def __init__(self, list_of_modules_to_winnow, reshape, in_place, verbose):
self._list_of_modules_to_winnow = list_of_modules_to_winnow
self._reshape = reshape
self._in_place = in_place
if verbose is True:
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Winnow, logging.INFO)
else:
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Winnow, logging.WARNING)
def test_setting_log_level(self):
logger.info("*** Testing test_setting_log_level() *** \n")
svd_logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Svd)
# The default logging level for SVD defined in default_logging_config.json is used.
logger.info(
"Log at the default log level for SVD defined in default_logging_config.json"
)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
svd_logger.warning("Testing Warning")
svd_logger.error("Testing Error")
svd_logger.critical("Testing Critical")
svd_logger.critical("****************************************\n")
# Change the default log level for SVD.
# Only CRITICAL level logs will be logged.
logger.info("Change SVD area's logging level to Critical")
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Svd,
logging.CRITICAL)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
svd_logger.warning("Testing Warning")
svd_logger.error("Testing Error")
svd_logger.critical("Testing Critical")
svd_logger.critical("****************************************\n")
# Change the default log level for SVD.
# All logs will be logged.
logger.info("Change SVD area's logging level to Critical")
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Svd,
logging.DEBUG)
svd_logger.debug("Testing Debug")
svd_logger.info("Testing Info")
svd_logger.warning("Testing Warning")
svd_logger.error("Testing Error")
svd_logger.critical("Testing Critical")
svd_logger.critical("****************************************\n")
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/'))
# SPDX-License-Identifier: BSD-3-Clause
#
# @@-COPYRIGHT-END-@@
# =============================================================================
""" This file contains unit tests for testing ModuleIdentifier modules. """
import unittest
import logging
import tensorflow as tf
from aimet_common.utils import AimetLogger
from aimet_tensorflow.common.module_identifier import StructureModuleIdentifier
from aimet_tensorflow.examples.test_models import keras_model, keras_model_functional, tf_slim_basic_model
logger = AimetLogger.get_area_logger(AimetLogger.LogAreas.Test)
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.Test, logging.DEBUG)
AimetLogger.set_area_logger_level(AimetLogger.LogAreas.ConnectedGraph,
logging.DEBUG)
class TestStructureModuleIdentifier(unittest.TestCase):
""" Test StructureModuleIdentifier module """
def test_get_op_info(self):
""" Test get_op_info() in StructureModuleIdentifier """
my_op_type_set = set()
current_module_set = set()
tf.compat.v1.reset_default_graph()
_ = keras_model()
module_identifier = StructureModuleIdentifier(
