My unpool max 2d #5826

sweetsky0901 · 2017-11-22T08:01:30Z

resolve #5787
当前只实现了unpool的2d的max

…dd it

… my_unpool_max_2d

wangkuiyi · 2017-11-22T23:02:09Z

paddle/operators/unpool_op.cc

+
+class Unpool2dOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  Unpool2dOpMaker(framework::OpProto* proto,  \


It seems that we don't need the trailing \ here.

表示连起来是一行来着。好吧，我删了它

wangkuiyi · 2017-11-22T23:02:58Z

paddle/operators/unpool_op.cc

+        "the number of channels, H and W is the height and "
+        "width of feature.");
+    AddAttr<std::vector<int>>("ksize",
+        "(vector ), the unpooling window size(height, width) "


(vector ) => (vector)

size(height => size (height

wangkuiyi · 2017-11-22T23:04:10Z

paddle/operators/unpool_op.cc

+        "of unpooling operator.");
+    AddAttr<std::vector<int>>("strides",
+        "(vector, default:{1, 1}), "
+        "strides(height, width) of unpooling operator.")


strides(height => strides (height

wangkuiyi · 2017-11-22T23:05:18Z

paddle/operators/unpool_op.cc

+                  framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+        "(Tensor) The input tensor of unpool operator. "


I am not sure should we have variable type in the parenthesis? Is this a convention?

这个是否有这项约定我不太确定从哪里看约定。
但是看了几个op都是这样写的

wangkuiyi · 2017-11-22T23:06:32Z

paddle/operators/unpool_op.cc

+    AddAttr<std::string>("unpoolingtype",
+        "(string), unpooling type, can be \"max\" for max-unpooling ")
+        .InEnum({"max"});
+    AddComment(R"DOC(


I'd thought that we no longer need that many "'s in C++'s DOC-string.

wangkuiyi · 2017-11-22T23:10:09Z

paddle/operators/unpool_op.h

+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;


Because the using in a header file might break existing namespace hierarchy if we include that header file (https://google.github.io/styleguide/cppguide.html#Namespaces), usually we don't do that.

抱歉，对这个的确没有太注意，这里也是copy了其他的op的写法。我把它去掉吧

wangkuiyi · 2017-11-22T23:10:43Z

paddle/operators/unpool_op.h

+      set_zero(context.device_context(), out, static_cast<T>(0));
+    }
+    switch (ksize.size()) {
+    case 2: {


There is only one case in this switch? It seems that we should use if instead of switch.

因为开始考虑要后面继续加3d的来着，ok，我先暂时去掉吧

wangkuiyi · 2017-11-22T23:10:57Z

paddle/operators/unpool_op.h

+      zero(device_ctx, in_x_grad, static_cast<T>(0.0));
+    }
+    switch (ksize.size()) {
+    case 2: {


Like in the above comment.

因为开始考虑要后面继续加3d的来着，ok，我先暂时去掉吧

wangkuiyi · 2017-11-22T23:11:38Z

paddle/operators/unpool_op.cc

+    std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+
+    PADDLE_ENFORCE(in_x_dims.size() == 4,
+                    "Unpooling intput should be 4-D.");


should be 4-D => must be of 4-dimensional

wangkuiyi · 2017-11-22T23:12:17Z

paddle/operators/unpool_op.h

+    math::SetConstant<Place, T> zero;
+    if (in_x_grad) {
+      in_x_grad->mutable_data<T>(context.GetPlace());
+      zero(device_ctx, in_x_grad, static_cast<T>(0.0));


0.0 => 0, in the case that T is int.

… my_unpool_max_2d

sweetsky0901

done thanks

sweetsky0901 · 2017-11-23T03:28:36Z

paddle/operators/unpool_op.h

+      zero(device_ctx, in_x_grad, static_cast<T>(0.0));
+    }
+    switch (ksize.size()) {
+    case 2: {


因为开始考虑要后面继续加3d的来着，ok，我先暂时去掉吧

sweetsky0901 · 2017-11-23T03:28:45Z

paddle/operators/unpool_op.h

+    math::SetConstant<Place, T> zero;
+    if (in_x_grad) {
+      in_x_grad->mutable_data<T>(context.GetPlace());
+      zero(device_ctx, in_x_grad, static_cast<T>(0.0));


sweetsky0901 · 2017-11-23T03:28:52Z

paddle/operators/unpool_op.h

+      set_zero(context.device_context(), out, static_cast<T>(0));
+    }
+    switch (ksize.size()) {
+    case 2: {


