Skip to content

GitLab

Commit 9302933b authored by Skirmantas Kligys's avatar Skirmantas Kligys Committed by Yongzhe Wang
Browse files

Update for tensorflow.contrib removal. (#8010) 

* Push conflicting earlier changes in files touched.

PiperOrigin-RevId: 268414407
PiperOrigin-RevId: 278006953
PiperOrigin-RevId: 281567252
PiperOrigin-RevId: 288363770

* Removes tf.contrib usage and replaces them with direct imports for lstm object detection.

PiperOrigin-RevId: 286065032

* Internal changes.

PiperOrigin-RevId: 288363770
parent 6b58625d
Hide whitespace changes
Inline Side-by-side
Showing with 118 additions and 89 deletions
research/lstm_object_detection/lstm/lstm_cells.py
View file @ 9302933b
......@@ -12,19 +12,18 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""BottleneckConvLSTMCell implementation."""
import tensorflow as tf
from tensorflow.contrib.framework.python.ops import variables
import lstm_object_detection.lstm.utils as lstm_utils
slim = tf.contrib.slim
_batch_norm = tf.contrib.layers.batch_norm
from tensorflow.contrib import layers as contrib_layers
from tensorflow.contrib import rnn as contrib_rnn
from tensorflow.contrib import slim
from tensorflow.contrib.framework.python.ops import variables as contrib_variables
import lstm_object_detection.lstm.utils as lstm_utils
class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
class BottleneckConvLSTMCell(contrib_rnn.RNNCell):
"""Basic LSTM recurrent network cell using separable convolutions.
The implementation is based on:
......@@ -58,15 +57,15 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
num_units: int, The number of channels in the LSTM cell.
forget_bias: float, The bias added to forget gates (see above).
activation: Activation function of the inner states.
flatten_state: if True, state tensor will be flattened and stored as
a 2-d tensor. Use for exporting the model to tfmini.
flatten_state: if True, state tensor will be flattened and stored as a 2-d
tensor. Use for exporting the model to tfmini.
clip_state: if True, clip state between [-6, 6].
output_bottleneck: if True, the cell bottleneck will be concatenated
to the cell output.
output_bottleneck: if True, the cell bottleneck will be concatenated to
the cell output.
pre_bottleneck: if True, cell assumes that bottlenecking was performing
before the function was called.
visualize_gates: if True, add histogram summaries of all gates
and outputs to tensorboard.
visualize_gates: if True, add histogram summaries of all gates and outputs
to tensorboard.
"""
self._filter_size = list(filter_size)
self._output_size = list(output_size)
......@@ -84,13 +83,12 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
@property
def state_size(self):
return tf.contrib.rnn.LSTMStateTuple(self._output_size + [self._num_units],
self._output_size + [self._num_units])
return contrib_rnn.LSTMStateTuple(self._output_size + [self._num_units],
self._output_size + [self._num_units])
@property
def state_size_flat(self):
return tf.contrib.rnn.LSTMStateTuple([self._param_count],
[self._param_count])
return contrib_rnn.LSTMStateTuple([self._param_count], [self._param_count])
@property
def output_size(self):
......@@ -103,6 +101,7 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
inputs: Input tensor at the current timestep.
state: Tuple of tensors, the state and output at the previous timestep.
scope: Optional scope.
Returns:
A tuple where the first element is the LSTM output and the second is
a LSTMStateTuple of the state at the current timestep.
......@@ -122,7 +121,7 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
if self._pre_bottleneck:
bottleneck = inputs
else:
bottleneck = tf.contrib.layers.separable_conv2d(
bottleneck = contrib_layers.separable_conv2d(
tf.concat([inputs, h], 3),
self._num_units,
self._filter_size,
......@@ -134,7 +133,7 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
if self._viz_gates:
slim.summaries.add_histogram_summary(bottleneck, 'bottleneck')
concat = tf.contrib.layers.separable_conv2d(
concat = contrib_layers.separable_conv2d(
bottleneck,
4 * self._num_units,
self._filter_size,
......@@ -165,7 +164,7 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
new_c = tf.reshape(new_c, [-1, self._param_count])
new_h = tf.reshape(new_h, [-1, self._param_count])
return output, tf.contrib.rnn.LSTMStateTuple(new_c, new_h)
return output, contrib_rnn.LSTMStateTuple(new_c, new_h)
def init_state(self, state_name, batch_size, dtype, learned_state=False):
"""Creates an initial state compatible with this cell.
