def main():
args = build_argparser().parse_args()
# load vocabulary file for model
vocab = load_vocab_file(args.vocab)
log.debug("Loaded vocab file from {}, get {} tokens".format(
args.vocab, len(vocab)))
# create tokenizer
tokenizer = Tokenizer(BPE(str(args.vocab), str(args.merges)))
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
tokenizer.decoder = decoders.ByteLevel()
log.info('OpenVINO Inference Engine')
log.info('\tbuild: {}'.format(get_version()))
ie = IECore()
# read IR
model_xml = args.model
model_bin = model_xml.with_suffix(".bin")
log.info('Reading model {}'.format(args.model))
ie_net = ie.read_network(model=model_xml, weights=model_bin)
# check input and output names
if len(ie_net.input_info) != 1:
raise RuntimeError(
'The demo expects model with single input, while provided {}'.
format(len(ie_net.input_info)))
if len(ie_net.outputs) != 1:
raise RuntimeError(
'The demo expects model with single output, while provided {}'.
format(len(ie_net.outputs)))
input_names = next(iter(ie_net.input_info))
output_names = next(iter(ie_net.outputs))
# load model to the device
ie_net_exec = ie.load_network(network=ie_net, device_name=args.device)
log.info('The model {} is loaded to {}'.format(args.model, args.device))
if args.input:
def prompts():
for prompt in args.input:
log.info("Input prompt: {}".format(prompt))
yield prompt
else:
def prompts():
while True:
yield input('Type input prompt (empty string to exit):')
# loop on user's or prepared prompts
for prompt in prompts():
if not prompt.strip():
break
# encode input
tokens = tokenizer.encode_batch([prompt])[0].ids
input_ids = np.array([tokens], dtype=np.int32)
# maximum number of tokens that can be processed by network at once
max_length = ie_net.input_info[input_names].input_data.shape[1]
eos_token_id = len(vocab) - 1
cur_input_len = input_ids.shape[-1]
# maximum number of tokens that will be generated
max_sample_token_num = args.max_sample_token_num + cur_input_len
t0 = time.perf_counter()
t_count = 0
while True:
# pad the rest of the request
pad_len = max_length - cur_input_len
model_input = np.concatenate(
(input_ids, [[eos_token_id] * pad_len]), axis=-1)
# create numpy inputs for IE
inputs = {
input_names: model_input,
}
# infer by IE
t_start = time.perf_counter()
res = ie_net_exec.infer(inputs=inputs)
t_end = time.perf_counter()
t_count += 1
log.info(
"Sequence of length {} is processed with {:0.2f} requests/sec ({:0.2} sec per request)"
.format(max_length, 1 / (t_end - t_start), t_end - t_start))
outputs = res[output_names]
next_token_logits = outputs[:, cur_input_len - 1, :]
# pre-process distribution
next_token_scores = process_logits(input_ids, next_token_logits,
eos_token_id)
if args.top_k > 0:
next_token_scores = get_top_k_logits(next_token_scores,
args.top_k)
if args.top_p < 1.0:
next_token_scores = get_top_p_logits(next_token_scores,
args.top_p)
# get next token id
probs = softmax(next_token_scores)
next_tokens = np.random.choice(probs.shape[-1],
1,
p=probs[0],
replace=True)
# update info for the next step
input_ids = np.concatenate((input_ids, [next_tokens]), axis=-1)
cur_input_len = input_ids.shape[-1]
if stop_criteria(input_ids, min(max_length, max_sample_token_num),
eos_token_id):
break
t1 = time.perf_counter()
text = tokenizer.decode_batch(input_ids)[0]
log.info(
"{} requests of {} length were processed in {:0.2f}sec ({:0.2}sec per request)"
.format(t_count, max_length, t1 - t0, (t1 - t0) / t_count))
# print result
log.info("GENERATED SEQUENCE: {}".format(text))
def main():
args = build_argparser().parse_args()
# load vocabulary file for model
vocab = load_vocab_file(args.vocab)
log.debug("Loaded vocab file from {}, get {} tokens".format(
args.vocab, len(vocab)))
# create tokenizer
tokenizer = Tokenizer(BPE.from_file(str(args.vocab), str(args.merges)))
tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
tokenizer.decoder = decoders.ByteLevel()
log.info('OpenVINO Runtime')
log.info('\tbuild: {}'.format(get_version()))
core = Core()
# read IR
log.info('Reading model {}'.format(args.model))
model = core.read_model(args.model)
# check number inputs and outputs
if len(model.inputs) != 1:
raise RuntimeError(
'The demo expects model with single input, while provided {}'.
format(len(model.inputs)))
if len(model.outputs) != 1:
raise RuntimeError(
'The demo expects model with single output, while provided {}'.
format(len(model.outputs)))
input_tensor = model.inputs[0].any_name
if not args.dynamic_shape and (
model.inputs[0].partial_shape.is_dynamic
or model.inputs[0].shape[1] != args.max_seq_len):
model.reshape({
input_tensor:
PartialShape([Dimension(1),
Dimension(args.max_seq_len)])
})
if args.dynamic_shape:
model.reshape({
input_tensor:
PartialShape([Dimension(1),
Dimension(0, args.max_seq_len)])
})
# load model to the device
compiled_model = core.compile_model(model, args.device)
output_tensor = compiled_model.outputs[0]
infer_request = compiled_model.create_infer_request()
log.info('The model {} is loaded to {}'.format(args.model, args.device))
if args.input:
def prompts():
for prompt in args.input:
log.info("Input prompt: {}".format(prompt))
yield prompt
else:
def prompts():
while True:
yield input('Type input prompt (empty string to exit):')
# loop on user's or prepared prompts
for prompt in prompts():
if not prompt.strip():
break
# encode input
tokens = tokenizer.encode_batch([prompt])[0].ids
input_ids = np.array([tokens], dtype=np.int32)
# maximum number of tokens that can be processed by network at once
max_length = args.max_seq_len
eos_token_id = len(vocab) - 1
cur_input_len = input_ids.shape[-1]
# maximum number of tokens that will be generated
max_sample_token_num = args.max_sample_token_num + cur_input_len
t0 = time.perf_counter()
t_count = 0
while True:
model_input = input_ids
if not args.dynamic_shape:
# pad the rest of the request
pad_len = max_length - cur_input_len
model_input = np.concatenate(
(input_ids, [[eos_token_id] * pad_len]), axis=-1)
# create numpy inputs for OpenVINO runtime
inputs = {
input_tensor: model_input,
}
# infer by OpenVINO runtime
t_start = time.perf_counter()
outputs = infer_request.infer(inputs)[output_tensor]
t_end = time.perf_counter()
t_count += 1
log.info(
"Sequence of length {} is processed with {:0.2f} requests/sec ({:0.2} sec per request)"
.format(model_input.shape[1], 1 / (t_end - t_start),
t_end - t_start))
next_token_logits = outputs[:, cur_input_len - 1, :]
# pre-process distribution
next_token_scores = process_logits(input_ids, next_token_logits,
eos_token_id)
if args.top_k > 0:
next_token_scores = get_top_k_logits(next_token_scores,
args.top_k)
if args.top_p < 1.0:
next_token_scores = get_top_p_logits(next_token_scores,
args.top_p)
# get next token id
probs = softmax(next_token_scores)
next_tokens = np.random.choice(probs.shape[-1],
1,
p=probs[0],
replace=True)
# update info for the next step
input_ids = np.concatenate((input_ids, [next_tokens]), axis=-1)
cur_input_len = input_ids.shape[-1]
if stop_criteria(input_ids, min(max_length, max_sample_token_num),
eos_token_id):
break
t1 = time.perf_counter()
text = tokenizer.decode_batch(input_ids)[0]
log.info(
"{} requests were processed in {:0.2f}sec ({:0.2}sec per request)".
format(t_count, t1 - t0, (t1 - t0) / t_count))
# print result
log.info("GENERATED SEQUENCE: {}".format(text))