SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Enable dynamic resolution input for Swin Transformer and variants (#30656)

Browse Source 
* add interpolation of positional encoding support to swin

* add style changes

* use default image processor and make size a dictionary

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* remove logits testing

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Refactor image size validation logic when interpolation is disabled

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* remove asserts in modeling

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* add dynamic resolution input support to swinv2

* change size to ensure interpolation encoding path is triggered

* set interpolate_pos_encoding default value to False

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* set interpolate_pos_encoding default value to False

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* set interpolate_pos_encoding default value to False

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* set interpolate_pos_encoding default value to False

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* set interpolate_pos_encoding default value to False

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* set interpolate_pos_encoding default value to False

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* set interpolate_pos_encoding default value to False

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* set interpolate_pos_encoding default value to False

* add dynamic resolution input to donut swin

* add dynamic resolution input to maskformer swin

---------

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
This commit is contained in:
Abhiroop Tejomay
2024-05-17 13:38:46 -04:00
committed by GitHub
parent b6eb708bf1
commit 481a957814
6 changed files with 291 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
tests/models/swin/test_modeling_swin.py
View File

@@ -493,6 +493,26 @@ class SwinModelIntegrationTest(unittest.TestCase):
expected_slice = torch.tensor([-0.0948, -0.6454, -0.0921]).to(torch_device)
self.assertTrue(torch.allclose(outputs.logits[0, :3], expected_slice, atol=1e-4))
@slow
def test_inference_interpolate_pos_encoding(self):
# Swin models have an `interpolate_pos_encoding` argument in their forward method,
# allowing to interpolate the pre-trained position embeddings in order to use
# the model on higher resolutions.
model = SwinModel.from_pretrained("microsoft/swin-tiny-patch4-window7-224").to(torch_device)
image_processor = self.default_image_processor
image = Image.open("./tests/fixtures/tests_samples/COCO/000000039769.png")
inputs = image_processor(images=image, size={"height": 481, "width": 481}, return_tensors="pt")
pixel_values = inputs.pixel_values.to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(pixel_values, interpolate_pos_encoding=True)
# verify the logits
expected_shape = torch.Size((1, 256, 768))
self.assertEqual(outputs.last_hidden_state.shape, expected_shape)
@require_torch
class SwinBackboneTest(unittest.TestCase, BackboneTesterMixin):