Add PSROIPooling Function #545

knorth55 · 2018-03-27T05:36:20Z

Related to #538

add psroi_pooling_2d
add test for psroi_pooling_2d
add doc for psroi_pooling_2d
add test to check CPU and GPU output is same.

yuyu2172 · 2018-03-27T06:20:06Z

chainercv/functions/psroi_pooling_2d.py

+    Args:
+        x (~chainer.Variable): Input variable. The shape is expected to be
+            4 dimentional: (n: batch, c: channel, h, height, w: width).
+        rois (~chainer.Variable): Input roi variable. The shape is expected to


I think it is not necessary to make the API of this function to follow chainer.functions.roi_pooling_2d.
Rather, it would be better to follow the preferred interface of ChainerCV.

Can you change the second input to rois and roi_indices as defined here? (bbox is yx order)

rois (array): A bounding box array containing coordinates of proposal boxes. This is a concatenation of bounding box arrays from multiple images in the batch. Its shape is :math:`(R', 4)`. Given :math:`R_i` proposed RoIs from the :math:`i` th image, :math:`R' = \\sum _{i=1} ^ N R_i`. roi_indices (array): An array containing indices of images to which bounding boxes correspond to. Its shape is :math:`(R',)`.

https://github.com/chainer/chainercv/blob/master/chainercv/links/model/faster_rcnn/faster_rcnn_vgg.py#L201

This would require some change to kernels.

yuyu2172 · 2018-03-27T06:25:25Z

tests/functions_tests/test_psroi_pooling_2d.py

+    def test_backward_gpu(self):
+        self.check_backward(cuda.to_gpu(self.x), cuda.to_gpu(self.rois),
+                            cuda.to_gpu(self.gy))
+


How about adding a test to check that CPU and GPU outputs are the same?

Related:
https://github.com/chainer/chainer/blob/master/tests/chainer_tests/functions_tests/array_tests/test_spatial_transformer_sampler.py#L119

yuyu2172 · 2018-03-27T06:26:26Z

chainercv/functions/psroi_pooling_2d.py

+    """Position Sensitive Region of Interest (ROI) pooling function.
+
+    This function computes position sensitive average of input spatial patch
+    with the given region of interest. Each ROI are splitted into


region of interest --> region of interests

Each ROI are --> Each ROI is

yuyu2172 · 2018-03-27T06:27:11Z

chainercv/functions/psroi_pooling_2d.py

+    This function computes position sensitive average of input spatial patch
+    with the given region of interest. Each ROI are splitted into
+    :math:`(group\_size, group\_size)` regions, and position sensitive values
+    in each region is inferenced.


infrenced --> computed

yuyu2172 · 2018-03-27T06:29:59Z

chainercv/functions/psroi_pooling_2d.py

+
+        bottom_data, bottom_rois = inputs
+        channels, height, width = bottom_data.shape[1:]
+        n_rois = bottom_rois.shape[0]


Could you follow ChainerCV's naming convention? (n_{Singular} instead of n_{Plural})
n_rois --> n_roi

yuyu2172 · 2018-03-27T06:31:18Z

chainercv/functions/psroi_pooling_2d.py

+            roi_width = max(xmax - xmin, 0.1)
+            roi_height = max(ymax - ymin, 0.1)
+
+            strided = 1. * channels / self.outd


Could you remove 1. and add from __future__ import division?

yuyu2172 · 2018-03-27T06:31:34Z

chainercv/functions/psroi_pooling_2d.py

+            roi_width = max(xmax - xmin, 0.1)
+            roi_height = max(ymax - ymin, 0.1)
+
+            strideh = 1. * roi_height / self.outh


yuyu2172 · 2018-03-27T06:33:10Z

chainercv/functions/psroi_pooling_2d.py

+        for i_roi in six.moves.range(n_rois):
+            idx, xmin, ymin, xmax, ymax = bottom_rois[i_roi]
+            idx = int(idx)
+            xmin = round(xmin * self.spatial_scale)


The variables are defined in xy, wh and yx, hw orders.
This is inconsistent. Could you reorder them to yx, hw?

can it be different from roi_pooling_2d in chainer master branch?

Yes
#545 (comment)

yuyu2172 · 2018-03-27T06:37:44Z

chainercv/functions/psroi_pooling_2d.py

@@ -0,0 +1,331 @@
+# Cuda Kernel Original work by Haozhi Qi (@Oh233)


Mention license

knorth55 · 2018-03-27T16:02:08Z

@yuyu2172 I'm sorry that I rebase the commits again.
I follow your review and update.
Other thing is that I updated variables such as outh -> out_h and sliceh -> slice_h.
Also, I renamed outd -> out_c.

knorth55 · 2018-03-28T15:22:43Z

tests/functions_tests/test_psroi_pooling_2d.py

+
+    @attr.gpu
+    @condition.retry(3)
+    def test_consistency_with_gpu(self):


I check that calculation output of GPU and CPU are same in this test.

yuyu2172 · 2018-03-29T00:20:10Z

chainercv/functions/psroi_pooling_2d.py

+            x_min = round(x_min * self.spatial_scale)
+            y_max = round(y_max * self.spatial_scale)
+            x_max = round(x_max * self.spatial_scale)
+            roi_width = max(x_max - x_min, 0.1)


Order of the two lines are opposite.

roi_height = max(y_max - y_min, 0.1) roi_width = max(x_max - x_min, 0.1)

yuyu2172 · 2018-03-29T00:21:27Z

chainercv/functions/psroi_pooling_2d.py

+            'float32 top_data',
+            '''
+            // pos in output filter
+            int pw = i % pooled_width;


yuyu2172 · 2018-03-29T00:21:35Z

chainercv/functions/psroi_pooling_2d.py

+                round(bottom_rois[n * 4 + 3])) * spatial_scale;
+
+            // Force too small ROIs to be 1x1
+            float roi_width = max(roi_end_w - roi_start_w, 0.1);  // avoid 0


yuyu2172 · 2018-03-29T00:21:47Z

chainercv/functions/psroi_pooling_2d.py

+            bool is_empty = (hend <= hstart) || (wend <= wstart);
+
+            // Compute c at bottom
+            int gw = floor(


yuyu2172 · 2018-03-29T00:21:53Z

chainercv/functions/psroi_pooling_2d.py

+                static_cast<float>(pw) * group_size / pooled_width);
+            int gh = floor(
+                static_cast<float>(ph) * group_size / pooled_height);
+            gw = min(max(gw, 0), group_size - 1);


yuyu2172 · 2018-03-29T00:23:04Z

chainercv/functions/psroi_pooling_2d.py

+                static_cast<float>(pw) * bin_size_w + roi_start_w);
+            int hstart = floor(
+                static_cast<float>(ph) * bin_size_h + roi_start_h);
+            int wend = ceil(


yuyu2172 · 2018-03-29T00:23:09Z

chainercv/functions/psroi_pooling_2d.py

+                static_cast<float>(ph + 1.0) * bin_size_h + roi_start_h);
+
+            // Add roi offsets and clip to input boundaries
+            wstart = min(max(wstart, 0), width);


yuyu2172 · 2018-03-29T00:23:16Z

chainercv/functions/psroi_pooling_2d.py

+            // Add roi offsets and clip to input boundaries
+            wstart = min(max(wstart, 0), width);
+            hstart = min(max(hstart, 0), height);
+            wend = min(max(wend, 0), width);


yuyu2172 · 2018-03-29T00:23:24Z

chainercv/functions/psroi_pooling_2d.py

+            bool is_empty = (hend <= hstart) || (wend <= wstart);
+
+            // Compute c at bottom
+            int gw = floor(


yuyu2172 · 2018-03-29T00:23:31Z

chainercv/functions/psroi_pooling_2d.py

+                static_cast<float>(pw) * group_size / pooled_width);
+            int gh = floor(
+                static_cast<float>(ph) * group_size / pooled_height);
+            gw = min(max(gw, 0), group_size - 1);


knorth55 added the feature label Mar 27, 2018

knorth55 self-assigned this Mar 27, 2018

knorth55 requested a review from yuyu2172 March 27, 2018 05:36

knorth55 force-pushed the psroipooling branch from f9f0117 to 94d1c3d Compare March 27, 2018 06:00

yuyu2172 reviewed Mar 27, 2018

View reviewed changes

chainercv/functions/psroi_pooling_2d.py Outdated

@@ -0,0 +1,331 @@

# Cuda Kernel Original work by Haozhi Qi (@Oh233)

Copy link

Member

yuyu2172 Mar 27, 2018

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Mention license

knorth55 force-pushed the psroipooling branch from 1f3b694 to 5476560 Compare March 27, 2018 15:59

knorth55 changed the title ~~Add PSROIPooling Function~~ [WIP] Add PSROIPooling Function Mar 27, 2018

knorth55 force-pushed the psroipooling branch 2 times, most recently from aa2c88b to e63e5e9 Compare March 28, 2018 15:16

knorth55 changed the title ~~[WIP] Add PSROIPooling Function~~ Add PSROIPooling Function Mar 28, 2018

knorth55 force-pushed the psroipooling branch from e63e5e9 to d2b19ca Compare March 28, 2018 15:19

knorth55 commented Mar 28, 2018

View reviewed changes

knorth55 force-pushed the psroipooling branch 4 times, most recently from 595e70e to 0d75610 Compare March 28, 2018 16:32

yuyu2172 reviewed Mar 29, 2018

View reviewed changes

knorth55 force-pushed the psroipooling branch from 0d75610 to 878f6c3 Compare March 29, 2018 09:50

knorth55 added 4 commits March 29, 2018 18:52

add psroi_pooling_2d function

62edac2

add tests for psroi_pooling_2d

6fa06f8

add doc for psroi_pooling_2d

5fcfc2d

add test_consistency for psroi_pooling_2d

f5fe9d9

knorth55 force-pushed the psroipooling branch from 878f6c3 to f5fe9d9 Compare March 29, 2018 09:52

yuyu2172 approved these changes Mar 29, 2018

View reviewed changes

yuyu2172 merged commit 50cf990 into chainer:master Mar 29, 2018

yuyu2172 added this to the v0.9 milestone Mar 29, 2018

knorth55 deleted the psroipooling branch March 29, 2018 15:59

yuyu2172 mentioned this pull request May 28, 2018

Add Instance Segmentation #538

Open

21 tasks

knorth55 mentioned this pull request Feb 18, 2019

add PS RoI max/average align #796

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Add PSROIPooling Function #545

Add PSROIPooling Function #545

knorth55 commented Mar 27, 2018 •

edited

Loading

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

knorth55 Mar 27, 2018

yuyu2172 Mar 27, 2018

yuyu2172 Mar 27, 2018

knorth55 commented Mar 27, 2018

knorth55 Mar 28, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

yuyu2172 Mar 29, 2018

		@@ -0,0 +1,331 @@
		# Cuda Kernel Original work by Haozhi Qi (@Oh233)

Add PSROIPooling Function #545

Add PSROIPooling Function #545

Conversation

knorth55 commented Mar 27, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

knorth55 commented Mar 27, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

knorth55 commented Mar 27, 2018 •

edited

Loading