modify Corr3dMM to support num_groups

theano/tensor/nnet/c_code/corr3d_gemm.c

@@ -127,7 +127,8 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
                         const int dilD = 1,
                         const int padH = 0,
                         const int padW = 0,
-                        const int padD = 0)
+                        const int padD = 0,
+                        const int numgroups=1)
 {
     if (PyArray_NDIM(bottom) != 5)
     {
@@ -178,7 +179,7 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
     const int kH = PyArray_DIMS(weight)[2];
     const int kW = PyArray_DIMS(weight)[3];
     const int kD = PyArray_DIMS(weight)[4];
-    if (nChannels != PyArray_DIMS(weight)[1]) {
+    if (nChannels != PyArray_DIMS(weight)[1] * numgroups) {
         PyErr_SetString(PyExc_ValueError,
                 "Corr3dMM images and kernel must have the same stack size\n");
         return NULL;
@@ -210,7 +211,7 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
                 "  weight shape: %%d %%d %%d %%d %%d\n"
                 "  top shape: %%ld %%ld %%ld %%ld %%ld (expected %%d %%d %%d %%d %%d)\n",
                 batchSize, nChannels, bottomHeight, bottomWidth, bottomDepth,
-                nFilters, nChannels, kH, kW, kD,
+                nFilters, nChannels / numgroups, kH, kW, kD,
                 PyArray_DIMS(top)[0], PyArray_DIMS(top)[1],
                 PyArray_DIMS(top)[2], PyArray_DIMS(top)[3], PyArray_DIMS(top)[4],
                 batchSize, nFilters, topHeight, topWidth, topDepth);
@@ -241,12 +242,16 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
     }
 
     // Define some useful variables
-    const int bottom_stride = PyArray_STRIDES(bottom)[0]/%(n_bytes)f;
-    const int top_stride = PyArray_STRIDES(top)[0]/%(n_bytes)f;
-    const int K_ = col_dim[1];
+    const int batch_bottom_stride = PyArray_STRIDES(bottom)[0]/%(n_bytes)f;
+    const int group_bottom_stride = (PyArray_STRIDES(bottom)[1] * nChannels / numgroups)/%(n_bytes)f;
+    const int batch_top_stride = PyArray_STRIDES(top)[0]/%(n_bytes)f;
+    const int group_top_stride = (PyArray_STRIDES(top)[1] * nFilters / numgroups)/%(n_bytes)f;
+    const int K_ = col_dim[1] / numgroups;
     const int N_ = col_dim[2];
-    const int col_stride = (K_ * N_);
-    const int M_ = nFilters;
+    const int col_stride = (K_ * N_ * numgroups);
+    const int group_col_stride = (K_ * N_);
+    const int group_weight_stride = (PyArray_STRIDES(weight)[0] * nFilters / numgroups)/%(n_bytes)f;
+    const int M_ = nFilters / numgroups;
     const %(c_float_type)s one = 1.0;
     const %(c_float_type)s zero = 0.0;
     char NTrans = 'N';
@@ -280,18 +285,21 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
         for (int n = 0; n < batchSize; ++n) {
             int tid = %(omp_get_thread_num)s;
             // First, im3d2col
-            im3d2col((%(float_type)s*)PyArray_DATA(bottom) + n * bottom_stride, nChannels,
-                     bottomHeight, bottomWidth, bottomDepth,
+            im3d2col((%(float_type)s*)PyArray_DATA(bottom) + n * batch_bottom_stride,
+                     nChannels, bottomHeight, bottomWidth, bottomDepth,
                      kH, kW, kD, dilH, dilW, dilD, padH, padW, padD, dH, dW, dD,
                      (%(float_type)s*)PyArray_DATA(col)+ tid * col_stride);
-            // Second, gemm
-            %(gemm)s(&NTrans, &NTrans,
-                   &N_, &M_, &K_,
-                   &one,
-                   (%(float_type)s*)PyArray_DATA(col)+ tid * col_stride, &N_,
-                   (%(float_type)s*)PyArray_DATA(weight), &K_,
-                   &zero,
-                   (%(float_type)s*)PyArray_DATA(top) + n * top_stride, &N_);
+
+            for ( int g = 0; g < numgroups; ++g){
+                // Second, gemm
+                %(gemm)s(&NTrans, &NTrans,
+                         &N_, &M_, &K_,
+                         &one,
+                         (%(float_type)s*)PyArray_DATA(col)+ tid * col_stride + g * group_col_stride, &N_,
+                         (%(float_type)s*)PyArray_DATA(weight) + g * group_weight_stride, &K_,
+                         &zero,
+                         (%(float_type)s*)PyArray_DATA(top) + n * batch_top_stride + g * group_top_stride, &N_);
+            }
         }
         // Restore to previous blas threads
         %(blas_set_num_threads)s(blas_threads_saved);
@@ -300,7 +308,7 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
         output = weight;
         npy_intp weight_dim[2];
         weight_dim[0] = (npy_intp)max_threads;
-        weight_dim[1] = (npy_intp)(M_ * K_);
+        weight_dim[1] = (npy_intp)(M_ * K_ * numgroups);
         PyArrayObject* local_weight = (PyArrayObject*)PyArray_ZEROS(2,
                                    weight_dim, PyArray_TYPE(weight), 0);
 
