fix bn fuse for fp16

Author: guschmue

tf2onnx/optimizer/back_to_back_optimizer.py

@@ -201,12 +201,13 @@ def _optimize_conv_batchnorm_fusion(g, node, consumer_nodes):
         if len(weights.shape) != 4:
             return []
 
-        bias = 0
         # optional bias value
         if len(node.inputs) > 2:
             if not node.inputs[2].is_const():
                 return []
             bias = node.inputs[2].get_tensor_value(as_list=False)
+        else:
+            bias = np.array(0, dtype=weights.dtype)
 
         # scale, offset, mean, var be const, otherwise skip
         if False in [node2.inputs[i].is_const() for i in [1, 2, 3, 4]]:
@@ -228,8 +229,8 @@ def _optimize_conv_batchnorm_fusion(g, node, consumer_nodes):
         weights_new = weights * scale_new
         weights_new = weights_new.transpose(3, 2, 0, 1)
         bias_new = (bias - mean) * scale_new + offset
-        bias_new_const = g.make_const(node.name + '_bias_fused_bn', bias_new)
-        weights_new_const = g.make_const(node.name + '_weights_fused_bn', weights_new)
+        bias_new_const = g.make_const(node.name + '_bias_fused_bn', bias_new.astype(bias.dtype))
+        weights_new_const = g.make_const(node.name + '_weights_fused_bn', weights_new.astype(weights.dtype))
         g.replace_inputs(node, [node.input[0], weights_new_const.output[0], bias_new_const.output[0]])
 
         # fuse conv and bn, delete bn