def dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, output_dir=None, quantization_dtype=None,
split_weights_by_layer=False,
weight_shard_size_bytes=1024 * 1024 * 4):
"""Converts a Keras HDF5 saved-model file to TensorFlow.js format.
Auto-detects saved_model versus weights-only and generates the correct
json in either case. This function accepts Keras HDF5 files in two formats:
- A weights-only HDF5 (e.g., generated with Keras Model's `save_weights()`
method),
- A topology+weights combined HDF5 (e.g., generated with
`tf.keras.model.save_model`).
Args:
h5_path: path to an HDF5 file containing keras model data as a `str`.
output_dir: Output directory to which the TensorFlow.js-format model JSON
file and weights files will be written. If the directory does not exist,
it will be created.
quantization_dtype: The quantized data type to store the weights in
(Default: `None`).
split_weights_by_layer: Whether to split the weights into separate weight
groups (corresponding to separate binary weight files) layer by layer
(Default: `False`).
weight_shard_size_bytes: Shard size (in bytes) of the weight files.
The size of each weight file will be <= this value.
Returns:
(model_json, groups)
model_json: a json dictionary (empty if unused) for model topology.
If `h5_path` points to a weights-only HDF5 file, this return value
will be `None`.
groups: an array of weight_groups as defined in tfjs weights_writer.
"""
if not os.path.exists(h5_path):
raise ValueError('Nonexistent path to HDF5 file: %s' % h5_path)
if os.path.isdir(h5_path):
raise ValueError(
'Expected path to point to an HDF5 file, but it points to a '
'directory: %s' % h5_path)
h5_file = h5py.File(h5_path, 'r')
if 'layer_names' in h5_file.attrs:
model_json = None
groups = conversion.h5_weights_to_tfjs_format(
h5_file, split_by_layer=split_weights_by_layer)
else:
model_json, groups = conversion.h5_merged_saved_model_to_tfjs_format(
h5_file, split_by_layer=split_weights_by_layer)
if output_dir:
if os.path.isfile(output_dir):
raise ValueError(
'Output path "%s" already exists as a file' % output_dir)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
conversion.write_artifacts(
model_json, groups, output_dir, quantization_dtype,
weight_shard_size_bytes=weight_shard_size_bytes)
return model_json, groups
def testConvertMergedModelFromSimpleModelSplitByLayer(self):
input_tensor = tf.keras.layers.Input((3, ))
dense1 = tf.keras.layers.Dense(4,
use_bias=True,
kernel_initializer='ones',
bias_initializer='zeros',
name='MergedDense30')(input_tensor)
output = tf.keras.layers.Dense(2,
use_bias=False,
kernel_initializer='ones',
name='MergedDense40')(dense1)
model = tf.keras.models.Model(inputs=[input_tensor], outputs=[output])
h5_path = os.path.join(self._tmp_dir, 'MyModelMerged.h5')
model.save(h5_path)
if six.PY3:
config_json = json.loads(model.to_json())
else:
config_json = json.loads(model.to_json(), encoding='utf8')
# Load the saved weights as a JSON string.
out, groups = conversion.h5_merged_saved_model_to_tfjs_format(
h5py.File(h5_path), split_by_layer=True)
saved_topology = out['model_config']
# check the model topology was stored
self.assertEqual(config_json['class_name'],
saved_topology['class_name'])
self.assertEqual(config_json['config'], saved_topology['config'])
# Check the loaded weights.
# Due to `split_by_layer=True`, there ought to be two weight groups,
# because the model has two layers.
self.assertEqual(2, len(groups))
self.assertEqual(tf.keras.__version__, out['keras_version'])
self.assertEqual('tensorflow', out['backend'])
self.assertEqual(2, len(groups[0]))
kernel1 = groups[0][0]
self.assertEqual('MergedDense30/kernel', kernel1['name'])
self.assertEqual('float32', kernel1['data'].dtype)
self.assertEqual((3, 4), kernel1['data'].shape)
self.assertTrue(np.allclose(np.ones([3, 4]), kernel1['data']))
bias1 = groups[0][1]
self.assertEqual('MergedDense30/bias', bias1['name'])
self.assertEqual('float32', bias1['data'].dtype)
self.assertEqual((4, ), bias1['data'].shape)
self.assertTrue(np.allclose(np.zeros([4]), bias1['data']))
self.assertEqual(1, len(groups[1]))
kernel2 = groups[1][0]
self.assertEqual('MergedDense40/kernel', kernel2['name'])
self.assertEqual('float32', kernel2['data'].dtype)
self.assertEqual((4, 2), kernel2['data'].shape)
self.assertTrue(np.allclose(np.ones([4, 2]), kernel2['data']))
def testConvertMergedModelFromSimpleModelNoSplitByLayer(self):
input_tensor = keras.layers.Input((3, ))
dense1 = keras.layers.Dense(4,
use_bias=True,
kernel_initializer='ones',
bias_initializer='zeros',
name='MergedDense10')(input_tensor)
output = keras.layers.Dense(2,
use_bias=False,
kernel_initializer='ones',
name='MergedDense20')(dense1)
model = keras.models.Model(inputs=[input_tensor], outputs=[output])
h5_path = os.path.join(self._tmp_dir, 'MyModelMerged.h5')
model.save(h5_path)
config_json = json.loads(model.to_json(), encoding='utf8')
# Load the saved weights as a JSON string.
out, groups = conversion.h5_merged_saved_model_to_tfjs_format(
h5py.File(h5_path))
saved_topology = out['model_config']
# check the model topology was stored
self.assertEqual(config_json['class_name'],
saved_topology['class_name'])
self.assertEqual(config_json['config'], saved_topology['config'])
# Check the loaded weights.
# By default, all weights of the model ought to be put in the same group.
self.assertEqual(1, len(groups))
self.assertEqual(keras.__version__, out['keras_version'])
self.assertEqual('tensorflow', out['backend'])
weight_group = groups[0]
self.assertEqual(3, len(weight_group))
kernel1 = weight_group[0]
self.assertEqual('MergedDense10/kernel', kernel1['name'])
self.assertEqual('float32', kernel1['data'].dtype)
self.assertEqual((3, 4), kernel1['data'].shape)
self.assertTrue(np.allclose(np.ones([3, 4]), kernel1['data']))
bias1 = weight_group[1]
self.assertEqual('MergedDense10/bias', bias1['name'])
self.assertEqual('float32', bias1['data'].dtype)
self.assertEqual((4, ), bias1['data'].shape)
self.assertTrue(np.allclose(np.zeros([4]), bias1['data']))
kernel2 = weight_group[2]
self.assertEqual('MergedDense20/kernel', kernel2['name'])
self.assertEqual('float32', kernel2['data'].dtype)
self.assertEqual((4, 2), kernel2['data'].shape)
self.assertTrue(np.allclose(np.ones([4, 2]), kernel2['data']))
