Problem with GlobalAveragePool operator inference

Required prerequisites

• Make sure you've read the documentation. Your issue may be addressed there.
• Search the issue tracker and discussions to verify that this hasn't already been reported. +1 or comment there if it has.

What commit version of aidge do you use

• aidge : 0.5.0
• aidge_backend_cpu : 0.5.0
• aidge_backend_cuda: 0.5.0
• aidge_core : 0.5.1
• aidge_export_cpp : 0.2.1
• aidge_onnx : 0.4.1
• aidge_quantization: 0.3.0

Problem description

The AvgPool2D and GlobalAveragePooling operators can be considered equivalent when in the case of an AvgPool2D operator, the kernel size is equal to the map size. During an annodin test on a 7x7 map, it turned out that the GlobalAveragePooling does not give the right result while the AvgPool2D does give the right result (see the following python code). The value calculated by hand gives 26.

Input map:
  {{   0,   8,  26,  35,  49,  45,  22},
   {   2,  24,  48,  66,  60,  46,  26},
   {   8,  41,  64,  68,  39,  18,   9},
   {  10,  48,  72,  76,  42,  14,   9},
   {   6,  29,  52,  65,  27,   7,   3},
   {   1,   9,  24,  31,  18,   7,   1},
   {   0,   0,   4,   6,   7,   1,   1}}

GlobalAveragePooling result:
   {{ 25}}

AvgPoolingD2 result:
   {{ 26}}

Reproducible example code

Python code for AvgPool2D operator

import aidge_core
import aidge_backend_cpu
import numpy as np

# Define a list of tensors
tensors = [
    np.array([[[[0, 8, 26, 35, 49, 45, 22],
                [2, 24, 48, 66, 60, 46, 26],
                [8, 41, 64, 68, 39, 18, 9],
                [10, 48, 72, 76, 42, 14, 9],
                [6, 29, 52, 65, 27, 7, 3],
                [1, 9, 24, 31, 18, 7, 1],
                [0, 0, 4, 6, 7, 1, 1]]]]),
]

# Loop through each tensor and perform the same operations
for tensor_array in tensors:
    tensor = aidge_core.Tensor(tensor_array)
    tensor.set_datatype(aidge_core.dtype.int32)

    op = aidge_core.AvgPooling2DOp([7, 7])
    print(op)
    op.set_datatype(aidge_core.dtype.int32)
    op.set_backend("cpu")

    op.associate_input(0, tensor)
    op.forward_dims()
    print(tensor.dims())
    print(tensor)
    op.forward()
    print(op.get_output(0))

Python code for GlobalAvgPool2D operator

import aidge_core
import aidge_backend_cpu
import numpy as np

# Define a list of tensors
tensors = [
    np.array([[[[0, 8, 26, 35, 49, 45, 22],
                [2, 24, 48, 66, 60, 46, 26],
                [8, 41, 64, 68, 39, 18, 9],
                [10, 48, 72, 76, 42, 14, 9],
                [6, 29, 52, 65, 27, 7, 3],
                [1, 9, 24, 31, 18, 7, 1],
                [0, 0, 4, 6, 7, 1, 1]]]]),
]

# Loop through each tensor and perform the same operations
for tensor_array in tensors:
    tensor = aidge_core.Tensor(tensor_array)
    tensor.set_datatype(aidge_core.dtype.int32)

    node = aidge_core.globalaveragepooling()
    op = node.get_operator()
    print(op)
    op.set_datatype(aidge_core.dtype.int32)
    op.set_backend("cpu")

    op.associate_input(0, tensor)
    op.forward_dims()
    print(tensor.dims())
    print(tensor)
    op.forward()
    print(op.get_output(0))

assigned to @mick94 and @axelfarr

changed the description

assigned to @axelfarr and @mick94

added Bug 🐛 label

moved from aidge_onnx#55 (moved)

changed milestone to %aidge v0.6.0

added TopicOperator label

added StatusTODO label

This has been fixed in !133 (merged), but is not yet merged in dev or main, but pending in !135 (merged).

assigned to @olivierbichler

@mick94 Everything is now merged: it should work!

closed

It solved the given example, but there still are cases with differences between the results given by the GlobalAvgPool and the AvgPool (I updated the example).

Yes, computation still wrong for integers: the returned type of stableMean was wrong. Fixed in !149 (merged).

changed the description

reopened

marked the checklist item Search the issue tracker and discussions to verify that this hasn't already been reported. +1 or comment there if it has. as incomplete

marked the checklist item Search the issue tracker and discussions to verify that this hasn't already been reported. +1 or comment there if it has. as completed

added StatusWork in Progress label and removed StatusTODO label

changed the description

mentioned in merge request !149 (merged)