Add Color transforms #275

chsasank · 2017-09-27T11:23:17Z

Please see #271 for the discussion.

Currently added Hue and saturation colour transforms based on tf code: https://github.com/tensorflow/tensorflow/blob/r1.3/tensorflow/python/ops/image_ops_impl.py

chsasank · 2017-09-27T12:40:31Z

@alykhantejani just to confirm, gamma correction should be done on V channel of HSV right?

Edit: looks like not. RGB channels are corrected together. Ref

Also * Change adjust_saturation to use pillow implementation * Documentation made clear

chsasank · 2017-09-27T14:31:25Z

I'm not using code from #27 because implementations are on tensors not on PIL images.

chsasank · 2017-09-27T20:15:55Z

Hey @alykhantejani, I have actually implemented all the transforms: brightness, contrast, saturation and gamma. I need to implement a transform class (and tests) using these functions and PR would be complete. I think this makes #27 redundant actually.

fmassa

LGTM, thanks!
I have only one comment that I think should be addressed (the overflow warning), the rest are minor details.
Also, could you add basic tests to those transforms?

torchvision/transforms.py

+
+    np_h = np.array(h, dtype='uint8')
+    # uint8 addition take cares of rotation across boundaries
+    np_h += np.uint8(hue_factor * 255)


torchvision/transforms.py

+
+    h, s, v = img.convert('HSV').split()
+
+    np_h = np.array(h, dtype='uint8')


torchvision/transforms.py

+
+    img = img.convert('RGB')
+
+    np_img = np.array(img)


torchvision/transforms.py

+        np.random.shuffle(transforms)
+        transform = Compose(transforms)
+
+        return transform


alykhantejani

Thanks @chsasank!

LGTM once we change the requirement for input images to be RGB.

torchvision/transforms.py

+
+
+def adjust_contrast(img, contrast_factor):
+    """Adjust brightness of an Image.


torchvision/transforms.py

+            2 will enhance the saturation by a factor of 2.
+
+    Returns:
+        Adjusted image(s), same shape and DType as `image`.


torchvision/transforms.py

+    """Adjust hue of an image.
+
+    `image` is an RGB image.  The image hue is adjusted by converting the
+    image to HSV and cyclically shifting the intensities in hue channel (H).


torchvision/transforms.py

+    Returns:
+        PIL.Image: Hue adjusted image.
+    """
+    if not(hue_factor <= 0.5 and hue_factor >= -0.5):


torchvision/transforms.py

+        PIL.Image: Hue adjusted image.
+    """
+    if not(hue_factor <= 0.5 and hue_factor >= -0.5):
+        raise ValueError('hue_factor {} out of range.'.format(hue_factor))


torchvision/transforms.py

+def adjust_hue(img, hue_factor):
+    """Adjust hue of an image.
+
+    `image` is an RGB image.  The image hue is adjusted by converting the


torchvision/transforms.py

+    np_img = 255 * gain * ((np_img / 255) ** gamma)
+    np_img = np.uint8(np.clip(np_img, 0, 255))
+
+    img = Image.fromarray(np_img, 'RGB')


chsasank · 2017-10-02T09:22:32Z

Can somebody guide me how to write tests for these? As in what do I check for etc.

alykhantejani · 2017-10-02T09:24:22Z

@chsasank you can use the TF tests for these transforms as reference, located here

torchvision/transforms.py

+    input_mode = img.mode
+    img = img.convert('RGB')
+
+    np_img = np.array(img, dtype='float32')


fmassa · 2017-10-04T22:31:44Z

Thanks a lot @chsasank !

Kaixhin · 2017-10-05T11:51:09Z

Quick! Thanks @chsasank! Would it be possible to also document these on the readme?

alykhantejani · 2017-10-05T11:57:15Z

@Kaixhin I think we actually should solve the root issue and get these docs autogenerated from the docstrings themselves, like it's done in the pytorch repo

Add adjust_hue and adjust_saturation

35c2c09

chsasank mentioned this pull request Sep 27, 2017

[Feature Request] More Image Transforms (Brightness, Contrast, Hue) #271

Closed

chsasank added 2 commits September 27, 2017 19:21

Add adjust_brightness, adjust_contrast

6e59812

Also * Change adjust_saturation to use pillow implementation * Documentation made clear

Add adjust_gamma

998c14e

alykhantejani mentioned this pull request Sep 27, 2017

fb.resnet like transforms #27

Closed

Add ColorJitter

6fbf634

fmassa requested changes Sep 29, 2017

View reviewed changes

chsasank added 2 commits September 29, 2017 18:47

Address review comments

26734c6

Fix documentation for ColorJitter

7d303d1

alykhantejani suggested changes Oct 1, 2017

View reviewed changes

Address review comments 2

fe1e15f

chsasank added 2 commits October 2, 2017 23:50

Fallback to adjust_hue in case of BW images

d670c07

Add tests

b08d89d

alykhantejani reviewed Oct 3, 2017

View reviewed changes

torchvision/transforms.py Outdated

input_mode = img.mode

img = img.convert('RGB')

np_img = np.array(img, dtype='float32')

This comment was marked as off-topic.

Sign in to view

alykhantejani approved these changes Oct 3, 2017

View reviewed changes

fix dtype

f04be26

fmassa merged commit 618072b into pytorch:master Oct 4, 2017

esube mentioned this pull request Oct 10, 2017

[Feature Request] Random Spatial Transforms #287

Open

pmeier mentioned this pull request Jul 3, 2023

fix PIL version check #7716

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add Color transforms #275

Add Color transforms #275

chsasank commented Sep 27, 2017

chsasank commented Sep 27, 2017 •

edited

Loading

chsasank commented Sep 27, 2017

chsasank commented Sep 27, 2017 •

edited

Loading

fmassa left a comment

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

alykhantejani left a comment

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

chsasank commented Oct 2, 2017

alykhantejani commented Oct 2, 2017

This comment was marked as off-topic.

fmassa commented Oct 4, 2017

Kaixhin commented Oct 5, 2017

alykhantejani commented Oct 5, 2017


		h, s, v = img.convert('HSV').split()

		np_h = np.array(h, dtype='uint8')



		def adjust_contrast(img, contrast_factor):
		"""Adjust brightness of an Image.

Add Color transforms #275

Add Color transforms #275

Conversation

chsasank commented Sep 27, 2017

chsasank commented Sep 27, 2017 • edited Loading

chsasank commented Sep 27, 2017

chsasank commented Sep 27, 2017 • edited Loading

fmassa left a comment

Choose a reason for hiding this comment

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

alykhantejani left a comment

Choose a reason for hiding this comment

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

This comment was marked as off-topic.

chsasank commented Oct 2, 2017

alykhantejani commented Oct 2, 2017

This comment was marked as off-topic.

fmassa commented Oct 4, 2017

Kaixhin commented Oct 5, 2017

alykhantejani commented Oct 5, 2017

chsasank commented Sep 27, 2017 •

edited

Loading

chsasank commented Sep 27, 2017 •

edited

Loading