因为开始考虑要后面继续加3d的来着，ok，我先暂时去掉吧

sweetsky0901 · 2017-11-23T03:30:12Z

paddle/operators/unpool_op.h

+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;


抱歉，对这个的确没有太注意，这里也是copy了其他的op的写法。我把它去掉吧

sweetsky0901 · 2017-11-23T03:33:36Z

paddle/operators/unpool_op.cc

+    std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+
+    PADDLE_ENFORCE(in_x_dims.size() == 4,
+                    "Unpooling intput should be 4-D.");


sweetsky0901 · 2017-11-23T03:35:58Z

paddle/operators/unpool_op.cc

+        "the number of channels, H and W is the height and "
+        "width of feature.");
+    AddAttr<std::vector<int>>("ksize",
+        "(vector ), the unpooling window size(height, width) "


sweetsky0901 · 2017-11-23T03:36:07Z

paddle/operators/unpool_op.cc

+
+class Unpool2dOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  Unpool2dOpMaker(framework::OpProto* proto,  \


表示连起来是一行来着。好吧，我删了它

sweetsky0901 · 2017-11-23T03:36:49Z

paddle/operators/unpool_op.cc

+                  framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X",
+        "(Tensor) The input tensor of unpool operator. "


这个是否有这项约定我不太确定从哪里看约定。
但是看了几个op都是这样写的

sweetsky0901 · 2017-11-23T03:41:03Z

paddle/operators/unpool_op.cc

+    AddAttr<std::string>("unpoolingtype",
+        "(string), unpooling type, can be \"max\" for max-unpooling ")
+        .InEnum({"max"});
+    AddComment(R"DOC(


sweetsky0901 · 2017-11-23T03:50:00Z

paddle/operators/unpool_op.cc

+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of UnpoolOp"
+                   "should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of UnpoolOp should not be null.");


这里没有很好的理解应该在什么地方还加空格，sorry

qingqing01 · 2017-11-24T11:06:47Z

paddle/operators/math/unpooling.h

+#pragma once
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/hostdevice.h"


remove this unused header file.

qingqing01 · 2017-11-24T11:17:55Z

paddle/operators/math/unpooling.h

+                  const framework::Tensor& indices,
+                  framework::Tensor * input_grad,
+                  const framework::Tensor& output,
+                  const framework::Tensor& output_grad);


Put the input_grad as the last parameter.

https://google.github.io/styleguide/cppguide.html#Function_Parameter_Ordering

When defining a function, parameter order is: inputs, then outputs.

qingqing01 · 2017-11-24T11:35:19Z

paddle/operators/math/unpooling.cc

+        for (int i = 0; i < input_feasize; ++i) {
+          int index =  indices_data[i];
+          PADDLE_ENFORCE(index < output_feasize, "err index in unpooling!");
+          output_data[index] = input_data[i];


这里为什么不是 output_data[index] += input_data[i];

另外，Unpool2dMaxFunctor可以不用写，直接复用MaxPool2dWithIndexGradFunctor就行。

这样会不会让代码结构上显得有点乱呢，我不太建议这么弄

qingqing01 · 2017-11-24T11:55:51Z

paddle/operators/math/unpooling.cu

+ * All tensors are in NCHW format.
+ */
+template <typename T>
+class Unpool2dMaxGradFunctor<platform::GPUPlace, T> {


同CPU，也可以复用MaxPool2dWithIndexGradFunctor，可以去掉 Unpool2dMaxFunctor。

这样会不会让代码结构上显得有点乱呢，我不太建议这么弄

qingqing01 · 2017-11-24T11:56:13Z

paddle/operators/unpool_op.cc

+        "(vector defalut:{0,0}), "
+        "paddings (height, width) of unpooling operator.")
+        .SetDefault({0, 0});
+    AddAttr<std::string>("unpoolingtype",


unpoolingtype -> unpooling_type

sweetsky0901

done thanks

sweetsky0901 · 2017-11-26T13:16:16Z

paddle/operators/math/unpooling.cc

+        for (int i = 0; i < input_feasize; ++i) {
+          int index =  indices_data[i];
+          PADDLE_ENFORCE(index < output_feasize, "err index in unpooling!");
+          output_data[index] = input_data[i];


