wolverinn / hevc-cu-depths-prediction-cnn Goto Github PK

Hello there,
I am recently studying on your code, the outcome of the CNN oriented network is fansinating, but I am wondering if you could show me some clues about the ResNet architecture. Is there any reference? I am currently facing the problem of training loss non-convergence, and I don't know why. Could you please give me more details about your network architecture?

Hello,

What is the version of HM you use?

I need help, I am trying to reimplement the code on the cpu but I am getting this error: "RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU."
I tried to add map_loction =torch.device('cpu') after model.load_state_dict(torch.load(paft.format(LOAD_DIR)) as shown below but I am still getting the same error
model.load_state_dict(torch.load(paft.format(LOAD_DIR)),map_location=torch.device('cpu'))

Hello,

Batch_size=1024 in your code, it is too large. Can you tell me the value of batch_size in your actual training model? Thank you very much!

hello,

Encoding time of using CNN didn't reduce the time, however, more than the time of using original HM. This why? I need your help. Thank you !!!

Hi there, your encoding approach sounds really interesting !

It would be very interesting to apply it to AV1 encoding (wich is extremely heavy/slow)...

...it would be also nice to "mixup" with other alternative methods such as @master-of-zen 's per scene quality encoding:

https://github.com/master-of-zen/Av1an

Hope tat inspires !

Hello, I would like to ask you how you embedded the CNN into the encoder. It would be very helpful if I could inform, I would be grateful.

Hi,

Are the vectors depicting 4x4 blocks in the 16x16 matrix from Left to right, top to bottom?
Also, can you please provide example text files that you used in your pipeline that was read by the HM encoder.