def test_model_train_cnn(self):
input_config = self._create_test_input_config(
input_config_pb2.EncoderType.CNN)
model_config = self._create_test_model_config()
test_model = recommendation_model.RecommendationModel(
input_config=input_config, model_config=model_config)
batch_size = 4
input_context_movie_id = tf.keras.layers.Input(shape=(10, ),
dtype=tf.int32,
batch_size=batch_size,
name='context_movie_id')
input_context_movie_rating = tf.keras.layers.Input(
shape=(10, ),
dtype=tf.float32,
batch_size=batch_size,
name='context_movie_rating')
input_label_movie_id = tf.keras.layers.Input(shape=(1, ),
dtype=tf.int32,
batch_size=batch_size,
name='label_movie_id')
inputs = {
'context_movie_id': input_context_movie_id,
'context_movie_rating': input_context_movie_rating,
'label_movie_id': input_label_movie_id
}
logits = test_model(inputs)
self.assertAllEqual([batch_size, 20], logits.shape.as_list())
def build_keras_model(params):
"""Construct and compile recommendation keras model."""
model = recommendation_model.RecommendationModel(params)
model.compile(optimizer=_get_optimizer(
learning_rate=FLAGS.learning_rate,
gradient_clip_norm=FLAGS.gradient_clip_norm),
loss=losses.GlobalSoftmax(),
metrics=_get_metrics(params['eval_top_k']))
return model
def export(checkpoint_path, export_dir, params):
"""Export savedmodel."""
model = recommendation_model.RecommendationModel(params)
checkpoint = tf.train.Checkpoint(model=model)
checkpoint.restore(checkpoint_path).run_restore_ops()
signatures = {
tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
model.serve.get_concrete_function(
input_context=tf.TensorSpec(shape=[FLAGS.max_history_length],
dtype=tf.dtypes.int32,
name='context'))
}
tf.saved_model.save(model, export_dir=export_dir, signatures=signatures)
return export_dir
def test_model_serve(self):
config = {
"context_embedding_dim": 128,
"label_embedding_dim": 32,
"hidden_layer_dim_ratios": [1, 0.5, 0.25],
"item_vocab_size": 16,
"encoder_type": "bow",
"num_predictions": 10
}
test_model = recommendation_model.RecommendationModel(config)
input_context = tf.constant([1, 2, 3, 4, 5])
outputs = test_model.serve(input_context)
self.assertAllEqual([10], outputs["top_prediction_ids"].shape.as_list())
self.assertAllEqual([10], outputs["top_prediction_scores"].shape.as_list())
def test_model_train(self):
config = {
"context_embedding_dim": 128,
"label_embedding_dim": 32,
"hidden_layer_dim_ratios": [1, 0.5, 0.25],
"item_vocab_size": 16,
"encoder_type": "bow"
}
batch_size = 128
test_model = recommendation_model.RecommendationModel(config)
input_context = tf.keras.layers.Input(
shape=(None,), dtype=tf.int32, batch_size=batch_size, name="context")
input_label = tf.keras.layers.Input(
shape=(1,), dtype=tf.int32, batch_size=batch_size, name="label")
inputs = {"context": input_context, "label": input_label}
logits = test_model(inputs)
self.assertAllEqual([batch_size, config["item_vocab_size"] + 1],
logits.shape.as_list())
def build_keras_model(input_config: input_config_pb2.InputConfig,
model_config: model_config_class.ModelConfig):
"""Construct and compile recommendation keras model.
Construct recommendation model according to input config and model config.
Compile the model with optimizer, loss function and eval metrics.
Args:
input_config: The configuration object(input_config_pb2.InputConfig) that
holds parameters for model input feature processing.
model_config: A ModelConfig object that holds parameters to set up the
model architecture.
Returns:
The compiled keras model.
"""
model = recommendation_model.RecommendationModel(input_config=input_config,
model_config=model_config)
compile_model(model, model_config.eval_top_k, FLAGS.learning_rate,
FLAGS.gradient_clip_norm)
return model
def save_model(checkpoint_path: str, export_dir: str,
input_config: input_config_pb2.InputConfig,
model_config: model_config_class.ModelConfig):
"""Export to savedmodel.
Args:
checkpoint_path: The path to the checkpoint that the model will be exported
based on.
export_dir: The directory to export models to.
input_config: The input config of the model.
model_config: The configuration to set up the model.
"""
model = recommendation_model.RecommendationModel(input_config=input_config,
model_config=model_config)
checkpoint = tf.train.Checkpoint(model=model)
checkpoint.restore(checkpoint_path).run_restore_ops()
input_specs = input_pipeline.get_serving_input_specs(input_config)
signatures = {
tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
model.serve.get_concrete_function(**input_specs)
}
tf.saved_model.save(model, export_dir=export_dir, signatures=signatures)
def build_keras_model(params, learning_rate, gradient_clip_norm): """Construct and compile recommendation keras model.""" model = recommendation_model.RecommendationModel(params) compile_model(model, params, learning_rate, gradient_clip_norm) return model
