Hi,
I have found some small issues and would like to provide fixes on:
1) src/data.py - VideoDataset
Sampling 16-frame sequence, taking every 2nd frame from shapes dataset cannot be done due to bug in data.py:102. With video_len = 32, self.video_length=16, self.every_nth=2 there are enough frames to sample video with stride 2.
In current version, frames from 0 to 15 are sampled, resulting into learning half of the full shape motion (from border to center).
2) src/models.py - VideoGenerator - sample_z_video
As suggested in #16, launching training with dim_z_category=0 results into crash.
Output:
{'--batches': '100000',
"please use transforms.Resize instead.")
'--dim_z_category': '0',
'--dim_z_content': '50',
'--dim_z_motion': '10',
'--every_nth': '2',
'--image_batch': '32',
'--image_dataset': '',
'--image_discriminator': 'PatchImageDiscriminator',
'--image_size': '64',
'--n_channels': '3',
'--noise_sigma': '0.1',
'--print_every': '100',
'--use_categories': False,
'--use_infogan': False,
'--use_noise': True,
'--video_batch': '32',
'--video_discriminator': 'PatchVideoDiscriminator',
'--video_length': '16',
'<dataset>': '../data/shapes',
'<log_folder>': '../logs/shapes'}
Total number of frames 256000
Traceback (most recent call last):
File "/home/vlad/PycharmProjects/temp/mocogan/src/train.py", line 130, in <module>
trainer.train(generator, image_discriminator, video_discriminator)
File "/home/vlad/PycharmProjects/temp/mocogan/src/trainers.py", line 262, in train
self.image_batch_size, use_categories=False)
File "/home/vlad/PycharmProjects/temp/mocogan/src/trainers.py", line 165, in train_discriminator
fake_batch, generated_categories = sample_fake(batch_size)
File "/home/vlad/PycharmProjects/temp/mocogan/src/trainers.py", line 236, in sample_fake_image_batch
return generator.sample_images(batch_size)
File "/home/vlad/PycharmProjects/temp/mocogan/src/models.py", line 282, in sample_images
z, z_category_labels = self.sample_z_video(num_samples * self.video_length * 2)
File "/home/vlad/PycharmProjects/temp/mocogan/src/models.py", line 258, in sample_z_video
z_category, z_category_labels = self.sample_z_categ(num_samples, video_len)
File "/home/vlad/PycharmProjects/temp/mocogan/src/models.py", line 234, in sample_z_categ
classes_to_generate = np.random.randint(self.dim_z_category, size=num_samples)
File "mtrand.pyx", line 993, in mtrand.RandomState.randint
ValueError: low >= high
Process finished with exit code 1
Fix: not using z_category in sampling in src/models.py:261
Hi, I have found some small issues and would like to provide fixes on:
1) src/data.py - VideoDataset Sampling 16-frame sequence, taking every 2nd frame from shapes dataset cannot be done due to bug in data.py:102. With video_len = 32, self.video_length=16, self.every_nth=2 there are enough frames to sample video with stride 2. In current version, frames from 0 to 15 are sampled, resulting into learning half of the full shape motion (from border to center).
2) src/models.py - VideoGenerator - sample_z_video As suggested in #16, launching training with dim_z_category=0 results into crash.
Output:
Fix: not using z_category in sampling in src/models.py:261
Hope it was helpful Regards, Vlad