lstm_op.h 14.8 KB
Newer Older
D
dangqingqing 已提交
1 2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

D
dangqingqing 已提交
3 4 5
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
D
dangqingqing 已提交
6

D
dangqingqing 已提交
7
http://www.apache.org/licenses/LICENSE-2.0
D
dangqingqing 已提交
8

D
dangqingqing 已提交
9 10 11 12 13
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. */
D
dangqingqing 已提交
14 15 16

#pragma once
#include "paddle/framework/op_registry.h"
17 18 19
#include "paddle/operators/math/lstm_compute.h"
#include "paddle/operators/math/math_function.h"
#include "paddle/operators/math/sequence2batch.h"
D
dangqingqing 已提交
20 21 22 23

namespace paddle {
namespace operators {

D
dangqingqing 已提交
24 25 26
using LoDTensor = framework::LoDTensor;
using Tensor = framework::Tensor;

Q
QI JUN 已提交
27 28
template <typename DeviceContext, typename T>
inline void ReorderInitState(const DeviceContext& ctx,
D
dangqingqing 已提交
29 30
                             const framework::Tensor& src, const size_t* index,
                             framework::Tensor* dst, bool indexed_src) {
Q
QI JUN 已提交
31
  math::CopyMatrixRowsFunctor<DeviceContext, T> row_shuffle;
D
dangqingqing 已提交
32 33 34 35
  dst->mutable_data<T>(src.dims(), ctx.GetPlace());
  row_shuffle(ctx, src, index, *dst, indexed_src);
}

Q
QI JUN 已提交
36
template <typename DeviceContext, typename T>
D
dangqingqing 已提交
37 38
class LSTMKernel : public framework::OpKernel<T> {
 public:
D
dangqingqing 已提交
39
  void Compute(const framework::ExecutionContext& ctx) const override {
D
dangqingqing 已提交
40 41 42
    auto* input = ctx.Input<LoDTensor>("Input");
    auto* weight = ctx.Input<Tensor>("Weight");
    auto* bias = ctx.Input<Tensor>("Bias");
43

44 45 46
    auto* hidden_t0 = ctx.Input<Tensor>("H0");
    auto* cell_t0 = ctx.Input<Tensor>("C0");

D
dangqingqing 已提交
47
    auto* batch_gate = ctx.Output<LoDTensor>("BatchGate");
48
    batch_gate->mutable_data<T>(ctx.GetPlace());
D
dangqingqing 已提交
49
    auto* hidden_out = ctx.Output<LoDTensor>("Hidden");
50
    hidden_out->mutable_data<T>(ctx.GetPlace());
D
dangqingqing 已提交
51
    auto* cell_out = ctx.Output<LoDTensor>("Cell");
52 53
    cell_out->mutable_data<T>(ctx.GetPlace());

54
    bool is_reverse = ctx.Attr<bool>("is_reverse");
Q
QI JUN 已提交
55 56
    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
    auto& device_ctx = ctx.template device_context<DeviceContext>();
D
dangqingqing 已提交
57
    to_batch(device_ctx, *input, *batch_gate, true, is_reverse);
58 59

    auto in_dims = input->dims();
Y
Yu Yang 已提交
60
    int frame_size = static_cast<int>(in_dims[1] / 4);
61
    framework::DDim dims({in_dims[0], frame_size});
D
dangqingqing 已提交
62

63
    if (bias) {
64 65 66
      Tensor b = *bias;
      b.Resize({bias->numel(), 1});
      Tensor gate_bias = b.Slice(0, 4 * frame_size);
Q
QI JUN 已提交
67
      math::RowwiseAdd<DeviceContext, T> add_bias;
68
      add_bias(device_ctx, *batch_gate, gate_bias, batch_gate);
69 70 71
    }

    math::LstmMetaValue<T> lstm_value;
D
dangqingqing 已提交
72
    if (bias && ctx.Attr<bool>("use_peepholes")) {
D
dangqingqing 已提交
73 74
      T* bias_data = const_cast<T*>(bias->data<T>());
      // the code style in LstmMetaValue will be updated later.
75

