pip install -r reuirements.txt
The directory structure is the standard layout as following.
/path/to/imagenet/
train/
class1/
img1.jpeg
class2/
img2.jpeg
val/
class1/
img3.jpeg
class/2
img4.jpeg
To training the same model as described in DeiT. You should use the following comman and specify the model.
python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --model deit_small_patch16_224 --batch-size 128 --data-path /path/to/imagenet
To distill the model use a teacher model in a hard distillation way, you should specify the teach model and whether use the distill token or the normal way. An example is given like:
python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --model deit_small_patch16_224 --batch-size 128 --data-path /path/to/imagenet --teacher_model /path/to/teacher_model_checkpoint [--distill_token]
To fine-tune a trained model on a higher resolution, you should specify the argument --input-size and --resume.
python -m torch.distributed.launch --nproc_per_node=8 --use_env finetune.py --model deit_small_patch16_224 --batch-size 128 --input-size 384 --data-path /path/to/imagenet --resume /path/to/trained_model
To use the 2D relative position, you should add the argument --relative_position to the training command. For example, to train deit small with relative position embedding, you should run:
python -m torch.distributed.launch --nproc_per_node=8 --use_env main.py --model deit_small_patch16_224 --batch-size 128 --data-path /path/to/imagenet --relative_position
The itp script is in the itp directory. To submit the job to cluster, you should specify the --cluster and the numbers of gpu you want to use --gpus. Here is an example:
python azureml_script.py --name ours_small_batch_size_128_8_gpus_lr_1e-3_warmup_epochs_20 --cluster itplabrr1cl1 --gpus 8
Implemented the polynomial lr scheduler and integrated it with timm LRscheduler class. To use the polynomial LR scheduler, please specify the argument --sched polynomial