Revert "[Feature] Fp16 training for resnet50 (#14850)"

This reverts commit 3d750f9c.

paddle/fluid/framework/details/multi_devices_graph_pass.cc

@@ -355,9 +355,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
           BuildStrategy::GradientScaleStrategy::kCustomized) {
         // TODO(paddle-dev): Why is there no input for this op_handle?
         auto loss_grad_name = node->Op()->OutputArgumentNames()[0];
-        auto out_dtype = all_vars_.at(loss_grad_name)->GetDataType();
-        CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0],
-                              out_dtype);
+        CreateScaleLossGradOp(&result, loss_grad_name, node->outputs[0]);
       }
       // This assumes the backward generating code will ensure IsScaleLossOp
       // is true only for the op that scale the final scalar loss.
@@ -660,13 +658,13 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(
 
 void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(
     ir::Graph *result, const std::string &loss_grad_name,
-    ir::Node *out_var_node, proto::VarType::Type dtype) const {
+    ir::Node *out_var_node) const {
   for (size_t i = 0; i < places_.size(); ++i) {
     // Insert ScaleCost OpHandle
     auto *dev_ctx = platform::DeviceContextPool::Instance().Get(places_[i]);
     auto *op_handle = new ScaleLossGradOpHandle(
         result->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation),
-        local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx, dtype);
+        local_scopes_.size(), local_scopes_[i], places_[i], dev_ctx);
     result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
 
     // FIXME: Currently ScaleLossGradOp only use device_count as scale

paddle/fluid/framework/details/multi_devices_graph_pass.h

@@ -68,8 +68,7 @@ class MultiDevSSAGraphBuilder : public ir::Pass {
 
   void CreateScaleLossGradOp(ir::Graph *result,
                              const std::string &loss_grad_name,
-                             ir::Node *out_var_node,
-                             proto::VarType::Type dtype) const;
+                             ir::Node *out_var_node) const;
 
   VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
                             int dst_dev_id) const;

paddle/fluid/framework/details/scale_loss_grad_op_handle.cc

@@ -22,66 +22,39 @@ namespace details {
 ScaleLossGradOpHandle::ScaleLossGradOpHandle(ir::Node *node, size_t num_dev,
                                              Scope *scope,
                                              platform::Place place,
-                                             platform::DeviceContext *dev_ctx,
-                                             proto::VarType::Type dtype)
+                                             platform::DeviceContext *dev_ctx)
     : OpHandleBase(node),
       coeff_(static_cast<float>(1.0 / num_dev)),
       scope_(scope),
-      place_(place),
-      out_dtype_(dtype) {
+      place_(place) {
   this->SetDeviceContext(place_, dev_ctx);
 }
 
 ScaleLossGradOpHandle::~ScaleLossGradOpHandle() {}
 
-struct ScaleLossGradFunctor {
-  float coeff_;
-  Tensor *out_;
-  platform::Place place_;
-  OpHandleBase *op_handle_;
-  proto::VarType::Type out_dtype_;
-  platform::DeviceContext *ctx_;
-
-  ScaleLossGradFunctor(float coeff, Tensor *out, platform::Place place,
-                       OpHandleBase *op_handle, proto::VarType::Type dtype,
-                       platform::DeviceContext *ctx)
-      : coeff_(coeff), out_(out), place_(place), out_dtype_(dtype), ctx_(ctx) {}
-
-  template <typename OutT>
-  void apply() const {
-    auto *out_data = out_->mutable_data<OutT>(place_);
-    if (platform::is_cpu_place(place_)) {
-      *out_data = static_cast<OutT>(coeff_);
-    } else {
-#ifdef PADDLE_WITH_CUDA
-      OutT cast_coeff = static_cast<OutT>(coeff_);
-      auto stream = static_cast<platform::CUDADeviceContext *>(ctx_)->stream();
-      memory::Copy(boost::get<platform::CUDAPlace>(place_), out_data,
-                   platform::CPUPlace(), &cast_coeff, SizeOfType(out_dtype_),
-                   stream);
-      VLOG(10) << place_ << "RUN Scale loss grad op";
-
-#endif
-    }
-  }
-};
-
 void ScaleLossGradOpHandle::RunImpl() {
   // Doesn't wait any event
   std::string var_name = static_cast<VarHandle *>(this->outputs_[0])->name_;
   auto &local_scope = *scope_->FindVar(kLocalExecScopeName)->Get<Scope *>();
 
