project_state.mp4

DL Plotting Tool

A repo for creating easily all the necessary plots by just letting the 'DLPlotter' collect all the data. Everything else will be handled. Writing useful plots can be very time-consuming let alone to present it in a proper way. With this tool you can save your time and invest it in optimizing your model. If you have your own plots you want to add - you can do this too.

How to:

• Collect the data for the Configuration Loss plot. (Hyperparameter configuration). Alternatively you can also just load a checkpoint.

        plotter = DLPlotter()     # add this line
        model = MyModel()
        ...
        total_loss = 0
        for epoch in range(5):
            for step, (x, y) in enumerate(loader):
                ...
                output = model(x)
                loss = loss_func(output, y)
                total_loss += loss.item()
                ...
        config = dict(lr=0.001, batch_size=64, ...)
        plotter.collect_parameter("exp001"", config, total_loss / (5 * len(loader))     # add this line
        plotter.construct()     # add this line

• Collect the data for the Loss plot. (Training / validation)

        plotter = DLPlotter()     # add this line
        model = MyModel()
        ...
        for epoch in range(5):
            for step, (x, y) in enumerate(loader):
                ...
                output = model(x)
                loss = loss_func(output, y)

                plotter.collect_loss("exp001", len(loaders), epoch, step, loss.item(), "train")     # add this line
                ...
        plotter.construct()     # add this line

• Collect the data for the Weight Distribution plot.

        plotter = DLPlotter()     # add this line
        model = MyModel()
        ...
        for epoch in range(5):
            for step, (x, y) in enumerate(loader):
                ...
                weights = dict(layer1=model.layer1.weight.detach().clone(),
                               layer2=model.layer2.weight.detach().clone(), ...)

                plotter.collect_weights("exp001", len(loader), epoch, step, weights)     # add this line
                ...
        plotter.construct()     # add this line

To Do:

1. Accuracy figures as a subclass of WindowFig needs to be done. You can find it in folder ./dlFigures/ .
   
2. Finish / create test files - you find them in ./pToolTest/ .
3. Improve Callbacks - update mechanism. You find them in constructor.py in the class DashStruct.
4. Loading data from checkpoints (pickle files) in DLPlotter class.
5. Adding custom plots also in DLPlotter class.
6. (I am atm fixing 2 bugs with the weight distribution and interpolation of loss figures.)