Trying to generate custom avalanche dataset

[caesar753]

Hello,
I'm trying to generate an Avalanche dataset from my image folder on Colab but when I run avalanche.benchmarks.generators.dataset_benchmark command I get this error

ValueError: Unsupported dataset: must have a valid targets field or has to be a Tensor Dataset with at least 2 Tensors

I found on the forum this discussion https://github.com/ContinualAI/avalanche/discussions/475 but I don't know how to add targets...

this is my code

from torch.optim import SGD
from torch.nn import CrossEntropyLoss
from torch.utils.data import DataLoader
from torch.utils.data.dataset import TensorDataset

import torchvision.utils
import torchvision.datasets as dsets
import torchvision.transforms as transforms

from avalanche.benchmarks.classic import SplitMNIST
from avalanche.evaluation.metrics import forgetting_metrics, accuracy_metrics, \
    loss_metrics, timing_metrics, cpu_usage_metrics, confusion_matrix_metrics, disk_usage_metrics
from avalanche.models import SimpleMLP
from avalanche.logging import InteractiveLogger, TextLogger, TensorboardLogger
from avalanche.training.plugins import EvaluationPlugin
from avalanche.training.supervised import Naive
import avalanche.benchmarks.generators as Generator
from avalanche.benchmarks.utils import AvalancheDataset

import numpy as np
import matplotlib.pyplot as plt


from google.colab import drive
drive.mount('/content/drive')
train_transform = transforms.Compose([
    transforms.RandomResizedCrop(299),
    transforms.RandomHorizontalFlip(),
    transforms.ToTensor(), # ToTensor : [0, 255] -> [0, 1]
    transforms.Normalize(mean = [0.485, 0.456, 0.406],
                         std = [0.229, 0.224, 0.225])
])

#defining function that transforms images for testing/validation (resizing), and normalize
test_transform = transforms.Compose([
    transforms.Resize((299, 299)),
    transforms.ToTensor(), # ToTensor : [0, 255] -> [0, 1]
    transforms.Normalize(mean = [0.485, 0.456, 0.406],
                         std = [0.229, 0.224, 0.225])
])


train_data = dsets.ImageFolder('/content/drive/MyDrive/myfolder_train', train_transform)
test_data = dsets.ImageFolder('/content/drive/MyDrive/my_folder_test', test_transform)

#dataloade for training
train_loader = DataLoader(train_data,
                          batch_size=5,
                          shuffle=True,
                          drop_last=True)

#dataloader for testing
test_loader = DataLoader(test_data, 
                         batch_size=5,
                         shuffle=True)

#generating an iterator
dataiter = iter(train_loader)
dataiter_test = iter(test_loader)

#calling next image and label with iterator
images, labels = next(dataiter)
images_test, labels_test = next(dataiter_test)

torch_dsets = TensorDataset(images, labels)
torch_dsets_test = TensorDataset(images_test, labels_test)

avalanche_dataset = AvalancheDataset(torch_dsets)
avalanche_dataset_test = AvalancheDataset(torch_dsets_test)
Generator.dataset_benchmark(torch_dsets, torch_dsets_test)

[vlomonaco]

Hi @caesar753 can you test it with avalanche v0.3.1 we just released?

[caesar753]

Hello,
I tried to updata the library, now I'm using 0.3.1 but the error is the same: i'm able to generate the TensorDataset and to transform it into an AvalancheDataset but when i'm trying to generate the dataset_bechmark it gives me the same error:

ValueError: Unsupported dataset: must have a valid targets field or has to be a Tensor Dataset with at least 2 Tensors

Thanks for your support!

[AntonioCarta]

In your case it's easier if you use make_tensor_classification_dataset instead of the AvalancheDataset.

[caesar753]

thanks for your replay, I used make_tensor_classification_dataset, the tensor of the dataset is now quite different but the error remains the same as above.

This is a sample of this tensor

from avalanche.benchmarks.utils import make_tensor_classification_dataset as MakeTensor
tensor_dset_labels = torch.tensor(labels)
tensor_dset_train = torch.tensor(images)
tensor_dset_eval = torch.tensor(images_eval)

maketensor_train = MakeTensor(tensor_dset_train, tensor_dset_labels)
maketensor_eval = MakeTensor(tensor_dset_eval, tensor_dset_labels)



print(maketensor_train[2])

[tensor([[[-1.2103, -1.1760, -1.1760,  ..., -1.0390, -1.0904, -1.0390],
         [-1.2103, -1.1932, -1.1760,  ..., -1.0048, -1.0219, -1.0219],
         [-1.1932, -1.1932, -1.1589,  ..., -1.0219, -1.0390, -1.0562],
         ...,
         [-1.8610, -1.8610, -1.8782,  ..., -1.5528, -1.5699, -1.5699],
         [-1.8610, -1.8610, -1.8953,  ..., -1.5528, -1.5528, -1.5699],
         [-1.8782, -1.8782, -1.8782,  ..., -1.5870, -1.5699, -1.5699]],

        [[ 0.1352,  0.1702,  0.1702,  ...,  0.4678,  0.4153,  0.4678],
         [ 0.1352,  0.1527,  0.1527,  ...,  0.4853,  0.4503,  0.4503],
         [ 0.1527,  0.1527,  0.1702,  ...,  0.4853,  0.4678,  0.4503],
         ...,
         [-1.3004, -1.3004, -1.3179,  ..., -0.6527, -0.6527, -0.6702],
         [-1.3004, -1.3004, -1.3004,  ..., -0.6352, -0.6352, -0.6527],
         [-1.3004, -1.3004, -1.3004,  ..., -0.6702, -0.6702, -0.6527]],

        [[ 1.4548,  1.4897,  1.5071,  ...,  1.7511,  1.7163,  1.7511],
         [ 1.4722,  1.4897,  1.5071,  ...,  1.7685,  1.7685,  1.7511],
         [ 1.4722,  1.4722,  1.4897,  ...,  1.7685,  1.7685,  1.7685],
         ...,
         [ 0.2522,  0.2696,  0.2348,  ...,  1.1237,  1.1062,  1.1062],
         [ 0.2696,  0.2696,  0.2348,  ...,  1.1237,  1.1237,  1.1062],
         [ 0.2522,  0.2522,  0.2173,  ...,  1.0888,  1.0888,  1.1062]]]), 2, 0]

and the error is, as before:

import avalanche.benchmarks.generators as Generator
Generator.dataset_benchmark(maketensor_train, maketensor_eval)

---------------------------------------------------------------------------

ValueError                                Traceback (most recent call last)

[<ipython-input-8-17554ccbc977>](https://localhost:8080/#) in <module>
----> 1 Generator.dataset_benchmark(maketensor_train, maketensor_eval)

5 frames

[/usr/local/lib/python3.8/dist-packages/avalanche/benchmarks/utils/classification_dataset.py](https://localhost:8080/#) in _select_targets(dataset, indices)
    797         found_targets = dataset.tensors[1]
    798     else:
--> 799         raise ValueError(
    800             "Unsupported dataset: must have a valid targets field "
    801             "or has to be a Tensor Dataset with at least 2 "

ValueError: Unsupported dataset: must have a valid targets field or has to be a Tensor Dataset with at least 2 Tensors

[AntonioCarta]

dataset_benchmark takes a list of datasets, not the datasets. This:

Generator.dataset_benchmark([maketensor_train], [maketensor_eval])

should create two parallel streams with exactly one experience each.

[caesar753]

Perfect! Now it works !

Thanks a lot for your support :-D