76 77 78
      lstm_value.check_ig = bias_data + 4 * frame_size;
      lstm_value.check_fg = lstm_value.check_ig + frame_size;
      lstm_value.check_og = lstm_value.check_fg + frame_size;
D
dangqingqing 已提交
79
    } else {
80 81 82
      lstm_value.check_ig = nullptr;
      lstm_value.check_fg = nullptr;
      lstm_value.check_og = nullptr;
D
dangqingqing 已提交
83
    }
84
    lstm_value.prev_state_value = nullptr;
85
    Tensor ordered_c0;
D
dangqingqing 已提交
86
    const size_t* order = batch_gate->lod()[2].data();
87
    if (cell_t0) {
D
dangqingqing 已提交
88 89 90
      // Since the batch computing for LSTM reorders the input sequence
      // according to their length. The initialized cell state also needs
      // to reorder.
Q
QI JUN 已提交
91 92
      ReorderInitState<DeviceContext, T>(device_ctx, *cell_t0, order,
                                         &ordered_c0, true);
93
      lstm_value.prev_state_value = ordered_c0.data<T>();
94
    }
95

D
dangqingqing 已提交
96 97
    // Use the local variable as here.
    LoDTensor batch_hidden, batch_cell;
98
    auto* batch_cell_pre_act = ctx.Output<LoDTensor>("BatchCellPreAct");
D
dangqingqing 已提交
99
    batch_hidden.mutable_data<T>(dims, ctx.GetPlace());
100
    batch_cell.mutable_data<T>(dims, ctx.GetPlace());
101
    batch_cell_pre_act->mutable_data<T>(dims, ctx.GetPlace());
102

D
dangqingqing 已提交
103
    auto batch_starts = batch_gate->lod()[0];
Y
Yu Yang 已提交
104
    size_t num_batch = batch_starts.size() - 1;
105 106 107
    auto gate_act = ctx.Attr<std::string>("gate_activation");
    auto cell_act = ctx.Attr<std::string>("cell_activation");
    auto cand_act = ctx.Attr<std::string>("candidate_activation");
108

Y
Yu Yang 已提交
109 110 111
    for (size_t n = 0; n < num_batch; n++) {
      int bstart = static_cast<int>(batch_starts[n]);
      int bend = static_cast<int>(batch_starts[n + 1]);
112

D
dangqingqing 已提交
113
      Tensor gate_t = batch_gate->Slice(bstart, bend);
D
dangqingqing 已提交
114
      Tensor out_t = batch_hidden.Slice(bstart, bend);
D
dangqingqing 已提交
115
      Tensor cell_t = batch_cell.Slice(bstart, bend);
116
      Tensor cell_pre_act_t = batch_cell_pre_act->Slice(bstart, bend);
117 118 119

      int cur_batch_size = bend - bstart;

120
      if (n > 0) {
Y
Yu Yang 已提交
121
        int pre_h_start = static_cast<int>(batch_starts[n - 1]);
D
dangqingqing 已提交
122
        int pre_h_end = pre_h_start + cur_batch_size;
D
dangqingqing 已提交
123
        auto pre_hidden_t = batch_hidden.Slice(pre_h_start, pre_h_end);
Q
QI JUN 已提交
124 125 126
        math::matmul<DeviceContext, T>(device_ctx, pre_hidden_t, false, *weight,
                                       false, static_cast<T>(1.0), &gate_t,
                                       static_cast<T>(1.0));
127
      } else if (hidden_t0) {
D
dangqingqing 已提交
128 129 130 131 132 133 134
        // If n == 0 and there is no initialized hidden state, that is to say
        // the H0 is zeros, the calculation W_h * H0 will be skiped.
        // If n == 0 and there is initialized hidden state, calculate W_h * H0.

        // Since the batch computing for LSTM reorders the input sequence
        // according to their length. The initialized hidden state also needs
        // to reorder.
135
        Tensor ordered_h0;
Q
QI JUN 已提交
136 137 138 139 140
        ReorderInitState<DeviceContext, T>(device_ctx, *hidden_t0, order,
                                           &ordered_h0, true);
        math::matmul<DeviceContext, T>(device_ctx, ordered_h0, false, *weight,
                                       false, static_cast<T>(1.0), &gate_t,
                                       static_cast<T>(1.0));
141 142
      }

143 144 145 146
      lstm_value.gate_value = gate_t.data<T>();
      lstm_value.output_value = out_t.data<T>();
      lstm_value.state_value = cell_t.data<T>();
      lstm_value.state_active_value = cell_pre_act_t.data<T>();
Q
QI JUN 已提交
147 148 149
      math::LstmUnitFunctor<DeviceContext, T>::compute(
          device_ctx, lstm_value, frame_size, cur_batch_size, gate_act,
          cell_act, cand_act);
150
      lstm_value.prev_state_value = lstm_value.state_value;
D
dangqingqing 已提交
151
    }
152

