chychen / air_writing Goto Github PK

When I'm trying to run the train_blstm.py, it is taking so much time to run
SparseTensor(indices=Tensor("decoder/CTCGreedyDecoder:0", shape=(?, 2), dtype=int64), values=Tensor("decoder/CTCGreedyDecoder:1", shape=(?,), dtype=int64), dense_shape=Tensor("decoder/CTCGreedyDecoder:2", shape=(2,), dtype=int64))
After this it doesn't show any progress.
Can you help me with this.
Thank you

Hey,
thanks for this amazing work.

i just have a small issue here at first there were some errors, i fixed them got the code to run and start training but it seems like it can't predict anything and the loss always high

Original val: then He had put his hands on her shoulders<b><b><b><b><b><b><b><b><b><b><b><b><b><b><b><b><b><b><b><b><b>

Decoded val:

1 epoches, 85 steps, mean loss: 617.738472, valid mean loss: 101.721208, time cost: 305.696634(sec/batch), levenshtein: 1.000000

this is the first epoch output so what is wrong here?

Thankyou.

Hi,
First of all, thank you very much for sharing your work.

Howerver, I read the origin paper which said some pre processing methods, such as skew correction, find baseline and the corpus line. Is there a function to complete it in your demo?

Thankyou

Line num 119
Shouldn't it be y_point[:] = [(point - y_min) * scale for point in y_point]
instead of x_point in the end

Hi,
First of all thanks for sharing your work.

There are some files missing which are UltraProcess.py and letters.txt.

I found that UltraProcess.py is nothing but the preprocess.py in src folder which creates data.npy and labels.npy but while executing read.py it was giving error that it couldn't locate letters.txt.

Appreciate your thoughts.