Running Keypoint RCNN on the new wellplate_v8 data

We tried running the KeypointRCNN model with different (hacky) loss functions, to emphasize issues where an individual point may be far off the ground truth. - max_keypoint_loss: compute the loss individually for all 5 keypoints, then take the max - hybrid_loss: average of max loss and average loss - wellplate_v8_data: non-modified version using cross entropy.

Testing out different batch sizes by doing gradient accumulation. The original model with batch_size = 2 is `learning_curve_338_early_stopping`. The other models are denoted by _{x}_{y} where x is the batch size and y the mini-batch size. It looks like the performance is getting worse for larger batch sizes.

To understand the impact of training set size, I retrained the Wellplate KeypointRCNN model for different sizes. I used an Adam optimizer and early stopping (patience = 2) - lr = 0.0001 - optimizer = torch.optim.Adam(params, lr=lr, weight_decay=0.0005) - no learning rate decay! Just looking at the validation pixel error (average Euclidean distance between predictions and target), it seems that we should expect further improvement from getting more data.