Access detector output as raw tensor

[aubreydacunha]

Hello,

I am working on training a model on my own dataset, but I am getting some odd behaviors. I can debug them myself, but it would be very helpful to actually see the raw tensor output from the model before it gets processed into bounding boxes. However, every time I try to find that in the code, I get lost in a tangle of multiple models with multiple base classes with multiple versions of forward (technically, one version of forward that dispatches to multiple implementations based on some non-intuitive configuration settings). It's a similar sort of issue to this.

I'm assuming there isn't a line anywhere as simple as

output = model(input)

but where would I look to actually see the input and output tensors?

If the answer is model-dependent, I am specifically fine-tuning a YOLOX model.

[jbwang1997]

In mmdet 2.x, it needs to modify the code to achieve the function. You can hack in the model and directly return the results of self.forward(feats) in bbox_head.

If you try the newest mmdet 3.0, you can directly get the raw outputs by using output = model(input, mode='tensor').