update image classification

.gitignore

@@ -114,6 +114,9 @@ venv.bak/
 bash
 data
 data_set
+detection/data
+segmentation/data
+output/
 work_dirs/
 workspace/
 tools/exp_bash/
@@ -129,4 +132,5 @@ tensorboard.sh
 *.json
 
 # temp
-moganet_tiny_sz224_2xb128_accu4_ep300/
+detection
+segmentation

INSTALL.md

@@ -14,17 +14,17 @@ Install [Pytorch](https://pytorch.org/)>=1.8.0, [torchvision](https://pytorch.or
 pip install torch==1.10.0+cu111 torchvision==0.11.0+cu111 torchaudio==0.10.0 -f https://download.pytorch.org/whl/torch_stable.html
 ```
 
-Clone this repo and install required packages:
+Clone this repo and install required packages. You can install [apex-amp](https://github.com/NVIDIA/apex) if you want to use fp16 with Pytorch<=1.6.0.
 ```
 git clone https://github.com/Westlake-AI/MogaNet
-pip install timm tensorboardX six
+pip install timm fvcore
 ```
 
-The results in the paper are produced with `torch==1.10.0+cu111 torchvision==0.11.0+cu111 timm==0.6.12`.
+The results in this repository are produced with `torch==1.10.0+cu111 torchvision==0.11.0+cu111 timm==0.6.12`, and we adopt amp fp16 for fast training.
 
 ## Dataset Preparation
 
-Download the [ImageNet-1K](http://image-net.org/) classification dataset and structure the data as follows:
+Download the [ImageNet-1K](http://image-net.org/) classification dataset ([train](https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_train.tar) and [val](https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_val.tar)) and structure the data as follows. You can extract ImageNet with this [script](https://gist.github.com/BIGBALLON/8a71d225eff18d88e469e6ea9b39cef4).
 ```
 │imagenet/
 ├──train/

LICENSE

@@ -186,7 +186,7 @@
       same "printed page" as the copyright notice for easier
       identification within third-party archives.
 
-   Copyright [yyyy] [name of copyright owner]
+   Copyright 2022-2023 CAIRI AI Lab.
 
    Licensed under the Apache License, Version 2.0 (the "License");
    you may not use this file except in compliance with the License.

README.md

@@ -44,7 +44,6 @@ We propose **MogaNet**, a new family of efficient ConvNets, to pursue informativ
 We plan to release implementations of MogaNet in a few months. Please watch us for the latest release. Currently, this repo is reimplemented according to our official implementations in [OpenMixup](https://github.com/Westlake-AI/openmixup/), and we are working on cleaning up experimental results and code implementations.
 
 - [x] ImageNet-1K Training Code
-- [ ] ImageNet-1K Fine-tuning Code
 - [ ] Downstream Transfer to Object Detection and Instance Segmentation on COCO
 - [ ] Downstream Transfer to Semantic Segmentation on ADE20K
 - [ ] Image Classification on Google Colab and Web Demo
@@ -57,7 +56,7 @@ Please check [INSTALL.md](INSTALL.md) for installation instructions.
 
 ## Training
 
-See [TRAINING.md](TRAINING.md) for ImageNet-1K training instructions, or refer to our [OpenMixup](https://github.com/Westlake-AI/openmixup/tree/main/configs/classification/imagenet/moganet/) implementations. We have released pre-trained models on [OpenMixup](https://github.com/Westlake-AI/openmixup/tree/main/configs/classification/imagenet/moganet/) in [moganet-in1k-weights](https://github.com/Westlake-AI/openmixup/releases/tag/moganet-in1k-weights).
+See [TRAINING.md](TRAINING.md) for ImageNet-1K training instructions, or refer to our [OpenMixup](https://github.com/Westlake-AI/openmixup/tree/main/configs/classification/imagenet/moganet/) implementations. We released pre-trained models on [OpenMixup](https://github.com/Westlake-AI/openmixup/tree/main/configs/classification/imagenet/moganet/) in [moganet-in1k-weights](https://github.com/Westlake-AI/openmixup/releases/tag/moganet-in1k-weights). We have also reproduced ImageNet results with this repo and released `args.yaml` / `summary.csv` / `model.pth.tar`.
 
 ## Results and Pre-trained Models
 

TRAINING.md

@@ -4,18 +4,18 @@ We provide ImageNet-1K training commands here. Please check [INSTALL.md](INSTALL
 
 ## ImageNet-1K Training 
 
-Taking MogaNet-T as an example, you can use the following command to run this experiment on a single machine: 
+Taking MogaNet-T as an example, you can use the following command to run this experiment on a single machine (8GPUs): 
 ```
-python -m torch.distributed.launch --nproc_per_node=8 main.py \
+python -m torch.distributed.launch --nproc_per_node=8 train.py \
 --model moganet_tiny --input_size 224 --drop_path 0.1 \
---batch_size 128 --lr 1e-3 --weight_decay 0.04 --update_freq 1 \
---aa rand-m7-mstd0.5-inc1 --mixup 0.1 \
---model_ema false --model_ema_eval false \
---data_path /path/to/imagenet-1k \
---output_dir /path/to/save_results
+--epochs 300 --batch_size 128 --lr 1e-3 --weight_decay 0.04 \
+--aa rand-m7-mstd0.5-inc1 --crop_pct 0.9 --mixup 0.1 \
+--amp --native_amp \
+--data_dir /path/to/imagenet-1k \
+--experiment /path/to/save_results
 ```
 
-- Here, the effective batch size = `--nproc_per_node` * `--batch_size` * `--update_freq`. In the example above, the effective batch size is `8*128*1 = 1024`. Running on one machine, we can increase `update_freq` and reduce `--batch_size` to avoid OOM issues while keeping the total batch size unchanged.
+- Here, the effective batch size = `--nproc_per_node` * `--batch_size`. In the example above, the effective batch size is `8*128 = 1024`. Running on one machine, we can reduce `--batch_size` and use `--amp` to avoid OOM issues while keeping the total batch size unchanged.
 
 To train other MogaNet variants, `--model` and `--drop_path` need to be changed. Examples with single-machine commands are given below:
 
@@ -27,12 +27,13 @@ MogaNet-XT
 Single-machine (8GPUs) with the input size of 224:
 
 ```
-python -m torch.distributed.launch --nproc_per_node=8 main.py \
+python -m torch.distributed.launch --nproc_per_node=8 train.py \
 --model moganet_xtiny --input_size 224 --drop_path 0.05 \
---batch_size 128 --lr 1e-3 --weight_decay 0.03 --update_freq 1 \
---aa rand-m7-mstd0.5-inc1 --mixup 0.1 \
---data_path /path/to/imagenet-1k \
---output_dir /path/to/save_results
+--epochs 300 --batch_size 128 --lr 1e-3 --weight_decay 0.03 \
+--aa rand-m7-mstd0.5-inc1 --crop_pct 0.9 --mixup 0.1 \
+--amp --native_amp \
+--data_dir /path/to/imagenet-1k \
+--experiment /path/to/save_results
 ```
 </details>
 
@@ -43,23 +44,25 @@ MogaNet-Tiny
 Single-machine (8GPUs) with the input size of 224:
 
 ```
-python -m torch.distributed.launch --nproc_per_node=8 main.py \
+python -m torch.distributed.launch --nproc_per_node=8 train.py \
 --model moganet_tiny --input_size 224 --drop_path 0.1 \
---batch_size 128 --lr 1e-3 --weight_decay 0.04 --update_freq 1 \
---aa rand-m7-mstd0.5-inc1 --mixup 0.1 \
---data_path /path/to/imagenet-1k \
---output_dir /path/to/save_results
+--epochs 300 --batch_size 128 --lr 1e-3 --weight_decay 0.04 \
+--aa rand-m7-mstd0.5-inc1 --crop_pct 0.9 --mixup 0.1 \
+--amp --native_amp \
+--data_dir /path/to/imagenet-1k \
+--experiment /path/to/save_results
 ```
 
