standalone_cifar.py

'''
Train a long conv model on sequential CIFAR10 / sequential MNIST with PyTorch for demonstration purposes.
This code borrows heavily from https://github.com/kuangliu/pytorch-cifar and is based on https://github.com/HazyResearch/state-spaces.

* Train standard sequential CIFAR:
    python -m standalone_cifar
* Train sequential CIFAR grayscale:
    python -m standalone_cifar --grayscale

It imports LongConv from the `standalone_long_convs.py` file.

The default CIFAR10 model trained by this file should get
90+% accuracy on the CIFAR10 val set.

Each epoch takes approximately 1m05s on an A100 GPU.
'''
import torch
import torch.nn as nn
import torch.optim as optim
import torch.backends.cudnn as cudnn

import torchvision
import torchvision.transforms as transforms

import os
import argparse

from tqdm.auto import tqdm

from standalone_long_convs import LongConvModel

parser = argparse.ArgumentParser(description='PyTorch CIFAR10 Training')
# Optimizer
parser.add_argument('--lr', default=0.01, type=float, help='Learning rate')
parser.add_argument('--weight_decay', default=0.05, type=float, help='Weight decay')
# Scheduler
parser.add_argument('--epochs', default=300, type=float, help='Training epochs')
# Dataset
parser.add_argument('--grayscale', action='store_true', help='Use grayscale CIFAR10')
# Dataloader
parser.add_argument('--num_workers', default=4, type=int, help='Number of workers to use for dataloader')
parser.add_argument('--batch_size', default=50, type=int, help='Batch size')
# Model
parser.add_argument('--n_layers', default=6, type=int, help='Number of layers')
parser.add_argument('--d_model', default=256, type=int, help='Model dimension')
parser.add_argument('--dropout', default=0.1, type=float, help='Model Dropout')
parser.add_argument('--kernel_dropout', default=0.2, type=float, help='Kernel Dropout')
parser.add_argument('--kernel_lr', default=0.001, type=float, help='Kernel Learning Rate')
parser.add_argument('--kernel_lam', default=0.001, type=float, help='Kernel Squash Parameter')
parser.add_argument('--prenorm', action='store_true', help='Prenorm')
# General
parser.add_argument('--resume', '-r', action='store_true', help='Resume from checkpoint')

args = parser.parse_args()

device = 'cuda' if torch.cuda.is_available() else 'cpu'
best_acc = 0  # best test accuracy
start_epoch = 0  # start from epoch 0 or last checkpoint epoch

# Data
print(f'==> Preparing data..')

def split_train_val(train, val_split):
    train_len = int(len(train) * (1.0-val_split))
    train, val = torch.utils.data.random_split(
        train,
        (train_len, len(train) - train_len),
        generator=torch.Generator().manual_seed(42),
    )
    return train, val

if args.grayscale:
    transform = transforms.Compose([
        transforms.Grayscale(),
        transforms.ToTensor(),
        transforms.Normalize(mean=122.6 / 255.0, std=61.0 / 255.0),
        transforms.Lambda(lambda x: x.view(1, 1024).t())
    ])
else:
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
        transforms.Lambda(lambda x: x.view(3, 1024).t())
    ])

# Train with no data augmentation
transform_train = transform_test = transform

trainset = torchvision.datasets.CIFAR10(
    root='./data/cifar/', train=True, download=True, transform=transform_train)
trainset, _ = split_train_val(trainset, val_split=0.1)

valset = torchvision.datasets.CIFAR10(
    root='./data/cifar/', train=True, download=True, transform=transform_test)
_, valset = split_train_val(valset, val_split=0.1)

testset = torchvision.datasets.CIFAR10(
    root='./data/cifar/', train=False, download=True, transform=transform_test)

d_input = 3 if not args.grayscale else 1
d_output = 10

# Dataloaders
trainloader = torch.utils.data.DataLoader(
    trainset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)
valloader = torch.utils.data.DataLoader(
    valset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers)
testloader = torch.utils.data.DataLoader(
    testset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers)

# Model
print('==> Building model..')
long_conv_args = {
    'kernel_dropout': args.kernel_dropout,
    'kernel_learning_rate': args.kernel_lr,
    'kernel_lam': args.kernel_lam,
}
model = LongConvModel(
    d_input=d_input,
    d_output=d_output,
    d_model=args.d_model,
    n_layers=args.n_layers,
    dropout=args.dropout,
    prenorm=args.prenorm,
    **long_conv_args,
)

model = model.to(device)
if device == 'cuda':
    cudnn.benchmark = True

if args.resume:
    # Load checkpoint.
    print('==> Resuming from checkpoint..')
    assert os.path.isdir('checkpoint'), 'Error: no checkpoint directory found!'
    checkpoint = torch.load('./checkpoint/ckpt.pth')
    model.load_state_dict(checkpoint['model'])
    best_acc = checkpoint['acc']
    start_epoch = checkpoint['epoch']

def setup_optimizer(model, lr, weight_decay, epochs):
    """
    Following S4, train the convolution layers with a smaller learning rate, with no weight decay.

