deconv1d dynamic weight

docs/developer-guide/operators.md

@@ -529,6 +529,7 @@ y = activation(x3, act_type, act_params)
 | 19        | output_pad_bottom| int | output_pad_right |           |
 | 20        | output_w      | int   | 0         |                   |
 | 21        | output_h      | int   | output_w  |                   |
+| 28        | dynamic_weight| int   | 0         |                   |
 
 | weight        | type  | shape                 |
 | ------------- | ----- | --------------------- |
@@ -558,6 +559,7 @@ y = activation(x3, act_type, act_params)
 | 15        | pad_right     | int   | pad_left  |                   |
 | 18        | output_pad_right| int | 0         |                   |
 | 20        | output_w      | int   | 0         |                   |
+| 28        | dynamic_weight| int   | 0         |                   |
 
 | weight        | type  | shape                 |
 | ------------- | ----- | --------------------- |
@@ -638,6 +640,7 @@ y = activation(x3, act_type, act_params)
 | 19        | output_pad_bottom| int | output_pad_right |           |
 | 20        | output_w      | int   | 0         |                   |
 | 21        | output_h      | int   | output_w  |                   |
+| 28        | dynamic_weight| int   | 0         |                   |
 
 | weight        | type  | shape                 |
 | ------------- | ----- | --------------------- |
@@ -668,6 +671,7 @@ y = activation(x3, act_type, act_params)
 | 15        | pad_right     | int   | pad_left  |                   |
 | 18        | output_pad_right| int | 0         |                   |
 | 20        | output_w      | int   | 0         |                   |
+| 28        | dynamic_weight| int   | 0         |                   |
 
 | weight        | type  | shape                 |
 | ------------- | ----- | --------------------- |

src/layer/deconvolution1d.cpp

@@ -39,11 +39,21 @@ int Deconvolution1D::load_param(const ParamDict& pd)
     activation_type = pd.get(9, 0);
     activation_params = pd.get(10, Mat());
 
+    dynamic_weight = pd.get(28, 0);
+
+    if (dynamic_weight)
+    {
+        one_blob_only = false;
+    }
+
     return 0;
 }
 
 int Deconvolution1D::load_model(const ModelBin& mb)
 {
+    if (dynamic_weight)
+        return 0;
+
     weight_data = mb.load(weight_data_size, 0);
     if (weight_data.empty())
         return -100;
@@ -143,6 +153,89 @@ int Deconvolution1D::forward(const Mat& bottom_blob, Mat& top_blob, const Option
     return 0;
 }
 
+int Deconvolution1D::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const
+{
+    const Mat& bottom_blob = bottom_blobs[0];
+    const Mat& _weight_data = bottom_blobs[1];
+    Mat& top_blob = top_blobs[0];
+
+    const int _num_input = bottom_blob.h;
+    const int _kernel_w = _weight_data.w;
+    const int _num_output = _weight_data.h * 1;
+
+    Mat weight_data_flattened;
+    flatten(_weight_data, weight_data_flattened, opt);
+    if (weight_data_flattened.empty())
+        return -100;
+
+    // transpose group-inch/group-outch/group-kw to group-outch/group-inch/group-kw
+    Mat weight_data_transposed;
+    {
+        weight_data_transposed.create(_kernel_w * _num_output * _num_input / 1, 4u, opt.workspace_allocator);
+        if (weight_data_transposed.empty())
+            return -100;
+
+        const int outch_g = _num_output / 1;
+        const int inch_g = _num_input / 1;
+        const int maxk = _kernel_w;
+
+        for (int g = 0; g < 1; g++)
+        {
+            // reorder weight from inch-outch to outch-inch
+            float* wg2 = (float*)weight_data_transposed + g * outch_g * inch_g * maxk;
+            const float* wg = (const float*)weight_data_flattened + g * inch_g * outch_g * maxk;
+            for (int i = 0; i < outch_g; i++)
+            {
+                for (int j = 0; j < inch_g; j++)
+                {
+                    for (int k = 0; k < maxk; k++)
+                    {
+                        wg2[(i * inch_g + j) * maxk + k] = wg[(j * outch_g + i) * maxk + k];
+                    }
+                }
+            }
+        }
+    }
+
+    Mat bias_data_flattened;
+    if (bias_term)
+    {
+        const Mat& _bias_data = bottom_blobs[2];
+        flatten(_bias_data, bias_data_flattened, opt);
+        if (bias_data_flattened.empty())
+            return -100;
+    }
+
+    const int w = bottom_blob.w;
+
+    const int kernel_extent_w = dilation_w * (_kernel_w - 1) + 1;
+
+    int outw = (w - 1) * stride_w + kernel_extent_w + output_pad_right;
+
+    Mat top_blob_bordered;
+    if (pad_left > 0 || pad_right > 0 || output_w > 0)
+    {
+        top_blob_bordered.create(outw, _num_output, 4u, opt.workspace_allocator);
+    }
+    else
+    {
+        top_blob_bordered = top_blob;
+        top_blob_bordered.create(outw, _num_output, 4u, opt.blob_allocator);
+    }
+    if (top_blob_bordered.empty())
+        return -100;
+
+    int ret = deconvolution1d(bottom_blob, top_blob_bordered, weight_data_transposed, bias_data_flattened, _kernel_w, stride_w, dilation_w, activation_type, activation_params, opt);
+    if (ret != 0)
+        return ret;
+
+    cut_padding(top_blob_bordered, top_blob, opt);
+    if (top_blob.empty())
+        return -100;
+
+    return 0;
+}
+
 void Deconvolution1D::cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const
 {
     if (pad_left > 0 || pad_right > 0)

src/layer/deconvolution1d.h

@@ -30,6 +30,8 @@ class Deconvolution1D : public Layer
 
     virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
 
+    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
+
 protected:
     void cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const;
 