......@@ -184,7 +183,8 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
self.state_size_flat if self._flatten_state else self.state_size)
# list of 2 zero tensors or variables tensors, depending on if
# learned_state is true
ret_flat = [(variables.model_variable(
# pylint: disable=g-long-ternary,g-complex-comprehension
ret_flat = [(contrib_variables.model_variable(
state_name + str(i),
shape=s,
dtype=dtype,
......@@ -202,8 +202,7 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
]
for s, r in zip(state_size, ret_flat):
r.set_shape([None] + s)
return tf.contrib.framework.nest.pack_sequence_as(
structure=[1, 1], flat_sequence=ret_flat)
return tf.nest.pack_sequence_as(structure=[1, 1], flat_sequence=ret_flat)
def pre_bottleneck(self, inputs, state, input_index):
"""Apply pre-bottleneck projection to inputs.
......@@ -217,6 +216,7 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
state: 4D Tensor with shape [batch_size x width x height x state_size].
input_index: integer index indicating which base features the inputs
correspoding to.
Returns:
inputs: pre-bottlenecked inputs.
Raises:
......@@ -227,12 +227,13 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
# returns the state as a tuple. This should not be an issue since we
# only need to modify state[1] during export, when state should be a
# list.
if not len(inputs.shape) == 4:
if len(inputs.shape) != 4:
raise ValueError('Expect rank 4 feature tensor.')
if not self._flatten_state and not len(state.shape) == 4:
if not self._flatten_state and len(state.shape) != 4:
raise ValueError('Expect rank 4 state tensor.')
if self._flatten_state and not len(state.shape) == 2:
if self._flatten_state and len(state.shape) != 2:
raise ValueError('Expect rank 2 state tensor when flatten_state is set.')
with tf.name_scope(None):
state = tf.identity(state, name='raw_inputs/init_lstm_h')
if self._flatten_state:
......@@ -242,7 +243,7 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
state = tf.reshape(state, [batch_size, height, width, -1])
with tf.variable_scope('conv_lstm_cell', reuse=tf.AUTO_REUSE):
scope_name = 'bottleneck_%d' % input_index
inputs = tf.contrib.layers.separable_conv2d(
inputs = contrib_layers.separable_conv2d(
tf.concat([inputs, state], 3),
self.output_size[-1],
self._filter_size,
......@@ -257,7 +258,7 @@ class BottleneckConvLSTMCell(tf.contrib.rnn.RNNCell):
return inputs
class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
class GroupedConvLSTMCell(contrib_rnn.RNNCell):
"""Basic LSTM recurrent network cell using separable convolutions.
The implementation is based on: https://arxiv.org/abs/1903.10172.
......@@ -297,21 +298,21 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
forget_bias: float, The bias added to forget gates (see above).
activation: Activation function of the inner states.
use_batch_norm: if True, use batch norm after convolution
flatten_state: if True, state tensor will be flattened and stored as
a 2-d tensor. Use for exporting the model to tfmini
flatten_state: if True, state tensor will be flattened and stored as a 2-d
tensor. Use for exporting the model to tfmini
groups: Number of groups to split the state into. Must evenly divide
num_units.
clip_state: if True, clips state between [-6, 6].
scale_state: if True, scales state so that all values are under 6 at all
times.
output_bottleneck: if True, the cell bottleneck will be concatenated
to the cell output.
output_bottleneck: if True, the cell bottleneck will be concatenated to
the cell output.
pre_bottleneck: if True, cell assumes that bottlenecking was performing
before the function was called.
is_quantized: if True, the model is in quantize mode, which requires
quantization friendly concat and separable_conv2d ops.
visualize_gates: if True, add histogram summaries of all gates
and outputs to tensorboard
visualize_gates: if True, add histogram summaries of all gates and outputs
to tensorboard
Raises:
ValueError: when both clip_state and scale_state are enabled.
......@@ -340,13 +341,12 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
@property
def state_size(self):
return tf.contrib.rnn.LSTMStateTuple(self._output_size + [self._num_units],
self._output_size + [self._num_units])
return contrib_rnn.LSTMStateTuple(self._output_size + [self._num_units],
self._output_size + [self._num_units])
@property
def state_size_flat(self):
return tf.contrib.rnn.LSTMStateTuple([self._param_count],
[self._param_count])
return contrib_rnn.LSTMStateTuple([self._param_count], [self._param_count])
@property
def output_size(self):
......@@ -370,6 +370,7 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
inputs: Input tensor at the current timestep.
state: Tuple of tensors, the state at the previous timestep.
scope: Optional scope.
