UserWarning and TypeError when using Keras v2.x+

[jungwon1413]

Environment Info

Operating Hardware: local Workstation with Nvidia 1080ti
Operating System: Ubuntu 16.04 LTS
Installed CUDA / CuDNN: (doesn't matter, but just in case)

• CUDA: 8.0
• CuDNN: 6.0

pip packages

• Keras 2.1.1
• Tensorflow 1.4.0,

Problem:

• When I install ba-dls-deepspeech with instruction following along, certain errors arises if I execute train.py.

Main Reason:

• This is when I use up-to-date Keras which has partially changed for layer functions used in model.py.

How I solved:

• Fixed certain layer arguments used in model.py
• Detailed fixes are explained in Implementing tf.Keras (Keras v2.x) #35

Log:

Using TensorFlow backend.
WARNING (theano.tensor.blas): Using NumPy C-API based implementation for BLAS functions.
2017-12-19 14:10:19,570 INFO    (data_generator) Reading description file: ba_ds_ETRI_train_corpus.jsonl for partition: train
2017-12-19 14:10:19,888 INFO    (data_generator) Reading description file: ba_ds_ETRI_test_corpus.jsonl for partition: validation
2017-12-19 14:10:21,062 INFO    (model) Building gru model
/media/clab/1TB_HDD/baidu_deepspeech/model.py:109: UserWarning: Update your `Conv1D` call to the Keras 2 API: `Conv1D(1000, 11, activation="relu", padding="valid", kernel_initializer="glorot_uniform", strides=2, name="conv1d")`
  activation='relu')(acoustic_input)
Traceback (most recent call last):
  File "train.py", line 155, in <module>
    args.sortagrad)
  File "train.py", line 123, in main
    model = compile_gru_model(recur_layers=3, nodes=1000, batch_norm=True)
  File "/media/clab/1TB_HDD/baidu_deepspeech/model.py", line 111, in compile_gru_model
    output = BatchNormalization(name='bn_conv_1d', mode=2)(conv_1d)
  File "/home/clab/.local/lib/python3.5/site-packages/keras/legacy/interfaces.py", line 30, in wrapper
    args, kwargs, converted = preprocessor(args, kwargs)
  File "/home/clab/.local/lib/python3.5/site-packages/keras/legacy/interfaces.py", line 446, in batchnorm_args_preprocessor
    raise TypeError('The `mode` argument of `BatchNormalization` '
TypeError: The `mode` argument of `BatchNormalization` no longer exists. `mode=1` and `mode=2` are no longer supported.

Log2:

Using TensorFlow backend.
WARNING (theano.tensor.blas): Using NumPy C-API based implementation for BLAS functions.
2017-12-19 15:39:07,863 INFO    (data_generator) Reading description file: ./corpus/train_100_corpus.jsonl for partition: train
2017-12-19 15:39:07,952 INFO    (data_generator) Reading description file: ./corpus/validate_clean_corpus.jsonl for partition: validation
2017-12-19 15:39:08,551 INFO    (model) Building gru model
/media/clab/1TB_HDD/baidu_deepspeech/model.py:109: UserWarning: Update your `Conv1D` call to the Keras 2 API: `Conv1D(1000, 11, padding="valid", kernel_initializer="glorot_uniform", activation="relu", name="conv1d", strides=2)`
  activation='relu')(acoustic_input)
/media/clab/1TB_HDD/baidu_deepspeech/model.py:118: UserWarning: Update your `GRU` call to the Keras 2 API: `GRU(1000, kernel_initializer="glorot_uniform", return_sequences=True, activation="relu", name="rnn_1")`
  return_sequences=True)(output)
/media/clab/1TB_HDD/baidu_deepspeech/model.py:118: UserWarning: Update your `GRU` call to the Keras 2 API: `GRU(1000, kernel_initializer="glorot_uniform", return_sequences=True, activation="relu", name="rnn_2")`
  return_sequences=True)(output)
/media/clab/1TB_HDD/baidu_deepspeech/model.py:118: UserWarning: Update your `GRU` call to the Keras 2 API: `GRU(1000, kernel_initializer="glorot_uniform", return_sequences=True, activation="relu", name="rnn_3")`
  return_sequences=True)(output)
/media/clab/1TB_HDD/baidu_deepspeech/model.py:125: UserWarning: Update your `Dense` call to the Keras 2 API: `Dense(29, activation="linear", kernel_initializer="glorot_uniform", name="dense")`
  output_dim, name='dense', activation='linear', init=initialization,
/media/clab/1TB_HDD/baidu_deepspeech/model.py:127: UserWarning: Update your `Model` call to the Keras 2 API: `Model(inputs=Tensor("ac..., outputs=Tensor("ti...)`
  model = Model(input=acoustic_input, output=network_output)
2017-12-19 15:39:12,264 INFO    (model) Building train_fn
Traceback (most recent call last):
  File "train.py", line 155, in <module>
    args.sortagrad)
  File "train.py", line 126, in main
    train_fn = compile_train_fn(model)
  File "/media/clab/1TB_HDD/baidu_deepspeech/model.py", line 35, in compile_train_fn
    network_output = network_output.dimshuffle((1, 0, 2))
AttributeError: 'Tensor' object has no attribute 'dimshuffle'

AttributeError: 'Tensor' object has no attribute 'dimshuffle' is not related with Keras, but rather with tensorflow's Tensor module.