch_to_deploy()` or `classificer.backbone.switch_to_deploy()` to switch to the deploy mode. For example:
+For example of using the reparameterized weights above:
-```python
-from mmcls.models import build_backbone
+```shell
+python ./tools/test.py ./configs/repvgg/repvgg-A0_deploy_in1k.py repvgg-A0_deploy.pth
+```
+
+You can get other deploy configs by modifying the [A0_deploy example](./repvgg-A0_deploy_in1k.py):
+
+```text
+# in repvgg-A0_deploy_in1k.py
+_base_ = '../repvgg-A0_8xb32_in1k.py' # basic A0 config
-backbone_cfg=dict(type='RepVGG',arch='A0'),
-backbone = build_backbone(backbone_cfg)
-backbone.switch_to_deploy()
+model = dict(backbone=dict(deploy=True)) # switch model into deploy mode
```
-or
+or add `--cfg-options model.backbone.deploy=True` in command as following：
-```python
-from mmcls.models import build_classifier
-
-cfg = dict(
- type='ImageClassifier',
- backbone=dict(
- type='RepVGG',
- arch='A0'),
- neck=dict(type='GlobalAveragePooling'),
- head=dict(
- type='LinearClsHead',
- num_classes=1000,
- in_channels=1280,
- loss=dict(type='CrossEntropyLoss', loss_weight=1.0),
- topk=(1, 5),
- ))
-
-classifier = build_classifier(cfg)
-classifier.backbone.switch_to_deploy()
+```shell
+python tools/test.py configs/repvgg/repvgg-A0_8xb32_in1k.py repvgg_A0_deploy.pth --cfg-options model.backbone.deploy=True
```
+
+
+
+
+## Results and models
+
+### ImageNet-1k
+
+| Model | Pretrain | Params(M)
（train\|deploy)
| Flops(G)
（train\|deploy)
| Top-1 (%) | Top-5 (%) | Config | Download |
+| :-------------------------: | :----------: | :-------------------------------------: | :--------------------------------------: | :-------: | :-------: | :-----------------------------: | :-------------------------------: |
+| repvgg-A0_8xb32_in1k | From scratch | 9.11 \| 8.31 | 1.53 \| 1.36 | 72.37 | 90.56 | [config](./repvgg-A0_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.log) |
+| repvgg-A1_8xb32_in1k | From scratch | 14.09 \| 12.79 | 2.65 \| 2.37 | 74.47 | 91.85 | [config](./repvgg-A1_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A1_8xb32_in1k_20221213-f81bf3df.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A1_8xb32_in1k_20221213-f81bf3df.log) |
+| repvgg-A2_8xb32_in1k | From scratch | 28.21 \| 25.5 | 5.72 \| 5.12 | 76.49 | 93.09 | [config](./repvgg-A2_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A2_8xb32_in1k_20221213-a8767caf.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A2_8xb32_in1k_20221213-a8767caf.log) |
+| repvgg-B0_8xb32_in1k | From scratch | 15.82 \| 14.34 | 3.43 \| 3.06 | 75.27 | 92.21 | [config](./repvgg-B0_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B0_8xb32_in1k_20221213-5091ecc7.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B0_8xb32_in1k_20221213-5091ecc7.log) |
+| repvgg-B1_8xb32_in1k | From scratch | 57.42 \| 51.83 | 13.20 \| 11.81 | 78.19 | 94.04 | [config](./repvgg-B1_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1_8xb32_in1k_20221213-d17c45e7.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1_8xb32_in1k_20221213-d17c45e7.log) |
+| repvgg-B1g2_8xb32_in1k | From scratch | 45.78 \| 41.36 | 9.86 \| 8.80 | 77.87 | 93.99 | [config](./repvgg-B1g2_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g2_8xb32_in1k_20221213-ae6428fd.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g2_8xb32_in1k_20221213-ae6428fd.log) |
+| repvgg-B1g4_8xb32_in1k | From scratch | 39.97 \| 36.13 | 8.19 \| 7.30 | 77.81 | 93.77 | [config](./repvgg-B1g4_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g4_8xb32_in1k_20221213-a7a4aaea.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g4_8xb32_in1k_20221213-a7a4aaea.log) |
+| repvgg-B2_8xb32_in1k | From scratch | 89.02 \| 80.32 | 20.5 \| 18.4 | 78.58 | 94.23 | [config](./repvgg-B2_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2_8xb32_in1k_20221213-d8b420ef.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2_8xb32_in1k_20221213-d8b420ef.log) |
+| repvgg-B2g4_8xb32_in1k | From scratch | 61.76 \| 55.78 | 12.7 \| 11.3 | 79.44 | 94.72 | [config](./repvgg-B2g4_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2g4_8xb32_in1k_20221213-0c1990eb.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2g4_8xb32_in1k_20221213-0c1990eb.log) |
+| repvgg-B3_8xb32_in1k | From scratch | 123.09 \| 110.96 | 29.2 \| 26.2 | 80.58 | 95.33 | [config](./repvgg-B3_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3_8xb32_in1k_20221213-927a329a.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3_8xb32_in1k_20221213-927a329a.log) |
+| repvgg-B3g4_8xb32_in1k | From scratch | 83.83 \| 75.63 | 18.0 \| 16.1 | 80.26 | 95.15 | [config](./repvgg-B3g4_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3g4_8xb32_in1k_20221213-e01cb280.pth) \| [log](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3g4_8xb32_in1k_20221213-e01cb280.log) |
+| repvgg-D2se_3rdparty_in1k\* | From scratch | 133.33 \| 120.39 | 36.6 \| 32.8 | 81.81 | 95.94 | [config](./repvgg-D2se_8xb32_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-D2se_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-cf3139b7.pth) |
+
+*Models with * are converted from the [official repo](https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L250). The config files of these models are only for inference. We don't ensure these config files' training accuracy and welcome you to contribute your reproduction results.*
+
## Citation
-```
+```bibtex
@inproceedings{ding2021repvgg,
title={Repvgg: Making vgg-style convnets great again},
author={Ding, Xiaohan and Zhang, Xiangyu and Ma, Ningning and Han, Jungong and Ding, Guiguang and Sun, Jian},
diff --git a/configs/repvgg/deploy/repvgg-A0_deploy_4xb64-coslr-120e_in1k.py b/configs/repvgg/deploy/repvgg-A0_deploy_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index 20787f286da..00000000000
--- a/configs/repvgg/deploy/repvgg-A0_deploy_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-A0_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-A1_deploy_4xb64-coslr-120e_in1k.py b/configs/repvgg/deploy/repvgg-A1_deploy_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index eea0da9c58c..00000000000
--- a/configs/repvgg/deploy/repvgg-A1_deploy_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-A1_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-A2_deploy_4xb64-coslr-120e_in1k.py b/configs/repvgg/deploy/repvgg-A2_deploy_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index 7b0cea7b7d5..00000000000
--- a/configs/repvgg/deploy/repvgg-A2_deploy_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-A2_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-B0_deploy_4xb64-coslr-120e_in1k.py b/configs/repvgg/deploy/repvgg-B0_deploy_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index 23a2898ac56..00000000000
--- a/configs/repvgg/deploy/repvgg-B0_deploy_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-B0_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-B1_deploy_4xb64-coslr-120e_in1k.py b/configs/repvgg/deploy/repvgg-B1_deploy_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index 24355edac7f..00000000000
--- a/configs/repvgg/deploy/repvgg-B1_deploy_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-B1_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-B1g2_deploy_4xb64-coslr-120e_in1k.py b/configs/repvgg/deploy/repvgg-B1g2_deploy_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index 579fcc47b9c..00000000000
--- a/configs/repvgg/deploy/repvgg-B1g2_deploy_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-B1g2_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-B1g4_deploy_4xb64-coslr-120e_in1k.py b/configs/repvgg/deploy/repvgg-B1g4_deploy_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index eab5d440374..00000000000
--- a/configs/repvgg/deploy/repvgg-B1g4_deploy_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-B1g4_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-B2_deploy_4xb64-coslr-120e_in1k.py b/configs/repvgg/deploy/repvgg-B2_deploy_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index 0681f14dc36..00000000000
--- a/configs/repvgg/deploy/repvgg-B2_deploy_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-B2_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-B2g4_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py b/configs/repvgg/deploy/repvgg-B2g4_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
deleted file mode 100644
index 8f1840145f7..00000000000
--- a/configs/repvgg/deploy/repvgg-B2g4_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-B2g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-B3_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py b/configs/repvgg/deploy/repvgg-B3_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
deleted file mode 100644
index e60b0678a9e..00000000000
--- a/configs/repvgg/deploy/repvgg-B3_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-B3g4_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py b/configs/repvgg/deploy/repvgg-B3g4_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
deleted file mode 100644
index 46f187789a3..00000000000
--- a/configs/repvgg/deploy/repvgg-B3g4_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-B3g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/deploy/repvgg-D2se_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py b/configs/repvgg/deploy/repvgg-D2se_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
deleted file mode 100644
index 66dff3b6d44..00000000000
--- a/configs/repvgg/deploy/repvgg-D2se_deploy_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = '../repvgg-D2se_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py'
-
-model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/metafile.yml b/configs/repvgg/metafile.yml
index 84fee5911c1..8c550729aea 100644
--- a/configs/repvgg/metafile.yml
+++ b/configs/repvgg/metafile.yml
@@ -14,57 +14,48 @@ Collections:
Version: v0.16.0
Models:
- - Name: repvgg-A0_3rdparty_4xb64-coslr-120e_in1k
+ - Name: repvgg-A0_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-A0_4xb64-coslr-120e_in1k.py
+ Config: configs/repvgg/repvgg-A0_8xb32_in1k.py
Metadata:
- FLOPs: 1520000000
- Parameters: 9110000
+ FLOPs: 1360233728
+ Parameters: 8309384
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 72.41
- Top 5 Accuracy: 90.50
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_3rdparty_4xb64-coslr-120e_in1k_20210909-883ab98c.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L196
- - Name: repvgg-A1_3rdparty_4xb64-coslr-120e_in1k
+ Top 1 Accuracy: 72.37
+ Top 5 Accuracy: 90.56
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A0_8xb32_in1k_20221213-60ae8e23.pth
+ - Name: repvgg-A1_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-A1_4xb64-coslr-120e_in1k.py
+ Config: configs/repvgg/repvgg-A1_8xb32_in1k.py
Metadata:
- FLOPs: 2640000000
- Parameters: 14090000
+ FLOPs: 2362750208
+ Parameters: 12789864
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 74.47
- Top 5 Accuracy: 91.85
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A1_3rdparty_4xb64-coslr-120e_in1k_20210909-24003a24.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L200
- - Name: repvgg-A2_3rdparty_4xb64-coslr-120e_in1k
+ Top 1 Accuracy: 74.23
+ Top 5 Accuracy: 91.80
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A1_8xb32_in1k_20221213-f81bf3df.pth
+ - Name: repvgg-A2_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-A2_4xb64-coslr-120e_in1k.py
+ Config: configs/repvgg/repvgg-A2_8xb32_in1k.py
Metadata:
- FLOPs: 28210000000
- Parameters: 5700000
+ FLOPs: 5115612544
+ Parameters: 25499944
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 76.48
- Top 5 Accuracy: 93.01
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A2_3rdparty_4xb64-coslr-120e_in1k_20210909-97d7695a.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L204
- - Name: repvgg-B0_3rdparty_4xb64-coslr-120e_in1k
+ Top 1 Accuracy: 76.49
+ Top 5 Accuracy: 93.09
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-A2_8xb32_in1k_20221213-a8767caf.pth
+ - Name: repvgg-B0_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-B0_4xb64-coslr-120e_in1k.py
+ Config: configs/repvgg/repvgg-B0_8xb32_in1k.py
Metadata:
FLOPs: 15820000000
Parameters: 3420000
@@ -72,130 +63,106 @@ Models:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 75.14
- Top 5 Accuracy: 92.42
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B0_3rdparty_4xb64-coslr-120e_in1k_20210909-446375f4.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L208
- - Name: repvgg-B1_3rdparty_4xb64-coslr-120e_in1k
+ Top 1 Accuracy: 75.27
+ Top 5 Accuracy: 92.21
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B0_8xb32_in1k_20221213-5091ecc7.pth
+ - Name: repvgg-B1_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-B1_4xb64-coslr-120e_in1k.py
+ Config: configs/repvgg/repvgg-B1_8xb32_in1k.py
Metadata:
- FLOPs: 57420000000
- Parameters: 13160000
+ FLOPs: 11813537792
+ Parameters: 51829480
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 78.37
- Top 5 Accuracy: 94.11
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1_3rdparty_4xb64-coslr-120e_in1k_20210909-750cdf67.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L212