Returns:
A tuple where the first element is the LSTM output and the second is
a LSTMStateTuple of the state at the current timestep.
......@@ -409,7 +410,7 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
if self._use_batch_norm:
b_x = lstm_utils.quantizable_separable_conv2d(
inputs,
self._num_units / self._groups,
self._num_units // self._groups,
self._filter_size,
is_quantized=self._is_quantized,
depth_multiplier=1,
......@@ -418,7 +419,7 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
scope='bottleneck_%d_x' % k)
b_h = lstm_utils.quantizable_separable_conv2d(
h_list[k],
self._num_units / self._groups,
self._num_units // self._groups,
self._filter_size,
is_quantized=self._is_quantized,
depth_multiplier=1,
......@@ -426,10 +427,14 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
normalizer_fn=None,
scope='bottleneck_%d_h' % k)
b_x = slim.batch_norm(
b_x, scale=True, is_training=self._is_training,
b_x,
scale=True,
is_training=self._is_training,
scope='BatchNorm_%d_X' % k)
b_h = slim.batch_norm(
b_h, scale=True, is_training=self._is_training,
b_h,
scale=True,
is_training=self._is_training,
scope='BatchNorm_%d_H' % k)
bottleneck = b_x + b_h
else:
......@@ -439,13 +444,13 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
bottleneck_concat = lstm_utils.quantizable_concat(
[inputs, h_list[k]],
axis=3,
is_training=self._is_training,
is_training=False,
is_quantized=self._is_quantized,
scope='bottleneck_%d/quantized_concat' % k)
bottleneck = lstm_utils.quantizable_separable_conv2d(
bottleneck_concat,
self._num_units / self._groups,
self._num_units // self._groups,
self._filter_size,
is_quantized=self._is_quantized,
depth_multiplier=1,
......@@ -455,7 +460,7 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
concat = lstm_utils.quantizable_separable_conv2d(
bottleneck,
4 * self._num_units / self._groups,
4 * self._num_units // self._groups,
self._filter_size,
is_quantized=self._is_quantized,
depth_multiplier=1,
......@@ -632,7 +637,7 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
with tf.name_scope(None):
new_c = tf.identity(new_c, name='raw_outputs/lstm_c')
new_h = tf.identity(new_h, name='raw_outputs/lstm_h')
states_and_output = tf.contrib.rnn.LSTMStateTuple(new_c, new_h)
states_and_output = contrib_rnn.LSTMStateTuple(new_c, new_h)
return output, states_and_output
......@@ -649,11 +654,12 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
Returns:
ret: the created initial state
"""
state_size = (self.state_size_flat if self._flatten_state
else self.state_size)
state_size = (
self.state_size_flat if self._flatten_state else self.state_size)
# list of 2 zero tensors or variables tensors,
# depending on if learned_state is true
ret_flat = [(variables.model_variable(
# pylint: disable=g-long-ternary,g-complex-comprehension
ret_flat = [(contrib_variables.model_variable(
state_name + str(i),
shape=s,
dtype=dtype,
......@@ -668,8 +674,7 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
for tensor in ret_flat]
for s, r in zip(state_size, ret_flat):
r = tf.reshape(r, [-1] + s)
ret = tf.contrib.framework.nest.pack_sequence_as(
structure=[1, 1], flat_sequence=ret_flat)
ret = tf.nest.pack_sequence_as(structure=[1, 1], flat_sequence=ret_flat)
return ret
def pre_bottleneck(self, inputs, state, input_index):
......@@ -684,6 +689,7 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
state: 4D Tensor with shape [batch_size x width x height x state_size].
input_index: integer index indicating which base features the inputs
correspoding to.
Returns:
inputs: pre-bottlenecked inputs.
