File size: 5,326 Bytes
be903e2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 | // Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2021 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.
#include "pass_level1.h"
#include "../utils.h"
namespace pnnx {
class LSTM : public FuseModulePass
{
public:
const char* match_type_str() const
{
return "__torch__.torch.nn.modules.rnn.LSTM";
}
const char* type_str() const
{
return "nn.LSTM";
}
void write(Operator* op, const std::shared_ptr<torch::jit::Graph>& graph, const torch::jit::Module& mod) const
{
// mod.dump(true, true, true);
//
// graph->dump();
const torch::jit::Node* lstm = find_node_by_kind(graph, "aten::lstm");
const torch::jit::Node* return_tuple = find_node_by_kind(graph, "prim::TupleConstruct");
if (return_tuple && return_tuple->inputs().size() == 3 && lstm->outputs().size() == 3
&& return_tuple->inputs()[0] == lstm->outputs()[1] && return_tuple->inputs()[1] == lstm->outputs()[2] && return_tuple->inputs()[2] == lstm->outputs()[0])
{
// mark the swapped output tuple
// we would restore the fine order in pass_level3/fuse_rnn_unpack
fprintf(stderr, "swapped detected !\n");
op->params["pnnx_rnn_output_swapped"] = 1;
}
// for (auto aa : lstm->schema().arguments())
// {
// fprintf(stderr, "arg %s\n", aa.name().c_str());
// }
const auto& weight_ih_l0 = mod.attr("weight_ih_l0").toTensor();
const auto& weight_hh_l0 = mod.attr("weight_hh_l0").toTensor();
op->params["input_size"] = weight_ih_l0.size(1);
op->params["hidden_size"] = weight_ih_l0.size(0) / 4;
op->params["num_layers"] = lstm->namedInput("num_layers");
op->params["bias"] = lstm->namedInput("has_biases");
op->params["batch_first"] = lstm->namedInput("batch_first");
op->params["bidirectional"] = lstm->namedInput("bidirectional");
op->params["proj_size"] = weight_ih_l0.size(0) / 4 == weight_hh_l0.size(1) ? 0 : weight_hh_l0.size(1);
const int num_layers = op->params["num_layers"].i;
const bool bias = op->params["bias"].b;
const bool bidirectional = op->params["bidirectional"].b;
const int proj_size = op->params["proj_size"].i;
for (int k = 0; k < num_layers; k++)
{
std::string weight_ih_lk_key = std::string("weight_ih_l") + std::to_string(k);
std::string weight_hh_lk_key = std::string("weight_hh_l") + std::to_string(k);
op->attrs[weight_ih_lk_key] = mod.attr(weight_ih_lk_key).toTensor();
op->attrs[weight_hh_lk_key] = mod.attr(weight_hh_lk_key).toTensor();
if (bias)
{
std::string bias_ih_lk_key = std::string("bias_ih_l") + std::to_string(k);
std::string bias_hh_lk_key = std::string("bias_hh_l") + std::to_string(k);
op->attrs[bias_ih_lk_key] = mod.attr(bias_ih_lk_key).toTensor();
op->attrs[bias_hh_lk_key] = mod.attr(bias_hh_lk_key).toTensor();
}
if (proj_size > 0)
{
std::string weight_hr_lk_key = std::string("weight_hr_l") + std::to_string(k);
op->attrs[weight_hr_lk_key] = mod.attr(weight_hr_lk_key).toTensor();
}
if (bidirectional)
{
std::string weight_ih_lk_reverse_key = std::string("weight_ih_l") + std::to_string(k) + "_reverse";
std::string weight_hh_lk_reverse_key = std::string("weight_hh_l") + std::to_string(k) + "_reverse";
op->attrs[weight_ih_lk_reverse_key] = mod.attr(weight_ih_lk_reverse_key).toTensor();
op->attrs[weight_hh_lk_reverse_key] = mod.attr(weight_hh_lk_reverse_key).toTensor();
if (bias)
{
std::string bias_ih_lk_reverse_key = std::string("bias_ih_l") + std::to_string(k) + "_reverse";
std::string bias_hh_lk_reverse_key = std::string("bias_hh_l") + std::to_string(k) + "_reverse";
op->attrs[bias_ih_lk_reverse_key] = mod.attr(bias_ih_lk_reverse_key).toTensor();
op->attrs[bias_hh_lk_reverse_key] = mod.attr(bias_hh_lk_reverse_key).toTensor();
}
if (proj_size > 0)
{
std::string weight_hr_lk_reverse_key = std::string("weight_hr_l") + std::to_string(k) + "_reverse";
op->attrs[weight_hr_lk_reverse_key] = mod.attr(weight_hr_lk_reverse_key).toTensor();
}
}
}
}
};
REGISTER_GLOBAL_PNNX_FUSE_MODULE_PASS(LSTM)
} // namespace pnnx
|