Questions about Train

[wanan0414]

I'm going to do a radar signal classification task, and using the pre-training weights from imagenet isn't appropriate. I also don't have enough dataset for pre-training here, is it possible to train directly with supervised learning using only a training set with 22000 data volume and a test set? Looking forward to your answer!

[ahatamiz]

Hi @wanan0414

Yes ! the codebase supports training on any datasets for classification tasks. Here's a command prompt assuming you are running on 8 GPUs (but it is also easy to extend for multi-node setup). You need to specify DATA_PATH_TRAIN and DATA_PATH_VAL for your train and validation (test) sets, respectively.

Although I recommend trying different learning rates LR and batch sizes BS for your setup.

Needless to say, any of our MambaVision variants can be specified such as mamba_vision_T.

#!/bin/bash

MODEL=mamba_vision_T
DATA_PATH_TRAIN="/my_dataset/train"
DATA_PATH_VAL="/my_dataset/val"
BS=256
EXP=my_experiment
LR=5e-4
WD=0.05
DR=0.2

torchrun --nproc_per_node=8 train.py --input-size 3 224 224 --crop-pct=0.875 \
--train-split=$DATA_PATH_TRAIN --val-split=$DATA_PATH_VAL --model $MODEL --amp --weight-decay ${WD} --drop-path ${DR} --batch-size $BS --tag $EXP --lr $LR

Hope this helps but let me know if there's any issue.

[ahatamiz]

The above also assumes that images of each class are placed within the same folder under both train or validation as shown in the following:

├── train
│   ├── class1
│   │   ├── img1.jpeg
│   │   ├── img2.jpeg
│   │   └── ...
│   ├── class2
│   │   ├── img3.jpeg
│   │   └── ...
│   └── ...
└── val
    ├── class1
    │   ├── img4.jpeg
    │   ├── img5.jpeg
    │   └── ...
    ├── class2
    │   ├── img6.jpeg
    │   └── ...
    └── ...

[wanan0414]

Hi @ahatamiz
Thank you very much for your reply! My dataset is in the same format as the train-val you mentioned. Thank you for providing the schematic code. A new question I have is in that case do I not use validate.py and just focus on the train.py file?

[ahatamiz]

Hi @wanan0414

In this case, validate.py can be used to test the model's performance on your test set, assuming it's different from your validation set.

[wanan0414]

Hello! Thank you for your previous answer.
Now I have a new question about the model parameters.
My image is 1×64×64 grayscale and the whole training set is only 22000 and test set 5500 . So I guess it might not be appropriate to use these parameters you gave directly.

def mamba_vision_T(pretrained=False, **kwargs):
    model_path = kwargs.pop("model_path", "/tmp/mamba_vision_T.pth.tar")
    depths = kwargs.pop("depths", [1, 3, 8, 4])
    num_heads = kwargs.pop("num_heads", [2, 4, 8, 16])
    window_size = kwargs.pop("window_size", [8, 8, 14, 7])
    dim = kwargs.pop("dim", 80)
    in_dim = kwargs.pop("in_dim", 32)
    mlp_ratio = kwargs.pop("mlp_ratio", 4)
    resolution = kwargs.pop("resolution", 224)
    drop_path_rate = kwargs.pop("drop_path_rate", 0.2)
    pretrained_cfg = resolve_pretrained_cfg('mamba_vision_T').to_dict()
    update_args(pretrained_cfg, kwargs, kwargs_filter=None)
    model = MambaVision(depths=[1, 3, 8, 4],
                        num_heads=[2, 4, 8, 16],
                        window_size=[8, 8, 14, 7],
                        dim=80,
                        in_dim=32,
                        mlp_ratio=4,
                        resolution=224,
                        drop_path_rate=0.2,
                        **kwargs)
    model.pretrained_cfg = pretrained_cfg
    model.default_cfg = model.pretrained_cfg
    if pretrained:
        if not Path(model_path).is_file():
            url = model.default_cfg['url']
            torch.hub.download_url_to_file(url=url, dst=model_path)
        model._load_state_dict(model_path)
    return model

Do you have some suggestions for me regarding the modification of the dim, in_dim, depths, etc. parameters in this part of the code?