Q
QI JUN 已提交
153
    math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
D
dangqingqing 已提交
154
    batch_hidden.set_lod(batch_gate->lod());
155
    // restore the output hidden in LoDTensor from the batch hidden
D
dangqingqing 已提交
156
    to_seq(device_ctx, batch_hidden, *hidden_out);
157

158
    batch_cell.set_lod(batch_gate->lod());
159
    // restore the output cell state in LoDTensor from the batch cell
D
dangqingqing 已提交
160
    to_seq(device_ctx, batch_cell, *cell_out);
D
dangqingqing 已提交
161
  }
D
dangqingqing 已提交
162 163
};

Q
QI JUN 已提交
164
template <typename DeviceContext, typename T>
D
dangqingqing 已提交
165 166
class LSTMGradKernel : public framework::OpKernel<T> {
 public:
D
dangqingqing 已提交
167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* input = ctx.Input<LoDTensor>("Input");
    auto* weight = ctx.Input<Tensor>("Weight");
    auto* bias = ctx.Input<Tensor>("Bias");

    auto* hidden_out = ctx.Input<LoDTensor>("Hidden");
    auto* cell_out = ctx.Input<LoDTensor>("Cell");

    auto* batch_gate = ctx.Input<LoDTensor>("BatchGate");
    auto* batch_cell_pre_act = ctx.Input<LoDTensor>("BatchCellPreAct");

    auto* hidden_g = ctx.Input<LoDTensor>(framework::GradVarName("Hidden"));

    auto* in_g = ctx.Output<LoDTensor>(framework::GradVarName("Input"));
    auto* weight_g = ctx.Output<Tensor>(framework::GradVarName("Weight"));
    auto* bias_g = ctx.Output<Tensor>(framework::GradVarName("Bias"));

184 185 186 187 188 189
    auto* h0 = ctx.Input<Tensor>("H0");
    auto* c0 = ctx.Input<Tensor>("C0");

    auto* h0_g = ctx.Output<Tensor>(framework::GradVarName("H0"));
    auto* c0_g = ctx.Output<Tensor>(framework::GradVarName("C0"));

Q
QI JUN 已提交
190 191
    auto& device_ctx = ctx.template device_context<DeviceContext>();
    math::SetConstant<DeviceContext, T> zero;
D
dangqingqing 已提交
192
    if (weight_g) {
193
      weight_g->mutable_data<T>(ctx.GetPlace());
D
dangqingqing 已提交
194 195 196
      zero(device_ctx, weight_g, static_cast<T>(0.0));
    }

D
dangqingqing 已提交
197 198 199
    // ordered_h0/c0 is the reordered hidden/cell initialization.
    // ordered_h0_g/c0_g is the reordered gradient of hidden/cell
    // initialization.
200 201 202
    Tensor ordered_h0, ordered_c0, ordered_h0_g, ordered_c0_g;
    const size_t* order = batch_gate->lod()[2].data();
    if (c0) {
Q
QI JUN 已提交
203 204
      ReorderInitState<DeviceContext, T>(device_ctx, *c0, order, &ordered_c0,
                                         true);
D
dangqingqing 已提交
205 206 207
    }
    if (c0 && c0_g) {
      ordered_c0_g.mutable_data<T>(c0_g->dims(), ctx.GetPlace());
208 209
    }

D
dangqingqing 已提交
210 211 212 213 214 215
    auto in_dims = input->dims();
    auto out_dims = hidden_g->dims();
    int frame_size = static_cast<int>(in_dims[1] / 4);
    PADDLE_ENFORCE_EQ(frame_size, out_dims[1]);

    math::LstmMetaValue<T> lstm_value;
D
dangqingqing 已提交
216
    if (bias && ctx.Attr<bool>("use_peepholes")) {
D
dangqingqing 已提交
217
      T* bias_data = const_cast<T*>(bias->data<T>());
218 219 220
      lstm_value.check_ig = bias_data + 4 * frame_size;
      lstm_value.check_fg = lstm_value.check_ig + frame_size;
      lstm_value.check_og = lstm_value.check_fg + frame_size;
D
dangqingqing 已提交
221
    } else {
222 223 224
      lstm_value.check_ig = nullptr;
      lstm_value.check_fg = nullptr;
      lstm_value.check_og = nullptr;
D
dangqingqing 已提交
225 226 227
    }

