evaluation on NYU looks horrible

[drsleep]

evaluation on NYU looks horrible

[DrSleep]

I did not run the models yet.

But I think I spotted the error: if you train using the script in this repo, then the model is being wrapped with nn.DataParallel - https://github.com/DrSleep/light-weight-refinenet/blob/master/src/train.py#L354
Now, when it happens, parameters' names in the state_dict have 'module.' before their original names.
In your inference script, you don't wrap the model with nn.DataParallel, hence the pre-trained weights are not actually loaded. You can see that by setting up strict=True

TL;DR in your inference script apply nn.DataParallel to your model before loading the weights.

[DrSleep]

that looks weird.

just to clarify: did you run the exactly same notebook code? and only changed the checkpoint file?
if so, I wonder if you can share the checkpoint file and the training log if possible.

[st2yang]

Hi Vladimir,

Thanks for your reply.

Yes I run the code adapted from the notebook (mainly just copied and pasted). I can produce the right result figure with provided weights file. And I only changed the weights file then the result figure becomes horrible.

I packed related files (infer.py adapted from example notebook, training log and weights file) here https://drive.google.com/open?id=1SOf1CFudGdZ0Tf7OHASoBQu3ta22BJo-
can you please take a look?

Thanks a lot!

[st2yang]

You're right. I made some mistakes in loading the weight file.

After fixing it, now I can get reasonably good prediction result (for ResNet50) as follows.

Thanks a lot for your help. I appreciate it a lot!

[xiaoyuamw]

Hi@st2yang

Can you tell me what you had changed to solve you issue? and the link "https://drive.google.com/open?id=1SOf1CFudGdZ0Tf7OHASoBQu3ta22BJo-" provided by you can't to be opened now, so i also don't know what you have changed!
Thank you very much.

Best Regards.

[st2yang]

@xiaoyuamw You can refer to Vladimir's response above or this repo https://github.com/choicelab/grasping-invisible.

As far I can remember, I just warp the model with nn. DataParallel