dnn(darknet-importer): add grouped convolutions, sigmoid, swish, scale_channels

[YashasSamaga]

resolves #15987

Merge with extra: opencv/opencv_extra#711

This pull request changes

• adds support for grouped convolutions
• adds new activations
  • sigmoid
  • swish
• adds support for scale channels

Prediction using OpenCV CPU backend:

Darknet Reference:

tvmonitor: 44%
dog: 78%
cat: 27%
bicycle: 58%
truck: 89%
car: 33%

force_builders=Custom,Custom Win,Custom Mac
build_image:Custom=ubuntu-openvino-2020.1.0:16.04
build_image:Custom Win=openvino-2020.1.0
build_image:Custom Mac=openvino-2020.1.0

test_modules:Custom=dnn,python2,python3,java
test_modules:Custom Win=dnn,python2,python3,java
test_modules:Custom Mac=dnn,python2,python3,java

buildworker:Custom=linux-1
# disabled due high memory usage: test_opencl:Custom=ON
test_opencl:Custom=OFF
test_bigdata:Custom=1
test_filter:Custom=*

[dkurt]

Please add single layer tests for each of newly added layers. You may take a look at #16358 as a reference.

find out why darknet output doesn't match with OCV CPU output

What about input image resolution? Have you used RGB or BGR input?

[YashasSamaga]

@dkurt For the convolution and activation tests, should I create a new cfg files for swish, logistic and grouped convolution OR add more layers to existing convolutional.cfg?

[dkurt]

@YashasSamaga, I think the best way if to check existing tests and replace some of duplicated activation to new ones.

[YashasSamaga]

scale_channels does not work with batches.

opencv/modules/dnn/src/layers/scale_layer.cpp

Lines 86 to 94 in 3efa783

	// There is a mode when we multiply a first blob by a second one
	// instead of trainable weights.
	Mat weights = blobs.empty() ? inputs[1] : (hasWeights ? blobs[0] : Mat());
	Mat bias = hasBias ? blobs.back().reshape(1, 1) : Mat();
	if (!weights.empty())
	weights = weights.reshape(1, 1);
	MatShape inpShape = shape(inpBlob);
	const int numWeights = !weights.empty() ? weights.total() : bias.total();
	CV_Assert(numWeights != 0);

weights is reshaped (assuming that batch size is one?).

opencv/modules/dnn/src/layers/scale_layer.cpp

Lines 98 to 104 in 3efa783

	int endAxis;
	for (endAxis = axis + 1; endAxis <= inpBlob.dims; ++endAxis)
	{
	if (total(inpShape, axis, endAxis) == numWeights)
	break;
	}
	CV_Assert(total(inpShape, axis, endAxis) == numWeights);

This code ignores the batch dimension and tries to match numWeights (i.e. weights.total()) to a subrange in the input tensor.

Example:

input shape: 2 3 3 3
weights input: 2 3 1 1

weights is reshaped to 6 1. The matching code tries to look for a subrange in the input (2 3 3 3) which totals to 6 but it fails to find any.

I can tweak the ScaleLayer to make it behave like Darknet's scale channels but I fear that it would break things and other backends may not be setup to behave like Darknet's scale channels.

Or I can add a new layer which behaves like Darknet's scale channels.

[dkurt]

@YashasSamaga, let's open a new issue for that observation. So for Scale layer there will be a new test with batch size 1 but for Swish and Logistic - let's try to include them to existing ones.

[YashasSamaga]

I have used a different branch name in opencv_extra PR. How can the CI be informed about it?

[alalek]

Need to adjust Inference Engine checks.
@l-bat Please suggest required changes.

[l-bat]

Please, rebase with updated 3.4. PR #16738 fixed your tests.