• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 // XLA implementations of Categorical op.
17 
18 #include "tensorflow/compiler/tf2xla/kernels/random_ops_util.h"
19 #include "tensorflow/compiler/tf2xla/shape_util.h"
20 #include "tensorflow/compiler/tf2xla/type_util.h"
21 #include "tensorflow/compiler/tf2xla/xla_helpers.h"
22 #include "tensorflow/compiler/tf2xla/xla_op_kernel.h"
23 #include "tensorflow/compiler/tf2xla/xla_op_registry.h"
24 #include "tensorflow/compiler/xla/client/lib/arithmetic.h"
25 #include "tensorflow/compiler/xla/client/lib/constants.h"
26 #include "tensorflow/compiler/xla/client/lib/prng.h"
27 #include "tensorflow/compiler/xla/client/xla_builder.h"
28 #include "tensorflow/compiler/xla/xla_data.pb.h"
29 #include "tensorflow/core/framework/op_kernel.h"
30 #include "tensorflow/core/framework/tensor.h"
31 #include "tensorflow/core/framework/tensor_shape.h"
32 #include "tensorflow/core/framework/types.pb.h"
33 
34 namespace tensorflow {
35 namespace {
36 
37 class CategoricalOp : public XlaOpKernel {
38  public:
CategoricalOp(OpKernelConstruction * ctx)39   explicit CategoricalOp(OpKernelConstruction* ctx)
40       : XlaOpKernel(ctx),
41         is_gpu_(ctx->device_type().type_string() == DEVICE_GPU_XLA_JIT) {}
42 
Compile(XlaOpKernelContext * ctx)43   void Compile(XlaOpKernelContext* ctx) override {
44     // Get the logits
45     const xla::XlaOp& logits = ctx->Input(0);
46     TensorShape logits_shape = ctx->InputShape(0);
47     int64_t num_samples;
48     OP_REQUIRES_OK(ctx,
49                    ctx->ConstantInputAsIntScalar(
50                        1, &num_samples, xla::ValueInferenceMode::kUpperBound));
51     OP_REQUIRES(ctx, TensorShapeUtils::IsMatrix(logits_shape),
52                 errors::InvalidArgument("logits should be a matrix, got shape ",
53                                         logits_shape.DebugString()));
54     OP_REQUIRES(ctx, num_samples >= 0,
55                 errors::InvalidArgument(
56                     "num_samples should be nonnegative, got ", num_samples));
57 
58     for (int i = 0; i < 2; i++) {
59       const int64_t dim = logits_shape.dim_size(i);
60       OP_REQUIRES(
61           ctx, static_cast<int>(dim) == dim,
62           errors::InvalidArgument("logits.shape = ", logits_shape.DebugString(),
63                                   " too large for int"));
64     }
65 
66     const int64_t batch_size = logits_shape.dim_size(0);
67     const int64_t num_classes = logits_shape.dim_size(1);
68 
69     xla::Shape uniform_shape;
70     int class_dimension;
71     bool num_samples_is_dynamic = false;
72     OP_REQUIRES_OK(
73         ctx, ctx->ResolveInputDynamismIntoPred(1, &num_samples_is_dynamic));
74     if (num_samples != 1 || num_samples_is_dynamic) {
75       std::array<int64, 3> uniform_shape_array = {
76           {batch_size, num_samples, num_classes}};
77       xla::PrimitiveType uniform_xla_type;
78       OP_REQUIRES_OK(ctx,
79                      DataTypeToPrimitiveType(input_type(0), &uniform_xla_type));
80       uniform_shape =
81           xla::ShapeUtil::MakeShape(uniform_xla_type, uniform_shape_array);
82       class_dimension = 2;
83     } else {
84       // Have a special case for when we only need one sample, because
85       // dimensions may be padded on architectures with tiled memory layouts, so
86       // if the num_classes or batch size is large then this can lead to
87       // expensive wasted memory.
88       std::array<int64, 2> uniform_shape_array = {{batch_size, num_classes}};
89       xla::PrimitiveType uniform_xla_type;
90       OP_REQUIRES_OK(ctx,
91                      DataTypeToPrimitiveType(input_type(0), &uniform_xla_type));
92       uniform_shape =
93           xla::ShapeUtil::MakeShape(uniform_xla_type, uniform_shape_array);
94       class_dimension = 1;
95     }
96     xla::PrimitiveType type;
97     OP_REQUIRES_OK(ctx, DataTypeToPrimitiveType(input_type(0), &type));
98     xla::XlaOp log_uniforms = GetLogUniforms(uniform_shape, type, ctx);
99 
100     if (num_samples_is_dynamic) {
101       // num_samples is dimension 1 in uniform_shape_array.
102       log_uniforms = xla::SetDimensionSize(log_uniforms, ctx->Input(1), 1);
103     }
104 
105     // Use Gumbel softmax trick to generate categorical samples.
106     // See:
107     // https://hips.seas.harvard.edu/blog/2013/04/06/the-gumbel-max-trick-for-discrete-distributions/
108     // TODO(b/68769470): Switch to using a cumulative sum approach.
109     auto softmax_entries =
110         xla::Sub(logits, log_uniforms,
111                  /*broadcast_dimensions=*/{0, class_dimension});
112 
113     xla::PrimitiveType xla_output_type;
114     OP_REQUIRES_OK(ctx,
115                    DataTypeToPrimitiveType(output_type(0), &xla_output_type));
116     xla::XlaOp argmax;
117     if (is_gpu_) {
118       argmax = xla::ArgMaxTwoPass(softmax_entries, xla_output_type,
119                                   /*axis=*/class_dimension);
120     } else {
121       argmax = xla::ArgMax(softmax_entries, xla_output_type,
122                            /*axis=*/class_dimension, /*stable=*/true);
123     }
124 
125     if (num_samples == 1 && !num_samples_is_dynamic) {
126       argmax = xla::Reshape(argmax, {batch_size, 1});
127     }
128 
129     ctx->SetOutput(0, argmax);
130   }
131 
GetLogUniforms(xla::Shape uniform_shape,xla::PrimitiveType type,XlaOpKernelContext * ctx)132   virtual xla::XlaOp GetLogUniforms(xla::Shape uniform_shape,
133                                     xla::PrimitiveType type,
134                                     XlaOpKernelContext* ctx) {
135     xla::XlaBuilder* builder = ctx->builder();
136     LOG_FIRST_N(WARNING, 1) << "Warning: Using tf.random.categorical with XLA"
137                                " compilation will ignore seeds.";
138     // We want a number in (0, 1) rather than [0, 1) or (0, 1]:
139     // * log(-log(0)) is ∞.
140     // * log(-log(1)) is -∞.
141     auto uniforms = xla::RngUniform(
142         xla::MinPositiveNormalValue(builder, type),
143         xla::One(builder, uniform_shape.element_type()), uniform_shape);
144     return xla::Log(-xla::Log(uniforms));
145   }
146 
147  private:
148   bool is_gpu_;
149   TF_DISALLOW_COPY_AND_ASSIGN(CategoricalOp);
150 };
151 
152 // TODO(b/68769717): Rename this sampler to Categorical.
153 REGISTER_XLA_OP(Name("Multinomial").CompileTimeConstantInput("num_samples"),
154                 CategoricalOp);
155 
156 class StatelessCategoricalOp : public CategoricalOp {
157  public:
StatelessCategoricalOp(OpKernelConstruction * ctx)158   explicit StatelessCategoricalOp(OpKernelConstruction* ctx)
159       : CategoricalOp(ctx),
160         device_type_string_(ctx->device_type().type_string()) {
161     OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_));
162   }
163 
GetLogUniforms(xla::Shape uniform_shape,xla::PrimitiveType type,XlaOpKernelContext * ctx)164   xla::XlaOp GetLogUniforms(xla::Shape uniform_shape, xla::PrimitiveType type,
165                             XlaOpKernelContext* ctx) override {
166     xla::XlaOp seed = ctx->Input(2);
167 
168     xla::XlaBuilder* builder = ctx->builder();
169     if (uniform_shape.element_type() == xla::BF16) {
170       uniform_shape.set_element_type(xla::F32);
171     }
172     // We want a number in (0, 1) rather than [0, 1) or (0, 1]:
173     // * log(-log(0)) is ∞.
174     // * log(-log(1)) is -∞.
175     xla::XlaOp uniforms = StatelessRngUniform(
176         device_type_string_, seed, uniform_shape,
177         xla::MinPositiveNormalValue(builder, uniform_shape.element_type()),
178         xla::One(builder, uniform_shape.element_type()));
179     return xla::ConvertElementType(xla::Log(-xla::Log(uniforms)), type);
180   }
181 
Compile(XlaOpKernelContext * ctx)182   void Compile(XlaOpKernelContext* ctx) override {
183     TensorShape seed_shape = ctx->InputShape(2);
184     OP_REQUIRES(ctx, seed_shape.dims() == 1 && seed_shape.dim_size(0) == 2,
185                 errors::InvalidArgument("seed must have shape [2], not ",
186                                         seed_shape.DebugString()));
187     CategoricalOp::Compile(ctx);
188   }
189 
190  private:
191   DataType dtype_;
192   string device_type_string_;
193 
194   TF_DISALLOW_COPY_AND_ASSIGN(StatelessCategoricalOp);
195 };
196 
197 REGISTER_XLA_OP(Name("StatelessMultinomial")
198                     .CompileTimeConstantInput("num_samples")
199                     .TypeConstraint("T", {DT_DOUBLE, DT_FLOAT, DT_BFLOAT16})
200                     .TypeConstraint("Tseed", DT_INT32),
201                 StatelessCategoricalOp);
202 
203 }  // anonymous namespace
204 }  // namespace tensorflow
205