def test_upsample_get_op_product_graph(self):
""" Test connected graph construction on a graph with upsample op
Need to perform one round of winnowing first to insert the upsample op """
tf.compat.v1.reset_default_graph()
sess = tf.compat.v1.Session()
module_zero_channels_list = []
_ = upsample_model()
init = tf.compat.v1.global_variables_initializer()
sess.run(init)
input_op_names = ['input_1']
output_op_names = ['upsample_model/Softmax']
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name("conv2d_3/Conv2D")
input_channels_to_winnow = [3, 5, 7]
module_mask_pair = (tf_op, input_channels_to_winnow)
module_zero_channels_list.append(module_mask_pair)
new_sess, _ = winnow.winnow_tf_model(sess, input_op_names, output_op_names,
list_of_modules_to_winnow=module_zero_channels_list,
reshape=True, in_place=True, verbose=True)
conn_graph = ConnectedGraph(tf.compat.v1.get_default_graph(), input_op_names, output_op_names)
self.assertEqual(18, len(conn_graph.get_all_ops()))
reduced_bn_1_op = conn_graph.get_op_from_module_name('reduced_batch_normalization_1/cond/FusedBatchNormV3_1')
self.assertTrue(reduced_bn_1_op.output.consumers[0].type == 'Upsample')
new_sess.close()
def test_reducing_vgg16_slim(self):
""" Test reducing vgg16 slim model """
tf.compat.v1.reset_default_graph()
sess = tf.compat.v1.Session()
module_zero_channels_list = []
inp = tf.compat.v1.placeholder(tf.float32, [1, 224, 224, 3])
_ = vgg.vgg_16(inp)
init = tf.compat.v1.global_variables_initializer()
sess.run(init)
input_op_names = ["Placeholder"]
output_op_names = ['vgg_16/fc8/squeezed']
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"vgg_16/fc7/Conv2D")
input_channels_to_winnow = [2, 3, 4]
module_mask_pair = (tf_op, input_channels_to_winnow)
module_zero_channels_list.append(module_mask_pair)
new_sess, ordered_modules_list = winnow.winnow_tf_model(
sess,
input_op_names,
output_op_names,
module_zero_channels_list,
reshape=True,
in_place=True,
verbose=True)
# Save and reload modified graph to allow changes to take effect
new_sess = save_and_load_graph('./saver', new_sess)
# _ = tf.compat.v1.summary.FileWriter('./reduced_graph', new_sess.graph)
with new_sess.graph.as_default():
inp = tf.random.uniform(shape=(1, 224, 224, 3))
inp_array = inp.eval(session=new_sess)
model_input = new_sess.graph.get_tensor_by_name("Placeholder:0")
model_output = new_sess.graph.get_tensor_by_name(
"vgg_16/fc8/squeezed:0")
# run through entire model to check no error is produced
_ = new_sess.run(model_output, feed_dict={model_input: inp_array})
self.assertEqual(4, len(ordered_modules_list))
new_sess.close()
sess.close()
def prune_model(self, layer_db: LayerDatabase,
layer_comp_ratio_list: List[LayerCompRatioPair],
cost_metric: CostMetric, trainer):
# sort all the layers in layer_comp_ratio_list based on occurrence
layer_comp_ratio_list = self._sort_on_occurrence(
layer_db.model, layer_comp_ratio_list)
# Copy the db
comp_layer_db = copy.deepcopy(layer_db)
