chqiwang / convseg Goto Github PK

Thanks.

请问那个train.py里面的TASK参数应该写什么呢？我看参数定义里面没有写默认值。谢谢！

When i run the code train.py

Problem:
TASK = import("args")
AttributeError: module 'args' has no attribute 'read_train_file'

precision?or recall ?or F?

Could you please show me how to segment sentence by the trained model? Thanks.

What does the max_length mean in
def data_iterator(inputs, batch_size, shuffle=True, max_length=200):

When I use word ebedding(word_window==4), I encountered this error ：

Traceback (most recent call last):
File "train.py", line 91, in
max_epoches=args.max_epoches)
File "/home/czy/Pycode/MasterPaper/code/convseg-master/tagger.py", line 440, in train
log_dir + '/dev', epoch)
File "/home/czy/Pycode/MasterPaper/code/convseg-master/cws.py", line 181, in evaluator
eval_lines = [l.rstrip() for l in codecs.open(temp_path, 'r', 'utf8')]
File "/home/czy/env/.virtualenvs/keras/lib/python2.7/codecs.py", line 896, in open
file = builtin.open(filename, mode, buffering)
IOError: [Errno 2] No such file or directory: 'model-wemb-msr/logs/dev/1.temp'

Can you tell me how to solve this problem? Thanks!

If can, how? Thanks.

Thanks for your code. But I have a question about the implementation af Batch Normalization.

if use_bn: w = layers.batch_norm(inputs=v, decay=0.9, is_training=is_train, center=True, scale=True, scope='BatchNorm_w_%d' % i) v = layers.batch_norm(inputs=w, decay=0.9, is_training=is_train, center=True, scale=True, scope='BatchNorm_v_%d' % i)
Where w is computed by the inputs=v and v is computed by the inputs=w. Is there a mistake? The correct code I think is
if use_bn: w = layers.batch_norm(inputs=w, decay=0.9, is_training=is_train, center=True, scale=True, scope='BatchNorm_w_%d' % i) v = layers.batch_norm(inputs=v, decay=0.9, is_training=is_train, center=True, scale=True, scope='BatchNorm_v_%d' % i)
Is it true?

Thanks for your brilliant job.

I am currently looking at building it on my own with pytorch.

Sorry that I am relatively new to NLP and thereby (maybe) immature on some issue. When the loss function looks like computationally expensive. Like, if I have a 28-char sentence and apply 4 tags as yours in the paper. I would definitely have 4^28 = 72057594037927936 results to take in. Is it really so expensive? Or there might be some form of optimization to avoid the daunting overhead?