valdata = Dataset(val_df,
char_vocab,
user_vocab,
max_len=params.max_len,
batch_size=params.batch_size)
model = Model(params)
saver = tf.train.Saver(tf.global_variables())
config = tf.ConfigProto(inter_op_parallelism_threads=args.threads,
intra_op_parallelism_threads=args.threads)
session = tf.Session(config=config)
session.run(tf.global_variables_initializer())
avg_loss = MovingAvg(0.97) # exponential moving average of the training loss
for idx in range(params.iters):
feed_dict = dataset.GetFeedDict(model)
feed_dict[model.dropout_keep_prob] = params.dropout
c, _ = session.run([model.avg_loss, model.train_op], feed_dict)
cc = avg_loss.Update(c)
if idx % 50 == 0 and idx > 0:
# test one batch from the validation set
val_c = session.run(model.avg_loss, valdata.GetFeedDict(model))
logging.info({'iter': idx, 'cost': cc, 'rawcost': c, 'valcost': val_c})
if idx % 2000 == 0: # save a model file every 2,000 minibatches
saver.save(session,
os.path.join(expdir, 'model.bin'),
write_meta_graph=True,
global_step=idx)
# gd = tf.graph_util.convert_variables_to_constants(sess, tf.get_default_graph().as_graph_def(), ['add'])
# with tf.gfile.GFile('./tmodel/model.pb', 'wb') as f:
type=int,
default=12,
help='how many threads to use in tensorflow')
args = parser.parse_args()
expdir = args.expdir
# 模型加载
metamodel = MetaModel(expdir)
model = metamodel.model
metamodel.MakeSessionAndRestore(args.threads)
# 数据加载
df = LoadData(args.data)
dataset = Dataset(df,
metamodel.char_vocab,
metamodel.user_vocab,
max_len=metamodel.params.max_len)
total_word_count = 0
total_log_prob = 0
print(len(dataset.df), dataset.batch_size) # 20999 24
for idx in range(0, int(len(dataset.df) / dataset.batch_size)):
feed_dict = dataset.GetFeedDict(model)
# 这里的session 是 获取的是 保存后的模型
c, words_in_batch = metamodel.session.run(
[model.avg_loss, model.words_in_batch], feed_dict)
# c是 total_loss, words_in_batch 一个batch里字数
total_word_count += words_in_batch # 整个字数
total_log_prob += float(c * words_in_batch)
print('整体损失值： {0}\t{1:.3f}'.format(
idx, np.exp(total_log_prob / total_word_count)))