Raises:
......@@ -696,12 +702,13 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
# list.
if not self._pre_bottleneck:
raise ValueError('Only applied when pre_bottleneck is set to true.')
if not len(inputs.shape) == 4:
if len(inputs.shape) != 4:
raise ValueError('Expect a rank 4 feature tensor.')
if not self._flatten_state and not len(state.shape) == 4:
if not self._flatten_state and len(state.shape) != 4:
raise ValueError('Expect rank 4 state tensor.')
if self._flatten_state and not len(state.shape) == 2:
if self._flatten_state and len(state.shape) != 2:
raise ValueError('Expect rank 2 state tensor when flatten_state is set.')
with tf.name_scope(None):
state = tf.identity(state, name='raw_inputs/init_lstm_h')
if self._flatten_state:
......@@ -739,6 +746,5 @@ class GroupedConvLSTMCell(tf.contrib.rnn.RNNCell):
# For exporting inference graph, we only mark the first timestep.
with tf.name_scope(None):
inputs = tf.identity(
inputs,
name='raw_outputs/base_endpoint_%d' % (input_index + 1))
inputs, name='raw_outputs/base_endpoint_%d' % (input_index + 1))
return inputs
research/lstm_object_detection/lstm/rnn_decoder_test.py
View file @ 9302933b
......@@ -19,13 +19,15 @@ from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
import numpy as np
import tensorflow as tf
from tensorflow.contrib import layers as contrib_layers
from tensorflow.contrib import rnn as contrib_rnn
from lstm_object_detection.lstm import rnn_decoder
class MockRnnCell(tf.contrib.rnn.RNNCell):
class MockRnnCell(contrib_rnn.RNNCell):
def __init__(self, input_size, num_units):
self._input_size = input_size
......@@ -45,7 +47,7 @@ class MockRnnCell(tf.contrib.rnn.RNNCell):
def pre_bottleneck(self, inputs, state, input_index):
with tf.variable_scope('bottleneck_%d' % input_index, reuse=tf.AUTO_REUSE):
inputs = tf.contrib.layers.separable_conv2d(
inputs = contrib_layers.separable_conv2d(
tf.concat([inputs, state], 3),
self._input_size,
self._filter_size,
......@@ -177,8 +179,6 @@ class MultiInputRnnDecoderTest(tf.test.TestCase):
(outputs, states, inputs_large, inputs_small, initial_state))
outputs_results = results[0]
states_results = results[1]
inputs_large_results = results[2]
inputs_small_results = results[3]
initial_states_results = results[4]
self.assertEqual(
outputs_results[0].shape,
......@@ -189,7 +189,7 @@ class MultiInputRnnDecoderTest(tf.test.TestCase):
(batch_size, width, height, num_units))
# The first step should always update state.
self.assertAllEqual(states_results[0][0],
np.multiply(initial_states_results[0], 2))
np.multiply(initial_states_results[0], 2))
self.assertAllEqual(states_results[0][1], initial_states_results[1])
def test_rnn_decoder_multiple_unroll(self):
......@@ -227,8 +227,6 @@ class MultiInputRnnDecoderTest(tf.test.TestCase):
(outputs, states, inputs_large, inputs_small, initial_state))
outputs_results = results[0]
states_results = results[1]
inputs_large_results = results[2]
inputs_small_results = results[3]
initial_states_results = results[4]
# The first step should always update state.
......@@ -281,8 +279,6 @@ class MultiInputRnnDecoderTest(tf.test.TestCase):
(outputs, states, inputs_large, inputs_small, initial_state))
outputs_results = results[0]
states_results = results[1]
inputs_large_results = results[2]
inputs_small_results = results[3]
initial_states_results = results[4]
for i in range(num_unroll):
......
research/lstm_object_detection/lstm/utils.py
View file @ 9302933b
......@@ -19,6 +19,8 @@ from __future__ import absolute_import
from __future__ import division
import tensorflow as tf
from tensorflow.contrib import framework as contrib_framework
from tensorflow.contrib import layers as contrib_layers
from tensorflow.python.training import moving_averages
......@@ -28,7 +30,7 @@ def _quant_var(
vars_collection=tf.GraphKeys.MOVING_AVERAGE_VARIABLES,
):
"""Create an var for storing the min/max quantization range."""
return tf.contrib.framework.model_variable(
return contrib_framework.model_variable(
name,
shape=[],
initializer=tf.constant_initializer(initializer_val),
......@@ -80,16 +82,21 @@ def quantizable_concat(inputs,
else:
concat_tensors = tf.concat(inputs, axis=axis)
tf.logging.info('concat_tensors: {}'.format(concat_tensors))
# TFLite requires that 0.0 is always in the [min; max] range.
range_min = tf.minimum(
tf.reduce_min(concat_tensors), 0.0, name='SafeQuantRangeMin')
range_max = tf.maximum(
tf.reduce_max(concat_tensors), 0.0, name='SafeQuantRangeMax')
