| 23 | return layer; |
| 24 | } |
| 25 | virtual bool onComputeSize(const MNN::Op* op, const std::vector<Tensor*>& inputs, |
| 26 | const std::vector<Tensor*>& outputs) const override { |
| 27 | MNN_ASSERT(inputs.size() >= 1); |
| 28 | MNN_ASSERT(1 == outputs.size()); |
| 29 | const Convolution2DCommon* layer = loadCommon(op); |
| 30 | int kX = layer->kernelX(); |
| 31 | int kY = layer->kernelY(); |
| 32 | auto outputCount = layer->outputCount(); |
| 33 | if (inputs.size() > 1 && outputCount == 0) { |
| 34 | // From TF's multi input convolution |
| 35 | outputCount = inputs[1]->length(0); |
| 36 | kX = inputs[1]->length(3); |
| 37 | kY = inputs[1]->length(2); |
| 38 | } |
| 39 | int kernel_width = layer->dilateX() * (kX - 1) + 1; |
| 40 | int kernel_height = layer->dilateY() * (kY - 1) + 1; |
| 41 | |
| 42 | int output_width = 1; |
| 43 | int output_height = 1; |
| 44 | |
| 45 | auto input = inputs[0]; |
| 46 | if (input->dimensions() <= 1) { |
| 47 | // Convolution is not valid for dimension <= 1 |
| 48 | return false; |
| 49 | } |
| 50 | |
| 51 | auto inputCount = layer->inputCount(); |
| 52 | bool depthwiseMatch = |
| 53 | inputCount == layer->outputCount() && |
| 54 | inputCount == layer->group() && |
| 55 | inputCount == input->channel(); |
| 56 | int commonChannelMatch = |
| 57 | inputCount == inputs[0]->channel() || // real relationship in express |
| 58 | (inputCount * layer->group() == input->channel()); // standard definition of group convolution |
| 59 | bool valid = inputCount == 0 || depthwiseMatch || commonChannelMatch; |
| 60 | |
| 61 | // For Tensorflow Group Convolution, the inputCount is the size of filter's input count |
| 62 | if (inputs.size() == 1 && !valid && OpType_Convolution == op->type()) { |
| 63 | input->printShape(); |
| 64 | MNN_ERROR( |
| 65 | "Error for compute convolution shape, inputCount:%d, outputCount:%d, KH:%d, KW:%d, group:%d\ninputChannel: %d, batch:%d, width:%d, height:%d. " |
| 66 | "Input data channel may be mismatch with filter channel count\n", |
| 67 | layer->inputCount(), outputCount, kY, kX, layer->group(), |
| 68 | input->channel(), input->batch(), input->width(), input->height()); |
| 69 | return false; |
| 70 | } |
| 71 | |
| 72 | if (layer->padMode() == PadMode_SAME) { |
| 73 | // Tensorflow padding mode SAME |
| 74 | output_width = ceil((float)input->width() / (float)layer->strideX()); |
| 75 | output_height = ceil((float)input->height() / (float)layer->strideY()); |
| 76 | } else if (layer->padMode() == PadMode_VALID) { |
| 77 | // Tensorflow padding mode VALID |
| 78 | output_width = ceil((float)(input->width() - kernel_width + 1) / (float)layer->strideX()); |
| 79 | output_height = ceil((float)(input->height() - kernel_height + 1) / (float)layer->strideY()); |
| 80 | } else { |
| 81 | // Pad_Caffe means User setted padding |
| 82 | if (nullptr != layer->pads()) { |
nothing calls this directly
no test coverage detected