Project Blog
CS298
Week 16 - Dec/01/2020
MoMs:
Week 15 - Nov/24/2020
MoMs:
- Discussed final results of the project, formalities for defense and report
TODO:
Week 14 - Nov/17/2020
MoMs:
- Implemented Pix2Pix GAN and trained to generate MRIs of the faces.
- The process of extracting MRIs of faces using the trained Pix2Pix GAN is started.
TODO:
- Start working on report
- Train then model using MRIs as features.
Week 13 - Nov/10/2020
MoMs:
- Implemented label smoothing, more runtime augmentation for regularization.
- Implemented code to generate ROC, more code fixes.
- Discussed xray based model designed.
TODO:
Week 12 - Nov/03/2020
MoMs:
- Trained the model for longer duration, the accuracy was about 65%
TODO:
- Explore ways to optimize and discussed about more ways to implement the model.
Week 11 - Oct/27/2020
MoMs:
- Implemented fake and real loss as compared to simple Binary cross entropy loss.
- Implemented AMP optimization to train in FP16 mode.
- Implemented frame-level classification. Each frames are passed individually to a model for classification. Efficient Net B0 is used for extracting features.
TODO:
- Train the model for longer duration.
- Write algorithm to aggregate results of frame based quantification to classify the full video.
Week 10 - Oct/20/2020
MoMs:
- Implemented optical flow based model. Accuracy was about 63%
TODO:
- Explore ways for optimization.
Week 9 - Oct/13/2020
MoMs:
- Implemented the new CNN model using efficient net B0 as feature extractor. The features as passes to a custom classifier, accuracy was about 60%
- Also implemented a custom CNN-based model and passed features of five frames as a single feature-map. Accuracy was about 60%
- Discussed about using optical flow as features and showed simple demo of how the optical flow looks for images.
TODO:
- Implement optical flow based model
Week 8 - Oct/06/2020
- No meeting due to personal emergency.
Week 7 - Sept/29/2020
MoMs:
- Implemented RNN based model, accuracy was about 53%
- Implemented the simple CNN based model. This model uses one frame from each video for classification, accuracy was about 55%
TODO:
- Implement another CNN based model and pass features of five frames as single feature-map
Week 6 - Sept/29/2020
MoMs:
- Implemented RNN based model, accuracy was about 53%
TODO:
- Implement another CNN based model and pass features of five frames as single feature-map
Week 5 - Sept/22/2020
MoMs:
- Json file for each video is created to locate faces in the video
- Faces are cropped from each frame of the video and saved on disk.
TODO:
- Implement a model. The model will extract CNN features from each frame and pass them RNN to find hidden representation of the faces in the video. Pass final hidden state to MLP for classification of DeepFake videos
- Start working on training and validation
Week 4 - Sept/15/2020
MoMs:
- Implemented Noise (gaussian, speckle, s&p, pepper, salt, poisson, localvar), blur, rotation, horizontal flip, rescale, brightness, and contrast methods of data augmentation
- Implemented static, rolling, and spontaneous methods of data distractions for random text and shapes (circle and rectangle). e.g. static-text method shows a random text in the video at random location. Rolling-shape method shows a random shape floating at either of the four methods. The four methods are left-to-right, right-to-left, top-to-bottom, bottom-to-top. The spontaneous-shape would show random shapes of random colors (red, green, blue, white, black) or random sizes are random frames
- Show demo of videos and a sample picture which shows all data augmentation and distraction applied
TODO:
- Keep running the process as it takes too long to apply on the dataset.
- Start working on training loops with a basic model
Week 3 - Sept/08/2020
MoMs:
- Show Work-in-progress (WIP) code for augmentation
- Implemented Adding various types of noise to videos
- Implemented code to detect face in video and draw rectangle around face in the video
TODO:
- Continue working on data augmentation and distraction
Week 2 - Sept/1/2020
MoMs:
- Gave presentation on DFDC dataset
TODO:
- Work on data augmentation
Week 1 - Aug/25/2020
MoMs:
- First meeting
- Review CS298 proposal
TODO:
- Update proposal with feedback
CS297
Week 16 - May/11/2020
MoMs:
Week 15 - May/04/2020
MoMs:
- Went over the report and received comments
TODO:
- Work on final report to fix the comments
Week 14 - Apr/28/2020
MoMs:
- Update on running deliverable #4 model for longer duration
- Presentation on paper [5]
TODO:
Week 13 - Apr/21/2020
MoMs:
TODO:
- Read paper [5] and make presentation
- Keep running model for deliverable #4
Week 12 - Apr/14/2020
MoMs:
TODO:
- Continue working on deliverable #4
Week 11 - Apr/07/2020
MoMs:
- Presented overview of paper [4]
TODO:
Week 10 - Mar/31/2020
Spring break
Week 9 - Mar/24/2020
MoMs:
- Presented overview of paper [3]
TODO:
- Make presentation on paper [4]
Week 8 - Mar/17/2020
MoMs:
TODO:
- Read papers [3] and [4]
- Make presentation on paper [3]
Week 7 - Mar/10/2020
MoMs:
- Presented overview of GANs
TODO:
Week 6 - Mar/03/2020
MoMs:
- Show demo of deliverable #2. (Basic AE (using plain ANN and CNN) and VAE)
TODO:
- Start on understanding GANs
- Make presentation on GANs
Week 5 - Feb/25/2020
MoMs:
- Presented Autoencoders (AEs)
- Show WIP demo of Basic AE
TODO:
- Work on deliverable #2
- Take samples from latent space to generate new characters.
Week 4 - Feb/18/2020
MoMs:
- Presented PyTorch Overview
- Demo of deliverable# 1
TODO:
- Make a presentation on Autoencoders
- Work on deliverable #2 and show WIP model of autoencoder
Week 3 - Feb/11/2020
MoMs:
- Presented introduction of Sanskit (Devnagari script)
- Show Work-in-progress (WIP) model of deliverable 1.
TODO:
- Make a presentation on PyTorch
- Complete deliverable #1.
Week 2 - Feb/4/2020
MoMs:
- Presented CS297 proposal and received feedback
- a/c setup and general discussion on direction of the project
TODO:
- Update proposal
- Presentation on Sanskrit
- PyTorch based CNN to recognize Sanskrit letters
Week 1 - Jan/28/2020
MoMs:
- First meeting
- Decided timings, Other formalities
TODO:
|