def testTensorflowjsToKerasConversionSucceeds(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
sequential_model = keras.models.Sequential([
keras.layers.Dense(3,
input_shape=(2, ),
use_bias=True,
kernel_initializer='ones',
name='Dense1'),
keras.layers.Dense(1,
use_bias=False,
kernel_initializer='ones',
name='Dense2')
])
h5_path = os.path.join(self._tmp_dir, 'SequentialModel.h5')
sequential_model.save(h5_path)
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, output_dir=self._tmp_dir)
old_model_json = sequential_model.to_json()
# Convert the tensorflowjs artifacts to a new H5 file.
new_h5_path = os.path.join(self._tmp_dir, 'new.h5')
converter.dispatch_tensorflowjs_to_keras_h5_conversion(
os.path.join(self._tmp_dir, 'model.json'), new_h5_path)
# Load the new H5 and compare the model JSONs.
with tf.Graph().as_default(), tf.compat.v1.Session():
new_model = keras.models.load_model(new_h5_path)
self.assertEqual(old_model_json, new_model.to_json())
def testConvertTfjsLayersModelWithUint8Quantization(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
model = self._createSimpleSequentialModel()
weights = model.get_weights()
total_weight_bytes = sum(np.size(w) for w in weights) * 4
# Save the keras model to a .h5 file.
h5_path = os.path.join(self._tmp_dir, 'model.h5')
model.save(h5_path)
# Convert the keras SavedModel to tfjs format.
tfjs_output_dir = os.path.join(self._tmp_dir, 'tfjs')
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, tfjs_output_dir)
weight_shard_size_bytes = int(total_weight_bytes * 2)
# Due to the shard size, there ought to be 1 shard after conversion.
# Convert the tfjs model to another tfjs model, with quantization.
sharded_model_path = os.path.join(self._tmp_dir, 'sharded_model')
converter.dispatch_tensorflowjs_to_tensorflowjs_conversion(
os.path.join(tfjs_output_dir, 'model.json'), sharded_model_path,
quantization_dtype=np.uint8,
weight_shard_size_bytes=weight_shard_size_bytes)
# Check the number of quantized files and their sizes.
weight_files = sorted(
glob.glob(os.path.join(sharded_model_path, 'group*.bin')))
self.assertEqual(len(weight_files), 1)
weight_file_size = os.path.getsize(weight_files[0])
# The size of the saved weight file should reflect the result of the
# uint16 quantization.
self.assertEqual(weight_file_size, total_weight_bytes / 4)
def testConvertTfjsLayersModelWithShardSizeGreaterThanTotalWeightSize(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
model = self._createSimpleSequentialModel()
weights = model.get_weights()
total_weight_bytes = sum(np.size(w) for w in weights) * 4
# Save the keras model to a .h5 file.
h5_path = os.path.join(self._tmp_dir, 'model.h5')
model.save(h5_path)
# Convert the keras SavedModel to tfjs format.
tfjs_output_dir = os.path.join(self._tmp_dir, 'tfjs')
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, tfjs_output_dir)
weight_shard_size_bytes = int(total_weight_bytes * 2)
# Due to the shard size, there ought to be 1 shard after conversion.
# Convert the tfjs model to another tfjs model, with a specified weight
# shard size.
sharded_model_path = os.path.join(self._tmp_dir, 'sharded_model')
converter.dispatch_tensorflowjs_to_tensorflowjs_conversion(
os.path.join(tfjs_output_dir, 'model.json'), sharded_model_path,
weight_shard_size_bytes=weight_shard_size_bytes)
# Check the number of sharded files and their sizes.
weight_files = sorted(
glob.glob(os.path.join(sharded_model_path, 'group*.bin')))
self.assertEqual(len(weight_files), 1)
weight_file_sizes = [os.path.getsize(f) for f in weight_files]
self.assertEqual(sum(weight_file_sizes), total_weight_bytes)
def testConvertSavedKerasModeltoTfLayersModelSharded(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
sequential_model = tf.keras.models.Sequential([
tf.keras.layers.Dense(3,
input_shape=(2, ),
use_bias=True,
kernel_initializer='ones',
name='Dense1')
])
h5_path = os.path.join(self._tmp_dir, 'SequentialModel.h5')
sequential_model.save(h5_path)
weights = sequential_model.get_weights()
total_weight_bytes = sum(np.size(w) for w in weights) * 4
# Due to the shard size, there ought to be 4 shards after conversion.
weight_shard_size_bytes = int(total_weight_bytes * 0.3)
