[Bugfix] Fix assertion error when using DenseNetFCN with grayscale input + Add activation parameter #52
Conversation
|
The tests are failing due to a "E303 too many blank lines (2)" in the PEP test which I have fixed. The other errors are due to Keras 2 code errors in test_deconvolution_3d and test_advanced_activations |
titu1994
changed the title
|
I've added an activation parameter, which allows use of sigmoid activation when training a network on binary segmentation masks. This is more suitable than a class wise loss over individual pixels for binary masks. |
| if include_top: | ||
| weights_path = get_file('densenet_40_12_th_dim_ordering_th_kernels.h5', | ||
| weights_path = get_file('densenet_40_12_th_data_format_th_kernels.h5', |
There was a problem hiding this comment.
did you upload files to the new path? Same applies to other weight files
There was a problem hiding this comment.
No I did not upload any new file to a new path. The path remains the same, just the name of the weight file changes when it is downloaded onto a local machine.
| # Determine proper input shape | ||
| min_size = 2 ** nb_dense_block | ||
| input_shape = _obtain_input_shape(input_shape, | ||
| default_size=32, | ||
| min_size=16, |
There was a problem hiding this comment.
should min_size be passed here?
There was a problem hiding this comment.
Seems the code did not update. Will fix in next commit.
| @@ -43,7 +43,7 @@ def DenseNet(depth=40, nb_dense_block=3, growth_rate=12, nb_filter=16, nb_layers | |||
| optionally loading weights pre-trained | |||
There was a problem hiding this comment.
Add activation parameter to DenseNet as well. Not only is it useful in regular DenseNet for the two class case, it will also be very useful when combined with Atrous DenseNet in #46.
| @@ -406,7 +415,7 @@ def __transition_up_block(ip, nb_filters, type='upsampling', output_shape=None, | |||
| elif type == 'subpixel': | |||
| x = Conv2D(nb_filters, (3, 3), padding="same", kernel_regularizer=l2(weight_decay), activation='relu', | |||
| use_bias=False, kernel_initializer='he_uniform')(ip) | |||
| #x = Convolution2D(nb_filters, 3, 3, activation="relu", border_mode='same', W_regularizer=l2(weight_decay), | |||
There was a problem hiding this comment.
These comments shouldn't be here, and all calls should use the updated Keras-2 API... You might just want to make use of the similar changes in #46
There was a problem hiding this comment.
I'm keeping Keras 1 code for now, since it is cross compatible with Keras 2. Not the other way around.
I need to use Keras 1.2.2 for the time being, since I need many features which were abruptly removed for random reasons from Keras 2 (such as Batch norm mode 2 which I need for GANs, fbeta_score metrics, and a few bugs which have come up in 2 are making me wait for the shift).
I will update the code when Keras 2 stabilizes (bug free), and fixes some of its mistakes like Mode 2 batchnorm. Until then, better to keep code compatible with both versions.
|
Someone merged the Keras 2 api code changes to this, and so I will rewrite this PR. Also, who merged the changes related to concatenate when the master Keras is showing bugs with concatenate in DenseNet?? |
Fixes a assertion error when gray scale image is input to DenseNetFCN.
The submitted code is basically a copy of the _check_input code in Keras, with the removal of a few parts which cause the crash.