/CS_74020_final_project

Primary LanguageJupyter Notebook

CS_74020_final_project

This repository contains the code for the final project for Graduate Center CS 74020 Fall 2021 Final Project

Authors:

  • Weicong Feng
  • Tugce Ozdemir
  • Gulya Radjapova
  • Nancy Sea

Instructions:

  1. Clone the repository
  2. Copy the contents of the final_project to Google Drive (to utilize the Google Collab environment) to the root folder of the Google Drive. All of the project documents should exist in final_project folder in the root of your drive.
  3. Download the contents of this folder and put them in the final_project folder in your drive. These files are too large to be hosted on Github but they contain the initial weights, and the experiment results in Experimentation folder.
  4. Open MatterportTrainMammals_final.ipynb in Google Collab
  5. Follow instructions in the notebook and run each portion.
  6. In the Load MaskRCNN and dependencies section, ensure that the ROOT_DIR variable set tp '/content/drive/MyDrive/final_project/' represents the location where you copied the repository.
  7. Continue running the notebook.
  8. In the Training section, ensure that the location of the original initialization via COCO is '/content/drive/MyDrive/final_project/mask_rcnn_coco.h5' in your drive.
  9. model.train() function allows you to train the model based on the configurations of interest. The resulting experiment results will be saved in the temp_log folder. Note that training on large amount of epochs and large size of training dataset will take quite a bit of time to train. To test that training is working, set epochs in model.train() to a small number.
  10. Copy the contents of the temp_log folder to the experiment folde in Google driver. For example:
  • !cp /content/temp_log/mammal20211219T1808/mask_rcnn_mammal_0010.h5 "/content/drive/MyDrive/final_project/Experimentation/12192021/"
  • !cp /content/temp_log/mammal20211219T1808/events.out.tfevents.1639937351.999d6dda7416 "/content/drive/MyDrive/final_project/Experimentation/12192021/" The timestamps associated with the model training will be different when you run this. Change mammal20211219T1808 to relevant folder.
  1. To replicate the loss analysis in the Visualizing & Understanding Losses section, ensure that the location of the experimentation folder is in /content/drive/MyDrive/final_project/Experimentation/. This folder will contain all the results of experiments we have conducted. We use tensor board to visualize losses.