@@ -322,22 +330,25 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
         for (int n = 0; n < batchSize; ++n) {
             int tid = %(omp_get_thread_num)s;
             // First, im2col
-            im3d2col((%(float_type)s*)PyArray_DATA(bottom) + n * bottom_stride, nChannels,
-                     bottomHeight, bottomWidth, bottomDepth,
+            im3d2col((%(float_type)s*)PyArray_DATA(bottom) + n * batch_bottom_stride,
+                     nChannels, bottomHeight, bottomWidth, bottomDepth,
                      kH, kW, kD, dilH, dilW, dilD, padH, padW, padD, dH, dW, dD,
                      (%(float_type)s*)PyArray_DATA(col)+ tid * col_stride);
-            // Second, gemm
-            // Note that we accumulate into weight. We do so by setting beta = 0
-            // for the first iteration and beta = 1 for subsequent ones. (This
-            // is faster than setting weight to all zeros before the loop.)
-            %(gemm)s(&Trans, &NTrans,
-                   &K_, &M_, &N_,
-                   &one,
-                   (%(float_type)s*)PyArray_DATA(col) + tid * col_stride, &N_,
-                   (%(float_type)s*)PyArray_DATA(top) + n * top_stride, &N_,
-                   (n == 0) ? &zero : &one,
-                   (%(float_type)s*)PyArray_DATA(local_weight) + 
-                   tid * weight_dim[1], &K_);
+
+            for ( int g = 0; g < numgroups; ++g){
+                // Second, gemm
+                // Note that we accumulate into weight. We do so by setting beta = 0
+                // for the first iteration and beta = 1 for subsequent ones. (This
+                // is faster than setting weight to all zeros before the loop.)
+                %(gemm)s(&Trans, &NTrans,
+                         &K_, &M_, &N_,
+                         &one,
+                         (%(float_type)s*)PyArray_DATA(col) + tid * col_stride + g * group_col_stride, &N_,
+                         (%(float_type)s*)PyArray_DATA(top) + n * batch_top_stride + g * group_top_stride, &N_,
+                         (n == 0) ? &zero : &one,
+                         (%(float_type)s*)PyArray_DATA(local_weight) + g * group_weight_stride +
+                         tid * weight_dim[1], &K_);
+            }
         }
         // Restore to previous blas threads
         %(blas_set_num_threads)s(blas_threads_saved);
@@ -370,20 +381,23 @@ PyArrayObject* corr3dMM(PyArrayObject* bottom,
         %(blas_set_num_threads)s(1);
         %(omp_flags)s
         for (int n = 0; n < batchSize; ++n) {
-            // gemm into columns
+
             int tid = %(omp_get_thread_num)s;
-            %(gemm)s(&NTrans, &Trans,
-                   &N_, &K_, &M_,
-                   &one,
-                   (%(float_type)s*)PyArray_DATA(top) + n * top_stride, &N_,
-                   (%(float_type)s*)PyArray_DATA(weight), &K_,
-                   &zero,
-                   (%(float_type)s*)PyArray_DATA(col) + tid * col_stride, &N_);
+            for ( int g = 0; g < numgroups; ++g){
+                // gemm into columns
+                %(gemm)s(&NTrans, &Trans,
+                         &N_, &K_, &M_,
+                         &one,
+                         (%(float_type)s*)PyArray_DATA(top) + n * batch_top_stride + g * group_top_stride, &N_,
+                         (%(float_type)s*)PyArray_DATA(weight) + g * group_weight_stride, &K_,
+                         &zero,
+                         (%(float_type)s*)PyArray_DATA(col) + tid * col_stride + g * group_col_stride, &N_);
+            }
             // col2im back to the data
             col2im3d((%(float_type)s*)PyArray_DATA(col) + tid * col_stride, nChannels,
                      bottomHeight, bottomWidth, bottomDepth,
                      kH, kW, kD, dilH, dilW, dilD, padH, padW, padD, dH, dW, dD,
-                     (%(float_type)s*)PyArray_DATA(bottom) + n * bottom_stride);
+                     (%(float_type)s*)PyArray_DATA(bottom) + n * batch_bottom_stride);
         }
         // Restore to previous blas threads
         %(blas_set_num_threads)s(blas_threads_saved);

theano/tensor/nnet/corr3d.py

@@ -51,10 +51,11 @@ class BaseCorr3dMM(gof.OpenMPOp):
                                                 ('DIRECTION_BACKPROP_INPUTS', 'backprop inputs')),  # 2
                              dH=int64, dW=int64, dD=int64,
                              dilH=int64, dilW=int64, dilD=int64,
-                             padH=int64, padW=int64, padD=int64)
+                             padH=int64, padW=int64, padD=int64,
+                             num_groups=int64)
 
     def __init__(self, border_mode="valid", subsample=(1, 1, 1),
-                 filter_dilation=(1, 1, 1), openmp=None):
+                 filter_dilation=(1, 1, 1), openmp=None, num_groups=1):
         super(BaseCorr3dMM, self).__init__(openmp=openmp)
         if isinstance(border_mode, integer_types):
             if border_mode < 0:
@@ -82,6 +83,9 @@ class BaseCorr3dMM(gof.OpenMPOp):
             raise ValueError("filter_dilation must have three elements")
         self.subsample = tuple(subsample)
         self.filter_dilation = tuple(filter_dilation)
+        if num_groups < 1:
+            raise ValueError("Number of groups should be greater than 0")
+        self.num_groups = num_groups
 
         if not theano.config.blas.ldflags:
             # Theano will use a NumPy C implementation of [sd]gemm_ instead.
@@ -127,11 +131,12 @@ class BaseCorr3dMM(gof.OpenMPOp):
     padD = property(lambda self: self.pad[2])
 
     def __str__(self):
-        return '%s{%s, %s, %s}' % (
+        return '%s{%s, %s, %s, %s}' % (
             self.__class__.__name__,
             self.border_mode,
             str(self.subsample),
-            str(self.filter_dilation))
+            str(self.filter_dilation),
+            str(self.num_groups))
 
     @staticmethod
     def as_common_dtype(in1, in2):
@@ -293,6 +298,7 @@ class BaseCorr3dMM(gof.OpenMPOp):
     int padH = %(params)s->padH;
     int padW = %(params)s->padW;
     int padD = %(params)s->padD;
+    int numgroups = %(params)s->num_groups;
 
     PyArrayObject * bottom = %(bottom)s;
     PyArrayObject * weights = %(weights)s;
@@ -428,7 +434,7 @@ class BaseCorr3dMM(gof.OpenMPOp):
         // output is weights: (num_filters, num_channels, height, width, depth)
         // height and width: weights = (bottom + 2*pad - (top - 1) * sample - 1) / dil + 1
         out_dim[0] = (npy_intp)PyArray_DIMS(top)[1];
-        out_dim[1] = (npy_intp)PyArray_DIMS(bottom)[1];
+        out_dim[1] = (npy_intp)PyArray_DIMS(bottom)[1] / numgroups;
         out_dim[2] = (npy_intp)kH;  // already inferred further above
         out_dim[3] = (npy_intp)kW;  // how convenient
         out_dim[4] = (npy_intp)kD;
@@ -454,7 +460,7 @@ class BaseCorr3dMM(gof.OpenMPOp):
         // output is bottom: (batchsize, num_channels, height, width, depth)
         // height and width: bottom = (top - 1) * sample + (weights-1)*dil + 1 - 2*pad
         out_dim[0] = (npy_intp)PyArray_DIMS(top)[0];
-        out_dim[1] = (npy_intp)PyArray_DIMS(weights)[1];
+        out_dim[1] = (npy_intp)PyArray_DIMS(weights)[1] * numgroups;
         out_dim[2] = (npy_intp)((%(height)s != -1) ? %(height)s : (PyArray_DIMS(top)[2] - 1) * dH + (PyArray_DIMS(weights)[2]-1)*dilH + 1 - 2*padH);
         out_dim[3] = (npy_intp)((%(width)s != -1) ? %(width)s : (PyArray_DIMS(top)[3] - 1) * dW + (PyArray_DIMS(weights)[3]-1)*dilW + 1 - 2*padW);
         out_dim[4] = (npy_intp)((%(depth)s != -1) ? %(depth)s : (PyArray_DIMS(top)[4] - 1) * dD + (PyArray_DIMS(weights)[4]-1)*dilD + 1 - 2*padD);
@@ -516,7 +522,8 @@ class BaseCorr3dMM(gof.OpenMPOp):
 
     // Call corr3dMM code
     out2 = corr3dMM(%(bottom)s, %(weights)s, %(top)s, direction,
-                    dH, dW, dD, dilH, dilW, dilD, padH, padW, padD);
+                    dH, dW, dD, dilH, dilW, dilD, padH, padW, padD,
+                    numgroups);
     if (out2==NULL){
        %(fail)s
     }
@@ -592,12 +599,14 @@ class Corr3dMM(BaseCorr3dMM):
         top, = grads
         d_bottom = Corr3dMM_gradInputs(self.border_mode,
                                        self.subsample,
-                                       self.filter_dilation)(weights, top,
-                                                             bottom.shape[-3:])
+                                       self.filter_dilation,
+                                       num_groups=self.num_groups)(weights, top,
+                                                                   bottom.shape[-3:])
         d_weights = Corr3dMM_gradWeights(self.border_mode,
                                          self.subsample,
-                                         self.filter_dilation)(bottom, top,
-                                                               weights.shape[-3:])
+                                         self.filter_dilation,
+                                         num_groups=self.num_groups)(bottom, top,
+                                                                     weights.shape[-3:])
         return d_bottom, d_weights
 
