run.py

#!/usr/bin/env python

import os
import json
import pprint as pp
import numpy as np

import torch
import torch.optim as optim
from tensorboard_logger import Logger as TbLogger

from options import get_options
from train import train_epoch, train_epoch_sl, validate, get_inner_model

from nets.attention_model import AttentionModel
from nets.nar_model import NARModel
from nets.critic_network import CriticNetwork
from nets.encoders.gat_encoder import GraphAttentionEncoder
from nets.encoders.gnn_encoder import GNNEncoder
from nets.encoders.mlp_encoder import MLPEncoder

from reinforce_baselines import NoBaseline, ExponentialBaseline, CriticBaseline, RolloutBaseline, WarmupBaseline

from utils import torch_load_cpu, load_problem

import warnings
warnings.filterwarnings("ignore", message="indexing with dtype torch.uint8 is now deprecated, please use a dtype torch.bool instead.")


def run(opts):
    """Top level method to run experiments for SL and RL
    """
    if opts.problem == 'tspsl':
        _run_sl(opts)
    else:
        _run_rl(opts)


def _run_rl(opts):

    # Pretty print the run args
    pp.pprint(vars(opts))

    # Set the random seed
    torch.manual_seed(opts.seed)
    np.random.seed(opts.seed)

    # Optionally configure tensorboard
    tb_logger = None
    if not opts.no_tensorboard:
        tb_logger = TbLogger(os.path.join(
            opts.log_dir, "{}_{}-{}".format(opts.problem, opts.min_size, opts.max_size), opts.run_name))

    os.makedirs(opts.save_dir)
    # Save arguments so exact configuration can always be found
    with open(os.path.join(opts.save_dir, "args.json"), 'w') as f:
        json.dump(vars(opts), f, indent=True)

    # Set the device
    opts.device = torch.device("cuda:0" if opts.use_cuda else "cpu")

    # Figure out what's the problem
    problem = load_problem(opts.problem)

    # Load data from load_path
    load_data = {}
    assert opts.load_path is None or opts.resume is None, "Only one of load path and resume can be given"
    load_path = opts.load_path if opts.load_path is not None else opts.resume
    if load_path is not None:
        print('\nLoading data from {}'.format(load_path))
        load_data = torch_load_cpu(load_path)

    # Initialize model
    model_class = {
        'attention': AttentionModel,
        'nar': NARModel,
        # 'pointer': PointerNetwork
    }.get(opts.model, None)
    assert model_class is not None, "Unknown model: {}".format(model_class)
    encoder_class = {
        'gnn': GNNEncoder,
        'gat': GraphAttentionEncoder,
        'mlp': MLPEncoder
    }.get(opts.encoder, None)
    assert encoder_class is not None, "Unknown encoder: {}".format(encoder_class)
    model = model_class(
        problem=problem,
        embedding_dim=opts.embedding_dim,
        encoder_class=encoder_class,
        n_encode_layers=opts.n_encode_layers,
        aggregation=opts.aggregation,
        aggregation_graph=opts.aggregation_graph,
        normalization=opts.normalization,
        learn_norm=opts.learn_norm,
        track_norm=opts.track_norm,
        gated=opts.gated,
        n_heads=opts.n_heads,
        tanh_clipping=opts.tanh_clipping,
        mask_inner=True,
        mask_logits=True,
        mask_graph=False,
        checkpoint_encoder=opts.checkpoint_encoder,
        shrink_size=opts.shrink_size
    ).to(opts.device)

    if opts.use_cuda and torch.cuda.device_count() > 1:
        model = torch.nn.DataParallel(model)
        
    # Compute number of network parameters
    print(model)
    nb_param = 0
    for param in model.parameters():
        nb_param += np.prod(list(param.data.size()))
    print('Number of parameters: ', nb_param)

    # Overwrite model parameters by parameters to load
    model_ = get_inner_model(model)
    model_.load_state_dict({**model_.state_dict(), **load_data.get('model', {})})

