Skip to content
This repository has been archived by the owner on Nov 17, 2023. It is now read-only.

Update documentation for deconvolution operation. #6184

Merged
merged 6 commits into from
May 26, 2017
Merged

Update documentation for deconvolution operation. #6184

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
b4bb2d3
Update documentation for deconvolution operation.
indhub May 10, 2017
58b4476
Add examples.
indhub May 24, 2017
fa181d4
Merge branch 'master' into doc/deconvolution
indhub May 25, 2017
f5cb859
Merge branch 'master' into doc/deconvolution
indhub May 25, 2017
20bb6be
Remove the example.
indhub May 25, 2017
2aebcfc
Merge branch 'master' into doc/deconvolution
indhub May 25, 2017
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 20 additions & 16 deletions src/operator/deconvolution-inl.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,26 +43,30 @@ struct DeconvolutionParam : public dmlc::Parameter<DeconvolutionParam> {
bool cudnn_off;
dmlc::optional<int> layout;
DMLC_DECLARE_PARAMETER(DeconvolutionParam) {
DMLC_DECLARE_FIELD(kernel).describe("deconvolution kernel size: (h, w) or (d, h, w)");
DMLC_DECLARE_FIELD(kernel).describe("Deconvolution kernel size: (h, w) or (d, h, w). "
"This is same as the kernel size used for the corresponding convolution");
DMLC_DECLARE_FIELD(stride).set_default(TShape())
.describe("deconvolution stride: (h, w) or (d, h, w)");
.describe("The stride used for the corresponding convolution: (h, w) or (d, h, w).");
DMLC_DECLARE_FIELD(dilate).set_default(TShape())
.describe("deconvolution dilate: (h, w) or (d, h, w)");
.describe("Dilation factor for each dimension of the input: (h, w) or (d, h, w).");
DMLC_DECLARE_FIELD(pad).set_default(TShape())
.describe("pad for deconvolution: (h, w) or (d, h, w). "
"A good number is : (kernel-1)/2. "
"If target_shape is set, "
"pad will be ignored and computed accordingly");
.describe("The amount of implicit zero padding added during convolution for each "
"dimension of the input: "
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is pading output instead?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The description says this is the padding that was used for the convolution input (not deconvolution input). Also, saying deconvolution pads the output could be misleading. Deconvolution doesn't actually set the borders to zero. Example:

>>> # The transpose of convolving a 4×4 kernel over a 5×5 input padded with a
... # 2 × 2 border of zeros using unit strides (i.e., i = 5, k = 4, s = 1 and p = 2).
... input_5x5 = mx.nd.normal(shape=[1, 16, 5, 5])
>>> kernel_4x4 = mx.nd.normal(shape=[16, 16, 4, 4])
>>> conv = mx.nd.Convolution(data=input_5x5, kernel=(4,4), weight=kernel_4x4, pad=(2,2), num_filter=16, no_bias=True)
>>> transpose = mx.nd.Deconvolution(data=conv, kernel=(4,4), weight=kernel_4x4, pad=(2,2), num_filter=16, no_bias=True)
>>> transpose[0][0].asnumpy()
array([[  13.21707153, -174.19221497,  295.35494995, -482.07144165,
         236.75886536],
       [-166.75314331,  161.54397583,  344.41635132, -123.3405838 ,
         -20.51438904],
       [-991.8168335 ,   14.92340851, -131.36260986, -170.46170044,
        -283.42138672],
       [ 165.29977417,  -47.57701492, -143.96827698, -194.85569763,
         -91.3994751 ],
       [   1.36075592,  244.04574585,  211.51608276,  -40.39221573,
        -149.74568176]], dtype=float32)

"(h, w) or (d, h, w). "
"``(kernel-1)/2`` is usually a good choice. "
"If `target_shape` is set, "
"`pad` will be ignored and a padding that will generate the target shape "
"will be used.");
DMLC_DECLARE_FIELD(adj).set_default(TShape())
.describe("adjustment for output shape: (h, w) or (d, h, w). "
"If target_shape is set, "
"ad will be ignored and computed accordingly");
.describe("Adjustment for output shape: (h, w) or (d, h, w). "
"If `target_shape` is set, "
"`adj` will be ignored and computed accordingly.");
DMLC_DECLARE_FIELD(target_shape).set_default(TShape())
.describe("output shape with target shape : (h, w) or (d, h, w)");
.describe("Shape of the output tensor: (h, w) or (d, h, w).");
DMLC_DECLARE_FIELD(num_filter).set_range(1, 100000)
.describe("deconvolution filter(channel) number");
.describe("Number of output filters.");
DMLC_DECLARE_FIELD(num_group).set_default(1)
.describe("number of groups partition");
.describe("Number of groups partition.");
DMLC_DECLARE_FIELD(workspace).set_default(512).set_range(0, 8192)
.describe("Maximum temporal workspace allowed for deconvolution (MB).");
DMLC_DECLARE_FIELD(no_bias).set_default(true)
Expand All @@ -72,7 +76,7 @@ struct DeconvolutionParam : public dmlc::Parameter<DeconvolutionParam> {
.add_enum("limited_workspace", deconv::kLimited)
.add_enum("fastest", deconv::kFastest)
.set_default(dmlc::optional<int>())
.describe("Whether to pick convolution algo by running performance test.");
.describe("Whether to pick convolution algorithm by running performance test.");
DMLC_DECLARE_FIELD(cudnn_off).set_default(false)
.describe("Turn off cudnn for this layer.");
DMLC_DECLARE_FIELD(layout)
Expand All @@ -82,8 +86,8 @@ struct DeconvolutionParam : public dmlc::Parameter<DeconvolutionParam> {
.add_enum("NHWC", mshadow::kNHWC)
.add_enum("NDHWC", mshadow::kNDHWC)
.set_default(dmlc::optional<int>())
.describe("Set layout for input, output and weight. Empty for\n "
"default layout: NCW for 1d, NCHW for 2d and NCDHW for 3d.");
.describe("Set layout for input, output and weight. Empty for "
"default layout, NCW for 1d, NCHW for 2d and NCDHW for 3d.");
}

template<size_t ndim>
Expand Down
13 changes: 9 additions & 4 deletions src/operator/deconvolution.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,11 +33,16 @@ Operator* DeconvolutionProp::CreateOperatorEx(Context ctx, std::vector<TShape> *
DMLC_REGISTER_PARAMETER(DeconvolutionParam);

MXNET_REGISTER_OP_PROPERTY(Deconvolution, DeconvolutionProp)
.add_argument("data", "NDArray-or-Symbol", "Input data to the DeconvolutionOp.")
.add_argument("weight", "NDArray-or-Symbol", "Weight matrix.")
.add_argument("bias", "NDArray-or-Symbol", "Bias parameter.")
.add_argument("data", "NDArray-or-Symbol", "Input tensor to the deconvolution operation.")
.add_argument("weight", "NDArray-or-Symbol", "Weights representing the kernel.")
.add_argument("bias", "NDArray-or-Symbol", "Bias added to the result after the deconvolution "
"operation.")
.add_arguments(DeconvolutionParam::__FIELDS__())
.describe("Applies deconvolution to input and adds a bias.");
.describe("Computes 2D transposed convolution (aka fractionally strided convolution) of the "
"input tensor. This operation can be seen as the gradient of Convolution operation with "
"respect to its input. Convolution usually reduces the size of the input. Transposed "
"convolution works the other way, going from a smaller input to a larger output while "
"preserving the connectivity pattern.");

} // namespace op
} // namespace mxnet