laviec / videoimagination Goto Github PK

Do you train on all UCF101 or do you filter by category?
Can you provide pre-trained models?

There is no value assigned for image_filters_2 in op.py

What should be its value?

In the 'volumetric_apply' function of 'ops.py' file, the return values are 'vid_1' and 'vid_2', but I find that 'vid_2' is not defined.

After making some changes to the source files, I was finally able to get this working. But am facing a few problems. I am using the 2D Shapes dataset for training.

1)It seems that the model is not training properly. The generated frames in tensorboard are very blurry even after ~194 Epochs.

2)The generator loss is ~0.78 and discriminator loss is ~ -1.58. The trend till now is that the G loss is increasing very slowly and the D loss is decreasing very slowly. These are that stats after 194 Epochs (Batch Size 64). Both should be going towards zero right ? When should I expect convergence ?

3. Further, strangely the learning starts at Epoch 34. Can't figure out why.

4. I can't get the validation part to work. I have the following error : "Variable generator/g_e0_conv/w already exists, disallowed. Did you mean to set reuse=True in VarScope?"

It is a wonderful work!
Can you release the dataset that you train?
Or can you tell how you generate the "tfrecords" data ?
If you can release the code that generate the data 'surfing.tfrecords',It will be very helpful for me!

"filenames = tf.gfile.Glob(self.data_dir+'/surfing_val.tfrecords')"

Hi ,
I would like to know if y'all have thought about implementation of the paper title CortexNet in your architecture?
https://arxiv.org/pdf/1706.02735v2.pdf
https://github.com/Atcold/pytorch-CortexNet