- - Name: repvgg-B1g2_3rdparty_4xb64-coslr-120e_in1k
+ Top 1 Accuracy: 78.19
+ Top 5 Accuracy: 94.04
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1_8xb32_in1k_20221213-d17c45e7.pth
+ - Name: repvgg-B1g2_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-B1g2_4xb64-coslr-120e_in1k.py
+ Config: configs/repvgg/repvgg-B1g2_8xb32_in1k.py
Metadata:
- FLOPs: 45780000000
- Parameters: 9820000
+ FLOPs: 8807794688
+ Parameters: 41360104
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 77.79
- Top 5 Accuracy: 93.88
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g2_3rdparty_4xb64-coslr-120e_in1k_20210909-344f6422.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L216
- - Name: repvgg-B1g4_3rdparty_4xb64-coslr-120e_in1k
+ Top 1 Accuracy: 77.87
+ Top 5 Accuracy: 93.99
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g2_8xb32_in1k_20221213-ae6428fd.pth
+ - Name: repvgg-B1g4_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-B1g4_4xb64-coslr-120e_in1k.py
+ Config: configs/repvgg/repvgg-B1g4_8xb32_in1k.py
Metadata:
- FLOPs: 39970000000
- Parameters: 8150000
+ FLOPs: 7304923136
+ Parameters: 36125416
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 77.58
- Top 5 Accuracy: 93.84
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g4_3rdparty_4xb64-coslr-120e_in1k_20210909-d4c1a642.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L220
- - Name: repvgg-B2_3rdparty_4xb64-coslr-120e_in1k
+ Top 1 Accuracy: 77.81
+ Top 5 Accuracy: 93.77
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B1g4_8xb32_in1k_20221213-a7a4aaea.pth
+ - Name: repvgg-B2_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-B2_4xb64-coslr-120e_in1k.py
+ Config: configs/repvgg/repvgg-B2_8xb32_in1k.py
Metadata:
- FLOPs: 89020000000
- Parameters: 20420000
+ FLOPs: 18374175232
+ Parameters: 80315112
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 78.78
- Top 5 Accuracy: 94.42
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2_3rdparty_4xb64-coslr-120e_in1k_20210909-bd6b937c.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L225
- - Name: repvgg-B2g4_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k
+ Top 1 Accuracy: 78.58
+ Top 5 Accuracy: 94.23
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2_8xb32_in1k_20221213-d8b420ef.pth
+ - Name: repvgg-B2g4_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-B2g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+ Config: configs/repvgg/repvgg-B2g4_8xb32_in1k.py
Metadata:
- FLOPs: 61760000000
- Parameters: 12630000
+ FLOPs: 11329464832
+ Parameters: 55777512
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 79.38
- Top 5 Accuracy: 94.68
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2g4_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-7b7955f0.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L229
- - Name: repvgg-B3_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k
+ Top 1 Accuracy: 79.44
+ Top 5 Accuracy: 94.72
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B2g4_8xb32_in1k_20221213-0c1990eb.pth
+ - Name: repvgg-B3_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+ Config: configs/repvgg/repvgg-B3_8xb32_in1k.py
Metadata:
- FLOPs: 123090000000
- Parameters: 29170000
+ FLOPs: 26206448128
+ Parameters: 110960872
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 80.52
- Top 5 Accuracy: 95.26
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-dda968bf.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L238
- - Name: repvgg-B3g4_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k
+ Top 1 Accuracy: 80.58
+ Top 5 Accuracy: 95.33
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3_8xb32_in1k_20221213-927a329a.pth
+ - Name: repvgg-B3g4_8xb32_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-B3g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+ Config: configs/repvgg/repvgg-B3g4_8xb32_in1k.py
Metadata:
- FLOPs: 83830000000
- Parameters: 17900000
+ FLOPs: 16062065152
+ Parameters: 75626728
Results:
- Dataset: ImageNet-1k
Task: Image Classification
Metrics:
- Top 1 Accuracy: 80.22
- Top 5 Accuracy: 95.10
- Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3g4_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k_20210909-4e54846a.pth
- Converted From:
- Weights: https://drive.google.com/drive/folders/1Avome4KvNp0Lqh2QwhXO6L5URQjzCjUq
- Code: https://github.com/DingXiaoH/RepVGG/blob/9f272318abfc47a2b702cd0e916fca8d25d683e7/repvgg.py#L238
- - Name: repvgg-D2se_3rdparty_4xb64-autoaug-lbs-mixup-coslr-200e_in1k
+ Top 1 Accuracy: 80.26
+ Top 5 Accuracy: 95.15
+ Weights: https://download.openmmlab.com/mmclassification/v0/repvgg/repvgg-B3g4_8xb32_in1k_20221213-e01cb280.pth
+ - Name: repvgg-D2se_3rdparty_in1k
In Collection: RepVGG
- Config: configs/repvgg/repvgg-D2se_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+ Config: configs/repvgg/repvgg-D2se_8xb32_in1k.py
Metadata:
- FLOPs: 133330000000
- Parameters: 36560000
+ FLOPs: 32838581760
+ Parameters: 120387572
Results:
- Dataset: ImageNet-1k
Task: Image Classification
diff --git a/configs/repvgg/repvgg-A0_4xb64-coslr-120e_in1k.py b/configs/repvgg/repvgg-A0_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index 8a93ed0a08c..00000000000
--- a/configs/repvgg/repvgg-A0_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,12 +0,0 @@
-_base_ = [
- '../_base_/models/repvgg-A0_in1k.py',
- '../_base_/datasets/imagenet_bs64_pil_resize.py',
- '../_base_/schedules/imagenet_bs256_coslr.py',
- '../_base_/default_runtime.py'
-]
-
-# schedule settings
-param_scheduler = dict(
- type='CosineAnnealingLR', T_max=120, by_epoch=True, begin=0, end=120)
-
-train_cfg = dict(by_epoch=True, max_epochs=120)
diff --git a/configs/repvgg/repvgg-A0_8xb32_in1k.py b/configs/repvgg/repvgg-A0_8xb32_in1k.py
new file mode 100644
index 00000000000..b767ae2a3e4
--- /dev/null
+++ b/configs/repvgg/repvgg-A0_8xb32_in1k.py
@@ -0,0 +1,33 @@
+_base_ = [
+ '../_base_/models/repvgg-A0_in1k.py',
+ '../_base_/datasets/imagenet_bs32_pil_resize.py',
+ '../_base_/schedules/imagenet_bs256_coslr.py',
+ '../_base_/default_runtime.py'
+]
+
+val_dataloader = dict(batch_size=256)
+test_dataloader = dict(batch_size=256)
+
+# schedule settings
+optim_wrapper = dict(
+ paramwise_cfg=dict(
+ bias_decay_mult=0.0,
+ custom_keys={
+ 'branch_3x3.norm': dict(decay_mult=0.0),
+ 'branch_1x1.norm': dict(decay_mult=0.0),
+ 'branch_norm.bias': dict(decay_mult=0.0),
+ }))
+
+# schedule settings
+param_scheduler = dict(
+ type='CosineAnnealingLR',
+ T_max=120,
+ by_epoch=True,
+ begin=0,
+ end=120,
+ convert_to_iter_based=True)
+
+train_cfg = dict(by_epoch=True, max_epochs=120)
+
+default_hooks = dict(
+ checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
diff --git a/configs/repvgg/repvgg-A0_deploy_in1k.py b/configs/repvgg/repvgg-A0_deploy_in1k.py
new file mode 100644
index 00000000000..16f0bbfcc7c
--- /dev/null
+++ b/configs/repvgg/repvgg-A0_deploy_in1k.py
@@ -0,0 +1,3 @@
+_base_ = '../repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(deploy=True))
diff --git a/configs/repvgg/repvgg-A1_4xb64-coslr-120e_in1k.py b/configs/repvgg/repvgg-A1_4xb64-coslr-120e_in1k.py
deleted file mode 100644
index 649020f2c6f..00000000000
--- a/configs/repvgg/repvgg-A1_4xb64-coslr-120e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = './repvgg-A0_4xb64-coslr-120e_in1k.py'
-
-model = dict(backbone=dict(arch='A1'))
diff --git a/configs/repvgg/repvgg-A1_8xb32_in1k.py b/configs/repvgg/repvgg-A1_8xb32_in1k.py
new file mode 100644
index 00000000000..fab5e586359
--- /dev/null
+++ b/configs/repvgg/repvgg-A1_8xb32_in1k.py
@@ -0,0 +1,3 @@
+_base_ = './repvgg-A0_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='A1'))
diff --git a/configs/repvgg/repvgg-A2_4xb64-coslr-120e_in1k.py b/configs/repvgg/repvgg-A2_8xb32_in1k.py
similarity index 58%
rename from configs/repvgg/repvgg-A2_4xb64-coslr-120e_in1k.py
rename to configs/repvgg/repvgg-A2_8xb32_in1k.py
index eedaf2d29b7..f6196f02fbf 100644
--- a/configs/repvgg/repvgg-A2_4xb64-coslr-120e_in1k.py
+++ b/configs/repvgg/repvgg-A2_8xb32_in1k.py
@@ -1,3 +1,3 @@
-_base_ = './repvgg-A0_4xb64-coslr-120e_in1k.py'
+_base_ = './repvgg-A0_8xb32_in1k.py'
model = dict(backbone=dict(arch='A2'), head=dict(in_channels=1408))
diff --git a/configs/repvgg/repvgg-B0_4xb64-coslr-120e_in1k.py b/configs/repvgg/repvgg-B0_8xb32_in1k.py
similarity index 58%
rename from configs/repvgg/repvgg-B0_4xb64-coslr-120e_in1k.py
rename to configs/repvgg/repvgg-B0_8xb32_in1k.py
index b3ce7ea27d2..9bbc4ab2259 100644
--- a/configs/repvgg/repvgg-B0_4xb64-coslr-120e_in1k.py
+++ b/configs/repvgg/repvgg-B0_8xb32_in1k.py
@@ -1,3 +1,3 @@
-_base_ = './repvgg-A0_4xb64-coslr-120e_in1k.py'
+_base_ = './repvgg-A0_8xb32_in1k.py'
model = dict(backbone=dict(arch='B0'), head=dict(in_channels=1280))
diff --git a/configs/repvgg/repvgg-B1_4xb64-coslr-120e_in1k.py b/configs/repvgg/repvgg-B1_8xb32_in1k.py
similarity index 58%
rename from configs/repvgg/repvgg-B1_4xb64-coslr-120e_in1k.py
rename to configs/repvgg/repvgg-B1_8xb32_in1k.py
index 30adea3dc8e..e08db3c4b81 100644
--- a/configs/repvgg/repvgg-B1_4xb64-coslr-120e_in1k.py
+++ b/configs/repvgg/repvgg-B1_8xb32_in1k.py
@@ -1,3 +1,3 @@
-_base_ = './repvgg-A0_4xb64-coslr-120e_in1k.py'
+_base_ = './repvgg-A0_8xb32_in1k.py'
model = dict(backbone=dict(arch='B1'), head=dict(in_channels=2048))
diff --git a/configs/repvgg/repvgg-B1g2_4xb64-coslr-120e_in1k.py b/configs/repvgg/repvgg-B1g2_8xb32_in1k.py
similarity index 59%
rename from configs/repvgg/repvgg-B1g2_4xb64-coslr-120e_in1k.py
rename to configs/repvgg/repvgg-B1g2_8xb32_in1k.py
index 2749db8d955..a1c53fded4e 100644
--- a/configs/repvgg/repvgg-B1g2_4xb64-coslr-120e_in1k.py
+++ b/configs/repvgg/repvgg-B1g2_8xb32_in1k.py
@@ -1,3 +1,3 @@
-_base_ = './repvgg-A0_4xb64-coslr-120e_in1k.py'
+_base_ = './repvgg-A0_8xb32_in1k.py'
model = dict(backbone=dict(arch='B1g2'), head=dict(in_channels=2048))
diff --git a/configs/repvgg/repvgg-B1g4_4xb64-coslr-120e_in1k.py b/configs/repvgg/repvgg-B1g4_8xb32_in1k.py
similarity index 59%
rename from configs/repvgg/repvgg-B1g4_4xb64-coslr-120e_in1k.py
rename to configs/repvgg/repvgg-B1g4_8xb32_in1k.py
index 2647690975d..0757b1e580e 100644
--- a/configs/repvgg/repvgg-B1g4_4xb64-coslr-120e_in1k.py
+++ b/configs/repvgg/repvgg-B1g4_8xb32_in1k.py
@@ -1,3 +1,3 @@
-_base_ = './repvgg-A0_4xb64-coslr-120e_in1k.py'
+_base_ = './repvgg-A0_8xb32_in1k.py'
model = dict(backbone=dict(arch='B1g4'), head=dict(in_channels=2048))
diff --git a/configs/repvgg/repvgg-B2_4xb64-coslr-120e_in1k.py b/configs/repvgg/repvgg-B2_8xb32_in1k.py
similarity index 58%
rename from configs/repvgg/repvgg-B2_4xb64-coslr-120e_in1k.py
rename to configs/repvgg/repvgg-B2_8xb32_in1k.py
index 4d215567f4d..b9a7d4ca557 100644
--- a/configs/repvgg/repvgg-B2_4xb64-coslr-120e_in1k.py
+++ b/configs/repvgg/repvgg-B2_8xb32_in1k.py
@@ -1,3 +1,3 @@
-_base_ = './repvgg-A0_4xb64-coslr-120e_in1k.py'
+_base_ = './repvgg-A0_8xb32_in1k.py'
model = dict(backbone=dict(arch='B2'), head=dict(in_channels=2560))
diff --git a/configs/repvgg/repvgg-B2g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py b/configs/repvgg/repvgg-B2g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
deleted file mode 100644
index 11331cf02f2..00000000000
--- a/configs/repvgg/repvgg-B2g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = './repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py'
-
-model = dict(backbone=dict(arch='B2g4'))
diff --git a/configs/repvgg/repvgg-B2g4_8xb32_in1k.py b/configs/repvgg/repvgg-B2g4_8xb32_in1k.py
new file mode 100644
index 00000000000..8b3397881d7
--- /dev/null
+++ b/configs/repvgg/repvgg-B2g4_8xb32_in1k.py
@@ -0,0 +1,3 @@
+_base_ = './repvgg-B3_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B2g4'), head=dict(in_channels=2560))
diff --git a/configs/repvgg/repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py b/configs/repvgg/repvgg-B3_8xb32_in1k.py
similarity index 54%
rename from configs/repvgg/repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
rename to configs/repvgg/repvgg-B3_8xb32_in1k.py
index 98bcad22da0..2d5d6e1358a 100644
--- a/configs/repvgg/repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+++ b/configs/repvgg/repvgg-B3_8xb32_in1k.py
@@ -1,10 +1,20 @@
_base_ = [
'../_base_/models/repvgg-B3_lbs-mixup_in1k.py',
- '../_base_/datasets/imagenet_bs64_pil_resize.py',
- '../_base_/schedules/imagenet_bs256_200e_coslr_warmup.py',
+ '../_base_/datasets/imagenet_bs32_pil_resize.py',
+ '../_base_/schedules/imagenet_bs256_coslr.py',
'../_base_/default_runtime.py'
]
+# schedule settings
+optim_wrapper = dict(
+ paramwise_cfg=dict(
+ bias_decay_mult=0.0,
+ custom_keys={
+ 'branch_3x3.norm': dict(decay_mult=0.0),
+ 'branch_1x1.norm': dict(decay_mult=0.0),
+ 'branch_norm.bias': dict(decay_mult=0.0),
+ }))
+
data_preprocessor = dict(
# RGB format normalization parameters
mean=[123.675, 116.28, 103.53],
@@ -21,8 +31,12 @@
dict(type='RandomResizedCrop', scale=224, backend='pillow'),
dict(type='RandomFlip', prob=0.5, direction='horizontal'),
dict(
- type='AutoAugment',
- policies='imagenet',
+ type='RandAugment',
+ policies='timm_increasing',
+ num_policies=2,
