def test_layer_database_with_dynamic_shape(self):
""" test layer database creation with different input shapes"""
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
graph = tf.Graph()
with graph.as_default():
# by default, model will be constructed in default graph
input_placeholder = tf.compat.v1.placeholder(tf.float32, [None, None, None, 3], 'input')
x = tf.keras.layers.Conv2D(8, (2, 2), padding='SAME')(input_placeholder)
x = tf.keras.layers.BatchNormalization(momentum=.3, epsilon=.65)(x)
x = tf.keras.layers.Conv2D(8, (1, 1), padding='SAME', activation=tf.nn.tanh)(x)
x = tf.keras.layers.BatchNormalization(momentum=.4, epsilon=.25)(x)
init = tf.compat.v1.global_variables_initializer()
# create session with graph
sess = tf.compat.v1.Session(graph=graph, config=config)
sess.run(init)
layer_db = LayerDatabase(model=sess, input_shape=(1, 224, 224, 3), working_dir=None, starting_ops=['input'],
ending_ops=['batch_normalization_1/cond/Merge'])
conv1_layer = layer_db.find_layer_by_name('conv2d/Conv2D')
conv2_layer = layer_db.find_layer_by_name('conv2d_1/Conv2D')
self.assertEqual(conv1_layer.output_shape, [1, 8, 224, 224])
self.assertEqual(conv2_layer.output_shape, [1, 8, 224, 224])
layer_db.destroy()
# 2) try with different input shape
# create another session with graph
sess = tf.compat.v1.Session(graph=graph, config=config)
sess.run(init)
batch_size = 32
layer_db = LayerDatabase(model=sess, input_shape=(batch_size, 28, 28, 3), working_dir=None,
starting_ops=['input'], ending_ops=['batch_normalization_1/cond/Merge'])
conv1_layer = layer_db.find_layer_by_name('conv2d/Conv2D')
conv2_layer = layer_db.find_layer_by_name('conv2d_1/Conv2D')
self.assertEqual(conv1_layer.output_shape, [32, 8, 28, 28])
self.assertEqual(conv2_layer.output_shape, [32, 8, 28, 28])
layer_db.destroy()
def test_layer_database_cpu_memory_Leak(self):
# create tf.compat.v1.Session and initialize the weights and biases with zeros
import psutil
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = ResNet50(weights=None)
init = tf.compat.v1.global_variables_initializer()
sess.run(init)
layer_db = LayerDatabase(model=sess, input_shape=(1, 28, 28, 1), working_dir=None)
mem_before = psutil.virtual_memory().available
for i in range(5):
copy_layer_db = copy.deepcopy(layer_db)
copy_layer_db.model.close()
mem_after = psutil.virtual_memory().available
memory_consumed = (mem_before - mem_after) / (1024 * 1024)
print('Memory consumed in MB:', memory_consumed, 'for iterations : ', 5)
self.assertTrue(memory_consumed < 200)
layer_db.model.close()
def test_layer_database_working_dir(self):
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
meta_path = str('./temp_working_dir/')
if not os.path.exists(meta_path):
os.mkdir(meta_path)
layer_db = LayerDatabase(model=sess, input_shape=(1, 28, 28, 1), working_dir=meta_path)
copy_layer_db = copy.deepcopy(layer_db)
shutil.rmtree(meta_path)
tf.compat.v1.reset_default_graph()
layer_db.model.close()
copy_layer_db.model.close()
def test_total_model_cost(self):
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
layer_database = LayerDatabase(model=sess,
input_shape=(1, 28, 28, 1),
working_dir=None)
cost_calc = cc.CostCalculator()
network_cost = cost_calc.compute_model_cost(layer_database)
self.assertEqual(800 + 51200 + 3211264 + 10240, network_cost.memory)
self.assertEqual(627200 + 10035200 + 3211264 + 10240, network_cost.mac)
tf.compat.v1.reset_default_graph()
sess.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def test_prune_model_2_layers(self):
""" Punning two layers with 0.5 comp-ratio in MNIST"""
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
# Create a layer database
orig_layer_db = LayerDatabase(model=sess,
input_shape=(1, 28, 28, 1),
working_dir=None)
conv1 = orig_layer_db.find_layer_by_name('conv2d/Conv2D')
conv2 = orig_layer_db.find_layer_by_name('conv2d_1/Conv2D')
layer_comp_ratio_list = [
LayerCompRatioPair(conv1, Decimal(0.5)),
LayerCompRatioPair(conv2, Decimal(0.5))
]
spatial_svd_pruner = SpatialSvdPruner()
comp_layer_db = spatial_svd_pruner.prune_model(orig_layer_db,
layer_comp_ratio_list,
CostMetric.mac,
trainer=None)
conv1_a = comp_layer_db.find_layer_by_name('conv2d_a/Conv2D')
conv1_b = comp_layer_db.find_layer_by_name('conv2d_b/Conv2D')
# Weights shape [kh, kw, Nic, Noc]
self.assertEqual([5, 1, 1, 2],
conv1_a.module.inputs[1].get_shape().as_list())
self.assertEqual([1, 5, 2, 32],
