We can merge all the trained models to create a best vote classifier.

In other words, we can create a model that outputs predictions that are an average of all the trained models. This will make the output more robust. We can do that simply by loading the saved models and predicting using them.

• For better readability and understanding of our dataset and model.

• Raw data viz

• Count plots of different categories

• Variability, Eigenimages

• Contrast Between Average Images

• Architecture of our model

@tirthasheshpatel can you please see and add some more transformation other than provided in the model.py to make the model training process harder. It will be helpful for us to avoid overfitting. I already added some and got some good results. Remember initially we were obtaining 98 % trining accuracy in the 1st epoch. Please add more..
https://pytorch.org/docs/stable/torchvision/transforms.html

I have uploaded our predictions on hidden test set given by Rupal mam. So, don't make it public unless all evaluations are complete!!

Result file name does not include the optimizer name
It just says Adam parameters.
Need to change it.

If we truly want to see how our model is working, accuracy is not a measure. I think we need to add F1-score, precision, and recall.

This is keep a track of the progress:

• Finalize the model : Currently it is pretrained googlenet. Other options are pretrained resnet50, pretrained resnet100, or pretrained inceptionnet.
• Data Preprocesing : data.py does some data cleaning but we may need to preprocess it too (IFF WE HAVE TIME)
• Training the Model :
  • Checkpoint at 90% (test set) accuracy
  • Checkpoint at (90+2)% (test set) accuracy
  • Checkpoint at (90+5)% (test set) accuracy
  • Checkpoint at (90+7)% (test set) accuracy

Any other tasks we need to track??

https://github.com/Shrey-Viradiya/Corona_Detection/blob/34c0f0a344fa17f2870d0d84cf5eef7d0093453b/model.py#L33

I think this line is not useful as below it is resizing it again. @tirthasheshpatel

As we are implementing Transfer Learning in our project, we should also build a model from scratch that will help us a lot to understand some basic but important concepts. Let me know your views on it. Or let's have a meeting!

Separating model objects in their own branch will make it easier to pull the repository as these large objects will not be included in the pull. We can then pull the required object from its respective branch. How does it sound?

@tirthasheshpatel Can you please code for trying different optimizers from the command line argument and saving the features with name in the result file?

The GoogleNet training has slowed down by more than 40 seconds because we are saving the training and testing accuracies and losses array at each epoch. We sould save them after every 5 or 6 epochs to save time.

We need to keep track of all the models we have trained so far. So, let's push them onto different branches and create a pull request. We will later merge them all together. Here are the models and their pull request:

• GoogleNet (@Shrey-Viradiya) #30
• Inception (@Shrey-Viradiya) #29
• ResNet50 (@Shrey-Viradiya) #37
• ResNet101 (@tirthasheshpatel) #36
• ResNet152 (@Shrey-Viradiya) #31
• VGG19 (@Shrey-Viradiya) #32

@tirth-hihoriya @Tirth1306 @Shrey-Viradiya

Our dataset has nested labels. If it is normal then okay, but if it is cases of pneumonia then whether it is bacteria or virus. If virus then COVID or other.

• Please look at this and suggest needful @tirthasheshpatel @tirth-hihoriya @Tirth1306.
• Right now I have only separated from pneumonia to normal only.

look at the table:

data.zip

• Code for making data directory is not written.

• Need to add Class Activation Mapping
• Additional model testing on other dataset from kaggle

We have no documentation of the trained models or our python files. @tirth-hihoriya had concerns. Let's keep documentation after we finish all the training and testing...