这样会不会让代码结构上显得有点乱呢，我不太建议这么弄

sweetsky0901 · 2017-11-26T13:16:27Z

paddle/operators/math/unpooling.cu

+ * All tensors are in NCHW format.
+ */
+template <typename T>
+class Unpool2dMaxGradFunctor<platform::GPUPlace, T> {


这样会不会让代码结构上显得有点乱呢，我不太建议这么弄

sweetsky0901 · 2017-11-26T13:23:09Z

paddle/operators/math/unpooling.h

+#pragma once
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/device_context.h"
+#include "paddle/platform/hostdevice.h"


sweetsky0901 · 2017-11-26T13:23:15Z

paddle/operators/math/unpooling.h

+                  const framework::Tensor& indices,
+                  framework::Tensor * input_grad,
+                  const framework::Tensor& output,
+                  const framework::Tensor& output_grad);


sweetsky0901 · 2017-11-26T13:23:28Z

paddle/operators/unpool_op.cc

+        "(vector defalut:{0,0}), "
+        "paddings (height, width) of unpooling operator.")
+        .SetDefault({0, 0});
+    AddAttr<std::string>("unpoolingtype",


… my_unpool_max_2d

NHZlX · 2017-11-27T05:41:01Z

paddle/operators/math/unpooling.cc

+    int input_feasize = input_height * input_width;
+    int output_feasize = output_height * output_width;
+    const T* input_data = input.data<T>();
+    const T * indices_data = indices.data<T>();


You specify indices type to int16 in https://github.com/sweetsky0901/Paddle/blob/27cf7f3376e2005522dbba8c964cb2e449b46380/python/paddle/v2/fluid/tests/test_unpool_op.py#L56 so, there must be wrong using const T * indices_data = indices.data<T>(); here, also, the paddle does not support int16 currently. So, you must specify int32 when test, and use const int * indices_data = indices.data<int>(); or add a new parameter type to the template, like here

done
系统默认对输入类型做了检查，必须一样，如果不一样，需要重写
protected:
framework::OpKernelType GetKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Inputframework::Tensor("X")->type()),
ctx.device_context());
}
之前为了逃避这个检查，用float传递的index，然后到具体计算再强转的。

不是这样的
在注册Op的时候一般注册两个(double/float)，那么在op运行的时候到底是用哪个呢（double的还是float），这时候需要调用GetKernelType函数来判断，如果X是float，那就调float的Op，如果是double，那就调double的Op

请问这些有文档吗？

… my_unpool_max_2d

sweetsky0901

done thanks

sweetsky0901 · 2017-11-27T07:00:29Z

paddle/operators/math/unpooling.cc

+    int input_feasize = input_height * input_width;
+    int output_feasize = output_height * output_width;
+    const T* input_data = input.data<T>();
+    const T * indices_data = indices.data<T>();


done
系统默认对输入类型做了检查，必须一样，如果不一样，需要重写
protected:
framework::OpKernelType GetKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
framework::ToDataType(ctx.Inputframework::Tensor("X")->type()),
ctx.device_context());
}
之前为了逃避这个检查，用float传递的index，然后到具体计算再强转的。

chengduoZH · 2017-11-27T06:08:57Z

paddle/operators/math/unpooling.cc

+      for (int c = 0; c < output_channels; ++c) {
+        for (int i = 0; i < input_feasize; ++i) {
+          int index =  indices_data[i];
+          PADDLE_ENFORCE(index < output_feasize, "err index in unpooling!");


这个可以写在循环外面，将indices转为EigenMatrix或者EigenVector，之后调用maximum方法返回indices中最大元素，然后判断是否越界。
CUDA实现也可以这样做，避免在cuda kernel里面做越界检测。

调用maximum实际上又多了很多次比较，不太赞成这样的写法

chengduoZH · 2017-11-27T06:13:06Z

paddle/operators/math/unpooling.cu

+                                  const int channels,
+                                  T* output_data,
+                                  const int output_height,
+                                  const int output_width) {


Please note the order of parameters.

chengduoZH · 2017-11-27T06:15:01Z

paddle/operators/unpool_op.cc

+        "(Tensor) The input tensor of unpool operator. "
+        "The format of input tensor is NCHW. Where N is batch size, C is the "
+        "number of channels, H and W is the height and width of feature.");
+    AddInput("Y",


I don't think the name of the variable is intuitive. It can be changed to Masks or Indices.

agree with @chengduoZH

chengduoZH · 2017-11-27T06:17:30Z

paddle/operators/unpool_op.cc

+  }
+};
+
+int OutputSize(int input_size, int ksize, int padding, int stride) {


Please add inline.

