Closed momopusheen closed 1 year ago
Maybe you can try to add one more resize augmentation
as the first augmentation in the config~ @momopusheen
dataloader.train = L(build_detection_train_loader)(
dataset=L(get_detection_dataset_dicts)(names="coco_2017_train"),
mapper=L(DetrDatasetMapper)(
augmentation=[
L(T.ResizeShortestEdge)(
short_edge_length=512,
max_size=512,
),
L(T.RandomFlip)(),
L(T.ResizeShortestEdge)(
short_edge_length=(480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800),
max_size=1333,
sample_style="choice",
),
],
augmentation_with_crop=[
L(T.ResizeShortestEdge)(
short_edge_length=512,
max_size=512,
),
L(T.RandomFlip)(),
L(T.ResizeShortestEdge)(
short_edge_length=(400, 500, 600),
sample_style="choice",
),
L(T.RandomCrop)(
crop_type="absolute_range",
crop_size=(384, 600),
),
L(T.ResizeShortestEdge)(
short_edge_length=(480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800),
max_size=1333,
sample_style="choice",
),
],
is_train=True,
mask_on=False,
img_format="RGB",
),
total_batch_size=16,
num_workers=4,
)
I'm not sure if this works for you, just a simple example, you can try to use a Resize augmentation to replace the ResizeShortEdge Augmentation, you can refer to d2's documentation to check the augmentation function there
I'm closing this issue~ Feel free to reopen it if needed~
Dear authors, thanks for your excellent work!
I have datas in various sizes. As I'm using a FPN-like structure, the input image size should be resized to [512, 512]
Now I realize this resize approach by modifying the detr_data_mapper.py
if self.augmentation_with_crop is None: image, transforms = T.apply_transform_gens(self.augmentation, image) else: if np.random.rand() > 0.5: image, transforms = T.apply_transform_gens([T.Resize((512, 512))], image)
However, in original code, it looks like several augmentation methods were set here:
`if np.random.rand() > 0.5:
image, transforms = T.apply_transform_gens(self.augmentation, image)`
Is it possible if I want to do the resize as well as using the augmentation? I mean, augment the resized [512, 512] images?
Thanks for your consideration!