keras-adversarial
keras-adversarial copied to clipboard
bstriner
Keras 2 issues
Hi everybody! Just letting you know there are probably a lot of new issues with the change to Keras 2. Will update the library soon and keep you all posted.
Cheers!
@agulli and @ErmiaAzarkhalili Keras-2 fixes are now in master branch.
I setup Travis. The GAN example is tested in python 2 and 3, keras 1 and 2 and tf and theano. Should be good for any configuration.
https://travis-ci.org/bstriner/keras-adversarial/builds/221544182
FYI, this is the block of renames for keras-2.
self._feed_loss_fns = self.loss_functions
self._feed_inputs = self.inputs
self._feed_input_names = self.input_names
self._feed_input_shapes = self.internal_input_shapes
self._feed_outputs = self.outputs
self._feed_output_names = self.output_names
self._feed_output_shapes = self.internal_output_shapes
self._feed_sample_weights = self.sample_weights
self._feed_sample_weight_modes = self.sample_weight_modes
In terms of examples, I fixed example_gan and example_gan_cifar10. Need to change all the calls to Dense, Conv2D, etc. to keras-2 APIs. Those changes should be obvious so no huge rush. I added a module named legacy so the examples can work in both Keras 1 and 2 without any warnings.
I'll probably leave this issue open until I get around to fixing the rest of the examples.
Please let me know if you guys run into any other compatibility issues.
Cheers
Thank you for the quick update with keras 2 ! I have convergence issues with bigan architecture, I'm quite sure the problem comes from BatchNormalization in the discriminator. The mode argument has been removed with keras 2 (I still don't understand the reason of this choice), and it seems critical to have a sample-wise normalization in the discriminator... do you know any workaround ?
Thanks !
with Tensorflow, example_gan and example_gan_convoluntional.py work correctly but there is apparently a problem with example_gan_cifar10
python example_gan_cifar10.py Using TensorFlow backend. Training: output/gan-cifar10/history.csv
Layer (type) Output Shape Param #
dense_1 (Dense) (None, 4096) 413696
batch_normalization_1 (Batch (None, 4096) 16384
reshape_1 (Reshape) (None, 4, 4, 256) 0
conv2d_1 (Conv2D) (None, 4, 4, 128) 819328
batch_normalization_2 (Batch (None, 4, 4, 128) 16
leaky_re_lu_1 (LeakyReLU) (None, 4, 4, 128) 0
up_sampling2d_1 (UpSampling2 (None, 8, 8, 128) 0
conv2d_2 (Conv2D) (None, 8, 8, 128) 409728
batch_normalization_3 (Batch (None, 8, 8, 128) 32
leaky_re_lu_2 (LeakyReLU) (None, 8, 8, 128) 0
up_sampling2d_2 (UpSampling2 (None, 16, 16, 128) 0
conv2d_3 (Conv2D) (None, 16, 16, 64) 204864
batch_normalization_4 (Batch (None, 16, 16, 64) 64
leaky_re_lu_3 (LeakyReLU) (None, 16, 16, 64) 0
up_sampling2d_3 (UpSampling2 (None, 32, 32, 64) 0
conv2d_4 (Conv2D) (None, 32, 32, 3) 4803
activation_1 (Activation) (None, 32, 32, 3) 0
Total params: 1,868,915.0 Trainable params: 1,860,667.0 Non-trainable params: 8,248.0
Layer (type) Output Shape Param #
conv2d_5 (Conv2D) (None, 32, 32, 64) 4864
max_pooling2d_1 (MaxPooling2 (None, 16, 16, 64) 0
leaky_re_lu_4 (LeakyReLU) (None, 16, 16, 64) 0
conv2d_6 (Conv2D) (None, 16, 16, 128) 204928
max_pooling2d_2 (MaxPooling2 (None, 8, 8, 128) 0
leaky_re_lu_5 (LeakyReLU) (None, 8, 8, 128) 0
conv2d_7 (Conv2D) (None, 8, 8, 256) 819456
max_pooling2d_3 (MaxPooling2 (None, 4, 4, 256) 0
leaky_re_lu_6 (LeakyReLU) (None, 4, 4, 256) 0
conv2d_8 (Conv2D) (None, 4, 4, 1) 6401
average_pooling2d_1 (Average (None, 1, 1, 1) 0
flatten_1 (Flatten) (None, 1) 0
activation_2 (Activation) (None, 1) 0
Total params: 1,035,649.0 Trainable params: 1,035,649.0 Non-trainable params: 0.0
Traceback (most recent call last):
File "example_gan_cifar10.py", line 147, in
I am still getting the issue of this: 'AdversarialModel' object has no attribute '_feed_output_shapes'
@JiteshPshah that attribute is definitely in the constructor. I'm wondering if you have an old copy somewhere. Maybe you have multiple environments and one has the wrong version.
try python -c 'import keras_adversarial; print(keras_adversarial.adversarial_model.__file__)'
See if that file has the _feed_output_shapes line that is in this link.
https://github.com/bstriner/keras-adversarial/blob/master/keras_adversarial/adversarial_model.py#L119
If that isn't the problem, please post a stack trace.
Cheers
I still get a lot of "value passed to parameter 'shape' has DataType float32 not in list of allowed values: int32, int64" errors within the Cifar10_GAN example, so had to replace the nch/2 and nch/4 values by integers 128 and 64 after sorting this I then get the same error as aguilli: "ValueError: Error when checking input: expected conv2d_5_input to have shape (None, 32, 32, 3) but got array with shape (50000, 32, 3, 32)"
Many people, including me, are getting the following error (there' a stackOverflow post relating to this) in example_gan_convolutional :
AttributeError: 'Model' object has no attribute 'internal_input_shapes'
The issue with 'internal_input_shapes' can be solved by downgrading Keras to 2.0.0. Batch normalization with mode = 1 still does not work. Bigan example uses such mode so the simplest workaround is to not pass any parameter to batch normalization: b1 = BatchNormalization()
I'm not yet familiar with Keras batch normalization functionality thus don't know if the above workaround will affect bigan's performance.
It would be very helpful if keras-adversarial would work with the latest version of Keras. Is such update planned at all? If yes, any idea about when it can be ready?
Thanks a lot, Dmitry
@JiteshPshah that attribute is definitely in the constructor. I'm wondering if you have an old copy somewhere. Maybe you have multiple environments and one has the wrong version.
try
python -c 'import keras_adversarial; print(keras_adversarial.adversarial_model.__file__)'See if that file has the
_feed_output_shapesline that is in this link.https://github.com/bstriner/keras-adversarial/blob/master/keras_adversarial/adversarial_model.py#L119
If that isn't the problem, please post a stack trace.
Cheers
That was the issue. My version of the file did not have those assignments which are meant for Keras2. I added those and it works. Is this the fix or do I have to do something to get this right?
