Add GIOU loss #477

fsx950223 · 2019-09-04T11:41:02Z

Implementation of GIOU loss

seanpmorgan · 2019-09-12T01:46:06Z

Hi @fsx950223 thank you for this contribution and #493 . Just wanted to let you know that we'll review ASAP, but are a bit backed up given the work that has been going into building packages for newly released TF2 packages.

Also, just for the future.. if you wouldn't mind opening an issue to discuss before submitting a PR that'd be much appreciated.

fsx950223 · 2019-09-12T01:52:18Z

Hi @fsx950223 thank you for this contribution and #493 . Just wanted to let you know that we'll review ASAP, but are a bit backed up given the work that has been going into building packages for newly released TF2 packages.

Also, just for the future.. if you wouldn't mind opening an issue to discuss before submitting a PR that'd be much appreciated.

OK

facaiy · 2019-09-18T06:14:41Z

@Squadrick Hi, Dheeraj, could you take a look if you have time? Thanks :-)

WindQAQ

Hi @fsx950223 , sorry for the late reply. I have taken a deeper look into the paper, but want to give you some general feedback on this. Also, if it's possible, could you elaborate how to compute GIoU in test cases or any other places? Thanks!

tensorflow_addons/losses/giou_loss.py

tensorflow_addons/losses/giou_loss_test.py

fsx950223 · 2019-10-31T03:02:48Z

Hi @fsx950223 , sorry for the late reply. I have taken a deeper look into the paper, but want to give you some general feedback on this. Also, if it's possible, could you elaborate how to compute GIoU in test cases or any other places? Thanks!

I will modify the code on the weekend.

fsx950223 · 2019-11-04T04:10:01Z

Changed. @WindQAQ

WindQAQ

LGTM. Only some documentation issues need to be addressed. Thank you!

tensorflow_addons/losses/giou_loss.py

WindQAQ · 2019-11-04T07:09:40Z

tensorflow_addons/losses/giou_loss.py

+    """
+    Args
+        b1: bbox.
+        b2: the other bbox.


Could we rename bbox to bounding box here? It should be clearer in docstring.

tensorflow_addons/losses/README.md

tensorflow_addons/losses/giou_loss.py

fsx950223 · 2019-11-04T07:26:43Z

Changed.

fsx950223 · 2019-11-13T02:38:22Z

Only works on matched bboxs, still have a lot of work to do. Does it need to support unmatched situation(N bboxs x M bboxs)?

fsx950223 · 2019-11-13T05:30:51Z

There are some tricks by using broadcasts to calculate unmatched gious.

WindQAQ

Very nice! Some doc needed to change. Thank you very much for the contribution!

WindQAQ · 2019-11-15T07:29:42Z

tensorflow_addons/losses/giou_loss.py

+    return 1 - giou
+
+
+def do_giou_calculate(b1, b2, mode='giou'):


Could we make this function private? Like _calculate_giou.

It's a common function, maybe I should migrate it to tfa.image module?

Umm, I would suggest leave it private here; let's move it in the future if there is any request!

WindQAQ · 2019-11-15T07:31:07Z

tensorflow_addons/losses/giou_loss.py

+    """Implements the GIoU loss function.
+
+    GIoU loss was first introduced in the
+    [Generalized Intersection over Union paper](https://giou.stanford.edu/GIoU.pdf).


Generalized Intersection over Union: A Metric and A Loss for Bounding Box Regression

WindQAQ · 2019-11-15T07:31:55Z

tensorflow_addons/losses/giou_loss.py

+
+    GIoU loss was first introduced in the
+    [Generalized Intersection over Union paper](https://giou.stanford.edu/GIoU.pdf).
+    GIoU is a enhance for model which use IOU in object detection.


NIT: ... which uses IoU in ...

WindQAQ · 2019-11-15T07:32:28Z

tensorflow_addons/losses/giou_loss.py

+    gl = tfa.losses.GIoU()
+    boxes1 = tf.constant([[4.0, 3.0, 7.0, 5.0], [5.0, 6.0, 10.0, 7.0]])
+    boxes2 = tf.constant([[3.0, 4.0, 6.0, 8.0], [14.0, 14.0, 15.0, 15.0]])
+    loss = gl(boxes1,boxes2)


Suggested change

loss = gl(boxes1,boxes2)

loss = gl(boxes1, boxes2)

WindQAQ · 2019-11-15T07:32:54Z

tensorflow_addons/losses/giou_loss.py

+
+    ```python
+    model = tf.keras.Model(inputs, outputs)
+    model.compile('sgd', loss=tf.keras.losses.GIoULoss())


Suggested change

model.compile('sgd', loss=tf.keras.losses.GIoULoss())

model.compile('sgd', loss=tfa.losses.GIoULoss())

WindQAQ · 2019-11-15T07:33:42Z

tensorflow_addons/losses/giou_loss.py

+    ```
+
+    Args:
+      mode: one of ['giou', 'iou'], decided to calculate giou loss or iou loss.


