Somshubra Majumdar comments

Results 245 comments of


                                            Somshubra Majumdar

why not make image_size as a variable instead of a constant?

You're right. This was very early code when it came out. It didn't follow Keras standards to model building. Since Keras now comes with InceptionV4 built into applications, this can...

both new and old weights don't load

I don't precisely remember, but it's most definitely Keras 1.x later releases.

AttributeError: 'function' object has no attribute 'get_shape'

What are the functions that you are defining ? There is no error for Tensorflow backend with TF din ordering. I have not yet set it up for Theano channels...

AttributeError: 'function' object has no attribute 'get_shape'

Hmm, i'll look into it. Its an issue with Lambda layers needing a shape predefined for them.

Adding coordconv to unet leads to stange results

Hmm. What I can see is that the input image to the unet coordinate conv model doesn't have a gradient like the one above it. Even considering that to not...

Converting to tensorflow.keras gives strange error

Could you provide a full stack trace. That would be more helpful.

Can I use this model for video classification?

The `non_local` block itself supports all types of convolutions, however you would need to write the 3d ResNet variant to actually use it.

Can I use this model for video classification?

Yup that's should be sufficient.

Small bug about NonLocalResNet34/50/.... ？

Would you like to add a PR for this ? Or should I merge it directly ?

I've computed histograms of the ground truth and predicted scores

I don't have the resources to train these models so I won't be able to improve them. I wonder whether the difference lies in the loss function or the amount...