def dump_checkpoint(
checkpoint_file, output_dir, shard_mb=4, remove_variables_regex=None,
quantization_dtype=None):
reader = tf.train.NewCheckpointReader(checkpoint_file)
var_to_shape_map = reader.get_variable_to_shape_map()
remove_variables_regex_re = (
re.compile(remove_variables_regex) if remove_variables_regex else None)
entries = []
for var_name, shape in var_to_shape_map.items():
if (remove_variables_regex_re and remove_variables_regex_re.match(var_name)
or var_name == 'global_step'):
print('Ignoring Regex Match: ' + var_name)
continue
if not shape:
print('Ignoring Scalar: ' + var_name)
continue
tensor = reader.get_tensor(var_name)
entries.append({'name': var_name, 'data': tensor})
print('Dumping %s (%r)' % (var_name, shape))
write_weights(
[entries],
output_dir,
write_manifest=True,
quantization_dtype=quantization_dtype,
shard_size_bytes=shard_mb * 1024 * 1024)
def extract_weights(graph_def, output_graph):
"""Takes a Python GraphDef object and extract the weights.
Args:
graph_def: tf.GraphDef tensorflow GraphDef proto object, which represents
the model topology
"""
constants = [node for node in graph_def.node if node.op == 'Const']
# removed the conditional inputs for constants
for const in constants:
del const.input[:]
print('Writing weight file ' + output_graph + '...')
const_manifest = []
path = os.path.dirname(output_graph)
graph = tf.Graph()
with tf.Session(graph=graph) as sess:
tf.import_graph_def(graph_def, name='')
for const in constants:
tensor = graph.get_tensor_by_name(const.name + ':0')
value = tensor.eval(session=sess)
if not isinstance(value, np.ndarray):
value = np.array(value)
const_manifest.append({'name': const.name, 'data': value})
# Remove the binary array from tensor and save it to the external file.
const.attr["value"].tensor.ClearField('tensor_content')
write_weights.write_weights([const_manifest], path)
file_io.atomic_write_string_to_file(os.path.abspath(output_graph),
graph_def.SerializeToString())
def test_non_grouped_weights_throws(self):
groups = [{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def test_bad_numpy_array_dtype_throws(self):
groups = [[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float64')
}]]
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def test_non_grouped_weights_throws(self):
groups = [{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def test_bad_weights_entry_throws_no_data(self):
groups = [[{
'name': 'weight1',
'nodata': np.array([1, 2, 3], 'float32')
}]]
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def test_bad_weights_entry_throws_no_data(self):
groups = [
[{
'name': 'weight1',
'nodata': np.array([1, 2, 3], 'float32')
}]
]
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def test_bad_numpy_array_dtype_throws(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float64')
}]
]
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def testReadWeightsWithIncorrectTypeInWeightsManifestRaisesError(self):
groups = [[{
'name': 'weight1',
'data': np.random.rand(1, 100).astype(np.float32)
}]]
write_weights.write_weights(groups, self._tmp_dir)
with self.assertRaises(ValueError):
read_weights.read_weights(groups[0][0], self._tmp_dir)
def test_duplicate_weight_name_throws(self):
groups = [[{
'name': 'duplicate',
'data': np.array([1, 2, 3], 'float32')
}], [{
'name': 'duplicate',
'data': np.array([4, 5, 6], 'float32')
}]]
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def testReadWeightsWithIncorrectTypeInWeightsManifestRaisesError(self):
groups = [
[{
'name': 'weight1',
'data': np.random.rand(1, 100).astype(np.float32)
}]
]
write_weights.write_weights(groups, self._tmp_dir)
with self.assertRaises(ValueError):
read_weights.read_weights(groups[0][0], self._tmp_dir)
def test_duplicate_weight_name_throws(self):
groups = [
[{
'name': 'duplicate',
'data': np.array([1, 2, 3], 'float32')
}], [{
'name': 'duplicate',
'data': np.array([4, 5, 6], 'float32')
}]
]
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def test_1_group_1_weight_bool(self):
groups = [[{
'name': 'weight1',
'data': np.array([True, False, True], 'bool')
}]]
manifest = write_weights.write_weights(groups,
TMP_DIR,
shard_size_bytes=4 * 4)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(manifest, [{
'paths': ['group1-shard1of1.bin'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'bool'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1.bin')
weight1 = np.fromfile(weights_path, 'bool')
np.testing.assert_array_equal(weight1,
np.array([True, False, True], 'bool'))
def test_1_group_2_packed_sharded_multi_dtype(self):
groups = [[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'int32')
}, {
'name': 'weight2',
'data': np.array([True, False, False, True], 'bool')
}, {
'name': 'weight3',
'data': np.array([4.1, 5.1], 'float32')
}]]
