experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Add simple transpose_conv tests with post-training quantization.

Browse Source 
PiperOrigin-RevId: 281864531
Change-Id: I57795cd589387783602abcdcf393e781888991bc
This commit is contained in:
Yunlu Li 2019-11-21 17:01:01 -08:00 committed by TensorFlower Gardener
parent 72e362bba7
commit 5686a3c261
2 changed files with 79 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
tensorflow/lite/testing/BUILD
View File

@ -509,6 +509,7 @@ edgetpu_ops = [
"sub",
"sum", # high error
"tanh",
"transpose_conv",
]
[gen_zipped_test_file(

108
tensorflow/lite/testing/op_tests/transpose_conv.py
View File

@ -34,22 +34,50 @@ def make_transpose_conv_tests(options):
"""Make a set of tests to do transpose_conv."""
# Tensorflow only supports equal strides
test_parameters = [{
"input_shape": [[1, 3, 4, 1], [1, 10, 10, 3], [3, 20, 20, 1]],
"filter_size": [[1, 1], [1, 2], [3, 3]],
"strides": [[1, 1, 1, 1], [1, 3, 3, 1]],
"padding": ["SAME", "VALID"],
"data_format": ["NHWC"],
"channel_multiplier": [1, 2],
}]
test_parameters = [
{
"input_shape": [[1, 3, 4, 1], [1, 10, 10, 3], [3, 20, 20, 1]],
"filter_size": [[1, 1], [1, 2], [3, 3]],
"strides": [[1, 1, 1, 1], [1, 3, 3, 1]],
"padding": ["SAME", "VALID"],
"data_format": ["NHWC"],
"channel_multiplier": [1, 2],
"output_shape": [[]],
"fully_quantize": [False]
},
# TODO(yunluli): Adding simple tests for now to unblock edgetpu debugging.
# Need to add more test cases.
{
"input_shape": [[1, 3, 3, 1]],
"filter_size": [[3, 3, 2, 1]],
"strides": [[1, 1, 1, 1]],
"padding": ["SAME"],
"data_format": ["NHWC"],
"channel_multiplier": [1],
"output_shape": [[1, 3, 3, 2]],
"fully_quantize": [True]
},
{
"input_shape": [[1, 3, 3, 1]],
"filter_size": [[3, 3, 2, 1]],
"strides": [[1, 2, 2, 1]],
"padding": ["SAME"],
"data_format": ["NHWC"],
"channel_multiplier": [1],
"output_shape": [[1, 6, 6, 2]],
"fully_quantize": [True]
}
]
def get_tensor_shapes(parameters):
input_shape = parameters["input_shape"]
filter_size = parameters["filter_size"]
filter_shape = filter_size + [
input_shape[3], parameters["channel_multiplier"]
]
return [input_shape, filter_shape]
if not parameters["fully_quantize"]:
filter_shape = filter_size + [
input_shape[3], parameters["channel_multiplier"]
]
return [input_shape, filter_shape]
return [input_shape, filter_size]
def build_graph(parameters):
"""Build a transpose_conv graph given `parameters`."""
@ -60,28 +88,48 @@ def make_transpose_conv_tests(options):
filter_input = tf.compat.v1.placeholder(
dtype=tf.float32, name="filter", shape=filter_shape)
conv_outputs = tf.nn.conv2d(
input_tensor,
filter_input,
strides=parameters["strides"],
padding=parameters["padding"],
data_format=parameters["data_format"])
out = tf.compat.v1.nn.conv2d_backprop_input(
input_shape,
filter_input,
conv_outputs,
strides=parameters["strides"],
padding=parameters["padding"],
data_format=parameters["data_format"])
input_tensors = [input_tensor, filter_input]
if not parameters["fully_quantize"]:
input_tensors = [input_tensor, filter_input]
conv_outputs = tf.nn.conv2d(
input_tensor,
filter_input,
strides=parameters["strides"],
padding=parameters["padding"],
data_format=parameters["data_format"])
out = tf.compat.v1.nn.conv2d_backprop_input(
input_shape,
filter_input,
conv_outputs,
strides=parameters["strides"],
padding=parameters["padding"],
data_format=parameters["data_format"])
else:
input_tensors = [input_tensor]
filter_input = create_tensor_data(
np.float32, filter_shape, min_value=-1, max_value=1)
out = tf.nn.conv2d_transpose(
input_tensor,
filter_input,
parameters["output_shape"],
strides=parameters["strides"],
padding=parameters["padding"],
data_format=parameters["data_format"])
return input_tensors, [out]
def build_inputs(parameters, sess, inputs, outputs):
input_shape, filter_shape = get_tensor_shapes(parameters)
values = [
create_tensor_data(np.float32, input_shape),
create_tensor_data(np.float32, filter_shape)
]
if not parameters["fully_quantize"]:
values = [
create_tensor_data(np.float32, input_shape),
create_tensor_data(np.float32, filter_shape)
]
else:
values = [
create_tensor_data(
np.float32, input_shape, min_value=-1, max_value=1),
]
return values, sess.run(outputs, feed_dict=dict(zip(inputs, values)))
make_zip_of_tests(options, test_parameters, build_graph, build_inputs)