# Otherwise we need to keep track of the moving averages of the min and
# of the elements of the input tensor max.
min_val = moving_averages.assign_moving_average(
min_var,
tf.reduce_min(concat_tensors),
range_min,
ema_decay,
name='AssignMinEma')
max_val = moving_averages.assign_moving_average(
max_var,
tf.reduce_max(concat_tensors),
range_max,
ema_decay,
name='AssignMaxEma')
tf.logging.info('min_val: {}'.format(min_val))
......@@ -144,7 +151,7 @@ def quantizable_separable_conv2d(inputs,
Tensor resulting from concatenation of input tensors
"""
if is_quantized:
outputs = tf.contrib.layers.separable_conv2d(
outputs = contrib_layers.separable_conv2d(
inputs,
None,
kernel_size,
......@@ -154,9 +161,9 @@ def quantizable_separable_conv2d(inputs,
normalizer_fn=None,
biases_initializer=None,
scope=scope)
outputs = tf.contrib.layers.bias_add(
outputs = contrib_layers.bias_add(
outputs, trainable=True, scope='%s_bias' % scope)
outputs = tf.contrib.layers.conv2d(
outputs = contrib_layers.conv2d(
outputs,
num_outputs, [1, 1],
activation_fn=activation_fn,
......@@ -164,7 +171,7 @@ def quantizable_separable_conv2d(inputs,
normalizer_fn=normalizer_fn,
scope=scope)
else:
outputs = tf.contrib.layers.separable_conv2d(
outputs = contrib_layers.separable_conv2d(
inputs,
num_outputs,
kernel_size,
......@@ -202,10 +209,13 @@ def quantize_op(inputs,
min_var = _quant_var('min', default_min)
max_var = _quant_var('max', default_max)
if is_training:
# TFLite requires that 0.0 is always in the [min; max] range.
range_min = tf.minimum(tf.reduce_min(inputs), 0.0, 'SafeQuantRangeMin')
range_max = tf.maximum(tf.reduce_max(inputs), 0.0, 'SafeQuantRangeMax')
min_val = moving_averages.assign_moving_average(
min_var, tf.reduce_min(inputs), ema_decay, name='AssignMinEma')
min_var, range_min, ema_decay, name='AssignMinEma')
max_val = moving_averages.assign_moving_average(
max_var, tf.reduce_max(inputs), ema_decay, name='AssignMaxEma')
max_var, range_max, ema_decay, name='AssignMaxEma')
inputs = tf.fake_quant_with_min_max_vars(inputs, min_val, max_val)
else:
inputs = tf.fake_quant_with_min_max_vars(inputs, min_var, max_var)
......
research/lstm_object_detection/lstm/utils_test.py
View file @ 9302933b
......@@ -89,6 +89,8 @@ class QuantizableUtilsTest(tf.test.TestCase):
op_names = [op.name for op in graph.get_operations()]
self.assertTrue(any('AssignMinEma' in name for name in op_names))
self.assertTrue(any('AssignMaxEma' in name for name in op_names))
self.assertTrue(any('SafeQuantRangeMin' in name for name in op_names))
self.assertTrue(any('SafeQuantRangeMax' in name for name in op_names))
def _check_no_min_max_vars(self, graph):
op_types = [op.type for op in graph.get_operations()]
......@@ -99,6 +101,8 @@ class QuantizableUtilsTest(tf.test.TestCase):
op_names = [op.name for op in graph.get_operations()]
self.assertFalse(any('AssignMinEma' in name for name in op_names))
self.assertFalse(any('AssignMaxEma' in name for name in op_names))
self.assertFalse(any('SafeQuantRangeMin' in name for name in op_names))
self.assertFalse(any('SafeQuantRangeMax' in name for name in op_names))
class QuantizableSeparableConv2dTest(tf.test.TestCase):
......
research/lstm_object_detection/models/lstm_ssd_interleaved_mobilenet_v2_feature_extractor.py
View file @ 9302933b
......@@ -16,6 +16,7 @@
"""LSTDInterleavedFeatureExtractor which interleaves multiple MobileNet V2."""
import tensorflow as tf
from tensorflow.contrib import slim
from tensorflow.python.framework import ops as tf_ops
from lstm_object_detection.lstm import lstm_cells
......@@ -28,8 +29,6 @@ from object_detection.utils import shape_utils
from nets.mobilenet import mobilenet
from nets.mobilenet import mobilenet_v2
slim = tf.contrib.slim
class LSTMSSDInterleavedMobilenetV2FeatureExtractor(
lstm_ssd_meta_arch.LSTMSSDInterleavedFeatureExtractor):
......
