Fit DenseNet examples to Keras-2 #53
Conversation
|
keras-team/keras#6034 may be relevant. Also a much more complete version of these changes is in https://github.com/farizrahman4u/keras-contrib/pull/46/files#diff-7c92da79e0a644cc9e206dab6de13bab |
|
LGTM, thanks! |
There was a problem hiding this comment.
@farizrahman4u please revert this change. It is broken. See: a6c7c91#r108550331
My changes will be ready for review today which are a superset of these changes.
| @@ -375,15 +376,16 @@ def __dense_block(x, nb_layers, nb_filter, growth_rate, bottleneck=False, dropou | |||
| x = __conv_block(x, growth_rate, bottleneck, dropout_rate, weight_decay) | |||
| x_list.append(x) | |||
|
|
|||
| x = merge(x_list, mode='concat', concat_axis=concat_axis) | |||
| x1 = concatenate(x_list, axis=concat_axis) | |||
There was a problem hiding this comment.
Not good! then all the convolutions will be the same! it is essential that all the concatenation be on x, in other words all new instances of x1 are a bug.
There was a problem hiding this comment.
Specifically, this is the source of keras-team/keras#6034
There was a problem hiding this comment.
Hmm, interesting. So I should wait for this PR to be merged, right?
There was a problem hiding this comment.
Unfortunately, this change both triggers an existing bug and adds a new one. Consider just utilizing/merging #46 it fixes everything mentioned here, and I'll be switching to concat calls once keras-team/keras#6034 or keras-team/keras#6035 is merged.
There was a problem hiding this comment.
Sure, I will take a look. Thanks.
This reverts commit fcd054f.
|
@ahundt Has you code be merged? If yes, then we don't need the revert any more, right? |
|
@kemaswill I'm still waiting on review for #46 |
|
Got it. |
No description provided.