def _create_rpc_call(self, sess, model_file, **kwargs):
model_base_dir = os.path.split(model_file)[0]
pid = model_file.split("-")[-1]
lengths = peek_lengths_key(model_file, self.feature_exporter_field_map)
pc = PreProcessorController(model_base_dir, pid, self.task.config_params['features'],
self.feature_exporter_field_map, lengths)
tf_example, preprocessed = pc.run()
# Create a dict of embedding names to sub-graph outputs to wire in as embedding inputs
embedding_inputs = {}
for feature in preprocessed:
embedding_inputs[feature] = preprocessed[feature]
model, classes, values = self._create_model(sess, model_file, **embedding_inputs)
sig_input = {x: tf.saved_model.utils.build_tensor_info(tf_example[x]) for x in pc.FIELD_NAMES}
if model.lengths is not None:
sig_input.update({model.lengths_key: tf.saved_model.utils.build_tensor_info(model.lengths)})
sig_output = SignatureOutput(classes, values)
sig_name = 'predict_text'
assets = create_assets(
model_file,
sig_input, sig_output, sig_name,
model.lengths_key,
return_labels=self.return_labels,
preproc=self.preproc_type(),
)
return sig_input, sig_output, sig_name, assets
def _create_rpc_call(self, sess, model_file, **kwargs):
model_base_dir = os.path.split(model_file)[0]
pid = model_file.split("-")[-1]
lengths = peek_lengths_key(model_file, self.feature_exporter_field_map)
pc = PreProcessorController(model_base_dir, pid,
self.task.config_params['features'],
self.feature_exporter_field_map, lengths)
tf_example, preprocessed = pc.run()
# Create a dict of embedding names to sub-graph outputs to wire in as embedding inputs
embedding_inputs = {}
for feature in preprocessed:
embedding_inputs[feature] = preprocessed[feature]
model, embedded = self._create_model(sess, model_file,
**embedding_inputs)
sig_input = {
x: tf.saved_model.utils.build_tensor_info(tf_example[x])
for x in pc.FIELD_NAMES
}
classes = tf.ones([1, 1], dtype=tf.uint8)
sig_output = SignatureOutput(classes, embedded)
sig_name = 'embed_text'
assets = create_assets(model_file,
sig_input,
sig_output,
sig_name,
None,
return_labels=False)
return sig_input, sig_output, sig_name, assets
def _create_rpc_call(self, sess, basename, **kwargs):
model, embedded = self._create_model(sess, basename)
predict_tensors = {}
for k, v in model.embeddings.items():
try:
predict_tensors[k] = tf.saved_model.utils.build_tensor_info(v.x)
except:
raise Exception('Unknown attribute in signature: {}'.format(v))
sig_input = predict_tensors
classes = tf.ones([1, 1], dtype=tf.uint8)
sig_output = SignatureOutput(classes, embedded)
sig_name = 'embed_text'
assets = create_assets(basename, sig_input, sig_output, sig_name, None, return_labels=False)
return sig_input, sig_output, sig_name, assets
def _create_rpc_call(self, sess, model_file, **kwargs):
model_base_dir = os.path.split(model_file)[0]
pid = model_file.split("-")[-1]
pc = PreProcessorController(model_base_dir, pid, self.task.config_params['features'],
self.feature_exporter_field_map)
tf_example, preprocessed = pc.run()
# Create a dict of embedding names to sub-graph outputs to wire in as embedding inputs
embedding_inputs = {}
for feature in preprocessed:
embedding_inputs[feature] = preprocessed[feature]
model, classes, values = self._create_model(sess, model_file, **embedding_inputs)
sig_input = {x: tf.saved_model.utils.build_tensor_info(tf_example[x]) for x in pc.FIELD_NAMES}
sig_input.update({model.lengths_key: tf.saved_model.utils.build_tensor_info(model.lengths)})
sig_output = SignatureOutput(classes, values)
sig_name = 'tag_text'
assets = create_assets(model_file, sig_input, sig_output, sig_name, model.lengths_key,
return_labels=self.return_labels)
return sig_input, sig_output, sig_name, assets
def _create_rpc_call(self, sess, model_file):
model_base_dir = os.path.split(model_file)[0]
pid = model_file.split("-")[-1]
pc = PreProcessorController(model_base_dir, pid,
self.task.config_params['features'])
model_params = self.task.config_params['model']
tf_example, preprocessed = pc.run()
for feature in preprocessed:
model_params[feature] = preprocessed[feature]
model, classes, values = self._create_model(sess, model_file)
sig_input = {
'tokens':
tf.saved_model.utils.build_tensor_info(tf_example[pc.FIELD_NAME]),
model.lengths_key:
tf.saved_model.utils.build_tensor_info(model.lengths)
}
sig_output = SignatureOutput(classes, values)
sig_name = 'tag_text'
assets = create_assets(model_file, sig_input, sig_output, sig_name,
model.lengths_key)
return sig_input, sig_output, sig_name, assets