research/lstm_object_detection/models/lstm_ssd_interleaved_mobilenet_v2_feature_extractor_test.py
View file @ 9302933b
......@@ -18,12 +18,12 @@
import itertools
import numpy as np
import tensorflow as tf
from tensorflow.contrib import slim
from tensorflow.contrib import training as contrib_training
from lstm_object_detection.models import lstm_ssd_interleaved_mobilenet_v2_feature_extractor
from object_detection.models import ssd_feature_extractor_test
slim = tf.contrib.slim
class LSTMSSDInterleavedMobilenetV2FeatureExtractorTest(
ssd_feature_extractor_test.SsdFeatureExtractorTestBase):
......@@ -233,13 +233,23 @@ class LSTMSSDInterleavedMobilenetV2FeatureExtractorTest(
'lstm_state_step': tf.zeros([1])
}
seq = {'dummy': tf.random_uniform([2, 1, 1, 1])}
stateful_reader1 = tf.contrib.training.SequenceQueueingStateSaver(
batch_size=1, num_unroll=1, input_length=2, input_key='',
input_sequences=seq, input_context={}, initial_states=init_state1,
stateful_reader1 = contrib_training.SequenceQueueingStateSaver(
batch_size=1,
num_unroll=1,
input_length=2,
input_key='',
input_sequences=seq,
input_context={},
initial_states=init_state1,
capacity=1)
stateful_reader2 = tf.contrib.training.SequenceQueueingStateSaver(
batch_size=1, num_unroll=1, input_length=2, input_key='',
input_sequences=seq, input_context={}, initial_states=init_state2,
stateful_reader2 = contrib_training.SequenceQueueingStateSaver(
batch_size=1,
num_unroll=1,
input_length=2,
input_key='',
input_sequences=seq,
input_context={},
initial_states=init_state2,
capacity=1)
image = tf.random_uniform([1, image_height, image_width, 3])
feature_extractor = self._create_feature_extractor(depth_multiplier,
......
research/lstm_object_detection/models/lstm_ssd_mobilenet_v1_feature_extractor.py
View file @ 9302933b
......@@ -16,6 +16,7 @@
"""LSTMSSDFeatureExtractor for MobilenetV1 features."""
import tensorflow as tf
from tensorflow.contrib import slim as contrib_slim
from tensorflow.python.framework import ops as tf_ops
from lstm_object_detection.lstm import lstm_cells
from lstm_object_detection.lstm import rnn_decoder
......@@ -26,7 +27,7 @@ from object_detection.utils import ops
from object_detection.utils import shape_utils
from nets import mobilenet_v1
slim = tf.contrib.slim
slim = contrib_slim
class LSTMSSDMobileNetV1FeatureExtractor(
......
research/lstm_object_detection/models/lstm_ssd_mobilenet_v1_feature_extractor_test.py
View file @ 9302933b
......@@ -17,11 +17,13 @@
import numpy as np
import tensorflow as tf
from tensorflow.contrib import slim as contrib_slim
from tensorflow.contrib import training as contrib_training
from lstm_object_detection.models import lstm_ssd_mobilenet_v1_feature_extractor as feature_extactor
from object_detection.models import ssd_feature_extractor_test
slim = tf.contrib.slim
slim = contrib_slim
class LstmSsdMobilenetV1FeatureExtractorTest(
......@@ -111,7 +113,7 @@ class LstmSsdMobilenetV1FeatureExtractorTest(
'lstm_state_step': tf.zeros([1])
}
seq = {'test': tf.random_uniform([3, 1, 1, 1])}
stateful_reader = tf.contrib.training.SequenceQueueingStateSaver(
stateful_reader = contrib_training.SequenceQueueingStateSaver(
batch_size=1,
num_unroll=1,
input_length=2,
......
research/lstm_object_detection/models/mobilenet_defs.py
View file @ 9302933b
......@@ -15,12 +15,13 @@
"""Definitions for modified MobileNet models used in LSTD."""
import tensorflow as tf
from tensorflow.contrib import slim as contrib_slim
from nets import mobilenet_v1
from nets.mobilenet import conv_blocks as mobilenet_convs
from nets.mobilenet import mobilenet
slim = tf.contrib.slim
slim = contrib_slim
def mobilenet_v1_lite_def(depth_multiplier, low_res=False):
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment