TorchUtils

TorchUtils is a pytorch lib with several useful tools and training tricks. (Work In Progress)

Install

git clone https://github.com/seefun/TorchUtils.git
cd TorchUtils

pip install -r requirements.txt
pip install .

Import

import torch_utils as tu

Seed All

SEED = 42
tu.tools.seed_everything(SEED)

Data Augmentation

import albumentations
from albumentations import pytorch as AT
train_transform = albumentations.Compose([
    albumentations.Resize(IMAGE_SIZE, IMAGE_SIZE),
    albumentations.HorizontalFlip(p=0.5),
    tu.dataset.randAugment(image_size=IMAGE_SIZE, N=2, M=12, p=0.9, mode='all', cut_out=False),
    albumentations.Normalize(),
    albumentations.CoarseDropout(max_holes=8, max_height=IMAGE_SIZE // 8, max_width=IMAGE_SIZE // 8, fill_value=0, p=0.25),
    AT.ToTensorV2(),
    ])

mixup_dataset = tu.dataset.MixupDataset(dataset, alpha=0.2, prob=0.2, mixup_to_cutmix=0.25) 
# 0.15 mixup and 0.05 cutmix

Model

fast build models with torch_utils:

model = tu.ImageModel(name='resnest50d', pretrained=True, 
                      pooling='concat', fc='multi-dropout', 
                      num_feature=2048, classes=1)
model.cuda()

using other libs along with torch_utils:

import timm

model = timm.create_model('tresnet_m', pretrained=True)
model.global_pool = tu.layers.FastGlobalConcatPool2d(flatten=True)
model.head = tu.layers.get_attention_fc(2048*2, 1) 
model.cuda()

from pytorchcv.model_provider import get_model as ptcv_get_model

model = ptcv_get_model('seresnext50_32x4d', pretrained=True)
model.features.final_pool = tu.layers.GeM() 
model.output = tu.layers.get_simple_fc(2048, 1)   
model.cuda()

segmentation models:

hrnet = tu.get_hrnet(name='hrnet_w18', out_channel=1, pretrained=True).cuda()
unet = tu.get_unet(name='resnest50d', out_channel=1, aspp=False, pretrained=True).cuda()

recommanded pretrained models:

ResNeSt
SEResNext-50
GPU-Efficient
swsl_ResNeXt
BiT/ResNetV2
TResNet
EfficientNet_ns
ResNext_WSL
MixNet
SKNet
SGENet
HRNet
Res2Net

recommanded github repos：

model utils:

# model summary
tu.summary(model, input_size=(batch_size, 3, 224, 224))
# macs and flops
tu.profile(model, input_shape=(batch_size, 3, 224, 224))

# 3 channels pretrained weights to 1 channel
weight_rgb = model.conv1.weight.data
weight_grey = weight_rgb.sum(dim=1, keepdim=True)
model.conv1 = nn.Conv2d(1, 64, kernel_size=xxx, stride=xxx, padding=xxx, bias=False)
model.conv1.weight.data = weight_grey

# 3 channels pretrained weights to 4 channel
weight_rgb = model.conv1.weight.data
weight_y = weight_rgb.mean(dim=1, keepdim=True)
weight_rgby = torch.cat([weight_rgb,weight_y], axis=1) * 3 / 4
model.conv1 = nn.Conv2d(4, 64, kernel_size=xxx, stride=xxx, padding=xxx, bias=False)
model.conv1.weight.data = weight_rgby

# 2D models to 3d models using ACSConv (advanced)
## using code in this repo: https://github.com/M3DV/ACSConv

Optimizer

optimizer_ranger = tu.Ranger(model_conv.parameters(), lr=LR)
# optimizer = torch.optim.AdamW(model_conv.parameters(), lr=LR, weight_decay=2e-4)

Criterion

# for example:
criterion = tu.SmoothBCEwLogits(smoothing=0.02)

# criterion = tu.LabelSmoothingCrossEntropy()

Find LR

lr_finder = tu.LRFinder(model, optimizer, criterion, device="cuda")
lr_finder.range_test(train_loader, end_lr=10, num_iter=500, accumulation_steps=1)
lr_finder.plot() # to inspect the loss-learning rate graph
lr_finder.reset() # to reset the model and optimizer to their initial state

LR Scheduler

scheduler = tu.get_flat_anneal_scheduler(optimizer, max_iter, warmup_iter=0, decay_start=0.5, anneal='cos', gamma=0.05)

# scheduler = tu.CosineAnnealingWarmUpRestarts(optimizer, T_0=T, T_mult=1, eta_max=LR, T_up=0, gamma=0.05)
# torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(optimizer, T_0, T_mult=1, eta_min=0, last_epoch=-1)
# torch.optim.lr_scheduler.OneCycleLR or tu.OneCycleScheduler

AMP

Ref: https://pytorch.org/docs/master/notes/amp_examples.html

seefun/TorchUtils