Feature/evaluator

doc/design/evaluator.md

@@ -0,0 +1,58 @@
+## Evaluator Design
+
+### The Problem
+
+During training or serving, we provide the evaluation function to measure the model performance, e.g., accuracy, precision. In the operator based framework design, the data go through the network pipeline batch by batch. As a result, inside the operator, we only can calculate one minibatch metrics. We need to provide a mechanism to calculate the metrics for each N pass/batch the user wanted.
+
+### Evaluator Design
+Currently, every operation is expressed in the graph. we divide the evaluator process into three steps.
+
+1. Initialize the metric state and add it into the block.
+
+2. Calculate the statistic of the metric state in every mini-batch. The single operator is only responsible for calculating necessary statistics for one mini-batch. For example, accuracy operator only calculate a minibatch data if run once.
+
+
+3. Merge the mini-batch statistics to form the evaluation result for multiple mini-batches. When it comes to distributed training/Multi-GPU training, aggregate the value from different devices.
+
+### Implementation
+This design is shown in python API. 
+Each metric operator need to caculate the metric statistic and return the batch aware states, Python side responsible for accumulate the states for each pass. 
+
+
+```python
+class Evaluator(object):
+    """
+    Evaluator Base class.
+    """
+    def __init__(self, name, **kwargs):
+       """
+       Different evaluator may has different metric states. E.g, Accuracy need two variables, total and right sample counts.
+       Auc need four variables, `true_positives`,
+         `true_negatives`, `false_positives` and `false_negatives`. So every evaluator should create its needed variables and append to main_program
+
+       The initialization of Evaluator should be responsible for:
+       create metric states and append to the main_program
+       """ 
+       pass
+
+    def _update_ops(self, input, label, **kwargs)
+       """
+       Add mini-batch evaluator caculate operators to the main_program.
+       Add increment operator to accumulate the metric states.
+       """
+
+
+    def reset(self, executor, reset_program=None):
+      """
+      Reset metric states at the begin of each pass/user specified batch number.
+      Execute the reset_program to reset the states.
+      """
+
+
+    def eval(self, executor, eval_program=None):
+      """
+      Merge the mini-batch statistics to form the evaluation result for multiple mini-batches.
+      Execute the eval_program and return the result.
+      """
+      return eval_result
+```

paddle/operators/accuracy_op.cc

@@ -30,6 +30,10 @@ class AccuracyOp : public framework::OperatorWithKernel {
                    "Input (Label) of accuracy op should not be null.");
     PADDLE_ENFORCE(ctx->HasOutput("Accuracy"),
                    "Output (Accuracy) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Correct"),
+                   "Output (Correct) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Total"),
+                   "Output (Total) of AccuracyOp should not be null.");
 
     auto inference_dim = ctx->GetInputDim("Out");
     auto label_dim = ctx->GetInputDim("Label");
@@ -43,6 +47,8 @@ class AccuracyOp : public framework::OperatorWithKernel {
                       " the same as label.");
 
     ctx->SetOutputDim("Accuracy", {1});
+    ctx->SetOutputDim("Correct", {1});
+    ctx->SetOutputDim("Total", {1});
     ctx->ShareLoD("Out", /*->*/ "Accuracy");
   }
 
@@ -66,6 +72,8 @@ class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("Label", "Label of the training data");
     // TODO(typhoonzero): AddInput("Weight", ...
     AddOutput("Accuracy", "The accuracy of current batch");
+    AddOutput("Correct", "The correct samples count of current batch");
+    AddOutput("Total", "The samples count of current batch");
 
     AddComment(R"DOC(
 Accuracy Operator. 

paddle/operators/accuracy_op.cu

@@ -24,7 +24,8 @@ using platform::PADDLE_CUDA_NUM_THREADS;
 template <int BlockSize>
 __global__ void AccuracyCudaKernel(const int N, const int D,
                                    const int64_t* Xdata,
-                                   const int64_t* labeldata, float* accuracy) {
+                                   const int64_t* labeldata, int* correct_data,
+                                   float* accuracy) {
   int count = 0;
   __shared__ int total[BlockSize];
 
@@ -43,6 +44,7 @@ __global__ void AccuracyCudaKernel(const int N, const int D,
   // reduce the count with init value 0, and output accuracy.
   int result = thrust::reduce(thrust::device, total, total + BlockSize, 0);
   if (threadIdx.x == 0) {
+    *correct_data = result;
     *accuracy = static_cast<float>(result) / static_cast<float>(N);
   }
 }
@@ -56,31 +58,48 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
     auto* inference = ctx.Input<Tensor>("Out");
     auto* indices = ctx.Input<Tensor>("Indices");
     auto* label = ctx.Input<Tensor>("Label");
+
     auto* accuracy = ctx.Output<Tensor>("Accuracy");
+    auto* correct = ctx.Output<Tensor>("Correct");
+    auto* total = ctx.Output<Tensor>("Total");
     // FIXME(typhoonzero): only support indices currently
     // if add support for output values, how to detect the data type?
     const int64_t* indices_data = indices->data<int64_t>();
     const int64_t* label_data = label->data<int64_t>();
+
+    int* correct_data = correct->mutable_data<int>(ctx.GetPlace());
+    int* total_data = total->mutable_data<int>(ctx.GetPlace());
     float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
 
-    size_t num_samples = inference->dims()[0];
+    int num_samples = static_cast<int>(inference->dims()[0]);
     size_t infer_width = inference->dims()[1];
     PADDLE_ENFORCE(cudaMemset(accuracy_data, 0, sizeof(float)));
+    // cudaMemset((void**)&correct_data, 0, sizeof(float));
 
     if (num_samples == 0) {
       return;
     }
+    cudaMemcpy(total_data, &num_samples, sizeof(int), cudaMemcpyHostToDevice);
 
     AccuracyCudaKernel<PADDLE_CUDA_NUM_THREADS><<<
         1, PADDLE_CUDA_NUM_THREADS, 0, ctx.cuda_device_context().stream()>>>(
-        num_samples, infer_width, indices_data, label_data, accuracy_data);
+        num_samples, infer_width, indices_data, label_data, correct_data,
+        accuracy_data);
+
+    int d_num_samples, d_num_correct;
+    float d_accuracy;
+    cudaMemcpy(&d_num_correct, correct_data, sizeof(int),
+               cudaMemcpyDeviceToHost);
+    cudaMemcpy(&d_num_samples, total_data, sizeof(int), cudaMemcpyDeviceToHost);
+    cudaMemcpy(&d_accuracy, accuracy_data, sizeof(float),
+               cudaMemcpyDeviceToHost);
   }
 };
 
 }  // namespace operators
 }  // namespace paddle
 
-// FIXME(typhoonzero): types of T is for infernece data.
-// label data is always int
+// FIXME(typhoonzero): types of T is for inference data.
+// label data is always int64
 REGISTER_OP_GPU_KERNEL(accuracy, paddle::operators::AccuracyOpCUDAKernel<float>,
                        paddle::operators::AccuracyOpCUDAKernel<double>);

paddle/operators/accuracy_op.h

@@ -29,7 +29,11 @@ class AccuracyKernel : public framework::OpKernel<T> {
     auto* indices = ctx.Input<Tensor>("Indices");
     auto* label = ctx.Input<Tensor>("Label");
     auto* accuracy = ctx.Output<Tensor>("Accuracy");
+    auto* correct = ctx.Output<Tensor>("Correct");
+    auto* total = ctx.Output<Tensor>("Total");
 
+    int* correct_data = correct->mutable_data<int>(ctx.GetPlace());
+    int* total_data = total->mutable_data<int>(ctx.GetPlace());
     float* accuracy_data = accuracy->mutable_data<float>(ctx.GetPlace());
 
     const int64_t* indices_data = indices->data<int64_t>();
@@ -55,7 +59,8 @@ class AccuracyKernel : public framework::OpKernel<T> {
       }
     }
 
-    // FIXME(typhoonzero): we don't accumulate the accuracy for now.
+    *correct_data = num_correct;
+    *total_data = num_samples;
     *accuracy_data =
         static_cast<float>(num_correct) / static_cast<float>(num_samples);
   }

paddle/operators/elementwise_add_op.cc

@@ -34,7 +34,13 @@ REGISTER_OP(elementwise_add, ops::ElementwiseOp, ops::ElementwiseAddOpMaker,
             elementwise_add_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_add,
-    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, float>,
+    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, double>,
+    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, int>,
+    ops::ElementwiseAddKernel<paddle::platform::CPUPlace, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_add_grad,
-    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, float>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, double>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, int>,
+    ops::ElementwiseAddGradKernel<paddle::platform::CPUPlace, int64_t>);

paddle/operators/elementwise_div_op.cc

@@ -35,7 +35,13 @@ REGISTER_OP(elementwise_div, ops::ElementwiseOp, ops::ElementwiseDivOpMaker,
             elementwise_div_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_div,
-    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, float>,
+    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, double>,
+    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, int>,
+    ops::ElementwiseDivKernel<paddle::platform::CPUPlace, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_div_grad,
-    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, float>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, double>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, int>,
+    ops::ElementwiseDivGradKernel<paddle::platform::CPUPlace, int64_t>);

paddle/operators/elementwise_mul_op.cc

@@ -37,8 +37,12 @@ REGISTER_OP(elementwise_mul, ops::ElementwiseOp, ops::ElementwiseMulOpMaker,
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul,
     ops::ElementwiseMulKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, double>);
+    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, double>,
+    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, int>,
+    ops::ElementwiseMulKernel<paddle::platform::CPUPlace, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_mul_grad,
     ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, float>,
-    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, double>);
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, double>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, int>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CPUPlace, int64_t>);

paddle/operators/elementwise_sub_op.cc

@@ -34,7 +34,13 @@ REGISTER_OP(elementwise_sub, ops::ElementwiseOp, ops::ElementwiseSubOpMaker,
             elementwise_sub_grad, ops::ElementwiseOpGrad);
 REGISTER_OP_CPU_KERNEL(
     elementwise_sub,
-    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, float>,
+    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, double>,
+    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, int>,
+    ops::ElementwiseSubKernel<paddle::platform::CPUPlace, int64_t>);
 REGISTER_OP_CPU_KERNEL(
     elementwise_sub_grad,
-    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, float>);
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, float>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, double>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, int>,
+    ops::ElementwiseSubGradKernel<paddle::platform::CPUPlace, int64_t>);