import tensorflow as tf
import numpy as np
curr_dir = os.path.dirname(os.path.realpath(__file__))
from qiznlp.common.modules.common_layers import mask_nonpad_from_embedding, add_timing_signal_1d, label_smoothing, softmax_cross_entropy
from qiznlp.common.modules.embedding import embedding
from qiznlp.common.modules.encoder import transformer_encoder, mean_pool, multi_head_attention_base_pool
import qiznlp.common.utils as utils
conf = utils.dict2obj({
'vocab_size': 14180,
'label_size': 16,
"embed_size": 300,
'hidden_size': 300,
'num_heads': 6,
'num_encoder_layers': 6,
'dropout_rate': 0.2,
'lr': 1e-3,
'pretrain_emb': None,
# 'pretrain_emb': np.load(f'{curr_dir}/pretrain_word_emb300.npy'),
})
class Model(object):
def __init__(self, build_graph=True, **kwargs):
self.conf = conf
self.run_model = kwargs.get('run_model', None) # acquire outside run_model instance
if build_graph:
# build placeholder
self.build_placeholder()
# build model
sys.path.append(curr_dir + '/..') # 添加上级目录即默认qiznlp根目录
from model.mch_model import Model as MCH_Model
try:
import horovod.tensorflow as hvd
# 示例：horovodrun -np 2 -H localhost:2 python run_mch.py
# 可根据local_rank设置shard的数据，保证各个gpu采样的数据不重叠。
except:
HVD_ENABLE = False
else:
HVD_ENABLE = True
conf = utils.dict2obj({
'early_stop_patience': None,
'just_save_best': False,
'n_epochs': 20,
'data_type': 'tfrecord',
# 'data_type': 'pkldata',
})
class Run_Model_Mch(Run_Model_Base):
def __init__(self,
model_name,
tokenize=None,
pbmodel_dir=None,
use_hvd=False):
# 维护sess graph config saver
self.model_name = model_name
if tokenize is None:
self.jieba = jieba.Tokenizer()
from qiznlp.common.modules.common_layers import shape_list, mask_nonpad_from_embedding, add_timing_signal_1d, get_timing_signal_1d, shift_right, split_heads
from qiznlp.common.modules.embedding import embedding, proj_logits
from qiznlp.common.modules.encoder import transformer_encoder, transformer_decoder, EncDecAttention
from qiznlp.common.modules.birnn import Bi_RNN, RNN
from qiznlp.common.modules.beam_search import beam_search, greedy_search, get_state_shape_invariants
import qiznlp.common.utils as utils
conf = utils.dict2obj({
'max_turn': 5,
'vocab_size': 4000,
'embed_size': 300,
'hidden_size': 300,
'num_heads': 6,
'num_encoder_layers': 6,
'num_decoder_layers': 6,
'dropout_rate': 0.2,
'lr': 1e-3,
'pretrain_emb': None,
'beam_size': 5,
'max_decode_len': 50,
'eos_id': 2,
})
class Model(object):
def __init__(self, build_graph=True, **kwargs):
self.conf = conf
self.run_model = kwargs.get('run_model', None) # acquire outside run_model instance
if build_graph:
# build placeholder
import os
import tensorflow as tf
import numpy as np
curr_dir = os.path.dirname(os.path.realpath(__file__))
from qiznlp.common.modules.common_layers import mask_nonpad_from_embedding
from qiznlp.common.modules.embedding import embedding
from qiznlp.common.modules.birnn import Bi_RNN
import qiznlp.common.utils as utils
conf = utils.dict2obj({
'vocab_size': 1142,
'embed_size': 300,
'birnn_hidden_size': 300,
'l2_reg': 0.0001,
'dropout_rate': 0.2,
'lr': 1e-3,
'pretrain_emb': None,
})
class Model(object):
def __init__(self, build_graph=True, **kwargs):
self.conf = conf
self.run_model = kwargs.get('run_model', None) # acquire outside run_model instance
if build_graph:
# build placeholder
self.build_placeholder()
# build model
self.model_name = kwargs.get('model_name', 'esim')
from qiznlp.common.modules.common_layers import shape_list, mask_nonpad_from_embedding, add_timing_signal_1d, get_timing_signal_1d, shift_right, split_heads
from qiznlp.common.modules.embedding import embedding, proj_logits
from qiznlp.common.modules.encoder import transformer_encoder, transformer_decoder, EncDecAttention
from qiznlp.common.modules.birnn import Bi_RNN
from qiznlp.common.modules.beam_search import beam_search, greedy_search, get_state_shape_invariants
import qiznlp.common.utils as utils
conf = utils.dict2obj({
'vocab_size': 4000,
'embed_size': 300,
'hidden_size': 300,
'num_heads': 6,
'num_encoder_layers': 6,
'num_decoder_layers': 6,
'dropout_rate': 0.2,
'lr': 1e-3,
'pretrain_emb': None,
'beam_size': 40,
'max_decode_len': 50,
'eos_id': 2,
'gamma': 1, # 多样性鼓励因子
'num_group': 1, # 分组beam_search
'top_k': 30 # 详见beam_search
})
class Model(object):
def __init__(self, build_graph=True, **kwargs):
self.conf = conf
self.run_model = kwargs.get('run_model', None) # acquire outside run_model instance
if build_graph:
import tensorflow as tf
import numpy as np
curr_dir = os.path.dirname(os.path.realpath(__file__))
from qiznlp.common.modules.common_layers import mask_nonpad_from_embedding
from qiznlp.common.modules.embedding import embedding
from qiznlp.common.modules.birnn import Bi_RNN
from qiznlp.common.modules.idcnn import IDCNN
import qiznlp.common.utils as utils
conf = utils.dict2obj({
'vocab_size': 1847,
'embed_size': 300,
'label_size': 30,
'birnn_hidden_size': 300,
'dropout_rate': 0.2,
'lr': 1e-3,
'pretrain_emb': None,
# 'pretrain_emb': np.load(f'{curr_dir}/pretrain_word_emb300.npy'),
})
class Model(object):
def __init__(self, build_graph=True, **kwargs):
self.conf = conf
self.run_model = kwargs.get('run_model',
None) # acquire outside run_model instance
if build_graph:
# build placeholder
self.build_placeholder()
# build model