+ total_level=10,
+ magnitude_level=7,
+ magnitude_std=0.5,
hparams=dict(pad_val=[round(x) for x in bgr_mean])),
dict(type='PackClsInputs'),
]
@@ -37,3 +51,17 @@
train_dataloader = dict(dataset=dict(pipeline=train_pipeline))
val_dataloader = dict(dataset=dict(pipeline=test_pipeline))
test_dataloader = dict(dataset=dict(pipeline=test_pipeline))
+
+# schedule settings
+param_scheduler = dict(
+ type='CosineAnnealingLR',
+ T_max=200,
+ by_epoch=True,
+ begin=0,
+ end=200,
+ convert_to_iter_based=True)
+
+train_cfg = dict(by_epoch=True, max_epochs=200)
+
+default_hooks = dict(
+ checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
diff --git a/configs/repvgg/repvgg-B3g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py b/configs/repvgg/repvgg-B3g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
deleted file mode 100644
index 67e3688c5ae..00000000000
--- a/configs/repvgg/repvgg-B3g4_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = './repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py'
-
-model = dict(backbone=dict(arch='B3g4'))
diff --git a/configs/repvgg/repvgg-B3g4_8xb32_in1k.py b/configs/repvgg/repvgg-B3g4_8xb32_in1k.py
new file mode 100644
index 00000000000..b0c5c00af84
--- /dev/null
+++ b/configs/repvgg/repvgg-B3g4_8xb32_in1k.py
@@ -0,0 +1,3 @@
+_base_ = './repvgg-B3_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='B3g4'))
diff --git a/configs/repvgg/repvgg-D2se_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py b/configs/repvgg/repvgg-D2se_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
deleted file mode 100644
index d235610f07d..00000000000
--- a/configs/repvgg/repvgg-D2se_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py
+++ /dev/null
@@ -1,3 +0,0 @@
-_base_ = './repvgg-B3_4xb64-autoaug-lbs-mixup-coslr-200e_in1k.py'
-
-model = dict(backbone=dict(arch='D2se'))
diff --git a/configs/repvgg/repvgg-D2se_8xb32_in1k.py b/configs/repvgg/repvgg-D2se_8xb32_in1k.py
new file mode 100644
index 00000000000..f532dcd7968
--- /dev/null
+++ b/configs/repvgg/repvgg-D2se_8xb32_in1k.py
@@ -0,0 +1,28 @@
+_base_ = './repvgg-B3_8xb32_in1k.py'
+
+model = dict(backbone=dict(arch='D2se'), head=dict(in_channels=2560))
+
+param_scheduler = [
+ # warm up learning rate scheduler
+ dict(
+ type='LinearLR',
+ start_factor=0.0001,
+ by_epoch=True,
+ begin=0,
+ end=5,
+ # update by iter
+ convert_to_iter_based=True),
+ # main learning rate scheduler
+ dict(
+ type='CosineAnnealingLR',
+ T_max=295,
+ eta_min=1.0e-6,
+ by_epoch=True,
+ begin=5,
+ end=300)
+]
+
+train_cfg = dict(by_epoch=True, max_epochs=300)
+
+default_hooks = dict(
+ checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
diff --git a/configs/revvit/README.md b/configs/revvit/README.md
new file mode 100644
index 00000000000..4d0035ea793
--- /dev/null
+++ b/configs/revvit/README.md
@@ -0,0 +1,50 @@
+# Reversible Vision Transformers
+
+> [Reversible Vision Tranformers](https://openaccess.thecvf.com/content/CVPR2022/papers/Mangalam_Reversible_Vision_Transformers_CVPR_2022_paper.pdf)
+
+
+
+## Introduction
+
+**RevViT** is initially described in [Reversible Vision Tranformers](https://openaccess.thecvf.com/content/CVPR2022/papers/Mangalam_Reversible_Vision_Transformers_CVPR_2022_paper.pdf), which introduce the reversible idea into vision transformer, to reduce the GPU memory footprint required for training.
+
+
+
+
+
+
+
+Show the paper's abstract
+
+
+
+
+We present Reversible Vision Transformers, a memory efficient architecture design for visual recognition. By decoupling the GPU memory footprint from the depth of the model, Reversible Vision Transformers enable memory efficient scaling of transformer architectures. We adapt two popular models, namely Vision Transformer and Multiscale Vision Transformers, to reversible variants and benchmark extensively across both model sizes and tasks of image classification, object detection and video classification. Reversible Vision Transformers achieve a reduced memory footprint of up to 15.5× at identical model complexity, parameters and accuracy, demonstrating the promise of reversible vision transformers as an efficient backbone for resource limited training regimes. Finally, we find that the additional computational burden of recomputing activations is more than overcome for deeper models, where throughput can increase up to 3.9× over their non-reversible counterparts.
+
+
+
+
+## Results and models
+
+| Model | Pretrain | Params(M) | Flops(G) | Top-1 (%) | Top-5 (%) | Config | Download |
+| :--------------------------: | :----------: | :-------: | :------: | :-------: | :-------: | :-------------------------------------: | :------------------------------------------------------------------------------------: |
+| revvit-small_3rdparty_in1k\* | From scratch | 22.43 | 4.58 | 79.87 | 94.90 | [config](./revvit-small_8xb256_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/revvit/revvit-base_3rdparty_in1k_20221213-87a7b0a5.pth) |
+| revvit-base_3rdparty_in1k\* | From scratch | 87.33 | 17.49 | 81.81 | 95.56 | [config](./revvit-base_8xb256_in1k.py) | [model](https://download.openmmlab.com/mmclassification/v0/revvit/revvit-small_3rdparty_in1k_20221213-a3a34f5c.pth) |
+
+*Models with * are converted from the [official repo](https://github.com/facebookresearch/SlowFast). The config files of these models are only for inference. We don't ensure these config files' training accuracy and welcome you to contribute your reproduction results.*
+
+## Citation
+
+```bibtex
+@inproceedings{mangalam2022reversible,
+ title={Reversible Vision Transformers},
+ author={Mangalam, Karttikeya and Fan, Haoqi and Li, Yanghao and Wu, Chao-Yuan and Xiong, Bo and Feichtenhofer, Christoph and Malik, Jitendra},
+ booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
+ pages={10830--10840},
+ year={2022}
+}
+```
diff --git a/configs/revvit/metafile.yml b/configs/revvit/metafile.yml
new file mode 100644
index 00000000000..6b5b5818125
--- /dev/null
+++ b/configs/revvit/metafile.yml
@@ -0,0 +1,48 @@
+Collections:
+ - Name: RevViT
+ Metadata:
+ Training Data: ImageNet-1k
+ Architecture:
+ - Vision Transformer
+ - Reversible
+ Paper:
+ URL: https://openaccess.thecvf.com/content/CVPR2022/papers/Mangalam_Reversible_Vision_Transformers_CVPR_2022_paper.pdf
+ Title: Reversible Vision Transformers
+ README: configs/revvit/README.md
+ Code:
+ Version: v1.0.0rc5
+ URL: https://github.com/open-mmlab/mmclassification/blob/1.0.0rc5/mmcls/models/backbones/revvit.py
+
+Models:
+ - Name: revvit-small_3rdparty_in1k
+ Metadata:
+ FLOPs: 4583427072
+ Parameters: 22435432
+ In Collection: RevViT
+ Results:
+ - Dataset: ImageNet-1k
+ Metrics:
+ Top 1 Accuracy: 79.87
+ Top 5 Accuracy: 94.90
+ Task: Image Classification
+ Weights: https://download.openmmlab.com/mmclassification/v0/revvit/revvit-base_3rdparty_in1k_20221213-87a7b0a5.pth
+ Config: configs/revvit/revvit-small_8xb256_in1k.py
+ Converted From:
+ Weights: https://dl.fbaipublicfiles.com/pyslowfast/rev/REV_VIT_S.pyth
+ Code: https://github.com/facebookresearch/SlowFast
+ - Name: revvit-base_3rdparty_in1k
+ Metadata:
+ FLOPs: 17490450432
+ Parameters: 87337192
+ In Collection: RevViT
+ Results:
+ - Dataset: ImageNet-1k
+ Metrics:
+ Top 1 Accuracy: 81.81
+ Top 5 Accuracy: 95.56
+ Task: Image Classification
+ Weights: https://download.openmmlab.com/mmclassification/v0/revvit/revvit-small_3rdparty_in1k_20221213-a3a34f5c.pth
+ Config: configs/revvit/revvit-base_8xb256_in1k.py
+ Converted From:
+ Weights: https://dl.fbaipublicfiles.com/pyslowfast/rev/REV_VIT_B.pyth
+ Code: https://github.com/facebookresearch/SlowFast
diff --git a/configs/revvit/revvit-base_8xb256_in1k.py b/configs/revvit/revvit-base_8xb256_in1k.py
new file mode 100644
index 00000000000..e4fde5c9487
--- /dev/null
+++ b/configs/revvit/revvit-base_8xb256_in1k.py
@@ -0,0 +1,6 @@
+_base_ = [
+ '../_base_/models/revvit/revvit-base.py',
+ '../_base_/datasets/imagenet_bs128_revvit_224.py',
+ '../_base_/schedules/imagenet_bs1024_adamw_revvit.py',
+ '../_base_/default_runtime.py'
+]
diff --git a/configs/revvit/revvit-small_8xb256_in1k.py b/configs/revvit/revvit-small_8xb256_in1k.py
new file mode 100644
index 00000000000..ec3904a3da8
--- /dev/null
+++ b/configs/revvit/revvit-small_8xb256_in1k.py
@@ -0,0 +1,6 @@
+_base_ = [
+ '../_base_/models/revvit/revvit-small.py',
+ '../_base_/datasets/imagenet_bs128_revvit_224.py',
+ '../_base_/schedules/imagenet_bs1024_adamw_revvit.py',
+ '../_base_/default_runtime.py'
+]
diff --git a/configs/tinyvit/README.md b/configs/tinyvit/README.md
new file mode 100644
index 00000000000..4e926422409
--- /dev/null
+++ b/configs/tinyvit/README.md
@@ -0,0 +1,34 @@
+# TinyViT
+
+> [TinyViT: Fast Pretraining Distillation for Small Vision Transformers](https://arxiv.org/abs/2207.10666)
+
+
+
+## Abstract
+
+
+
+Vision transformer (ViT) recently has drawn great attention in computer vision due to its remarkable model capability. However, most prevailing ViT models suffer from huge number of parameters, restricting their applicability on devices with limited resources. To alleviate this issue, we propose TinyViT, a new family of tiny and efficient small vision transformers pretrained on large-scale datasets with our proposed fast distillation framework. The central idea is to transfer knowledge from large pretrained models to small ones, while enabling small models to get the dividends of massive pretraining data. More specifically, we apply distillation during pretraining for knowledge transfer. The logits of large teacher models are sparsified and stored in disk in advance to save the memory cost and computation overheads. The tiny student transformers are automatically scaled down from a large pretrained model with computation and parameter constraints. Comprehensive experiments demonstrate the efficacy of TinyViT. It achieves a top-1 accuracy of 84.8% on ImageNet-1k with only 21M parameters, being comparable to SwinB pretrained on ImageNet-21k while using 4.2 times fewer parameters. Moreover, increasing image resolutions, TinyViT can reach 86.5% accuracy, being slightly better than Swin-L while using only 11% parameters. Last but not the least, we demonstrate a good transfer ability of TinyViT on various downstream tasks.
+
+
+
+
+
+
+
+Show the result
+
+
+
+**LB A**
+
+
+
+**LB B**
+
+
+
+
+
+
+
+## Reproduce / 复现
+
+For detailed self-supervised pretrain code, please refer to [MMSelfSup](https://github.com/open-mmlab/mmselfsup/tree/dev-1.x/projects/fgia_accv2022_1st).
+For detailed finetuning and inference code, please refer to [this repo](https://github.com/Ezra-Yu/ACCV2022_FGIA_1st).
+
+## Description
+
+### Overview of Our Solution
+
+
+
+### Our Model
+
+- ViT(MAE-pre-train) # Pretrained from [**MMSelfSup**](https://github.com/open-mmlab/mmselfsup/tree/dev-1.x/projects/fgia_accv2022_1st).
+- Swin-v2(SimMIM-pre-train) # From [MMCls-swin_transformer_v2](https://github.com/open-mmlab/mmclassification/tree/dev-1.x/configs/swin_transformer_v2).
+
+\*\*The architectures we use \*\*
+
+- ViT + CE-loss + post-LongTail-Adjusment
+- ViT + SubCenterArcFaceWithAdvMargin(CE)
+- Swin-B + SubCenterArcFaceWithAdvMargin(SoftMax-EQL)
+- Swin-L + SubCenterArcFaceWithAdvMargin(SoftMAx-EQL)
+
+## bag of tricks paper and code
+
+- [MAE](https://github.com/open-mmlab/mmselfsup/tree/dev-1.x/configs/selfsup/mae) | [Config](https://github.com/Ezra-Yu/ACCV_workshop/tree/master/configs/vit)
+- [Swinv2](https://github.com/open-mmlab/mmclassification/tree/dev-1.x/configs/swin_transformer_v2) | [Config](https://github.com/Ezra-Yu/ACCV_workshop/tree/master/configs/swin)
+- [ArcFace](https://arxiv.org/abs/1801.07698) | [Code](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/src/models/arcface_head.py)
+- [SubCenterArcFaceWithAdvMargin](https://paperswithcode.com/paper/sub-center-arcface-boosting-face-recognition) | [Code](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/src/models/arcface_head.py)
+- [Post-LT-adjusment](https://paperswithcode.com/paper/long-tail-learning-via-logit-adjustment) | [Code](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/src/models/linear_head_lt.py)
+- [SoftMaxEQL](https://paperswithcode.com/paper/the-equalization-losses-gradient-driven) | [Code](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/src/models/eql.py)
+- FlipTTA [Code](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/src/models/tta_classifier.py)
+- clean dataset
+- self-emsemble: [Uniform-model-soup](https://arxiv.org/abs/2203.05482) | [code](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/tools/model_soup.py)
+- [pseudo](https://lilianweng.github.io/posts/2021-12-05-semi-supervised/) | [Code](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/tools/creat_pseudo.py)
+- bagging-emsemble [Code](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/tools/emsemble.py),
+- post-process: [re-distribute-label](https://github.com/Ezra-Yu/ACCV_workshop/blob/master/tools/re-distribute-label.py);
+
+
+
+
+
+#### Used but no improvements
+
+1. Using retrieval paradigm to solve this classification task;
+2. Using EfficientNetv2 backbone.
+
+#### Not used but worth to do
+
+1. Try [DiVE](https://arxiv.org/abs/2103.15042) algorithm to improve performance in long tail dataset;
+2. Use SimMIM to pre-train Swin-v2 on the competition dataset;
+3. refine the re-distribute-label tool.
diff --git a/requirements/mminstall.txt b/requirements/mminstall.txt
index 18016ab9e65..3b8103c0321 100644
--- a/requirements/mminstall.txt
+++ b/requirements/mminstall.txt
@@ -1,2 +1,2 @@
mmcv>=2.0.0rc1,<=2.0.0
-mmengine>=0.2.0,<1.0.0
+mmengine>=0.4.0,<1.0.0
diff --git a/requirements/runtime.txt b/requirements/runtime.txt
index 9f814b0eab9..e0acc881b7e 100644
--- a/requirements/runtime.txt
+++ b/requirements/runtime.txt
@@ -1,4 +1,5 @@
matplotlib
+modelindex
numpy
packaging
rich
diff --git a/tests/__init__.py b/tests/__init__.py
new file mode 100644
index 00000000000..ef101fec61e
--- /dev/null
+++ b/tests/__init__.py
@@ -0,0 +1 @@
+# Copyright (c) OpenMMLab. All rights reserved.
diff --git a/tests/data/dataset/multi-task.json b/tests/data/dataset/multi-task.json
new file mode 100644
index 00000000000..bf96384a0e6
--- /dev/null
+++ b/tests/data/dataset/multi-task.json
@@ -0,0 +1,40 @@
+{
+ "metainfo": {
+ "tasks": [
+ "gender",
+ "wear"
+ ]
+ },
+ "data_list": [
+ {
+ "img_path": "a/1.JPG",
+ "gt_label": {
+ "gender": 0
+ }
+ },
+ {
+ "img_path": "b/2.jpeg",
+ "gt_label": {
+ "gender": 0,
+ "wear": [
+ 1,
+ 0,
+ 1,
+ 0
+ ]
+ }
+ },
+ {
+ "img_path": "b/subb/3.jpg",
+ "gt_label": {
+ "gender": 1,
+ "wear": [
+ 0,
+ 1,
+ 0,
+ 1
+ ]
+ }
+ }
+ ]
+}
diff --git a/tests/test_datasets/test_datasets.py b/tests/test_datasets/test_datasets.py
index f637fb9580d..eb2fab213e5 100644
--- a/tests/test_datasets/test_datasets.py
+++ b/tests/test_datasets/test_datasets.py
@@ -79,7 +79,7 @@ def test_repr(self):
else:
self.assertIn('The `CLASSES` meta info is not set.', repr(dataset))
- self.assertIn("Haven't been initialized", repr(dataset))
+ self.assertIn('Haven\'t been initialized', repr(dataset))
dataset.full_init()
self.assertIn(f'Number of samples: \t{len(dataset)}', repr(dataset))
@@ -452,7 +452,7 @@ def test_extra_repr(self):
cfg = {**self.DEFAULT_ARGS, 'lazy_init': True}
dataset = dataset_class(**cfg)
- self.assertIn(f'Prefix of data: \t{dataset.data_prefix["root"]}',
+ self.assertIn(f"Prefix of data: \t{dataset.data_prefix['root']}",
repr(dataset))
@classmethod
@@ -597,7 +597,7 @@ def test_load_data_list(self):
}
dataset = dataset_class(**cfg)
- self.assertIn("Haven't been initialized", repr(dataset))
+ self.assertIn('Haven\'t been initialized', repr(dataset))
dataset.full_init()
self.assertIn(f'Number of samples: \t{len(dataset)}', repr(dataset))
@@ -770,7 +770,7 @@ def test_extra_repr(self):
cfg = {**self.DEFAULT_ARGS, 'lazy_init': True}
dataset = dataset_class(**cfg)
- self.assertIn(f'Prefix of data: \t{dataset.data_prefix["root"]}',
+ self.assertIn(f"Prefix of data: \t{dataset.data_prefix['root']}",
repr(dataset))
@classmethod
@@ -874,3 +874,70 @@ def test_extra_repr(self):
@classmethod
def tearDownClass(cls):
cls.tmpdir.cleanup()
+
+
+class TestMultiTaskDataset(TestCase):
+ DATASET_TYPE = 'MultiTaskDataset'
+
+ DEFAULT_ARGS = dict(
+ data_root=ASSETS_ROOT,
+ ann_file=osp.join(ASSETS_ROOT, 'multi-task.json'),
+ pipeline=[])
+
+ def test_metainfo(self):
+ dataset_class = DATASETS.get(self.DATASET_TYPE)
+
+ # Test default behavior
+ dataset = dataset_class(**self.DEFAULT_ARGS)
+ metainfo = {'tasks': ['gender', 'wear']}
+ self.assertDictEqual(dataset.metainfo, metainfo)
+ self.assertFalse(dataset.test_mode)
+
+ def test_parse_data_info(self):
+ dataset_class = DATASETS.get(self.DATASET_TYPE)
+ dataset = dataset_class(**self.DEFAULT_ARGS)
+
+ data = dataset.parse_data_info({
+ 'img_path': 'a.jpg',
+ 'gt_label': {
+ 'gender': 0
+ }
+ })
+ self.assertDictContainsSubset(
+ {
+ 'img_path': os.path.join(ASSETS_ROOT, 'a.jpg'),
+ 'gt_label': {
+ 'gender': 0
+ }
+ }, data)
+ np.testing.assert_equal(data['gt_label']['gender'], 0)
+
+ # Test missing path
+ with self.assertRaisesRegex(AssertionError, 'have `img_path` field'):
+ dataset.parse_data_info(
+ {'gt_label': {
+ 'gender': 0,
+ 'wear': [1, 0, 1, 0]
+ }})
+
+ def test_repr(self):
+ dataset_class = DATASETS.get(self.DATASET_TYPE)
+ dataset = dataset_class(**self.DEFAULT_ARGS)
+
+ task_doc = ('For 2 tasks\n gender \n wear ')
+ self.assertIn(task_doc, repr(dataset))
+
+ def test_load_data_list(self):
+ dataset_class = DATASETS.get(self.DATASET_TYPE)
+
+ # Test default behavior
+ dataset = dataset_class(**self.DEFAULT_ARGS)
+
+ data = dataset.load_data_list(self.DEFAULT_ARGS['ann_file'])
+ self.assertIsInstance(data, list)
+ np.testing.assert_equal(len(data), 3)
+ np.testing.assert_equal(data[0]['gt_label'], {'gender': 0})
+ np.testing.assert_equal(data[1]['gt_label'], {
+ 'gender': 0,
+ 'wear': [1, 0, 1, 0]
+ })
diff --git a/tests/test_datasets/test_transforms/test_formatting.py b/tests/test_datasets/test_transforms/test_formatting.py
index 0d271b3b0c6..6806b0b8060 100644
--- a/tests/test_datasets/test_transforms/test_formatting.py
+++ b/tests/test_datasets/test_transforms/test_formatting.py
@@ -10,7 +10,7 @@
from PIL import Image
from mmcls.registry import TRANSFORMS
-from mmcls.structures import ClsDataSample
+from mmcls.structures import ClsDataSample, MultiTaskDataSample
from mmcls.utils import register_all_modules
register_all_modules()
@@ -51,9 +51,8 @@ def test_transform(self):
# Test without `img` and `gt_label`
del data['img']
del data['gt_label']
- with self.assertWarnsRegex(Warning, 'Cannot get "img"'):
- results = transform(copy.deepcopy(data))
- self.assertNotIn('gt_label', results['data_samples'])
+ results = transform(copy.deepcopy(data))
+ self.assertNotIn('gt_label', results['data_samples'])
def test_repr(self):
cfg = dict(type='PackClsInputs', meta_keys=['flip', 'img_shape'])
@@ -130,3 +129,54 @@ def test_repr(self):
cfg = dict(type='Collect', keys=['img'])
transform = TRANSFORMS.build(cfg)
self.assertEqual(repr(transform), "Collect(keys=['img'])")
+
+
+class TestPackMultiTaskInputs(unittest.TestCase):
+
+ def test_transform(self):
+ img_path = osp.join(osp.dirname(__file__), '../../data/color.jpg')
+ data = {
+ 'sample_idx': 1,
+ 'img_path': img_path,
+ 'ori_shape': (300, 400),
+ 'img_shape': (300, 400),
+ 'scale_factor': 1.0,
+ 'flip': False,
+ 'img': mmcv.imread(img_path),
+ 'gt_label': {
+ 'task1': 1,
+ 'task3': 3
+ },
+ }
+
+ cfg = dict(type='PackMultiTaskInputs', )
+ transform = TRANSFORMS.build(cfg)
+ results = transform(copy.deepcopy(data))
+ self.assertIn('inputs', results)
+ self.assertIsInstance(results['inputs'], torch.Tensor)
+ self.assertIn('data_samples', results)
+ self.assertIsInstance(results['data_samples'], MultiTaskDataSample)
+ self.assertIn('flip', results['data_samples'].task1.metainfo_keys())
+ self.assertIsInstance(results['data_samples'].task1.gt_label,
+ LabelData)
+
+ # Test grayscale image
+ data['img'] = data['img'].mean(-1)
+ results = transform(copy.deepcopy(data))
+ self.assertIn('inputs', results)
+ self.assertIsInstance(results['inputs'], torch.Tensor)
+ self.assertEqual(results['inputs'].shape, (1, 300, 400))
+
+ # Test without `img` and `gt_label`
+ del data['img']
+ del data['gt_label']
+ results = transform(copy.deepcopy(data))
+ self.assertNotIn('gt_label', results['data_samples'])
+
+ def test_repr(self):
+ cfg = dict(type='PackMultiTaskInputs', meta_keys=['img_shape'])
+ transform = TRANSFORMS.build(cfg)
+ rep = 'PackMultiTaskInputs(task_handlers={},'
+ rep += ' multi_task_fields=(\'gt_label\',),'
+ rep += ' meta_keys=[\'img_shape\'])'
+ self.assertEqual(repr(transform), rep)
diff --git a/tests/test_evaluation/test_metrics/test_multi_task_metrics.py b/tests/test_evaluation/test_metrics/test_multi_task_metrics.py
new file mode 100644
index 00000000000..29e4d96d414
--- /dev/null
+++ b/tests/test_evaluation/test_metrics/test_multi_task_metrics.py
@@ -0,0 +1,134 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from unittest import TestCase
+
+import torch
+
+from mmcls.evaluation.metrics import MultiTasksMetric
+from mmcls.structures import ClsDataSample
+
+
+class MultiTaskMetric(TestCase):
+ data_pred = [
+ {
+ 'task0': torch.tensor([0.7, 0.0, 0.3]),
+ 'task1': torch.tensor([0.5, 0.2, 0.3])
+ },
+ {
+ 'task0': torch.tensor([0.0, 0.0, 1.0]),
+ 'task1': torch.tensor([0.0, 0.0, 1.0])
+ },
+ ]
+ data_gt = [{'task0': 0, 'task1': 2}, {'task1': 2}]
+
+ preds = []
+ for i, pred in enumerate(data_pred):
+ sample = {}
+ for task_name in pred:
+ task_sample = ClsDataSample().set_pred_score(pred[task_name])
+ if task_name in data_gt[i]:
+ task_sample.set_gt_label(data_gt[i][task_name])
+ task_sample.set_field(True, 'eval_mask', field_type='metainfo')
+ else:
+ task_sample.set_field(
+ False, 'eval_mask', field_type='metainfo')
+ sample[task_name] = task_sample.to_dict()
+
+ preds.append(sample)
+ data2 = zip([
+ {
+ 'task0': torch.tensor([0.7, 0.0, 0.3]),
+ 'task1': {
+ 'task10': torch.tensor([0.5, 0.2, 0.3]),
+ 'task11': torch.tensor([0.4, 0.3, 0.3])
+ }
+ },
+ {
+ 'task0': torch.tensor([0.0, 0.0, 1.0]),
+ 'task1': {
+ 'task10': torch.tensor([0.1, 0.6, 0.3]),
+ 'task11': torch.tensor([0.5, 0.2, 0.3])
+ }
+ },
+ ], [{
+ 'task0': 0,
+ 'task1': {
+ 'task10': 2,
+ 'task11': 0
+ }
+ }, {
+ 'task0': 2,
+ 'task1': {
+ 'task10': 1,
+ 'task11': 0
+ }
+ }])
+
+ pred2 = []
+ for score, label in data2:
+ sample = {}
+ for task_name in score:
+ if type(score[task_name]) != dict:
+ task_sample = ClsDataSample().set_pred_score(score[task_name])
+ task_sample.set_gt_label(label[task_name])
+ sample[task_name] = task_sample.to_dict()
+ sample[task_name]['eval_mask'] = True
+ else:
+ sample[task_name] = {}
+ sample[task_name]['eval_mask'] = True
+ for task_name2 in score[task_name]:
+ task_sample = ClsDataSample().set_pred_score(
+ score[task_name][task_name2])
+ task_sample.set_gt_label(label[task_name][task_name2])
+ sample[task_name][task_name2] = task_sample.to_dict()
+ sample[task_name][task_name2]['eval_mask'] = True
+
+ pred2.append(sample)
+
+ pred3 = [{'task0': {'eval_mask': False}, 'task1': {'eval_mask': False}}]
+ task_metrics = {
+ 'task0': [dict(type='Accuracy', topk=(1, ))],
+ 'task1': [
+ dict(type='Accuracy', topk=(1, 3)),
+ dict(type='SingleLabelMetric', items=['precision', 'recall'])
+ ]
+ }
+ task_metrics2 = {
+ 'task0': [dict(type='Accuracy', topk=(1, ))],
+ 'task1': [
+ dict(
+ type='MultiTasksMetric',
+ task_metrics={
+ 'task10': [
+ dict(type='Accuracy', topk=(1, 3)),
+ dict(type='SingleLabelMetric', items=['precision'])
+ ],
+ 'task11': [dict(type='Accuracy', topk=(1, ))]
+ })
+ ]
+ }
+
+ def test_evaluate(self):
+ """Test using the metric in the same way as Evalutor."""
