A modification of Matterport's Mask RCNN for Nucleus along with mAP
• The dataset = download nuclei_datasets.tar.gz from kaggle which needs to be copied to the root directory
• main.py file takes an argument based on the requirement
• python3 main.py train trains the model from scratch using the resnet50 weights for the FCN layers. The model is stored in the logs file. • python3 main.py test tests the model on the test folder images, and stores the output along with a submit file in the results folder inside the root directory. If you do not want to train the model yourself, use the weights I have saved after 40 epochs in the logs folder.
• python3 main.py mAP-val plots the Precision Recall curve for the 25 validation images from the training set manually taken by nucleus.VAL_IMAGE_IDS
• python3 main.py mAP-train plots the PR curve for the raining data. But if you want to check the PR curve for a partial amount (Takes less time to plot and less processing capacity), input a lower number. If not, input the number of files in training data.