m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Add fp16 backward support (#14202)

Browse Source 
* add fp16 backward support
test=develop

* add sum_op fp16 test

* disable test_dist_save_load
test=develop

* add check_grad for sum

* add unit test for softmax_grad fp16
test=develop

* add scale_op unit test

* add mul_grad_op unit test for fp16

* add cross_entropy_grad and eman_grad unit test for fp16
test=develop

* fix cross_entropy unit test

* add pool2d fp16 unit test

* refine conv2d fp16 unit test
test=develop

* refine activation unit test
test=develop

* fix ci
test=develop

* follow zhihong's comment, copy from https://github.com/PaddlePaddle/Paddle/pull/12796
test=develop
revert-14324-fix_vlog
chengduo 7 years ago committed by GitHub
parent 0953cd3e16
commit a9b5d42dd4
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
31 changed files with 767 additions and 961 deletions
Show Stats Download Patch File Download Diff File

4
paddle/fluid/operators/activation_op.cu
Unescape View File

@ -26,6 +26,8 @@ namespace plat = paddle::platform;
act_type##_grad, ops::ActivationGradKernel<plat::CUDADeviceContext, \
ops::grad_functor<float>>, \
ops::ActivationGradKernel<plat::CUDADeviceContext, \
ops::grad_functor<double>>);
ops::grad_functor<double>>, \
ops::ActivationGradKernel<plat::CUDADeviceContext, \
ops::grad_functor<plat::float16>>);
FOR_EACH_KERNEL_FUNCTOR(REGISTER_ACTIVATION_CUDA_KERNEL);

5
paddle/fluid/operators/activation_op.h
Unescape View File

@ -333,8 +333,7 @@ struct SqrtGradFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out, typename dOut,
typename dX>
void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
const Out out_conj = Eigen::numext::conj(out);
dx.device(d) = static_cast<T>(0.5) * dout / out_conj;
dx.device(d) = static_cast<T>(0.5) * dout / out;
}
};
@ -740,7 +739,7 @@ struct PowGradFunctor : public BaseActivationFunctor<T> {
typename dX>
void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
dx.device(d) = dout * static_cast<T>(factor) *
x.pow(static_cast<T>(factor - static_cast<T>(1)));
x.pow(static_cast<T>(factor) - static_cast<T>(1));
}
};

21
paddle/fluid/operators/batch_norm_op.cu.cc
Unescape View File

@ -219,8 +219,8 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
auto *d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
d_x->mutable_data<T>(ctx.GetPlace());
d_scale->mutable_data<T>(ctx.GetPlace());
d_bias->mutable_data<T>(ctx.GetPlace());
d_scale->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
d_bias->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
if ((N * H * W * D) == 1) {
@ -272,8 +272,10 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
const auto *saved_mean = ctx.Input<Tensor>("SavedMean");
const auto *saved_var = ctx.Input<Tensor>("SavedVariance");
const void *saved_mean_data = saved_mean->template data<T>();
const void *saved_var_data = saved_var->template data<T>();
const void *saved_mean_data =
saved_mean->template data<BatchNormParamType<T>>();
const void *saved_var_data =
saved_var->template data<BatchNormParamType<T>>();
CUDNN_ENFORCE(platform::dynload::cudnnBatchNormalizationBackward(
dev_ctx.cudnn_handle(), mode_, CudnnDataType<T>::kOne(),
@ -281,10 +283,10 @@ class BatchNormGradKernel<platform::CUDADeviceContext, T>
CudnnDataType<T>::kZero(), data_desc_, x->template data<T>(),
data_desc_, d_y->template data<T>(), data_desc_,
d_x->template mutable_data<T>(ctx.GetPlace()), bn_param_desc_,
scale->template data<T>(),
d_scale->template mutable_data<T>(ctx.GetPlace()),
d_bias->template mutable_data<T>(ctx.GetPlace()), epsilon,
saved_mean_data, saved_var_data));
scale->template data<BatchNormParamType<T>>(),
d_scale->template mutable_data<BatchNormParamType<T>>(ctx.GetPlace()),
d_bias->template mutable_data<BatchNormParamType<T>>(ctx.GetPlace()),
epsilon, saved_mean_data, saved_var_data));
// clean when exit.
CUDNN_ENFORCE(platform::dynload::cudnnDestroyTensorDescriptor(data_desc_));
@ -304,4 +306,5 @@ REGISTER_OP_CUDA_KERNEL(
ops::BatchNormKernel<plat::CUDADeviceContext, plat::float16>);
REGISTER_OP_CUDA_KERNEL(
batch_norm_grad, ops::BatchNormGradKernel<plat::CUDADeviceContext, float>,
ops::BatchNormGradKernel<plat::CUDADeviceContext, double>);
ops::BatchNormGradKernel<plat::CUDADeviceContext, double>,
ops::BatchNormGradKernel<plat::CUDADeviceContext, plat::float16>);

