Skip to content

Latest commit

 

History

History

pretrain

README.md

Preparation for ImageNet-1k pretraining

See /INSTALL.md to prepare pip dependencies and the ImageNet dataset.

Note: for neural network definitions, we directly use timm.models.ResNet and official ConvNeXt.

Tutorial for pretraining your own CNN model

See /pretrain/models/custom.py. Your todo list is:

Then run the experiment with --model=your_convnet.

Tutorial for pretraining on your own dataset

See the comment of build_dataset_to_pretrain in line55 of /pretrain/utils/imagenet.py. Your todo list:

  • Define a subclass of torch.utils.data.Dataset for your own unlabeled dataset, to replace our ImageNetDataset.
  • Use args.data_path and args.input_size to help build your dataset, with --data_path=... --input_size=... to specify them.
  • Note the batch size --bs is the total batch size of all GPU, which may need to be adjusted based on your dataset size. FYI: we use --bs=4096 for ImageNet, which contains 1.28 million images.

If your dataset is relatively small, you can try --init_weight=/path/to/res50_withdecoder_1kpretrained_spark_style.pth to do your pretraining from our pretrained weights, rather than form scratch.

Debug on 1 GPU (without DistributedDataParallel)

Use a small batch size --bs=32 for avoiding OOM.

python3 main.py --exp_name=debug --data_path=/path/to/imagenet --model=resnet50 --bs=32

Pretraining Any Model on ImageNet-1k (224x224)

For pretraining, run /pretrain/main.py with torchrun. It is required to specify the ImageNet data folder (--data_path), your experiment name & log dir (--exp_name and --exp_dir, automatically created if not exists), and the model name (--model, valid choices see the keys of 'pretrain_default_model_kwargs' in /pretrain/models/__init__.py line34).

We use the same pretraining configurations (lr, batch size, etc.) for all models (ResNets and ConvNeXts) in 224 pretraining. Their names and default values are in /pretrain/utils/arg_util.py line23-44. All these default configurations (like batch size 4096) would be used, unless you specify some like --bs=512.

Note: the batch size --bs is the total batch size of all GPU, and the learning rate --base_lr is the base lr. The actual lr would be lr = base_lr * bs / 256, as in /pretrain/utils/arg_util.py line131. So do not use --lr to specify a lr (that will be ignored)

Here is an example to pretrain a ResNet50 on an 8-GPU single machine (we use DistributedDataParallel), overwriting the default batch size to 512:

$ cd /path/to/SparK/pretrain
$ torchrun --nproc_per_node=8 --nnodes=1 --node_rank=0 --master_addr=localhost --master_port=<some_port> main.py \
  --data_path=/path/to/imagenet --exp_name=<your_exp_name> --exp_dir=/path/to/logdir \
  --model=resnet50 --bs=512

For multiple machines, change the --nnodes, --node_rank, --master_address and --master_port to your configurations. E.g.:

$ torchrun --nproc_per_node=8 --nnodes=<your_nnodes> --node_rank=<rank_starts_from_0> --master_address=<some_address> --master_port=<some_port> main.py \
  ...

Pretraining ConvNeXt-Large on ImageNet-1k (384x384)

For 384 pretraining we use a larger mask ratio (0.75), a half batch size (2048), and a double base learning rate (4e-4):

$ cd /path/to/SparK/pretrain
$ torchrun --nproc_per_node=8 --nnodes=<your_nnodes> --node_rank=<rank_starts_from_0> --master_address=<some_address> --master_port=<some_port> main.py \
  --data_path=/path/to/imagenet --exp_name=<your_exp_name> --exp_dir=/path/to/logdir \
  --model=convnext_large --input_size=384 --mask=0.75 --bs=2048 --base_lr=4e-4

Logging

See files in your --exp_dir to track your experiment:

  • <model>_withdecoder_1kpretrained_spark_style.pth: saves model and optimizer states, current epoch, current reconstruction loss, etc.; can be used to resume pretraining; can also be used for visualization in /pretrain/viz_reconstruction.ipynb

  • <model>_1kpretrained_timm_style.pth: can be used for downstream finetuning

  • pretrain_log.txt: records some important information such as:

    • git_commit_id: git version
    • cmd: the command of this experiment

    It also reports the loss and remaining pretraining time.

  • tensorboard_log/: saves a lot of tensorboard logs including loss values, learning rates, gradient norms and more things. Use tensorboard --logdir /path/to/this/tensorboard_log/ --port 23333 for viz.

  • stdout_backup.txt and stderr_backup.txt: backups stdout/stderr.

Resuming

Specify --resume_from=path/to/<model>_withdecoder_1kpretrained_spark_style.pth to resume pretraining. Note this is different from --init_weight:

  • --resume_from will load three things: model weights, optimizer states, and current epoch, so it is used to resume some interrupted experiment (will start from that 'current epoch').
  • --init_weight ONLY loads the model weights, so it's just like a model initialization (will start from epoch 0).

Regarding sparse convolution

We do not use sparse convolutions in this pytorch implementation, due to their limited optimization on modern hardware. As can be found in /pretrain/encoder.py, we use masked dense convolution to simulate submanifold sparse convolution. We also define some sparse pooling or normalization layers in /pretrain/encoder.py. All these "sparse" layers are implemented through pytorch built-in operators.

Some details: how we mask images and how to set the patch size

In SparK, the mask patch size equals to the downsample ratio of the CNN model (so there is no configuration like --patch_size=32).

Here is the reason: when we do mask, we:

  1. first generate the binary mask for the smallest resolution feature map, i.e., generate the _cur_active or active_b1ff in /pretrain/spark.py line86-87, which is a torch.BoolTensor shaped as [B, 1, fmap_h, fmap_w], and would be used to mask the smallest feature map.
  2. then progressively upsample it (i.e., expand its 2nd and 3rd dimensions by calling repeat_interleave(..., dim=2) and repeat_interleave(..., dim=3) in /pretrain/encoder.py line16), to mask those feature maps (x in line21) with larger resolutions .

So if you want a patch size of 16 or 8, you should actually define a new CNN model with a downsample ratio of 16 or 8. See Tutorial for pretraining your own CNN model (above).