pip install -r requirements.txt

Codebase developed on Python-2.7.15

We need to prepare the raw SVHN dataset

1. Go to the cloned multi_digit_classification_attention folder and run the following command:

cd SVHN
mkdir dataset

2. Download the SVHN dataset and extract the train and test SVHN data into the dataset folder inside multi_digit_classification_attention folder.

3. Select which type to data to curate and run the following command:

python gen_crop_dataset.py --dataset_type=<train/test>

4. Select which type to data to generate attention mask and run the following command:

python gen_attn_truth.py --dataset_type=<train/test>

1. To train detection model run the following command:

python train.py

2. To train classification model run the following command:

python train_classify_net.py

To test and visualize results run the following command:

jupyter notebook

and open and run:

> evaluate_and_viz.ipynb

1. To train full model for both detection and classification run the following command:

python train_end2end.py

2. To test and visualize results run the following command:

jupyter notebook

and open and run:

> evaluate_and_viz_end2end.ipynb

We need to generate raw CAPTCHA dataset

1. Go to the cloned multi_digit_classification_attention folder and run the following command:

cd other/
mkdir dataset

2. Generate dataset by running the following command:

python gen_captcha_dataset.py

3. Move the generated dataset into CAPTCHA folder.

1. To train detection model run the following command:

python train.py

2. To train classification model run the following command:

python train_classify_net.py

To test and visualize results run the following command:

jupyter notebook

and open and run:

> evaluate_and_viz.ipynb

1. To train full model for both detection and classification run the following command:

python train_end2end.py

2. To test and visualize results run the following command:

jupyter notebook

and open and run:

> evaluate_and_viz_end2end.ipynb

Any improvements would be appreciated, send a merge request if you would like to contribute.