You need to prepare list files of image paths and labels for training/validation/testing. The format is
path/to/image0.png 2
path/to/image1.png 0
path/to/image2.png 0
path/to/image3.png 1
The pretrained AlexNet model is split to 3 files because Github has a limitation of 100MB/file. To merge them, please run
$ ./prepare_pretrained_alexnet.sh
Please modify N_CLASSES in globals.py for your data.
To train at the first time, run
$ mkdir tmp
$ python train.py --train_dir tmp --caffemodel alexnet_imagenet.npy
To fine-tune, run
# N is your checkpoint iteration
$ python train.py --train_dir tmp/ --weights tmp/model.ckpt-N --learning-rate=0.001
# N is your checkpoint iteration
$ python test.py --weights tmp/model.ckpt-N