/usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators
Edit: /usr/local/lib64/python3.6/site-packages/torch/include/caffe2/operators/order_switch_ops.h (2149B)
#ifndef CAFFE2_OPERATORS_ORDER_SWITCH_OPS_H_
#define CAFFE2_OPERATORS_ORDER_SWITCH_OPS_H_
#include
#include "caffe2/core/operator.h"
#include "caffe2/utils/math.h"
namespace caffe2 {
// Note(Yangqing): I think it is possible to do a more general swapaxes operator
// but I am a little afraid of going down that general path. Only implementing
// the two actually needed ones here.
template
class NHWC2NCHWOp final : public Operator {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
USE_SIMPLE_CTOR_DTOR(NHWC2NCHWOp);
bool RunOnDevice() override {
const auto& X = Input(0);
const int ndim = X.dim();
CAFFE_ENFORCE_GE(ndim, 3);
const int N = X.dim32(0);
const int C = X.dim32(ndim - 1);
std::vector Y_dims(ndim);
Y_dims[0] = N;
Y_dims[1] = C;
int HxW = 1;
for (int i = 2; i < ndim; ++i) {
Y_dims[i] = X.dim32(i - 1);
HxW *= Y_dims[i];
}
auto* Y = Output(0, Y_dims, at::dtype());
if (X.numel() <= 0) {
return true;
}
math::NHWC2NCHW(
N,
C,
HxW,
X.template data(),
Y->template mutable_data(),
&context_);
return true;
}
};
template
class NCHW2NHWCOp final : public Operator {
public:
USE_OPERATOR_CONTEXT_FUNCTIONS;
USE_SIMPLE_CTOR_DTOR(NCHW2NHWCOp);
bool RunOnDevice() override {
const auto& X = Input(0);
const int ndim = X.dim();
CAFFE_ENFORCE_GE(ndim, 3);
const int N = X.dim32(0);
const int C = X.dim32(1);
std::vector Y_dims(ndim);
Y_dims[0] = N;
Y_dims[ndim - 1] = C;
int HxW = 1;
for (int i = 1; i < ndim - 1; ++i) {
Y_dims[i] = X.dim32(i + 1);
HxW *= Y_dims[i];
}
auto* Y = Output(0, Y_dims, at::dtype());
if (X.numel() <= 0) {
return true;
}
math::NCHW2NHWC(
N,
C,
HxW,
X.template data(),
Y->template mutable_data(),
&context_);
return true;
}
};
} // namespace caffe2
#endif // CAFFE2_OPERATORS_ORDER_SWITCH_OPS_H_