def predict_static(args, batch_generator):
test_prog = fluid.Program()
with fluid.program_guard(test_prog):
paddle.manual_seed(SEED)
paddle.framework.random._manual_program_seed(SEED)
# define input and reader
input_field_names = util.encoder_data_input_fields + util.fast_decoder_data_input_fields
input_descs = util.get_input_descs(args, 'test')
input_slots = [{
"name": name,
"shape": input_descs[name][0],
"dtype": input_descs[name][1]
} for name in input_field_names]
input_field = util.InputField(input_slots)
feed_list = input_field.feed_list
loader = fluid.io.DataLoader.from_generator(
feed_list=feed_list, capacity=10)
# define model
transformer = Transformer(
args.src_vocab_size, args.trg_vocab_size, args.max_length + 1,
args.n_layer, args.n_head, args.d_key, args.d_value, args.d_model,
args.d_inner_hid, args.prepostprocess_dropout,
args.attention_dropout, args.relu_dropout, args.preprocess_cmd,
args.postprocess_cmd, args.weight_sharing, args.bos_idx,
args.eos_idx)
out_ids, out_scores = transformer.beam_search(
*feed_list,
bos_id=args.bos_idx,
eos_id=args.eos_idx,
beam_size=args.beam_size,
max_len=args.max_out_len)
# This is used here to set dropout to the test mode.
test_prog = test_prog.clone(for_test=True)
# define the executor and program for training
exe = fluid.Executor(place)
util.load(test_prog,
os.path.join(args.save_static_model_path, "transformer"), exe)
loader.set_batch_generator(batch_generator, places=place)
step_idx = 0
speed_list = []
for feed_dict in loader:
seq_ids, seq_scores = exe.run(
test_prog,
feed=feed_dict,
fetch_list=[out_ids.name, out_scores.name],
return_numpy=True)
if step_idx % args.print_step == 0:
if step_idx == 0:
logging.info(
"Static Predict: step_idx: %d, 1st seq_id: %d, 1st seq_score: %.2f,"
% (step_idx, seq_ids[0][0][0], seq_scores[0][0]))
avg_batch_time = time.time()
else:
speed = args.print_step / (time.time() - avg_batch_time)
speed_list.append(speed)
logging.info(
"Static Predict: step_idx: %d, 1st seq_id: %d, 1st seq_score: %.2f, speed: %.3f steps/s"
% (step_idx, seq_ids[0][0][0], seq_scores[0][0], speed))
avg_batch_time = time.time()
step_idx += 1
if step_idx == STEP_NUM:
break
logging.info("Static Predict: avg_speed: %.4f steps/s" %
(np.mean(speed_list)))
return seq_ids, seq_scores
def predict_dygraph(args, batch_generator):
with fluid.dygraph.guard(place):
paddle.manual_seed(SEED)
paddle.framework.random._manual_program_seed(SEED)
# define data loader
test_loader = fluid.io.DataLoader.from_generator(capacity=10)
test_loader.set_batch_generator(batch_generator, places=place)
# define model
transformer = Transformer(
args.src_vocab_size, args.trg_vocab_size, args.max_length + 1,
args.n_layer, args.n_head, args.d_key, args.d_value, args.d_model,
args.d_inner_hid, args.prepostprocess_dropout,
args.attention_dropout, args.relu_dropout, args.preprocess_cmd,
args.postprocess_cmd, args.weight_sharing, args.bos_idx,
args.eos_idx)
# load the trained model
model_dict, _ = util.load_dygraph(
os.path.join(args.save_dygraph_model_path, "transformer"))
# to avoid a longer length than training, reset the size of position
# encoding to max_length
model_dict["encoder.pos_encoder.weight"] = position_encoding_init(
args.max_length + 1, args.d_model)
model_dict["decoder.pos_encoder.weight"] = position_encoding_init(
args.max_length + 1, args.d_model)
transformer.load_dict(model_dict)
# set evaluate mode
transformer.eval()
step_idx = 0
speed_list = []
for input_data in test_loader():
(src_word, src_pos, src_slf_attn_bias, trg_word,
trg_src_attn_bias) = input_data
seq_ids, seq_scores = transformer.beam_search(
src_word,
src_pos,
src_slf_attn_bias,
trg_word,
trg_src_attn_bias,
bos_id=args.bos_idx,
eos_id=args.eos_idx,
beam_size=args.beam_size,
max_len=args.max_out_len)
seq_ids = seq_ids.numpy()
seq_scores = seq_scores.numpy()
if step_idx % args.print_step == 0:
if step_idx == 0:
logging.info(
"Dygraph Predict: step_idx: %d, 1st seq_id: %d, 1st seq_score: %.2f"
% (step_idx, seq_ids[0][0][0], seq_scores[0][0]))
avg_batch_time = time.time()
else:
speed = args.print_step / (time.time() - avg_batch_time)
speed_list.append(speed)
logging.info(
"Dygraph Predict: step_idx: %d, 1st seq_id: %d, 1st seq_score: %.2f, speed: %.3f steps/s"
% (step_idx, seq_ids[0][0][0], seq_scores[0][0], speed))
avg_batch_time = time.time()
step_idx += 1
if step_idx == STEP_NUM:
break
logging.info("Dygraph Predict: avg_speed: %.4f steps/s" %
(np.mean(speed_list)))
return seq_ids, seq_scores