conv1_b.module.inputs[1].get_shape().as_list())
conv2_a = comp_layer_db.find_layer_by_name('conv2d_1_a/Conv2D')
conv2_b = comp_layer_db.find_layer_by_name('conv2d_1_b/Conv2D')
self.assertEqual([5, 1, 32, 53],
conv2_a.module.inputs[1].get_shape().as_list())
self.assertEqual([1, 5, 53, 64],
conv2_b.module.inputs[1].get_shape().as_list())
for layer in comp_layer_db:
print("Layer: " + layer.name)
print(" Module: " + str(layer.module.name))
tf.compat.v1.reset_default_graph()
sess.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def test_prune_model_tf_slim(self):
""" Punning a model with tf slim api """
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
x = tf.compat.v1.placeholder(tf.float32, [1, 32, 32, 3])
_ = tf_slim_basic_model(x)
sess.run(tf.compat.v1.global_variables_initializer())
conn_graph_orig = ConnectedGraph(sess.graph, ['Placeholder'],
['tf_slim_model/Softmax'])
num_ops_orig = len(conn_graph_orig.get_all_ops())
# Create a layer database
orig_layer_db = LayerDatabase(model=sess,
input_shape=(1, 32, 32, 3),
working_dir=None)
conv1 = orig_layer_db.find_layer_by_name('Conv_1/Conv2D')
conv1_bias = BiasUtils.get_bias_as_numpy_data(orig_layer_db.model,
conv1.module)
layer_comp_ratio_list = [LayerCompRatioPair(conv1, Decimal(0.5))]
spatial_svd_pruner = SpatialSvdPruner()
comp_layer_db = spatial_svd_pruner.prune_model(orig_layer_db,
layer_comp_ratio_list,
CostMetric.mac,
trainer=None)
# Check that svd added these ops
_ = comp_layer_db.model.graph.get_operation_by_name('Conv_1_a/Conv2D')
_ = comp_layer_db.model.graph.get_operation_by_name('Conv_1_b/Conv2D')
conn_graph_new = ConnectedGraph(comp_layer_db.model.graph,
['Placeholder'],
['tf_slim_model/Softmax'])
num_ops_new = len(conn_graph_new.get_all_ops())
self.assertEqual(num_ops_orig + 1, num_ops_new)
bias_add_op = comp_layer_db.model.graph.get_operation_by_name(
'Conv_1_b/BiasAdd')
conv_1_b_op = comp_layer_db.model.graph.get_operation_by_name(
'Conv_1_b/Conv2D')
self.assertEqual(
conn_graph_new._module_identifier.get_op_info(bias_add_op),
conn_graph_new._module_identifier.get_op_info(conv_1_b_op))
self.assertTrue(
np.array_equal(
conv1_bias,
BiasUtils.get_bias_as_numpy_data(comp_layer_db.model,
conv_1_b_op)))
def test_layer_database_with_mnist(self):
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
# initialize the weights and biases with appropriate initializer
sess.run(tf.compat.v1.global_variables_initializer())
layer_db = LayerDatabase(model=sess, input_shape=(1, 28, 28, 1), working_dir=None)
layers = list(layer_db._compressible_layers.values())
# check output shapes
self.assertEqual(layers[0].output_shape, [None, 32, 28, 28])
self.assertEqual(layers[1].output_shape, [None, 64, 14, 14])
self.assertEqual(layers[2].output_shape, [None, 1024, 1, 1])
self.assertEqual(layers[3].output_shape, [None, 10, 1, 1])
# check weight shapes
# layer weight_shape is in common format [Noc, Nic, k_h, k_w]
self.assertEqual(layers[0].weight_shape, [32, 1, 5, 5])
self.assertEqual(layers[1].weight_shape, [64, 32, 5, 5])
self.assertEqual(layers[2].weight_shape, [1024, 3136])
self.assertEqual(layers[3].weight_shape, [10, 1024])
# tensorflow weights are stored in [k_h, k_w, Nic, Noc]
self.assertEqual(layers[0].module.inputs[1].eval(session=layer_db.model).shape, (5, 5, 1, 32))
self.assertEqual(layers[1].module.inputs[1].eval(session=layer_db.model).shape, (5, 5, 32, 64))
self.assertEqual(layers[2].module.inputs[1].eval(session=layer_db.model).shape, (3136, 1024))
self.assertEqual(layers[3].module.inputs[1].eval(session=layer_db.model).shape, (1024, 10))
tf.compat.v1.reset_default_graph()
sess.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def test_calculate_spatial_svd_cost_all_layers(self):
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
layer_database = LayerDatabase(model=sess,
input_shape=(1, 28, 28, 1),
working_dir=None)
# Compress all layers by 50%
# Create a list of tuples of (layer, comp_ratio)
layer_ratio_list = []
for layer in layer_database:
if layer.module.type == 'Conv2D':
layer_ratio_list.append(LayerCompRatioPair(
layer, Decimal(0.5)))
else:
layer_ratio_list.append(LayerCompRatioPair(layer, None))
compressed_cost = cc.SpatialSvdCostCalculator.calculate_compressed_cost(
layer_database, layer_ratio_list, CostMetric.mac)
self.assertEqual(8466464, compressed_cost.mac)
tf.compat.v1.reset_default_graph()