 Single-machine (8GPUs) with the input size of 256:
 
 ```
-python -m torch.distributed.launch --nproc_per_node=8 main.py \
+python -m torch.distributed.launch --nproc_per_node=8 train.py \
 --model moganet_tiny --input_size 256 --drop_path 0.1 \
---batch_size 128 --lr 1e-3 --weight_decay 0.04 --update_freq 1 \
---aa rand-m7-mstd0.5-inc1 --mixup 0.1 \
---data_path /path/to/imagenet-1k \
---output_dir /path/to/save_results
+--epochs 300 --batch_size 128 --lr 1e-3 --weight_decay 0.04 \
+--aa rand-m7-mstd0.5-inc1 --crop_pct 0.9 --mixup 0.1 \
+--amp --native_amp \
+--data_dir /path/to/imagenet-1k \
+--experiment /path/to/save_results
 ```
 </details>
 
@@ -70,12 +73,13 @@ MogaNet-Small
 Single-machine (8GPUs) with the input size of 224 with EMA (you can evaluate it without EMA):
 
 ```
-python -m torch.distributed.launch --nproc_per_node=8 main.py \
+python -m torch.distributed.launch --nproc_per_node=8 train.py \
 --model moganet_small --input_size 224 --drop_path 0.1 \
---batch_size 128 --lr 1e-3 --weight_decay 0.05 --update_freq 1 \
---model_ema true --model_ema_eval true \
---data_path /path/to/imagenet-1k \
---output_dir /path/to/save_results
+--epochs 300 --batch_size 128 --lr 1e-3 --weight_decay 0.05 \
+--crop_pct 0.9 \
+--model_ema --model_ema_decay 0.9999 \
+--data_dir /path/to/imagenet-1k \
+--experiment /path/to/save_results
 ```
 </details>
 
@@ -86,12 +90,13 @@ MogaNet-Base
 Single-machine (8GPUs) with the input size of 224 with EMA:
 
 ```
-python -m torch.distributed.launch --nproc_per_node=8 main.py \
+python -m torch.distributed.launch --nproc_per_node=8 train.py \
 --model moganet_base --input_size 224 --drop_path 0.2 \
---batch_size 128 --lr 1e-3 --weight_decay 0.05 --update_freq 1 \
---model_ema true --model_ema_eval true \
---data_path /path/to/imagenet-1k \
---output_dir /path/to/save_results
+--epochs 300 --batch_size 128 --lr 1e-3 --weight_decay 0.05 \
+--crop_pct 0.9 \
+--model_ema --model_ema_decay 0.9999 \
+--data_dir /path/to/imagenet-1k \
+--experiment /path/to/save_results
 ```
 </details>
 
@@ -102,11 +107,12 @@ MogaNet-Large
 Single-machine (8GPUs) with the input size of 224 with EMA:
 
 ```
-python -m torch.distributed.launch --nproc_per_node=8 main.py \
+python -m torch.distributed.launch --nproc_per_node=8 train.py \
 --model moganet_large --input_size 224 --drop_path 0.3 \
---batch_size 128 --lr 1e-3 --weight_decay 0.05 --update_freq 1 \
---model_ema true --model_ema_eval true \
---data_path /path/to/imagenet-1k \
---output_dir /path/to/save_results
+--epochs 300 --batch_size 128 --lr 1e-3 --weight_decay 0.05 \
+--crop_pct 0.9 \
+--model_ema --model_ema_decay 0.9999 \
+--data_dir /path/to/imagenet-1k \
+--experiment /path/to/save_results
 ```
 </details>