Cuda Memory Full during Feature extraction

I am working on video and getting frames and using the pretrained model, I tried predict features. I am trying to use GPU and end up having OOM error. Algo looks like

for loop --> for each video
—with torch.no_grad()
-----process video n fetch frames
-----batch frames
-----ft = model(frames)
-----x =

However, memory is full after processing few videos. Should I use
-----x =
-----del ft, model

I really don’t understand how things are working with tensors in Pytorch. Please let me understand the concept. Thanks