I had a pretrained Dense-121 model from Pytorch and export it as onnx format. After that I load the onnx file and used prepare function from onnx_tf.backend. I finally used