This model maps handwritten digit(mnist) to fine 'DIGITAL' digit. We create sythetic IMAGE label for each input image with time-new-roman.ttf font using PILLOW libary. Our model is a multi-task learner with main objective being to map input digit image to fine-grained image while classifiying the input acts as secondary objective.
To make it work with your workstation, follow the following steps:
1. Clone the repo.
2. Setup your enviroment with requirement.txt file.
Basic Info:
- sythetic_label.py was used to generate synthetic images
- data_preparation.py generates train, test image. Models keep part of it training data for validation
- For tensorboard visualization run "tensorboard --logdir="./data/logs" from base directory
- To view model architecture, have a look on ./data/model_architecture.png
With config.py, You can change the outcome of the main.py file as your liking. Change the following CONSTANT to change:
1. TRAINING: For training
2. TESTING : For Testing
3. DEMO : Demo with image in DEMO_DIR
4. EVALUATE_PERFORMANCE : To Evaluate Performance and generate evaluation metrices & Plot
5. VISULIZE_2ND_LAST_LAYER : To visualize 2nd last layer Activation
6. POSTPROCESS : To Remove noise(But image has staircase effect)
7. PRINT_SUMMARY : To print model summary-
8. TESTING_VIS: To visualize testing- [Default: 256; this number of arbitarily selected images will be saved in the OUTPUT_DIRS during testing]
[Change it to 0 to turn it off]
9. You can also save concatenated image of input & output image in utils.py function
Other parameters you can change:
Training epoch : TRAINING_EPOCH = 100
Pretrainded Directory : PRETRAINED_DIR = './data/saved_model/model.h5' [for Testing or Retrianing]
Learing Rate : LEARNING_RATE = .0001 [Learning rate of Nadam Optimizer]
Batch size : BATCH_SIZE = 128 [11GB VRAM]
Put your image here to visualize : DEMO_DIR = './data/demo_images/'
Demo output directory during testing : OUTPUT_DIRS = './data/output'
FONT_DIR = './data/times-new-roman.ttf' [Font of Digital Output]
FONT_SIZE = 30 [Font size of Digital Output]
INFERENCE_BATCH_SIZE = 256 [Double of Training Batch Size]
LOSS_WEIGHTS = {'classification':.1,'reconstruction':9} [Weight of Classification & Reconstruction Loss]
VAL_SPLIT = .8 [ 20% of data will be kept for validation]
LAYER_VISUALIZATION_DIR = 'second_last_layer_vis' [Layer Visualization Output]
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent