def test_harmonize_dtypes(self):
"""harmonize_dtypes should change tensor dtype to node dtype"""
graph_def = rewrite.GraphDef()
const_nodes = [
# tensor "a" is float32
testutils.node_proto_from_json('{"name":"a","op":"Const","attr":'
'{"dtype":{"type":"DT_FLOAT"}}}'),
# tensor "b" is int64
testutils.node_proto_from_json('{"name":"b","op":"Const","attr":'
'{"dtype":{"type":"DT_INT64"}}}')
]
graph_def.node.extend(const_nodes)
weight_dict = {
# weight "a" matches tensor "a"
'a':
convert_to_tensor(
np.arange(9., dtype=np.float32).reshape((1, 3, 3))),
# weight "b" is int32 (must be widened to match node)
'b':
convert_to_tensor(np.arange(4, dtype=np.int32)),
# no matching node for weight "c"
'c':
convert_to_tensor(np.array(23, dtype=np.int64))
}
result = rewrite.harmonize_dtypes(graph_def, weight_dict)
# existing should be unchanged if matching
self.assertEqual(result['a'].numpy().dtype, np.float32)
# existing should be altered to match node
self.assertEqual(result['b'].numpy().dtype, np.int64)
# non-existing should be unchanged
self.assertEqual(result['c'].numpy().dtype, np.int64)
def test_is_fused_op_without_activation(self):
"""is_fused_op should return True if op is fused with BiasAdd only
and no activation function is given
"""
fused_matmul = testutils.node_proto_from_json(
'{"name":"model/output/BiasAdd","op":"_FusedMatMul",'
'"input":["model/dense/BiasAdd",'
'"model/output/MatMul/ReadVariableOp",'
'"model/output/BiasAdd/ReadVariableOp"],"device":"/device:CPU:0",'
'"attr":{"transpose_b":{"b":false},"T":{"type":"DT_FLOAT"},'
'"num_args":{"i": "1"},"epsilon":{"f": 0},'
'"fused_ops":{"list":{"s":["Qmlhc0FkZA=="]}},'
'"transpose_a":{"b":false}}}')
self.assertTrue(
rewrite.is_fused_op(fused_matmul, 'MatMul', activation=''))
fused_conv2d = testutils.node_proto_from_json(
'{"name":"/model/batch_normalization_v1_8/FusedBatchNormV3",'
'"op":"_FusedConv2D",'
'"input":["model/depthwise","model/weights","model/bn_offset"],'
'"device":"/device:CPU:0",'
'"attr":{"fused_ops":{"list":{"s":["Qmlhc0FkZA=="]}},'
'"dilations":{"list":{"i":["1","1","1","1"]}},'
'"T":{"type": "DT_FLOAT"},'
'"strides":{"list":{"i": ["1","1","1","1"]}},'
'"data_format":{"s":"TkhXQw=="},'
'"explicit_paddings":{"list":{}},'
'"num_args":{"i":"1"},'
'"epsilon":{"f":0},'
'"padding":{"s":"VkFMSUQ="}}}')
self.assertTrue(
rewrite.is_fused_op(fused_conv2d, 'Conv2D', activation=None))
def test_copy_op_attrs(self):
"""copy_op_attrs should only copy attrs supported by the target node"""
# copy_op_attrs is used to transfer attrs from a fused op node
# (e.g. _FusedConv2D) to a standalone op (e.g. Conv2D)
# any additional attrs of the fused op need to be ignored
fused_op_str = '{"name":"model/conv2d/BiasAdd",'\
+ '"op":"_FusedConv2D","input":["input",'\
+ '"model/conv2d/Conv2D/ReadVariableOp",'\
+ '"model/conv2d/BiasAdd/ReadVariableOp",'\
+ '"model/p_re_lu/Neg"],"device":"/device:CPU:0",' \
+ '"attr":{"dilations":{"list":{"i":["1","1","1","1"]}},'\
+ '"T":{"type":"DT_FLOAT"},"data_format":{"s":"TkhXQw=="},'\
+ '"strides":{"list":{"i":["1","1","1","1"]}},'\
+ '"use_cudnn_on_gpu":{"b":true},'\
+ '"explicit_paddings":{"list":{}},'\
+ '"num_args":{"i":"2"},"epsilon":{"f":0},'\
+ '"padding":{"s":"VkFMSUQ="},'\
+ '"fused_ops":{"list":{"s":["Qmlhc0FkZA==","UHJlbHU="]}}}}'
fused_op = testutils.node_proto_from_json(fused_op_str)
node = rewrite.make_op_node('Conv2D', fused_op.input[0:2])
rewrite.copy_op_attrs(source=fused_op, target=node)
op_def = rewrite.get_op_def(node.op)
allowed = set(attr.name for attr in op_def.attr)
forbidden = any(attr for attr in node.attr if attr not in allowed)
self.assertFalse(forbidden)
# randomply check for some of the expected attributes
self.assertTrue('padding' in node.attr)
self.assertTrue('strides' in node.attr)
def test_is_fused_op(self):
"""is_fused_op should be true if op is fused with BiasAdd+Activation"""
missing_activation = testutils.node_proto_from_json(
'{"name":"model/output/BiasAdd","op":"_FusedMatMul",'
'"input":["model/dense/BiasAdd",'
'"model/output/MatMul/ReadVariableOp",'
'"model/output/BiasAdd/ReadVariableOp"],"device":"/device:CPU:0",'
'"attr":{"transpose_b":{"b":false},"T":{"type":"DT_FLOAT"},'
'"num_args":{"i": "1"},"epsilon":{"f": 0},'
'"fused_ops":{"list":{"s":["Qmlhc0FkZA=="]}},'
'"transpose_a":{"b":false}}}')
self.assertFalse(
rewrite.is_fused_op(missing_activation, 'MatMul', b'Relu'))
fused_matmul = testutils.node_proto_from_json(
'{"name":"model/dense/BiasAdd","op":"_FusedMatMul",'
'"input":["model/flatten/Reshape",'
'"model/dense/MatMul/ReadVariableOp",'
'"model/dense/BiasAdd/ReadVariableOp","model/p_re_lu_2/Neg"],'
'"device":"/device:CPU:0","attr":{"transpose_b":{"b":false},'
'"T":{"type":"DT_FLOAT"},"num_args":{"i":"2"},"epsilon":{"f":0},'
'"fused_ops":{"list":{"s":["Qmlhc0FkZA==","UHJlbHU="]}},'
'"transpose_a":{"b":false}}}')
self.assertTrue(rewrite.is_fused_op(fused_matmul, 'MatMul', b'Prelu'))
def test_validate_supported_ops_given_invalid_graph(self):
"""validate_supported_ops should raise ValueError for unsupported op"""
# case 1: unsupported op node
graph_def = rewrite.GraphDef()
unsupported_op = testutils.node_proto_from_json(
'{"name":"model/p_re_lu_1/Relu","op":"Prelu","input":'
'["model/add/add","model/p_re_lu_1/Neg"]}')
graph_def.node.extend([unsupported_op])
self.assertRaises(ValueError,
lambda: rewrite.validate_supported_ops(graph_def))
# case 2: unsupported fused op
unsupported_fused_op = testutils.node_proto_from_json(
'{"name":"model/dense/BiasAdd","op":"_FusedMatMul",'
'"input":["model/flatten/Reshape",'
'"model/dense/MatMul/ReadVariableOp",'
'"model/dense/BiasAdd/ReadVariableOp","model/p_re_lu_2/Neg"],'
'"device":"/device:CPU:0","attr":{"transpose_b":{"b":false},'
'"T":{"type":"DT_FLOAT"},"num_args":{"i":"2"},"epsilon":{"f":0},'
'"fused_ops":{"list":{"s":["Qmlhc0FkZA==","UHJlbHU="]}},'
'"transpose_a":{"b":false}}}')
graph_def = rewrite.GraphDef()
graph_def.node.extend([unsupported_fused_op])
self.assertRaises(ValueError,
lambda: rewrite.validate_supported_ops(graph_def))