Merge pull request #2263 from mozilla/remove-unneeded-ops

Remove use of StridedSlice and update op/kernel deps (Fixes #2179)
This commit is contained in:
Reuben Morais 2019-07-22 13:39:01 +00:00 committed by GitHub
commit b68bfdbb6e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
8 changed files with 29 additions and 29 deletions

View File

@ -141,11 +141,12 @@ def rnn_impl_static_rnn(x, seq_length, previous_state, reuse):
return output, output_state
def create_model(batch_x, seq_length, dropout, reuse=False, previous_state=None, overlap=True, rnn_impl=rnn_impl_lstmblockfusedcell):
def create_model(batch_x, batch_size, seq_length, dropout, reuse=False, previous_state=None, overlap=True, rnn_impl=rnn_impl_lstmblockfusedcell):
layers = {}
# Input shape: [batch_size, n_steps, n_input + 2*n_input*n_context]
batch_size = tf.shape(batch_x)[0]
if not batch_size:
batch_size = tf.shape(batch_x)[0]
# Create overlapping feature windows if needed
if overlap:
@ -206,7 +207,7 @@ def create_model(batch_x, seq_length, dropout, reuse=False, previous_state=None,
# Conveniently, this loss function is implemented in TensorFlow.
# Thus, we can simply make use of this implementation to define our loss.
def calculate_mean_edit_distance_and_loss(iterator, dropout, reuse):
def calculate_mean_edit_distance_and_loss(iterator, dropout, batch_size, reuse):
r'''
This routine beam search decodes a mini-batch and calculates the loss and mean edit distance.
Next to total and average loss it returns the mean edit distance,
@ -221,7 +222,7 @@ def calculate_mean_edit_distance_and_loss(iterator, dropout, reuse):
rnn_impl = rnn_impl_lstmblockfusedcell
# Calculate the logits of the batch
logits, _ = create_model(batch_x, batch_seq_len, dropout, reuse=reuse, rnn_impl=rnn_impl)
logits, _ = create_model(batch_x, batch_size, batch_seq_len, dropout, reuse=reuse, rnn_impl=rnn_impl)
# Compute the CTC loss using TensorFlow's `ctc_loss`
total_loss = tfv1.nn.ctc_loss(labels=batch_y, inputs=logits, sequence_length=batch_seq_len)
@ -266,7 +267,7 @@ def create_optimizer():
# on which all operations within the tower execute.
# For example, all operations of 'tower 0' could execute on the first GPU `tf.device('/gpu:0')`.
def get_tower_results(iterator, optimizer, dropout_rates):
def get_tower_results(iterator, optimizer, dropout_rates, batch_size):
r'''
With this preliminary step out of the way, we can for each GPU introduce a
tower for which's batch we calculate and return the optimization gradients
@ -288,7 +289,7 @@ def get_tower_results(iterator, optimizer, dropout_rates):
with tf.name_scope('tower_%d' % i):
# Calculate the avg_loss and mean_edit_distance and retrieve the decoded
# batch along with the original batch's labels (Y) of this tower
avg_loss = calculate_mean_edit_distance_and_loss(iterator, dropout_rates, reuse=i > 0)
avg_loss = calculate_mean_edit_distance_and_loss(iterator, dropout_rates, batch_size, reuse=i > 0)
# Allow for variables to be re-used by the next tower
tfv1.get_variable_scope().reuse_variables()
@ -435,7 +436,7 @@ def train():
# Building the graph
optimizer = create_optimizer()
gradients, loss = get_tower_results(iterator, optimizer, dropout_rates)
gradients, loss = get_tower_results(iterator, optimizer, dropout_rates, FLAGS.train_batch_size)
# Average tower gradients across GPUs
avg_tower_gradients = average_gradients(gradients)
@ -626,6 +627,7 @@ def create_inference_graph(batch_size=1, n_steps=16, tflite=False):
rnn_impl = rnn_impl_lstmblockfusedcell
logits, layers = create_model(batch_x=input_tensor,
batch_size=batch_size,
seq_length=seq_length if not FLAGS.export_tflite else None,
dropout=no_dropout,
previous_state=previous_state,