current_sess = comp_layer_db.model
# Dictionary to map original layer name to list of most recent pruned layer name and output mask.
# Masks remain at the original length and specify channels winnowed after each round of winnower.
orig_layer_name_to_pruned_name_and_mask_dict = {}
# Dictionary to map most recent pruned layer name to the original layer name
pruned_name_to_orig_name_dict = {}
# List to hold original layers to reconstruct
layers_to_reconstruct = []
detached_op_names = set()
# Prune layers which have comp ratios less than 1
for layer_comp_ratio in layer_comp_ratio_list:
orig_layer = layer_db.find_layer_by_name(
layer_comp_ratio.layer.name)
if layer_comp_ratio.comp_ratio is not None and layer_comp_ratio.comp_ratio < 1.0:
# 1) channel selection
prune_indices = self._select_inp_channels(
orig_layer, layer_comp_ratio.comp_ratio)
if not prune_indices:
continue
# 2) Winnowing the model
current_sess, ordered_modules_list = winnow.winnow_tf_model(
current_sess,
self._input_op_names,
self._output_op_names,
[(orig_layer.module, prune_indices)],
reshape=self._allow_custom_downsample_ops,
in_place=True,
verbose=False)
if not ordered_modules_list:
continue
layers_to_reconstruct.append(orig_layer)
# Update dictionaries with new info about pruned ops and new masks
self._update_pruned_ops_and_masks_info(
ordered_modules_list,
orig_layer_name_to_pruned_name_and_mask_dict,
pruned_name_to_orig_name_dict, detached_op_names)
# Save and reload modified graph to allow changes to take effect
# Need to initialize uninitialized variables first since only newly winnowed conv ops are initialized during
# winnow_tf_model, and all other newly winnowed ops are not.
with current_sess.graph.as_default():
initialize_uninitialized_vars(current_sess)
current_sess = save_and_load_graph('./saver', current_sess)
comp_layer_db.update_database(current_sess,
detached_op_names,
update_model=True)
# Perform reconstruction
self._reconstruct_layers(layers_to_reconstruct,
orig_layer_name_to_pruned_name_and_mask_dict,
layer_db, comp_layer_db)
return comp_layer_db
def calculate_compressed_cost(
self, layer_db: LayerDatabase,
layer_comp_ratio_list: List[LayerCompRatioPair]) -> Cost:
"""
Calculate cost of a compressed model given a set of layers and corresponding comp-ratios
:param layer_db: Layer database for original model
:param layer_comp_ratio_list: List of (layer + comp-ratio) pairs
:return: Estimated cost of the compressed model
"""
# sort all the layers in layer_comp_ratio_list based on occurrence
layer_comp_ratio_list = self._sort_on_occurrence(
layer_db.model, layer_comp_ratio_list)
detached_op_names = set()
# Copy the db
comp_layer_db = copy.deepcopy(layer_db)
current_sess = comp_layer_db.model
for layer_comp_ratio in layer_comp_ratio_list:
orig_layer = layer_db.find_layer_by_name(
layer_comp_ratio.layer.name)
comp_ratio = layer_comp_ratio.comp_ratio
if comp_ratio is not None and comp_ratio < 1.0:
# select input channels of conv2d op to winnow
prune_indices = self._select_inp_channels(
orig_layer, comp_ratio)
if not prune_indices:
continue
