The training script load the data from DocSet files.
The preprocessed miniimagenet dataset can be downloaded from here (Extract code: knvt).
miniimagenet/
train.ds
valid.ds
test.ds
After you installed the docset python package, you can run docset train.ds
in your terminal to view the file format:
train.ds
Count: 36000, Size: 1.0 GB, Avg: 29.5 KB/sample
Sample 0
"filename": "n01614925_1001.JPEG"
"image": jpeg_image(size=(256, 353, 3))
"wnid": "n01614925"
"label": 22
Sample 1
"filename": "n01614925_10153.JPEG"
"image": jpeg_image(size=(256, 419, 3))
"wnid": "n01614925"
"label": 22
...
Sample 35999
"filename": "n03424325_9995.JPEG"
"image": jpeg_image(size=(383, 256, 3))
"wnid": "n03424325"
"label": 570
The following code train the model using on GPU:0
CUDA_VISIBLE_DEVICES=0 python3 train.py --data-path ~/data/miniimagenet/