@@ -50,6 +52,8 @@ class Deconvolution1D : public Layer
     int activation_type;
     Mat activation_params;
 
+    int dynamic_weight;
+
     // model
     Mat weight_data;
     Mat bias_data;

src/layer/deconvolutiondepthwise1d.cpp

@@ -40,11 +40,21 @@ int DeconvolutionDepthWise1D::load_param(const ParamDict& pd)
     activation_type = pd.get(9, 0);
     activation_params = pd.get(10, Mat());
 
+    dynamic_weight = pd.get(28, 0);
+
+    if (dynamic_weight)
+    {
+        one_blob_only = false;
+    }
+
     return 0;
 }
 
 int DeconvolutionDepthWise1D::load_model(const ModelBin& mb)
 {
+    if (dynamic_weight)
+        return 0;
+
     weight_data = mb.load(weight_data_size, 0);
     if (weight_data.empty())
         return -100;
@@ -195,6 +205,89 @@ int DeconvolutionDepthWise1D::forward(const Mat& bottom_blob, Mat& top_blob, con
     return 0;
 }
 
+int DeconvolutionDepthWise1D::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const
+{
+    const Mat& bottom_blob = bottom_blobs[0];
+    const Mat& _weight_data = bottom_blobs[1];
+    Mat& top_blob = top_blobs[0];
+
+    const int _num_input = bottom_blob.h;
+    const int _kernel_w = _weight_data.w;
+    const int _num_output = _weight_data.h * group;
+
+    Mat weight_data_flattened;
+    flatten(_weight_data, weight_data_flattened, opt);
+    if (weight_data_flattened.empty())
+        return -100;
+
+    // transpose group-inch/group-outch/group-kw to group-outch/group-inch/group-kw
+    Mat weight_data_transposed;
+    {
+        weight_data_transposed.create(_kernel_w * _num_output * _num_input / group, 4u, opt.workspace_allocator);
+        if (weight_data_transposed.empty())
+            return -100;
+
+        const int outch_g = _num_output / group;
+        const int inch_g = _num_input / group;
+        const int maxk = _kernel_w;
+
+        for (int g = 0; g < group; g++)
+        {
+            // reorder weight from inch-outch to outch-inch
+            float* wg2 = (float*)weight_data_transposed + g * outch_g * inch_g * maxk;
+            const float* wg = (const float*)weight_data_flattened + g * inch_g * outch_g * maxk;
+            for (int i = 0; i < outch_g; i++)
+            {
+                for (int j = 0; j < inch_g; j++)
+                {
+                    for (int k = 0; k < maxk; k++)
+                    {
+                        wg2[(i * inch_g + j) * maxk + k] = wg[(j * outch_g + i) * maxk + k];
+                    }
+                }
+            }
+        }
+    }
+
+    Mat bias_data_flattened;
+    if (bias_term)
+    {
+        const Mat& _bias_data = bottom_blobs[2];
+        flatten(_bias_data, bias_data_flattened, opt);
+        if (bias_data_flattened.empty())
+            return -100;
+    }
+
+    const int w = bottom_blob.w;
+
+    const int kernel_extent_w = dilation_w * (_kernel_w - 1) + 1;
+
+    int outw = (w - 1) * stride_w + kernel_extent_w + output_pad_right;
+
+    Mat top_blob_bordered;
+    if (pad_left > 0 || pad_right > 0 || output_w > 0)
+    {
+        top_blob_bordered.create(outw, _num_output, 4u, opt.workspace_allocator);
+    }
+    else
+    {
+        top_blob_bordered = top_blob;
+        top_blob_bordered.create(outw, _num_output, 4u, opt.blob_allocator);
+    }
+    if (top_blob_bordered.empty())
+        return -100;
+
+    int ret = deconvolutiondepthwise1d(bottom_blob, top_blob_bordered, weight_data_transposed, bias_data_flattened, _kernel_w, stride_w, dilation_w, group, activation_type, activation_params, opt);
+    if (ret != 0)
+        return ret;
+
+    cut_padding(top_blob_bordered, top_blob, opt);
+    if (top_blob.empty())
+        return -100;
+
+    return 0;
+}
+
 void DeconvolutionDepthWise1D::cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const
 {
     if (pad_left > 0 || pad_right > 0)

src/layer/deconvolutiondepthwise1d.h

@@ -30,6 +30,8 @@ class DeconvolutionDepthWise1D : public Layer
 
     virtual int forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const;
 
+    virtual int forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const;
+
 protected:
     void cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const;
 
@@ -51,6 +53,8 @@ class DeconvolutionDepthWise1D : public Layer
     int activation_type;
     Mat activation_params;
 
+    int dynamic_weight;
+
     // model
     Mat weight_data;
     Mat bias_data;