-  auto *tensor = local_scope.FindVar(var_name)->GetMutable<LoDTensor>();
-  tensor->Resize(make_ddim({1}));
+  float *tmp = local_scope.FindVar(var_name)
+                   ->GetMutable<LoDTensor>()
+                   ->mutable_data<float>(make_ddim({1}), place_);
 
+  if (platform::is_cpu_place(place_)) {
+    *tmp = coeff_;
+  } else {
 #ifdef PADDLE_WITH_CUDA
-  ScaleLossGradFunctor func(coeff_, tensor, place_, this, out_dtype_,
-                            this->dev_ctxes_.at(place_));
-  this->RunAndRecordEvent([&] { framework::VisitDataType(out_dtype_, func); });
-#else
-  ScaleLossGradFunctor func(coeff_, tensor, place_, this, out_dtype_, nullptr);
-  framework::VisitDataType(out_dtype_, func);
+    this->RunAndRecordEvent([&] {
+      auto stream = static_cast<platform::CUDADeviceContext *>(
+                        this->dev_ctxes_.at(place_))
+                        ->stream();
+      memory::Copy(boost::get<platform::CUDAPlace>(place_), tmp,
+                   platform::CPUPlace(), &coeff_, sizeof(float), stream);
+      VLOG(10) << place_ << "RUN Scale loss grad op";
+    });
 #endif
+  }
 }
 
 std::string ScaleLossGradOpHandle::Name() const { return "Scale LossGrad"; }

paddle/fluid/framework/details/scale_loss_grad_op_handle.h

@@ -26,8 +26,8 @@ namespace details {
 
 struct ScaleLossGradOpHandle : public OpHandleBase {
   ScaleLossGradOpHandle(ir::Node *node, size_t num_dev, Scope *scope,
-                        platform::Place place, platform::DeviceContext *context,
-                        proto::VarType::Type dtype);
+                        platform::Place place,
+                        platform::DeviceContext *context);
 
   ~ScaleLossGradOpHandle() final;
 
@@ -40,7 +40,6 @@ struct ScaleLossGradOpHandle : public OpHandleBase {
   float coeff_;
   Scope *scope_;
   platform::Place place_;
-  proto::VarType::Type out_dtype_;
 };
 
 }  // namespace details

paddle/fluid/operators/elementwise/elementwise_div_op.cu

@@ -12,23 +12,18 @@ 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 "paddle/fluid/operators/elementwise/elementwise_div_op.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
 
 REGISTER_OP_CUDA_KERNEL(
     elementwise_div,
     ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext,
-                              paddle::platform::float16>,
     ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, double>,
     ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, int>,
     ops::ElementwiseDivKernel<paddle::platform::CUDADeviceContext, int64_t>);
 REGISTER_OP_CUDA_KERNEL(
     elementwise_div_grad,
     ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext,
-                                  paddle::platform::float16>,
     ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, double>,
     ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext, int>,
     ops::ElementwiseDivGradKernel<paddle::platform::CUDADeviceContext,

paddle/fluid/operators/elementwise/elementwise_mul_op.cu

@@ -12,21 +12,19 @@ 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 "paddle/fluid/operators/elementwise/elementwise_mul_op.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
-namespace plat = paddle::platform;
 
 REGISTER_OP_CUDA_KERNEL(
-    elementwise_mul, ops::ElementwiseMulKernel<plat::CUDADeviceContext, float>,
-    ops::ElementwiseMulKernel<plat::CUDADeviceContext, double>,
-    ops::ElementwiseMulKernel<plat::CUDADeviceContext, int>,
-    ops::ElementwiseMulKernel<plat::CUDADeviceContext, int64_t>,
-    ops::ElementwiseMulKernel<plat::CUDADeviceContext, plat::float16>);
+    elementwise_mul,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseMulKernel<paddle::platform::CUDADeviceContext, int64_t>);
 REGISTER_OP_CUDA_KERNEL(
     elementwise_mul_grad,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, float>,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, double>,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, int>,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, int64_t>,
-    ops::ElementwiseMulGradKernel<plat::CUDADeviceContext, plat::float16>);
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, float>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, double>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext, int>,
+    ops::ElementwiseMulGradKernel<paddle::platform::CUDADeviceContext,
+                                  int64_t>);

paddle/fluid/operators/fill_zeros_like_op.cu.cc

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "paddle/fluid/operators/fill_zeros_like_op.h"
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
@@ -23,6 +22,4 @@ REGISTER_OP_CUDA_KERNEL(
     ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, int64_t>,
     ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, float>,
     ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, double>,
-    ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext,
-                             paddle::platform::float16>,
     ops::FillZerosLikeKernel<paddle::platform::CUDADeviceContext, bool>);

paddle/fluid/operators/metrics/accuracy_op.cu

@@ -16,7 +16,6 @@ limitations under the License. */
 #include <thrust/reduce.h>
 #include "paddle/fluid/operators/metrics/accuracy_op.h"
 #include "paddle/fluid/platform/cuda_primitives.h"
-#include "paddle/fluid/platform/float16.h"
 #include "paddle/fluid/platform/gpu_info.h"
 
 namespace paddle {
@@ -95,7 +94,6 @@ class AccuracyOpCUDAKernel : public framework::OpKernel<T> {
 