+
+ # Test with score (use score instead of label if score exists)
+ metric = MultiTasksMetric(self.task_metrics)
+ metric.process(None, self.preds)
+ results = metric.evaluate(2)
+ self.assertIsInstance(results, dict)
+ self.assertAlmostEqual(results['task0_accuracy/top1'], 100)
+ self.assertGreater(results['task1_single-label/precision'], 0)
+
+ # Test nested
+ metric = MultiTasksMetric(self.task_metrics2)
+ metric.process(None, self.pred2)
+ results = metric.evaluate(2)
+ self.assertIsInstance(results, dict)
+ self.assertGreater(results['task1_task10_single-label/precision'], 0)
+ self.assertGreater(results['task1_task11_accuracy/top1'], 0)
+
+ # Test with without any ground truth value
+ metric = MultiTasksMetric(self.task_metrics)
+ metric.process(None, self.pred3)
+ results = metric.evaluate(2)
+ self.assertIsInstance(results, dict)
+ self.assertEqual(results['task0_Accuracy'], 0)
diff --git a/tests/test_models/test_backbones/test_beit.py b/tests/test_models/test_backbones/test_beit.py
index 02968bdec5f..cb9512b6b2a 100644
--- a/tests/test_models/test_backbones/test_beit.py
+++ b/tests/test_models/test_backbones/test_beit.py
@@ -75,6 +75,7 @@ def test_structure(self):
self.assertEqual(layer.attn.embed_dims, 768)
self.assertEqual(layer.attn.num_heads, 12)
self.assertEqual(layer.ffn.feedforward_channels, 3072)
+ self.assertFalse(layer.ffn.add_identity)
self.assertAlmostEqual(layer.ffn.dropout_layer.drop_prob, dpr)
dpr += dpr_inc
diff --git a/tests/test_models/test_backbones/test_efficientnet_v2.py b/tests/test_models/test_backbones/test_efficientnet_v2.py
new file mode 100644
index 00000000000..240688b099b
--- /dev/null
+++ b/tests/test_models/test_backbones/test_efficientnet_v2.py
@@ -0,0 +1,150 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import pytest
+import torch
+from torch.nn.modules import GroupNorm
+from torch.nn.modules.batchnorm import _BatchNorm
+
+from mmcls.models.backbones import EfficientNetV2
+
+
+def is_norm(modules):
+ """Check if is one of the norms."""
+ if isinstance(modules, (GroupNorm, _BatchNorm)):
+ return True
+ return False
+
+
+def check_norm_state(modules, train_state):
+ """Check if norm layer is in correct train state."""
+ for mod in modules:
+ if isinstance(mod, _BatchNorm):
+ if mod.training != train_state:
+ return False
+ return True
+
+
+def test_efficientnet_v2_backbone():
+ with pytest.raises(TypeError):
+ # pretrained must be a string path
+ model = EfficientNetV2()
+ model.init_weights(pretrained=0)
+
+ with pytest.raises(AssertionError):
+ # arch must in arc_settings
+ EfficientNetV2(arch='others')
+
+ with pytest.raises(ValueError):
+ # frozen_stages must less than 8
+ EfficientNetV2(arch='b1', frozen_stages=12)
+
+ # Test EfficientNetV2
+ model = EfficientNetV2()
+ model.init_weights()
+ model.train()
+ x = torch.rand((1, 3, 224, 224))
+ model(x)
+
+ # Test EfficientNetV2 with first stage frozen
+ frozen_stages = 7
+ model = EfficientNetV2(arch='b0', frozen_stages=frozen_stages)
+ model.init_weights()
+ model.train()
+ for i in range(frozen_stages):
+ layer = model.layers[i]
+ for mod in layer.modules():
+ if isinstance(mod, _BatchNorm):
+ assert mod.training is False
+ for param in layer.parameters():
+ assert param.requires_grad is False
+
+ # Test EfficientNetV2 with norm eval
+ model = EfficientNetV2(norm_eval=True)
+ model.init_weights()
+ model.train()
+ assert check_norm_state(model.modules(), False)
+
+ # Test EfficientNetV2 forward with 'b0' arch
+ out_channels = [32, 16, 32, 48, 96, 112, 192, 1280]
+ model = EfficientNetV2(arch='b0', out_indices=(0, 1, 2, 3, 4, 5, 6, 7))
+ model.init_weights()
+ model.train()
+
+ imgs = torch.randn(1, 3, 224, 224)
+ feat = model(imgs)
+ assert len(feat) == 8
+ assert feat[0].shape == torch.Size([1, out_channels[0], 112, 112])
+ assert feat[1].shape == torch.Size([1, out_channels[1], 112, 112])
+ assert feat[2].shape == torch.Size([1, out_channels[2], 56, 56])
+ assert feat[3].shape == torch.Size([1, out_channels[3], 28, 28])
+ assert feat[4].shape == torch.Size([1, out_channels[4], 14, 14])
+ assert feat[5].shape == torch.Size([1, out_channels[5], 14, 14])
+ assert feat[6].shape == torch.Size([1, out_channels[6], 7, 7])
+ assert feat[6].shape == torch.Size([1, out_channels[6], 7, 7])
+
+ # Test EfficientNetV2 forward with 'b0' arch and GroupNorm
+ out_channels = [32, 16, 32, 48, 96, 112, 192, 1280]
+ model = EfficientNetV2(
+ arch='b0',
+ out_indices=(0, 1, 2, 3, 4, 5, 6, 7),
+ norm_cfg=dict(type='GN', num_groups=2, requires_grad=True))
+ for m in model.modules():
+ if is_norm(m):
+ assert isinstance(m, GroupNorm)
+ model.init_weights()
+ model.train()
+
+ imgs = torch.randn(1, 3, 64, 64)
+ feat = model(imgs)
+ assert len(feat) == 8
+ assert feat[0].shape == torch.Size([1, out_channels[0], 32, 32])
+ assert feat[1].shape == torch.Size([1, out_channels[1], 32, 32])
+ assert feat[2].shape == torch.Size([1, out_channels[2], 16, 16])
+ assert feat[3].shape == torch.Size([1, out_channels[3], 8, 8])
+ assert feat[4].shape == torch.Size([1, out_channels[4], 4, 4])
+ assert feat[5].shape == torch.Size([1, out_channels[5], 4, 4])
+ assert feat[6].shape == torch.Size([1, out_channels[6], 2, 2])
+ assert feat[7].shape == torch.Size([1, out_channels[7], 2, 2])
+
+ # Test EfficientNetV2 forward with 'm' arch
+ out_channels = [24, 24, 48, 80, 160, 176, 304, 512, 1280]
+ model = EfficientNetV2(arch='m', out_indices=(0, 1, 2, 3, 4, 5, 6, 7, 8))
+ model.init_weights()
+ model.train()
+
+ imgs = torch.randn(1, 3, 64, 64)
+ feat = model(imgs)
+ assert len(feat) == 9
+ assert feat[0].shape == torch.Size([1, out_channels[0], 32, 32])
+ assert feat[1].shape == torch.Size([1, out_channels[1], 32, 32])
+ assert feat[2].shape == torch.Size([1, out_channels[2], 16, 16])
+ assert feat[3].shape == torch.Size([1, out_channels[3], 8, 8])
+ assert feat[4].shape == torch.Size([1, out_channels[4], 4, 4])
+ assert feat[5].shape == torch.Size([1, out_channels[5], 4, 4])
+ assert feat[6].shape == torch.Size([1, out_channels[6], 2, 2])
+ assert feat[7].shape == torch.Size([1, out_channels[7], 2, 2])
+ assert feat[8].shape == torch.Size([1, out_channels[8], 2, 2])
+
+ # Test EfficientNetV2 forward with 'm' arch and GroupNorm
+ out_channels = [24, 24, 48, 80, 160, 176, 304, 512, 1280]
+ model = EfficientNetV2(
+ arch='m',
+ out_indices=(0, 1, 2, 3, 4, 5, 6, 7, 8),
+ norm_cfg=dict(type='GN', num_groups=2, requires_grad=True))
+ for m in model.modules():
+ if is_norm(m):
+ assert isinstance(m, GroupNorm)
+ model.init_weights()
+ model.train()
+
+ imgs = torch.randn(1, 3, 64, 64)
+ feat = model(imgs)
+ assert len(feat) == 9
+ assert feat[0].shape == torch.Size([1, out_channels[0], 32, 32])
+ assert feat[1].shape == torch.Size([1, out_channels[1], 32, 32])
+ assert feat[2].shape == torch.Size([1, out_channels[2], 16, 16])
+ assert feat[3].shape == torch.Size([1, out_channels[3], 8, 8])
+ assert feat[4].shape == torch.Size([1, out_channels[4], 4, 4])
+ assert feat[5].shape == torch.Size([1, out_channels[5], 4, 4])
+ assert feat[6].shape == torch.Size([1, out_channels[6], 2, 2])
+ assert feat[7].shape == torch.Size([1, out_channels[7], 2, 2])
+ assert feat[8].shape == torch.Size([1, out_channels[8], 2, 2])
diff --git a/tests/test_models/test_backbones/test_mixmim.py b/tests/test_models/test_backbones/test_mixmim.py
new file mode 100644
index 00000000000..e21d143c9b5
--- /dev/null
+++ b/tests/test_models/test_backbones/test_mixmim.py
@@ -0,0 +1,40 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from copy import deepcopy
+from unittest import TestCase
+
+import torch
+
+from mmcls.models.backbones import MixMIMTransformer
+
+
+class TestMixMIM(TestCase):
+
+ def setUp(self):
+ self.cfg = dict(arch='b', drop_rate=0.0, drop_path_rate=0.1)
+
+ def test_structure(self):
+
+ # Test custom arch
+ cfg = deepcopy(self.cfg)
+
+ model = MixMIMTransformer(**cfg)
+ self.assertEqual(model.embed_dims, 128)
+ self.assertEqual(sum(model.depths), 24)
+ self.assertIsNotNone(model.absolute_pos_embed)
+
+ num_heads = [4, 8, 16, 32]
+ for i, layer in enumerate(model.layers):
+ self.assertEqual(layer.blocks[0].num_heads, num_heads[i])
+ self.assertEqual(layer.blocks[0].ffn.feedforward_channels,
+ 128 * (2**i) * 4)
+
+ def test_forward(self):
+ imgs = torch.randn(1, 3, 224, 224)
+
+ cfg = deepcopy(self.cfg)
+ model = MixMIMTransformer(**cfg)
+ outs = model(imgs)
+ self.assertIsInstance(outs, tuple)
+ self.assertEqual(len(outs), 1)
+ averaged_token = outs[-1]
+ self.assertEqual(averaged_token.shape, (1, 1024))
diff --git a/tests/test_models/test_backbones/test_revvit.py b/tests/test_models/test_backbones/test_revvit.py
new file mode 100644
index 00000000000..d50bffc11ac
--- /dev/null
+++ b/tests/test_models/test_backbones/test_revvit.py
@@ -0,0 +1,141 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import os
+import tempfile
+from copy import deepcopy
+from unittest import TestCase
+
+import torch
+from mmengine.runner import load_checkpoint, save_checkpoint
+
+from mmcls.models.backbones import RevVisionTransformer
+from .utils import timm_resize_pos_embed
+
+
+class TestRevVisionTransformer(TestCase):
+
+ def setUp(self):
+ self.cfg = dict(
+ arch='b', img_size=224, patch_size=16, drop_path_rate=0.1)
+
+ def test_structure(self):
+ # Test invalid default arch
+ with self.assertRaisesRegex(AssertionError, 'not in default archs'):
+ cfg = deepcopy(self.cfg)
+ cfg['arch'] = 'unknown'
+ RevVisionTransformer(**cfg)
+
+ # Test invalid custom arch
+ with self.assertRaisesRegex(AssertionError, 'Custom arch needs'):
+ cfg = deepcopy(self.cfg)
+ cfg['arch'] = {
+ 'num_layers': 24,
+ 'num_heads': 16,
+ 'feedforward_channels': 4096
+ }
+ RevVisionTransformer(**cfg)
+
+ # Test custom arch
+ cfg = deepcopy(self.cfg)
+ cfg['arch'] = {
+ 'embed_dims': 128,
+ 'num_layers': 24,
+ 'num_heads': 16,
+ 'feedforward_channels': 1024
+ }
+ model = RevVisionTransformer(**cfg)
+ self.assertEqual(model.embed_dims, 128)
+ self.assertEqual(model.num_layers, 24)
+ for layer in model.layers:
+ self.assertEqual(layer.attn.num_heads, 16)
+ self.assertEqual(layer.ffn.feedforward_channels, 1024)
+
+ # Test out_indices
+ # TODO: to be implemented, current only support last layer
+ cfg = deepcopy(self.cfg)
+ cfg['out_indices'] = {1: 1}
+ with self.assertRaisesRegex(AssertionError, "get "):
+ RevVisionTransformer(**cfg)
+ cfg['out_indices'] = [13]
+ with self.assertRaisesRegex(AssertionError, 'Invalid out_indices 13'):
+ RevVisionTransformer(**cfg)
+
+ # Test model structure
+ cfg = deepcopy(self.cfg)
+ model = RevVisionTransformer(**cfg)
+ self.assertEqual(len(model.layers), 12)
+ dpr_inc = 0.1 / (12 - 1)
+ dpr = 0
+ for layer in model.layers:
+ self.assertEqual(layer.attn.embed_dims, 768)
+ self.assertEqual(layer.attn.num_heads, 12)
+ self.assertEqual(layer.ffn.feedforward_channels, 3072)
+ # self.assertAlmostEqual(layer.attn.out_drop.drop_prob, dpr)
+ # self.assertAlmostEqual(layer.ffn.dropout_layer.drop_prob, dpr)
+ dpr += dpr_inc
+
+ def test_init_weights(self):
+ # test weight init cfg
+ cfg = deepcopy(self.cfg)
+ cfg['init_cfg'] = [
+ dict(
+ type='Kaiming',
+ layer='Conv2d',
+ mode='fan_in',
+ nonlinearity='linear')
+ ]
+ model = RevVisionTransformer(**cfg)
+ ori_weight = model.patch_embed.projection.weight.clone().detach()
+ # The pos_embed is all zero before initialize
+ self.assertTrue(torch.allclose(model.pos_embed, torch.tensor(0.)))
+
+ model.init_weights()
+ initialized_weight = model.patch_embed.projection.weight
+ self.assertFalse(torch.allclose(ori_weight, initialized_weight))
+ self.assertFalse(torch.allclose(model.pos_embed, torch.tensor(0.)))
+
+ # test load checkpoint
+ pretrain_pos_embed = model.pos_embed.clone().detach()
+ tmpdir = tempfile.gettempdir()
+ checkpoint = os.path.join(tmpdir, 'test.pth')
+ save_checkpoint(model.state_dict(), checkpoint)
+ cfg = deepcopy(self.cfg)
+ model = RevVisionTransformer(**cfg)
+ load_checkpoint(model, checkpoint, strict=True)
+ self.assertTrue(torch.allclose(model.pos_embed, pretrain_pos_embed))
+
+ # test load checkpoint with different img_size
+ cfg = deepcopy(self.cfg)
+ cfg['img_size'] = 384
+ model = RevVisionTransformer(**cfg)
+ load_checkpoint(model, checkpoint, strict=True)
+ resized_pos_embed = timm_resize_pos_embed(pretrain_pos_embed,
+ model.pos_embed)
+ self.assertTrue(torch.allclose(model.pos_embed, resized_pos_embed))
+
+ os.remove(checkpoint)
+
+ def test_forward(self):
+ imgs = torch.randn(1, 3, 224, 224)
+
+ cfg = deepcopy(self.cfg)
+ cfg['with_cls_token'] = False
+ cfg['output_cls_token'] = False
+ model = RevVisionTransformer(**cfg)
+ outs = model(imgs)
+ self.assertIsInstance(outs, tuple)
+ self.assertEqual(len(outs), 1)
+ patch_token = outs[-1]
+ self.assertEqual(patch_token.shape, (1, 768 * 2))
+
+ # Test forward with dynamic input size
+ imgs1 = torch.randn(1, 3, 224, 224)
+ imgs2 = torch.randn(1, 3, 256, 256)
+ imgs3 = torch.randn(1, 3, 256, 309)
+ cfg = deepcopy(self.cfg)
+ model = RevVisionTransformer(**cfg)
+ for imgs in [imgs1, imgs2, imgs3]:
+ outs = model(imgs)
+ self.assertIsInstance(outs, tuple)
+ self.assertEqual(len(outs), 1)
+ avg_token = outs[-1]
+ self.assertEqual(avg_token.shape, (1, 768 * 2))
diff --git a/tests/test_models/test_backbones/test_tinyvit.py b/tests/test_models/test_backbones/test_tinyvit.py
new file mode 100644
index 00000000000..50b0218a293
--- /dev/null
+++ b/tests/test_models/test_backbones/test_tinyvit.py
@@ -0,0 +1,80 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+# Copyright (c) OpenMMLab. All rights reserved.
+import pytest
+import torch
+
+from mmcls.models.backbones import TinyViT
+
+
+def test_assertion():
+ with pytest.raises(AssertionError):
+ TinyViT(arch='unknown')
+
+ with pytest.raises(AssertionError):
+ # MobileViT out_indices should be valid depth.
+ TinyViT(out_indices=-100)
+
+
+def test_tinyvit():
+
+ # Test forward
+ model = TinyViT(arch='5m')
+ model.init_weights()
+ model.train()
+
+ imgs = torch.randn(1, 3, 224, 224)
+ feat = model(imgs)
+ assert len(feat) == 1
+ assert feat[0].shape == torch.Size([1, 320])
+
+ # Test forward with multiple outputs
+ model = TinyViT(arch='5m', out_indices=(0, 1, 2, 3))
+ feat = model(imgs)
+ assert len(feat) == 4
+ assert feat[0].shape == torch.Size([1, 128])
+ assert feat[1].shape == torch.Size([1, 160])
+ assert feat[2].shape == torch.Size([1, 320])
+ assert feat[3].shape == torch.Size([1, 320])
+
+ # Test with custom arch
+ model = TinyViT(
+ arch={
+ 'depths': [2, 3, 4, 5],
+ 'channels': [64, 128, 256, 448],
+ 'num_heads': [4, 4, 4, 4]
+ },
+ out_indices=(0, 1, 2, 3))
+ model.init_weights()
+ model.train()
+
+ imgs = torch.randn(1, 3, 224, 224)
+ feat = model(imgs)
+ assert len(feat) == 4
+ assert feat[0].shape == torch.Size([1, 128])
+ assert feat[1].shape == torch.Size([1, 256])
+ assert feat[2].shape == torch.Size([1, 448])
+ assert feat[3].shape == torch.Size([1, 448])
+
+ # Test without gap before final norm
+ model = TinyViT(
+ arch='21m', out_indices=(0, 1, 2, 3), gap_before_final_norm=False)
+
+ imgs = torch.randn(1, 3, 224, 224)
+ feat = model(imgs)
+ assert len(feat) == 4
+
+ assert feat[0].shape == torch.Size([1, 192, 28, 28])
+ assert feat[1].shape == torch.Size([1, 384, 14, 14])
+ assert feat[2].shape == torch.Size([1, 576, 7, 7])
+ assert feat[3].shape == torch.Size([1, 576, 7, 7])
+
+ # Test frozen_stages
+ model = TinyViT(arch='11m', out_indices=(0, 1, 2, 3), frozen_stages=2)
+ model.init_weights()
+ model.train()
+
+ for i in range(2):
+ assert not model.stages[i].training
+
+ for i in range(2, 4):
+ assert model.stages[i].training
diff --git a/tests/test_models/test_backbones/test_vision_transformer.py b/tests/test_models/test_backbones/test_vision_transformer.py
index b51fe306f88..5f0fbbf5a4a 100644
--- a/tests/test_models/test_backbones/test_vision_transformer.py
+++ b/tests/test_models/test_backbones/test_vision_transformer.py
@@ -73,6 +73,12 @@ def test_structure(self):
self.assertAlmostEqual(layer.ffn.dropout_layer.drop_prob, dpr)
dpr += dpr_inc
+ # Test model structure: prenorm
+ cfg = deepcopy(self.cfg)
+ cfg['pre_norm'] = True
+ model = VisionTransformer(**cfg)
+ self.assertNotEqual(model.pre_norm.__class__, torch.nn.Identity)
+
def test_init_weights(self):
# test weight init cfg
cfg = deepcopy(self.cfg)
diff --git a/tests/test_models/test_heads.py b/tests/test_models/test_heads.py
index 0b1f72f1db8..85fdd7aa1fa 100644
--- a/tests/test_models/test_heads.py
+++ b/tests/test_models/test_heads.py
@@ -10,7 +10,7 @@
from mmengine import is_seq_of
from mmcls.registry import MODELS
-from mmcls.structures import ClsDataSample
+from mmcls.structures import ClsDataSample, MultiTaskDataSample
from mmcls.utils import register_all_modules
register_all_modules()
@@ -484,6 +484,142 @@ def test_forward(self):
head(feats)
+class TestMultiTaskHead(TestCase):
+ DEFAULT_ARGS = dict(
+ type='MultiTaskHead', # <- Head config, depends on #675
+ task_heads={
+ 'task0': dict(type='LinearClsHead', num_classes=3),
+ 'task1': dict(type='LinearClsHead', num_classes=6),
+ },
+ in_channels=10,
+ loss=dict(type='CrossEntropyLoss', loss_weight=1.0),
+ )
+
+ DEFAULT_ARGS2 = dict(
+ type='MultiTaskHead', # <- Head config, depends on #675
+ task_heads={
+ 'task0':
+ dict(
+ type='MultiTaskHead',
+ task_heads={
+ 'task00': dict(type='LinearClsHead', num_classes=3),
+ 'task01': dict(type='LinearClsHead', num_classes=6),
+ }),
+ 'task1':
+ dict(type='LinearClsHead', num_classes=6)
+ },
+ in_channels=10,
+ loss=dict(type='CrossEntropyLoss', loss_weight=1.0),
+ )
+
+ def test_forward(self):
+ head = MODELS.build(self.DEFAULT_ARGS)
+ # return the last item (same as pre_logits)
+ feats = (torch.rand(4, 10), )
+ outs = head(feats)
+ self.assertEqual(outs['task0'].shape, (4, 3))
+ self.assertEqual(outs['task1'].shape, (4, 6))
+ self.assertTrue(isinstance(outs, dict))
+
+ def test_loss(self):
+ feats = (torch.rand(4, 10), )
+ data_samples = []
+
+ for _ in range(4):
+ data_sample = MultiTaskDataSample()
+ for task_name in self.DEFAULT_ARGS['task_heads']:
+ task_sample = ClsDataSample().set_gt_label(1)
+ data_sample.set_field(task_sample, task_name)
+ data_samples.append(data_sample)
+ # with cal_acc = False
+ head = MODELS.build(self.DEFAULT_ARGS)
+
+ losses = head.loss(feats, data_samples)
+ self.assertEqual(
+ losses.keys(),
+ {'task0_loss', 'task0_mask_size', 'task1_loss', 'task1_mask_size'})
+ self.assertGreater(losses['task0_loss'].item(), 0)
+ self.assertGreater(losses['task1_loss'].item(), 0)
+
+ def test_predict(self):
+ feats = (torch.rand(4, 10), )
+ data_samples = []
+
+ for _ in range(4):
+ data_sample = MultiTaskDataSample()
+ for task_name in self.DEFAULT_ARGS['task_heads']:
+ task_sample = ClsDataSample().set_gt_label(1)
+ data_sample.set_field(task_sample, task_name)
+ data_samples.append(data_sample)
+ head = MODELS.build(self.DEFAULT_ARGS)
+ # with without data_samples
+ predictions = head.predict(feats)
+ self.assertTrue(is_seq_of(predictions, MultiTaskDataSample))
+ for pred in predictions:
+ self.assertIn('task0', pred)
+ task0_sample = predictions[0].task0
+ self.assertTrue(type(task0_sample.pred_label.score), 'torch.tensor')
+
+ # with with data_samples
+ predictions = head.predict(feats, data_samples)
+ self.assertTrue(is_seq_of(predictions, MultiTaskDataSample))
+ for sample, pred in zip(data_samples, predictions):
+ self.assertIs(sample, pred)
+ self.assertIn('task0', pred)
+
+ def test_loss_empty_data_sample(self):
+ feats = (torch.rand(4, 10), )
+ data_samples = []
+
+ for _ in range(4):
+ data_sample = MultiTaskDataSample()
+ data_samples.append(data_sample)
+ # with cal_acc = False
+ head = MODELS.build(self.DEFAULT_ARGS)
+ losses = head.loss(feats, data_samples)
+ self.assertEqual(
+ losses.keys(),
+ {'task0_loss', 'task0_mask_size', 'task1_loss', 'task1_mask_size'})
+ self.assertEqual(losses['task0_loss'].item(), 0)
+ self.assertEqual(losses['task1_loss'].item(), 0)
+
+ def test_nested_multi_task_loss(self):
+
+ head = MODELS.build(self.DEFAULT_ARGS2)
+ # return the last item (same as pre_logits)
+ feats = (torch.rand(4, 10), )
+ outs = head(feats)
+ self.assertEqual(outs['task0']['task01'].shape, (4, 6))
+ self.assertTrue(isinstance(outs, dict))
+ self.assertTrue(isinstance(outs['task0'], dict))
+
+ def test_nested_invalid_sample(self):
+ feats = (torch.rand(4, 10), )
+ gt_label = {'task0': 1, 'task1': 1}
+ head = MODELS.build(self.DEFAULT_ARGS2)
+ data_sample = MultiTaskDataSample()
+ for task_name in gt_label:
+ task_sample = ClsDataSample().set_gt_label(gt_label[task_name])
+ data_sample.set_field(task_sample, task_name)
+ with self.assertRaises(Exception):
+ head.loss(feats, data_sample)
+
+ def test_nested_invalid_sample2(self):
+ feats = (torch.rand(4, 10), )
+ gt_label = {'task0': {'task00': 1, 'task01': 1}, 'task1': 1}
+ head = MODELS.build(self.DEFAULT_ARGS)
+ data_sample = MultiTaskDataSample()
+ task_sample = ClsDataSample().set_gt_label(gt_label['task1'])
+ data_sample.set_field(task_sample, 'task1')
+ data_sample.set_field(MultiTaskDataSample(), 'task0')
+ for task_name in gt_label['task0']:
+ task_sample = ClsDataSample().set_gt_label(
+ gt_label['task0'][task_name])
+ data_sample.task0.set_field(task_sample, task_name)
+ with self.assertRaises(Exception):
+ head.loss(feats, data_sample)
+
+
class TestArcFaceClsHead(TestCase):
DEFAULT_ARGS = dict(type='ArcFaceClsHead', in_channels=10, num_classes=5)
diff --git a/tests/test_models/test_tta.py b/tests/test_models/test_tta.py
new file mode 100644
index 00000000000..2affb190a30
--- /dev/null
+++ b/tests/test_models/test_tta.py
@@ -0,0 +1,67 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+from copy import deepcopy
+from unittest import TestCase
+
+import torch
+from mmengine import ConfigDict
+
+from mmcls.models import AverageClsScoreTTA, ImageClassifier
+from mmcls.registry import MODELS
+from mmcls.structures import ClsDataSample
+from mmcls.utils import register_all_modules
+
+register_all_modules()
+
+
+class TestAverageClsScoreTTA(TestCase):
+ DEFAULT_ARGS = dict(
+ type='AverageClsScoreTTA',
+ module=dict(
+ type='ImageClassifier',
+ backbone=dict(type='ResNet', depth=18),
+ neck=dict(type='GlobalAveragePooling'),
+ head=dict(
+ type='LinearClsHead',
+ num_classes=10,
+ in_channels=512,
+ loss=dict(type='CrossEntropyLoss'))))
+
+ def test_initialize(self):
+ model: AverageClsScoreTTA = MODELS.build(self.DEFAULT_ARGS)
+ self.assertIsInstance(model.module, ImageClassifier)
+
+ def test_forward(self):
+ inputs = torch.rand(1, 3, 224, 224)
+ model: AverageClsScoreTTA = MODELS.build(self.DEFAULT_ARGS)
+
+ # The forward of TTA model should not be called.