sess.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def test_prune_conv_no_bias(self):
""" Test spatial svd on a conv layer with no bias """
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
inputs = tf.keras.Input(shape=(
32,
32,
3,
))
x = tf.keras.layers.Conv2D(32, (3, 3), use_bias=False)(inputs)
_ = tf.keras.layers.Flatten()(x)
sess.run(tf.compat.v1.global_variables_initializer())
# Create a layer database
orig_layer_db = LayerDatabase(model=sess,
input_shape=(1, 32, 32, 3),
working_dir=None)
conv_op = orig_layer_db.find_layer_by_name('conv2d/Conv2D')
layer_comp_ratio_list = [LayerCompRatioPair(conv_op, Decimal(0.5))]
spatial_svd_pruner = SpatialSvdPruner()
comp_layer_db = spatial_svd_pruner.prune_model(orig_layer_db,
layer_comp_ratio_list,
CostMetric.mac,
trainer=None)
# Check that svd added these ops
_ = comp_layer_db.model.graph.get_operation_by_name('conv2d_a/Conv2D')
conv2d_b_op = comp_layer_db.model.graph.get_operation_by_name(
'conv2d_b/Conv2D')
reshape_op = comp_layer_db.model.graph.get_operation_by_name(
'flatten/Reshape')
self.assertEqual(conv2d_b_op, reshape_op.inputs[0].op)
def test_layer_database_destroy(self):
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
init = tf.compat.v1.global_variables_initializer()
sess.run(init)
layer_db = LayerDatabase(model=sess, input_shape=(1, 28, 28, 1), working_dir=None)
layer_db.destroy()
self.assertRaises(RuntimeError, lambda: sess.run(init))
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def test_per_layer_eval_scores(self):
pruner = unittest.mock.MagicMock()
eval_func = unittest.mock.MagicMock()
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
# Create a layer database
layer_db = LayerDatabase(model=sess,
input_shape=(1, 28, 28, 1),
working_dir=None)
layer1 = layer_db.find_layer_by_name('conv2d/Conv2D')
layer_db.mark_picked_layers([layer1])
eval_func.side_effect = [90, 80, 70, 60, 50, 40, 30, 20, 10]
url, process = start_bokeh_server_session(8006)
bokeh_session = BokehServerSession(url=url, session_id="compression")
# Instantiate child
greedy_algo = comp_ratio_select.GreedyCompRatioSelectAlgo(
layer_db=layer_db,
pruner=pruner,
cost_calculator=SpatialSvdCostCalculator(),
eval_func=eval_func,
eval_iterations=20,
cost_metric=CostMetric.mac,
target_comp_ratio=0.5,
num_candidates=10,
use_monotonic_fit=True,
saved_eval_scores_dict=None,
comp_ratio_rounding_algo=None,
use_cuda=False,
bokeh_session=bokeh_session)
progress_bar = ProgressBar(1,
"eval scores",
"green",
bokeh_session=bokeh_session)
data_table = DataTable(num_columns=3,
num_rows=1,
column_names=[
'0.1', '0.2', '0.3', '0.4', '0.5', '0.6',
'0.7', '0.8', '0.9'
],
row_index_names=[layer1.name],
bokeh_session=bokeh_session)
pruner.prune_model.return_value = layer_db
eval_dict = greedy_algo._compute_layerwise_eval_score_per_comp_ratio_candidate(
data_table, progress_bar, layer1)
self.assertEqual(90, eval_dict[Decimal('0.1')])
tf.compat.v1.reset_default_graph()
sess.close()
bokeh_session.server_session.close("test complete")
os.killpg(os.getpgid(process.pid), signal.SIGTERM)
def test_select_per_layer_comp_ratios_with_spatial_svd_pruner(self):
pruner = SpatialSvdPruner()
eval_func = unittest.mock.MagicMock()
rounding_algo = unittest.mock.MagicMock()
eval_func.side_effect = [
10, 20, 30, 40, 50, 60, 70, 80, 90, 11, 21, 31, 35, 40, 45, 50, 55,
60
]
rounding_algo.round.side_effect = [
0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.1, 0.2, 0.3, 0.4,
0.5, 0.6, 0.7, 0.8, 0.9
]
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
# Create a layer database
layer_db = LayerDatabase(model=sess,
input_shape=(1, 28, 28, 1),
working_dir=None)
selected_layers = [
layer for layer in layer_db if layer.module.type == 'Conv2D'
]
layer_db.mark_picked_layers(selected_layers)
url, process = start_bokeh_server_session(8006)
bokeh_session = BokehServerSession(url=url, session_id="compression")
# Instantiate child
greedy_algo = comp_ratio_select.GreedyCompRatioSelectAlgo(
layer_db=layer_db,
pruner=pruner,
cost_calculator=SpatialSvdCostCalculator(),
eval_func=eval_func,
eval_iterations=20,
cost_metric=CostMetric.mac,
target_comp_ratio=Decimal(0.4),
num_candidates=10,
use_monotonic_fit=True,
saved_eval_scores_dict=None,
comp_ratio_rounding_algo=rounding_algo,
use_cuda=False,
bokeh_session=bokeh_session)
layer_comp_ratio_list, stats = greedy_algo.select_per_layer_comp_ratios(
)