 // FIXME(typhoonzero): types of T is for inference data.
 // label data is always int64
-REGISTER_OP_CUDA_KERNEL(
-    accuracy, paddle::operators::AccuracyOpCUDAKernel<float>,
-    paddle::operators::AccuracyOpCUDAKernel<double>,
-    paddle::operators::AccuracyOpCUDAKernel<paddle::platform::float16>);
+REGISTER_OP_CUDA_KERNEL(accuracy,
+                        paddle::operators::AccuracyOpCUDAKernel<float>,
+                        paddle::operators::AccuracyOpCUDAKernel<double>);

paddle/fluid/operators/optimizers/momentum_op.cu

@@ -14,11 +14,8 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/operators/optimizers/momentum_op.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace ops = paddle::operators;
 REGISTER_OP_CUDA_KERNEL(
     momentum, ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, float>,
-    ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, double>,
-    ops::MomentumOpKernel<paddle::platform::CUDADeviceContext,
-                          paddle::platform::float16>);
+    ops::MomentumOpKernel<paddle::platform::CUDADeviceContext, double>);

paddle/fluid/operators/optimizers/momentum_op.h

@@ -237,8 +237,7 @@ class SparseMomentumFunctor<T, UseNesterov> {
   inline HOSTDEVICE void operator()(size_t i) {
     auto row_idx =
         math::BinarySearch<int64_t>(rows_, row_height_, i / row_numel_);
-    T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_]
-                       : static_cast<T>(0);
+    T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0;
     // put memory access in register
     const T p = p_[i];
     const T lr = lr_[0];
@@ -283,8 +282,7 @@ class SparseMomentumFunctor<T, NoNesterov> {
   inline HOSTDEVICE void operator()(size_t i) {
     auto row_idx =
         math::BinarySearch<int64_t>(rows_, row_height_, i / row_numel_);
-    T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_]
-                       : static_cast<T>(0);
+    T g = row_idx >= 0 ? g_[row_idx * row_numel_ + i % row_numel_] : 0;
     // put memory access in register
     const T p = p_[i];
     const T lr = lr_[0];

paddle/fluid/operators/top_k_op.cu

@@ -16,7 +16,6 @@ limitations under the License. */
 #include "paddle/fluid/operators/top_k_op.h"
 #include "paddle/fluid/platform/assert.h"
 #include "paddle/fluid/platform/cuda_device_function.h"
-#include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace operators {
@@ -151,7 +150,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair<T> topk[], int* beam,
         if (k < MaxLength - (*beam)) {
           topk[k] = topk[k + *beam];
         } else {
-          topk[k].set(-static_cast<T>(INFINITY), -1);
+          topk[k].set(-INFINITY, -1);
         }
       }
       if (!(*is_empty)) {
@@ -161,7 +160,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair<T> topk[], int* beam,
     }
 
     *max = topk[MaxLength - 1];
-    if ((*max).v == -static_cast<T>(1)) *is_empty = true;
+    if ((*max).v == -1) *is_empty = true;
     *beam = 0;
   }
 }
@@ -182,7 +181,7 @@ __device__ __forceinline__ void ThreadGetTopK(Pair<T> topk[], int* beam,
         if (k < MaxLength - *beam) {
           topk[k] = topk[k + *beam];
         } else {
-          topk[k].set(-static_cast<T>(INFINITY), -1);
+          topk[k].set(-INFINITY, -1);
         }
       }
       if (!(*is_empty)) {
@@ -279,7 +278,7 @@ __global__ void KeMatrixTopK(T* output, int output_stride, int64_t* indices,
     bool firststep = true;
 
     for (int j = 0; j < MaxLength; j++) {
-      topk[j].set(-static_cast<T>(INFINITY), -1);
+      topk[j].set(-INFINITY, -1);
     }
     while (top_num) {
       ThreadGetTopK<T, MaxLength, BlockSize>(
@@ -363,7 +362,5 @@ class TopkOpCUDAKernel : public framework::OpKernel<T> {
 }  // namespace operators
 }  // namespace paddle
 
