holmdk / video-prediction-using-pytorch Goto Github PK

Thanks for your great work! I wonder if I want to use my dataset to train model, how to rewrite the dataloader？

I have not been able to run the main code due to an error thrown by the Pillow library, inside the coding of the torchvision library. I have been able to install all the exact versions of the libraries provided, except for pytorch lightning which does not show support for the stated version so I installed the latest available for a Conda environment.

You can find here a screenshot of the error given. If you could please provide some insight, this would be very helpful. Thank you!

Hi, How do you evaluate this model?

Hello!
Thank you for the nice code.
Is it possible to change the input size of the images given to the network?
Let's say: not 64x64, but 200x100?
How the structure of the network should be changed?
Thank you for the answer in advance!

It would be nice to have some results on the README or the blog.

hello, I came from your blog:) - https://holmdk.github.io/2020/04/02/video_prediction.html

Can we say input sequence is encoded through encoder even though dimension of last hidden state (12,10,64,64,64) of encoder is bigger than dimension of input sequence (12,10,1,64,64)?

I get the following error when trying to run the main.py

File "main.py", line 127, in MovingMNISTLightning
@pl.data_loader
AttributeError: module 'pytorch_lightning' has no attribute 'data_loader'

Hi!
I fed my own dataset to train the model. The shape of each image in my dataset is 48 * 48 * 1, My images are not binary images but gray images with their value in [0.0, 1.0].
I noticed that the backgroud color of x[6:], y_hat are different from each other, and the ground-truth imagey.

Would you please help me explain the reason for the issue?
Thanks.

Hi,

I would like to ask about using this model with a data set of multi-channel images. Which parts of the code should be changed to adjust this code for multi-channel images?

When looking at the prediction of evaluation the outputs of the model are just blank.
Were you able to train the model to predict expected frames?
The loss goes down but the model seems to have converged to optimally predicting blanks.