[TinyYolo v2] Bug in maxpool?

[owulveryck]

This commit allows the model tiny Yolo v2 to be compiled and executed With Gorgonia.

Sadly the execution does not give the expected result:

➜  model_zoo_executor git:(tiny-yolov2) ✗  export MODELDIR=~/Documents/tiny_yolov2
➜  model_zoo_executor git:(tiny-yolov2) ✗ go run main.go -model $MODELDIR/model.onnx -input $MODELDIR/test_data_set_0/input_0.pb -output $MODELDIR/test_data_set_0/output_0.pb

        Error Trace:    main.go:72
                                                proc.go:200
                                                asm_amd64.s:1337
        Error:          Max difference between -0.17929432 and 0.056231752 allowed is 0.005, but difference was -0.23552606999874115
        Messages:       the two tensors should be equal.
exit status 1

According to this blog post the architecture should be:

Layer         kernel  stride  output shape
---------------------------------------------
Input                          (416, 416, 3)
Convolution    3×3      1      (416, 416, 16)
MaxPooling     2×2      2      (208, 208, 16)
Convolution    3×3      1      (208, 208, 32)
MaxPooling     2×2      2      (104, 104, 32)
Convolution    3×3      1      (104, 104, 64)
MaxPooling     2×2      2      (52, 52, 64)
Convolution    3×3      1      (52, 52, 128)
MaxPooling     2×2      2      (26, 26, 128)
Convolution    3×3      1      (26, 26, 256)
MaxPooling     2×2      2      (13, 13, 256)
Convolution    3×3      1      (13, 13, 512)
MaxPooling     2×2      1      (13, 13, 512)
Convolution    3×3      1      (13, 13, 1024)
Convolution    3×3      1      (13, 13, 1024)
Convolution    1×1      1      (13, 13, 125)
---------------------------------------------

After setting some logs, the architecture of the decoded network is:

+Convolution             (3, 3)          [1 1]           (1, 16, 416, 416)
+MaxPooling              (2, 2)          [2 2]           (1, 16, 208, 208)
+Convolution             (3, 3)          [1 1]           (1, 32, 208, 208)
+MaxPooling              (2, 2)          [2 2]           (1, 32, 104, 104)
+Convolution             (3, 3)          [1 1]           (1, 64, 104, 104)
+MaxPooling              (2, 2)          [2 2]           (1, 64, 52, 52)
+Convolution             (3, 3)          [1 1]           (1, 128, 52, 52)
+MaxPooling              (2, 2)          [2 2]           (1, 128, 26, 26)
+Convolution             (3, 3)          [1 1]           (1, 256, 26, 26)
+MaxPooling              (2, 2)          [2 2]           (1, 256, 13, 13)
+Convolution             (3, 3)          [1 1]           (1, 512, 13, 13)
-MaxPooling              (2, 2)          [1 1]           (1, 512, 14, 14)
-Convolution             (3, 3)          [1 1]           (1, 1024, 14, 14)
-Convolution             (3, 3)          [1 1]           (1, 1024, 14, 14)
-Convolution             (1, 1)          [1 1]           (1, 125, 14, 14)

The last layer using the Maxpool operator does not give the correct output size.
The padding used is computed from the auto_pad argument but seems ok (padding is [1,1]).

It requires more investigation; maybe a bug in Gorgonia.

Note : the computation is slow, but Make it work, then Make it fast

cc @chewxy

[owulveryck]

Maybe related: PR 274 from Gorgonia (from @lynic)

[chewxy]

Does yolo use overlapping pooling? i.e. maxpool with a stride of other than size

[owulveryck]

@chewxy the stride of the operator is [1 1] (the third column)

[owulveryck]

I think I got it; the ONNX spec specifies that a ceil_mode attribute can change the behavior of the shape calculus:

ceil_mode : int (default is 0)
Whether to use ceil or floor (default) to compute the output shape.

The calcShape method in Gorgonia does not take this into account; the shape calculation is probably different than what onnx expects.

I will do different tests, but If this is the case, I will have to find a smart solution to implement this without breaking Gorgonia's API 🧐

[owulveryck]

I did not notice that the implementation of the SAME_UPPER was breaking the unit tests of gorgonnx:

--- FAIL: TestONNX (0.04s)
    --- FAIL: TestONNX/TestMaxpool2dSameUpper (0.00s)
        test_structure.go:78:
                Error Trace:    test_structure.go:135
                Error:          Max difference between 1.7640524 and 0.978738 allowed is 1e-06, but difference was 0.7853143811225891
                Messages:       the two tensors should be equal.
    --- FAIL: TestONNX/TestMaxpool2dPrecomputedSameUpper (0.00s)
        test_structure.go:78:
                Error Trace:    test_structure.go:135
                Error:          Max difference between 1 and 7 allowed is 1e-06, but difference was -6
                Messages:       the two tensors should be equal.
FAIL

Let's fix this before going further.

[owulveryck]

Probably related to ONNX issue #1113

[owulveryck]

The problem linked to #74 and because the padding is asymmetric;
this is not implemented in Gorgonia.

[owulveryck]

Should be fixed with PR #80 and PR #295 from Gorgonia

[owulveryck]

Tiny yolo v2 is working, let's make it fast now!