-REGISTER_OP_CUDA_KERNEL(
-    top_k, paddle::operators::TopkOpCUDAKernel<float>,
-    paddle::operators::TopkOpCUDAKernel<double>,
-    paddle::operators::TopkOpCUDAKernel<paddle::platform::float16>);
+REGISTER_OP_CUDA_KERNEL(top_k, paddle::operators::TopkOpCUDAKernel<float>,
+                        paddle::operators::TopkOpCUDAKernel<double>);

paddle/fluid/platform/nccl_helper.h

@@ -23,7 +23,6 @@
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/platform/dynload/nccl.h"
 #include "paddle/fluid/platform/enforce.h"
-#include "paddle/fluid/platform/float16.h"
 
 #define NCCL_ID_VARNAME "NCCLID"
 
@@ -39,8 +38,6 @@ inline ncclDataType_t ToNCCLDataType(framework::proto::VarType::Type type) {
     return ncclInt;
   } else if (type == framework::proto::VarType::INT64) {
     return ncclInt64;
-  } else if (type == framework::proto::VarType::FP16) {
-    return ncclFloat16;
   } else {
     PADDLE_THROW("Not supported");
   }

python/paddle/fluid/data_feeder.py

@@ -44,8 +44,6 @@ class DataToLoDTensorConverter(object):
             self.dtype = 'int64'
         elif dtype == core.VarDesc.VarType.FP64:
             self.dtype = 'float64'
-        elif dtype == core.VarDesc.VarType.FP16:
-            self.dtype = 'float16'
         elif dtype == core.VarDesc.VarType.INT32:
             self.dtype = 'int32'
         elif dtype == core.VarDesc.VarType.UINT8:

python/paddle/fluid/initializer.py

@@ -18,7 +18,6 @@ from . import framework
 import numpy as np
 import contextlib
 from .core import VarDesc
-from . import unique_name
 
 __all__ = [
     'Constant', 'Uniform', 'Normal', 'TruncatedNormal', 'Xavier', 'Bilinear',
@@ -208,39 +207,16 @@ class UniformInitializer(Initializer):
         # Initialization Ops should be prepended and not appended
         if self._seed == 0:
             self._seed = block.program.random_seed
-
-        # to be compatible of fp16 initalizers
-        if var.dtype == VarDesc.VarType.FP16:
-            out_dtype = VarDesc.VarType.FP32
-            out_var = block.create_var(
-                name=unique_name.generate(".".join(['gaussian_random', 'tmp'])),
-                shape=var.shape,
-                dtype=out_dtype,
-                type=VarDesc.VarType.LOD_TENSOR,
-                persistable=False)
-        else:
-            out_dtype = var.dtype
-            out_var = var
-
         op = block._prepend_op(
             type="uniform_random",
-            outputs={"Out": out_var},
+            outputs={"Out": var},
             attrs={
                 "shape": var.shape,
-                "dtype": out_dtype,
+                "dtype": int(var.dtype),
                 "min": self._low,
                 "max": self._high,
                 "seed": self._seed
             })
-
-        if var.dtype == VarDesc.VarType.FP16:
-            block.append_op(
-                type="cast",
-                inputs={"X": out_var},
-                outputs={"Out": var},
-                attrs={"in_dtype": out_var.dtype,
-                       "out_dtype": var.dtype})
-
         var.op = op
         return op
 
@@ -285,39 +261,17 @@ class NormalInitializer(Initializer):
         # Initialization Ops should be prepended and not appended
         if self._seed == 0:
             self._seed = block.program.random_seed
-
-        # to be compatible of fp16 initalizers
-        if var.dtype == VarDesc.VarType.FP16:
-            out_dtype = VarDesc.VarType.FP32
-            out_var = block.create_var(
-                name=unique_name.generate(".".join(['gaussian_random', 'tmp'])),
-                shape=var.shape,
-                dtype=out_dtype,
-                type=VarDesc.VarType.LOD_TENSOR,
-                persistable=False)
-        else:
-            out_dtype = var.dtype
-            out_var = var
-
         op = block._prepend_op(
             type="gaussian_random",
-            outputs={"Out": out_var},
+            outputs={"Out": var},
             attrs={
                 "shape": var.shape,
-                "dtype": out_dtype,
+                "dtype": int(var.dtype),
                 "mean": self._mean,
                 "std": self._std_dev,
                 "seed": self._seed,
                 "use_mkldnn": False
             })
-
-        if var.dtype == VarDesc.VarType.FP16:
-            block.append_op(
-                type="cast",
-                inputs={"X": out_var},
-                outputs={"Out": var},
-                attrs={"in_dtype": out_var.dtype,
-                       "out_dtype": var.dtype})
         var.op = op
         return op
 

python/paddle/fluid/layers/learning_rate_scheduler.py

@@ -63,18 +63,14 @@ def noam_decay(d_model, warmup_steps):
     Returns:
         The decayed learning rate.
     """
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter(1)
 
-    def _lr_schedule(dtype):
-        with default_main_program()._lr_schedule_guard():
-            global_step = _decay_step_counter(1)
+        a = global_step**-0.5
+        b = (warmup_steps**-1.5) * global_step
+        lr_value = (d_model**-0.5) * nn.elementwise_min(a, b)
 
-            a = global_step**-0.5
-            b = (warmup_steps**-1.5) * global_step
-            lr_value = (d_model**-0.5) * nn.elementwise_min(a, b)
-
-        return lr_value
-
-    return _lr_schedule
+    return lr_value
 
 
 def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
@@ -113,19 +109,15 @@ def exponential_decay(learning_rate, decay_steps, decay_rate, staircase=False):
           sgd_optimizer.minimize(avg_cost)
 
     """
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter()
 
-    def _lr_schedule(dtype):
-        with default_main_program()._lr_schedule_guard():
-            global_step = _decay_step_counter()
-
-            div_res = global_step / decay_steps
-            if staircase:
-                div_res = ops.floor(div_res)
-            decayed_lr = learning_rate * (decay_rate**div_res)
+        div_res = global_step / decay_steps
+        if staircase:
+            div_res = ops.floor(div_res)
+        decayed_lr = learning_rate * (decay_rate**div_res)
 
-            return decayed_lr
-
-    return _lr_schedule
+        return decayed_lr
 
 
 def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
@@ -146,19 +138,15 @@ def natural_exp_decay(learning_rate, decay_steps, decay_rate, staircase=False):
     Returns:
         The decayed learning rate
     """
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter()
 
-    def _lr_schedule(dtype):
-        with default_main_program()._lr_schedule_guard():
-            global_step = _decay_step_counter()
-
-            div_res = global_step / decay_steps
-            if staircase:
-                div_res = ops.floor(div_res)
-            decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res)
-
-            return decayed_lr
+        div_res = global_step / decay_steps
+        if staircase:
+            div_res = ops.floor(div_res)
+        decayed_lr = learning_rate * ops.exp(-1 * decay_rate * div_res)
 
-    return _lr_schedule
+        return decayed_lr
 
 
 def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
@@ -196,20 +184,16 @@ def inverse_time_decay(learning_rate, decay_steps, decay_rate, staircase=False):
                     staircase=True))
           sgd_optimizer.minimize(avg_cost)
     """
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter()
 
-    def _lr_schedule(dtype):
-        with default_main_program()._lr_schedule_guard():
-            global_step = _decay_step_counter()
+        div_res = global_step / decay_steps
+        if staircase:
+            div_res = ops.floor(div_res)
 
-            div_res = global_step / decay_steps
-            if staircase:
-                div_res = ops.floor(div_res)
+        decayed_lr = learning_rate / (1 + decay_rate * div_res)
 
-            decayed_lr = learning_rate / (1 + decay_rate * div_res)
-
-            return decayed_lr
-
-    return _lr_schedule
+        return decayed_lr
 
 
 def polynomial_decay(learning_rate,
@@ -240,33 +224,28 @@ def polynomial_decay(learning_rate,
     Returns:
         Variable: The decayed learning rate
     """
+    with default_main_program()._lr_schedule_guard():
+        global_step = _decay_step_counter()
 
-    def _lr_schedule(dtype, decay_steps=decay_steps):
-        with default_main_program()._lr_schedule_guard():
-            global_step = _decay_step_counter()
-
-            if cycle:
-                div_res = ops.ceil(global_step / decay_steps)
-                zero_var = tensor.fill_constant(
-                    shape=[1], dtype=dtype, value=0.0)
-                one_var = tensor.fill_constant(
-                    shape=[1], dtype=dtype, value=1.0)
-
-                with control_flow.Switch() as switch:
-                    with switch.case(global_step == zero_var):
-                        tensor.assign(input=one_var, output=div_res)
-                decay_steps = decay_steps * div_res
-            else:
-                decay_steps_var = tensor.fill_constant(
-                    shape=[1], dtype=dtype, value=float(decay_steps))
-                global_step = nn.elementwise_min(
-                    x=global_step, y=decay_steps_var)
+        if cycle:
+            div_res = ops.ceil(global_step / decay_steps)
+            zero_var = tensor.fill_constant(
+                shape=[1], dtype='float32', value=0.0)
+            one_var = tensor.fill_constant(
+                shape=[1], dtype='float32', value=1.0)
 
-            decayed_lr = (learning_rate - end_learning_rate) * \
-                ((1 - global_step / decay_steps) ** power) + end_learning_rate
-            return decayed_lr
+            with control_flow.Switch() as switch:
+                with switch.case(global_step == zero_var):
+                    tensor.assign(input=one_var, output=div_res)
+            decay_steps = decay_steps * div_res
+        else:
+            decay_steps_var = tensor.fill_constant(
+                shape=[1], dtype='float32', value=float(decay_steps))
+            global_step = nn.elementwise_min(x=global_step, y=decay_steps_var)
 
-    return _lr_schedule
+        decayed_lr = (learning_rate - end_learning_rate) * \
+            ((1 - global_step / decay_steps) ** power) + end_learning_rate
+        return decayed_lr
 
 
 def piecewise_decay(boundaries, values):
@@ -294,42 +273,38 @@ def piecewise_decay(boundaries, values):
 
 
     """
-
-    def _lr_schedule(dtype):
-        with default_main_program()._lr_schedule_guard():
-            if len(values) - len(boundaries) != 1:
-                raise ValueError("len(values) - len(boundaries) should be 1")
-
-            global_step = _decay_step_counter()
-
-            lr = tensor.create_global_var(
-                shape=[1],
-                value=0.0,
-                dtype='float32',
-                persistable=True,
-                name="learning_rate")
-
-            with control_flow.Switch() as switch:
-                for i in range(len(boundaries)):
-                    boundary_val = tensor.fill_constant(
-                        shape=[1],
-                        dtype='float32',
-                        value=float(boundaries[i]),
-                        force_cpu=True)
-                    value_var = tensor.fill_constant(
-                        shape=[1], dtype='float32', value=float(values[i]))
-                    with switch.case(global_step < boundary_val):
-                        tensor.assign(value_var, lr)
-                last_value_var = tensor.fill_constant(
+    with default_main_program()._lr_schedule_guard():
+        if len(values) - len(boundaries) != 1:
+            raise ValueError("len(values) - len(boundaries) should be 1")
+
+        global_step = _decay_step_counter()
+
+        lr = tensor.create_global_var(
+            shape=[1],
+            value=0.0,
+            dtype='float32',
+            persistable=True,
+            name="learning_rate")
+
+        with control_flow.Switch() as switch:
+            for i in range(len(boundaries)):
+                boundary_val = tensor.fill_constant(
                     shape=[1],
                     dtype='float32',
-                    value=float(values[len(values) - 1]))
-                with switch.default():
-                    tensor.assign(last_value_var, lr)
-
-        return lr
+                    value=float(boundaries[i]),
+                    force_cpu=True)
+                value_var = tensor.fill_constant(
+                    shape=[1], dtype='float32', value=float(values[i]))
+                with switch.case(global_step < boundary_val):
+                    tensor.assign(value_var, lr)
+            last_value_var = tensor.fill_constant(
+                shape=[1],
+                dtype='float32',
+                value=float(values[len(values) - 1]))
+            with switch.default():
+                tensor.assign(last_value_var, lr)
 
-    return _lr_schedule
+    return lr
 
 
 def append_LARS(params_grads, learning_rate, weight_decay):

python/paddle/fluid/layers/nn.py

@@ -2798,10 +2798,6 @@ def batch_norm(input,
     helper = LayerHelper('batch_norm', **locals())
     dtype = helper.input_dtype()
 
-    # use fp32 for bn parameter
-    if dtype == core.VarDesc.VarType.FP16:
-        dtype = core.VarDesc.VarType.FP32
-
     input_shape = input.shape
     if data_layout == 'NCHW':
         channel_num = input_shape[1]
@@ -2836,7 +2832,7 @@ def batch_norm(input,
             trainable=False,
             do_model_average=do_model_average_for_mean_and_var),
         shape=param_shape,
-        dtype=dtype)
+        dtype=input.dtype)
     mean.stop_gradient = True
 
     variance = helper.create_parameter(
@@ -2846,7 +2842,7 @@ def batch_norm(input,
             trainable=False,
             do_model_average=do_model_average_for_mean_and_var),
         shape=param_shape,
-        dtype=dtype)
+        dtype=input.dtype)
     variance.stop_gradient = True
 