+ with self.assertRaisesRegex(NotImplementedError, 'will not be called'):
+ model(inputs)
+
+ def test_test_step(self):
+ cfg = ConfigDict(deepcopy(self.DEFAULT_ARGS))
+ cfg.module.data_preprocessor = dict(
+ mean=[127.5, 127.5, 127.5], std=[127.5, 127.5, 127.5])
+ model: AverageClsScoreTTA = MODELS.build(cfg)
+
+ img1 = torch.randint(0, 256, (1, 3, 224, 224))
+ img2 = torch.randint(0, 256, (1, 3, 224, 224))
+ data1 = {
+ 'inputs': img1,
+ 'data_samples': [ClsDataSample().set_gt_label(1)]
+ }
+ data2 = {
+ 'inputs': img2,
+ 'data_samples': [ClsDataSample().set_gt_label(1)]
+ }
+ data_tta = {
+ 'inputs': [img1, img2],
+ 'data_samples': [[ClsDataSample().set_gt_label(1)],
+ [ClsDataSample().set_gt_label(1)]]
+ }
+
+ score1 = model.module.test_step(data1)[0].pred_label.score
+ score2 = model.module.test_step(data2)[0].pred_label.score
+ score_tta = model.test_step(data_tta)[0].pred_label.score
+
+ torch.testing.assert_allclose(score_tta, (score1 + score2) / 2)
diff --git a/tests/test_structures/test_datasample.py b/tests/test_structures/test_datasample.py
index ee45c3f24a3..e02c95fc787 100644
--- a/tests/test_structures/test_datasample.py
+++ b/tests/test_structures/test_datasample.py
@@ -5,7 +5,7 @@
import torch
from mmengine.structures import LabelData
-from mmcls.structures import ClsDataSample
+from mmcls.structures import ClsDataSample, MultiTaskDataSample
class TestClsDataSample(TestCase):
@@ -122,3 +122,20 @@ def test_set_pred_score(self):
with self.assertRaisesRegex(AssertionError, 'but got 2'):
data_sample.set_pred_score(
torch.tensor([[0.1, 0.1, 0.6, 0.1, 0.1]]))
+
+
+class TestMultiTaskDataSample(TestCase):
+
+ def test_multi_task_data_sample(self):
+ gt_label = {'task0': {'task00': 1, 'task01': 1}, 'task1': 1}
+ data_sample = MultiTaskDataSample()
+ task_sample = ClsDataSample().set_gt_label(gt_label['task1'])
+ data_sample.set_field(task_sample, 'task1')
+ data_sample.set_field(MultiTaskDataSample(), 'task0')
+ for task_name in gt_label['task0']:
+ task_sample = ClsDataSample().set_gt_label(
+ gt_label['task0'][task_name])
+ data_sample.task0.set_field(task_sample, task_name)
+ self.assertIsInstance(data_sample.task0, MultiTaskDataSample)
+ self.assertIsInstance(data_sample.task1, ClsDataSample)
+ self.assertIsInstance(data_sample.task0.task00, ClsDataSample)
diff --git a/tools/model_converters/clip_to_mmcls.py b/tools/model_converters/clip_to_mmcls.py
new file mode 100644
index 00000000000..6c179cfbf5e
--- /dev/null
+++ b/tools/model_converters/clip_to_mmcls.py
@@ -0,0 +1,74 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import argparse
+import os.path as osp
+from collections import OrderedDict
+
+import mmengine
+import torch
+from mmengine.runner import CheckpointLoader
+
+
+def convert_clip(ckpt):
+ new_ckpt = OrderedDict()
+
+ for k, v in list(ckpt.items()):
+ new_v = v
+ if k.startswith('head'):
+ new_k = k.replace('head.', 'head.layers.head.')
+ new_ckpt[new_k] = new_v
+ continue
+ elif k.startswith('patch_embed'):
+ if 'proj.' in k:
+ new_k = k.replace('proj.', 'projection.')
+ else:
+ new_k = k
+ elif k.startswith('norm_pre'):
+ new_k = k.replace('norm_pre', 'pre_norm')
+ elif k.startswith('blocks'):
+ new_k = k.replace('blocks.', 'layers.')
+ if 'norm1' in k:
+ new_k = new_k.replace('norm1', 'ln1')
+ elif 'norm2' in k:
+ new_k = new_k.replace('norm2', 'ln2')
+ elif 'mlp.fc1' in k:
+ new_k = new_k.replace('mlp.fc1', 'ffn.layers.0.0')
+ elif 'mlp.fc2' in k:
+ new_k = new_k.replace('mlp.fc2', 'ffn.layers.1')
+ elif k.startswith('norm'):
+ new_k = k.replace('norm', 'ln1')
+ else:
+ new_k = k
+
+ if not new_k.startswith('head'):
+ new_k = 'backbone.' + new_k
+ new_ckpt[new_k] = new_v
+ return new_ckpt
+
+
+'${ls /mnt/lustre/lirongjie/tmp/clip_ckpt/download_ckpt}'
+
+
+def main():
+ parser = argparse.ArgumentParser(
+ description='Convert keys in pretrained clip models to mmcls style.')
+ parser.add_argument('src', help='src model path or url')
+ # The dst path must be a full path of the new checkpoint.
+ parser.add_argument('dst', help='save path')
+ args = parser.parse_args()
+
+ checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
+
+ if 'state_dict' in checkpoint:
+ state_dict = checkpoint['state_dict']
+ else:
+ state_dict = checkpoint
+
+ weight = convert_clip(state_dict)
+ mmengine.mkdir_or_exist(osp.dirname(args.dst))
+ torch.save(weight, args.dst)
+
+ print('Done!!')
+
+
+if __name__ == '__main__':
+ main()
diff --git a/tools/model_converters/convnext_to_mmcls.py b/tools/model_converters/convnext_to_mmcls.py
new file mode 100644
index 00000000000..da24aca82ae
--- /dev/null
+++ b/tools/model_converters/convnext_to_mmcls.py
@@ -0,0 +1,62 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import argparse
+import os.path as osp
+from collections import OrderedDict
+
+import mmengine
+import torch
+from mmengine.runner import CheckpointLoader
+
+
+def convert_convnext(ckpt):
+
+ new_ckpt = OrderedDict()
+
+ for k, v in list(ckpt.items()):
+ new_v = v
+ if k.startswith('head'):
+ new_k = k.replace('head.', 'head.fc.')
+ new_ckpt[new_k] = new_v
+ continue
+ elif k.startswith('stages'):
+ if 'dwconv' in k:
+ new_k = k.replace('dwconv', 'depthwise_conv')
+ elif 'pwconv' in k:
+ new_k = k.replace('pwconv', 'pointwise_conv')
+ else:
+ new_k = k
+ elif k.startswith('norm'):
+ new_k = k.replace('norm', 'norm3')
+ else:
+ new_k = k
+
+ if not new_k.startswith('head'):
+ new_k = 'backbone.' + new_k
+ new_ckpt[new_k] = new_v
+ return new_ckpt
+
+
+def main():
+ parser = argparse.ArgumentParser(
+ description='Convert keys in pretrained van models to mmcls style.')
+ parser.add_argument('src', help='src model path or url')
+ # The dst path must be a full path of the new checkpoint.
+ parser.add_argument('dst', help='save path')
+ args = parser.parse_args()
+
+ checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
+
+ if 'model' in checkpoint:
+ state_dict = checkpoint['model']
+ else:
+ state_dict = checkpoint
+
+ weight = convert_convnext(state_dict)
+ mmengine.mkdir_or_exist(osp.dirname(args.dst))
+ torch.save(dict(state_dict=weight), args.dst)
+
+ print('Done!!')
+
+
+if __name__ == '__main__':
+ main()
diff --git a/tools/model_converters/efficientnetv2_to_mmcls.py b/tools/model_converters/efficientnetv2_to_mmcls.py
new file mode 100644
index 00000000000..b6ae4ec1c8f
--- /dev/null
+++ b/tools/model_converters/efficientnetv2_to_mmcls.py
@@ -0,0 +1,99 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+"""convert the weights of efficientnetv2 in
+timm(https://github.com/rwightman/pytorch-image-models) to mmcls format."""
+import argparse
+import os.path as osp
+
+import mmengine
+import torch
+from mmengine.runner import CheckpointLoader
+
+
+def convert_from_efficientnetv2_timm(param):
+ # main change_key
+ param_lst = list(param.keys())
+ op = str(int(param_lst[-9][7]) + 2)
+ new_key = dict()
+ for name in param_lst:
+ data = param[name]
+ if 'blocks' not in name:
+ if 'conv_stem' in name:
+ name = name.replace('conv_stem', 'backbone.layers.0.conv')
+ if 'bn1' in name:
+ name = name.replace('bn1', 'backbone.layers.0.bn')
+ if 'conv_head' in name:
+ # if efficientnet-v2_s/base/b1/b2/b3，op = 7，
+ # if for m/l/xl , op = 8
+ name = name.replace('conv_head', f'backbone.layers.{op}.conv')
+ if 'bn2' in name:
+ name = name.replace('bn2', f'backbone.layers.{op}.bn')
+ if 'classifier' in name:
+ name = name.replace('classifier', 'head.fc')
+ else:
+ operator = int(name[7])
+ if operator == 0:
+ name = name[:7] + str(operator + 1) + name[8:]
+ name = name.replace('blocks', 'backbone.layers')
+ if 'conv' in name:
+ name = name.replace('conv', 'conv')
+ if 'bn1' in name:
+ name = name.replace('bn1', 'bn')
+ elif operator < 3:
+ name = name[:7] + str(operator + 1) + name[8:]
+ name = name.replace('blocks', 'backbone.layers')
+ if 'conv_exp' in name:
+ name = name.replace('conv_exp', 'conv1.conv')
+ if 'conv_pwl' in name:
+ name = name.replace('conv_pwl', 'conv2.conv')
+ if 'bn1' in name:
+ name = name.replace('bn1', 'conv1.bn')
+ if 'bn2' in name:
+ name = name.replace('bn2', 'conv2.bn')
+ else:
+ name = name[:7] + str(operator + 1) + name[8:]
+ name = name.replace('blocks', 'backbone.layers')
+ if 'conv_pwl' in name:
+ name = name.replace('conv_pwl', 'linear_conv.conv')
+ if 'conv_pw' in name:
+ name = name.replace('conv_pw', 'expand_conv.conv')
+ if 'conv_dw' in name:
+ name = name.replace('conv_dw', 'depthwise_conv.conv')
+ if 'bn1' in name:
+ name = name.replace('bn1', 'expand_conv.bn')
+ if 'bn2' in name:
+ name = name.replace('bn2', 'depthwise_conv.bn')
+ if 'bn3' in name:
+ name = name.replace('bn3', 'linear_conv.bn')
+ if 'se.conv_reduce' in name:
+ name = name.replace('se.conv_reduce', 'se.conv1.conv')
+ if 'se.conv_expand' in name:
+ name = name.replace('se.conv_expand', 'se.conv2.conv')
+ new_key[name] = data
+ return new_key
+
+
+def main():
+ parser = argparse.ArgumentParser(
+ description='Convert pretrained efficientnetv2 '
+ 'models in timm to mmcls style.')
+ parser.add_argument('src', help='src model path or url')
+ # The dst path must be a full path of the new checkpoint.
+ parser.add_argument('dst', help='save path')
+ args = parser.parse_args()
+
+ checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
+
+ if 'state_dict' in checkpoint:
+ state_dict = checkpoint['state_dict']
+ else:
+ state_dict = checkpoint
+
+ weight = convert_from_efficientnetv2_timm(state_dict)
+ mmengine.mkdir_or_exist(osp.dirname(args.dst))
+ torch.save(weight, args.dst)
+
+ print('Done!!')
+
+
+if __name__ == '__main__':
+ main()
diff --git a/tools/model_converters/eva_to_mmcls.py b/tools/model_converters/eva_to_mmcls.py
new file mode 100644
index 00000000000..14ac24ec4f1
--- /dev/null
+++ b/tools/model_converters/eva_to_mmcls.py
@@ -0,0 +1,75 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import argparse
+import os.path as osp
+from collections import OrderedDict
+
+import mmengine
+import torch
+from mmengine.runner import CheckpointLoader
+
+
+def convert_eva(ckpt):
+
+ new_ckpt = OrderedDict()
+
+ for k, v in list(ckpt.items()):
+ if 'decoder' in k or 'mask_token' in k:
+ continue
+ new_v = v
+ if k.startswith('head'):
+ new_k = k.replace('head.', 'head.fc.')
+ new_ckpt[new_k] = new_v
+ continue
+ elif k.startswith('patch_embed'):
+ if 'proj.' in k:
+ new_k = k.replace('proj.', 'projection.')
+ else:
+ new_k = k
+ elif k.startswith('blocks'):
+ new_k = k.replace('blocks.', 'layers.')
