Revision history [back]

Use optimize_for_inference.py tool to remove unused nodes. Your model receives 784-dimensional input vector then reshapes it to 28x28 image. We also remove this reshape to prevent conflicts between TensorFlow's NHWC and OpenCV's NCHW data layouts:

python ~/tensorflow/tensorflow/python/tools/optimize_for_inference.py \
  --input frozen_graph.pb \
  --output frozen_graph_opt.pb \
  --input_names "Reshape" \
  --output_names "softmax" \
  --frozen_graph

Test:

graph = 'frozen_graph_opt.pb'
cvNet = cv.dnn.readNet(graph)

with tf.gfile.FastGFile(graph) as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

with tf.Session() as sess:
    sess.graph.as_default()
    tf.import_graph_def(graph_def, name='')

    np.random.seed(234)
    inp = np.random.standard_normal([1, 28, 28, 1]).astype(np.float32)
    out = sess.run(sess.graph.get_tensor_by_name('softmax:0'),
                   feed_dict={'Reshape:0': inp})

    cvNet.setInput(inp.transpose(0, 3, 1, 2))
    cvOut = cvNet.forward()

    print np.max(np.abs(cvOut - out))

gives 1.1920929e-07 maximal absolute difference.