Tzu-Wei Sung

Thanks for clarification! Our use case: In time sequence modeling, the input has dynamic spatial dim (batch, sequence, feature), we do multiple dilated conv to capture larger receptive field (finally...

I am sorry to ask why this kind of inputs does not work correctly. It should work properly as least for dynamic batch dim.

Hi @bjtho08, thanks for reporting the issue. It seems that we only support 2D `y_true` and `y_pred` shaped like `[batch_size, num_classes]`.

> @WindQAQ do you know what kind of F1 impl we have in TF/Model? No, but I remember that they asked us if we can migrate F1 score very long...

> @WindQAQ yes, I realized over the weekend as I was working on isolating the issue. I made a local copy, where I reshape `y_true` and `y_pred` using `tf.reshape(tf.cast(y_true, self.dtype),...

/cc @Susmit-A for visibility. Seems that pytorch will have overlapping kernel when the dimension is not divisible by output size. I wonder if this is achievable with python ops. https://colab.research.google.com/drive/1KMSD4nMv_-P59TsbSxQksndBZ6oNyyri?usp=sharing

Hi @Susmit-A, I'd say we should try to support non-divisible case. If not, is there any different between adaptive and original pooling? We can always infer the strides and kernel...

For divisible case, rounding up or down do not matter, which means current adaptive pooling is just a variant for `tf.nn.{avg,max}_pool`. We should do some benchmarks to see if that...