vuron/adept: Платформа для обучения и вывода нейронных сетей, с API приближенным к PyTorch.

apps/python/resnet/train.py

84 строки3 KB

kolkir

Fix memroy acces in batchnorm, update resnet training script

08 мар 2025, 13:46

08 мар 2025, 13:461476fed

import argparse
import sys
from tqdm import tqdm
from pathlib import Path
import time

from resnet import ResNet18

from adept.data import CIFAR10, DataLoader
from adept import device_t, dtype_t, Variable, no_grad, Tensor, TensorProperties, Shape
from adept.optim import Adam
from adept.loss import cross_entropy_with_logits
from adept.serialize import FileInput, FileOutput

device = device_t.CPU
dtype = dtype_t.Float32

epochs = 5
batch_size = 8
lr = 0.01
num_classes = 10


def main():
    parser = argparse.ArgumentParser(
        prog="ResNet trainer", description="Script trains ResNet with CIFAR-10 dataset"
    )
    parser.add_argument("dataset_path", help="path to the CIFAR-10 root folder")
    parser.add_argument("-c", "--checkpoint", type=str, help="checkpoint file name")
    args = parser.parse_args(args=None if sys.argv[1:] else ["--help"])

    train_dataset = CIFAR10(args.dataset_path, train=True)
    test_dataset = CIFAR10(args.dataset_path, train=False)

    test_dataloader = DataLoader(test_dataset, batch_size, True)
    train_dataloader = DataLoader(train_dataset, batch_size, True)

    resnet = ResNet18(num_classes)

    optimizer = Adam(resnet.parameters(), lr)

    if args.checkpoint and Path(args.checkpoint).exists():
        input = FileInput(args.checkpoint)
        resnet.load(input)
        # TODO: load Adam params

    for epoch in tqdm(range(epochs), unit="epoch"):
        resnet.train()
        pbar = tqdm(train_dataloader, unit="batch")
        for b_i, batch in enumerate(pbar):
            x, y = batch

            out = resnet.forward(Variable(x, requires_grad=False))

            loss = cross_entropy_with_logits(
                out.squeeze(1), Variable(y, requires_grad=False).squeeze(1)
            )
            if b_i % 64 == 0:
                pbar.set_postfix(loss=loss.data().float_at([0, 0]))

            loss.backward()
            optimizer.step()
            optimizer.zero_grad()

        # save checkpoint
        checkpoint_path = f"checkpoint_{epoch}_{int(time.time()*1000.0)}.pt"
        output = FileOutput(checkpoint_path)
        resnet.save(output)
        # TODO: save Adam params

        # test
        resnet.eval()
        with no_grad():
            total_loss = Tensor.zero(TensorProperties(Shape([1, 1]), device, dtype))
            for b_i, batch in enumerate(train_dataloader):
                x, y = batch
                out = resnet.forward(Variable(x))
                loss = cross_entropy_with_logits(out.squeeze(1), Variable(y).squeeze(1))
                total_loss += loss.data()
            pbar.set_postfix(test_loss=total_loss.float_at([0, 0]) / b_i)


if __name__ == "__main__":
    main()