    # Initialize baseline
    if opts.baseline == 'exponential':
        baseline = ExponentialBaseline(opts.exp_beta)
    
    elif opts.baseline == 'critic' or opts.baseline == 'critic_lstm':
        assert problem.NAME == 'tsp', "Critic only supported for TSP"
        baseline = CriticBaseline(
            (
                CriticNetwork(
                    embedding_dim=opts.embedding_dim,
                    encoder_class=encoder_class,
                    n_encode_layers=opts.n_encode_layers,
                    aggregation=opts.aggregation,
                    normalization=opts.normalization,
                    learn_norm=opts.learn_norm,
                    track_norm=opts.track_norm,
                    gated=opts.gated,
                    n_heads=opts.n_heads
                )
            ).to(opts.device)
        )
        
        print(baseline.critic)
        nb_param = 0
        for param in baseline.get_learnable_parameters():
            nb_param += np.prod(list(param.data.size()))
        print('Number of parameters (BL): ', nb_param)
        
    elif opts.baseline == 'rollout':
        baseline = RolloutBaseline(model, problem, opts)
    
    else:
        assert opts.baseline is None, "Unknown baseline: {}".format(opts.baseline)
        baseline = NoBaseline()

    if opts.bl_warmup_epochs > 0:
        baseline = WarmupBaseline(baseline, opts.bl_warmup_epochs, warmup_exp_beta=opts.exp_beta)

    # Load baseline from data, make sure script is called with same type of baseline
    if 'baseline' in load_data:
        baseline.load_state_dict(load_data['baseline'])

    # Initialize optimizer
    optimizer = optim.Adam(
        [{'params': model.parameters(), 'lr': opts.lr_model}]
        + (
            [{'params': baseline.get_learnable_parameters(), 'lr': opts.lr_critic}]
            if len(baseline.get_learnable_parameters()) > 0
            else []
        )
    )

    # Load optimizer state
    if 'optimizer' in load_data:
        optimizer.load_state_dict(load_data['optimizer'])
        for state in optimizer.state.values():
            for k, v in state.items():
                if torch.is_tensor(v):
                    state[k] = v.to(opts.device)

    # Initialize learning rate scheduler, decay by lr_decay once per epoch!
    lr_scheduler = optim.lr_scheduler.LambdaLR(optimizer, lambda epoch: opts.lr_decay ** epoch)

    # Load/generate datasets
    val_datasets = []
    for val_filename in opts.val_datasets:
        val_datasets.append(
            problem.make_dataset(
                filename=val_filename, batch_size=opts.batch_size, num_samples=opts.val_size, 
                neighbors=opts.neighbors, knn_strat=opts.knn_strat, supervised=True, nar=False
            ))

    if opts.resume:
        epoch_resume = int(os.path.splitext(os.path.split(opts.resume)[-1])[0].split("-")[1])

        torch.set_rng_state(load_data['rng_state'])
        if opts.use_cuda:
            torch.cuda.set_rng_state_all(load_data['cuda_rng_state'])
        # Set the random states
        # Dumping of state was done before epoch callback, so do that now (model is loaded)
        baseline.epoch_callback(model, epoch_resume)
        print("Resuming after {}".format(epoch_resume))
        opts.epoch_start = epoch_resume + 1

    # Start training loop
    for epoch in range(opts.epoch_start, opts.epoch_start + opts.n_epochs):
        train_epoch(
            model,
            optimizer,
            baseline,
            lr_scheduler,
            epoch,
            val_datasets,
            problem,
            tb_logger,
            opts
        )


def _run_sl(opts):

    # Pretty print the run args
    pp.pprint(vars(opts))

    # Set the random seed
    torch.manual_seed(opts.seed)
    np.random.seed(opts.seed)

    # Optionally configure tensorboard
    tb_logger = None
    if not opts.no_tensorboard:
        tb_logger = TbLogger(os.path.join(
            opts.log_dir, "{}_{}-{}".format(opts.problem, opts.min_size, opts.max_size), opts.run_name))

    os.makedirs(opts.save_dir)
    # Save arguments so exact configuration can always be found
    with open(os.path.join(opts.save_dir, "args.json"), 'w') as f:
        json.dump(vars(opts), f, indent=True)

    # Set the device
    opts.device = torch.device("cuda:0" if opts.use_cuda else "cpu")

    # Figure out what's the problem
    problem = load_problem(opts.problem)
    assert opts.problem == 'tspsl', "Only TSP is supported for supervised learning"