original_cost = SpatialSvdCostCalculator.compute_model_cost(layer_db)
for layer in layer_db:
if layer not in selected_layers:
layer_comp_ratio_list.append(LayerCompRatioPair(layer, None))
compressed_cost = SpatialSvdCostCalculator.calculate_compressed_cost(
layer_db, layer_comp_ratio_list, CostMetric.mac)
actual_compression_ratio = compressed_cost.mac / original_cost.mac
self.assertTrue(
math.isclose(Decimal(0.3), actual_compression_ratio, abs_tol=0.8))
print('\n')
for pair in layer_comp_ratio_list:
print(pair)
tf.compat.v1.reset_default_graph()
sess.close()
bokeh_session.server_session.close("test complete")
os.killpg(os.getpgid(process.pid), signal.SIGTERM)
def test_eval_scores_with_spatial_svd_pruner(self):
pruner = SpatialSvdPruner()
eval_func = unittest.mock.MagicMock()
eval_func.side_effect = [
90, 80, 70, 60, 50, 40, 30, 20, 10, 91, 81, 71, 61, 51, 41, 31, 21,
11
]
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
# Create a layer database
layer_db = LayerDatabase(model=sess,
input_shape=(1, 28, 28, 1),
working_dir=None)
layer1 = layer_db.find_layer_by_name('conv2d/Conv2D')
layer2 = layer_db.find_layer_by_name('conv2d_1/Conv2D')
layer_db.mark_picked_layers([layer1, layer2])
url, process = start_bokeh_server_session(8006)
bokeh_session = BokehServerSession(url=url, session_id="compression")
# Instantiate child
greedy_algo = comp_ratio_select.GreedyCompRatioSelectAlgo(
layer_db=layer_db,
pruner=pruner,
cost_calculator=SpatialSvdCostCalculator(),
eval_func=eval_func,
eval_iterations=20,
cost_metric=CostMetric.mac,
target_comp_ratio=0.5,
num_candidates=10,
use_monotonic_fit=True,
saved_eval_scores_dict=None,
comp_ratio_rounding_algo=None,
use_cuda=False,
bokeh_session=bokeh_session)
dict = greedy_algo._compute_eval_scores_for_all_comp_ratio_candidates()
print()
print(dict)
self.assertEqual(90, dict['conv2d/Conv2D'][Decimal('0.1')])
self.assertEqual(51, dict['conv2d_1/Conv2D'][Decimal('0.5')])
self.assertEqual(21, dict['conv2d_1/Conv2D'][Decimal('0.8')])
tf.compat.v1.reset_default_graph()
sess.close()
bokeh_session.server_session.close("test complete")
os.killpg(os.getpgid(process.pid), signal.SIGTERM)
def test_calculate_channel_pruning_cost_all_layers(self):
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# model will be constructed in default graph
_ = mnist(data_format='channels_last')
# initialize the weights and biases with appropriate initializer
sess.run(tf.compat.v1.global_variables_initializer())
meta_path = str('./temp_working_dir/')
if not os.path.exists(meta_path):
os.mkdir(meta_path)
layer_db = LayerDatabase(model=sess,
input_shape=(1, 28, 28, 1),
working_dir=meta_path)
# Compress all layers by 50%
# Create a list of tuples of (layer, comp_ratio)
layer_ratio_list = []
# Unfortunately in mnist we can only input channel prune conv2d_1/Conv2D
for layer in layer_db:
if layer.module.name == 'conv2d_1/Conv2D':
layer_ratio_list.append(
LayerCompRatioPair(layer, Decimal('0.5')))
else:
layer_ratio_list.append(LayerCompRatioPair(layer, None))
inp_op_names = ['reshape_input']
output_op_names = ['dense_1/BiasAdd']
data_set = unittest.mock.MagicMock()
batch_size = unittest.mock.MagicMock()
num_reconstruction_samples = unittest.mock.MagicMock()
pruner = InputChannelPruner(
input_op_names=inp_op_names,
output_op_names=output_op_names,
data_set=data_set,
batch_size=batch_size,
num_reconstruction_samples=num_reconstruction_samples,
allow_custom_downsample_ops=True)
cost_calculator = ChannelPruningCostCalculator(pruner)
compressed_cost = cost_calculator.calculate_compressed_cost(
layer_db, layer_ratio_list, CostMetric.mac)
self.assertEqual(8552704, compressed_cost.mac)
self.assertEqual(3247504, compressed_cost.memory)
# delete the meta and the checkpoint files
shutil.rmtree(meta_path)
layer_db.model.close()
def test_calculate_channel_pruning_cost_two_layers(self):
"""
test compressed model cost using two layers
:return:
"""
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# model will be constructed in default graph
test_models.single_residual()
init = tf.compat.v1.global_variables_initializer()
# initialize the weights and biases with appropriate initializer
sess.run(init)
meta_path = str('./temp_working_dir/')
if not os.path.exists(meta_path):
os.mkdir(meta_path)
layer_db = LayerDatabase(model=sess,
input_shape=None,
working_dir=meta_path)
# Create a list of tuples of (layer, comp_ratio)
layer_ratio_list = []
layer_names = ['conv2d_2/Conv2D', 'conv2d_3/Conv2D']
for layer in layer_db:
if layer.module.name in layer_names:
layer_ratio_list.append(LayerCompRatioPair(layer, 0.5))
else:
layer_ratio_list.append(LayerCompRatioPair(layer, None))
input_op_names = ['input_1']
output_op_names = ['single_residual/Softmax']
data_set = unittest.mock.MagicMock()
batch_size = unittest.mock.MagicMock()
num_reconstruction_samples = unittest.mock.MagicMock()
pruner = InputChannelPruner(
input_op_names=input_op_names,
output_op_names=output_op_names,
data_set=data_set,
batch_size=batch_size,
num_reconstruction_samples=num_reconstruction_samples,
allow_custom_downsample_ops=True)
cost_calculator = ChannelPruningCostCalculator(pruner)
compressed_cost = cost_calculator.calculate_compressed_cost(
layer_db, layer_ratio_list, CostMetric.mac)
self.assertEqual(108544, compressed_cost.mac)
self.assertEqual(1264, compressed_cost.memory)
# delete the meta and the checkpoint files
shutil.rmtree(meta_path)
layer_db.model.close()
def test_prune_layer(self):
""" Pruning single layer with 0.5 comp-ratio in MNIST"""
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
# Create a layer database
orig_layer_db = LayerDatabase(model=sess,
input_shape=(1, 28, 28, 1),
working_dir=None)
# Copy the db
comp_layer_db = copy.deepcopy(orig_layer_db)
conv1 = comp_layer_db.find_layer_by_name('conv2d/Conv2D')
# before the splitting
bias_op = get_succeeding_bias_op(conv1.module)
for consumer in bias_op.outputs[0].consumers():
self.assertEqual(consumer.name, "conv2d/Relu")
spatial_svd_pruner = SpatialSvdPruner()
spatial_svd_pruner._prune_layer(orig_layer_db, comp_layer_db, conv1,
0.5, CostMetric.mac)
conv2d_a_op = comp_layer_db.model.graph.get_operation_by_name(
'conv2d_a/Conv2D')
conv2d_b_op = comp_layer_db.model.graph.get_operation_by_name(
'conv2d_b/Conv2D')
conv2d_a_weight = WeightTensorUtils.get_tensor_as_numpy_data(
comp_layer_db.model, conv2d_a_op)
conv2d_b_weight = WeightTensorUtils.get_tensor_as_numpy_data(
comp_layer_db.model, conv2d_b_op)
conv1_a = comp_layer_db.find_layer_by_name('conv2d_a/Conv2D')
conv1_b = comp_layer_db.find_layer_by_name('conv2d_b/Conv2D')
# [Noc, Nic, kh, kw]
self.assertEqual([2, 1, 5, 1], conv1_a.weight_shape)
self.assertEqual([32, 2, 1, 5], conv1_b.weight_shape)
# after the splitting
bias_op = get_succeeding_bias_op(conv1_b.module)
for consumer in bias_op.outputs[0].consumers():
self.assertEqual(consumer.name, "conv2d/Relu")
# original layer should be not there in the database
self.assertRaises(
KeyError,
lambda: comp_layer_db.find_layer_by_name('conv2d/Conv2D'))
# check if the layer replacement is done correctly
orig_conv_op = comp_layer_db.model.graph.get_operation_by_name(
'conv2d/Conv2D')
bias_op = get_succeeding_bias_op(orig_conv_op)
# consumers list should be empty
consumers = [consumer for consumer in bias_op.outputs[0].consumers()]
self.assertEqual(len(consumers), 0)
# Check that weights loaded during svd pruning will stick after save and load
new_sess = save_and_load_graph('./temp_meta/', comp_layer_db.model)
conv2d_a_op = comp_layer_db.model.graph.get_operation_by_name(
'conv2d_a/Conv2D')
conv2d_b_op = comp_layer_db.model.graph.get_operation_by_name(
'conv2d_b/Conv2D')
conv2d_a_weight_after_save_load = WeightTensorUtils.get_tensor_as_numpy_data(
comp_layer_db.model, conv2d_a_op)
conv2d_b_weight_after_save_load = WeightTensorUtils.get_tensor_as_numpy_data(
comp_layer_db.model, conv2d_b_op)
self.assertTrue(
np.array_equal(conv2d_a_weight, conv2d_a_weight_after_save_load))
self.assertTrue(
np.array_equal(conv2d_b_weight, conv2d_b_weight_after_save_load))
tf.compat.v1.reset_default_graph()
sess.close()
new_sess.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def test_sort_on_occurrence(self):
"""
Test sorting of ops based on occurrence
"""
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Winnow,
logging.INFO)
tf.compat.v1.reset_default_graph()
orig_g = tf.Graph()
with orig_g.as_default():
_ = VGG16(weights=None,
input_shape=(224, 224, 3),
include_top=False)
orig_init = tf.compat.v1.global_variables_initializer()
# create sess with graph
orig_sess = tf.compat.v1.Session(graph=orig_g)
orig_sess.run(orig_init)
# create layer database
layer_db = LayerDatabase(model=orig_sess,
input_shape=(1, 224, 224, 3),
working_dir=None)
block1_conv2 = layer_db.model.graph.get_operation_by_name(
'block1_conv2/Conv2D')
