概述

PyTorchImageModels，简称 timm，是一个巨大的 PyTorch 代码集合，包括了一系列：

• image models
• layers
• utilities
• optimizers
• schedulers
• data-loaders / augmentations
• training / validation scripts

旨在将各种 SOTA 模型整合在一起，并具有复现 ImageNet 训练结果的能力。虽然模型架构是 timm 的重点，但它还包括许多数据增强 (data augmentations)、正则化技术 (regularization techniques)、优化器 (optimizers) 和学习率策略 (learning rate schedulers) 的实现。

作者github链接：rwightman - Overview (opens new window) timm库链接：rwightman/pytorch-image-models (opens new window)

timm库特点

所有的模型都有默认的API

• accessing/changing the classifier - get_classifier and reset_classifier
• 只对features做前向传播 - forward_features

所有模型都支持多尺度特征提取 (feature pyramids) (通过create_model函数)：

• create_model(name, features_only=True, out_indices=..., output_stride=...)

out_indices 指定返回哪个feature maps to return, 从0开始，out_indices[i]对应着 C(i + 1) feature level。

output_stride 通过dilated convolutions控制网络的output stride。大多数网络默认 stride 32 。

所有的模型都有一致的pretrained weight loader，adapts last linear if necessary。

训练方式支持

• NVIDIA DDP w/ a single GPU per process, multiple processes with APEX present (AMP mixed-precision optional)

• PyTorch DistributedDataParallel w/ multi-gpu, single process (AMP disabled as it crashes when enabled)

• PyTorch w/ single GPU single process (AMP optional)

动态的全局池化方式可以选择：average pooling, max pooling, average + max, or concat([average, max])，默认是adaptive average。

Schedulers：

Schedulers 包括step,cosinew/ restarts,tanhw/ restarts,plateau 。

Optimizer：

• rmsprop_tf adapted from PyTorch RMSProp by myself. Reproduces much improved Tensorflow RMSProp behaviour.
• radam by https://github.com/LiyuanLucasLiu/RAdam (opens new window) (https://arxiv.org/abs/1908.03265 (opens new window))
• novograd by https://github.com/convergence-lab/novograd (opens new window) (https://arxiv.org/abs/1905.11286 (opens new window))
• lookahead adapted from impl by https://github.com/alphadl/lookahead.pytorch (opens new window) (https://arxiv.org/abs/1907.08610 (opens new window))
• fused<name> optimizers by name with https://github.com/NVIDIA/apex/tree/master/apex/optimizers (opens new window) installed
• adamp and sgdp by https://github.com/clovaai (opens new window) (https://arxiv.org/abs/2006.08217 (opens new window))
• adafactor adapted from https://github.com/pytorch/fairseq/blob/master/fairseq/optim/adafactor.py (opens new window) (https://arxiv.org/abs/1804.04235 (opens new window))
• adahessian by https://github.com/davda54/ada-hessian (opens new window) (https://arxiv.org/abs/2006.00719 (opens new window))