Skip to content

Latest commit

 

History

106 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DL_Final_Project

Deep Learning (NYU) Final Project

Updated Plan (4/16/24 12:30PM)
    - connect SimVP with Slot attention
    - Define autoregressive script to utilize data best
    - generate masks on unlabeled data with Slot attention, use SimVP to predict mask directly

Successes
        -SimVP is trained on our image data -- loss is reasonable & frames are predicted on hidden data
Downfalls
        -SimVP does not train well on mask data -- need to figure this out to predict masks directly
        -Slot Attention training proved harder than anticipated -- it is up and running now but not on our data

Next steps
        -Train Slot Attention on our train set -- use it to predict masks on unlabeled data
        -Use slot attention to predict masks of predicted frames -- this is baseline solution for task (expected to be bad)
        -Work on training SimVP on mask data -- why is the loss so high? is the data imported correctly?
        -Look into other mask predictors if Slot Attention proves too finicky -- Mask RCNN is next choice


Current Plan (4/9/24 10:54AM):
    - Get skeleton code for RAFT (by 4/10)
    - Train RAFT on pairs of labeled data in train set (by 4/12)
        - Split train set into train_train & train_val
        - Treat each video in train_train as mini-batch of 21 pairs of images
        - Train optical flow on train_train set; validate on train_val set
    - Test trained RAFT on validation set (by 4/14)
        - With pre-trained RAFT, shift task to predict last frame given first 11
            - Use validation set as training set; train_val set as validation set (these examples have not been seen in training)
            - From the 11th frame, predict optical flow of frames 12-22; use last optical flow projection as prediction mask
            - Train on ground truth mask and prediction mask IoU

Dataset stored at /scratch/pdt9929/DL_Final_Project/dataset
    - Let me know if permissions aren't set right

About

Deep Learning (NYU) Final Project

Resources

Stars

0 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages