backprop_derivative.hpp 8.61 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
//*****************************************************************************
// Copyright 2017-2018 Intel Corporation
//
// 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
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// 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.
//*****************************************************************************
Scott Cyphers's avatar
Scott Cyphers committed
16 17 18 19

#pragma once

#include <memory>
20
#include <unordered_map>
Scott Cyphers's avatar
Scott Cyphers committed
21

22
#include "ngraph/autodiff/adjoints.hpp"
23
#include "ngraph/graph_util.hpp"
Robert Kimball's avatar
Robert Kimball committed
24
#include "ngraph/log.hpp"
25
#include "ngraph/type/element_type.hpp"
Robert Kimball's avatar
Robert Kimball committed
26
#include "ngraph/util.hpp"
27
#include "util/all_close.hpp"
28
#include "util/test_tools.hpp"
Scott Cyphers's avatar
Scott Cyphers committed
29 30 31 32 33 34

namespace ngraph
{
    class Node;
    class Function;

35 36 37 38
    static std::unordered_map<std::shared_ptr<Function>, std::shared_ptr<Function>> s_df_map;
    static std::unordered_map<std::shared_ptr<Function>, std::shared_ptr<Function>> s_clone_fwd_map;
    static std::unordered_map<std::shared_ptr<Function>, std::shared_ptr<Function>> s_clone_bwd_map;

Scott Cyphers's avatar
Scott Cyphers committed
39 40 41 42
    namespace runtime
    {
        class Backend;
        class Manager;
43
    }
Scott Cyphers's avatar
Scott Cyphers committed
44 45 46

    namespace autodiff
    {
47
        template <typename T>
48
        std::vector<std::shared_ptr<runtime::Tensor>>
49
            get_autodiff(runtime::Backend* backend,
50
                         std::shared_ptr<Function>& df,
51
                         const std::vector<std::shared_ptr<runtime::Tensor>>& df_input_args,
52
                         const std::vector<std::shared_ptr<op::Parameter>>& indep_params)
Robert Kimball's avatar
Robert Kimball committed
53
        {
54 55 56 57
            // df/dX* = f'(c, ...)
            // using X* to denote all x "of interest" (represented by indep_params)

            // return value for this function
58
            std::vector<std::shared_ptr<runtime::Tensor>> results;
Robert Kimball's avatar
Robert Kimball committed
59

60
            // adjoint
61 62
            auto c_arg = df_input_args[0];
            auto y_shape = c_arg->get_shape();
Robert Kimball's avatar
Robert Kimball committed
63

64
            // df/dX* arguments
65
            std::vector<std::shared_ptr<runtime::Tensor>> df_output_args;
66 67 68

            // for each x "of interest"
            for (auto x : indep_params)
Robert Kimball's avatar
Robert Kimball committed
69
            {
70
                // add df/dx to df/dX* arguments
71
                auto x_shape = x->get_shape();
72
                df_output_args.push_back(backend->create_tensor<T>(x_shape));
73 74 75 76 77

                // each element of y has a derivative with respect to each element of x
                // hence, create a y by x sized tensor for this result
                auto y_by_x_shape = y_shape;
                y_by_x_shape.insert(y_by_x_shape.end(), x_shape.begin(), x_shape.end());
78
                results.push_back(backend->create_tensor<T>(y_by_x_shape));
Robert Kimball's avatar
Robert Kimball committed
79 80
            }

81
            // create storage for results
82
            std::vector<std::vector<T>> result_vect;
83
            std::vector<typename std::vector<T>::iterator> result_pos;
Robert Kimball's avatar
Robert Kimball committed
84 85
            for (auto result : results)
            {
86
                result_vect.push_back(read_vector<T>(result));
Robert Kimball's avatar
Robert Kimball committed
87 88 89
                result_pos.push_back(result_vect.back().begin());
            }

90
            // get adjoint and force to all elements to zero
91
            auto c_vec = read_vector<T>(c_arg);
92
            fill(c_vec.begin(), c_vec.end(), 0);
93 94 95

            // for each element of the adjoint
            // same as saying for each element of y
Robert Kimball's avatar
Robert Kimball committed
96 97
            for (size_t i = 0; i < c_vec.size(); i++)
            {
98
                // set a single adjoint element
Robert Kimball's avatar
Robert Kimball committed
99
                c_vec[i] = 1;
100
                write_vector(c_arg, c_vec);
101 102

                // call modified df/dX* = f'(c, cached)
103
                backend->call_with_validate(df, df_output_args, df_input_args);
104 105

                // reset the adjoint element
Robert Kimball's avatar
Robert Kimball committed
106
                c_vec[i] = 0;
107
                write_vector(c_arg, c_vec);
108 109 110

                // for each result
                // same as saying for each x "of interest"
Robert Kimball's avatar
Robert Kimball committed
111 112
                for (size_t j = 0; j < results.size(); j++)
                {
113 114 115
                    // copy df/dx to storage for this element of y
                    auto dfdx = read_vector<T>(df_output_args[j]);
                    result_pos[j] = std::copy(dfdx.begin(), dfdx.end(), result_pos[j]);
Robert Kimball's avatar
Robert Kimball committed
116 117 118
                }
            }

119
            // copy storage to results and return
Robert Kimball's avatar
Robert Kimball committed
120 121
            for (size_t j = 0; j < results.size(); j++)
            {
122
                write_vector(results[j], result_vect[j]);
Robert Kimball's avatar
Robert Kimball committed
123 124 125
            }
            return results;
        }
126 127

        template <typename T>
128
        std::vector<std::shared_ptr<runtime::Tensor>>
129
            backprop_derivative(runtime::Backend* backend,
130 131 132
                                const std::shared_ptr<Function>& f,
                                const std::vector<std::shared_ptr<runtime::Tensor>>& f_input_args,
                                const std::vector<std::shared_ptr<op::Parameter>>& indep_params)
133 134 135 136 137 138 139 140 141
        {
            // y = f(X)
            // using X (upper case) to denote all paramenters of f (represented by f_input_args)
            // using x (lower case) to denote an individual paramemter of f
            // using X* to denote all x "of interest" (represented by indep_params)
            Shape y_shape = f->get_output_shape(0);

            // adjoint
            auto c_param = std::make_shared<op::Parameter>(element::from<T>(), y_shape);
142
            auto c_arg = backend->create_tensor<T>(y_shape);
143 144 145 146

            // df/dX*
            std::vector<std::shared_ptr<Node>> df_output_params;

147 148
            Adjoints adjoints(NodeVector{f->get_output_op(0)}, NodeVector{c_param});

149 150 151 152
            // for each x "of interest"
            for (auto x : indep_params)
            {
                // add df/dx to df/dX*
153
                df_output_params.push_back(adjoints.backprop_node(x));
154 155 156 157 158 159 160
            }

            // (c, X)
            std::vector<std::shared_ptr<op::Parameter>> df_input_params = f->get_parameters();
            df_input_params.insert(df_input_params.begin(), c_param);

            // df/dX* = f'(c, X)
161 162 163 164 165
            if (!s_df_map[f])
            {
                s_df_map[f] = std::make_shared<Function>(df_output_params, df_input_params);
            }
            auto df = s_df_map[f];
166 167

            // (c, X) arguments
168
            std::vector<std::shared_ptr<runtime::Tensor>> df_input_args = f_input_args;
169 170 171
            df_input_args.insert(df_input_args.begin(), c_arg);

            // call f'(c,X) to get df/dX*
172
            auto dfdx = get_autodiff<T>(backend, df, df_input_args, indep_params);
173 174 175 176

            // create fprop cache
            // creates modified forward function -> (y, cached) = f(x)
            // creates modified backward function -> df/dX* = f'(c, cached)
177
            auto fprop_cache = cache_fprop(f, df);
178 179

            // (y, cached) arguments
180
            std::vector<std::shared_ptr<runtime::Tensor>> mod_f_output_args;
181
            mod_f_output_args.push_back(backend->create_tensor<T>(y_shape));
182 183

            // (c, cached) arguments
184
            std::vector<std::shared_ptr<runtime::Tensor>> mod_df_input_args = df_input_args;
185 186 187 188

            // add cached nodes to both modified f output and modified f' input arguments
            for (auto node : fprop_cache.fprop_output_nodes)
            {
189
                auto tv = backend->create_tensor(node->get_element_type(), node->get_shape());
190 191 192 193 194
                mod_f_output_args.push_back(tv);
                mod_df_input_args.push_back(tv);
            }

            // compile and run modified (y, cached) = f(x)
195 196 197 198 199 200
            if (!s_clone_fwd_map[f])
            {
                s_clone_fwd_map[f] = clone_function(*fprop_cache.fprop);
            }
            auto clone_fwd = s_clone_fwd_map[f];

201
            backend->call_with_validate(clone_fwd, mod_f_output_args, f_input_args);
202 203

            // call modfied f'(c, cached) to get df/dX*
204 205 206 207 208
            if (!s_clone_bwd_map[f])
            {
                s_clone_bwd_map[f] = clone_function(*fprop_cache.bprop);
            }
            auto clone_bwd = s_clone_bwd_map[f];
209
            auto cache_dfdx = get_autodiff<T>(backend, clone_bwd, mod_df_input_args, indep_params);
210

211 212
            const T numpy_atol = static_cast<const T>(1e-5f);
            const T numpy_rtol = static_cast<const T>(1e-8f);
213 214 215 216 217 218 219 220 221
            auto close = ngraph::test::all_close<T>(dfdx, cache_dfdx, numpy_atol, numpy_rtol);
            if (!close)
            {
                throw ngraph_error(
                    "Derivatives mismatch between cache and non-cache bprop functions");
            }

            return dfdx;
        }
Scott Cyphers's avatar
Scott Cyphers committed
222 223
    }
}