     # create output

python/paddle/fluid/optimizer.py

@@ -50,21 +50,17 @@ class Optimizer(object):
 
     def __init__(self, learning_rate, regularization=None, name=None):
         if not isinstance(learning_rate, float) and \
-                not isinstance(learning_rate, framework.Variable) and \
-                not callable(learning_rate):
-            raise TypeError(
-                "learning rate should be float or Variable or callable(dtype)")
+                not isinstance(learning_rate, framework.Variable):
+            raise TypeError("learning rate should be float or Variable")
         self._name = name
         self.regularization = regularization
         self._learning_rate = learning_rate
         # the learning rate type should be inferenced from loss
         self._dtype = None
         # each program should have a independent learning rate
-        # program -> Variable(learning_rate) or:
-        # program -> callable(return learning_rate Variable)
+        # program -> Variable(learning_rate)
         self._learning_rate_map = dict()
-        if isinstance(self._learning_rate, framework.Variable) or \
-            callable(self._learning_rate):
+        if isinstance(self._learning_rate, framework.Variable):
             self._learning_rate_map[framework.default_main_program(
             )] = self._learning_rate
         # Dictionary of accumulators. Some optimizer subclasses need to
@@ -79,11 +75,6 @@ class Optimizer(object):
 
         if isinstance(lr, framework.Variable):
             return
-        elif callable(lr):
-            dtype = 'float32' if self._dtype is None else self._dtype
-            self._learning_rate_map[framework.default_main_program()] = lr(
-                dtype)
-            return
         else:
             if not isinstance(self._learning_rate, float):
                 raise TypeError(

python/paddle/fluid/tests/unittests/op_test.py

@@ -368,8 +368,6 @@ class OpTest(unittest.TestCase):
                 place = core.CUDAPlace(0)
                 if core.is_float16_supported(place):
                     return [place]
-                else:
-                    return []
             else:
                 return []
         places = [fluid.CPUPlace()]

python/paddle/fluid/tests/unittests/test_accuracy_op.py

@@ -22,10 +22,8 @@ from op_test import OpTest
 class TestAccuracyOp(OpTest):
     def setUp(self):
         self.op_type = "accuracy"
-        self.dtype = np.float32
-        self.init_dtype()
         n = 8192
-        infer = np.random.random((n, 1)).astype(self.dtype)
+        infer = np.random.random((n, 1)).astype("float32")
         indices = np.random.randint(0, 2, (n, 1))
         label = np.random.randint(0, 2, (n, 1))
         self.inputs = {'Out': infer, 'Indices': indices, "Label": label}
@@ -36,25 +34,14 @@ class TestAccuracyOp(OpTest):
                     num_correct += 1
                     break
         self.outputs = {
-            'Accuracy': np.array([num_correct / float(n)]).astype(self.dtype),
+            'Accuracy': np.array([num_correct / float(n)]).astype("float32"),
             'Correct': np.array([num_correct]).astype("int32"),
             'Total': np.array([n]).astype("int32")
         }
 
-    def init_dtype(self):
-        pass
-
     def test_check_output(self):
         self.check_output()
 
 
-class TestAccuracyOpFp16(TestAccuracyOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-    def test_check_output(self):
-        self.check_output(atol=1e-3)
-
-
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_elementwise_div_op.py

@@ -21,16 +21,14 @@ from op_test import OpTest
 class ElementwiseDivOp(OpTest):
     def setUp(self):
         self.op_type = "elementwise_div"
-        self.dtype = np.float32
-        self.init_dtype()
         """ Warning
         CPU gradient check error!
         'X': np.random.random((32,84)).astype("float32"),
         'Y': np.random.random((32,84)).astype("float32")
         """
         self.inputs = {
-            'X': np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype),
-            'Y': np.random.uniform(0.1, 1, [13, 17]).astype(self.dtype)
+            'X': np.random.uniform(0.1, 1, [13, 17]).astype("float32"),
+            'Y': np.random.uniform(0.1, 1, [13, 17]).astype("float32")
         }
         self.outputs = {'Out': np.divide(self.inputs['X'], self.inputs['Y'])}
 
@@ -48,9 +46,6 @@ class ElementwiseDivOp(OpTest):
         self.check_grad(
             ['X'], 'Out', max_relative_error=0.05, no_grad_set=set('Y'))
 
-    def init_dtype(self):
-        pass
-
 
 class TestElementwiseDivOp_scalar(ElementwiseDivOp):
     def setUp(self):
@@ -131,21 +126,5 @@ class TestElementwiseDivOp_broadcast_3(ElementwiseDivOp):
         }
 
