dit-doclaynet / README.md
jzju's picture
update readme
2fe6dfe
metadata
library_name: transformers
pipeline_tag: image-segmentation
tags:
  - vision
  - image-segmentation
  - dit
datasets:
  - ds4sd/DocLayNet-v1.1
widget:
  - src: >-
      https://upload.wikimedia.org/wikipedia/commons/c/c3/LibreOffice_Writer_6.3.png
    example_title: Wiki

Trained for 4 epochs.

Usage:

image_processor = AutoImageProcessor.from_pretrained("microsoft/dit-large")
model = BeitForSemanticSegmentation.from_pretrained("jzju/dit-doclaynet")
image = Image.open('img.png').convert('RGB')
inputs = image_processor(images=image, return_tensors="pt")
outputs = model(**inputs)
# logits are of shape (batch_size, num_labels, height, width)
logits = outputs.logits
out = logits[0].detach()
out.size()
for i in range(11):
    plt.imshow(out[i])
    plt.show()

Labels:

1: Caption
2: Footnote
3: Formula
4: List-item
5: Page-footer
6: Page-header
7: Picture
8: Section-header
9: Table
10: Text
11: Title

Data label convert:

model = BeitForSemanticSegmentation.from_pretrained("microsoft/dit-base", num_labels=11)
ds = load_dataset("ds4sd/DocLayNet-v1.1")
mask = np.zeros([11, 1025, 1025])
for b, c in zip(d["bboxes"], d["category_id"]):
    b = [np.clip(int(bb), 0, 1025) for bb in b]
    mask[c - 1][b[1]:b[1]+b[3], b[0]:b[0]+b[2]] = 1
mask = [cv2.resize(a, dsize=(56, 56), interpolation=cv2.INTER_AREA) for a in mask]
d["label"] = np.stack(mask)