 
@@ -691,11 +700,13 @@ class Corr3dMM_gradWeights(BaseCorr3dMM):
         weights, = grads
         d_bottom = Corr3dMM_gradInputs(self.border_mode,
                                        self.subsample,
-                                       self.filter_dilation)(weights, top,
-                                                             bottom.shape[-3:])
+                                       self.filter_dilation,
+                                       num_groups=self.num_groups)(weights, top,
+                                                                   bottom.shape[-3:])
         d_top = Corr3dMM(self.border_mode,
                          self.subsample,
-                         self.filter_dilation)(bottom, weights)
+                         self.filter_dilation,
+                         num_groups=self.num_groups)(bottom, weights)
         d_height_width_depth = ((theano.gradient.DisconnectedType()(),) * 3
                                 if len(inp) == 5 else ())
         return (d_bottom, d_top) + d_height_width_depth
@@ -738,8 +749,12 @@ class Corr3dMM_gradInputs(BaseCorr3dMM):
                                   as_tensor_variable(shape[1]).astype('int64'),
                                   as_tensor_variable(shape[2]).astype('int64')]
 
-        broadcastable = [topgrad.type.broadcastable[0], kern.type.broadcastable[1],
-                         False, False, False]
+        if self.num_groups > 1:
+            broadcastable = [topgrad.type.broadcastable[0], False,
+                             False, False, False]
+        else:
+            broadcastable = [topgrad.type.broadcastable[0], kern.type.broadcastable[1],
+                             False, False, False]
         dtype = kern.type.dtype
         return Apply(self, [kern, topgrad] + height_width_depth,
                      [TensorType(dtype, broadcastable)()])
@@ -807,12 +822,14 @@ class Corr3dMM_gradInputs(BaseCorr3dMM):
         bottom, = grads
         d_weights = Corr3dMM_gradWeights(self.border_mode,
                                          self.subsample,
-                                         self.filter_dilation)(bottom,
-                                                               top,
-                                                               weights.shape[-3:])
+                                         self.filter_dilation,
+                                         num_groups=self.num_groups)(bottom,
+                                                                     top,
+                                                                     weights.shape[-3:])
         d_top = Corr3dMM(self.border_mode,
                          self.subsample,
-                         self.filter_dilation)(bottom, weights)
+                         self.filter_dilation,
+                         num_groups=self.num_groups)(bottom, weights)
         d_height_width_depth = ((theano.gradient.DisconnectedType()(),) * 3
                                 if len(inp) == 5 else ())
         return (d_weights, d_top) + d_height_width_depth

theano/tensor/nnet/opt.py

@@ -114,7 +114,8 @@ def local_abstractconv3d_gemm(node):
         kern = kern[:, :, ::-1, ::-1, ::-1]
     rval = Corr3dMM(border_mode=node.op.border_mode,
                     subsample=node.op.subsample,
-                    filter_dilation=node.op.filter_dilation)(img, kern)
+                    filter_dilation=node.op.filter_dilation,
+                    num_groups=node.op.num_groups)(img, kern)
     copy_stack_trace(node.outputs[0], rval)
 
     return [rval]
@@ -163,7 +164,8 @@ def local_abstractconv3d_gradweight_gemm(node):
 
     rval = Corr3dMM_gradWeights(border_mode=node.op.border_mode,
                                 subsample=node.op.subsample,
-                                filter_dilation=node.op.filter_dilation)(img, topgrad, shape)
+                                filter_dilation=node.op.filter_dilation,
+                                num_groups=node.op.num_groups)(img, topgrad, shape)
     copy_stack_trace(node.outputs[0], rval)
 
     # need to flip the kernel if necessary
@@ -219,8 +221,9 @@ def local_abstractconv3d_gradinputs_gemm(node):
         kern = kern[:, :, ::-1, ::-1, ::-1]
     rval = Corr3dMM_gradInputs(border_mode=node.op.border_mode,
                                subsample=node.op.subsample,
-                               filter_dilation=node.op.filter_dilation)(kern, topgrad,
-                                                                        shape)
+                               filter_dilation=node.op.filter_dilation,
+                               num_groups=node.op.num_groups)(kern, topgrad,
+                                                              shape)
     copy_stack_trace(node.outputs[0], rval)
 
     return [rval]