pytorch-hrvvi-ext 1.4.13

HrvvI's extension to PyTorch

Overview

pytorch-hrvvi-ext is my extension to PyTorch, which contains many "out of the box" tools to facilitate my everyday study. It is very easy to use them and integrate them to your projects. I will call it hutil below because of import hutil.

Install

pip3 install -U --no-cache-dir --index-url https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple pytorch-hrvvi-ext

Hightlights

Trainer

Trainer is written on ignite, providing the following features:

• Train your network in few lines without writing loops explicitly.
• Automatic gpu support like Keras
• Metric for both CV and NLP (Loss, Accuracy, Top-K Accuracy, mAP, BLEU)
• Checkpoints of the whole trainer by epochs or metrics
• Send metric history to WeChat

Datasets

hutil contains many datasets wrapped by me providing torchvison.datasets style API. Some of them is much easier to train than VOC or COCO and more suitable for BEGINNERS in object detection. Now it contains the following datasets:

• CaptchaDetectionOnline: generate captcha image and bounding boxes of chars online
• SVHNDetection: SVHN dataset for object detection
• CocoDetection: unreleased dataset of torchvison with hutil's transforms
• VOCDetection: unreleased dataset of torchvison with hutil's transforms

Transforms

Transoforms in hutil transform inputs and targets of datasets simultaneously, which is more flexible than torchvison.transforms and makes it easier to do data augmentation for object detection with torchvision.transforms style API. The following transoforms is provided now:

• Resize
• CenterCrop
• ToPercentCoords
• Compose
• InputTransform
• TargetTransform

Others

• train_test_split: Split a dataset to a train set and a test set with different (or same) transforms
• Fullset: Transform your dataset to hutil' style dataset

Examples

CIFAR10

# Data Preparation

train_transforms = InputTransform(
    Compose([
        RandomCrop(32, padding=4),
        RandomHorizontalFlip(),
        ToTensor(),
        Normalize((0.4914, 0.4822, 0.4465), (0.247, 0.243, 0.261)),
    ])
)

test_transform = InputTransform(
    Compose([
        ToTensor(),
        Normalize((0.4914, 0.4822, 0.4465), (0.247, 0.243, 0.261)),
    ])
)

data_home = gpath("datasets/CIFAR10")
ds = CIFAR10(data_home, train=True, download=True)
ds_train, ds_val = train_test_split(
    ds, test_ratio=0.04,
    transform=train_transforms,
    test_transform=test_transform,
)
ds_test = CIFAR10(data_home, train=False, download=True)


# Define network, loss and optimizer

net = ResNet(WideSEBasicBlock, [4,4,4], k=2)
net.apply(init_weights(nonlinearity='relu'))
criterion = nn.CrossEntropyLoss()
optimizer = SGD(net.parameters(), lr=1e-1, momentum=0.9, dampening=0, weight_decay=5e-4, nesterov=True)
lr_scheduler = MultiStepLR(optimizer, [40, 80, 110], gamma=0.2)


# Define metrics

metrics = {
    'loss': Loss(),
    'acc': Accuracy(),
}

# Put it together with Trainer

trainer = Trainer(net, criterion, optimizer, lr_scheduler, metrics=metrics, save_path=gpath("models"), name="CIFAR10-SE-WRN28-2")

# Show number of parameters

summary(net, (3,32,32))

# Define batch size

train_loader = DataLoader(ds_train, batch_size=32, shuffle=True, num_workers=1, pin_memory=True)
test_loader = DataLoader(ds_test, batch_size=128)
val_loader = DataLoader(ds_val, batch_size=128)

# Train and save good models by val loss (lower is better) after first 40 epochs

trainer.fit(train_loader, 100, val_loader=val_loader, save_by_metric='-val_loss', patience=40)

CaptchaDetectionOnline

letters = "0123456789abcdefghijkmnopqrstuvwxyzABDEFGHJKMNRT"
NUM_CLASSES = len(letters) + 1
WIDTH = 128
HEIGHT = 48
LOCATIONS = [
    (8, 3),
    (4, 2),
]
ASPECT_RATIOS = [
    (1, 2, 1/2),
    (1, 2, 1/2),
]
ASPECT_RATIOS = [torch.tensor(ars) for ars in ASPECT_RATIOS]
NUM_FEATURE_MAPS = len(ASPECT_RATIOS)
SCALES = compute_scales(NUM_FEATURE_MAPS, 0.2, 0.9)
DEFAULT_BOXES = [
    compute_default_boxes(lx, ly, scale, ars)
    for (lx, ly), scale, ars in zip(LOCATIONS, SCALES, ASPECT_RATIOS)
]


# Define captcha dataset

fonts = [
    gpath("fonts/msyh.ttf"),
    gpath("fonts/sfsl0800.pfb.ttf"),
    gpath("fonts/SimHei.ttf"),
    gpath("fonts/Times New Roman.ttf"),
]

font_sizes = (28, 32, 36, 40, 44, 48)
image = ImageCaptcha(
    WIDTH, HEIGHT, fonts=fonts, font_sizes=font_sizes)

transform = Compose([
    ToPercentCoords(),
    ToTensor(),
    SSDTransform(SCALES, DEFAULT_BOXES, NUM_CLASSES),
])

test_transform = Compose([
    ToTensor(),
])

ds_train = CaptchaDetectionOnline(
    image, size=50000, letters=letters, rotate=20, transform=transform)
ds_val = CaptchaDetectionOnline(
    image, size=1000, letters=letters, rotate=20, transform=test_transform, online=False)


# Define network, loss and optimizer

out_channels = [
    (NUM_CLASSES + 4) * len(ars)
    for ars in ASPECT_RATIOS
]
net = DSOD([3, 4, 4, 4], 36, out_channels=out_channels, reduction=1)
net.apply(init_weights(nonlinearity='relu'))
criterion = SSDLoss(NUM_CLASSES)
optimizer = Adam(net.parameters(), lr=3e-4)
lr_scheduler = MultiStepLR(optimizer, [40, 70, 100], gamma=0.1)


# Define metrics for training and testing

metrics = {
    'loss': TrainLoss(),
}
test_metrics = {
    'mAP': MeanAveragePrecision(
        SSDInference(
            width=WIDTH, height=HEIGHT,
            f_default_boxes=[ cuda(d) for d in DEFAULT_BOXES ],
            num_classes=NUM_CLASSES,
        )
    )
}

# Put it together with Trainer

trainer = Trainer(net, criterion, optimizer, lr_scheduler,
                  metrics=metrics, evaluate_metrics=test_metrics,
                  save_path=gpath("models"), name="DSOD-CAPTCHA-48")

# Show numbers of parameters

summary(net, (3,HEIGHT, WIDTH))


# Define batch size

train_loader = DataLoader(
    ds_train, batch_size=32, shuffle=True, num_workers=1, pin_memory=True)
val_loader = DataLoader(
    ds_val, batch_size=32, collate_fn=box_collate_fn)

# Train and save by val mAP (higher is better) after first 10 epochs

trainer.fit(train_loader, 15, val_loader=val_loader, save_by_metric='val_mAP', patience=10)