Error in TransposedConv2d

[Shashi456]

Refer to comment here.

Even when trying to write test cases for #174 and the transposed conv2d layer, I found a similar error and i reported it there. I think we need to add preconditions and check the layer logic thoroughly.

[Shashi456]

The error in question :

Trying to use the TransposedConv2D layer but getting strange error. Below is a small example that when I run either in Colab or locally with a recent build, I get the following error. Any ideas??

[1, 4, 4, 256]
0.69032013
Fatal error: Conv2DCustomBackpropFilter: filter and out_backprop must have the same out_depth: file /swift-base/swift/stdlib/public/TensorFlow/CompilerRuntime.swift, line 2123
Current stack trace:
0 libswiftCore.so 0x00007f06347314a0 swift_stdlib_reportFatalErrorInFile + 115
1 libswiftCore.so 0x00007f063467930c + 3035916
2 libswiftCore.so 0x00007f06346793fe + 3036158
3 libswiftCore.so 0x00007f06344c06c2 + 1230530
4 libswiftCore.so 0x00007f0634646292 + 2826898
5 libswiftCore.so 0x00007f06344bfba9 + 1227689
6 libswiftTensorFlow.so 0x00007f0631aae572 + 599410
7 libswiftTensorFlow.so 0x00007f0631aaccc0 checkOk(:file:line:) + 508
8 libswiftTensorFlow.so 0x00007f0631ad1ad0 TFCCheckOk(:) + 81
9 libswiftTensorFlow.so 0x00007f0631ad1ac0 _swift_tfc_CheckOk + 9
Current stack trace:
frame #12: 0x00007f0661d90ac9 $__lldb_expr76AD__$s14__lldb_expr_7510TensorFlow0C0VySfGAB16TransposedConv2DVXEfU___adjoint_src_0_wrt_0 at <Cell 14>:14:15 frame #20: 0x00007f0661d8e9a0 $__lldb_expr76main at <Cell 14>:13:38
frame #33: 0x00007f0661d98be7 $lldb_expr72AD__$s14__lldb_expr_7110TensorFlow0C0VySfGAB16TransposedConv2DVXEfU___primal_src_0_wrt_0(layer=<unavailable>) at <Cell 13>:14:15 frame #34: 0x00007f0661d99b84 $__lldb_expr72AD$s14__lldb_expr_7110TensorFlow0C0VySfGAB16TransposedConv2DVXEfU___vjp_src_0_wrt_0 at <Cell 13>:13:56
frame #42: 0x00007f0661d979a0 $lldb_expr72main at <Cell 13>:13:38 frame #55: 0x00007f0661da2be7 $__lldb_expr68AD$s14__lldb_expr_6710TensorFlow0C0VySfGAB16TransposedConv2DVXEfU___primal_src_0_wrt_0(layer=) at <Cell 12>:14:15
frame #56: 0x00007f0661da3b84 $__lldb_expr68AD__$s14__lldb_expr_6710TensorFlow0C0VySfGAB16TransposedConv2DVXEfU___vjp_src_0_wrt_0 at <Cell 12>:13:56 frame #64: 0x00007f0661da19a0 $__lldb_expr68main at <Cell 12>:13:38

import TensorFlow

let tensorShape = TensorShape([4, 4, 256, 64])
let seed = (Int64.random(in: Int64.min..<Int64.max), Int64.random(in: Int64.min..<Int64.max))
let transposedConv2D = TransposedConv2D(
filter: Tensor(glorotUniform: tensorShape, seed: seed),
bias: Tensor(zeros: TensorShape([128])),
activation: identity,
strides: (1, 1),
padding: .valid)

let tensor = Tensor(randomUniform: TensorShape([1,1,1,64]))
let (loss, grads) = transposedConv2D.valueWithGradient { layer -> Tensor in
let out = layer(tensor)
print(out.shape)
let loss = out.sum()
print(loss)
return loss
}

[sjaz24]

@Shashi456 do you think it has to do with this. Shouldn't channels (c) be filter.shape[3]. I'm looking at the original merged code, so not sure???

@differentiable
    public func applied(to input: Tensor<Float>, in _: Context) -> Tensor<Float> {
        let batchSize = input.shape[0]
        let w = (input.shape[1] - (1 * paddingIndex)) * strides.0 + (filter.shape[0] * paddingIndex)
        let h = (input.shape[2] - (1 * paddingIndex)) * strides.1 + (filter.shape[1] * paddingIndex)
        let c = filter.shape[2]
        let newShape = Tensor<Int32>([batchSize, w, h, c])
        return activation(input.conv2DBackpropInput(shape: newShape, filter: filter,
                                                         strides: (1, strides.0, strides.1, 1),
                                                         padding: padding) + bias)
    }

[sjaz24]

Note this other issue with this layer. https://bugs.swift.org/browse/TF-540

[Shashi456]

The solution also seems to be here, https://groups.google.com/a/tensorflow.org/forum/m/#!msg/swift/UUPwV01sZrE/LszG6T7dBQAJ

@sjaz24, would you be up to fix the error and write a test for the layer? I think it's the perfect opportunity for a first contribution to the repo :) , if not I'll put a PR up later today.

[sjaz24]

I'll give it a go at fixing it.

[Shashi456]

@sjaz24 maybe you could take a look at the tests I wrote in #174, the filter dimensions and other dimensions were inspired by running a transposed layer in tensor flow, I ran into the filter depth and out backprop error there as well and just played around with the dimensions until they worked. Maybe that could be another starting point.

[Shashi456]

@rxwei I had some doubts, when i'm trying to solve this, Should i open a PR and we could continue the conversation and discuss the solution there?

[eaplatanios]

@Shashi456 Feel free to open a PR and continue the conversation there.