block2_conv1 = layer_db.model.graph.get_operation_by_name(
'block2_conv1/Conv2D')
block2_conv2 = layer_db.model.graph.get_operation_by_name(
'block2_conv2/Conv2D')
block5_conv3 = layer_db.model.graph.get_operation_by_name(
'block5_conv3/Conv2D')
# output shape in NCHW format
block1_conv2_output_shape = block1_conv2.outputs[0].shape
block2_conv1_output_shape = block2_conv1.outputs[0].shape
block2_conv2_output_shape = block2_conv2.outputs[0].shape
block5_conv3_output_shape = block5_conv3.outputs[0].shape
# keeping compression ratio = None for all layers
layer_comp_ratio_list = [
LayerCompRatioPair(
Layer(model=layer_db.model,
op=block5_conv3,
output_shape=block5_conv3_output_shape), None),
LayerCompRatioPair(
Layer(model=layer_db.model,
op=block2_conv2,
output_shape=block2_conv2_output_shape), None),
LayerCompRatioPair(
Layer(model=layer_db.model,
op=block1_conv2,
output_shape=block1_conv2_output_shape), None),
LayerCompRatioPair(
Layer(model=layer_db.model,
op=block2_conv1,
output_shape=block2_conv1_output_shape), None)
]
input_op_names = ['input_1']
output_op_names = ['block5_pool/MaxPool']
dataset = unittest.mock.MagicMock()
batch_size = unittest.mock.MagicMock()
num_reconstruction_samples = unittest.mock.MagicMock()
cp = InputChannelPruner(
input_op_names=input_op_names,
output_op_names=output_op_names,
data_set=dataset,
batch_size=batch_size,
num_reconstruction_samples=num_reconstruction_samples,
allow_custom_downsample_ops=True)
sorted_layer_comp_ratio_list = cp._sort_on_occurrence(
layer_db.model, layer_comp_ratio_list)
self.assertEqual(sorted_layer_comp_ratio_list[0].layer.module,
block1_conv2)
self.assertEqual(sorted_layer_comp_ratio_list[1].layer.module,
block2_conv1)
self.assertEqual(sorted_layer_comp_ratio_list[2].layer.module,
block2_conv2)
self.assertEqual(sorted_layer_comp_ratio_list[3].layer.module,
block5_conv3)
self.assertEqual(len(sorted_layer_comp_ratio_list), 4)
layer_db.model.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def test_prune_model(self):
"""
Test end-to-end prune_model with VGG16-imagenet
"""
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Winnow,
logging.INFO)
tf.compat.v1.reset_default_graph()
batch_size = 1
input_data = np.random.rand(100, 224, 224, 3)
dataset = tf.data.Dataset.from_tensor_slices(input_data)
dataset = dataset.batch(batch_size=batch_size)
orig_g = tf.Graph()
with orig_g.as_default():
_ = VGG16(weights=None,
input_shape=(224, 224, 3),
include_top=False)
orig_init = tf.compat.v1.global_variables_initializer()
input_op_names = ['input_1']
output_op_names = ['block5_pool/MaxPool']
# create sess with graph
orig_sess = tf.compat.v1.Session(graph=orig_g)
# initialize all the variables in VGG16
orig_sess.run(orig_init)
# create layer database
layer_db = LayerDatabase(model=orig_sess,
input_shape=(1, 224, 224, 3),
working_dir=None)
block1_conv2 = layer_db.model.graph.get_operation_by_name(
'block1_conv2/Conv2D')
block2_conv1 = layer_db.model.graph.get_operation_by_name(
'block2_conv1/Conv2D')
block2_conv2 = layer_db.model.graph.get_operation_by_name(
'block2_conv2/Conv2D')
# output shape in NCHW format
block1_conv2_output_shape = block1_conv2.outputs[0].shape
block2_conv1_output_shape = block2_conv1.outputs[0].shape
block2_conv2_output_shape = block2_conv2.outputs[0].shape
# keeping compression ratio = 0.5 for all layers
layer_comp_ratio_list = [
LayerCompRatioPair(
Layer(model=layer_db.model,
op=block1_conv2,
output_shape=block1_conv2_output_shape), 0.5),
LayerCompRatioPair(
Layer(model=layer_db.model,
op=block2_conv1,
output_shape=block2_conv1_output_shape), 0.5),
LayerCompRatioPair(
Layer(model=layer_db.model,
op=block2_conv2,
output_shape=block2_conv2_output_shape), 0.5)
]
cp = InputChannelPruner(input_op_names=input_op_names,
output_op_names=output_op_names,
data_set=dataset,
batch_size=batch_size,
num_reconstruction_samples=20,
allow_custom_downsample_ops=True)
comp_layer_db = cp.prune_model(
layer_db=layer_db,
layer_comp_ratio_list=layer_comp_ratio_list,
cost_metric=CostMetric.mac,
trainer=None)
pruned_block1_conv2 = comp_layer_db.find_layer_by_name(
'reduced_reduced_block1_conv2/Conv2D')
pruned_block2_conv1 = comp_layer_db.find_layer_by_name(
'reduced_reduced_block2_conv1/Conv2D')
pruned_block2_conv2 = comp_layer_db.find_layer_by_name(
'reduced_block2_conv2/Conv2D')
# input channels = 64 * 0.5 = 32
# output channels = 64 * 0.5 = 32
self.assertEqual(pruned_block1_conv2.weight_shape[1], 32)
self.assertEqual(pruned_block1_conv2.weight_shape[0], 32)
# input channels = 64 * 0.5 = 32
# output channels = 128 * 0.5 = 64
self.assertEqual(pruned_block2_conv1.weight_shape[1], 32)
self.assertEqual(pruned_block2_conv1.weight_shape[0], 64)
# input channels = 128 * 0.5 = 64
# output channels = 128
self.assertEqual(pruned_block2_conv2.weight_shape[1], 64)
self.assertEqual(pruned_block2_conv2.weight_shape[0], 128)
layer_db.model.close()
comp_layer_db.model.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def test_datasampling_and_reconstruction(self):
"""
Test data sampling and reconstruction logic
"""
tf.compat.v1.reset_default_graph()
batch_size = 1
input_data = np.random.rand(100, 224, 224, 3)
dataset = tf.data.Dataset.from_tensor_slices(input_data)
dataset = dataset.batch(batch_size=batch_size)
orig_g = tf.Graph()
with orig_g.as_default():
_ = VGG16(weights=None,
input_shape=(224, 224, 3),
include_top=False)
orig_init = tf.compat.v1.global_variables_initializer()
input_op_names = ['input_1']
output_op_names = ['block5_pool/MaxPool']
# create sess with graph
orig_sess = tf.compat.v1.Session(graph=orig_g)
# initialize all the variables in VGG16
orig_sess.run(orig_init)
# create layer database
layer_db = LayerDatabase(model=orig_sess,
input_shape=(1, 224, 224, 3),
working_dir=None)
conv_layer = layer_db.find_layer_by_name('block1_conv1/Conv2D')
comp_layer_db = copy.deepcopy(layer_db)
comp_conv_layer = comp_layer_db.find_layer_by_name(
'block1_conv1/Conv2D')
# get the weights before reconstruction in original model
before_recon_weights_orig_model = layer_db.model.run(
conv_layer.module.inputs[1])
# get the weights before reconstruction in pruned model
before_recon_weights_pruned_model = comp_layer_db.model.run(
comp_conv_layer.module.inputs[1])
# weight should be exactly same before reconstruction in original and pruned layer database
self.assertTrue(
np.array_equal(before_recon_weights_orig_model,
before_recon_weights_pruned_model))
cp = InputChannelPruner(input_op_names=input_op_names,
output_op_names=output_op_names,
data_set=dataset,
batch_size=batch_size,
num_reconstruction_samples=50,
allow_custom_downsample_ops=True)
num_in_channels = comp_conv_layer.weight_shape[0]
cp._data_subsample_and_reconstruction(orig_layer=conv_layer,
pruned_layer=comp_conv_layer,
output_mask=[1] *
num_in_channels,
orig_layer_db=layer_db,
comp_layer_db=comp_layer_db)
# get the weights after reconstruction
after_recon_weights_pruned_model = comp_layer_db.model.run(
comp_conv_layer.module.inputs[1])
# weight should not be same before and after reconstruction
self.assertFalse(
np.array_equal(before_recon_weights_orig_model,
after_recon_weights_pruned_model))
layer_db.model.close()
comp_layer_db.model.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))
def create_spatial_svd_algo(cls,
sess: tf.compat.v1.Session,
working_dir: str,
eval_callback: EvalFunction,
eval_iterations,
input_shape: Union[Tuple, List[Tuple]],
cost_metric: CostMetric,
params: SpatialSvdParameters,
bokeh_session=None) -> CompressionAlgo:
"""
Factory method to construct SpatialSvdCompressionAlgo
:param sess: Model, represented by a tf.compat.v1.Session, to compress
:param working_dir: path to store temp meta and checkpoint files
:param eval_callback: Evaluation callback for the model
:param eval_iterations: Evaluation iterations
:param input_shape: tuple or list of tuples of input shape to the model
:param cost_metric: Cost metric (mac or memory)
:param params: Spatial SVD compression parameters
:param bokeh_session: The Bokeh Session to display plots
:return: An instance of SpatialSvdCompressionAlgo
"""
# pylint: disable=too-many-arguments
# pylint: disable=too-many-locals
# Rationale: Factory functions unfortunately need to deal with a lot of parameters
# Create a layer database
layer_db = LayerDatabase(sess,
input_shape,
working_dir,
starting_ops=params.input_op_names,
ending_ops=params.output_op_names)
use_cuda = False
# Create a pruner
pruner = SpatialSvdPruner()
cost_calculator = SpatialSvdCostCalculator()
comp_ratio_rounding_algo = RankRounder(params.multiplicity,
cost_calculator)
# Create a comp-ratio selection algorithm
if params.mode == SpatialSvdParameters.Mode.auto:
greedy_params = params.mode_params.greedy_params
comp_ratio_select_algo = GreedyCompRatioSelectAlgo(
layer_db,
pruner,
cost_calculator,
eval_callback,
eval_iterations,
cost_metric,
greedy_params.target_comp_ratio,
greedy_params.num_comp_ratio_candidates,
greedy_params.use_monotonic_fit,
greedy_params.saved_eval_scores_dict,
comp_ratio_rounding_algo,
use_cuda,
bokeh_session=bokeh_session)
layer_selector = ConvNoDepthwiseLayerSelector()
modules_to_ignore = params.mode_params.modules_to_ignore
else:
# Convert (module,comp-ratio) pairs to (layer,comp-ratio) pairs
layer_comp_ratio_pairs = cls._get_layer_pairs(
layer_db, params.mode_params.list_of_module_comp_ratio_pairs)
comp_ratio_select_algo = ManualCompRatioSelectAlgo(
layer_db,
layer_comp_ratio_pairs,
comp_ratio_rounding_algo,
cost_metric=cost_metric)
layer_selector = ManualLayerSelector(layer_comp_ratio_pairs)
modules_to_ignore = []
# Create the overall Spatial SVD compression algorithm
spatial_svd_algo = CompressionAlgo(layer_db, comp_ratio_select_algo,
pruner, eval_callback,
layer_selector, modules_to_ignore,
cost_calculator, use_cuda)
return spatial_svd_algo
def test_layer_database_deepcopy(self):
# create tf.compat.v1.Session and initialize the weights and biases with zeros
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
# create session with graph
sess = tf.compat.v1.Session(graph=tf.Graph(), config=config)
with sess.graph.as_default():
# by default, model will be constructed in default graph
_ = mnist_tf_model.create_model(data_format='channels_last')
sess.run(tf.compat.v1.global_variables_initializer())
layer_db = LayerDatabase(model=sess, input_shape=(1, 28, 28, 1), working_dir=None)
layers = list(layer_db._compressible_layers.values())
# set the picked_for_compression for first two layers
layers[0].picked_for_compression = True
layers[1].picked_for_compression = True
layer_db_copy = copy.deepcopy(layer_db)
layers_copy = list(layer_db_copy._compressible_layers.values())
# should be True
self.assertTrue(layers_copy[0].picked_for_compression)
self.assertTrue(layers_copy[1].picked_for_compression)
# should be False
self.assertFalse(layers_copy[2].picked_for_compression)
self.assertFalse(layers_copy[3].picked_for_compression)
# op reference should be different
self.assertNotEqual(layers[0].module, layers_copy[0].module)
self.assertNotEqual(layers[1].module, layers_copy[1].module)
self.assertNotEqual(layers[2].module, layers_copy[2].module)
self.assertNotEqual(layers[3].module, layers_copy[3].module)
# layer reference should be different
self.assertNotEqual(layers[0], layers_copy[0])
self.assertNotEqual(layers[1], layers_copy[1])
self.assertNotEqual(layers[2], layers_copy[2])
self.assertNotEqual(layers[3], layers_copy[3])
# op name should be same for both original and copy
self.assertEqual(layers[0].name, layers_copy[0].name)
self.assertEqual(layers[1].name, layers_copy[1].name)
self.assertEqual(layers[2].name, layers_copy[2].name)
self.assertEqual(layers[3].name, layers_copy[3].name)
# session should be different
self.assertNotEqual(layer_db.model, layer_db_copy.model)
# check output shapes for layers in layer_db_copy
self.assertEqual(layers_copy[0].output_shape, [None, 32, 28, 28])
self.assertEqual(layers_copy[1].output_shape, [None, 64, 14, 14])
self.assertEqual(layers_copy[2].output_shape, [None, 1024, 1, 1])
self.assertEqual(layers_copy[3].output_shape, [None, 10, 1, 1])
# check weight shapes for layers in layer_db_copy
self.assertEqual(layers_copy[0].weight_shape, [32, 1, 5, 5])
self.assertEqual(layers_copy[1].weight_shape, [64, 32, 5, 5])
self.assertEqual(layers_copy[2].weight_shape, [1024, 3136])
self.assertEqual(layers_copy[3].weight_shape, [10, 1024])
# check the weight elements are equal in both data bases
self.assertTrue(np.array_equal(layers[0].module.inputs[1].eval(session=layer_db.model),
layers_copy[0].module.inputs[1].eval(session=layer_db_copy.model)))
self.assertTrue(np.array_equal(layers[1].module.inputs[1].eval(session=layer_db.model),
layers_copy[1].module.inputs[1].eval(session=layer_db_copy.model)))
self.assertTrue(np.array_equal(layers[2].module.inputs[1].eval(session=layer_db.model),
layers_copy[2].module.inputs[1].eval(session=layer_db_copy.model)))
self.assertTrue(np.array_equal(layers[3].module.inputs[1].eval(session=layer_db.model),
layers_copy[3].module.inputs[1].eval(session=layer_db_copy.model)))
tf.compat.v1.reset_default_graph()
sess.close()
# delete temp directory
shutil.rmtree(str('./temp_meta/'))