Is it possbile to performe ocr task using RNNSharp ? #1

BackT0TheFuture · 2015-07-31T15:47:25Z

Hi there,

recently I wanna do some tests about sequence labelling(OCR without segment) via RNN

I googled and found this project , Thanks for your efforts

I have hundreds of handwritten word images and the corresponding word

Would you like to give me some instruction about this problem

any advice will be welcomeed. thanks in advance !

Best regards,

zhongkaifu · 2015-08-05T01:01:26Z

Yes. RNN has good performance on OCR task, such as hand writing recognition. Generally, you need to segment handwritten word images at first, and then recognize each word in the image.

The feature selection is the key important part. You can design and combine your feature manually, and RNN can also help you to generate features automatically, such as embedding input pixels into vectors.

With reasonable feature set, you can choose a classifier to detect which word is in the image.

BackT0TheFuture · 2015-08-05T05:17:04Z

I'm a newbie on Machine Learning but very interested in it.
do you mean the segmentation is needed before ocr using RNN ?
I'm not good at c#, is there some document for RnnSharp ?
would you like to make a little ocr task demo using RNNSHARP when you are free?
Thanks !

Best regards,

Signed-off-by: Zhongkai Fu <[email protected]>

#2. Fix BPTT initialization bug

#2. Speed up training performance

#2. Update: Train can be ended early if current PPL is larger than the previous one Signed-off-by: Zhongkai Fu <[email protected]>

#2. Using error token ratio to verify validated set performance

#2. Add dropout for LSTM

…running validation #2. Support model vector quantization reduce model size to 1/4 original #3. Refactoring code and speed up training #4. Fixing feature extracting bug

#2. Improve BiRNN learning process #3. Support to train model without validated corpus

#2. Optimize LSTM encoding to improve performance significantly #3. Apply dynamtic learning rate

#2. Improve encoding performance by SIMD instructions

…ce by SIMD instructions" This reverts commit 1a3070c.

#2. Execute CRF forward-backward in parallel

#2. Update readme file

#2. Normalize LSTM cell value in weights updating

…m input layer. #2. Refactoring dropout layer and output layer #3. Refactoring layer initialization

#2. Fix bug in softmax output layer when computing hidden layer value #3. Refactoring code

#2. Refactor configuration file and command line parameter #3. use SIMD for backward pass in output layer

#2. Refactoring code

…o encoder is used. #2. For seq2seq autoencoder, concatenate first top hidden layer and last top hidden layer as final encoder output for decoder.

… is worse than LSTM #2. Fix backward bug in Dropout layer #3. Refactoring code

…hidden layer is more than 1 #2. Improve training part of bi-directional RNN. We don't re-run forward before updating weights #3. Fix bugs in Dropout layer #4. Change hidden layer settings in configuration file. #5. Refactoring code

#2. Add incremental training

#2. Refactoring code #3. Make RNNDecoder thread-safe

#2. Delete useless code

#2. Code refactoring #3. Performance improvement

#2. Improve training performnce ~ 300% up #3. Fix learning rate update bug #4. Apply SIMD instruction to update error in layers #5. Code refactoring

zhongkaifu added a commit that referenced this issue Nov 14, 2015

#1. Refactoring code

7dc851e

Signed-off-by: Zhongkai Fu <[email protected]>

zhongkaifu added a commit that referenced this issue Dec 2, 2015

#1. Improve SimpleRNN training performance

5b0de99

#2. Fix BPTT initialization bug

zhongkaifu added a commit that referenced this issue Dec 2, 2015

#1. Remove beta regularization

a4dcfcf

zhongkaifu added a commit that referenced this issue Dec 3, 2015

#1. Support dropout for SimpleRNN (both forward and bi-directional)

debffdb

#2. Speed up training performance

zhongkaifu added a commit that referenced this issue Dec 7, 2015

#1. Bug fix: Failed to load bi-directional model from file in test model

c18b865

#2. Update: Train can be ended early if current PPL is larger than the previous one Signed-off-by: Zhongkai Fu <[email protected]>

zhongkaifu added a commit that referenced this issue Dec 24, 2015

#1. Improve logger

0838a69

#2. Using error token ratio to verify validated set performance

zhongkaifu added a commit that referenced this issue Jan 9, 2016

#1. Bug fix: output layer is not cleaned before calculating new values

1d2b3be

#2. Add dropout for LSTM

zhongkaifu added a commit that referenced this issue Feb 1, 2016

#1. Refactoring code and improve BiRNN training performance

a50c8ed

zhongkaifu mentioned this issue Feb 12, 2016

Poor performance in Sequential Tagging #4

Closed

zhongkaifu added a commit that referenced this issue Feb 15, 2016

#1. Fix RNNSharp crashing bug when dense feature isn't used

1dc9759

#2. Improve BiRNN learning process #3. Support to train model without validated corpus