    math::LstmMetaGrad<T> lstm_grad;
D
dangqingqing 已提交
228

D
dangqingqing 已提交
229
    if (bias && bias_g) {
D
dangqingqing 已提交
230
      bias_g->mutable_data<T>(ctx.GetPlace());
231
      zero(device_ctx, bias_g, static_cast<T>(0.0));
D
dangqingqing 已提交
232 233 234
    }
    if (bias && bias_g && ctx.Attr<bool>("use_peepholes")) {
      T* bias_g_data = bias_g->data<T>();
235 236 237
      lstm_grad.check_ig_grad = bias_g_data + 4 * frame_size;
      lstm_grad.check_fg_grad = lstm_grad.check_ig_grad + frame_size;
      lstm_grad.check_og_grad = lstm_grad.check_fg_grad + frame_size;
D
dangqingqing 已提交
238
    } else {
239 240 241
      lstm_grad.check_ig_grad = nullptr;
      lstm_grad.check_fg_grad = nullptr;
      lstm_grad.check_og_grad = nullptr;
D
dangqingqing 已提交
242 243
    }

Q
QI JUN 已提交
244
    math::LoDTensor2BatchFunctor<DeviceContext, T> to_batch;
D
dangqingqing 已提交
245

D
dangqingqing 已提交
246
    auto ToBatch = [&batch_gate, &to_batch](
Q
QI JUN 已提交
247
        const DeviceContext& ctx, const framework::LoDTensor& src,
D
dangqingqing 已提交
248 249 250 251 252
        const framework::DDim& dims, framework::LoDTensor& dst) {
      dst.mutable_data<T>(dims, ctx.GetPlace());
      dst.set_lod(batch_gate->lod());
      to_batch(ctx, src, dst, false);
    };
D
dangqingqing 已提交
253

D
dangqingqing 已提交
254 255 256 257
    LoDTensor batch_hidden, batch_hidden_g, batch_cell;
    ToBatch(device_ctx, *hidden_out, out_dims, batch_hidden);
    ToBatch(device_ctx, *hidden_g, out_dims, batch_hidden_g);
    ToBatch(device_ctx, *cell_out, out_dims, batch_cell);
D
dangqingqing 已提交
258

D
dangqingqing 已提交
259
    LoDTensor batch_cell_g, batch_gate_g;
D
dangqingqing 已提交
260
    batch_cell_g.mutable_data<T>(out_dims, ctx.GetPlace());
261
    // TODO(qingqing) support the case output cell has gradient.
D
dangqingqing 已提交
262
    // to_batch(device_ctx, *cell_g, batch_cell_g, false);
263
    zero(device_ctx, &batch_cell_g, static_cast<T>(0.0));
D
dangqingqing 已提交
264 265 266
    batch_gate_g.mutable_data<T>(batch_gate->dims(), ctx.GetPlace());
    batch_gate_g.set_lod(batch_gate->lod());

267 268 269
    auto gate_act = ctx.Attr<std::string>("gate_activation");
    auto cell_act = ctx.Attr<std::string>("cell_activation");
    auto cand_act = ctx.Attr<std::string>("candidate_activation");
D
dangqingqing 已提交
270 271 272

    auto batch_starts = batch_gate->lod()[0];
    size_t num_batch = batch_starts.size() - 1;
273
    for (int n = static_cast<int>(num_batch) - 1; n >= 0; n--) {
D
dangqingqing 已提交
274 275 276 277 278 279
      int bstart = static_cast<int>(batch_starts[n]);
      int bend = static_cast<int>(batch_starts[n + 1]);

      Tensor gate = batch_gate->Slice(bstart, bend);
      Tensor cell = batch_cell.Slice(bstart, bend);
      Tensor cell_pre_act = batch_cell_pre_act->Slice(bstart, bend);
280 281 282
      lstm_value.gate_value = gate.data<T>();
      lstm_value.state_value = cell.data<T>();
      lstm_value.state_active_value = cell_pre_act.data<T>();
D
dangqingqing 已提交
283 284 285 286

      Tensor out_g = batch_hidden_g.Slice(bstart, bend);
      Tensor gate_g = batch_gate_g.Slice(bstart, bend);
      Tensor cell_g = batch_cell_g.Slice(bstart, bend);
287 288 289
      lstm_grad.state_grad = cell_g.data<T>();
      lstm_grad.gate_grad = gate_g.data<T>();
      lstm_grad.output_grad = out_g.data<T>();
D
dangqingqing 已提交
290

