wwjwhen / watch-attend-and-parse-tensorflow-version Goto Github PK

Hi, I cannot clone offline_train.pkl and offline_test.pkl from your project. When I try to fetch them with git-lfs I get an error:
Error downloading object: data/offline-test.pkl (e936420): Smudge error: Error downloading
data/offline-test.pkl
(e9364201cdbce24d71b61d47c52c1c497df95af6e909fb6cca4d335ddc8b6318): batch response:
This repository is over its data quota. Purchase more data packs to restore access.

I've googled this problem but I can't solve it. I think it might because you don't have enough space on git-lfs and maybe you need to purchase more space.
Can you send me these two files if it's not too much trouble? Or can I use these two files from jianshuZhang 's project? I wonder if they are the same.
Thanks!

Hi mate,
I converted your code into jupyter notebook. I tried running it and after some updates in epoch 0, it gave error that 'anno' is not defined in the get_sample() function. Could you please help me with this issue?

the datasets is small ,but when the datasets is like100k,the model will gonna keep the high accuracy?

Hello, @wwjwhen ，I found a uninitialized variable “argmax”， I now set to True? Does it OK?

I use python 3.7 and TF 1.6
The code has 2 errors:

in model-single-GPU.py line 383, in
lambda: tf.nn.embedding_lookup(wap.embed_matrix, sample_y)
NameError: name 'wap' is not defined

--> I changed wap.embed_matrix to self.embed_matrix

in model-single-GPU.py line 465, in
input_dict = {
anno:ctx,
infer_y:next_w,
alpha_past:next_alpha_past,
h_pre:next_state,
if_trainning:training
}
NameError: name 'anno' is not defined

--> I added global definition for variables
global anno, infer_y, h_pre, alpha_past, if_trainning

Finally, I can run the code, but it is not coverage.
WER and cost in the validation set dose not reduce.
Do you have any suggestion?

Thank you alot.

Hello,

Query 1:
In your experiment a total of 181 epochs were run to train the model.
In our case it will take around 2 weeks to run 181 epochs.
Did you also spend so much time in training or can the speed be increased by using some GPU settings?

Query 2:
Due to certain constraints our model training gets interrupted every 6 hours.
Could you tell us a way to save the model after each epoch during training, so that if training
gets interrupted, we could reload the saved model and continue training from that point.