Merge remote-tracking branch 'upstream/master' into add/conv3d

Author: Jakub Kaczmarzyk

.npmignore

@@ -5,6 +5,7 @@
 .DS_Store
 .babelrc
 .idea/
+src/**/*_test.ts
 demo/
 docs/
 scripts/

README.md

@@ -171,25 +171,32 @@ the fetch() method.
 
 Currently TensorFlow.js only supports a limited set of TensorFlow Ops. See the
 [full list](./docs/supported_ops.md).
-If your model uses an unsupported ops, the `tensorflowjs_converter` script will fail and
+If your model uses unsupported ops, the `tensorflowjs_converter` script will fail and
 produce a list of the unsupported ops in your model. Please file issues to let us
 know what ops you need support with.
 
-## Loading the weights only
+## Manual forward pass and direct weights loading
 
-If you prefer to load the weights only, you can use follow code snippet.
+If you want to manually write the forward pass with the ops API, you can load the weights directly as a map from weight names to tensors:
 
-```typescript
+```js
 import * as tf from '@tensorflow/tfjs';
 
 const modelUrl = "https://example.org/model/model.json";
 
-const model = await fetch(modelUrl);
-this.weightManifest = (await model.json())['weightsManifest'];
+const response = await fetch(modelUrl);
+this.weightManifest = (await response.json())['weightsManifest'];
 const weightMap = await tf.io.loadWeights(
         this.weightManifest, "https://example.org/model");
 ```
 
+`weightMap` maps a weight name to a tensor. You can use it to manually implement the forward pass of the model:
+
+```js
+const input = tf.tensor(...);
+tf.matMul(weightMap['fc1/weights'], input).add(weightMap['fc1/bias']);
+```
+
 ## FAQ
 
 1. What TensorFlow models does the converter currently support?
@@ -221,6 +228,20 @@ tensorflowjs_converter \
 
 The time of first call also includes the compilation time of WebGL shader programs for the model. After the first call the shader programs are cached, which makes the subsequent calls much faster. You can warm up the cache by calling the predict method with an all zero inputs, right after the completion of the model loading.
 
+6. I have a model converted with a previous version of TensorFlow.js converter (0.15.x), that is in .pb format. How do I convert it to the new JSON format?
+
+You can use the built-in migration tool to convert the models generated by previous versions. Here are the steps:
+
+```
+git clone git@github.com:tensorflow/tfjs-converter.git
+cd tfjs-converter
+yarn
+yarn ts-node tools/pb2json_converter.ts pb_model_directory/ json_model_directory/
+```
+
+`pb_model_directory` is the directory where the model generated by previous version is located.
+`json_model_directory` is the destination directory for the converted model.
+
 ## Development
 
 To build **TensorFlow.js converter** from source, we need to clone the project and prepare
@@ -234,7 +255,7 @@ $ yarn # Installs dependencies.
 
 We recommend using [Visual Studio Code](https://code.visualstudio.com/) for
 development. Make sure to install
-[TSLint VSCode extension](https://marketplace.visualstudio.com/items?itemName=eg2.tslint)
+[TSLint VSCode extension](https://marketplace.visualstudio.com/items?itemName=ms-vscode.vscode-typescript-tslint-plugin)
 and the npm [clang-format](https://github.com/angular/clang-format) `1.2.2` or later
 with the
 [Clang-Format VSCode extension](https://marketplace.visualstudio.com/items?itemName=xaver.clang-format)

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tensorflow/tfjs-converter",
-  "version": "1.0.0",
+  "version": "1.0.2",
   "description": "Tensorflow model converter for javascript",
   "main": "dist/src/index.js",
   "jsnext:main": "dist/tf-converter.esm.js",
@@ -14,10 +14,10 @@
   },
   "license": "Apache-2.0",
   "peerDependencies": {
-    "@tensorflow/tfjs-core": "1.0.0"
+    "@tensorflow/tfjs-core": "1.0.2"
   },
   "devDependencies": {
-    "@tensorflow/tfjs-core": "1.0.0",
+    "@tensorflow/tfjs-core": "1.0.2",
     "@types/jasmine": "~2.8.6",
     "@types/long": "~3.0.32",
     "@types/node-fetch": "1.6.9",

python/tensorflowjs/converters/BUILD

@@ -53,6 +53,7 @@ py_test(
     ],
 )
 
+py_test(
     name = "tf_saved_model_conversion_v2_test",
     srcs = ["tf_saved_model_conversion_v2_test.py"],
     srcs_version = "PY2AND3",
@@ -68,10 +69,11 @@ py_library(
     data = ["//tensorflowjs:op_list_jsons"],
     srcs_version = "PY2AND3",
     deps = [