5
paddle/fluid/operators/conv_cudnn_op.cu.cc
Unescape View File

@ -143,9 +143,11 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
cudnn_conv_desc, CUDNN_TENSOR_OP_MATH));
// Currently tensor core is only enabled using this algo
algo = CUDNN_CONVOLUTION_FWD_ALGO_IMPLICIT_PRECOMP_GEMM;
VLOG(5) << "use cudnn_tensor_op_math";
} else {
CUDNN_ENFORCE(platform::dynload::cudnnSetConvolutionMathType(
cudnn_conv_desc, CUDNN_DEFAULT_MATH));
VLOG(5) << "NOT use cudnn_tensor_op_math";
}
#endif
@ -361,7 +363,8 @@ REGISTER_OP_KERNEL(conv2d, CUDNN, plat::CUDAPlace,
paddle::operators::CUDNNConvOpKernel<plat::float16>);
REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace,
paddle::operators::CUDNNConvGradOpKernel<float>,
paddle::operators::CUDNNConvGradOpKernel<double>);
paddle::operators::CUDNNConvGradOpKernel<double>,
paddle::operators::CUDNNConvGradOpKernel<plat::float16>);
REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace,
paddle::operators::CUDNNConvOpKernel<float>,

13
paddle/fluid/operators/cross_entropy_op.cu
Unescape View File

@ -13,12 +13,17 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/cross_entropy_op.h"
#include "paddle/fluid/platform/float16.h"
namespace plat = paddle::platform;
namespace ops = paddle::operators;
using CUDACtx = paddle::platform::CUDADeviceContext;
REGISTER_OP_CUDA_KERNEL(cross_entropy,
ops::CrossEntropyOpKernel<CUDACtx, float>,
ops::CrossEntropyOpKernel<CUDACtx, double>);
REGISTER_OP_CUDA_KERNEL(cross_entropy_grad,
ops::CrossEntropyGradientOpKernel<CUDACtx, float>,
ops::CrossEntropyGradientOpKernel<CUDACtx, double>);
ops::CrossEntropyOpKernel<CUDACtx, double>,
ops::CrossEntropyOpKernel<CUDACtx, plat::float16>);
REGISTER_OP_CUDA_KERNEL(
cross_entropy_grad, ops::CrossEntropyGradientOpKernel<CUDACtx, float>,
ops::CrossEntropyGradientOpKernel<CUDACtx, double>,
ops::CrossEntropyGradientOpKernel<CUDACtx, plat::float16>);

3
paddle/fluid/operators/elementwise_add_op.cu
Unescape View File

@ -30,4 +30,5 @@ REGISTER_OP_CUDA_KERNEL(
ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, float>,
ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, double>,
ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int>,
ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int64_t>);
ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, int64_t>,
ops::ElementwiseAddGradKernel<plat::CUDADeviceContext, plat::float16>);

4
paddle/fluid/operators/elementwise_op_function.h
Unescape View File

@ -365,7 +365,7 @@ static __global__ void ElemwiseGradBroadcast1CUDAKernel(
int j = blockIdx.x;
int i = threadIdx.x;
int tid = threadIdx.x;
T val = 0;
T val(0);
do {
int x_offset = i * w + j;
@ -433,7 +433,7 @@ static __global__ void ElemwiseGradBroadcast2CUDAKernel(
int tid = threadIdx.x;
int j = blockIdx.x;
T val = 0;
T val(0);
int ttid = tid;
while (true) {

22
paddle/fluid/operators/math/cross_entropy.cu
Unescape View File

@ -21,6 +21,16 @@ namespace operators {
namespace math {
namespace {
__device__ __forceinline__ float real_log(float x) { return logf(x); }
__device__ __forceinline__ double real_log(double x) { return log(x); }
__device__ __forceinline__ platform::float16 real_log(
const platform::float16& val) {
return static_cast<platform::float16>(hlog(static_cast<half>(val)));
}
template <typename T>
__global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label,
const int N, const int D,
@ -29,8 +39,8 @@ __global__ void CrossEntropyKernel(T* Y, const T* X, const int64_t* label,
i += blockDim.x * gridDim.x) {
PADDLE_ASSERT(label[i] >= 0 && label[i] < D || label[i] == ignore_index);
Y[i] = ignore_index == label[i]
? 0
: -math::TolerableValue<T>()(log(X[i * D + label[i]]));
? static_cast<T>(0)
: -math::TolerableValue<T>()(real_log(X[i * D + label[i]]));
}
}
@ -38,12 +48,12 @@ template <typename T>
__global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label,
const int class_num) {
int tid = threadIdx.x;
T val = 0;
T val(0);
int idx = blockIdx.x * class_num + tid;
int end = blockIdx.x * class_num + class_num;
for (; idx < end; idx += blockDim.x) {
val += math::TolerableValue<T>()(std::log(X[idx])) * label[idx];
val += math::TolerableValue<T>()(real_log(X[idx])) * label[idx];
}
val = paddle::platform::reduceSum(val, tid, blockDim.x);
@ -53,8 +63,6 @@ __global__ void SoftCrossEntropyKernel(T* Y, const T* X, const T* label,
}
} // namespace
using Tensor = framework::Tensor;
template <typename T>
class CrossEntropyFunctor<platform::CUDADeviceContext, T> {
public:
@ -89,6 +97,8 @@ class CrossEntropyFunctor<platform::CUDADeviceContext, T> {
template class CrossEntropyFunctor<platform::CUDADeviceContext, float>;
template class CrossEntropyFunctor<platform::CUDADeviceContext, double>;
template class CrossEntropyFunctor<platform::CUDADeviceContext,
platform::float16>;
} // namespace math
} // namespace operators
} // namespace paddle

21
paddle/fluid/operators/math/cross_entropy.h
Unescape View File

@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <limits>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/hostdevice.h"
@ -33,6 +34,26 @@ struct TolerableValue {
}
};
// NOTE(dzh): float16 value clip behave different.
// 1. Our ValueClipping has a hardcore threshold 1e20
// for float number. 1e20 will resulting in overflow in float16.
// 2. float16 should expose the the real number overflow to python.
// because mixed-training depends the inf/nan value to determine
// if the scale value will be adjusted.
// Also. In standard implementation of cross entropy, other
// framework not has the ValueClipping.
template <>
struct TolerableValue<platform::float16> {
HOSTDEVICE platform::float16 operator()(const platform::float16& x) const {
if (platform::isfinite(x))
return x;
else if (x > static_cast<platform::float16>(0))
return std::numeric_limits<platform::float16>::max();
else
return std::numeric_limits<platform::float16>::min();
}
};
template <typename DeviceContext, typename T>
class CrossEntropyFunctor {
public:

15
paddle/fluid/operators/math/selected_rows_functor.cu
Unescape View File

@ -18,6 +18,7 @@ limitations under the License. */
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/platform/cuda_primitives.h"
#include "paddle/fluid/platform/float16.h"
namespace paddle {
namespace operators {
@ -118,7 +119,7 @@ struct SelectedRowsAddTensor<platform::CUDADeviceContext, T> {
auto* out_data = output->data<T>();
SetConstant<platform::CUDADeviceContext, T> functor;
functor(context, output, 0.0);
functor(context, output, static_cast<T>(0));
const int block_size = 256;
dim3 threads(block_size, 1);
@ -136,6 +137,9 @@ struct SelectedRowsAddTensor<platform::CUDADeviceContext, T> {
template struct SelectedRowsAddTensor<platform::CUDADeviceContext, float>;
template struct SelectedRowsAddTensor<platform::CUDADeviceContext, double>;
template struct SelectedRowsAdd<platform::CUDADeviceContext, platform::float16>;
template struct SelectedRowsAddTensor<platform::CUDADeviceContext,
platform::float16>;
template <typename T>
struct SelectedRowsAddTo<platform::CUDADeviceContext, T> {
@ -175,6 +179,8 @@ template struct SelectedRowsAddTo<platform::CUDADeviceContext, float>;
template struct SelectedRowsAddTo<platform::CUDADeviceContext, double>;
template struct SelectedRowsAddTo<platform::CUDADeviceContext, int>;
template struct SelectedRowsAddTo<platform::CUDADeviceContext, int64_t>;
template struct SelectedRowsAddTo<platform::CUDADeviceContext,
platform::float16>;
namespace {
template <typename T, int block_size>
@ -227,6 +233,8 @@ template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, float>;
template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, double>;
template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, int>;
template struct SelectedRowsAddToTensor<platform::CUDADeviceContext, int64_t>;
template struct SelectedRowsAddToTensor<platform::CUDADeviceContext,
platform::float16>;
namespace scatter {
@ -287,7 +295,7 @@ struct MergeAdd<platform::CUDADeviceContext, T> {
context.GetPlace());
math::SetConstant<platform::CUDADeviceContext, T> constant_functor;
constant_functor(context, out.mutable_value(), 0.0);
constant_functor(context, out.mutable_value(), static_cast<T>(0));
auto* out_data = out.mutable_value()->data<T>();
auto* input_data = input.value().data<T>();
@ -347,7 +355,7 @@ struct MergeAdd<platform::CUDADeviceContext, T> {
context.GetPlace());
math::SetConstant<platform::CUDADeviceContext, T> constant_functor;
constant_functor(context, out.mutable_value(), 0.0);
constant_functor(context, out.mutable_value(), static_cast<T>(0));
auto* out_data = out.mutable_value()->data<T>();
@ -374,6 +382,7 @@ template struct MergeAdd<platform::CUDADeviceContext, float>;
template struct MergeAdd<platform::CUDADeviceContext, double>;
template struct MergeAdd<platform::CUDADeviceContext, int>;
template struct MergeAdd<platform::CUDADeviceContext, int64_t>;
template struct MergeAdd<platform::CUDADeviceContext, platform::float16>;
template <typename T, int block_size>
__global__ void UpdateToTensorKernel(const T* selected_rows,

3
paddle/fluid/operators/math/softmax.cu
Unescape View File

@ -96,12 +96,15 @@ template class SoftmaxCUDNNFunctor<float>;
template class SoftmaxCUDNNFunctor<double>;
template class SoftmaxGradCUDNNFunctor<float>;
template class SoftmaxGradCUDNNFunctor<double>;
template class SoftmaxGradCUDNNFunctor<platform::float16>;
template class SoftmaxFunctor<platform::CUDADeviceContext, platform::float16>;
template class SoftmaxFunctor<platform::CUDADeviceContext, float>;
template class SoftmaxFunctor<platform::CUDADeviceContext, double>;
template class SoftmaxGradFunctor<platform::CUDADeviceContext, float>;
template class SoftmaxGradFunctor<platform::CUDADeviceContext, double>;
template class SoftmaxGradFunctor<platform::CUDADeviceContext,
platform::float16>;
} // namespace math
} // namespace operators

8
paddle/fluid/operators/mean_op.cu
Unescape View File

@ -15,11 +15,15 @@ limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/fluid/operators/mean_op.h"
#include "paddle/fluid/platform/float16.h"
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_CUDA_KERNEL(
mean, ops::MeanKernel<paddle::platform::CUDADeviceContext, float>,
ops::MeanKernel<paddle::platform::CUDADeviceContext, double>);
ops::MeanKernel<paddle::platform::CUDADeviceContext, double>,
ops::MeanKernel<paddle::platform::CUDADeviceContext, plat::float16>);
REGISTER_OP_CUDA_KERNEL(
mean_grad, ops::MeanGradKernel<paddle::platform::CUDADeviceContext, float>,
ops::MeanGradKernel<paddle::platform::CUDADeviceContext, double>);
ops::MeanGradKernel<paddle::platform::CUDADeviceContext, double>,
ops::MeanGradKernel<paddle::platform::CUDADeviceContext, plat::float16>);