# Winnow the selected op and modify it's upstream affected ops
current_sess, ordered_modules_list = winnow.winnow_tf_model(
current_sess,
self._input_op_names,
self._output_op_names,
[(orig_layer.module, prune_indices)],
reshape=self._allow_custom_downsample_ops,
in_place=True,
verbose=False)
if not ordered_modules_list:
continue
# Get all the detached op names from updated session graph
for orig_op_name, _, _, _ in ordered_modules_list:
detached_op_names.add(orig_op_name)
# update layer database by excluding the detached ops
comp_layer_db.update_database(current_sess,
detached_op_names,
update_model=False)
# calculate the cost of this model
compressed_model_cost = CostCalculator.compute_model_cost(
comp_layer_db)
# close the session associated with compressed layer database
comp_layer_db.model.close()
return compressed_model_cost
def test_reducing_vgg16(self):
""" This test winnows a VGG16 model"""
tf.compat.v1.reset_default_graph()
sess = tf.compat.v1.Session()
module_zero_channels_list = []
_ = VGG16(weights=None)
init = tf.compat.v1.global_variables_initializer()
sess.run(init)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"block5_conv1/Conv2D")
input_channels_to_winnow = [3, 5, 7]
module_mask_pair = (tf_op, input_channels_to_winnow)
module_zero_channels_list.append(module_mask_pair)
tf_op_2 = tf.compat.v1.get_default_graph().get_operation_by_name(
"block3_conv1/Conv2D")
input_channels_to_winnow_2 = [11, 13, 15, 17]
module_mask_pair_2 = (tf_op_2, input_channels_to_winnow_2)
module_zero_channels_list.append(module_mask_pair_2)
tf_op_3 = tf.compat.v1.get_default_graph().get_operation_by_name(
"block2_conv2/Conv2D")
input_channels_to_winnow_3 = [1, 2, 3, 4, 5]
module_mask_pair_3 = (tf_op_3, input_channels_to_winnow_3)
module_zero_channels_list.append(module_mask_pair_3)
tf_op_4 = tf.compat.v1.get_default_graph().get_operation_by_name(
"block2_conv1/Conv2D")
input_channels_to_winnow_4 = [20, 21, 22, 23]
module_mask_pair_4 = (tf_op_4, input_channels_to_winnow_4)
module_zero_channels_list.append(module_mask_pair_4)
input_op_names = ["input_1"]
output_op_names = ['predictions/Softmax']
new_sess, ordered_modules_list = winnow.winnow_tf_model(
sess,
input_op_names,
output_op_names,
module_zero_channels_list,
reshape=True,
in_place=True,
verbose=True)
# Save and reload modified graph to allow changes to take effect
new_sess = save_and_load_graph('./saver', new_sess)
# uncomment the following to generate tensorboard viewable file
# _ = tf.compat.v1.summary.FileWriter('./reduced_graph', new_sess.graph)
# Check certain weight indices to ensure that weights were reduced correctly
b4c3_kernel = new_sess.graph.get_tensor_by_name(
"block4_conv3/kernel/Read/"
"ReadVariableOp:0").eval(session=new_sess)
red_b4c3_kernel = new_sess.graph.get_tensor_by_name(
"reduced_block4_conv3/kernel/"
"Read/ReadVariableOp:0").eval(session=new_sess)
self.assertEqual(red_b4c3_kernel.shape, (3, 3, 512, 509))
self.assertEqual(red_b4c3_kernel[0][0][0][2], b4c3_kernel[0][0][0][2])
self.assertEqual(np.sum(red_b4c3_kernel[0][0][0][5:]),
np.sum(b4c3_kernel[0][0][0][8:]))
# Test that evaluating the new session uses the newly reduced modules.
# Do so by first evaluating a tensor in a module coming after a set of reduced modules.
# Zero out weights and biases of one of the original unreduced modules preceding the tensor.
# Reevaluate the tensor and expect to see no change, since the original unreduced module should not be used
# anymore.
# Then zero out weights and biases of one of the newly reduced modules.
# Finally reevaluate the same tensor as before. This time, we expect to see the result be zero.
with new_sess.graph.as_default():
inp = tf.random.uniform(shape=(1, 224, 224, 3))
inp_array = inp.eval(session=new_sess)
model_input = new_sess.graph.get_tensor_by_name("input_1:0")
# run through entire model to check no error is produced
model_output = new_sess.graph.get_tensor_by_name(
"predictions/Softmax:0")
_ = new_sess.run(model_output, feed_dict={model_input: inp_array})
self.assertEqual(13, len(ordered_modules_list))
new_sess.close()
sess.close()
def test_reducing_inceptionV3(self):
""" Test module reduction in inceptionV3 """
tf.compat.v1.reset_default_graph()
sess = tf.compat.v1.Session()
module_zero_channels_list = []
_ = InceptionV3(weights=None)
init = tf.compat.v1.global_variables_initializer()
sess.run(init)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"conv2d_12/Conv2D")
input_channels_to_winnow = [0, 1, 64, 128, 224]
module_mask_pair = (tf_op, input_channels_to_winnow)
module_zero_channels_list.append(module_mask_pair)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"conv2d_13/Conv2D")
input_channels_to_winnow_1 = [0, 64, 65, 66, 128, 224]
module_mask_pair = (tf_op, input_channels_to_winnow_1)
module_zero_channels_list.append(module_mask_pair)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"conv2d_15/Conv2D")
input_channels_to_winnow_2 = [0, 64, 128, 129, 130, 131, 224]
module_mask_pair = (tf_op, input_channels_to_winnow_2)
module_zero_channels_list.append(module_mask_pair)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"conv2d_18/Conv2D")
input_channels_to_winnow_3 = [0, 64, 128, 224, 225, 226, 227, 228]
module_mask_pair = (tf_op, input_channels_to_winnow_3)
module_zero_channels_list.append(module_mask_pair)
input_op_names = ["input_1"]
output_op_names = ['predictions/Softmax']
new_sess, ordered_modules_list = winnow.winnow_tf_model(
sess,
input_op_names,
output_op_names,
module_zero_channels_list,
reshape=True,
in_place=True,
verbose=True)
# Save and reload modified graph to allow changes to take effect
# Need to initialize uninitialized variables first since only newly winnowed conv ops are initialized during
# winnow_tf_model, and all other newly winnowed ops are not.
with new_sess.graph.as_default():
initialize_uninitialized_vars(new_sess)
new_sess = save_and_load_graph('./saver', new_sess)
# _ = tf.compat.v1.summary.FileWriter('./reduced_graph', new_sess.graph)
with new_sess.graph.as_default():
inp = tf.random.uniform(shape=(1, 299, 299, 3))
inp_array = inp.eval(session=new_sess)
model_input = new_sess.graph.get_tensor_by_name("input_1:0")
model_output = new_sess.graph.get_tensor_by_name(
"predictions/Softmax:0")
# check that reduced tensor shapes are as expected
reduced_conv2d_12_input = new_sess.graph.get_operation_by_name(
"reduced_conv2d_12/Conv2D").inputs[0]
reduced_conv2d_13_input = new_sess.graph.get_operation_by_name(
"reduced_conv2d_13/Conv2D").inputs[0]
reduced_conv2d_15_input = new_sess.graph.get_operation_by_name(
"reduced_conv2d_15/Conv2D").inputs[0]
reduced_conv2d_18_input = new_sess.graph.get_operation_by_name(
"reduced_conv2d_18/Conv2D").inputs[0]
reduced_conv2d_5_output = new_sess.graph.get_tensor_by_name(
"reduced_conv2d_5/Conv2D:0")
reduced_conv2d_7_output = new_sess.graph.get_tensor_by_name(
"reduced_conv2d_7/Conv2D:0")
reduced_conv2d_10_output = new_sess.graph.get_tensor_by_name(
"reduced_conv2d_10/Conv2D:0")
reduced_conv2d_11_output = new_sess.graph.get_tensor_by_name(
"reduced_conv2d_11/Conv2D:0")
self.assertEqual(251, reduced_conv2d_12_input.shape.as_list()[-1])
self.assertEqual(250, reduced_conv2d_13_input.shape.as_list()[-1])
