def forward_request(self, model_info, inputs, outputs=None, options=None):
translate_fn = model_info
tokens, lengths = utils.pad_lists(inputs, padding_value='')
outputs = translate_fn(tokens=tf.constant(tokens, dtype=tf.string),
length=tf.constant(lengths, dtype=tf.int32))
batch_predictions = outputs['tokens'].numpy()
batch_lengths = outputs['length'].numpy()
batch_log_probs = outputs['log_probs'].numpy()
batch_outputs = []
for predictions, lengths, log_probs in zip(batch_predictions,
batch_lengths,
batch_log_probs):
outputs = []
for prediction, length, log_prob in zip(predictions, lengths,
log_probs):
prediction = prediction[:length].tolist()
prediction = [token.decode('utf-8') for token in prediction]
score = float(log_prob)
outputs.append(
serving.TranslationOutput(prediction, score=score))
batch_outputs.append(outputs)
return batch_outputs
def forward_request(self, batch_inputs, info, timeout=None):
channel = grpc.insecure_channel("localhost:%s" % info['port'])
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
max_length = max(len(src) for src in batch_inputs)
tokens, lengths = utils.pad_lists(batch_inputs,
padding_value='',
max_length=max_length)
batch_size = len(lengths)
predict_request = predict_pb2.PredictRequest()
predict_request.model_spec.name = info['model_name']
predict_request.inputs['tokens'].CopyFrom(
tf.make_tensor_proto(tokens,
dtype=tf.string,
shape=(batch_size, max_length)))
predict_request.inputs['length'].CopyFrom(
tf.make_tensor_proto(lengths, dtype=tf.int32,
shape=(batch_size, )))
try:
future = stub.Predict.future(predict_request, timeout)
result = future.result()
except ExpirationError as e:
logger.error('%s', e)
return None
lengths = tf.make_ndarray(result.outputs['length'])
predictions = tf.make_ndarray(result.outputs['tokens'])
log_probs = tf.make_ndarray(result.outputs['log_probs'])
batch_outputs = []
for hypotheses, length, log_prob in zip(predictions, lengths,
log_probs):
outputs = []
for i, prediction in enumerate(hypotheses):
prediction_length = length[i] - 1 # Ignore </s>.
prediction = prediction[0:prediction_length].tolist()
prediction = [tf.compat.as_text(pred) for pred in prediction]
score = float(log_prob[i])
outputs.append(
serving.TranslationOutput(prediction, score=score))
batch_outputs.append(outputs)
return batch_outputs