CUDNN_STATUS_BAD_PARAM when trying to perform inference on a LSTM Seq2Seq with masked inputs

问题

I'm using keras layers on tensorflow 2.0 to build a simple LSTM-based Seq2Seq model for text generation.

versions I'm using: Python 3.6.9, Tensorflow 2.0.0, CUDA 10.0, CUDNN 7.6.1, Nvidia driver version 410.78.

I'm aware of the criteria needed by TF to delegate to CUDNNLstm when a GPU is present (I do have a GPU and my model/data fill all these criteria).

Training goes smoothly (with a warning message, see the end of this post) and I can verify that CUDNNLstm is being used.

However, when I try to call encoder_model.predict(input_sequence) at inference time, I get the following error message:

UnknownError:  [_Derived_]  CUDNN_STATUS_BAD_PARAM
in tensorflow/stream_executor/cuda/cuda_dnn.cc(1424): 'cudnnSetRNNDataDescriptor( data_desc.get(), data_type, layout, max_seq_length, batch_size, data_size, seq_lengths_array, (void*)&padding_fill)'
     [[{{node cond/then/_0/CudnnRNNV3}}]]
     [[lstm/StatefulPartitionedCall]] [Op:__inference_keras_scratch_graph_91878]

Function call stack:
keras_scratch_graph -> keras_scratch_graph -> keras_scratch_graph

Here is the training code: (both source_sequences and target_sequences are right-padded sequences and the embedding matrices are pretrained Glove embeddings)

# Define an input sequence and process it.
encoder_inputs = tf.keras.layers.Input(shape=(24,))
encoder_embedding_layer = tf.keras.layers.Embedding(
  VOCABULARY_SIZE_1,
  EMBEDDING_DIMS,
  embeddings_initializer=initializers.Constant(encoder_embedding_matrix),
  mask_zero=True)
encoder_embedding = encoder_embedding_layer(encoder_inputs)

_, state_h, state_c = tf.keras.layers.LSTM(
  EMBEDDING_DIMS,
  implementation=1,
  return_state=True)(encoder_embedding)

encoder_states = [state_h, state_c]

decoder_inputs = tf.keras.layers.Input(shape=(24,))
decoder_embedding_layer = tf.keras.layers.Embedding(
  VOCABULARY_SIZE_2,
  EMBEDDING_DIMS,
  embeddings_initializer=initializers.Constant(decoder_embedding_matrix),
  mask_zero=True)
decoder_embedding = decoder_embedding_layer(decoder_inputs)

decoder_lstm = tf.keras.layers.LSTM(
    EMBEDDING_DIMS, 
    return_sequences=True, 
    return_state=True,
    implementation=1)

decoder_outputs, _, _ = decoder_lstm(decoder_embedding, initial_state=encoder_states)

decoder_dense = tf.keras.layers.Dense(VOCABULARY_SIZE_TITLE, activation='softmax')

output = decoder_dense(decoder_outputs)

model = tf.keras.models.Model([encoder_inputs, decoder_inputs], output)

model.compile(optimizer='rmsprop', loss='sparse_categorical_crossentropy')
model.summary()

model.fit([source_sequences, target_sequences], decoder_target_data,
    batch_size=32,
    epochs=10,
    validation_split=0.0,
    verbose=2)

These are the inference models:

encoder_model = tf.keras.models.Model(encoder_inputs, encoder_states)

decoder_state_input_h = tf.keras.layers.Input(shape=(input_dimension ,))
decoder_state_input_c = tf.keras.layers.Input(shape=(input_dimension ,))

decoder_states_inputs = [decoder_state_input_h, decoder_state_input_c]

decoder_outputs, state_h, state_c = decoder_lstm_layer(
        decoder_embedding_layer , initial_state=decoder_states_inputs)

decoder_states = [state_h, state_c]

decoder_outputs = output_layer(decoder_outputs)
decoder_model = tf.keras.models.Model(
        [decoder_inputs] + decoder_states_inputs,
        [decoder_outputs] + decoder_states)

When I call predict() on the encoder_model, I get CUDNN_STATUS_BAD_PARAM

Inference code (where error gets triggered)

# build the initial state with a right-padded input sequence
#### CUDNN_STATUS_BAD_PARAM is TRIGGERED ON THIS LINE!!! ######## <<<<<<<<<
state = encoder_model.predict(masked_input_sequence)

empty_target_sequence = np.zeros((1,1))
# this signals the Start of sequence
empty_target_sequence[0,0] = titles_word_index[sos_token]

decoder_outputs, h, c = decoder_model.predict([empty_target_sequence] + state)

Things I have tried

create masks explicitly (encoder_embedding_layer.compute_mask()) and add them as parameters every time I call an LSTM layer, for example:

encoder_embedding = encoder_embedding_layer(encoder_inputs)

encoder_mask = encoder_embedding_layer.compute_mask(encoder_inputs)

_, state_h, state_c = tf.keras.layers.LSTM(
  EMBEDDING_DIMS,
  return_state=True)(encoder_embedding,mask=encoder_mask)

not use initializers for the embedding layers to see if the problem was there

P.S.: forcing the training to take place on a CPU makes the error go away but I need to train it on GPU otherwise it would take ages to complete.

P.S.: This seems to be the very same error I have: Masking LSTM: OP_REQUIRES failed at cudnn_rnn_ops.cc:1498 : Unknown: CUDNN_STATUS_BAD_PARAM

P.S.: when I call method supports_masking on model,encoder_model and decoder_model, all of them return False for some reason.

P.S.: Like I said, training is done with no (apparent) errors but if I look at the Jupyter output log on the command line, I can see the following warning message during training:

2019-11-16 19:48:20.144265: W 
tensorflow/core/grappler/optimizers/implementation_selector.cc:310] Skipping optimization due to error while loading function libraries: 
Invalid argument: Functions '__inference___backward_cudnn_lstm_with_fallback_47598_49057' and 
'__inference___backward_cudnn_lstm_with_fallback_47598_49057_specialized_for_StatefulPartitionedCall_1_at___inference_distributed_function_52868'
 both implement 'lstm_d41d5ccb-14be-4a74-b5e8-cc4f63c5bb02' but their signatures do not match.

来源：https://stackoverflow.com/questions/58895694/cudnn-status-bad-param-when-trying-to-perform-inference-on-a-lstm-seq2seq-with-m

标签

python

tensorflow

lstm

tensorflow2.0

tf.keras