twcmchang / mlds2017 Goto Github PK

I tried your code in hw3, and found some errors when running the project.

1. It seems that the images are not transformed to 64x64 in the function 'get_image()' in utils.py. So I add image transform in the function ' get_image()'.

2. In embed.py, tags containing 'hair' or 'eye' (i.e. 18175 images) are selected for training with the following code:
   tag_dict_in_use_1 = dict([(k,v) for k,v in tag_dict_in_use.items() if not v == ' and '])
   but I think it should be:
   tag_dict_in_use_1 = dict([(k,v) for k,v in tag_dict_in_use.items() if not v == ' '])
   Could you please check if my revision is correct?

3. I tried the code with the above revision, but I found the discriminator and generator loss both increases rapidly when running the code with WGAN_v2 (i.e., the loss goes to 10000 when epoch>200). However, DCGAN and WGAN worked.
   I want to ask if there are any other different settings except the args when running the code with WGAN_v2?

4. In your report, there are 3 cases with 18175 images, 11569 images and 33431 images, respectively. But I didn't find the tag processing code corresponding the 3 cases. Could you please tell how to process the images and tags, and run the code in case 2 and case 3?

Thank you very much!