# Shard size is smaller than the sum of the weights so they get packed and
# then sharded. The two buffers will get sharded into 3 files, with shapes
# [2], [2], and [1]. The second shard is a mixed dtype.
manifest = write_weights.write_weights(groups,
TMP_DIR,
shard_size_bytes=2 * 4)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(manifest, [{
'paths': [
'group1-shard1of3.bin', 'group1-shard2of3.bin',
'group1-shard3of3.bin'
],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'int32'
}, {
'name': 'weight2',
'shape': [4],
'dtype': 'bool'
}, {
'name': 'weight3',
'shape': [2],
'dtype': 'float32'
}]
}])
shard_1_path = os.path.join(TMP_DIR, 'group1-shard1of3.bin')
shard_1 = np.fromfile(shard_1_path, 'int32')
np.testing.assert_array_equal(shard_1, np.array([1, 2], 'int32'))
# Shard 2 has a mixed dtype so we parse the bytes directly.
shard_2_path = os.path.join(TMP_DIR, 'group1-shard2of3.bin')
with open(shard_2_path, 'rb') as f:
shard_2_bytes = f.read()
self.assertEqual(len(shard_2_bytes), 8)
shard_2_int = np.frombuffer(shard_2_bytes[:4], 'int32')
np.testing.assert_array_equal(shard_2_int, np.array([3], 'int32'))
shard_2_bool = np.frombuffer(shard_2_bytes[4:], 'bool')
np.testing.assert_array_equal(
shard_2_bool, np.array([True, False, False, True], 'bool'))
shard_3_path = os.path.join(TMP_DIR, 'group1-shard3of3.bin')
shard_3 = np.fromfile(shard_3_path, 'float32')
np.testing.assert_array_equal(shard_3, np.array([4.1, 5.1], 'float32'))
def test_1_group_1_weight(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]
]
manifest_json = write_weights.write_weights(
groups, TMP_DIR, shard_size_bytes=4 * 4)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of1'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1')
weight1 = np.fromfile(weights_path, 'float32')
np.testing.assert_array_equal(weight1, np.array([1, 2, 3], 'float32'))
def test_1_group_1_weight_bool(self):
groups = [
[{
'name': 'weight1',
'data': np.array([True, False, True], 'bool')
}]
]
manifest_json = write_weights.write_weights(
groups, TMP_DIR, shard_size_bytes=4 * 4)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of1'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'bool'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1')
weight1 = np.fromfile(weights_path, 'bool')
np.testing.assert_array_equal(
weight1, np.array([True, False, True], 'bool'))
def test_no_write_manfest(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]
]
manifest_json = write_weights.write_weights(
groups, TMP_DIR, write_manifest=False)
manifest = json.loads(manifest_json)
self.assertFalse(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json exists, but expected not to exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of1'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1')
weight1 = np.fromfile(weights_path, 'float32')
np.testing.assert_array_equal(weight1, np.array([1, 2, 3], 'float32'))
def test_1_group_1_weight_sharded(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]
]
# Shard size is smaller than the size of the array so it gets split between
# multiple files.
manifest_json = write_weights.write_weights(
groups, TMP_DIR, shard_size_bytes=2 * 4)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of2', 'group1-shard2of2'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32'
}]
}])
shard_1_path = os.path.join(TMP_DIR, 'group1-shard1of2')
shard_1 = np.fromfile(shard_1_path, 'float32')
np.testing.assert_array_equal(shard_1, np.array([1, 2], 'float32'))
shard_2_path = os.path.join(TMP_DIR, 'group1-shard2of2')
shard_2 = np.fromfile(shard_2_path, 'float32')
np.testing.assert_array_equal(shard_2, np.array([3], 'float32'))
def test_1_group_2_weights_packed(self):
groups = [[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}, {
'name': 'weight2',
'data': np.array([4, 5], 'float32')
}]]
# Shard size is larger than the sum of the two weights so they get packed.
manifest_json = write_weights.write_weights(groups,
TMP_DIR,
shard_size_bytes=8 * 4)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(manifest, [{
'paths': ['group1-shard1of1'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32'
}, {
'name': 'weight2',
'shape': [2],
'dtype': 'float32'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1')
weights = np.fromfile(weights_path, 'float32')
np.testing.assert_array_equal(weights,
np.array([1, 2, 3, 4, 5], 'float32'))
def test_no_write_manfest(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]
]
manifest_json = write_weights.write_weights(
groups, TMP_DIR, write_manifest=False)
manifest = json.loads(manifest_json)
self.assertFalse(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json exists, but expected not to exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of1'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1')
weight1 = np.fromfile(weights_path, 'float32')
np.testing.assert_array_equal(weight1, np.array([1, 2, 3], 'float32'))
def testReadCyrillicStringUnicodeAndEncoded(self):
groups = [
[{
'name': 'weight1',
# String is stored as unicode.
'data': np.array([u'здраво'], 'object')
}, {
'name': 'weight2',
# String is stored encoded.
'data': np.array([u'поздрав'.encode('utf-8')], 'object')
}]
]
manifest = write_weights.write_weights(groups, self._tmp_dir)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(manifest, self._tmp_dir)
self.assertEqual(1, len(read_output))
group = read_output[0]
self.assertEqual(2, len(group))
weight1 = group[0]
self.assertEqual('weight1', weight1['name'])
np.testing.assert_array_equal(
weight1['data'],
np.array([u'здраво'.encode('utf-8')], 'object'))
weight2 = group[1]
self.assertEqual('weight2', weight2['name'])
np.testing.assert_array_equal(
weight2['data'],
np.array([u'поздрав'.encode('utf-8')], 'object'))
def test_1_group_1_weight_string_empty(self):
groups = [[{'name': 'weight1', 'data': np.array([''], 'object')}]]
manifest = write_weights.write_weights(groups,
TMP_DIR,
shard_size_bytes=4 * 1024 *
1024)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(manifest, [{
'paths': ['group1-shard1of1.bin'],
'weights': [{
'name': 'weight1',
'shape': [1],
'dtype': 'string'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1.bin')
with open(weights_path, 'rb') as f:
weight_bytes = f.read()
self.assertEqual(len(weight_bytes), 4)
size = np.frombuffer(weight_bytes[:4], 'uint32')[0]
self.assertEqual(size, 0) # Empty string.
def write_artifacts(topology,
weights,
output_graph,
tf_version,
quantization_dtype=None):
"""Writes weights and topology to the output_dir.
If `topology` is Falsy (e.g., `None`), only emit weights to output_dir.
Args:
topology: tf.GraphDef TensorFlow GraphDef proto object, which represents
the model topology.
weights: an array of weight groups (as defined in tfjs write_weights).
output_graph: the output file name to hold all the contents.
tf_version: Tensorflow version of the input graph.
quantization_dtype: An optional numpy dtype to quantize weights to for
compression. Only np.uint8 and np.uint16 are supported.
"""
model_json = {
common.FORMAT_KEY: common.TFJS_GRAPH_MODEL_FORMAT,
# TODO(piyu): Add tensorflow version below by using `meta_info_def`.
common.GENERATED_BY_KEY: tf_version,
common.CONVERTED_BY_KEY: common.get_converted_by(),
}
model_json[common.ARTIFACT_MODEL_TOPOLOGY_KEY] = topology or None
weights_manifest = write_weights.write_weights(
weights, os.path.dirname(output_graph), write_manifest=False,
quantization_dtype=quantization_dtype)
assert isinstance(weights_manifest, list)
model_json[common.ARTIFACT_WEIGHTS_MANIFEST_KEY] = weights_manifest
with open(output_graph, 'wt') as f:
json.dump(model_json, f)
def write_artifacts(self, topology, weights, output_dir):
"""Writes weights and topology to the output_dir.
If `topology` is Falsy (e.g., `None`), only emit weights to output_dir.
Args:
topology: a JSON dictionary, representing the Keras config.
weights: an array of weight groups (as defined in tfjs write_weights).
output_dir: the directory to hold all the contents.
"""
# TODO(cais, nielsene): This method should allow optional arguments of
# `write_weights.write_weights` (e.g., shard size) and forward them.
# We write the topology after since write_weights makes no promises about
# preserving directory contents.
if os.path.isfile(output_dir):
raise ValueError(
'Path "%d" already exists as a file (not a directory).' %
output_dir)
model_json = {}
model_json[ARTIFACT_MODEL_TOPOLOGY_KEY] = topology or None
weights_manifest = write_weights.write_weights(weights,
output_dir,
write_manifest=False)
if not isinstance(weights_manifest, list):
weights_manifest = json.loads(weights_manifest)
assert isinstance(weights_manifest, list)
model_json[ARTIFACT_WEIGHTS_MANIFEST_KEY] = weights_manifest
model_json_path = os.path.join(output_dir,
ARTIFACT_MODEL_JSON_FILE_NAME)
with open(model_json_path, 'wt') as f:
json.dump(model_json, f)
def test_1_group_1_weight_sharded(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]
]
# Shard size is smaller than the size of the array so it gets split between
# multiple files.
manifest_json = write_weights.write_weights(
groups, TMP_DIR, shard_size_bytes=2 * 4)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of2', 'group1-shard2of2'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32'
}]
}])
shard_1_path = os.path.join(TMP_DIR, 'group1-shard1of2')
shard_1 = np.fromfile(shard_1_path, 'float32')
np.testing.assert_array_equal(shard_1, np.array([1, 2], 'float32'))
shard_2_path = os.path.join(TMP_DIR, 'group1-shard2of2')
shard_2 = np.fromfile(shard_2_path, 'float32')
np.testing.assert_array_equal(shard_2, np.array([3], 'float32'))
def testReadOneGroupEmptyStrings(self):
groups = [[{
'name': 'weight1',
'data': np.array(['', ''], 'object')
}, {
'name': 'weight2',
'data': np.array([], 'object')
}, {
'name': 'weight3',
'data': np.array([[]], 'object')
}]]
manifest = write_weights.write_weights(groups, self._tmp_dir)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(manifest, self._tmp_dir)
self.assertEqual(1, len(read_output))
group = read_output[0]
self.assertEqual(3, len(group))
weight1 = group[0]
self.assertEqual('weight1', weight1['name'])
np.testing.assert_array_equal(
weight1['data'],
np.array([u''.encode('utf-8'), u''.encode('utf-8')], 'object'))
weight2 = group[1]
self.assertEqual('weight2', weight2['name'])
np.testing.assert_array_equal(weight2['data'], np.array([], 'object'))
weight3 = group[2]
self.assertEqual('weight3', weight3['name'])
np.testing.assert_array_equal(weight3['data'], np.array([[]],
'object'))
def write_artifacts(topology,
weights,
output_graph,
quantization_dtype=None):
"""Writes weights and topology to the output_dir.
If `topology` is Falsy (e.g., `None`), only emit weights to output_dir.
Args:
topology: tf.GraphDef TensorFlow GraphDef proto object, which represents
the model topology.
weights: an array of weight groups (as defined in tfjs write_weights).
output_graph: the output file name to hold all the contents.
quantization_dtype: An optional numpy dtype to quantize weights to for
compression. Only np.uint8 and np.uint16 are supported.
"""
model_json = {}
model_json[ARTIFACT_MODEL_TOPOLOGY_KEY] = topology or None
weights_manifest = write_weights.write_weights(
weights, os.path.dirname(output_graph), write_manifest=False,
quantization_dtype=quantization_dtype)
assert isinstance(weights_manifest, list)
model_json[ARTIFACT_WEIGHTS_MANIFEST_KEY] = weights_manifest
with open(output_graph, 'wt') as f:
json.dump(model_json, f)
def write_artifacts(topology,
weights,
output_graph,
tf_version,
signature_def,
quantization_dtype_map=None,
weight_shard_size_bytes=1024 * 1024 * 4,
initializer_graph_def=None,
metadata=None):
"""Writes weights and topology to the output_dir.
If `topology` is Falsy (e.g., `None`), only emit weights to output_dir.
Args:
topology: tf.GraphDef TensorFlow GraphDef proto object, which represents
the model topology.
weights: an array of weight groups (as defined in tfjs write_weights).
output_graph: the output file name to hold all the contents.
tf_version: Tensorflow version of the input graph.
signature_def: the SignatureDef of the inference graph.
quantization_dtype_map: A mapping from dtype
(`uint8`, `uint16`, `float16`) to weights names. The weight mapping
supports wildcard substitution.
weight_shard_size_bytes: Shard size (in bytes) of the weight files.
The size of each weight file will be <= this value.
initializer_graph_def: tf.GraphDef proto object for initializer graph.
metadata: User defined metadata map.
"""
model_json = {
common.FORMAT_KEY:
common.TFJS_GRAPH_MODEL_FORMAT,
# TODO(piyu): Add tensorflow version below by using `meta_info_def`.
common.GENERATED_BY_KEY:
tf_version,
common.CONVERTED_BY_KEY:
common.get_converted_by(),
common.SIGNATURE_KEY:
MessageToDict(signature_def),
}
model_json[common.ARTIFACT_MODEL_TOPOLOGY_KEY] = topology or None
if metadata:
model_json[common.USER_DEFINED_METADATA_KEY] = metadata
if initializer_graph_def and initializer_graph_def.node:
model_json[common.ARTIFACT_MODEL_INITIALIZER] = MessageToDict(
initializer_graph_def)
weights_manifest = write_weights.write_weights(
weights,
os.path.dirname(output_graph),
write_manifest=False,
quantization_dtype_map=quantization_dtype_map,
shard_size_bytes=weight_shard_size_bytes)
assert isinstance(weights_manifest, list)
model_json[common.ARTIFACT_WEIGHTS_MANIFEST_KEY] = weights_manifest
with tf.io.gfile.GFile(output_graph, 'w') as f:
json.dump(model_json, f)
def write_artifacts(topology,
weights,
output_dir,
quantization_dtype_map=None,
weight_shard_size_bytes=1024 * 1024 * 4,
metadata=None):
"""Writes weights and topology to the output_dir.
If `topology` is Falsy (e.g., `None`), only emit weights to output_dir.
Args:
topology: a JSON dictionary, representing the Keras config.
weights: an array of weight groups (as defined in tfjs write_weights).
output_dir: the directory to hold all the contents.
quantization_dtype_map: (Optional) A mapping from dtype
(`uint8`, `uint16`, `float16`) to weights names. The weight mapping
supports wildcard substitution.
weight_shard_size_bytes: Shard size (in bytes) of the weight files.
The size of each weight file will be <= this value.
metadata: User defined metadata map.
"""
# TODO(cais, nielsene): This method should allow optional arguments of
# `write_weights.write_weights` (e.g., shard size) and forward them.
# We write the topology after since write_weights makes no promises about
# preserving directory contents.
if not (isinstance(weight_shard_size_bytes, int)
and weight_shard_size_bytes > 0):
raise ValueError(
'Expected weight_shard_size_bytes to be a positive integer, '
'but got %s' % weight_shard_size_bytes)
if os.path.isfile(output_dir):
raise ValueError(
'Path "%d" already exists as a file (not a directory).' %
output_dir)
model_json = {
common.FORMAT_KEY: common.TFJS_LAYERS_MODEL_FORMAT,
common.GENERATED_BY_KEY: _get_generated_by(topology),
common.CONVERTED_BY_KEY: common.get_converted_by()
}
if metadata:
model_json[common.USER_DEFINED_METADATA_KEY] = metadata
model_json[common.ARTIFACT_MODEL_TOPOLOGY_KEY] = topology or None
weights_manifest = write_weights.write_weights(
weights,
output_dir,
write_manifest=False,
quantization_dtype_map=quantization_dtype_map,
shard_size_bytes=weight_shard_size_bytes)
assert isinstance(weights_manifest, list)
model_json[common.ARTIFACT_WEIGHTS_MANIFEST_KEY] = weights_manifest
model_json_path = os.path.join(output_dir,
common.ARTIFACT_MODEL_JSON_FILE_NAME)
with open(model_json_path, 'wt') as f:
json.dump(model_json, f)
def extract_weights(graph_def, output_graph, quantization_dtype=None):
"""Takes a Python GraphDef object and extract the weights.
Args:
graph_def: tf.GraphDef TensorFlow GraphDef proto object, which represents
the model topology.
quantization_dtype: An optional numpy dtype to quantize weights to for
compression. Only np.uint8 and np.uint16 are supported.
"""
constants = [node for node in graph_def.node if node.op == 'Const']
constInputs = {}
# removed the conditional inputs for constants
for const in constants:
constInputs[const.name] = const.input[:]
del const.input[:]
print('Writing weight file ' + output_graph + '...')
const_manifest = []
path = os.path.dirname(output_graph)
graph = tf.Graph()
with tf.Session(graph=graph) as sess:
tf.import_graph_def(graph_def, name='')
for const in constants:
tensor = graph.get_tensor_by_name(const.name + ':0')
value = tensor.eval(session=sess)
if not isinstance(value, np.ndarray):
value = np.array(value)
# Restore the conditional inputs
const_manifest.append({'name': const.name, 'data': value})
const.input[:] = constInputs[const.name]
# Remove the binary array from tensor and save it to the external file.
for field_name in CLEARED_TENSOR_FIELDS:
const.attr["value"].tensor.ClearField(field_name)
write_weights.write_weights([const_manifest],
path,
quantization_dtype=quantization_dtype)
file_io.atomic_write_string_to_file(os.path.abspath(output_graph),
graph_def.SerializeToString())
def extract_weights(graph_def,
output_graph,
quantization_dtype=None):
"""Takes a Python GraphDef object and extract the weights.
Args:
graph_def: tf.GraphDef TensorFlow GraphDef proto object, which represents
the model topology.
quantization_dtype: An optional numpy dtype to quantize weights to for
compression. Only np.uint8 and np.uint16 are supported.
"""
constants = [node for node in graph_def.node if node.op == 'Const']
constInputs = {}
# removed the conditional inputs for constants
for const in constants:
constInputs[const.name] = const.input[:]
del const.input[:]
print('Writing weight file ' + output_graph + '...')
const_manifest = []
path = os.path.dirname(output_graph)
graph = tf.Graph()
with tf.Session(graph=graph) as sess:
tf.import_graph_def(graph_def, name='')
for const in constants:
tensor = graph.get_tensor_by_name(const.name + ':0')
value = tensor.eval(session=sess)
if not isinstance(value, np.ndarray):
value = np.array(value)
# Restore the conditional inputs
const_manifest.append({'name': const.name, 'data': value})
const.input[:] = constInputs[const.name]
# Remove the binary array from tensor and save it to the external file.
const.attr["value"].tensor.ClearField('tensor_content')
write_weights.write_weights(
[const_manifest], path, quantization_dtype=quantization_dtype)
file_io.atomic_write_string_to_file(
os.path.abspath(output_graph), graph_def.SerializeToString())
def dump_checkpoint(
checkpoint_file, output_dir, remove_variables_regex=None,
quantization_dtype=None):
reader = tf.train.NewCheckpointReader(checkpoint_file)
var_to_shape_map = reader.get_variable_to_shape_map()
remove_variables_regex_re = (
re.compile(remove_variables_regex) if remove_variables_regex else None)
entries = []
for var_name in var_to_shape_map:
if (remove_variables_regex_re and remove_variables_regex_re.match(var_name)
or var_name == 'global_step'):
print('Ignoring ' + var_name)
continue
tensor = reader.get_tensor(var_name)
entries.append({'name': var_name, 'data': tensor})
print('Dumping ' + var_name)
write_weights([entries], output_dir, quantization_dtype=quantization_dtype)
def test_1_group_1_weight_string(self):
groups = [[{
'name':
'weight1',
'data':
np.array([['здраво', 'end'], ['test', 'a']], 'object')
}]]
manifest = write_weights.write_weights(groups,
TMP_DIR,
shard_size_bytes=4 * 1024 *
1024)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(manifest, [{
'paths': ['group1-shard1of1.bin'],
'weights': [{
'name': 'weight1',
'shape': [2, 2],
'dtype': 'string'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1.bin')
with open(weights_path, 'rb') as f:
weight_bytes = f.read()
self.assertEqual(len(weight_bytes), 36)
# 'здраво'
size = np.frombuffer(weight_bytes[:4], 'uint32')[0]
self.assertEqual(size, 12) # 6 cyrillic chars (2 bytes each).
string = weight_bytes[4:16].decode('utf-8')
self.assertEqual(string, u'здраво')
# 'end'
size = np.frombuffer(weight_bytes[16:20], 'uint32')[0]
self.assertEqual(size, 3) # 3 ascii chars.
string = weight_bytes[20:23].decode('utf-8')
self.assertEqual(string, u'end')
# 'test'
size = np.frombuffer(weight_bytes[23:27], 'uint32')[0]
self.assertEqual(size, 4) # 4 ascii chars.
string = weight_bytes[27:31].decode('utf-8')
self.assertEqual(string, u'test')
# 'a'
size = np.frombuffer(weight_bytes[31:35], 'uint32')[0]
self.assertEqual(size, 1) # 4 ascii chars.
string = weight_bytes[35:36].decode('utf-8')
self.assertEqual(string, u'a')
def testReadOneGroup(self):
groups = [[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]]
manifest = write_weights.write_weights(groups, self._tmp_dir)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(manifest, self._tmp_dir)
self.assertEqual(1, len(read_output))
self.assertEqual(1, len(read_output[0]))
self.assertEqual('weight1', read_output[0][0]['name'])
self.assertTrue(
np.allclose(groups[0][0]['data'], read_output[0][0]['data']))
def testReadOneGroupWithShards(self):
groups = [[{
'name': 'weight1',
'data': np.random.rand(1, 100).astype(np.float32)
}]]
manifest = write_weights.write_weights(groups, self._tmp_dir)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(manifest, self._tmp_dir)
self.assertEqual(1, len(read_output))
self.assertEqual(1, len(read_output[0]))
self.assertEqual('weight1', read_output[0][0]['name'])
self.assertTrue(
np.allclose(groups[0][0]['data'], read_output[0][0]['data']))
def test_quantize_group(self):
groups = [[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}, {
'name': 'weight2',
'data': np.array([4, 5], 'int32')
}]]
manifest_json = write_weights.write_weights(
groups,
TMP_DIR,
shard_size_bytes=8 * 4,
quantization_dtype=np.uint8)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
q, s, m = zip(
quantization.quantize_weights(groups[0][0]['data'], np.uint8),
quantization.quantize_weights(groups[0][1]['data'], np.uint8))
self.assertEqual(manifest, [{
'paths': ['group1-shard1of1'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32',
'quantization': {
'min': m[0],
'scale': s[0],
'dtype': 'uint8'
}
}, {
'name': 'weight2',
'shape': [2],
'dtype': 'int32',
'quantization': {
'min': m[1],
'scale': s[1],
'dtype': 'uint8'
}
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1')
weights = np.fromfile(weights_path, 'uint8')
np.testing.assert_array_equal(weights, np.concatenate([q[0], q[1]]))
def write_artifacts(topology,
weights,
output_graph,
tf_version,
signature_def,
quantization_dtype=None,
weight_shard_size_bytes=1024 * 1024 * 4):
"""Writes weights and topology to the output_dir.
If `topology` is Falsy (e.g., `None`), only emit weights to output_dir.
Args:
topology: tf.GraphDef TensorFlow GraphDef proto object, which represents
the model topology.
weights: an array of weight groups (as defined in tfjs write_weights).
output_graph: the output file name to hold all the contents.
tf_version: Tensorflow version of the input graph.
signature_def: the SignatureDef of the inference graph.
quantization_dtype: An optional numpy dtype to quantize weights to for
compression. Only np.uint8 and np.uint16 are supported.
weight_shard_size_bytes: Shard size (in bytes) of the weight files.
The size of each weight file will be <= this value.
"""
model_json = {
common.FORMAT_KEY:
common.TFJS_GRAPH_MODEL_FORMAT,
# TODO(piyu): Add tensorflow version below by using `meta_info_def`.
common.GENERATED_BY_KEY:
tf_version,
common.CONVERTED_BY_KEY:
common.get_converted_by(),
common.USER_DEFINED_METADATA_KEY: {
common.SIGNATURE_KEY: MessageToDict(signature_def)
}
}
model_json[common.ARTIFACT_MODEL_TOPOLOGY_KEY] = topology or None
weights_manifest = write_weights.write_weights(
weights,
os.path.dirname(output_graph),
write_manifest=False,
quantization_dtype=quantization_dtype,
shard_size_bytes=weight_shard_size_bytes)
assert isinstance(weights_manifest, list)
model_json[common.ARTIFACT_WEIGHTS_MANIFEST_KEY] = weights_manifest
with tf.io.gfile.GFile(output_graph, 'w') as f:
json.dump(model_json, f)
def testReadOneGroupFlattened(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}]
]
manifest_json = write_weights.write_weights(groups, self._tmp_dir)
manifest = json.loads(manifest_json)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(
manifest, self._tmp_dir, flatten=True)
self.assertEqual(1, len(read_output))
self.assertEqual('weight1', read_output[0]['name'])
self.assertTrue(np.allclose(groups[0][0]['data'], read_output[0]['data']))
def testReadQuantizedWeights(self):
groups = [[{
'name': 'weight1',
'data': np.array([0, 1, 2, 3], 'float32')
}]]
manifest_json = write_weights.write_weights(
groups, self._tmp_dir, quantization_dtype=np.uint8)
manifest = json.loads(manifest_json)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(manifest, self._tmp_dir)
self.assertEqual(1, len(read_output))
self.assertEqual(1, len(read_output[0]))
self.assertEqual('weight1', read_output[0][0]['name'])
self.assertTrue(
np.allclose(groups[0][0]['data'], read_output[0][0]['data']))
def testReadOneGroupString(self):
groups = [[{
'name': 'weight1',
'data': np.array([['test', 'a'], ['b', 'c']], 'object')
}]]
manifest = write_weights.write_weights(groups, self._tmp_dir)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(manifest, self._tmp_dir)
self.assertEqual(1, len(read_output))
self.assertEqual(1, len(read_output[0]))
self.assertEqual('weight1', read_output[0][0]['name'])
np.testing.assert_array_equal(
read_output[0][0]['data'],
np.array([[u'test'.encode('utf-8'), u'a'.encode('utf-8')],
[u'b'.encode('utf-8'), u'c'.encode('utf-8')]], 'object'))
def test_quantize_group(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}, {
'name': 'weight2',
'data': np.array([4, 5], 'int32')
}]
]
manifest_json = write_weights.write_weights(
groups, TMP_DIR, shard_size_bytes=8 * 4, quantization_dtype=np.uint8)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
q, s, m = zip(
quantization.quantize_weights(groups[0][0]['data'], np.uint8),
quantization.quantize_weights(groups[0][1]['data'], np.uint8))
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of1'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32',
'quantization': {
'min': m[0], 'scale': s[0], 'dtype': 'uint8'
}
}, {
'name': 'weight2',
'shape': [2],
'dtype': 'int32',
'quantization': {
'min': m[1], 'scale': s[1], 'dtype': 'uint8'
}
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1')
weights = np.fromfile(weights_path, 'uint8')
np.testing.assert_array_equal(weights, np.concatenate([q[0], q[1]]))
def testReadOneGroupWithShards(self):
groups = [
[{
'name': 'weight1',
'data': np.random.rand(1, 100).astype(np.float32)
}]
]
manifest_json = write_weights.write_weights(groups, self._tmp_dir)
manifest = json.loads(manifest_json)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(manifest, self._tmp_dir)
self.assertEqual(1, len(read_output))
self.assertEqual(1, len(read_output[0]))
self.assertEqual('weight1', read_output[0][0]['name'])
self.assertTrue(
np.allclose(groups[0][0]['data'], read_output[0][0]['data']))
def testReadQuantizedWeights(self):
groups = [
[{
'name': 'weight1',
'data': np.array([0, 1, 2, 3], 'float32')
}]
]
manifest_json = write_weights.write_weights(
groups, self._tmp_dir, quantization_dtype=np.uint8)
manifest = json.loads(manifest_json)
# Read the weights using `read_weights`.
read_output = read_weights.read_weights(manifest, self._tmp_dir)
self.assertEqual(1, len(read_output))
self.assertEqual(1, len(read_output[0]))
self.assertEqual('weight1', read_output[0][0]['name'])
self.assertTrue(
np.allclose(groups[0][0]['data'], read_output[0][0]['data']))
def write_artifacts(self,
topology,
weights,
output_dir,
quantization_dtype=None):
"""Writes weights and topology to the output_dir.
If `topology` is Falsy (e.g., `None`), only emit weights to output_dir.
Args:
topology: a JSON dictionary, representing the Keras config.
weights: an array of weight groups (as defined in tfjs write_weights).
output_dir: the directory to hold all the contents.
quantization_dtype: An optional numpy dtype to quantize weights to for
compression. Only np.uint8 and np.uint16 are supported.
"""
# TODO(cais, nielsene): This method should allow optional arguments of
# `write_weights.write_weights` (e.g., shard size) and forward them.
# We write the topology after since write_weights makes no promises about
# preserving directory contents.
if os.path.isfile(output_dir):
raise ValueError(
'Path "%d" already exists as a file (not a directory).' % output_dir)
model_json = {}
model_json[ARTIFACT_MODEL_TOPOLOGY_KEY] = topology or None
weights_manifest = write_weights.write_weights(
weights, output_dir, write_manifest=False,
quantization_dtype=quantization_dtype)
if not isinstance(weights_manifest, list):
weights_manifest = json.loads(weights_manifest)
assert isinstance(weights_manifest, list)
model_json[ARTIFACT_WEIGHTS_MANIFEST_KEY] = weights_manifest
model_json_path = os.path.join(output_dir, ARTIFACT_MODEL_JSON_FILE_NAME)
with open(model_json_path, 'wt') as f:
json.dump(model_json, f)
def test_1_group_2_weights_packed(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}, {
'name': 'weight2',
'data': np.array([4, 5], 'float32')
}]
]
# Shard size is larger than the sum of the two weights so they get packed.
manifest_json = write_weights.write_weights(
groups, TMP_DIR, shard_size_bytes=8 * 4)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of1'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32'
}, {
'name': 'weight2',
'shape': [2],
'dtype': 'float32'
}]
}])
weights_path = os.path.join(TMP_DIR, 'group1-shard1of1')
weights = np.fromfile(weights_path, 'float32')
np.testing.assert_array_equal(weights, np.array([1, 2, 3, 4, 5], 'float32'))
def test_1_group_2_packed_sharded_multi_dtype(self):
groups = [
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'int32')
}, {
'name': 'weight2',
'data': np.array([True, False, False, True], 'bool')
}, {
'name': 'weight3',
'data': np.array([4.1, 5.1], 'float32')
}]
]
# Shard size is smaller than the sum of the weights so they get packed and
# then sharded. The two buffers will get sharded into 3 files, with shapes
# [2], [2], and [1]. The second shard is a mixed dtype.
manifest_json = write_weights.write_weights(
groups, TMP_DIR, shard_size_bytes=2 * 4)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of3',
'group1-shard2of3',
'group1-shard3of3'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'int32'
}, {
'name': 'weight2',
'shape': [4],
'dtype': 'bool'
}, {
'name': 'weight3',
'shape': [2],
'dtype': 'float32'
}]
}])
shard_1_path = os.path.join(TMP_DIR, 'group1-shard1of3')
shard_1 = np.fromfile(shard_1_path, 'int32')
np.testing.assert_array_equal(shard_1, np.array([1, 2], 'int32'))
# Shard 2 has a mixed dtype so we parse the bytes directly.
shard_2_path = os.path.join(TMP_DIR, 'group1-shard2of3')
with open(shard_2_path, 'rb') as f:
shard_2_bytes = f.read()
shard_2_int = np.frombuffer(shard_2_bytes[:4], 'int32')
np.testing.assert_array_equal(shard_2_int, np.array([3], 'int32'))
shard_2_bool = np.frombuffer(shard_2_bytes[4:], 'bool')
np.testing.assert_array_equal(
shard_2_bool, np.array([True, False, False, True], 'bool'))
shard_3_path = os.path.join(TMP_DIR, 'group1-shard3of3')
shard_3 = np.fromfile(shard_3_path, 'float32')
np.testing.assert_array_equal(shard_3, np.array([4.1, 5.1], 'float32'))
def test_2_groups_4_weights_sharded_packed(self):
groups = [
# Group 1
[{
'name': 'weight1',
'data': np.array([1, 2, 3], 'float32')
}, {
'name': 'weight2',
'data': np.array([[4, 5], [6, 7]], 'float32')
}],
# Group 2
[{
'name': 'weight3',
'data': np.array([1, 2, 3, 4], 'int32')
}, {
'name': 'weight4',
'data': np.array([[1.1, 1.2], [1.3, 1.4], [1.5, 1.6]], 'float32')
}]
]
manifest_json = write_weights.write_weights(
groups, TMP_DIR, shard_size_bytes=4 * 4)
manifest = json.loads(manifest_json)
self.assertTrue(
os.path.isfile(os.path.join(TMP_DIR, 'weights_manifest.json')),
'weights_manifest.json does not exist')
self.assertEqual(
manifest,
[{
'paths': ['group1-shard1of2', 'group1-shard2of2'],
'weights': [{
'name': 'weight1',
'shape': [3],
'dtype': 'float32'
}, {
'name': 'weight2',
'shape': [2, 2],
'dtype': 'float32'
}]
}, {
'paths': ['group2-shard1of3',
'group2-shard2of3',
'group2-shard3of3'],
'weights': [{
'name': 'weight3',
'shape': [4],
'dtype': 'int32'
}, {
'name': 'weight4',
'shape': [3, 2],
'dtype': 'float32'
}]
}])
group0_shard_1_path = os.path.join(TMP_DIR, 'group1-shard1of2')
group0_shard_1 = np.fromfile(group0_shard_1_path, 'float32')
np.testing.assert_array_equal(
group0_shard_1, np.array([1, 2, 3, 4], 'float32'))
group0_shard_2_path = os.path.join(TMP_DIR, 'group1-shard2of2')
group0_shard_2 = np.fromfile(group0_shard_2_path, 'float32')
np.testing.assert_array_equal(
group0_shard_2, np.array([5, 6, 7], 'float32'))
group1_shard_1_path = os.path.join(TMP_DIR, 'group2-shard1of3')
group1_shard_1 = np.fromfile(group1_shard_1_path, 'int32')
np.testing.assert_array_equal(
group1_shard_1, np.array([1, 2, 3, 4], 'int32'))
group2_shard_2_path = os.path.join(TMP_DIR, 'group2-shard2of3')
group2_shard_2 = np.fromfile(group2_shard_2_path, 'float32')
np.testing.assert_array_equal(
group2_shard_2, np.array([1.1, 1.2, 1.3, 1.4], 'float32'))
group2_shard_3_path = os.path.join(TMP_DIR, 'group2-shard3of3')
group2_shard_3 = np.fromfile(group2_shard_3_path, 'float32')
np.testing.assert_array_equal(
group2_shard_3, np.array([1.5, 1.6], 'float32'))
def test_no_weights_groups_throws(self):
groups = None
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
def test_empty_groups_throws(self):
groups = []
with self.assertRaises(Exception):
write_weights.write_weights(groups, TMP_DIR)
