Skip to content

Latest commit

 

History

History
89 lines (77 loc) · 4.71 KB

log.md

File metadata and controls

89 lines (77 loc) · 4.71 KB
Raw

Log

  • 22-Aug:

    • Explored some repositories and did some background study on hand writing extraction.

  • 23-Aug:

    • Progress of "chars in boxes" can be seen below:
    • 23-08--1.jpg
      • Here position of vertical line is choosen randomly from gaussian distribution having mean as midpoint of char_spacing.
    • 23-08--2.jpg
      • Here position of vertical line is choosen as midpoint of randomised char_spacing from gaussian distribution.
    • Trying to code such that the boxes are of fixed width.
    • 23-08--3.jpg
      • Here width of boxes are fixed and assumed characters is written in center deterministically
    • 23-08--4.jpg
      • Here width of boxes are fixed and assumed characters is written at center with error parameterised by gaussian distribution.

  • 24-Aug:

    • Written tfrecord_helper.py to manage operations of images abd labels generated in this code.
    • Had a issue with eager_execution as bytes_list can't be converted to string in TFRecord. So written an generator-wrapper to solve this issue.
    • Carefully, handpicked 108 handwritten fonts to induce good amound of variance in the dataset. The variations include position, thickness etc.
    • Added apply_all_sides in Line.py to get borders on all sides. Included Line dropouts and padding.
    • Added 30 more fonts and removed some

  • 25-Aug:

    • Generated sample dataset using 130+ hand-fonts.
    • Tested some sample with pre-trained models on IAM Dataset, the accuracy is very low. Almost, the ouput seems random.
    • Removed Light fonts.
    • Adjusted data to the required format and trained the samples on https://github.com/githubharald/SimpleHTR repository.
    • It consists of 5 CNN layers, 2 RNN (LSTM) layers and the CTC loss and decoding layer.

  • 26-Aug:

    • Written shell scripts for converting the annotation formats.
    • Encountered few errors with Tf 1.15, those issues were addressed with some tweaks in tensorflow git issues. It doesn't feel lasting and good implemention in tf v1.
    • I then used another git repo (ps://github.com/FLming/CRNN.tf2) modified it accordingly.
    • Trained CRNN model with tf2 with 80 epochs consisting of 95-5 split of ~50k images.
    • The complete tensorboard logs are uploaded in git.
    • After 30-40 epochs the model saturated
    • From the logs, it is evident that LSTMs didn't learn much. Hence, It didn't rely much on context,this might be due to randomly generated characters in words.
    • CNNs,BNs learnt well. The catch here is, it is more of a Computer Vision problem rather than NLP centric.
    • Choosen width of input image from analysing the distribution plot of training set. Width distplot
    • To handle outliers of width and height added zero padding when required while resizing to retain aspect ratio of image
    • Trained Public dataset on pre-trained weights (trained ~50k images)
    • The model is overfitting after val_loss (CTC loss) reaching ~6 and val_accuracy (60-65%) on public dataset Loss
      • CTC loss Loss
      • Word_Accuracy
    • Currently tuning hyper parameters for better results.

  • 27-Aug:

    • Modifyied model architecture after analysing the tensorboard.
    • Generated another ~50k synthetic images to train the model.
    • Tuning hyper parameters.
    • Implemented flask app on heroku with dummy_detect function.

  • 28-Aug:

    • Tuned Hyper Params.
    • Re-modified custom architecture, results weren't favourable.
    • Due to memory constraints in heroku (500 Mb), Compressed the model from hdf5 to savedmodel
    • Had issue with Functional Models while conversion in tf2.2, so upgraded to tf2.3 (as of Colab)
    • Though the compressed slug size was 476 Mb which eventually blew up to ~800Mb while running dyno in heroku cloud.
    • Ran heroku app locally successfully.

  • 29-Aug:

    • Successfully built model on CRNN with deployment.
    • Started implemeting Attention mechanism.
    • Qualitatively gave a good result on pre-trained model.
    • Quantative analysis on metrics are being done.

  • 30-Aug: