add the example symbol of modulated deformable RoIpooling layer

DCNv2_op/example_symbol.py

@@ -1,7 +1,9 @@
+import mx
+
 """
 demo symbol of using modulated deformable convolution
 """
-def modulated_deformable_conv(data, name, num_filter, stride, lr_mult=1):
+def modulated_deformable_conv(data, name, num_filter, stride, lr_mult=0.1):
     weight_var = mx.sym.Variable(name=name+'_conv2_offset_weight', init=mx.init.Zero(), lr_mult=lr_mult)
     bias_var = mx.sym.Variable(name=name+'_conv2_offset_bias', init=mx.init.Zero(), lr_mult=lr_mult)
     conv2_offset = mx.symbol.Convolution(name=name + '_conv2_offset', data=data, num_filter=27,
@@ -14,3 +16,49 @@ def modulated_deformable_conv(data, name, num_filter, stride, lr_mult=1):
                        num_filter=num_filter, pad=(1, 1), kernel=(3, 3), stride=stride, 
                        num_deformable_group=1, no_bias=True)
     return conv2
+
+"""
+demo symbol of using modulated deformable RoI pooling
+"""
+def modulated_deformable_roi_pool(data, rois, spatial_scale, imfeat_dim=256, deform_fc_dim=1024, roi_size=7, trans_std=0.1):
+    roi_align = mx.contrib.sym.DeformablePSROIPooling(name='roi_align', 
+                        data=data,
+                        rois=rois,
+                        group_size=1,
+                        pooled_size=roi_size, 
+                        sample_per_part=2,
+                        no_trans=True, 
+                        part_size=roi_size, 
+                        output_dim=imfeat_dim,
+                        spatial_scale=spatial_scale)
+
+    feat_deform = mx.symbol.FullyConnected(name='fc_deform_1', data=roi_align, num_hidden=deform_fc_dim)
+    feat_deform = mx.sym.Activation(data=feat_deform, act_type='relu', name='fc_deform_1_relu')
+
+    feat_deform = mx.symbol.FullyConnected(name='fc_deform_2', data=feat_deform, num_hidden=deform_fc_dim)
+    feat_deform = mx.sym.Activation(data=feat_deform, act_type='relu', name='fc_deform_2_relu')
+
+    feat_deform = mx.symbol.FullyConnected(name='fc_deform_3', data=feat_deform, num_hidden=roi_size * roi_size * 3)
+
+    roi_offset = mx.sym.slice_axis(feat_deform, axis=1, begin=0, end=roi_size * roi_size * 2)
+    roi_offset = mx.sym.reshape(roi_offset, shape=(-1, 2, roi_size, roi_size))
+
+    roi_mask = mx.sym.slice_axis(feat_deform, axis=1, begin=roi_size * roi_size * 2, end=None)
+    roi_mask_sigmoid = mx.sym.Activation(roi_mask, act_type='sigmoid')
+    roi_mask_sigmoid = mx.sym.reshape(roi_mask_sigmoid, shape=(-1, 1, roi_size, roi_size))
+
+    deform_roi_pool = mx.contrib.sym.DeformablePSROIPooling(name='deform_roi_pool', 
+                        data=data,
+                        rois=rois, 
+                        trans=roi_offset,
+                        group_size=1,
+                        pooled_size=roi_size, 
+                        sample_per_part=2,
+                        no_trans=False, 
+                        part_size=roi_size, 
+                        output_dim=imfeat_dim,
+                        spatial_scale=spatial_scale, 
+                        trans_std=trans_std)
+
+    modulated_deform_roi_pool = mx.sym.broadcast_mul(deform_roi_pool, roi_mask_sigmoid)
+    return modulated_deform_roi_pool