mhjabreel / charcnn Goto Github PK

hi,
Why you set the input layer size equal to 1014 and in your another code repository CharCnn_Keras, you
put a config.json file, is that a config file of caffee or torch? thanks.

boundary and values.

Hi,
When I run the training.py, I faced a problem like this
Traceback (most recent call last):
File "training.py", line 63, in
learning_rate = tf.train.piecewise_constant(global_step, boundaries, values)
File "/Users/chestersima/anaconda3/lib/python3.6/site-packages/tensorflow/python/training/learning_rate_decay.py", line 168, in piecewise_constant
"The length of boundaries should be 1 less than the length of values")

The source code is below:
boundaries = []
br = config.training.base_rate
values = [br]
for i in range(1, 10):
values.append(br / (2 ** i))
boundaries.append(15000 * i)
values.append(br / (2 ** (i + 1)))
print(values)
print(boundaries)
learning_rate = tf.train.piecewise_constant(global_step, boundaries, values)

I read the source code and find that the length of values is 11, however, the length of boundaries is 9. So how to work out this problem? Values has a initial item in config.py, and after the iteration, values append one more item. What's the point at doing so?

Hello,
I cloned the repo and ran the script. The model showed 20% accuracy after 16,000 steps. Do I just have to run the model longer ?

Hi,
I am trying to run CharCNN on 4 GeForce GTX 1080.I am struggling with 2 problems.

1. The 4 GPUs seem to have unbalanced load.When I ran nvidia-smi,the result is shown as following
   
   what can I do to make full use of all GPUs?
2. I have been running CharCNN for several days.But the accuracy remains around 0.3 without any advancement.I just ran python training.py without any change to your code.The current status is shown as following:
   
   Do I have to change some parameters or the type of optimizer? Have you ever tried to run CharCNN with better performance?

Thank you very much!

got errors that pointed to a different name for jaccard scoring module.

ImportError: cannot import name 'jaccard_similarity_score' from 'sklearn.metrics' (/opt/conda/lib/python3.7/site-packages/sklearn/metrics/__init__.py)

looks like the last version to use the format you use is 0.20.4
this worked for me and the program executed perfectly.

pip uninstall scikit-learn
Found existing installation: scikit-learn 0.23.2
Uninstalling scikit-learn-0.23.2:
  Would remove:
    /opt/conda/lib/python3.7/site-packages/scikit_learn-0.23.2.dist-info/*
    /opt/conda/lib/python3.7/site-packages/sklearn/*
Proceed (y/n)? y
  Successfully uninstalled scikit-learn-0.23.2
daveccampbell@tensorflow-115-vm:~/projects$ pip install scikit-learn==0.20.4

when i user CharCNN torch implement, chinese alphabate list is too large. How do you solve this problem in tensorflow implement?

If I want freeze the model to pb ，which node is output_node?

I have followed other issues and comments to improve the accuracy of model but its not going above 25% neither on default not my datasets of proteins. I have tried Adam as well as Momentum optimizer with large epochs.

seq_len set 50 is error??why?

My system slows down everytime I run the code, when I looked it online, I found an issue regarding memory leak in Tensorflow. tensorflow/tensorflow#10408 , The least I could understand was that should you not use sess.close() and tf.reset_default_graph() in training.py after the training loop at the end.

hello, I am interested in your repository, but can you provide some pre-trained models for AG, Sougou, DBP and other datasets mentioned in your paper.

Thank you very much!

Is this the same data that is used in Zhang et al. or did you create your own subsample of 4 classes from the AG corpus?

Hi,

I've tried to reproduce Zhang et al.'s result using your code (downloaded and run as is) but obtain very low accuracy numbers (25%) not at all near the paper's reported numbers.
Have you managed to reproduce their numbers?

do the csvs happen to be available anywhere - public dataset somewhere?
https://github.com/mhjabreel/CharCNN/tree/master/data/ag_news_csv

they are mentioned here:
https://github.com/johnb30/py_crepe

i could reassemble them given the text files in /data, but would hope to not introduce some oddity by mishandling double-quotes or something like that.

Hi ,

Change in training.py char_cnn.drop_prop to char_cnn.drop_keep_prob and
Change in training.py CharCNN class attribute "full_layers" to "fully_layers" .