search

lliuz / ARFlow

The official PyTorch implementation of the paper "Learning by Analogy: Reliable Supervision from Transformations for Unsupervised Optical Flow Estimation".
MIT License
251 stars 50 forks source link

Loss doesn't go down #5

Closed leogogogo closed 4 years ago

leogogogo commented 4 years ago

Hi, read your paper and very impressive, and thank you for sharing your code. I'm trying your code recently, maybe not dig into too much details yet, just plainly try to reproduce the fine-tune with Sintel datasets. The loss doesn't go down, and stays around 0.7, and the evaluation in epoch 68 is "EPE_0: 3.19 EPE_1: 4.22", is this normal? Because all I have done is to download official Sintel dataset and try the command "python3 train.py -c sintel_ft_ar.json", and I also use "correlation_native".

lliuz commented 4 years ago

Hi @leogogogo, Your result is more like for directly fine-tuning without ar. I wonder what's the version of the code you used. Because of my carelessness when I sort and public this code, training with ar has a bug, and I have updated the training code a few days ago. Please check your code and let me know whether you are using the latest version.

leogogogo commented 4 years ago

Hi @lliuz Thank you for your quick answer, the code was from few days ago. I'll try the latest version and keep updating my result.

lliuz commented 4 years ago

BTW, when I was training, I did not pay much attention to the training loss, since the unsupervised loss is too noisy for each sample. You can just focus on the validation results to determine when to stop training.

Besides, you can email me if you have any other questions.

leogogogo commented 4 years ago

Got it, thank you for your help and hints!