| 1 |
""" |
| 2 |
Fast training script for CIFAR-10 using FFCV. |
| 3 |
For tutorial, see https://docs.ffcv.io/ffcv_examples/cifar10.html. |
| 4 |
|
| 5 |
First, from the same directory, run: |
| 6 |
|
| 7 |
`python write_datasets.py --data.train_dataset [TRAIN_PATH] \ |
| 8 |
--data.val_dataset [VAL_PATH]` |
| 9 |
|
| 10 |
to generate the FFCV-formatted versions of CIFAR. |
| 11 |
|
| 12 |
Then, simply run this to train models with default hyperparameters: |
| 13 |
|
| 14 |
`python train_cifar.py --config-file default_config.yaml` |
| 15 |
|
| 16 |
You can override arguments as follows: |
| 17 |
|
| 18 |
`python train_cifar.py --config-file default_config.yaml \ |
| 19 |
--training.lr 0.2 --training.num_workers 4 ... [etc]` |
| 20 |
|
| 21 |
or by using a different config file. |
| 22 |
""" |
| 23 |
from argparse import ArgumentParser |
| 24 |
from typing import List |
| 25 |
import time |
| 26 |
import numpy as np |
| 27 |
from tqdm import tqdm |
| 28 |
|
| 29 |
import torch as ch |
| 30 |
from torch.cuda.amp import GradScaler, autocast |
| 31 |
from torch.nn import CrossEntropyLoss, Conv2d, BatchNorm2d |
| 32 |
from torch.optim import SGD, lr_scheduler |
| 33 |
import torchvision |
| 34 |
|
| 35 |
from fastargs import get_current_config, Param, Section |
| 36 |
from fastargs.decorators import param |
| 37 |
from fastargs.validation import And, OneOf |
| 38 |
|
| 39 |
from ffcv.fields import IntField, RGBImageField |
| 40 |
from ffcv.fields.decoders import IntDecoder, SimpleRGBImageDecoder |
| 41 |
from ffcv.loader import Loader, OrderOption |
| 42 |
from ffcv.pipeline.operation import Operation |
| 43 |
from ffcv.transforms import RandomHorizontalFlip, Cutout, \ |
| 44 |
RandomTranslate, Convert, ToDevice, ToTensor, ToTorchImage |
| 45 |
from ffcv.transforms.common import Squeeze |
| 46 |
from ffcv.writer import DatasetWriter |
| 47 |
|
| 48 |
Section('training', 'Hyperparameters').params( |
| 49 |
lr=Param(float, 'The learning rate to use', required=True), |
| 50 |
epochs=Param(int, 'Number of epochs to run for', required=True), |
| 51 |
lr_peak_epoch=Param(int, 'Peak epoch for cyclic lr', required=True), |
| 52 |
batch_size=Param(int, 'Batch size', default=512), |
| 53 |
momentum=Param(float, 'Momentum for SGD', default=0.9), |
| 54 |
weight_decay=Param(float, 'l2 weight decay', default=5e-4), |
| 55 |
label_smoothing=Param(float, 'Value of label smoothing', default=0.1), |
| 56 |
num_workers=Param(int, 'The number of workers', default=8), |
| 57 |
lr_tta=Param(bool, 'Test time augmentation by averaging with horizontally flipped version', default=True) |
| 58 |
) |
| 59 |
|
| 60 |
Section('data', 'data related stuff').params( |
| 61 |
train_dataset=Param(str, '.dat file to use for training', required=True), |
| 62 |
val_dataset=Param(str, '.dat file to use for validation', required=True), |
| 63 |
) |
| 64 |
|
| 65 |
@param('data.train_dataset') |
| 66 |
@param('data.val_dataset') |
| 67 |
@param('training.batch_size') |
| 68 |
@param('training.num_workers') |
| 69 |
def make_dataloaders(train_dataset=None, val_dataset=None, batch_size=None, num_workers=None): |
| 70 |
paths = { |
| 71 |
'train': train_dataset, |
| 72 |
'test': val_dataset |
| 73 |
|
| 74 |
} |
| 75 |
|
| 76 |
start_time = time.time() |
| 77 |
CIFAR_MEAN = [125.307, 122.961, 113.8575] |
| 78 |
CIFAR_STD = [51.5865, 50.847, 51.255] |
| 79 |
loaders = {} |
| 80 |
|
| 81 |
for name in ['train', 'test']: |
| 82 |
label_pipeline: List[Operation] = [IntDecoder(), ToTensor(), ToDevice(ch.device('cuda:0')), Squeeze()] |
| 83 |
image_pipeline: List[Operation] = [SimpleRGBImageDecoder()] |
| 84 |
if name == 'train': |
| 85 |
image_pipeline.extend([ |
| 86 |
RandomHorizontalFlip(), |
| 87 |
RandomTranslate(padding=2, fill=tuple(map(int, CIFAR_MEAN))), |
| 88 |
Cutout(4, tuple(map(int, CIFAR_MEAN))), |
| 89 |
]) |
| 90 |
image_pipeline.extend([ |
| 91 |
ToTensor(), |
| 92 |
ToDevice(ch.device('cuda:0'), non_blocking=True), |
| 93 |
ToTorchImage(), |
| 94 |
Convert(ch.float16), |
| 95 |
torchvision.transforms.Normalize(CIFAR_MEAN, CIFAR_STD), |
| 96 |
]) |
| 97 |
|
| 98 |
ordering = OrderOption.RANDOM if name == 'train' else OrderOption.SEQUENTIAL |
| 99 |
|
| 100 |
loaders[name] = Loader(paths[name], batch_size=batch_size, num_workers=num_workers, |
| 101 |
order=ordering, drop_last=(name == 'train'), |
| 102 |
pipelines={'image': image_pipeline, 'label': label_pipeline}) |
| 103 |
|
| 104 |
return loaders, start_time |
| 105 |
|
| 106 |
# Model (from KakaoBrain: https://github.com/wbaek/torchskeleton) |
| 107 |
class Mul(ch.nn.Module): |
| 108 |
def __init__(self, weight): |
| 109 |
super(Mul, self).__init__() |
| 110 |
self.weight = weight |
| 111 |
def forward(self, x): return x * self.weight |
| 112 |
|
| 113 |
class Flatten(ch.nn.Module): |
| 114 |
def forward(self, x): return x.view(x.size(0), -1) |
| 115 |
|
| 116 |
class Residual(ch.nn.Module): |
| 117 |
def __init__(self, module): |
| 118 |
super(Residual, self).__init__() |
| 119 |
self.module = module |
| 120 |
def forward(self, x): return x + self.module(x) |
| 121 |
|
| 122 |
def conv_bn(channels_in, channels_out, kernel_size=3, stride=1, padding=1, groups=1): |
| 123 |
return ch.nn.Sequential( |
| 124 |
ch.nn.Conv2d(channels_in, channels_out, kernel_size=kernel_size, |
| 125 |
stride=stride, padding=padding, groups=groups, bias=False), |
| 126 |
ch.nn.BatchNorm2d(channels_out), |
| 127 |
ch.nn.ReLU(inplace=True) |
| 128 |
) |
| 129 |
|
| 130 |
def construct_model(): |
| 131 |
num_class = 10 |
| 132 |
model = ch.nn.Sequential( |
| 133 |
conv_bn(3, 64, kernel_size=3, stride=1, padding=1), |
| 134 |
conv_bn(64, 128, kernel_size=5, stride=2, padding=2), |
| 135 |
Residual(ch.nn.Sequential(conv_bn(128, 128), conv_bn(128, 128))), |
| 136 |
conv_bn(128, 256, kernel_size=3, stride=1, padding=1), |
| 137 |
ch.nn.MaxPool2d(2), |
| 138 |
Residual(ch.nn.Sequential(conv_bn(256, 256), conv_bn(256, 256))), |
| 139 |
conv_bn(256, 128, kernel_size=3, stride=1, padding=0), |
| 140 |
ch.nn.AdaptiveMaxPool2d((1, 1)), |
| 141 |
Flatten(), |
| 142 |
ch.nn.Linear(128, num_class, bias=False), |
| 143 |
Mul(0.2) |
| 144 |
) |
| 145 |
model = model.to(memory_format=ch.channels_last).cuda() |
| 146 |
return model |
| 147 |
|
| 148 |
@param('training.lr') |
| 149 |
@param('training.epochs') |
| 150 |
@param('training.momentum') |
| 151 |
@param('training.weight_decay') |
| 152 |
@param('training.label_smoothing') |
| 153 |
@param('training.lr_peak_epoch') |
| 154 |
def train(model, loaders, lr=None, epochs=None, label_smoothing=None, |
| 155 |
momentum=None, weight_decay=None, lr_peak_epoch=None): |
| 156 |
opt = SGD(model.parameters(), lr=lr, momentum=momentum, weight_decay=weight_decay) |
| 157 |
iters_per_epoch = len(loaders['train']) |
| 158 |
# Cyclic LR with single triangle |
| 159 |
lr_schedule = np.interp(np.arange((epochs+1) * iters_per_epoch), |
| 160 |
[0, lr_peak_epoch * iters_per_epoch, epochs * iters_per_epoch], |
| 161 |
[0, 1, 0]) |
| 162 |
scheduler = lr_scheduler.LambdaLR(opt, lr_schedule.__getitem__) |
| 163 |
scaler = GradScaler() |
| 164 |
loss_fn = CrossEntropyLoss(label_smoothing=label_smoothing) |
| 165 |
|
| 166 |
for _ in range(epochs): |
| 167 |
for ims, labs in tqdm(loaders['train']): |
| 168 |
opt.zero_grad(set_to_none=True) |
| 169 |
with autocast(): |
| 170 |
out = model(ims) |
| 171 |
loss = loss_fn(out, labs) |
| 172 |
|
| 173 |
scaler.scale(loss).backward() |
| 174 |
scaler.step(opt) |
| 175 |
scaler.update() |
| 176 |
scheduler.step() |
| 177 |
|
| 178 |
@param('training.lr_tta') |
| 179 |
def evaluate(model, loaders, lr_tta=False): |
| 180 |
model.eval() |
| 181 |
with ch.no_grad(): |
| 182 |
for name in ['train', 'test']: |
| 183 |
total_correct, total_num = 0., 0. |
| 184 |
for ims, labs in tqdm(loaders[name]): |
| 185 |
with autocast(): |
| 186 |
out = model(ims) |
| 187 |
if lr_tta: |
| 188 |
out += model(ims.flip(-1)) |
| 189 |
total_correct += out.argmax(1).eq(labs).sum().cpu().item() |
| 190 |
total_num += ims.shape[0] |
| 191 |
print(f'{name} accuracy: {total_correct / total_num * 100:.1f}%') |
| 192 |
|
| 193 |
if __name__ == "__main__": |
| 194 |
config = get_current_config() |
| 195 |
parser = ArgumentParser(description='Fast CIFAR-10 training') |
| 196 |
config.augment_argparse(parser) |
| 197 |
# Also loads from args.config_path if provided |
| 198 |
config.collect_argparse_args(parser) |
| 199 |
config.validate(mode='stderr') |
| 200 |
config.summary() |
| 201 |
|
| 202 |
loaders, start_time = make_dataloaders() |
| 203 |
model = construct_model() |
| 204 |
train(model, loaders) |
| 205 |
print(f'Total time: {time.time() - start_time:.5f}') |
| 206 |
evaluate(model, loaders) |