暂时不改

chengduoZH · 2017-11-27T06:27:28Z

paddle/operators/unpool_op.cc

+
+    PADDLE_ENFORCE(in_x_dims.size() == 4,
+                    "Unpooling intput must be of 4-dimensional.");
+    for (int i = 0; i < 4; ++i) {


You can directly use "==" operator to compare in_x_dims and in_y_dims .

Paddle/paddle/framework/dim.h

Lines 42 to 44 in 33fa2df

bool operator==(const Dim<i>& o) const {

return (head == o.head) && (tail == o.tail);

}

期待等能有一些关键数据类型的一些说明文档之后批量的改

PADDLE_ENFORCE_EQ(in_x_dims, in_y_dims);

即可。

chengduoZH · 2017-11-27T06:33:43Z

paddle/operators/unpool_op.h

+    math::SetConstant<Place, T> zero;
+    if (in_x_grad) {
+      in_x_grad->mutable_data<T>(context.GetPlace());
+      zero(device_ctx, in_x_grad, static_cast<T>(0));


You can directly use this function.

现在这么写，是不是也可以？

chengduoZH · 2017-11-27T06:34:52Z

paddle/operators/unpool_op.h

+        context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    framework::Tensor* in_x_grad =
+      context.Output<framework::Tensor>(framework::GradVarName("X"));
+    std::string unpooling_type = context.Attr<std::string>("unpoolingtype");


unpoolingtype -> unpooling_type

chengduoZH · 2017-11-27T07:17:55Z

paddle/operators/unpool_op.h

+      in_x_grad->mutable_data<T>(context.GetPlace());
+      zero(device_ctx, in_x_grad, static_cast<T>(0));
+    }
+    math::Unpool2dMaxGradFunctor<Place, T, T2> unpool2d_max_backward;


Pooling 有maxPool, avgPool，当前Unpooling只有maxPool的上采样，是否应该也加上avgPool的上采样 @qingqing01 @NHZlX

暂时可以不用考虑 avg 的吧，有需求的时候再进行添加

chengduoZH · 2017-11-27T07:21:30Z

paddle/operators/math/unpooling.h

+namespace operators {
+namespace math {
+
+template <typename Place, typename T, typename T2>


typename T, typename T2 - > typename T1, typename T2

chengduoZH · 2017-11-27T07:22:56Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+                    hidx = (index - index % out_W) / out_W
+                    widx = index % out_W
+                    out[nidx, cidx, int(hidx), int(widx)] = \
+                            input[nidx, cidx, h, w]


我觉得python代码可以写的更简单一点

因为是测试代码，就先以准确就好了

sweetsky0901

done

sweetsky0901 · 2017-11-27T07:36:03Z

paddle/operators/math/unpooling.cc

+    int input_feasize = input_height * input_width;
+    int output_feasize = output_height * output_width;
+    const T* input_data = input.data<T>();
+    const T * indices_data = indices.data<T>();


请问这些有文档吗？

sweetsky0901 · 2017-11-27T07:40:05Z

paddle/operators/math/unpooling.cc

+      for (int c = 0; c < output_channels; ++c) {
+        for (int i = 0; i < input_feasize; ++i) {
+          int index =  indices_data[i];
+          PADDLE_ENFORCE(index < output_feasize, "err index in unpooling!");


调用maximum实际上又多了很多次比较，不太赞成这样的写法

sweetsky0901 · 2017-11-27T07:46:32Z

paddle/operators/unpool_op.cc

+  }
+};
+
+int OutputSize(int input_size, int ksize, int padding, int stride) {


暂时不改

sweetsky0901 · 2017-11-27T07:47:53Z

paddle/operators/unpool_op.cc

+
+    PADDLE_ENFORCE(in_x_dims.size() == 4,
+                    "Unpooling intput must be of 4-dimensional.");
+    for (int i = 0; i < 4; ++i) {