    The rest of the model can be trained with a higher learning rate (e.g. 0.004, 0.01)
    and weight decay (if desired).
    """

    # All parameters in the model
    all_parameters = list(model.parameters())

    # General parameters don't contain the special _optim key
    params = [p for p in all_parameters if not hasattr(p, "_optim")]

    # Create an optimizer with the general parameters
    optimizer = optim.AdamW(params, lr=lr, weight_decay=weight_decay)

    # Add parameters with special hyperparameters
    hps = [getattr(p, "_optim") for p in all_parameters if hasattr(p, "_optim")]
    hps = [
        dict(s) for s in sorted(list(dict.fromkeys(frozenset(hp.items()) for hp in hps)))
    ]  # Unique dicts
    for hp in hps:
        params = [p for p in all_parameters if getattr(p, "_optim", None) == hp]
        optimizer.add_param_group(
            {"params": params, **hp}
        )

    # Create a lr scheduler
    # scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, patience=patience, factor=0.2)
    scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, epochs)

    # Print optimizer info
    keys = sorted(set([k for hp in hps for k in hp.keys()]))
    for i, g in enumerate(optimizer.param_groups):
        group_hps = {k: g.get(k, None) for k in keys}
        print(' | '.join([
            f"Optimizer group {i}",
            f"{len(g['params'])} tensors",
        ] + [f"{k} {v}" for k, v in group_hps.items()]))

    return optimizer, scheduler

criterion = nn.CrossEntropyLoss()
optimizer, scheduler = setup_optimizer(
    model, lr=args.lr, weight_decay=args.weight_decay, epochs=args.epochs
)

###############################################################################
# Everything after this point is standard PyTorch training!
###############################################################################

# Training
def train():
    model.train()
    train_loss = 0
    correct = 0
    total = 0
    pbar = tqdm(enumerate(trainloader))
    for batch_idx, (inputs, targets) in pbar:
        inputs, targets = inputs.to(device), targets.to(device)
        optimizer.zero_grad()
        outputs = model(inputs)
        loss = criterion(outputs, targets)
        loss.backward()
        optimizer.step()

        train_loss += loss.item()
        _, predicted = outputs.max(1)
        total += targets.size(0)
        correct += predicted.eq(targets).sum().item()

        pbar.set_description(
            'Batch Idx: (%d/%d) | Loss: %.3f | Train Acc: %.3f%% (%d/%d)' %
            (batch_idx, len(trainloader), train_loss/(batch_idx+1), 100.*correct/total, correct, total)
        )


def eval(epoch, dataloader, checkpoint=False):
    global best_acc
    model.eval()
    eval_loss = 0
    correct = 0
    total = 0
    with torch.no_grad():
        pbar = tqdm(enumerate(dataloader))
        for batch_idx, (inputs, targets) in pbar:
            inputs, targets = inputs.to(device), targets.to(device)
            outputs = model(inputs)
            loss = criterion(outputs, targets)

            eval_loss += loss.item()
            _, predicted = outputs.max(1)
            total += targets.size(0)
            correct += predicted.eq(targets).sum().item()

            pbar.set_description(
                'Batch Idx: (%d/%d) | Loss: %.3f | Eval Acc: %.3f%% (%d/%d)' %
                (batch_idx, len(dataloader), eval_loss/(batch_idx+1), 100.*correct/total, correct, total)
            )

    # Save checkpoint.
    if checkpoint:
        acc = 100.*correct/total
        if acc > best_acc:
            state = {
                'model': model.state_dict(),
                'acc': acc,
                'epoch': epoch,
            }
            if not os.path.isdir('checkpoint'):
                os.mkdir('checkpoint')
            torch.save(state, './checkpoint/ckpt.pth')
            best_acc = acc

        return acc

pbar = tqdm(range(start_epoch, args.epochs))
for epoch in pbar:
    if epoch == 0:
        pbar.set_description('Epoch: %d' % (epoch))
    else:
        pbar.set_description('Epoch: %d | Val acc: %1.3f' % (epoch, val_acc))
    train()
    val_acc = eval(epoch, valloader, checkpoint=True)
    eval(epoch, testloader)
    scheduler.step()