-        ":tf_saved_model_conversion",
         "//tensorflowjs:expect_numpy_installed",
         "//tensorflowjs:expect_tensorflow_installed",
+        "//tensorflowjs:version",
         "//tensorflowjs:write_weights",
+        "//tensorflowjs/converters:common",
     ],
 )
 

python/tensorflowjs/converters/keras_h5_conversion.py

@@ -101,6 +101,7 @@ def _check_version(h5file):
     raise ValueError(
         'Expected Keras version 2; got Keras version %s' % keras_version)
 
+
 def _initialize_output_dictionary(h5file):
   """Prepopulate required fields for all data foramts.
 
@@ -127,6 +128,36 @@ def _ensure_json_dict(item):
   return item if isinstance(item, dict) else json.loads(as_text(item))
 
 
+# https://github.com/tensorflow/tfjs/issues/1255, b/124791387
+# In tensorflow version 1.13 and some alpha and nightly-preview versions,
+# the following layers have different class names in their serialization.
+# This issue should be fixed in later releases. But we include the logic
+# to translate them anyway, for users who use those versions of tensorflow.
+_CLASS_NAME_MAP = {
+    'BatchNormalizationV1': 'BatchNormalization',
+    'UnifiedGRU': 'GRU',
+    'UnifiedLSTM': 'LSTM'
+}
+
+
+def translate_class_names(input_object):
+  """Perform class name replacement.
+
+  Beware that this method modifies the input object in-place.
+  """
+  if not isinstance(input_object, dict):
+    return
+  for key in input_object:
+    value = input_object[key]
+    if key == 'class_name' and value in _CLASS_NAME_MAP:
+      input_object[key] = _CLASS_NAME_MAP[value]
+    elif isinstance(value, dict):
+      translate_class_names(value)
+    elif isinstance(value, (tuple, list)):
+      for item in value:
+        translate_class_names(item)
+
+
 def h5_merged_saved_model_to_tfjs_format(h5file, split_by_layer=False):
   """Load topology & weight values from HDF5 file and convert.
 
@@ -156,6 +187,7 @@ def h5_merged_saved_model_to_tfjs_format(h5file, split_by_layer=False):
 
   model_json['model_config'] = _ensure_json_dict(
       h5file.attrs['model_config'])
+  translate_class_names(model_json['model_config'])
   if 'training_config' in h5file.attrs:
     model_json['training_config'] = _ensure_json_dict(
         h5file.attrs['training_config'])

python/tensorflowjs/converters/keras_h5_conversion_test.py

@@ -459,6 +459,54 @@ def testSavedModelRaisesErrorIfArtifactsDirExistsAsAFile(self):
         ValueError, r'already exists as a file'):
       conversion.save_keras_model(model, artifacts_dir)
 
+  def testTranslateBatchNormalizationV1ClassName(self):
+    # The config JSON of a model consisting of a "BatchNormalizationV1" layer.
+    # pylint: disable=line-too-long
+    json_object = json.loads(
+        '{"class_name": "Sequential", "keras_version": "2.2.4-tf", "config": {"layers": [{"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "GlorotUniform", "config": {"dtype": "float32", "seed": null}}, "name": "dense", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "dtype": "float32", "activation": "relu", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {"dtype": "float32"}}, "units": 10, "batch_input_shape": [null, 3], "use_bias": true, "activity_regularizer": null}}, {"class_name": "BatchNormalizationV1", "config": {"beta_constraint": null, "gamma_initializer": {"class_name": "Ones", "config": {"dtype": "float32"}}, "moving_mean_initializer": {"class_name": "Zeros", "config": {"dtype": "float32"}}, "name": "batch_normalization_v1", "dtype": "float32", "trainable": true, "moving_variance_initializer": {"class_name": "Ones", "config": {"dtype": "float32"}}, "beta_initializer": {"class_name": "Zeros", "config": {"dtype": "float32"}}, "scale": true, "axis": [1], "epsilon": 0.001, "gamma_constraint": null, "gamma_regularizer": null, "beta_regularizer": null, "momentum": 0.99, "center": true}}, {"class_name": "Dense", "config": {"kernel_initializer": {"class_name": "GlorotUniform", "config": {"dtype": "float32", "seed": null}}, "name": "dense_1", "kernel_constraint": null, "bias_regularizer": null, "bias_constraint": null, "dtype": "float32", "activation": "linear", "trainable": true, "kernel_regularizer": null, "bias_initializer": {"class_name": "Zeros", "config": {"dtype": "float32"}}, "units": 1, "use_bias": true, "activity_regularizer": null}}], "name": "sequential"}, "backend": "tensorflow"}')
+    # pylint: enable=line-too-long
+    conversion.translate_class_names(json_object)
+    # Some class names should not have been changed be translate_class_names().
+    self.assertEqual(json_object['class_name'], 'Sequential')
+    self.assertEqual(json_object['keras_version'], '2.2.4-tf')
+    self.assertEqual(json_object['config']['layers'][0]['class_name'], 'Dense')
+    # The translation should have happend:
+    #     BatchNormalizationV1 --> BatchNormalization.
+    self.assertEqual(
+        json_object['config']['layers'][1]['class_name'], 'BatchNormalization')
+    self.assertEqual(json_object['config']['layers'][2]['class_name'], 'Dense')
+
+    # Assert that converted JSON can be reconstituted as a model object.
+    model = keras.models.model_from_json(json.dumps(json_object))
+    self.assertTrue(isinstance(model, keras.Sequential))
+    self.assertEqual(model.input_shape, (None, 3))
+    self.assertEqual(model.output_shape, (None, 1))
+    self.assertEqual(model.layers[0].units, 10)
+    self.assertEqual(model.layers[2].units, 1)
+
+  def testTranslateUnifiedGRUAndLSTMClassName(self):
+    # The config JSON of a model consisting of a "UnifiedGRU" layer
+    # and a "UnifiedLSTM" layer.
+    # pylint: disable=line-too-long
+    json_object = json.loads(
+        '{"class_name": "Sequential", "keras_version": "2.2.4-tf", "config": {"layers": [{"class_name": "UnifiedGRU", "config": {"recurrent_activation": "sigmoid", "dtype": "float32", "trainable": true, "recurrent_initializer": {"class_name": "Orthogonal", "config": {"seed": null, "gain": 1.0}}, "use_bias": true, "bias_regularizer": null, "return_state": false, "unroll": false, "activation": "tanh", "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 10, "batch_input_shape": [null, 4, 3], "activity_regularizer": null, "recurrent_dropout": 0.0, "kernel_initializer": {"class_name": "GlorotUniform", "config": {"seed": null}}, "kernel_constraint": null, "time_major": false, "dropout": 0.0, "stateful": false, "reset_after": true, "recurrent_regularizer": null, "name": "unified_gru", "bias_constraint": null, "go_backwards": false, "implementation": 1, "kernel_regularizer": null, "return_sequences": true, "recurrent_constraint": null}}, {"class_name": "UnifiedLSTM", "config": {"recurrent_activation": "sigmoid", "dtype": "float32", "trainable": true, "recurrent_initializer": {"class_name": "Orthogonal", "config": {"seed": null, "gain": 1.0}}, "use_bias": true, "bias_regularizer": null, "return_state": false, "unroll": false, "activation": "tanh", "bias_initializer": {"class_name": "Zeros", "config": {}}, "units": 2, "unit_forget_bias": true, "activity_regularizer": null, "recurrent_dropout": 0.0, "kernel_initializer": {"class_name": "GlorotUniform", "config": {"seed": null}}, "kernel_constraint": null, "time_major": false, "dropout": 0.0, "stateful": false, "recurrent_regularizer": null, "name": "unified_lstm", "bias_constraint": null, "go_backwards": false, "implementation": 1, "kernel_regularizer": null, "return_sequences": false, "recurrent_constraint": null}}], "name": "sequential"}, "backend": "tensorflow"}')
+    # pylint: enable=line-too-long
+    conversion.translate_class_names(json_object)
+    # Some class names should not have been changed be translate_class_names().
+    self.assertEqual(json_object['class_name'], 'Sequential')
+    self.assertEqual(json_object['keras_version'], '2.2.4-tf')
+    # The translation should have happend:
+    #     UnifiedGRU --> GRU.
+    #     UnifiedLSTM --> LSTM.
+    self.assertEqual(json_object['config']['layers'][0]['class_name'], 'GRU')
+    self.assertEqual(json_object['config']['layers'][1]['class_name'], 'LSTM')
+
+    # Assert that converted JSON can be reconstituted as a model object.
+    model = keras.models.model_from_json(json.dumps(json_object))
+    self.assertTrue(isinstance(model, keras.Sequential))
+    self.assertEqual(model.input_shape, (None, 4, 3))
+    self.assertEqual(model.output_shape, (None, 2))
+
 
 if __name__ == '__main__':
   unittest.main()

python/tensorflowjs/version.py

@@ -1,4 +1,4 @@
 # @license See the LICENSE file.
 
 # This code is auto-generated, do not modify this file!
-version = '1.0.0'
+version = '1.0.2'

src/version.ts

@@ -1,5 +1,5 @@
 /** @license See the LICENSE file. */
 
 // This code is auto-generated, do not modify this file!
-const version = '1.0.0';
+const version = '1.0.2';
 export {version};