def test_set_and_freeze_op_mode(self):
""" Create QuantSim for a CPU model, test set and freeze op mode """
tf.compat.v1.reset_default_graph()
with tf.device('/cpu:0'):
_ = keras_model()
init = tf.compat.v1.global_variables_initializer()
session = tf.compat.v1.Session()
session.run(init)
sim = QuantizationSimModel(session, ['conv2d_input'],
['keras_model/Softmax'],
use_cuda=False)
quantizer = sim.quantizer_config(
'conv2d/Conv2D/ReadVariableOp_quantized')
op_mode = int(libpymo.TensorQuantizerOpMode.oneShotQuantizeDequantize)
quantizer.set_op_mode(op_mode)
quantizer.freeze_encoding()
self.assertEqual(op_mode, quantizer.get_op_mode())
new_op_mode = int(libpymo.TensorQuantizerOpMode.passThrough)
quantizer.set_op_mode(new_op_mode)
self.assertNotEqual(new_op_mode, quantizer.get_op_mode())
self.assertEqual(op_mode, quantizer.get_op_mode())
session.close()
def test_set_and_freeze_param_encodings(self):
""" Test set and freeze parameter encodings functionality """
tf.compat.v1.reset_default_graph()
with tf.device('/cpu:0'):
_ = keras_model()
init = tf.compat.v1.global_variables_initializer()
session = tf.compat.v1.Session()
session.run(init)
sim = QuantizationSimModel(session, ['conv2d_input'],
['keras_model/Softmax'],
use_cuda=False)
param_encodings = {
'conv2d/Conv2D/ReadVariableOp:0': [{
'bitwidth': 4,
'is_symmetric': False,
'max': 0.14584073424339294,
'min': -0.12761062383651733,
'offset': -7.0,
'scale': 0.01823008991777897
}]
}
# export encodings to JSON file
encoding_file_path = os.path.join('./', 'dummy.encodings')
with open(encoding_file_path, 'w') as encoding_fp:
json.dump(param_encodings, encoding_fp, sort_keys=True, indent=4)
sim.set_and_freeze_param_encodings(encoding_path='./dummy.encodings')
quantizer = sim.quantizer_config(
'conv2d/Conv2D/ReadVariableOp_quantized')
encoding = param_encodings['conv2d/Conv2D/ReadVariableOp:0'][0]
encoding_max = quantizer.get_variable_from_op(
QuantizeOpIndices.encoding_max)
encoding_min = quantizer.get_variable_from_op(
QuantizeOpIndices.encoding_min)
self.assertEqual(encoding_min, encoding.get('min'))
self.assertEqual(encoding_max, encoding.get('max'))
self.assertEqual(int(libpymo.TensorQuantizerOpMode.quantizeDequantize),
quantizer.get_op_mode())
self.assertEqual(quantizer.is_encoding_valid(), True)
session.close()
# Delete encodings JSON file
if os.path.exists("./dummy.encodings"):
os.remove("./dummy.encodings")
def test_keras_model_get_op_product_graph(self):
""" Test connected graph construction on keras model """
tf.compat.v1.reset_default_graph()
_ = keras_model()
conn_graph = ConnectedGraph(tf.compat.v1.get_default_graph(), ['conv2d_input'], ['keras_model/Softmax'])
self.assertTrue(validate_branch_ops(conn_graph))
self.assertTrue(validate_product_tensor_lists(conn_graph))
self.assertEqual(0, conn_graph.branch_count)
self.assertEqual(11, len(conn_graph.get_all_ops()))
# 10 products from inter module connections
# 14 products from parameters
self.assertEqual(24, len(conn_graph.get_all_products()))
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(
tf.compat.v1.get_default_graph(), ["conv2d_input"],
set(tf.compat.v1.get_default_graph().get_operations()))
for op_info in module_identifier.op_to_module_dict.values():
my_op_type_set.add(op_info.op_type)
current_module_set.add(op_info.module_name)
# Only identifies 2 conv2d, 2 fusedbatchnorm, flatten, and dense
self.assertEqual(6, len(current_module_set))
self.assertEqual(4, len(my_op_type_set))
def test_set_and_freeze_encoding(self):
""" Create QuantSim for a CPU model, test set and freeze encoding """
tf.compat.v1.reset_default_graph()
with tf.device('/cpu:0'):
_ = keras_model()
init = tf.compat.v1.global_variables_initializer()
session = tf.compat.v1.Session()
session.run(init)
sim = QuantizationSimModel(session, ['conv2d_input'],
['keras_model/Softmax'],
use_cuda=False)
quantizer = sim.quantizer_config(
'conv2d/Conv2D/ReadVariableOp_quantized')
encoding = quantizer.compute_encoding(8, False)
print(encoding.max, encoding.min)
# Set and freeze encoding
quantizer.set_encoding(encoding)
quantizer.freeze_encoding()
old_encoding_min = quantizer.get_variable_from_op(
QuantizeOpIndices.encoding_min)
old_encoding_max = quantizer.get_variable_from_op(
QuantizeOpIndices.encoding_max)
self.assertEqual(encoding.min, old_encoding_min)
self.assertEqual(encoding.max, old_encoding_max)
self.assertEqual(quantizer.is_encoding_valid(), True)
# Try updating encoding min and max with new values, but values can not be changed
encoding.min = -0.4
encoding.max = 0.6
quantizer.set_encoding(encoding)
self.assertEqual(
old_encoding_min,
quantizer.get_variable_from_op(QuantizeOpIndices.encoding_min))
self.assertEqual(
old_encoding_max,
quantizer.get_variable_from_op(QuantizeOpIndices.encoding_max))
session.close()
def test_get_encoding(self):
""" Create QuantSim for a CPU model, test get encoding """
tf.compat.v1.reset_default_graph()
with tf.device('/cpu:0'):
_ = keras_model()
init = tf.compat.v1.global_variables_initializer()
session = tf.compat.v1.Session()
session.run(init)
sim = QuantizationSimModel(session, ['conv2d_input'],
['keras_model/Softmax'],
use_cuda=False)
quantizer = sim.quantizer_config(
'conv2d/Conv2D/ReadVariableOp_quantized')
self.assertRaises(AssertionError, lambda: quantizer.get_encoding())
session.close()