alokwhitewolf / guided-attention-inference-network Goto Github PK

Thanks for your code.
I had a question of the set_init_grad code reference here:

On line 61, you use xp.random.choice(label,1). Can you explain why are you using random choice when you know the ground truth label index? I did not see that in the original paper and was curious to know if it improved your accuracy or performance. @alokwhitewolf

Thanks a lot for your great work.
I have read your paper and now try to reproduce your work.
For this purpose, do I need to install any deep-learning libraries, such as pytorch or tensorflow?
Thanks a lot.

Thanks for nice implementation!
I found you commented 'assert image.shape[0] == 1, "Batchsize of only 1 is allowed for now"' in the code,
and wonder why you limited the batch size to 1.
Is there any operation that force you to limit batchsize = 1 ?
Thanks!

Line 86-88 in model/fcn8.py, layer name is unmatched.

Hi, thank you for your work.
Do you have any suggestions on this behavior ?

I ran the training for the classificator (without GAIN) the minimal amount of iterations just to have a snapshot of the model, same with GAIN (I've tried to use the previously pretrained model, and also without pretraining, same accuracy, no change).

The accuracy for both with and without GAIN is about 72%, although maybe it is explainable, why the result is the same (not trained enough), but isn't there should be any heating (in the meaning of the heating maps) just with 72% accuracy ?

Is this expected or anomalous behavior ?
I didn't use any pretrained models of FCN8, because I didn't find them available now, so I can look only at those results.
If it is anomalous, what can be the reason, where should I look for the problem for your opinion ?

Thank you in advance, appreciate your help.

Thank you for sharing the implementation. I tried to reproduce by running 'train_classifier.py' which produces classifier_model_14640, then running train_GAIN.py with the classifier_model_14640 as the pretrained FCN8 model. However it throw out this error 'ValueError: Loss is nan.' at [line 206, fcn8.py]. All variables after conv2 seem to be all nan. Please could you have a look at this problem? Many thanks.

The pretrained model is not available. I tried following the code "python3 train_classifier.py --device 0" but the file is not there for usage.

Thanks for this great implement. I go through your codes, and find that this repo has implemented the classification loss and attention mining loss in the paper. Do you have the plan to add the attention loss (directly compare the l2 distance of the attention map and pixel-wise mask) in the paper? Meanwhile, I notice that you add the segmentation loss in this repo. This segmentation loss is helpful for generating an accurate attention map?

In visualize.py file, some variables are assigned but never used, for example, trainer in line 47 and undefined variable in line 49 etc.

Could you please check this out?

After training classifer, it generates snapshot. When i use this snapshot to train gain, this error occurs. Maybe fc6/b is bias of conv layer, which is not in classifier model. In classifier, it is fully connected layer named fc6_cl.
I am new to chainer, how to load snapshot to Gain model?