3
paddle/fluid/operators/mean_op.h
Unescape View File

@ -55,8 +55,7 @@ class MeanGradKernel : public framework::OpKernel<T> {
IG->mutable_data<T>(context.GetPlace());
T ig_size = static_cast<T>(IG->numel());
Eigen::DSizes<int, 1> bcast(ig_size);
Eigen::DSizes<int, 1> bcast(static_cast<int>(ig_size));
EigenVector<T>::Flatten(*IG).device(
*context.template device_context<DeviceContext>().eigen_device()) =
(EigenVector<T>::From(*OG) / ig_size).broadcast(bcast);

7
paddle/fluid/operators/mul_op.cu.cc
Unescape View File

@ -20,6 +20,7 @@ namespace plat = paddle::platform;
REGISTER_OP_CUDA_KERNEL(mul, ops::MulKernel<plat::CUDADeviceContext, float>,
ops::MulKernel<plat::CUDADeviceContext, double>,
ops::MulKernel<plat::CUDADeviceContext, plat::float16>);
REGISTER_OP_CUDA_KERNEL(mul_grad,
ops::MulGradKernel<plat::CUDADeviceContext, float>,
ops::MulGradKernel<plat::CUDADeviceContext, double>);
REGISTER_OP_CUDA_KERNEL(
mul_grad, ops::MulGradKernel<plat::CUDADeviceContext, float>,
ops::MulGradKernel<plat::CUDADeviceContext, double>,
ops::MulGradKernel<plat::CUDADeviceContext, plat::float16>);

3
paddle/fluid/operators/pool_cudnn_op.cu.cc
Unescape View File

@ -178,7 +178,8 @@ REGISTER_OP_KERNEL(pool2d, CUDNN, plat::CUDAPlace,
ops::PoolCUDNNOpKernel<plat::float16>);
REGISTER_OP_KERNEL(pool2d_grad, CUDNN, plat::CUDAPlace,
ops::PoolCUDNNGradOpKernel<float>,
ops::PoolCUDNNGradOpKernel<double>);
ops::PoolCUDNNGradOpKernel<double>,
ops::PoolCUDNNGradOpKernel<plat::float16>);
REGISTER_OP_KERNEL(pool3d, CUDNN, plat::CUDAPlace,
ops::PoolCUDNNOpKernel<float>,

6
paddle/fluid/operators/scale_op.cu
Unescape View File

@ -13,6 +13,8 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/scale_op.h"
#include "paddle/fluid/platform/float16.h"
namespace plat = paddle::platform;
REGISTER_OP_CUDA_KERNEL(
scale,
@ -20,4 +22,6 @@ REGISTER_OP_CUDA_KERNEL(
paddle::operators::ScaleKernel<paddle::platform::CUDADeviceContext, double>,
paddle::operators::ScaleKernel<paddle::platform::CUDADeviceContext, int>,
paddle::operators::ScaleKernel<paddle::platform::CUDADeviceContext,
int64_t>);
int64_t>,
paddle::operators::ScaleKernel<paddle::platform::CUDADeviceContext,
plat::float16>);

3
paddle/fluid/operators/softmax_cudnn_op.cu.cc
Unescape View File

@ -80,4 +80,5 @@ REGISTER_OP_KERNEL(softmax, CUDNN, plat::CUDAPlace,
ops::SoftmaxCUDNNKernel<plat::float16>);
REGISTER_OP_KERNEL(softmax_grad, CUDNN, plat::CUDAPlace,
ops::SoftmaxGradCUDNNKernel<float>,
ops::SoftmaxGradCUDNNKernel<double>);
ops::SoftmaxGradCUDNNKernel<double>,
ops::SoftmaxGradCUDNNKernel<plat::float16>);

3
paddle/fluid/operators/softmax_op.cu.cc
Unescape View File

@ -23,4 +23,5 @@ REGISTER_OP_CUDA_KERNEL(
ops::SoftmaxKernel<plat::CUDADeviceContext, plat::float16>);
REGISTER_OP_CUDA_KERNEL(
softmax_grad, ops::SoftmaxGradKernel<plat::CUDADeviceContext, float>,
ops::SoftmaxGradKernel<plat::CUDADeviceContext, double>);
ops::SoftmaxGradKernel<plat::CUDADeviceContext, double>,
ops::SoftmaxGradKernel<plat::CUDADeviceContext, plat::float16>);

