comsaint / dlh_project Goto Github PK

Restructure the train model and validate model code, to modulize and for speed up.

Add Global Config and notebook for preprocessing data

The current dataset used is CheXpert. Need to migrate it to NIH Chest Xray, including the data access part.

Previously in @AsadBinImtiaz 's branch, the Dataset object generates multiple labels (e.g. [1, 0, 0, 0, 1, 1, ..., 0]) for each image, which is good for multi-label classification. Unfortunately, I overwrote the work due to time constraint but implemented setting for binary classification instead. Now it is a good time to implement the multi-label setting.

Tasks:

1. Edit data_processing.py and dataset.py so that the data loader produces a label vector (instead of 1 label).
2. Change the loss function to binary cross-entropy loss.
3. Modify the training and evaluation functions in train_model.py to adapt for new loss and metric (Macro-average ROC?). Also, pay attention to the last classification layer, making sure it is not using Softmax.
4. (Question) how to handle class weight and stratify strategy?
5. (Question) how to implement early stopping?

Notes:

1. Do not confuse multi-class with multi-label classification. link
2. Good reference on loss function. link

There are some utilities which are useful for the heatmap:

1. A plotting function which overlays a bounding box on top of an image:
   

Input: image, x/y coordinates of the bounding box.

2. Evaluation metric which compares a given bounding box with ground truth. Research is required.

Part of the model's architecture leverages CapsNet in conjunction with CNN. Possible code: here

Training on custom VM is too slow and expensive. Can we migrate our code to use cloud AI service?
E.g. Colab example, GCP AI Platform

Add e.g. TensorBoard to visualization training process.

Based on the starter code, replace the simple model with state-of-the-art ones e.g. DenseNet, Inception etc., to evaluate both the speed and accuracy.

1. train-valsplit the training data
2. use the val split as validation inside training loop
3. rename the current val_data_loader to test_data_loader.

ref: code example

Concatinate capsnet with other models and get improved results

Reference: paper1 paper2