View File

@ -57,6 +57,7 @@ def evaluate(test_csvs, create_model, try_loading):
# One rate per layer
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x,
batch_size=FLAGS.test_batch_size,
seq_length=batch_x_len,
dropout=no_dropout)

View File

@ -30,24 +30,24 @@ KENLM_INCLUDES = [
"kenlm",
]
OPENFST_SOURCES_PLATFORM = select({
OPENFST_SOURCES_PLATFORM = select({
"//tensorflow:windows": glob(["ctcdecode/third_party/openfst-1.6.9-win/src/lib/*.cc"]),
"//conditions:default": glob(["ctcdecode/third_party/openfst-1.6.7/src/lib/*.cc"]),
})
DECODER_SOURCES = glob([
"ctcdecode/*.h",
"ctcdecode/*.cpp",
"ctcdecode/*.cpp",
], exclude=["ctcdecode/*_wrap.cpp"]) + OPENFST_SOURCES_PLATFORM + KENLM_SOURCES
OPENFST_INCLUDES_PLATFORM = select({
OPENFST_INCLUDES_PLATFORM = select({
"//tensorflow:windows": ["ctcdecode/third_party/openfst-1.6.9-win/src/include"],
"//conditions:default": ["ctcdecode/third_party/openfst-1.6.7/src/include"],
})
DECODER_INCLUDES = [
".",
"ctcdecode/third_party/ThreadPool",
"ctcdecode/third_party/ThreadPool",
] + OPENFST_INCLUDES_PLATFORM + KENLM_INCLUDES
LINUX_LINKOPTS = [
@ -77,7 +77,7 @@ tf_cc_shared_object(
"tfmodelstate.h",
"tfmodelstate.cc"
]}),
copts = select({
copts = select({
# -fvisibility=hidden is not required on Windows, MSCV hides all declarations by default
"//tensorflow:windows": ["/w"],
# -Wno-sign-compare to silent a lot of warnings from tensorflow itself,
@ -107,28 +107,26 @@ tf_cc_shared_object(
### => Trying to be more fine-grained
### Use bin/ops_in_graph.py to list all the ops used by a frozen graph.
### CPU only build, libdeepspeech.so file size reduced by ~50%
"//tensorflow/core/kernels:dense_update_ops", # Assign (remove once prod model no longer depends on it)
"//tensorflow/core/kernels:spectrogram_op", # AudioSpectrogram
"//tensorflow/core/kernels:bias_op", # BiasAdd
"//tensorflow/contrib/rnn:lstm_ops_kernels", # BlockLSTM
"//tensorflow/core/kernels:cast_op", # Cast
"//tensorflow/core/kernels:concat_op", # ConcatV2
"//tensorflow/core/kernels:constant_op", # Const, Placeholder
"//tensorflow/core/kernels:shape_ops", # ExpandDims, Shape
"//tensorflow/core/kernels:gather_nd_op", # GatherNd
"//tensorflow/core/kernels:identity_op", # Identity
"//tensorflow/core/kernels:immutable_constant_op", # ImmutableConst (used in memmapped models)
"//tensorflow/core/kernels:deepspeech_cwise_ops", # Less, Minimum
"//tensorflow/core/kernels:deepspeech_cwise_ops", # Less, Minimum, Mul
"//tensorflow/core/kernels:matmul_op", # MatMul
"//tensorflow/core/kernels:reduction_ops", # Max
"//tensorflow/core/kernels:mfcc_op", # Mfcc
"//tensorflow/core/kernels:no_op", # NoOp
"//tensorflow/core/kernels:pack_op", # Pack
"//tensorflow/core/kernels:constant_op", # Placeholder
"//tensorflow/core/kernels:sequence_ops", # Range
"//tensorflow/core/kernels:relu_op", # Relu
"//tensorflow/core/kernels:reshape_op", # Reshape
"//tensorflow/core/kernels:shape_ops", # Shape
"//tensorflow/core/kernels:slice_op", # Slice, needed by StridedSlice
"//tensorflow/core/kernels:softmax_op", # Softmax
"//tensorflow/core/kernels:strided_slice_op", # StridedSlice
"//tensorflow/core/kernels:tile_ops", # Tile
"//tensorflow/core/kernels:transpose_op", # Transpose
# And we also need the op libs for these ops used in the model:
@ -139,7 +137,6 @@ tf_cc_shared_object(
"//tensorflow/core:no_op_op_lib", # NoOp
"//tensorflow/core:nn_ops_op_lib", # Relu, Softmax, BiasAdd
# And op libs for these ops brought in by dependencies of dependencies to silence unknown OpKernel warnings:
"//tensorflow/core:state_ops_op_lib", # Assign, AssignSub, AssignAnd, Variable, VariableV2
"//tensorflow/core:bitwise_ops_op_lib", # BitwiseAnd, BitwiseOr, BitwiseXor, LeftShift, RightShift
"//tensorflow/core:random_ops_op_lib", # RandomGammaGrad
"//tensorflow/core:dataset_ops_op_lib", # UnwrapDatasetVariant, WrapDatasetVariant

View File

@ -38,8 +38,8 @@ then:
DEEPSPEECH_ARTIFACTS_ROOT: https://queue.taskcluster.net/v1/task/${linux_arm64_build}/artifacts/public
DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package_cpu}/artifacts/public
DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pbmm
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pbmm
PIP_DEFAULT_TIMEOUT: "60"
PIP_EXTRA_INDEX_URL: "https://lissyx.github.io/deepspeech-python-wheels/"
EXTRA_PYTHON_CONFIGURE_OPTS: "--with-fpectl" # Required by Debian Stretch

View File

@ -43,8 +43,8 @@ then:
DEEPSPEECH_ARTIFACTS_TFLITE_ROOT: https://queue.taskcluster.net/v1/task/${darwin_amd64_tflite}/artifacts/public
DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package_cpu}/artifacts/public
DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pbmm
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pbmm
EXPECTED_TENSORFLOW_VERSION: "${build.tensorflow_git_desc}"
command:

View File

@ -43,8 +43,8 @@ then:
DEEPSPEECH_ARTIFACTS_TFLITE_ROOT: https://queue.taskcluster.net/v1/task/${linux_amd64_tflite}/artifacts/public
DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package_cpu}/artifacts/public
DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pbmm
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pbmm
DECODER_ARTIFACTS_ROOT: https://queue.taskcluster.net/v1/task/${linux_amd64_ctc}/artifacts/public
PIP_DEFAULT_TIMEOUT: "60"
EXPECTED_TENSORFLOW_VERSION: "${build.tensorflow_git_desc}"

View File

@ -38,8 +38,8 @@ then:
DEEPSPEECH_ARTIFACTS_ROOT: https://queue.taskcluster.net/v1/task/${linux_rpi3_build}/artifacts/public
DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package_cpu}/artifacts/public
DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pbmm
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pbmm
PIP_DEFAULT_TIMEOUT: "60"
PIP_EXTRA_INDEX_URL: "https://www.piwheels.org/simple"
EXTRA_PYTHON_CONFIGURE_OPTS: "--with-fpectl" # Required by Raspbian Stretch / PiWheels

View File

@ -45,8 +45,8 @@ then:
DEEPSPEECH_ARTIFACTS_TFLITE_ROOT: https://queue.taskcluster.net/v1/task/${win_amd64_tflite}/artifacts/public
DEEPSPEECH_NODEJS: https://queue.taskcluster.net/v1/task/${node_package_cpu}/artifacts/public
DEEPSPEECH_TEST_MODEL: https://queue.taskcluster.net/v1/task/${training}/artifacts/public/output_graph.pb
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.0/output_graph.pbmm
DEEPSPEECH_PROD_MODEL: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pb
DEEPSPEECH_PROD_MODEL_MMAP: https://github.com/reuben/DeepSpeech/releases/download/v0.6.0-alpha.4/output_graph.pbmm
EXPECTED_TENSORFLOW_VERSION: "${build.tensorflow_git_desc}"
TC_MSYS_VERSION: 'MSYS_NT-6.3'
MSYS: 'winsymlinks:nativestrict'