Sharing Ideas on Continuous Expansion of FAQ

[dolaameng]

Following the discussion on slack, I am trying to compile a list of questions from the last 3 days for the purpose of finding trends and updating FAQ. The current idea is to compile a summary page for issues during a period, say weekly, and picking candidates to add to FAQ.

You are welcome to share your thoughts, specially on

• 1. whether this is useful
• 2. how to go about it by using community efforts - commitment from specific people might not be a long term solution.
     I have spent almost 2 hours compiling the questions from last 3 days - somehow longer than I expected.
• 3. collection of questions fr om both issue pages and gitter, slack groups
• 4. formats of this sharing

For this experiment, I have put questions in the following categories.

• _bug_: bug reporting
• _documentation_: chance for revision of documentation
• _configuration_: installation, configuration, integration with other tools
• _performance_: need suggestions for performance improvments
• _customization_: lego-building new layers/models by Keras
• _enhancement_: feature requests, modifications for Keras itself

See the compilation below.

Bug

Gradients may become inf/nan through merge(mode = 'cos') #3941

• Description: Suggestion of changing merge(mode='cos') implementation to avoid inf/nan gradients error

Compatibility models keras 1.0.4 and 1.1.0 #3957

• Description: Models trained on Keras 1.0.4 cannot be loaded in 1.1.0

fail to load model with multiple outputs that have multiple metrics #3958

• Description: load_model doesn't work for multiple output with multiple metrics

Cifar-10 example not converging when image_dim_ordering == 'tf' #3959

• Description: 'cifar-10' example doesn't work with tf image_dim_ordering

load_model fails to load optimizer #3964

• Description: load_model fails to load optimizer but model_from_json and load_weights work fine

None type on input size does not seem to work. #3965

• Description: Specifying None in input_shape for Convolution2D with tf backend throws error

Documentation

Dimension conventions #3925

• Description: Explaination of 'th' and 'tf' image_dim_ordering for image and sequence data

Unable to Call model.predict_classes() #3938

• Description: calling predict_classes on a non-sequential model throws an error

Keras MaxPooling2D gives ValueError: Negative dimension size caused by subtracting 2 from 1 for 'MaxPool_x' #3945

• Description: "Negative Dimension" error by MaxPooling2D without explicitly specifying dim_ordering

flow_from_directory seems to find no images #3946

• Description: ImageDataGenerator.flow_from_directory didn't work if there is no subdir created

Input, InputLayer, Embedding, Merge does not import #3951

• Description: importing from keras.layers.core doesn't work any more

Configuration

Theano backend will not work on AWS Ubuntu:14.04 #3936

• Description: Configuring Keras on AWS with Flask, with different backend supports

Performance

Memory not released from the gpu #3939

• Description: GPU runs out of memory and gc.collect() doesn't work

Out of memory after 4 epochs #3954

• Description: GPU runs out of memory and gc.collect() doesn't work

How to normalize input #3943

• Description: Effects of normalizing inputs in sequence classification by RNN

BatchNorm with statistics from whole dataset #3949

• Description: BatchNormalization has a high momentum value which influences performances

keras 1.1.0 conv problems? #3956

• Description: Performance change of Convolution2D, MaxPooling2D from 1.0.6 to 1.1.0

Customization

SRGAN - Bypass check_array_lengths(X, Y, W) in training.py for different input and output batch sizes #3940

• Description: Effective way of implementing SRGAN

Implementing Custom Bilinear tensor product layer in keras #3944

• Description: Implementing Bilinear tensor product layer

error making simple FFT layer #3950

• Description: How to implement 1D FFT layer in Keras

How to create multiple outputs of sigmoid activation with shared layers #3960

• Description: Replacing softmax with n sigmoid activations

How to parallelize fit_generator? (PicklingError) #3962

• Description: Parallelizing data geneator

Enhancement

GlobalPooling for 3D inputs #3942

• Description: Suggestion of adding GlobalMaxPooling3D and GlobalAveragePooling3D

Feature Request: Output transformation in ImageDataGenerator #3953

• Description: adding "output transformation" feature in ImageDataGeneator