(op: Operation, values: List[torch.Tensor], ctx: TorchBackendContext = None, **kwargs)
| 760 | |
| 761 | |
| 762 | def AveragePool_forward(op: Operation, values: List[torch.Tensor], ctx: TorchBackendContext = None, **kwargs) -> torch.Tensor: |
| 763 | ASSERT_NUM_OF_INPUT(op=op, values=values, min_num_of_input=1, max_num_of_input=1) |
| 764 | process_attribute(op.attributes, values[0].shape[2:]) |
| 765 | |
| 766 | [x] = values |
| 767 | onnx_pads = GET_ATTRIBUTE_FROM_OPERATION(op=op, attribute='pads', default=0) |
| 768 | stride = GET_ATTRIBUTE_FROM_OPERATION(op=op, attribute='strides', default=None) |
| 769 | ceil_mode = bool(GET_ATTRIBUTE_FROM_OPERATION(op=op, attribute='ceil_mode', default=False)) |
| 770 | if op.type == 'GlobalAveragePool': kernel_size = x.size()[2:] |
| 771 | else: kernel_size = GET_ATTRIBUTE_FROM_OPERATION(op=op, attribute='kernel_shape', compulsive=True) |
| 772 | |
| 773 | ndim = x.ndim |
| 774 | # pool 1d |
| 775 | if ndim == 3: |
| 776 | torch_pads = convert_onnx_pads_to_torch(onnx_pads=onnx_pads, mode='1d') |
| 777 | if isinstance(torch_pads, list) and len(torch_pads) != 1: |
| 778 | x = F.pad(x, torch_pads) |
| 779 | torch_pads = 0 |
| 780 | output = F.avg_pool1d( |
| 781 | x, kernel_size=kernel_size, padding=torch_pads, |
| 782 | stride=stride, ceil_mode=ceil_mode) |
| 783 | return output |
| 784 | # pool 2d |
| 785 | if ndim == 4: |
| 786 | # onnx pads format[top, left, bottom, right] to torch pads format[left, right, top, bottom] |
| 787 | if isinstance(onnx_pads, list) and len(onnx_pads) == 4: |
| 788 | p_left, p_right, p_top, p_bottom = onnx_pads[1], onnx_pads[3], onnx_pads[0], onnx_pads[2] |
| 789 | # torch does not support padding contains 4 value, there is a fix of it. |
| 790 | if p_left == p_right and p_top == p_bottom: |
| 791 | onnx_pads = [p_top, p_left] |
| 792 | else: |
| 793 | x = F.pad(x, pad=[p_left, p_right, p_top, p_bottom]) |
| 794 | onnx_pads = 0 |
| 795 | |
| 796 | output = F.avg_pool2d( |
| 797 | x, kernel_size=kernel_size, |
| 798 | padding=onnx_pads, stride=stride, ceil_mode=ceil_mode) |
| 799 | return output |
| 800 | # pool 3d |
| 801 | elif ndim == 5: |
| 802 | torch_pads = convert_onnx_pads_to_torch(onnx_pads=onnx_pads, mode='3d') |
| 803 | if isinstance(torch_pads, list) and len(torch_pads) != 3: |
| 804 | x = F.pad(x, torch_pads) |
| 805 | torch_pads = 0 |
| 806 | output = F.avg_pool3d( |
| 807 | x, kernel_size=kernel_size, padding=torch_pads, |
| 808 | stride=stride, ceil_mode=ceil_mode) |
| 809 | return output |
| 810 | else: |
| 811 | raise ValueError(f'Operation {op.name} is invalid, {ndim}-d input is not supported.') |
| 812 | |
| 813 | return output |
| 814 | |
| 815 | |
| 816 | def AdaptiveAvgPool2d_forward(op: Operation, values: List[torch.Tensor], ctx: TorchBackendContext = None, **kwargs) -> torch.Tensor: |
nothing calls this directly
no test coverage detected