NIT: ... calculate GIoU or IoU loss.

Squadrick

@fsx950223 What happens when y_pred is integer instead of floating point?

Squadrick · 2019-11-19T12:13:19Z

tensorflow_addons/losses/README.md

@@ -5,6 +5,7 @@
 |:---------- |:----------- |:------------- |
 | contrastive |  @WindQAQ | [email protected] |
 | focal_loss | @SSaishruthi  | [email protected] |
+| giou_loss | @who who who  | [email protected] |


Could you put your GitHub username @fsx950223 instead of @who who who?

Squadrick · 2019-11-19T12:14:50Z

tensorflow_addons/losses/giou_loss.py

+    [Generalized Intersection over Union:
+    A Metric and A Loss for Bounding Box Regression]
+    (https://giou.stanford.edu/GIoU.pdf).
+    GIoU is a enhance for model which use IoU in object detection.


Nit: a enhance -> an enhacement
Nit: model -> models

Squadrick · 2019-11-19T12:16:00Z

tensorflow_addons/losses/giou_loss.py

+    """
+    Args:
+        y_true: true targets tensor. The coordinates of the each bounding
+        box in boxes are encoded as [y_min, x_min, y_max, x_max].


Add extra four space at the start of this line.

Squadrick · 2019-11-19T12:16:15Z

tensorflow_addons/losses/giou_loss.py

+        y_true: true targets tensor. The coordinates of the each bounding
+        box in boxes are encoded as [y_min, x_min, y_max, x_max].
+        y_pred: predictions tensor. The coordinates of the each bounding
+        box in boxes are encoded as [y_min, x_min, y_max, x_max].


Squadrick · 2019-11-19T12:16:47Z

tensorflow_addons/losses/giou_loss.py

+    y_true = tf.cast(y_true, y_pred.dtype)
+    giou = _calculate_giou(y_pred, y_true, mode)
+
+    # compute the final loss and return


Comment not required. It's obvious what the code is doing.

Squadrick · 2019-11-19T12:17:00Z

tensorflow_addons/losses/giou_loss.py

+    """
+    Args:
+        b1: bounding box. The coordinates of the each bounding box in boxes are
+        encoded as [y_min, x_min, y_max, x_max].


4 spaces at start.

Squadrick · 2019-11-19T12:17:13Z

tensorflow_addons/losses/giou_loss.py

+        b1: bounding box. The coordinates of the each bounding box in boxes are
+        encoded as [y_min, x_min, y_max, x_max].
+        b2: the other bounding box. The coordinates of the each bounding box
+        in boxes are encoded as [y_min, x_min, y_max, x_max].


Squadrick

@fsx950223 Thanks for the changes. Just a few small nits.

tensorflow_addons/losses/giou_loss.py

Squadrick · 2019-11-20T09:05:38Z

tensorflow_addons/losses/giou_loss.py

@@ -86,20 +86,21 @@ def giou_loss(y_true, y_pred, mode='giou'):
    if mode not in ['giou', 'iou']:
        raise ValueError("Value of mode should be 'iou' or 'giou'")
    y_pred = tf.convert_to_tensor(y_pred)
+    if y_pred.dtype.is_floating is not True:


This can be written as:

if not y_pred.dtype.is_floating:

Squadrick

@fsx950223 LGTM!

cc: @WindQAQ

WindQAQ

Thanks for the contribution!

fsx950223 requested review from facaiy and WindQAQ as code owners September 4, 2019 11:41

googlebot added the cla: yes label Sep 4, 2019

fsx950223 changed the title ~~WIP: Add GIOU loss~~ [WIP] Add GIOU loss Sep 5, 2019

fsx950223 changed the title ~~[WIP] Add GIOU loss~~ Add GIOU loss Sep 5, 2019

fsx950223 force-pushed the loss branch from 7038cec to 672549f Compare September 5, 2019 02:46

Squadrick added the losses label Sep 8, 2019

seanpmorgan added the kokoro:force-run label Sep 8, 2019

kokoro-team removed the kokoro:force-run label Sep 8, 2019

fsx950223 force-pushed the loss branch 2 times, most recently from 54732a3 to 4568ba4 Compare September 9, 2019 01:35

seanpmorgan added the kokoro:force-run label Sep 9, 2019

kokoro-team removed the kokoro:force-run label Sep 9, 2019

seanpmorgan added the kokoro:force-run label Sep 9, 2019

kokoro-team removed the kokoro:force-run label Sep 9, 2019

facaiy requested a review from a team September 18, 2019 06:13

WindQAQ requested changes Oct 30, 2019

View reviewed changes

tensorflow_addons/losses/giou_loss.py Outdated Show resolved Hide resolved

tensorflow_addons/losses/giou_loss.py Outdated Show resolved Hide resolved

tensorflow_addons/losses/giou_loss.py Show resolved Hide resolved

tensorflow_addons/losses/giou_loss_test.py Outdated Show resolved Hide resolved

fsx950223 force-pushed the loss branch from 4568ba4 to 311c64c Compare November 4, 2019 04:08

WindQAQ requested changes Nov 4, 2019

View reviewed changes

fsx950223 added 5 commits November 5, 2019 01:53

Add GIOU loss

ca4197d

Refact giou calculate

27ae8ea

Fix doc

ae1a020

Update Readme.md

1606430

Format code

eb7458d

fsx950223 added 2 commits November 13, 2019 02:57

add one bbox test

ec11713

add different shapes test case

8d4b3dd

fsx950223 changed the title ~~[WIP]: Add GIOU loss~~ Add GIOU loss Nov 13, 2019

format code

55f4d0b

seanpmorgan added the kokoro:force-run label Nov 13, 2019

kokoro-team removed the kokoro:force-run label Nov 13, 2019

WindQAQ requested changes Nov 15, 2019

View reviewed changes

fsx950223 added 2 commits November 17, 2019 06:10

fix docs

f830058

make private

e42ccd5

fsx950223 force-pushed the loss branch from 57ff353 to e42ccd5 Compare November 19, 2019 11:28

Squadrick suggested changes Nov 19, 2019

View reviewed changes

seanpmorgan added the kokoro:force-run label Nov 19, 2019

kokoro-team removed the kokoro:force-run label Nov 19, 2019

fsx950223 added 2 commits November 20, 2019 07:16

add interger test

a03a320

format code

4b915d1

Squadrick suggested changes Nov 20, 2019

View reviewed changes

change expression

f8ae33c

Squadrick added the kokoro:force-run label Nov 20, 2019

kokoro-team removed the kokoro:force-run label Nov 20, 2019

Squadrick approved these changes Nov 20, 2019

View reviewed changes

Squadrick added the awaiting testing (then merge) label Nov 20, 2019

WindQAQ added the kokoro:force-run label Nov 21, 2019

kokoro-team removed the kokoro:force-run label Nov 21, 2019

WindQAQ self-requested a review November 21, 2019 16:08

WindQAQ approved these changes Nov 21, 2019

View reviewed changes

WindQAQ merged commit 5d7dee6 into tensorflow:master Nov 21, 2019

fsx950223 mentioned this pull request Jan 22, 2020

add ciou and diou #914

Closed

bhack mentioned this pull request Jun 16, 2020

[Object_Detection] Add IOUSimilarity layer. keras-team/keras-cv#3

Closed

	model.compile('sgd', loss=tf.keras.losses.GIoULoss())
	model.compile('sgd', loss=tfa.losses.GIoULoss())

Add GIOU loss #477

Add GIOU loss #477

Uh oh!

Conversation

fsx950223 commented Sep 4, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

seanpmorgan commented Sep 12, 2019

Uh oh!

fsx950223 commented Sep 12, 2019

Uh oh!

facaiy commented Sep 18, 2019

Uh oh!

WindQAQ left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

fsx950223 commented Oct 31, 2019

Uh oh!

fsx950223 commented Nov 4, 2019

Uh oh!

WindQAQ left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

fsx950223 commented Nov 4, 2019

Uh oh!

fsx950223 commented Nov 13, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fsx950223 commented Nov 13, 2019

Uh oh!

WindQAQ left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Squadrick left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

fsx950223 commented Sep 4, 2019 •

edited

Loading

fsx950223 commented Nov 13, 2019 •

edited

Loading

Squadrick left a comment •

edited

Loading

Squadrick left a comment •

edited

Loading