Rohan_Kapoor-IITB-Assignment-Jul-Dec2020-Batch2

Handwritten Form Reader Web App

0. my_run.sh

To predict for a folder containing images simply execute my_run.sh. This will print the predictions in the output file.

my_run.sh path/of/folder output/file/path/filename.txt

1. Create new environment

If using anaconda then create new environment using this in the conda prompt else you can use the navigator as well

conda create -n envname python=3.7

If not using anaconda then see this link https://packaging.python.org/guides/installing-using-pip-and-virtual-environments/

Note: Python version is 3.7. Now everything will be done in this environment.

To activate and use the environment

conda activate envname

2. Clone this repository

Either Install git-bash from https://git-scm.com/downloads or

pip install git

Clone this repository using

git clone repository_address

After that change location to this repository (Required for all operations)

cd Rohan_Kapoor_9599023170-IITB-Assignment-Jul-Dec2020-Batch2 

3. Install the required packages

pip install -r requirements.txt

4. Generating images

For generating images go to trdg folder in TextRecognitionDataGenerator and then run 'run.py' file

cd TextRecognitionDataGenerator/trdg
python run.py -c 100

This will generate 100 images in out folder in this directory. To check options do

python run.py -h

You can add more fonts in the fonts folder, more background images, texts, dictionaries etc according to the requirements
Go to the official documentation https://textrecognitiondatagenerator.readthedocs.io/en/latest/index.html for more details

5. Preparing the dataset in tfrecords format

First you need to prepare the annotations.txt file which is just a simple text file containing the locations of all the images in the dataset and their corresponding labels for eg

c:/Users/rkcha/TextRecognitionDataGenerator/trdg/out/11.jpg 7hjLcQ
c:/Users/rkcha/TextRecognitionDataGenerator/trdg/out/12.jpg Yx5vNVfg
c:/Users/rkcha/TextRecognitionDataGenerator/trdg/out/13.jpg DtbngV3Rs

Then use the following commands to prepare tfrecords

aocr dataset location/of/datasets/annotations-training.txt location/of/datasets/training.tfrecords
aocr dataset location/of/datasets/annotations-testing.txt location/of/datasets/testing.tfrecords

To check more options use

aocr dataset -h

6. For training the model

First copy the dataset in the tfrecords format in the dataset directory of the repository

Use the following command to train a fresh model and use the option --modcnn if you want to train using the modified cnn architecture

aocr train ./dataset/nameoffile.tfrecords
aocr train --modcnn ./dataset/nameoffile.tfrecords

For checking the options available use

aocr train -h

Here you can change number of epochs, batch size, image size etc.

7. Testing the model

Copy the data in the dataset folder of the repository
To test modified model use

aocr test --model-dir ./checkpoint_mod_model --modcnn ./dataset/nameoffile.tfrecords

To test original model use

aocr test --model-dir ./checkpoint_orig_model ./dataset/nameoffile.tfrecords

To test any other model (use --modcnn if trained using modified cnn)

aocr test path/of/checkpoints ./dataset/filename.tfrecords
aocr test --modcnn path/of/checkpoints ./dataset/filename.tfrecords
 

8. To use the web app

python app.py

and go to link http://127.0.0.1:5000/ then give the url of the image