5
paddle/fluid/operators/sum_op.cu
Unescape View File

@ -11,10 +11,13 @@ limitations under the License. */
#define EIGEN_USE_GPU
#include "paddle/fluid/operators/sum_op.h"
#include "paddle/fluid/platform/float16.h"
namespace ops = paddle::operators;
namespace plat = paddle::platform;
REGISTER_OP_CUDA_KERNEL(
sum, ops::SumKernel<paddle::platform::CUDADeviceContext, float>,
ops::SumKernel<paddle::platform::CUDADeviceContext, double>,
ops::SumKernel<paddle::platform::CUDADeviceContext, int>,
ops::SumKernel<paddle::platform::CUDADeviceContext, int64_t>);
ops::SumKernel<paddle::platform::CUDADeviceContext, int64_t>,
ops::SumKernel<paddle::platform::CUDADeviceContext, plat::float16>);

2
paddle/fluid/operators/sum_op.h
Unescape View File

@ -61,7 +61,7 @@ class SumKernel : public framework::OpKernel<T> {
if (start != 2) {
math::SetConstant<DeviceContext, T> constant_functor;
constant_functor(context.template device_context<DeviceContext>(),
out, 0.0);
out, static_cast<T>(0));
}
}

17
python/paddle/fluid/tests/unittests/op_test.py
Unescape View File

@ -54,14 +54,6 @@ def get_numeric_gradient(place,
def product(dim):
return six.moves.reduce(lambda a, b: a * b, dim, 1)
def get_output():
sum = []
op.run(scope, place)
for output_name in output_names:
sum.append(
np.array(scope.find_var(output_name).get_tensor()).mean())
return np.array(sum).sum() / len(output_names)
tensor_to_check = scope.find_var(input_to_check).get_tensor()
tensor_size = product(tensor_to_check.shape())
tensor_to_check_dtype = tensor_to_check._dtype()
@ -77,6 +69,15 @@ def get_numeric_gradient(place,
raise ValueError("Not supported data type " + str(
tensor_to_check_dtype))
def get_output():
sum = []
op.run(scope, place)
for output_name in output_names:
sum.append(
np.array(scope.find_var(output_name).get_tensor()).astype(
tensor_to_check_dtype).mean())
return tensor_to_check_dtype(np.array(sum).sum() / len(output_names))
gradient_flat = np.zeros(shape=(tensor_size, ), dtype=tensor_to_check_dtype)
def __get_elem__(tensor, i):

627
python/paddle/fluid/tests/unittests/test_activation_op.py
View File

File diff suppressed because it is too large Load Diff

151
python/paddle/fluid/tests/unittests/test_conv2d_op.py
Unescape View File

@ -223,106 +223,81 @@ class TestWithInput1x1Filter1x1(TestConv2dOp):
#----------------Conv2dCUDNN----------------
class TestCUDNN(TestConv2dOp):
def init_kernel_type(self):
self.use_cudnn = True
class TestFP16CUDNN(TestConv2dOp):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_output_with_place(place, atol=2e-2)
def create_test_cudnn_class(parent, cls_name):
@unittest.skipIf(not core.is_compiled_with_cuda(),
"core is not compiled with CUDA")
class TestCUDNNCase(parent):
def init_kernel_type(self):
self.use_cudnn = True
cls_name = "{0}".format(cls_name)
TestCUDNNCase.__name__ = cls_name
globals()[cls_name] = TestCUDNNCase
class TestCUDNNWithPad(TestWithPad):
def init_kernel_type(self):
self.use_cudnn = True
class TestFP16CUDNNWithPad(TestWithPad):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_output_with_place(place, atol=2e-2)
class TestCUDNNWithStride(TestWithStride):
def init_kernel_type(self):
self.use_cudnn = True
class TestFP16CUDNNWithStride(TestWithStride):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_output_with_place(place, atol=2e-2)
create_test_cudnn_class(TestConv2dOp, "TestPool2DCUDNNOp")
create_test_cudnn_class(TestWithPad, "TestPool2DCUDNNOpCase1")
create_test_cudnn_class(TestWithStride, "TestPool2DCUDNNOpCase2")
create_test_cudnn_class(TestWithGroup, "TestPool2DCUDNNOpCase3")
create_test_cudnn_class(TestWith1x1, "TestPool2DCUDNNOpCase4")
create_test_cudnn_class(TestWithInput1x1Filter1x1, "TestPool2DCUDNNOpCase4")
class TestCUDNNWithGroup(TestWithGroup):
def init_kernel_type(self):
self.use_cudnn = True
class TestFP16CUDNNWithGroup(TestWithGroup):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_output_with_place(place, atol=2e-2)
#----------------Conv2dCUDNN----------------
class TestCUDNNWith1x1(TestWith1x1):
def init_kernel_type(self):
self.use_cudnn = True
def create_test_cudnn_fp16_class(parent, cls_name, grad_check=True):
@unittest.skipIf(not core.is_compiled_with_cuda(),
"core is not compiled with CUDA")
class TestConv2DCUDNNFp16(parent):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
class TestFP16CUDNNWith1x1(TestWith1x1):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
if core.is_compiled_with_cuda():
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_output_with_place(place, atol=2e-2)
def test_check_output(self):
if core.is_compiled_with_cuda():
def test_check_grad_no_filter(self):
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_output_with_place(place, atol=2e-2)
class TestCUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
def init_kernel_type(self):
self.use_cudnn = True
class TestFP16CUDNNWithInput1x1Filter1x1(TestWithInput1x1Filter1x1):
def init_kernel_type(self):
self.use_cudnn = True
self.dtype = np.float16
def test_check_output(self):
if core.is_compiled_with_cuda():
if core.is_float16_supported(place) and grad_check:
self.check_grad_with_place(
place, ['Input'],
'Output',
max_relative_error=0.02,
no_grad_set=set(['Filter']))
def test_check_grad_no_input(self):
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_output_with_place(place, atol=2e-2)
if core.is_float16_supported(place) and grad_check:
self.check_grad_with_place(
place, ['Filter'],
'Output',
max_relative_error=0.02,
no_grad_set=set(['Input']))
cls_name = "{0}".format(cls_name)
TestConv2DCUDNNFp16.__name__ = cls_name
globals()[cls_name] = TestConv2DCUDNNFp16
create_test_cudnn_fp16_class(
TestConv2dOp, "TestPool2DCUDNNFp16Op", grad_check=False)
create_test_cudnn_fp16_class(
TestWithPad, "TestPool2DCUDNNFp16OpCase1", grad_check=False)
create_test_cudnn_fp16_class(
TestWithStride, "TestPool2DCUDNNFp16OpCase2", grad_check=False)
create_test_cudnn_fp16_class(
TestWithGroup, "TestPool2DCUDNNFp16OpCase3", grad_check=False)
create_test_cudnn_fp16_class(
TestWith1x1, "TestPool2DCUDNNFp16OpCase4", grad_check=False)
create_test_cudnn_fp16_class(
TestWithInput1x1Filter1x1, "TestPool2DCUDNNFp16OpCase4", grad_check=False)
# -------TestDepthwiseConv
class TestDepthwiseConv(TestConv2dOp):

338
python/paddle/fluid/tests/unittests/test_cross_entropy_op.py
View File

File diff suppressed because it is too large Load Diff

26
python/paddle/fluid/tests/unittests/test_mean_op.py
Unescape View File

@ -17,14 +17,20 @@ from __future__ import print_function
import unittest
import numpy as np
from op_test import OpTest
import paddle.fluid.core as core
class TestMeanOp(OpTest):
def setUp(self):
self.op_type = "mean"
self.inputs = {'X': np.random.random((10, 10)).astype("float32")}
self.dtype = np.float32
self.init_dtype_type()
self.inputs = {'X': np.random.random((10, 10)).astype(self.dtype)}
self.outputs = {'Out': np.mean(self.inputs["X"])}
def init_dtype_type(self):
pass
def test_check_output(self):
self.check_output()
@ -32,5 +38,23 @@ class TestMeanOp(OpTest):
self.check_grad(['X'], 'Out')
@unittest.skipIf(not core.is_compiled_with_cuda(),
"core is not compiled with CUDA")
class TestFP16MeanOp(TestMeanOp):
def init_dtype_type(self):
self.dtype = np.float16
def test_check_output(self):
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_output_with_place(place, atol=2e-3)
def test_checkout_grad(self):
place = core.CUDAPlace(0)
if core.is_float16_supported(place):
self.check_grad_with_place(
place, ['X'], 'Out', max_relative_error=0.8)
if __name__ == "__main__":
unittest.main()

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save