self.assertEqual(249, reduced_conv2d_15_input.shape.as_list()[-1])
self.assertEqual(248, reduced_conv2d_18_input.shape.as_list()[-1])
self.assertEqual(63, reduced_conv2d_5_output.shape.as_list()[-1])
self.assertEqual(63, reduced_conv2d_7_output.shape.as_list()[-1])
self.assertEqual(95, reduced_conv2d_10_output.shape.as_list()[-1])
self.assertEqual(31, reduced_conv2d_11_output.shape.as_list()[-1])
self.assertEqual(17, len(ordered_modules_list))
# run through entire model to check no error is produced
_ = new_sess.run(model_output, feed_dict={model_input: inp_array})
new_sess.close()
sess.close()
def test_reducing_resnet_50(self):
""" Test module reduction in resnet_50 """
tf.compat.v1.reset_default_graph()
sess = tf.compat.v1.Session()
module_zero_channels_list = []
_ = ResNet50(weights=None)
init = tf.compat.v1.global_variables_initializer()
sess.run(init)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"conv2_block1_1_conv/Conv2D")
input_channels_to_winnow_1 = [3, 5, 7]
module_mask_pair = (tf_op, input_channels_to_winnow_1)
module_zero_channels_list.append(module_mask_pair)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"conv2_block1_0_conv/Conv2D")
input_channels_to_winnow_2 = [3, 5, 7, 8]
module_mask_pair = (tf_op, input_channels_to_winnow_2)
module_zero_channels_list.append(module_mask_pair)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"conv3_block1_1_conv/Conv2D")
input_channels_to_winnow_3 = [3, 5, 7]
module_mask_pair = (tf_op, input_channels_to_winnow_3)
module_zero_channels_list.append(module_mask_pair)
tf_op = tf.compat.v1.get_default_graph().get_operation_by_name(
"conv3_block1_0_conv/Conv2D")
input_channels_to_winnow_4 = [3, 5, 7, 8]
module_mask_pair = (tf_op, input_channels_to_winnow_4)
module_zero_channels_list.append(module_mask_pair)
input_op_names = ["input_1"]
output_op_names = ['probs/Softmax']
new_sess, ordered_modules_list = winnow.winnow_tf_model(
sess,
input_op_names,
output_op_names,
module_zero_channels_list,
reshape=True,
in_place=True,
verbose=True)
# Save and reload modified graph to allow changes to take effect
# Need to initialize uninitialized variables first since only newly winnowed conv ops are initialized during
# winnow_tf_model, and all other newly winnowed ops are not.
with new_sess.graph.as_default():
initialize_uninitialized_vars(new_sess)
new_sess = save_and_load_graph('./saver', new_sess)
# _ = tf.compat.v1.summary.FileWriter('./reduced_graph', new_sess.graph)
with new_sess.graph.as_default():
inp = tf.random.uniform(shape=(1, 224, 224, 3))
inp_array = inp.eval(session=new_sess)
model_input = new_sess.graph.get_tensor_by_name("input_1:0")
model_output = new_sess.graph.get_tensor_by_name("probs/Softmax:0")
# check that reduced tensor shapes are as expected
reduced_conv3_block1_1_input = new_sess.graph.get_operation_by_name(
"reduced_conv3_block1_1_conv/"
"Conv2D").inputs[0]
reduced_conv3_block1_0_input = new_sess.graph.get_operation_by_name(
"reduced_conv3_block1_0_conv/"
"Conv2D").inputs[0]
reduced_conv2_block3_3_output = new_sess.graph.get_tensor_by_name(
"reduced_conv2_block3_3_conv/"
"Conv2D:0")
reduced_conv2_block1_1_input = new_sess.graph.get_operation_by_name(
"reduced_conv2_block1_1_conv/"
"Conv2D").inputs[0]
reduced_conv2_block1_0_input = new_sess.graph.get_operation_by_name(
"reduced_conv2_block1_0_conv/"
"Conv2D").inputs[0]
reduced_conv1_output = new_sess.graph.get_tensor_by_name(
"reduced_conv1_conv/Conv2D:0")
self.assertEqual(253, reduced_conv3_block1_1_input.shape.as_list()[-1])
self.assertEqual(252, reduced_conv3_block1_0_input.shape.as_list()[-1])
self.assertEqual(253,
reduced_conv2_block3_3_output.shape.as_list()[-1])
self.assertEqual(61, reduced_conv2_block1_1_input.shape.as_list()[-1])
self.assertEqual(60, reduced_conv2_block1_0_input.shape.as_list()[-1])
self.assertEqual(61, reduced_conv1_output.shape.as_list()[-1])
# run through entire model to check no error is produced
_ = new_sess.run(model_output, feed_dict={model_input: inp_array})
self.assertEqual(11, len(ordered_modules_list))
new_sess.close()
sess.close()