# Convert Keras model to tfjs_layers_model format.
output_dir = os.path.join(self._tmp_dir, 'sharded_tfjs')
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path,
output_dir,
weight_shard_size_bytes=weight_shard_size_bytes)
weight_files = sorted(
glob.glob(os.path.join(output_dir, 'group*.bin')))
self.assertEqual(len(weight_files), 4)
weight_file_sizes = [os.path.getsize(f) for f in weight_files]
self.assertEqual(sum(weight_file_sizes), total_weight_bytes)
self.assertEqual(weight_file_sizes[0], weight_file_sizes[1])
self.assertEqual(weight_file_sizes[0], weight_file_sizes[2])
self.assertLess(weight_file_sizes[3], weight_file_sizes[0])
def testWrongConverterRaisesCorrectErrorMessage(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
model = self._createSimpleSequentialModel()
keras.experimental.export_saved_model(model, self._tmp_dir)
# Convert the keras SavedModel to tfjs format.
tfjs_output_dir = os.path.join(self._tmp_dir, 'tfjs')
# Use wrong dispatcher.
with self.assertRaisesRegexp( # pylint: disable=deprecated-method
ValueError,
r'Expected path to point to an HDF5 file, but it points to a '
r'directory'):
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
self._tmp_dir, tfjs_output_dir)
def testConvertSavedKerasModelNoSplitByLayer(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
input_tensor = keras.layers.Input((3,))
dense1 = keras.layers.Dense(
4, use_bias=True, kernel_initializer='ones', bias_initializer='zeros',
name='MergedDense1')(input_tensor)
output = keras.layers.Dense(
2, use_bias=False,
kernel_initializer='ones', name='MergedDense2')(dense1)
model = keras.models.Model(inputs=[input_tensor], outputs=[output])
h5_path = os.path.join(self._tmp_dir, 'MyModelMerged.h5')
model.save(h5_path)
# Load the saved weights as a JSON string.
model_json, groups = (
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, output_dir=self._tmp_dir))
# check the model topology was stored
self.assertIsInstance(model_json['model_config'], dict)
self.assertIsInstance(model_json['model_config']['config'], dict)
self.assertIn('layers', model_json['model_config']['config'])
# Check the loaded weights.
self.assertEqual(keras.__version__, model_json['keras_version'])
self.assertEqual('tensorflow', model_json['backend'])
self.assertEqual(1, len(groups))
self.assertEqual(3, len(groups[0]))
# contents of weights are verified in tests of the library code
# Check the content of the output directory.
output_json = json.load(
open(os.path.join(self._tmp_dir, 'model.json'), 'rt'))
self.assertEqual(model_json, output_json['modelTopology'])
self.assertIsInstance(output_json['weightsManifest'], list)
self.assertTrue(glob.glob(os.path.join(self._tmp_dir, 'group*-*')))
def testConvertWeightsFromSequentialModel(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
sequential_model = keras.models.Sequential([
keras.layers.Dense(
3, input_shape=(2,), use_bias=True, kernel_initializer='ones',
name='Dense1'),
keras.layers.Dense(
1, use_bias=False, kernel_initializer='ones', name='Dense2')])
h5_path = os.path.join(self._tmp_dir, 'SequentialModel.h5')
sequential_model.save_weights(h5_path)
# Load the saved weights as a JSON string.
model_json, groups = (
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, output_dir=self._tmp_dir))
self.assertIsNone(model_json)
# Check the loaded weights.
self.assertEqual(1, len(groups))
self.assertEqual(3, len(groups[0]))
# contents of weights are verified in tests of the library code
# Check the content of the output directory.
output_json = json.load(
open(os.path.join(self._tmp_dir, 'model.json'), 'rt'))
self.assertEqual(model_json, output_json['modelTopology'])
self.assertIsInstance(output_json['weightsManifest'], list)
self.assertTrue(glob.glob(os.path.join(self._tmp_dir, 'group*-*')))
def testWeightsOnly(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
input_tensor = tf.keras.layers.Input((3, ))
dense1 = tf.keras.layers.Dense(4,
use_bias=True,
kernel_initializer='ones',
bias_initializer='zeros',
name='MyDense1')(input_tensor)
output = tf.keras.layers.Dense(2,
use_bias=False,
kernel_initializer='ones',
name='MyDense2')(dense1)
model = tf.keras.models.Model(inputs=[input_tensor],
outputs=[output])
h5_path = os.path.join(self._tmp_dir, 'MyModel.h5')
model.save_weights(h5_path)
# Load the saved weights as a JSON string.
model_json, groups = (
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, output_dir=self._tmp_dir))
self.assertIsNone(model_json)
# Check the loaded weights.
self.assertEqual(1, len(groups))
self.assertEqual(3, len(groups[0]))
# contents of weights are verified in tests of the library code
# Check the content of the output directory.
output_json = json.load(
open(os.path.join(self._tmp_dir, 'model.json'), 'rt'))
self.assertEqual(model_json, output_json['modelTopology'])
self.assertIsInstance(output_json['weightsManifest'], list)
self.assertTrue(glob.glob(os.path.join(self._tmp_dir, 'group*-*')))
def testTensorflowjsToKerasConversionFailsOnExistingDirOutputPath(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
sequential_model = keras.models.Sequential([
keras.layers.Dense(
3, input_shape=(2,), use_bias=True, kernel_initializer='ones',
name='Dense1'),
keras.layers.Dense(
1, use_bias=False, kernel_initializer='ones', name='Dense2')])
h5_path = os.path.join(self._tmp_dir, 'SequentialModel.h5')
sequential_model.save(h5_path)
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, output_dir=self._tmp_dir)
with self.assertRaisesRegexp( # pylint: disable=deprecated-method
ValueError, r'but received an existing directory'):
converter.dispatch_tensorflowjs_to_keras_h5_conversion(
os.path.join(self._tmp_dir, 'model.json'), self._tmp_dir)
def testOutpuDirAsAnExistingFileLeadsToValueError(self):
output_path = os.path.join(self._tmp_dir, 'foo_model')
with open(output_path, 'wt') as f:
f.write('\n')
with tf.Graph().as_default(), tf.compat.v1.Session():
sequential_model = keras.models.Sequential([
keras.layers.Dense(
3, input_shape=(2,), use_bias=True, kernel_initializer='ones',
name='Dense1')])
h5_path = os.path.join(self._tmp_dir, 'SequentialModel.h5')
sequential_model.save_weights(h5_path)
with self.assertRaisesRegexp( # pylint: disable=deprecated-method
ValueError, r'already exists as a file'):
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, output_dir=output_path)
def testConvertModelForNonexistentDirCreatesDir(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
output_dir = os.path.join(self._tmp_dir, 'foo_model')
sequential_model = keras.models.Sequential([
keras.layers.Dense(
3, input_shape=(2,), use_bias=True, kernel_initializer='ones',
name='Dense1')])
h5_path = os.path.join(self._tmp_dir, 'SequentialModel.h5')
sequential_model.save_weights(h5_path)
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, output_dir=output_dir)
# Check the content of the output directory.
output_json = json.load(
open(os.path.join(output_dir, 'model.json'), 'rt'))
self.assertIsNone(output_json['modelTopology'])
self.assertIsInstance(output_json['weightsManifest'], list)
self.assertTrue(glob.glob(os.path.join(output_dir, 'group*-*')))
def testConvertTfjsLayersModelToKerasSavedModel(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
model = self._createSimpleSequentialModel()
# Save the keras model to a .h5 file.
h5_path = os.path.join(self._tmp_dir, 'model.h5')
model.save(h5_path)
# Convert the keras SavedModel to tfjs format.
tfjs_output_dir = os.path.join(self._tmp_dir, 'tfjs')
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, tfjs_output_dir)
# Convert the tfjs LayersModel to tf.keras SavedModel.
keras_saved_model_dir = os.path.join(self._tmp_dir, 'saved_model')
converter.dispatch_tensorflowjs_to_keras_saved_model_conversion(
os.path.join(tfjs_output_dir, 'model.json'), keras_saved_model_dir)
# Check the files of the keras SavedModel.
files = glob.glob(os.path.join(keras_saved_model_dir, '*'))
self.assertIn(os.path.join(keras_saved_model_dir, 'saved_model.pb'), files)
self.assertIn(os.path.join(keras_saved_model_dir, 'variables'), files)
self.assertIn(os.path.join(keras_saved_model_dir, 'assets'), files)
def testTfjsLayer2TfjsLayersConversionWithExistingFilePathFails(self):
with tf.Graph().as_default(), tf.compat.v1.Session():
model = self._createSimpleSequentialModel()
# Save the keras model to a .h5 file.
h5_path = os.path.join(self._tmp_dir, 'model.h5')
model.save(h5_path)
# Convert the keras SavedModel to tfjs format.
tfjs_output_dir = os.path.join(self._tmp_dir, 'tfjs')
converter.dispatch_keras_h5_to_tfjs_layers_model_conversion(
h5_path, tfjs_output_dir)
# Convert the tfjs model to another tfjs model, with a specified weight
# shard size.
sharded_model_path = os.path.join(self._tmp_dir, 'sharded_model')
with open(sharded_model_path, 'wt') as f:
# Create a fie at the path to elicit the error.
f.write('hello')
with self.assertRaisesRegexp( # pylint: disable=deprecated-method
ValueError, r'already exists as a file'):
converter.dispatch_tensorflowjs_to_tensorflowjs_conversion(
os.path.join(tfjs_output_dir, 'model.json'), sharded_model_path)
