Changing batch size for inference

[cinjon]

How do I change the batch size when running inference? No matter what I change cfg.SOLVER.IMS_PER_BATCH, it's always doing just one image:

        data_loader = build_detection_test_loader(cfg, tag)
        evaluator = COCOEvaluator(tag,
                                  tasks=('bbox',),
                                  distributed=True,
                                  output_dir=output_dir)
        results_i = inference_on_dataset(model, data_loader, evaluator)

[Wilann]

@cinjon Did you manage to find a way to do this?

[cinjon]

Yes, it's right here:

detectron2/detectron2/data/build.py

Line 473 in 730ccef

batch_sampler = torchdata.sampler.BatchSampler(sampler, 1, drop_last=False)

.

[Wilann]

Thank you so much for your reply! If possible, could you post your full dataloading code? I'm a bit confused.

[cinjon]

I don't quite understand your request. If you change that to have a larger number than 1, it will increase the inference batch side when using the test dataloader function.

[Wilann]

I'm trying to use a standard PyTorch dataloader (with a simple custom dataset) to load the data like so:

class Frames_Dataset(Dataset):
    
    def __init__(self, frames):

        self.preprocess = transforms.Compose([
            transforms.ToPILImage(),
            transforms.Resize(512),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
        ])

        self.frames = frames

    def __getitem__(self, index):
        frame_preprocessed = self.preprocess(self.frames[index])
        frame_detectron2_format = [{'image': frame} for frame in frame_preprocessed]
        return frame_detectron2_format

    def __len__(self):
        return len(self.frames)

frames = ... # list of frames of shape (8393, 3, 512, 512)
frames_dataset = Frames_Dataset(frames)
frames_dataloader = DataLoader(frames_dataset, batch_size=100, shuffle=False, num_workers=4)

cfg = get_cfg()
config_name = "COCO-InstanceSegmentation/mask_rcnn_R_101_FPN_3x.yaml"

cfg.merge_from_file(model_zoo.get_config_file(config_name))

cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.9

cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url(config_name)

predictor = DefaultPredictor(cfg)

model = build_model(cfg)
model.eval()

for batch_num, batch in numerate(frames_dataloader):
    outputs = predictor(batch)

Can I not do this? In this part in the __getitem__ ([{'image': frame} for frame in frame_preprocessed]) I'm trying to format my input required specified in the docs