 
-class TestElementwiseDivOpFp16(ElementwiseDivOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-    def test_check_grad_normal(self):
-        self.check_grad(['X', 'Y'], 'Out', max_relative_error=1)
-
-    def test_check_grad_ingore_x(self):
-        self.check_grad(
-            ['Y'], 'Out', max_relative_error=1, no_grad_set=set("X"))
-
-    def test_check_grad_ingore_y(self):
-        self.check_grad(
-            ['X'], 'Out', max_relative_error=1, no_grad_set=set('Y'))
-
-
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py

@@ -135,10 +135,5 @@ class TestElementwiseMulOp_broadcast_3(ElementwiseMulOp):
         }
 
 
-class TestElementwiseMulOpFp16(ElementwiseMulOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_fill_zeros_like_op.py

@@ -22,22 +22,12 @@ from op_test import OpTest
 class TestFillZerosLikeOp(OpTest):
     def setUp(self):
         self.op_type = "fill_zeros_like"
-        self.dtype = np.float32
-        self.init_dtype()
-        self.inputs = {'X': np.random.random((219, 232)).astype(self.dtype)}
+        self.inputs = {'X': np.random.random((219, 232)).astype("float32")}
         self.outputs = {'Out': np.zeros_like(self.inputs["X"])}
 
-    def init_dtype(self):
-        pass
-
     def test_check_output(self):
         self.check_output()
 
 
-class TestFillZerosLikeOpFp16(TestFillZerosLikeOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-
 if __name__ == "__main__":
     unittest.main()

python/paddle/fluid/tests/unittests/test_learning_rate_scheduler.py

@@ -97,7 +97,7 @@ class TestLearningRateDecay(unittest.TestCase):
         startup_prog = fluid.Program()
 
         with fluid.program_guard(main_prog, startup_prog):
-            decayed_lr = fluid_decay_fn(**kwargs)("float32")
+            decayed_lr = fluid_decay_fn(**kwargs)
 
         place = fluid.CPUPlace()
         exe = fluid.Executor(place)

python/paddle/fluid/tests/unittests/test_momentum_op.py

@@ -24,13 +24,11 @@ from op_test import OpTest
 class TestMomentumOp1(OpTest):
     def setUp(self):
         self.op_type = "momentum"
-        self.dtype = np.float32
-        self.init_dtype()
 
-        param = np.random.random((123, 321)).astype(self.dtype)
-        grad = np.random.random((123, 321)).astype(self.dtype)
-        velocity = np.zeros((123, 321)).astype(self.dtype)
-        learning_rate = np.array([0.001]).astype(self.dtype)
+        param = np.random.random((123, 321)).astype("float32")
+        grad = np.random.random((123, 321)).astype("float32")
+        velocity = np.zeros((123, 321)).astype("float32")
+        learning_rate = np.array([0.001]).astype("float32")
         mu = 0.0001
         use_nesterov = False
 
@@ -52,21 +50,10 @@ class TestMomentumOp1(OpTest):
 
         self.outputs = {'ParamOut': param_out, 'VelocityOut': velocity_out}
 
-    def init_dtype(self):
-        pass
-
     def test_check_output(self):
         self.check_output()
 
 
-class TestMomentumOpFp16(TestMomentumOp1):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-    def test_check_output(self):
-        self.check_output(atol=1e-3)
-
-
 class TestMomentumOp2(OpTest):
     '''Test Momentum with default values for attributes
     '''

python/paddle/fluid/tests/unittests/test_top_k_op.py

@@ -23,11 +23,8 @@ class TestTopkOp(OpTest):
     def setUp(self):
         self.set_args()
         self.op_type = "top_k"
-        self.dtype = np.float32
-        self.init_dtype()
-
         k = self.top_k
-        input = np.random.random((self.row, k)).astype(self.dtype)
+        input = np.random.random((self.row, k)).astype("float32")
         output = np.ndarray((self.row, k))
         indices = np.ndarray((self.row, k)).astype("int64")
 
@@ -41,9 +38,6 @@ class TestTopkOp(OpTest):
 
         self.outputs = {'Out': output, 'Indices': indices}
 
-    def init_dtype(self):
-        pass
-
     def set_args(self):
         self.row = 32
         self.top_k = 1
@@ -52,11 +46,6 @@ class TestTopkOp(OpTest):
         self.check_output()
 
 
-class TestTopkOpFp16(TestTopkOp):
-    def init_dtype(self):
-        self.dtype = np.float16
-
-
 class TestTopkOp3d(OpTest):
     def setUp(self):
         self.op_type = "top_k"