How to retrieve the loss function of FasterRCNN for object detection?

vision
1

Context: I’m using the pretrained model fasterrcnn_resnet50_fpn. If the model is on train mode, it works like this:

model.train()
model(image, targets) # pass BOTH the image and the targets

# returns the loss
{'loss_classifier': tensor(2.0448, device='cuda:0', grad_fn=<NllLossBackward0>),
 'loss_box_reg': tensor(0.6015, device='cuda:0', grad_fn=<DivBackward0>),
 'loss_objectness': tensor(1.4119, device='cuda:0',
        grad_fn=<BinaryCrossEntropyWithLogitsBackward0>),
 'loss_rpn_box_reg': tensor(0.0668, device='cuda:0', grad_fn=<DivBackward0>)}

While if the model is on eval mode, it works like this:

model.eval()
model(image) # pass ONLY the image

# Returns the predicted targets, but not the loss
{'boxes': tensor([[142.8750, 135.8750, 176.1250, 169.1250],
        [130.8750, 116.8750, 164.1250, 150.1250],
        [ 91.1250, 180.1250, 113.8750, 202.8750],
        [131.1250, 145.1250, 153.8750, 167.8750],
        [ 65.1250, 103.1250,  87.8750, 125.8750],
        [ -3.8750,  91.1250,  18.8750, 113.8750],
        [208.1250,  21.1250, 230.8750,  43.8750]]), 'labels': tensor([3, 3, 4, 4, 4, 4, 4])}

What I want to be able to do is find the loss function that fasterrcnn_resnet50_fpn, such that I could pass the predicted and actual targets into the function to get the losses.

e.g.

image, actual_targets = X_train[0]
model.eval()
pred_targets = model(image)

loss_fn(pred_targets, actual_targets)

Does anyone know how I could do this?

2

The losses are calculated here in the GeneralizedRCNN.forward method so you might be able to reimplement the forward method and pass the targets to during the validation pass, too.

3

What I’m more looking for is a function to compare two sets of targets. I wouldn’t want to change the model as such, ideally I’d be able to run the loss function without running the model.

Is there any way I could do this instead?