MNN/source/shape/ShapeRNNSequenceGRU.cpp

//
//  ShapeRNNSequenceGRU.cpp
//  MNN
//
//  Created by MNN on 2019/03/19.
//  Copyright © 2018, Alibaba Group Holding Limited
//

#include "shape/SizeComputer.hpp"
#include "core/TensorUtils.hpp"
namespace MNN {

class RNNSequenceGRUComputer : public SizeComputer {
public:
    virtual bool onComputeSize(const MNN::Op* op, const std::vector<Tensor*>& inputs,
                               const std::vector<Tensor*>& outputs) const override {

        auto outputSize = outputs.size();
        MNN_ASSERT(6 <= inputs.size());
        MNN_ASSERT(1 <= outputSize);
        auto input  = inputs[0]; // typically onnx input shape: {sequenceLength, batchSize, inputLength}
        auto output = outputs[0];

        const auto rnnParam     = op->main_as_RNNParam();
        const int numUnits      = rnnParam->numUnits();
        bool keepAllOutputs     = rnnParam->keepAllOutputs();
        bool isBidirectionalRNN = rnnParam->isBidirectionalRNN();

        // input->printShape();

        // MNN_ASSERT(2 == rnnParam->fwGateWeight()->dims()->size());
        // MNN_ASSERT(2 * numUnits == rnnParam->fwGateWeight()->dims()->data()[1]);
        // MNN_ASSERT((input->length(2) + numUnits) == rnnParam->fwGateWeight()->dims()->data()[0]);
        output->buffer().type = halide_type_of<float>();
        TensorUtils::getDescribe(output)->dimensionFormat = TensorUtils::getDescribe(input)->dimensionFormat;



        if (keepAllOutputs) {
            TensorUtils::setShape(output, {input->length(0), isBidirectionalRNN + 1, input->length(1), numUnits});
            // output shape: {sequenceLength, numDirection, batchSize, inputLength}
            // !!caution: onnx model graph some time would squeeze the ‘1 dim’ in output 'numDirection', we should keep numDirection index at 1,
            // but, the typical memory layout of input tensor in CPURNNSequenceGRU.cpp is {batch, sequenceLength, inputLength},
            // there is mismatch here when batch or sequence is not 1
            output->buffer().type = input->buffer().type;
            if (outputSize > 1) {
                auto YHOutput = outputs[1];
                TensorUtils::setShape(YHOutput, {1, isBidirectionalRNN + 1, input->length(1), numUnits});
                YHOutput->buffer().type = input->buffer().type;
                TensorUtils::getDescribe(YHOutput)->dimensionFormat = TensorUtils::getDescribe(input)->dimensionFormat;
            }
        } else { // only keep the last hidden layer sequence
            TensorUtils::setShape(output, {1, isBidirectionalRNN + 1, input->length(1), numUnits});
            output->buffer().type = input->buffer().type;
        }

        // output->printShape();

        return true;
    }
};

REGISTER_SHAPE(RNNSequenceGRUComputer, OpType_RNNSequenceGRU);
} // namespace MNN