+ if 'norm1' in k:
+ new_k = new_k.replace('norm1', 'ln1')
+ elif 'norm2' in k:
+ new_k = new_k.replace('norm2', 'ln2')
+ elif 'mlp.fc1' in k:
+ new_k = new_k.replace('mlp.fc1', 'ffn.layers.0.0')
+ elif 'mlp.fc2' in k:
+ new_k = new_k.replace('mlp.fc2', 'ffn.layers.1')
+ elif 'fc_norm' in k:
+ new_k = k.replace('fc_norm', 'ln2')
+ elif k.startswith('norm'):
+ # for mim pretrain
+ new_k = k.replace('norm', 'ln2')
+ else:
+ new_k = k
+
+ if not new_k.startswith('head'):
+ new_k = 'backbone.' + new_k
+ new_ckpt[new_k] = new_v
+ return new_ckpt
+
+
+def main():
+ parser = argparse.ArgumentParser(
+ description='Convert keys in pretrained van models to mmcls style.')
+ parser.add_argument('src', help='src model path or url')
+ # The dst path must be a full path of the new checkpoint.
+ parser.add_argument('dst', help='save path')
+ args = parser.parse_args()
+
+ checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
+
+ if 'model' in checkpoint:
+ state_dict = checkpoint['model']
+ else:
+ state_dict = checkpoint
+
+ weight = convert_eva(state_dict)
+ mmengine.mkdir_or_exist(osp.dirname(args.dst))
+ torch.save(weight, args.dst)
+
+ print('Done!!')
+
+
+if __name__ == '__main__':
+ main()
diff --git a/tools/model_converters/mixmimx_to_mmcls.py b/tools/model_converters/mixmimx_to_mmcls.py
new file mode 100644
index 00000000000..dcf9858b59b
--- /dev/null
+++ b/tools/model_converters/mixmimx_to_mmcls.py
@@ -0,0 +1,98 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import argparse
+import os.path as osp
+from collections import OrderedDict
+
+import mmengine
+import torch
+from mmengine.runner import CheckpointLoader
+
+
+def correct_unfold_reduction_order(x: torch.Tensor):
+ out_channel, in_channel = x.shape
+ x = x.reshape(out_channel, 4, in_channel // 4)
+ x = x[:, [0, 2, 1, 3], :].transpose(1, 2).reshape(out_channel, in_channel)
+ return x
+
+
+def correct_unfold_norm_order(x):
+ in_channel = x.shape[0]
+ x = x.reshape(4, in_channel // 4)
+ x = x[[0, 2, 1, 3], :].transpose(0, 1).reshape(in_channel)
+ return x
+
+
+def convert_mixmim(ckpt):
+
+ new_ckpt = OrderedDict()
+
+ for k, v in list(ckpt.items()):
+ new_v = v
+
+ if k.startswith('patch_embed'):
+ new_k = k.replace('proj', 'projection')
+
+ elif k.startswith('layers'):
+ if 'norm1' in k:
+ new_k = k.replace('norm1', 'ln1')
+ elif 'norm2' in k:
+ new_k = k.replace('norm2', 'ln2')
+ elif 'mlp.fc1' in k:
+ new_k = k.replace('mlp.fc1', 'ffn.layers.0.0')
+ elif 'mlp.fc2' in k:
+ new_k = k.replace('mlp.fc2', 'ffn.layers.1')
+ else:
+ new_k = k
+
+ elif k.startswith('norm') or k.startswith('absolute_pos_embed'):
+ new_k = k
+
+ elif k.startswith('head'):
+ new_k = k.replace('head.', 'head.fc.')
+
+ else:
+ raise ValueError
+
+ # print(new_k)
+ if not new_k.startswith('head'):
+ new_k = 'backbone.' + new_k
+
+ if 'downsample' in new_k:
+ print('Covert {} in PatchMerging from timm to mmcv format!'.format(
+ new_k))
+
+ if 'reduction' in new_k:
+ new_v = correct_unfold_reduction_order(new_v)
+ elif 'norm' in new_k:
+ new_v = correct_unfold_norm_order(new_v)
+
+ new_ckpt[new_k] = new_v
+
+ return new_ckpt
+
+
+def main():
+ parser = argparse.ArgumentParser(
+ description='Convert keys in pretrained van models to mmcls style.')
+ parser.add_argument('src', help='src model path or url')
+ # The dst path must be a full path of the new checkpoint.
+ parser.add_argument('dst', help='save path')
+ args = parser.parse_args()
+
+ checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
+
+ if 'model' in checkpoint:
+ state_dict = checkpoint['model']
+ else:
+ state_dict = checkpoint
+
+ weight = convert_mixmim(state_dict)
+ # weight = convert_official_mixmim(state_dict)
+ mmengine.mkdir_or_exist(osp.dirname(args.dst))
+ torch.save(weight, args.dst)
+
+ print('Done!!')
+
+
+if __name__ == '__main__':
+ main()
diff --git a/tools/model_converters/revvit_to_mmcls.py b/tools/model_converters/revvit_to_mmcls.py
new file mode 100644
index 00000000000..8b5f1dd5edb
--- /dev/null
+++ b/tools/model_converters/revvit_to_mmcls.py
@@ -0,0 +1,104 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import argparse
+import os.path as osp
+from collections import OrderedDict
+
+import mmengine
+import torch
+from mmengine.runner import CheckpointLoader
+
+
+def convert_revvit(ckpt):
+
+ new_ckpt = OrderedDict()
+
+ for k, v in list(ckpt.items()):
+ new_v = v
+ if k.startswith('head.projection'):
+ new_k = k.replace('head.projection', 'head.fc')
+ new_ckpt[new_k] = new_v
+ continue
+ elif k.startswith('patch_embed'):
+ if 'proj.' in k:
+ new_k = k.replace('proj.', 'projection.')
+ else:
+ new_k = k
+ elif k.startswith('rev_backbone'):
+ new_k = k.replace('rev_backbone.', '')
+ if 'F.norm' in k:
+ new_k = new_k.replace('F.norm', 'ln1')
+ elif 'G.norm' in k:
+ new_k = new_k.replace('G.norm', 'ln2')
+ elif 'F.attn' in k:
+ new_k = new_k.replace('F.attn', 'attn')
+ elif 'G.mlp.fc1' in k:
+ new_k = new_k.replace('G.mlp.fc1', 'ffn.layers.0.0')
+ elif 'G.mlp.fc2' in k:
+ new_k = new_k.replace('G.mlp.fc2', 'ffn.layers.1')
+ elif k.startswith('norm'):
+ new_k = k.replace('norm', 'ln1')
+ else:
+ new_k = k
+
+ if not new_k.startswith('head'):
+ new_k = 'backbone.' + new_k
+ new_ckpt[new_k] = new_v
+
+ tmp_weight_dir = []
+ tmp_bias_dir = []
+ final_ckpt = OrderedDict()
+ for k, v in list(new_ckpt.items()):
+ if 'attn.q.weight' in k:
+ tmp_weight_dir.append(v)
+ elif 'attn.k.weight' in k:
+ tmp_weight_dir.append(v)
+ elif 'attn.v.weight' in k:
+ tmp_weight_dir.append(v)
+ new_k = k.replace('attn.v.weight', 'attn.qkv.weight')
+ final_ckpt[new_k] = torch.cat(tmp_weight_dir, dim=0)
+ tmp_weight_dir = []
+ elif 'attn.q.bias' in k:
+ tmp_bias_dir.append(v)
+ elif 'attn.k.bias' in k:
+ tmp_bias_dir.append(v)
+ elif 'attn.v.bias' in k:
+ tmp_bias_dir.append(v)
+ new_k = k.replace('attn.v.bias', 'attn.qkv.bias')
+ final_ckpt[new_k] = torch.cat(tmp_bias_dir, dim=0)
+ tmp_bias_dir = []
+ else:
+ final_ckpt[k] = v
+
+ # add pos embed for cls token
+ if k == 'backbone.pos_embed':
+ v = torch.cat([torch.ones_like(v).mean(dim=1, keepdim=True), v],
+ dim=1)
+ final_ckpt[k] = v
+
+ return final_ckpt
+
+
+def main():
+ parser = argparse.ArgumentParser(
+ description='Convert keys in pretrained van models to mmcls style.')
+ parser.add_argument('src', help='src model path or url')
+ # The dst path must be a full path of the new checkpoint.
+ parser.add_argument('dst', help='save path')
+ args = parser.parse_args()
+
+ checkpoint = CheckpointLoader.load_checkpoint(args.src, map_location='cpu')
+
+ if 'model_state' in checkpoint:
+ state_dict = checkpoint['model_state']
+ else:
+ state_dict = checkpoint
+
+ weight = convert_revvit(state_dict)
+ mmengine.mkdir_or_exist(osp.dirname(args.dst))
+ torch.save(weight, args.dst)
+
+ print('Done!!')
+
+
+if __name__ == '__main__':
+ main()
diff --git a/tools/model_converters/tinyvit_to_mmcls.py b/tools/model_converters/tinyvit_to_mmcls.py
new file mode 100644
index 00000000000..e2791998eaf
--- /dev/null
+++ b/tools/model_converters/tinyvit_to_mmcls.py
@@ -0,0 +1,61 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import argparse
+from pathlib import Path
+
+import torch
+
+
+def convert_weights(weight):
+ """Weight Converter.
+
+ Converts the weights from timm to mmcls
+ Args:
+ weight (dict): weight dict from timm
+ Returns:
+ Converted weight dict for mmcls
+ """
+ result = dict()
+ result['meta'] = dict()
+ temp = dict()
+ mapping = {
+ 'c.weight': 'conv2d.weight',
+ 'bn.weight': 'bn2d.weight',
+ 'bn.bias': 'bn2d.bias',
+ 'bn.running_mean': 'bn2d.running_mean',
+ 'bn.running_var': 'bn2d.running_var',
+ 'bn.num_batches_tracked': 'bn2d.num_batches_tracked',
+ 'layers': 'stages',
+ 'norm_head': 'norm3',
+ }
+
+ weight = weight['model']
+
+ for k, v in weight.items():
+ # keyword mapping
+ for mk, mv in mapping.items():
+ if mk in k:
+ k = k.replace(mk, mv)
+
+ if k.startswith('head.'):
+ temp['head.fc.' + k[5:]] = v
+ else:
+ temp['backbone.' + k] = v
+
+ result['state_dict'] = temp
+ return result
+
+
+if __name__ == '__main__':
+ parser = argparse.ArgumentParser(description='Convert model keys')
+ parser.add_argument('src', help='src detectron model path')
+ parser.add_argument('dst', help='save path')
+ args = parser.parse_args()
+ dst = Path(args.dst)
+ if dst.suffix != '.pth':
+ print('The path should contain the name of the pth format file.')
+ exit(1)
+ dst.parent.mkdir(parents=True, exist_ok=True)
+
+ original_model = torch.load(args.src, map_location='cpu')
+ converted_model = convert_weights(original_model)
+ torch.save(converted_model, args.dst)
diff --git a/tools/test.py b/tools/test.py
index 0fd56510453..61004d8c187 100644
--- a/tools/test.py
+++ b/tools/test.py
@@ -56,6 +56,13 @@ def parse_args():
'--no-pin-memory',
action='store_true',
help='whether to disable the pin_memory option in dataloaders.')
+ parser.add_argument(
+ '--tta',
+ action='store_true',
+ help='Whether to enable the Test-Time-Aug (TTA). If the config file '
+ 'has `tta_pipeline` and `tta_model` fields, use them to determine the '
+ 'TTA transforms and how to merge the TTA results. Otherwise, use flip '
+ 'TTA by averaging classification score.')
parser.add_argument(
'--launcher',
choices=['none', 'pytorch', 'slurm', 'mpi'],
@@ -105,7 +112,27 @@ def merge_args(cfg, args):
else:
cfg.test_evaluator = [cfg.test_evaluator, dump_metric]
- # set dataloader args
+ # -------------------- TTA related args --------------------
+ if args.tta:
+ if 'tta_model' not in cfg:
+ cfg.tta_model = dict(type='mmcls.AverageClsScoreTTA')
+ if 'tta_pipeline' not in cfg:
+ test_pipeline = cfg.test_dataloader.dataset.pipeline
+ cfg.tta_pipeline = deepcopy(test_pipeline)
+ flip_tta = dict(
+ type='TestTimeAug',
+ transforms=[
+ [
+ dict(type='RandomFlip', prob=1.),
+ dict(type='RandomFlip', prob=0.)
+ ],
+ [test_pipeline[-1]],
+ ])
+ cfg.tta_pipeline[-1] = flip_tta
+ cfg.model = ConfigDict(**cfg.tta_model, module=cfg.model)
+ cfg.test_dataloader.dataset.pipeline = cfg.tta_pipeline
+
+ # ----------------- Default dataloader args -----------------
default_dataloader_cfg = ConfigDict(
pin_memory=True,
collate_fn=dict(type='default_collate'),
diff --git a/tools/visualizations/vis_cam.py b/tools/visualizations/vis_cam.py
index 83241cae36e..61b5fa5716a 100644
--- a/tools/visualizations/vis_cam.py
+++ b/tools/visualizations/vis_cam.py
@@ -10,6 +10,7 @@
import numpy as np
from mmcv.transforms import Compose
from mmengine.config import Config, DictAction
+from mmengine.dataset import default_collate
from mmengine.utils import to_2tuple
from torch.nn import BatchNorm1d, BatchNorm2d, GroupNorm, LayerNorm
@@ -276,7 +277,7 @@ def main():
transforms = Compose(cfg.test_dataloader.dataset.pipeline)
data = transforms({'img_path': args.img})
src_img = copy.deepcopy(data['inputs']).numpy().transpose(1, 2, 0)
- data = model.data_preprocessor(data, False)
+ data = model.data_preprocessor(default_collate([data]), False)
# build target layers
if args.target_layers:
@@ -306,7 +307,7 @@ def main():
# calculate cam grads and show|save the visualization image
grayscale_cam = cam(
- data['inputs'].unsqueeze(0),
+ data['inputs'],
targets,
eigen_smooth=args.eigen_smooth,
aug_smooth=args.aug_smooth)