dmlc / keras

Deep Learning library for Python. Convnets, recurrent neural networks, and more. Runs on MXNet, Theano or TensorFlow.
http://keras.io/
Other
125 stars 34 forks source link

Failing test 06/21/2017 #65

Closed kevinthesun closed 7 years ago

kevinthesun commented 7 years ago

Require sparse support:

Test cases
test_sparse_mlp
test_sparse_dot
test_sparse_concat

Require rnn symbolic loop implementation:

Test cases
test_dynamic_behavior[SimpleRNN]
test_dynamic_behavior[GRU]
test_dynamic_behavior[LSTM]
imdb_lstm
conv_lstm

Utility function issue, not major blocker:

Test cases Reason
test_value_manipulation Print tensor should return tensor
test_gradient Require implementation of symbol gradient
test_function Require updating variables
test_switch Require implementation of switching between two operations
test_random_binomial Require implementation of random binomial
test_ctc Require implementation of ctc_batch_cost
test_map Require implementation of mappinf function over elements
test_foldl Require implementation of reducing elements using function
test_foldr Require implementation of reducing elements using function
test_Eigenvalue_reg Not supported

Function incompatible with mxnet:

Test cases Reason
test_nn_operations MXNet Categorical_crossentropy doesn't support from_logits
test_arange Keras requires that when start >= stop and step > 0, this function should return an empty sequence. Currently mxnet returns error
test_batchnorm_mode_0_or_2 Currently keras mxnet doesn't support batchnorm mode 2
test_shared_batchnorm Currently keras mxnet doesn't support batchnorm mode 2
variational_autoencoder_deconv Too big target shape for mx.sym.deconvolution operator
mnist_acgan For mxnet backend, parameters are not shared between concatenated model and other separate models. So user needs to directly train generator instead of combine generator and discriminator to one model.
test_sequential_model_saving Optimizer states not preserved when saving/loading model
mnist_net2net mxnet doesn't support set weights to larger shape

Test case image ordering issue. Passing after modifying test cases. These are actually not failing tests:

Test cases
test_image_classification
test_conv2d
test_conv3d
test_atrous_conv_2d
test_averagepooling_2d
test_zero_padding_3d
test_TimeDistributed