padding problem about onnx2tflite HOT 7 CLOSED

That's a good question.
Because tensorflow convolution padding is inconsistent with pytorch, tensorflow 'SAME' padding may pad more pixels on the right than left.
You can see https://stackoverflow.com/questions/52975843/comparing-conv2d-with-padding-between-tensorflow-and-pytorch.
If you get a good to fix this problem, welcome PR.

from onnx2tflite.

thanks, may be solved with some trick.

from onnx2tflite.

Hi
I do some tests.
because network has many 3x3 conv with same padding.
I change the code with below code on mobileone. it can reduce many pad and get the same result, reduce the peak of mobile running time. so we can use this trick^^.

  if pads is not None and max(pads) == 1 and max(strides) == 1:
        self.conv = keras.layers.DepthwiseConv2D(
            kernel_size, strides, "SAME", use_bias=False if bias is None else True,
            weights=[weights] if bias is None else [weights, bias],
            dilation_rate=dilations,
            activation=None,
            kernel_initializer='zeros',
            bias_initializer='zeros'
        )
        self.pad =None

from onnx2tflite.

Hi I do some tests. because network has many 3x3 conv with same padding. I change the code with below code on mobileone. it can reduce many pad and get the same result, reduce the peak of mobile running time. so we can use this trick^^.

  if pads is not None and max(pads) == 1 and max(strides) == 1:
        self.conv = keras.layers.DepthwiseConv2D(
            kernel_size, strides, "SAME", use_bias=False if bias is None else True,
            weights=[weights] if bias is None else [weights, bias],
            dilation_rate=dilations,
            activation=None,
            kernel_initializer='zeros',
            bias_initializer='zeros'
        )
        self.pad =None

Nice idea, thanks for your test.
Can you evaluate it on other types of convolution, such as keras.layers.Conv2D and keras.layers.Conv2DTranspose.
It's pleasure for your PR.

from onnx2tflite.

DepthwiseConv2D should be Consistent with Conv2D. As to Conv2DTranspose, it need to be tested.

from onnx2tflite.

Thanks for your nice work again, I will test it as soon~

from onnx2tflite.

I was tested it on this model.
Your trick is work well for conv\depthwise conv\group conv, but not with ConvTranspose, thanks for your contribution again.

import torch
import torch.nn as nn

class testmodel(nn.Module):
    def __init__(self) -> None:
        super(testmodel, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 3, 2, 1)
        self.conv2 = nn.Conv2d(3, 6, 3, 1, 1)
        self.depthwiseConv1 = nn.Conv2d(3, 6, 3, 2, 1, groups=3)
        self.depthwiseConv2 = nn.Conv2d(3, 6, 3, 1, 1, groups=3)
        self.convTranspose1 = nn.ConvTranspose2d(6, 3, 2, 2, 0)
        self.convTranspose2 = nn.ConvTranspose2d(6, 30, 2, 2, 0)
        self.group_conv1 = nn.Conv2d(30, 30, 3, 1, 1, groups=30)

    def forward(self, X):
        x1 = self.conv1(X)
        x2 = self.conv2(X)

        x3 = self.depthwiseConv1(X)
        x4 = self.depthwiseConv2(X)
        
        x5 = self.convTranspose1(x1)
        x6 = self.convTranspose1(x3)

        out = x2+x4+torch.cat((x5, x6), dim=1)
        out = self.convTranspose2(out)
        out = self.group_conv1(out)
        return out

model = testmodel()
X = torch.randn(1, 3, 224, 224)
torch.onnx.export(model, X, './models/opt_test.onnx', opset_version=12)

from onnx2tflite.