ImageClassification

1. Introduction

This repo is only used for image classification task, such as imagenet. Include ddp training and inference, calculate the real acc and so on.

2. Enveriments

• python 3.7+
• pytorch 1.7.1
• pillow
• apex
• opencv-python

You can see this repo to find how to install the apex

3. Training & Inference

• dataset prepare

  /data/home/imagenet/xxx.jpeg, 0
  /data/home/imagenet/xxx.jpeg, 1
  ...
  /data/home/imagenet/xxx.jpeg, 999
  

• training
  1. Only used FP16 with bn FP32

     #!/bin/bash
     OMP_NUM_THREADS=1
     MKL_NUM_THREADS=1
     export OMP_NUM_THREADS
     export MKL_NUM_THREADS
     cd ImageClassification;
     CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -W ignore -m torch.distributed.launch --nproc_per_node 8 train_lanuch.py \
     --batch_size 512 \
     --num_workers 48 \
     --lr 1.6 \
     --max_epochs 90 \
     --warmup_epochs 5 \
     --num-classes 1000 \
     --input_size 256 \
     --crop_size 224 \
     --FP16 1 \
     --mode O2 \
     --apex 0 \
     --amp 0 \
     --train_file $train_file \
     --val_file $val_file \
     --log-dir $log_dir \
     --checkpoints-path $ckpt_dir

  2. Use Apex training

     #!/bin/bash
     OMP_NUM_THREADS=1
     MKL_NUM_THREADS=1
     export OMP_NUM_THREADS
     export MKL_NUM_THREADS
     cd ImageClassification;
     CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -W ignore -m torch.distributed.launch --nproc_per_node 8 train_lanuch.py \
     --batch_size 512 \
     --num_workers 48 \
     --lr 1.6 \
     --max_epochs 90 \
     --warmup_epochs 5 \
     --num-classes 1000 \
     --input_size 256 \
     --crop_size 224 \
     --FP16 0 \
     --mode O1 \
     --apex 1 \
     --amp 0 \
     --train_file $train_file \
     --val_file $val_file \
     --log-dir $log_dir \
     --checkpoints-path $ckpt_dir

  3. Use pytorch amp training

     #!/bin/bash
     OMP_NUM_THREADS=1
     MKL_NUM_THREADS=1
     export OMP_NUM_THREADS
     export MKL_NUM_THREADS
     cd ImageClassification;
     CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -W ignore -m torch.distributed.launch --nproc_per_node 8 train_lanuch.py \
     --batch_size 512 \
     --num_workers 48 \
     --lr 1.6 \
     --max_epochs 90 \
     --warmup_epochs 5 \
     --num-classes 1000 \
     --input_size 256 \
     --crop_size 224 \
     --FP16 1 \
     --mode O2 \
     --apex 0 \
     --amp 1 \
     --train_file $train_file \
     --val_file $val_file \
     --log-dir $log_dir \
     --checkpoints-path $ckpt_dir

• inference

  #!/bin/bash
  cd ImageClassification;
  CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 python -W ignore test.py \
  --dist-url 'tcp://127.0.0.1:9966' \
  --dist-backend 'nccl' \
  --multiprocessing-distributed=1 \
  --world-size=1 \
  --rank=0 \
  --test_file $test_file \
  --batch-size 128 \
  --num-workers 48 \
  --num-classes 1000 \
  --swin 0 \
  --checkpoints-path $ckpt_path \
  --save_folder $logits_folder

• calculate acc

python utils/calculate_acc.py --logits_file $logits_folder