Analysis of Deep Image Colourization Techniques

Authors

1. Harsh Rao Dhanyamraju @HarshRaoD
2. Ang Boon Leng @jimmysqqr
3. Kshitij Parashar @xitij27

Inference

Setup

1. Please download the model checkpoints here
2. Create a new virtual environment
3. pip install -r requirements.txt

Running Tests

1. Make sure you have Trained_Colourization_Models.py and Model_Testing.ipynb in the same directory
2. You dont need to download the Coco Datset for inference you can use the Images in the Sample_Images Directory
3. Start a Jupyter Server and Model_Testing.ipynb
4. Create a Custom Dataset:

import Trained_Colourization_Models as tcm
test_dataset = tcm.CustomDataset(<Your-Path-here>,'test')

5. Load the testing Image

ti2 = tcm.Testing_Image(test_dataset, filename=<Your-file-name-here>)
# file name should be in the same directory as specified in test_dataset

6. Load a Model Runner and generate output by passing the Testing_Image Object

model_runner = tcm.Default_Model_Runner()
output_img = model_runner.get_image_output(ti2)

7. Visualise the output

plt.imshow(output_img)  # For model outputs
plt.imshow(ti2.get_gray(), cmap='gray')  # For Input Images
plt.imshow(ti2.get_rgb())  # For ground truth Images

Training

1. Create a new virtual environment
2. pip install -r requirements.txt
3. Navigate to the directory with the training script cd Training_Scripts/<experiment_dir>
4. Create a new directory /Models to store the model checkpoints created during training
5. Change the necessary configurations in the Configuration class and the data paths in the code
   1. set load_model_to_train = False
   2. set data path in CustomDataset
6. Run the file to begin training python training_script.py

Resume Training

1. Navigate to the directory with the training script cd Training_Scripts/<experiment_dir>
2. Change the necessary configurations in the Configuration class and the data paths in the code
   1. load_model_file_name: path to checkpoint file
   2. set load_model_to_train = True
   3. set starting_epoch = (current checkpoint epoch + 1)
   4. set data path in CustomDataset
3. Run the file to resume training python training_script.py

Downloading COCO Dataset

1. You can run DataAnalysis.ipynb to download the COCO Dataset and reduce the size of all images (to get training data)
2. You'll need a valid kaggle api key and atleast 55 GB of free space