期待等能有一些关键数据类型的一些说明文档之后批量的改

sweetsky0901 · 2017-11-27T07:48:54Z

paddle/operators/unpool_op.h

+    if (output_data) {
+      math::SetConstant<Place, T> set_zero;
+      set_zero(context.device_context(), out, static_cast<T>(0));
+    }


习惯对指针都加判断了，按道理应该判断！的，然后抛异常。

sweetsky0901 · 2017-11-27T07:51:51Z

paddle/operators/unpool_op.h

+        context.Input<framework::Tensor>(framework::GradVarName("Out"));
+    framework::Tensor* in_x_grad =
+      context.Output<framework::Tensor>(framework::GradVarName("X"));
+    std::string unpooling_type = context.Attr<std::string>("unpoolingtype");


sweetsky0901 · 2017-11-27T07:52:44Z

paddle/operators/unpool_op.h

+    math::SetConstant<Place, T> zero;
+    if (in_x_grad) {
+      in_x_grad->mutable_data<T>(context.GetPlace());
+      zero(device_ctx, in_x_grad, static_cast<T>(0));


现在这么写，是不是也可以？

sweetsky0901 · 2017-11-27T07:53:11Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+                    hidx = (index - index % out_W) / out_W
+                    widx = index % out_W
+                    out[nidx, cidx, int(hidx), int(widx)] = \
+                            input[nidx, cidx, h, w]


因为是测试代码，就先以准确就好了

qingqing01 · 2017-11-27T14:28:22Z

paddle/operators/unpool_op.cc

+          ksize  – Size of the max pooling window.
+          stride – Stride of the max pooling window.
+                   "It is set to kernel_size by default.
+          padding – Padding that was added to the input"


line 41 - line 48已经解释了ksize, stride, padding的含义，这里不用重复解释。Doc里可以给出参数文献，output shape的计算公式等。

比如参考 PyTorch: http://pytorch.org/docs/master/nn.html?highlight=unpool#torch.nn.MaxUnpool2d

qingqing01 · 2017-11-27T14:31:12Z

paddle/operators/unpool_op.cc

+
+    PADDLE_ENFORCE(in_x_dims.size() == 4,
+                    "Unpooling intput must be of 4-dimensional.");
+    for (int i = 0; i < 4; ++i) {


PADDLE_ENFORCE_EQ(in_x_dims, in_y_dims);

即可。

qingqing01 · 2017-11-27T15:31:38Z

paddle/operators/math/unpooling.cu

+    int bsize = input_height * input_width * channels;
+    int csize = input_height * input_width;
+    int out_bsize = output_height * output_width * channels;
+    int out_csize = output_height * output_width;


感觉命名有点随意。

int in_c_stride = input_height * input_width; int in_n_stride = in_c_stride * channels; int out_c_stride = output_height * output_width; int out_n_stride = out_c_stride * channels;

数据是NCHW，这里就用了n_stride, c_stride，其他更好的命名也可以。但bsize, csize等觉得有点随意。

qingqing01 · 2017-11-27T15:32:34Z

paddle/operators/math/unpooling.cu

+      int cidx = boffset / csize;
+      int out_offset = bidx * out_bsize + cidx * out_csize;
+      int out_index = indices_data[i];
+      PADDLE_ASSERT(out_index < (output_height * output_width));


PADDLE_ASSERT(out_index < out_c_stride);

qingqing01 · 2017-11-27T15:34:34Z

paddle/operators/math/unpooling.cu

+    int bsize = input_height * input_width * channels;
+    int csize = input_height * input_width;
+    int out_bsize = output_height * output_width * channels;
+    int out_csize = output_height * output_width;


qingqing01 · 2017-11-27T15:47:01Z

paddle/operators/unpool_op.cc

+        "(Tensor) The input tensor of unpool operator. "
+        "The format of input tensor is NCHW. Where N is batch size, C is the "
+        "number of channels, H and W is the height and width of feature.");
+    AddInput("Y",


agree with @chengduoZH

qingqing01 · 2017-11-27T15:50:21Z

paddle/operators/unpool_op.cu.cc

+   distributed under the License is distributed on an "AS IS" BASIS,
+   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. */


不要有缩进，参考 paddle/operators/accuracy_op.h

qingqing01 · 2017-11-27T15:52:39Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+def unpool2dmax_forward_naive(input, indices, ksize, strides, paddings):
+    s0, s1, s2, s3 = input.shape
+    out_H=(s2 - 1) * strides[0] - 2 * paddings[0] + ksize[0]
+    out_W=(s2 - 1) * strides[1] - 2 * paddings[1] + ksize[1]


