def load_model(self):
data = common.middle_load(os.path.join(self.model_path, "corpus"))
self.dict = data["word2idx"]
self.turn_size = data["turn_size"]
self.max_context_len = data["max_context_len"]
tf.reset_default_graph()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
saver = tf.train.import_meta_graph(
os.path.join(self.model_path, "model.meta"))
saver.restore(sess, tf.train.latest_checkpoint(self.model_path))
graph = tf.get_default_graph()
self.src_max_len = graph.get_tensor_by_name(
"init_variables/src_max_len:0")
self.batch_size = graph.get_tensor_by_name(
"init_variables/batch_size:0")
self.tgt_max_len = graph.get_tensor_by_name(
"init_variables/tgt_max_len:0")
self.src = graph.get_tensor_by_name("init_variables/src_tensor:0")
self.postion = graph.get_tensor_by_name("init_variables/src_postion:0")
self.turns = graph.get_tensor_by_name("init_variables/turns_tensor:0")
self.dropout_rate = graph.get_tensor_by_name(
"init_variables/dropout_keep_prob:0")
self.enc_output = graph.get_tensor_by_name("enc_output:0")
self.tgt = graph.get_tensor_by_name("init_variables/tgt_tensor:0")
self.pre_enc_output = graph.get_tensor_by_name(
"init_variables/pre_enc_output:0")
self.distributes = graph.get_tensor_by_name("pre_distributes:0")
self.session = sess
parser.add_argument('--dist_rate', type=float, default=0.4)
parser.add_argument('--dist_encode_rate', type=float, default=0.5)
parser.add_argument('--not_use_pretrain', action='store_true')
parser.add_argument('--use_debug', action='store_true')
args = parser.parse_args()
random.seed(args.seed)
np.random.seed(args.seed)
tf.compat.v1.set_random_seed(args.seed)
args.turn_size = 6
os.environ["CUDA_VISIBLE_DEVICES"] = args.cuda_device
data = middle_load(args.data)
args.vocab_size = len(data["word2idx"])
args.max_context_len = data["max_len"]
args.use_pretrain = not args.not_use_pretrain
training_data = DataLoader(data["train"]["src_texts"],
data["train"]["src_turn"],
data["train"]["tgt_indexs"],
data["train"]["tgt_texts"],
data["train"]["eos_indexs"],
data["train"]["src_context"],
data["train"]["tgt_context"],
batch_size=args.batch_size)
validation_data = DataLoader(data["valid"]["src_texts"],
self.tgt_texts[_start:_start + _bsz])
tgt_indexs_tensor = index_pairs(self.tgt_indexs[_start:_start + _bsz],
tgt_max_len)
turns_tensor = self.src_turn[_start:_start + _bsz]
eos_indexs = self.eos_indexs[_start:_start + _bsz]
src_context = self.src_context[_start:_start + _bsz]
return (src_tensor, src_postion, turns_tensor), (
tgt_tensor, tgt_postion
), tgt_indexs_tensor, src_max_len, eos_indexs, tgt_max_len, src_context
if __name__ == "__main__":
import common
corpus = common.middle_load("data/corpus")
dl = DataLoader(corpus["train"]["src_texts"], corpus["train"]["src_turn"],
corpus["train"]["tgt_indexs"],
corpus["train"]["tgt_texts"],
corpus["train"]["eos_indexs"],
corpus["train"]["src_context"], 4)
(src_tensor, src_postion, turns_tensor), (
tgt_tensor, tgt_postion
), tgt_indexs_tensor, src_max_len, eos_indexs, tgt_max_len, src_context = next(
dl)
print(tgt_max_len)
print(tgt_indexs_tensor.tolist())
ei = eos_indexs[0]
tgt_ei = tgt_indexs_tensor[0] == 0
tgt_max_len)
turns_tensor, _, _ = pad_to_longest(self.src_turn[_start:_start +
_bsz])
eos_indexs = self.eos_indexs[_start:_start + _bsz]
src_context = self.src_context[_start:_start + _bsz]
tgt_context = self.tgt_context[_start:_start + _bsz]
return (src_tensor, src_postion, turns_tensor), (
tgt_tensor, tgt_postion
), tgt_indexs_tensor, src_max_len, eos_indexs, tgt_max_len, src_context, tgt_context
if __name__ == "__main__":
from common import middle_load, middle_save, set_logger
data = middle_load("data/corpus")
training_data = DataLoader(data["train"]["src_texts"],
data["train"]["src_turn"],
data["train"]["tgt_indexs"],
data["train"]["tgt_texts"],
data["train"]["eos_indexs"],
data["train"]["src_context"],
data["train"]["tgt_context"],
batch_size=8)
(src_tensor, src_postion, turns_tensor), (
tgt_tensor, tgt_postion
), tgt_indexs_tensor, src_max_len, eos_indexs, tgt_max_len, src_context, tgt_context = next(
training_data)
def load_model(self):
data = common.middle_load(os.path.join(self.model_path, "corpus"))
self.dict = data["word2idx"]
self.turn_size = data["turn_size"]
self.max_context_len = data["max_context_len"]
tvars, checkpoint_name)
tf.train.init_from_checkpoint(checkpoint_name, assignment_map)
tmp_g = tf.get_default_graph().as_graph_def()
config = tf.ConfigProto(allow_soft_placement=True)
with tf.Session(config=config) as sess:
sess.run(tf.global_variables_initializer())
tmp_g = tf.graph_util.convert_variables_to_constants(
sess, tmp_g, output_tensors)
tmp_g = optimize_for_inference(
tmp_g, input_tensors, output_tensors, input_datatype, False)
with tf.gfile.GFile(f"{graph_output_dir}/graph", 'wb') as f:
f.write(tmp_g.SerializeToString())
if __name__ == "__main__":
from common import middle_load
from model import Transformer
class CorpusArgs(object):
def __init__(self, _args):
for k, v in _args.items():
self.__setattr__(k, v)
data = middle_load("weights/corpus")
cargs = CorpusArgs(data["args"])
graph_optimize(cargs, Transformer, "weights", "weights/model", ["enc_output", "pre_distributes"], [tf.int32.as_datatype_enum, tf.int32.as_datatype_enum, tf.int32.as_datatype_enum, tf.int64.as_datatype_enum, tf.int64.as_datatype_enum, tf.int64.as_datatype_enum, tf.float32.as_datatype_enum, tf.int64.as_datatype_enum, tf.int64.as_datatype_enum, tf.float32.as_datatype_enum], [
"init_variables/src_max_len", "init_variables/batch_size", "init_variables/tgt_max_len", "init_variables/src_tensor", "init_variables/src_postion", "init_variables/turns_tensor", "init_variables/dropout_keep_prob", "init_variables/tgt_tensor", "init_variables/tgt_postion", "init_variables/pre_enc_output"])