    # Load data from load_path
    load_data = {}
    assert opts.load_path is None or opts.resume is None, "Only one of load path and resume can be given"
    load_path = opts.load_path if opts.load_path is not None else opts.resume
    if load_path is not None:
        print('\nLoading data from {}'.format(load_path))
        load_data = torch_load_cpu(load_path)

    # Initialize model
    model_class = {
        'attention': AttentionModel,
        'nar': NARModel,
        # 'pointer': PointerNetwork
    }.get(opts.model, None)
    assert model_class is not None, "Unknown model: {}".format(model_class)
    encoder_class = {
        'gnn': GNNEncoder,
        'gat': GraphAttentionEncoder,
        'mlp': MLPEncoder
    }.get(opts.encoder, None)
    assert encoder_class is not None, "Unknown encoder: {}".format(encoder_class)
    model = model_class(
        problem=problem,
        embedding_dim=opts.embedding_dim,
        encoder_class=encoder_class,
        n_encode_layers=opts.n_encode_layers,
        aggregation=opts.aggregation,
        aggregation_graph=opts.aggregation_graph,
        normalization=opts.normalization,
        learn_norm=opts.learn_norm,
        track_norm=opts.track_norm,
        gated=opts.gated,
        n_heads=opts.n_heads,
        tanh_clipping=opts.tanh_clipping,
        mask_inner=True,
        mask_logits=True,
        mask_graph=False,
        checkpoint_encoder=opts.checkpoint_encoder,
        shrink_size=opts.shrink_size
    ).to(opts.device)

    if opts.use_cuda and torch.cuda.device_count() > 1:
        model = torch.nn.DataParallel(model)
        
    # Compute number of network parameters
    print(model)
    nb_param = 0
    for param in model.parameters():
        nb_param += np.prod(list(param.data.size()))
    print('Number of parameters: ', nb_param)

    # Overwrite model parameters by parameters to load
    model_ = get_inner_model(model)
    model_.load_state_dict({**model_.state_dict(), **load_data.get('model', {})})

    # Initialize optimizer
    optimizer = optim.Adam([{'params': model.parameters(), 'lr': opts.lr_model}])

    # Load optimizer state
    if 'optimizer' in load_data:
        optimizer.load_state_dict(load_data['optimizer'])
        for state in optimizer.state.values():
            for k, v in state.items():
                if torch.is_tensor(v):
                    state[k] = v.to(opts.device)

    # Initialize learning rate scheduler, decay by lr_decay once per epoch!
    lr_scheduler = optim.lr_scheduler.LambdaLR(optimizer, lambda epoch: opts.lr_decay ** epoch)

    # Load/generate datasets
    train_dataset = problem.make_dataset(
        filename=opts.train_dataset, batch_size=opts.batch_size, num_samples=opts.epoch_size, 
        neighbors=opts.neighbors, knn_strat=opts.knn_strat, supervised=True, nar=(opts.model == 'nar')
    )
    opts.epoch_size = train_dataset.size  # Training set size might be different from specified epoch size
    val_datasets = []
    for val_filename in opts.val_datasets:
        val_datasets.append(
            problem.make_dataset(
                filename=val_filename, batch_size=opts.batch_size, num_samples=opts.val_size, 
                neighbors=opts.neighbors, knn_strat=opts.knn_strat, supervised=True, nar=False
            ))

    if opts.resume:
        epoch_resume = int(os.path.splitext(os.path.split(opts.resume)[-1])[0].split("-")[1])

        torch.set_rng_state(load_data['rng_state'])
        if opts.use_cuda:
            torch.cuda.set_rng_state_all(load_data['cuda_rng_state'])
        # Set the random states
        print("Resuming after {}".format(epoch_resume))
        opts.epoch_start = epoch_resume + 1

    # Start training loop
    for epoch in range(opts.epoch_start, opts.epoch_start + opts.n_epochs):
        train_epoch_sl(
            model,
            optimizer,
            lr_scheduler,
            epoch,
            train_dataset,
            val_datasets,
            problem,
            tb_logger,
            opts
        )


if __name__ == "__main__":
    run(get_options())