请安装格式检查工具！！！( 上个PR就说过)，这里的格式不符合Python格式。
= 前后有空格。

另外，python里不要大小写混合，这个python文件里好几处，请一并修改。

我已经装了这个工具了！！！！！但是就是没检查出来。
done

qingqing01 · 2017-11-27T15:58:58Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+                        arg = x_masked.argmax()
+                        indices[nidx, cidx, i, j] = \
+                                (r_start + arg / self.ksize[1]) * W + \
+                                c_start + arg % self.ksize[1]


这里不用自己写max-pooling的forward，可以调用python/paddle/v2/fluid/tests/test_pool_max_op.py里的函数max_pool2D_forward_naive即可。

那单测文件不是互相依赖了吗？

qingqing01 · 2017-11-27T15:59:58Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+                        indices[nidx, cidx, i, j] = \
+                                (r_start + arg / self.ksize[1]) * W + \
+                                c_start + arg % self.ksize[1]
+        output = self.Unpool2d_forward_naive(input, indices, self.ksize, \


self.Unpool2d_forward_naive -> unpool2d_forward_naive

remove line 72.

这个是为了准备将来写avg的测试兼容用的

sweetsky0901

done

sweetsky0901 · 2017-11-27T23:47:28Z

paddle/operators/math/unpooling.cu

+    int bsize = input_height * input_width * channels;
+    int csize = input_height * input_width;
+    int out_bsize = output_height * output_width * channels;
+    int out_csize = output_height * output_width;


sweetsky0901 · 2017-11-27T23:49:04Z

paddle/operators/math/unpooling.cu

+      int cidx = boffset / csize;
+      int out_offset = bidx * out_bsize + cidx * out_csize;
+      int out_index = indices_data[i];
+      PADDLE_ASSERT(out_index < (output_height * output_width));


sweetsky0901 · 2017-11-27T23:49:13Z

paddle/operators/math/unpooling.cu

+    int bsize = input_height * input_width * channels;
+    int csize = input_height * input_width;
+    int out_bsize = output_height * output_width * channels;
+    int out_csize = output_height * output_width;


sweetsky0901 · 2017-11-27T23:49:18Z

paddle/operators/math/unpooling.cu

+      int cidx = boffset / csize;
+      int out_offset = bidx * out_bsize + cidx * out_csize;
+      int out_index = indices_data[i];
+      PADDLE_ASSERT(out_index < (output_height * output_width));


sweetsky0901 · 2017-11-27T23:51:56Z

paddle/operators/math/unpooling.cu

+    const T2 * indices_data = indices.data<T2>();
+    T* output_data = output->mutable_data<T>(context.GetPlace());
+    int nthreads = batch_size * output_channels * input_height * input_width;
+    int blocks = (nthreads + 1024 - 1) / 1024;


sweetsky0901 · 2017-11-28T00:01:35Z

paddle/operators/unpool_op.cc

+
+    PADDLE_ENFORCE(in_x_dims.size() == 4,
+                    "Unpooling intput must be of 4-dimensional.");
+    for (int i = 0; i < 4; ++i) {


sweetsky0901 · 2017-11-28T00:02:30Z

paddle/operators/unpool_op.cu.cc

+   distributed under the License is distributed on an "AS IS" BASIS,
+   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. */


sweetsky0901 · 2017-11-28T00:11:27Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+def unpool2dmax_forward_naive(input, indices, ksize, strides, paddings):
+    s0, s1, s2, s3 = input.shape
+    out_H=(s2 - 1) * strides[0] - 2 * paddings[0] + ksize[0]
+    out_W=(s2 - 1) * strides[1] - 2 * paddings[1] + ksize[1]


我已经装了这个工具了！！！！！但是就是没检查出来。
done

sweetsky0901 · 2017-11-28T00:12:52Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+                        indices[nidx, cidx, i, j] = \
+                                (r_start + arg / self.ksize[1]) * W + \
+                                c_start + arg % self.ksize[1]
+        output = self.Unpool2d_forward_naive(input, indices, self.ksize, \


这个是为了准备将来写avg的测试兼容用的

sweetsky0901 · 2017-11-28T00:13:08Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+                        arg = x_masked.argmax()
+                        indices[nidx, cidx, i, j] = \
+                                (r_start + arg / self.ksize[1]) * W + \
+                                c_start + arg % self.ksize[1]


那单测文件不是互相依赖了吗？

… my_unpool_max_2d

qingqing01

希望下个PR继续完成对3D的支持，并修改这个PR遗留问题。

qingqing01 · 2017-11-29T13:08:30Z

python/paddle/v2/fluid/tests/test_unpool_op.py

+                        arg = x_masked.argmax()
+                        indices[nidx, cidx, i, j] = \
+                                (r_start + arg / self.ksize[1]) * wsize + \
+                                c_start + arg % self.ksize[1]


这里max-pooling的计算，请复用 paddle/v2/fluid/tests/test_pool_max_op.py 或 paddle/v2/fluid/tests/test_pool2d_op.py的代码。

wanghaox and others added 17 commits November 11, 2017 19:59

for resolve conflicts

fef617a

Merge branch 'sweetsky0901-my_maxout_op' into develop

9127840

paddle/operators/math/CMakeLists.txt maybe del sequence_pooling and a…

4748073

…dd it

Merge branch 'sweetsky0901-my_maxout_op2' into develop

6665c49

add unpool

bc45335

merge cmakelist

f638f91

add unpool2d make ok

45a8c9d

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

ab03daa

… my_unpool_max_2d

add test

200f07c

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

822f283

… my_unpool_max_2d

test unpool ok cpu

90f664d

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

abb3357

… my_unpool_max_2d

gpu test ok unpool2dmax

e2a5905

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

8ba8237

… my_unpool_max_2d

del printf

47bd0bb

format code

0112c5d

format test code

e553d57

wangkuiyi reviewed Nov 22, 2017

View reviewed changes

wangkuiyi requested a review from qingqing01 November 22, 2017 23:14

sweetsky0901 added 2 commits November 23, 2017 11:53

modify for code review by wangyi

66b8436

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

ee4a5d2

… my_unpool_max_2d

sweetsky0901 commented Nov 23, 2017

View reviewed changes

qingqing01 mentioned this pull request Nov 24, 2017

Roi pool operator #5831

Merged

qingqing01 reviewed Nov 24, 2017

View reviewed changes

qingqing01 requested a review from chengduoZH November 24, 2017 11:57

modify for code review by qingqing

c218961

sweetsky0901 commented Nov 26, 2017

View reviewed changes

sweetsky0901 added 3 commits November 26, 2017 21:24

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

a38bbc8

… my_unpool_max_2d

add unpool_op.h modify

cfd7721

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

27cf7f3

… my_unpool_max_2d

chengduoZH requested a review from NHZlX November 27, 2017 04:34

NHZlX reviewed Nov 27, 2017

View reviewed changes

sweetsky0901 added 2 commits November 27, 2017 14:54

modify for type check rewrite

20654cf

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

022b48e

… my_unpool_max_2d

sweetsky0901 commented Nov 27, 2017

View reviewed changes

chengduoZH reviewed Nov 27, 2017

View reviewed changes

sweetsky0901 commented Nov 27, 2017

View reviewed changes

modify for code review zcd

f9c2a5c

qingqing01 reviewed Nov 27, 2017

View reviewed changes

sweetsky0901 commented Nov 28, 2017

View reviewed changes

sweetsky0901 added 12 commits November 28, 2017 08:15

modify for code review by qingqing 2nd

57e68e5

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

ee0a794

… my_unpool_max_2d

modify for del T2 and doc update

6fc9a9f

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

821899c

… my_unpool_max_2d

format code

d9673ca

format code

bd56138

format code

c52ed8d

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

2d42fa7

… my_unpool_max_2d

format code

d2ee3c9

format code

3206094

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

5b449b6

… my_unpool_max_2d

format ..

4ffb73f

qingqing01 approved these changes Nov 29, 2017

View reviewed changes

sweetsky0901 merged commit 966a6ce into PaddlePaddle:develop Nov 29, 2017

	bool operator==(const Dim<i>& o) const {
	return (head == o.head) && (tail == o.tail);
	}

My unpool max 2d #5826

My unpool max 2d #5826

Conversation

sweetsky0901 commented Nov 22, 2017

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sweetsky0901 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sweetsky0901 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

NHZlX Nov 27, 2017 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

sweetsky0901 left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

NHZlX Nov 27, 2017 •

edited

Loading