Batch inference

[davidmasek]

Hello,
I've been using the model for some experiments with code based on the demo/ctw1500_detection.py. I would like to move from VisualizationDemo because it does not support batch inference. Looking through the source code I tried something like:

from detectron2.engine.defaults import DefaultPredictor

predictor = DefaultPredictor(cfg)

height, width = bgr_img.shape[:2]
image = predictor.transform_gen.get_transform(bgr_img).apply_image(bgr_img)
image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1))
inputs = {"image": image, "height": height, "width": width}

with torch.no_grad():
    predictions_single = predictor.model([inputs]) # works OK
    predictions_batch = predictor.model([inputs, inputs]) # crashes

Inference on single image is fine, but giving multiple images crashes inside mutil_path_fuse_module.py, specifically on feature_fuse = char_context + x + global_context with

RuntimeError: The size of tensor a (2) must match the size of tensor b (4) at non-singleton dimension 0

What should I do to get batch inference?

Thanks.

[davidmasek]

So I think I solved this?

Inside of Mutil_Path_Fuse_Module class the forward method just uses proposals[0] instead of iterating over them when there are multiple. However this looks like it would never work, so now I'm confused a bit (since I assume the model was used for batch inference at least in training).

@ying09 could you comment on this? Thanks.