zhongkaifu added a commit that referenced this issue Feb 24, 2016

#1. Support SIMD instruction to speed up encoding and decoding

e390b85

#2. Optimize LSTM encoding to improve performance significantly #3. Apply dynamtic learning rate

zhongkaifu added a commit that referenced this issue Feb 25, 2016

#1. Fix Forward-LSTM crash bug

1a3070c

#2. Improve encoding performance by SIMD instructions

zhongkaifu added a commit that referenced this issue Feb 25, 2016

Revert "#1. Fix Forward-LSTM crash bug #2. Improve encoding performan…

513cb0c

…ce by SIMD instructions" This reverts commit 1a3070c.

zhongkaifu added a commit that referenced this issue Mar 9, 2016

#1. LSTM weights vectorization for SIMD instruction

8df8dd9

#2. Execute CRF forward-backward in parallel

zhongkaifu added a commit that referenced this issue Mar 9, 2016

#1. Code refactoring

e699d62

#2. Update readme file

zhongkaifu added a commit that referenced this issue Mar 9, 2016

#1. Code refactoring

2d58c7a

#2. Normalize LSTM cell value in weights updating

zhongkaifu added a commit that referenced this issue Mar 9, 2016

#1. Adding bias cell for LSTM

37a29ef

My-Khan mentioned this issue May 26, 2016

Model structure #22

Closed

zhongkaifu added a commit that referenced this issue Jul 8, 2016

#1. Provide sparse features for each layer. The sparse feature is fro…

b277d06

…m input layer. #2. Refactoring dropout layer and output layer #3. Refactoring layer initialization

zhongkaifu mentioned this issue Jul 15, 2016

Older version? #25

Closed

zhongkaifu added a commit that referenced this issue Nov 30, 2016

#1. Support sequence-to-sequence model

eb1cc30

#2. Fix bug in softmax output layer when computing hidden layer value #3. Refactoring code

zhongkaifu added a commit that referenced this issue Dec 22, 2016

#1. Convert numeric type from double to float

0d6259f

#2. Refactor configuration file and command line parameter #3. use SIMD for backward pass in output layer

zhongkaifu added a commit that referenced this issue Jan 6, 2017

#1. Improve LSTM layer performance ~ 25% by SIMD instruction.

a0d3aa6

#2. Refactoring code

zhongkaifu added a commit that referenced this issue Feb 5, 2017

#1. Retire BPTTLayer, since it's an old algorithm and its performance…

d80f402

… is worse than LSTM #2. Fix backward bug in Dropout layer #3. Refactoring code

zhongkaifu added a commit that referenced this issue Feb 19, 2017

#1.Fixed Dropout crashed bug in ForwardRNN

fe3bb58

#2. Add incremental training

zhongkaifu added a commit that referenced this issue Mar 8, 2017

#1. Improve encoder and decoder performance ~ 100% up

6778c27

#2. Refactoring code #3. Make RNNDecoder thread-safe

zhongkaifu added a commit that referenced this issue Mar 21, 2017

#1. Fix the bug that incremental training is crashed.

df4406b

#2. Delete useless code

zhongkaifu closed this as completed Mar 21, 2017

zhongkaifu added a commit that referenced this issue Apr 22, 2017

#1. Bugs fixed

4db5e03

#2. Code refactoring #3. Performance improvement

zhongkaifu added a commit that referenced this issue May 3, 2017

#1. Support mini batch for training

5043eb3

#2. Improve training performnce ~ 300% up #3. Fix learning rate update bug #4. Apply SIMD instruction to update error in layers #5. Code refactoring

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Is it possbile to performe ocr task using RNNSharp ? #1

Is it possbile to performe ocr task using RNNSharp ? #1

BackT0TheFuture commented Jul 31, 2015

zhongkaifu commented Aug 5, 2015

BackT0TheFuture commented Aug 5, 2015

Is it possbile to performe ocr task using RNNSharp ? #1

Is it possbile to performe ocr task using RNNSharp ? #1

Comments

BackT0TheFuture commented Jul 31, 2015

zhongkaifu commented Aug 5, 2015

BackT0TheFuture commented Aug 5, 2015