291
      if (n > 0) {
D
dangqingqing 已提交
292 293 294
        int bstart_pre = static_cast<int>(batch_starts[n - 1]);
        Tensor cell_pre = batch_cell.Slice(bstart_pre, bstart);
        Tensor cell_pre_g = batch_cell_g.Slice(bstart_pre, bstart);
295 296
        lstm_value.prev_state_value = cell_pre.data<T>();
        lstm_grad.prev_state_grad = cell_pre_g.data<T>();
D
dangqingqing 已提交
297
      } else {
298 299
        lstm_value.prev_state_value = c0 ? ordered_c0.data<T>() : nullptr;
        lstm_grad.prev_state_grad = c0_g ? ordered_c0_g.data<T>() : nullptr;
D
dangqingqing 已提交
300 301 302
      }

      int cur_batch_size = bend - bstart;
Q
QI JUN 已提交
303
      math::LstmUnitGradFunctor<DeviceContext, T>::compute(
D
dangqingqing 已提交
304 305 306
          device_ctx, lstm_value, lstm_grad, frame_size, cur_batch_size,
          gate_act, cell_act, cand_act);

307
      if (n > 0) {
D
dangqingqing 已提交
308 309 310
        int pre_h_start = static_cast<int>(batch_starts[n - 1]);
        int pre_h_end = pre_h_start + cur_batch_size;
        auto pre_hidden_g = batch_hidden_g.Slice(pre_h_start, pre_h_end);
Q
QI JUN 已提交
311 312 313
        math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight, true,
                                       static_cast<T>(1.0), &pre_hidden_g,
                                       static_cast<T>(1.0));
D
dangqingqing 已提交
314 315 316
        if (weight_g) {
          /* backward weight */
          auto pre_hidden = batch_hidden.Slice(pre_h_start, pre_h_end);
Q
QI JUN 已提交
317 318 319
          math::matmul<DeviceContext, T>(device_ctx, pre_hidden, true, gate_g,
                                         false, static_cast<T>(1.0), weight_g,
                                         static_cast<T>(1.0));
D
dangqingqing 已提交
320
        }
321 322
      } else {
        if (h0 && weight_g) {
Q
QI JUN 已提交
323 324 325 326 327
          ReorderInitState<DeviceContext, T>(device_ctx, *h0, order,
                                             &ordered_h0, true);
          math::matmul<DeviceContext, T>(device_ctx, ordered_h0, true, gate_g,
                                         false, static_cast<T>(1.0), weight_g,
                                         static_cast<T>(1.0));
328 329 330
        }
        if (h0 && h0_g) {
          ordered_h0_g.mutable_data<T>(h0_g->dims(), ctx.GetPlace());
Q
QI JUN 已提交
331 332 333
          math::matmul<DeviceContext, T>(device_ctx, gate_g, false, *weight,
                                         true, static_cast<T>(1.0),
                                         &ordered_h0_g, static_cast<T>(0.0));
334
        }
D
dangqingqing 已提交
335 336 337
      }
    }

Q
QI JUN 已提交
338
    math::Batch2LoDTensorFunctor<DeviceContext, T> to_seq;
D
dangqingqing 已提交
339 340
    if (in_g) {
      /* backward data */
341
      in_g->mutable_data<T>(ctx.GetPlace());
D
dangqingqing 已提交
342 343 344 345
      to_seq(device_ctx, batch_gate_g, *in_g);
    }
    if (bias && bias_g) {
      /* backward bias */
346 347 348
      Tensor b_g = *bias_g;
      b_g.Resize({bias_g->numel(), 1});
      Tensor gate_bias_g = b_g.Slice(0, 4 * frame_size);
Q
QI JUN 已提交
349
      math::ColwiseSum<DeviceContext, T> col_sum;
350
      col_sum(device_ctx, batch_gate_g, &gate_bias_g);
D
dangqingqing 已提交
351
    }
352 353

    if (h0 && h0_g) {
Q
QI JUN 已提交
354 355
      ReorderInitState<DeviceContext, T>(device_ctx, ordered_h0_g, order, h0_g,
                                         false);
356 357
    }
    if (c0 && c0_g) {
Q
QI JUN 已提交
358 359
      ReorderInitState<DeviceContext, T>(device_ctx, ordered_c0_g, order, c0_g,
                                         false);
360
    }
D
dangqingqing 已提交
361
  }
D
dangqingqing 已提交
362 363 364 365
};

}  // namespace operators
}  // namespace paddle