Opencv dnn import error for keras pretrained Vgg16 model

edit

@MennoK, You may see that TensorFlow makes a flattening through Reshape layer using dynamically calculated shape:

import tensorflow as tf

# Read the graph.
with tf.gfile.FastGFile('model.pb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

with tf.Session() as sess:
    # Restore session
    sess.graph.as_default()
    tf.import_graph_def(graph_def, name='')
    tf.summary.FileWriter('logs', graph_def)

OpenCV can't do it. There are two possible ways: replace a flatten node by reshape during graph creation or help OpenCV to import it using text graph.

Optimize

Run optimize_for_inference.py tool to remove some extra nodes from the graph:

python ~/tensorflow/tensorflow/python/tools/optimize_for_inference.py \
  --input model.pb \
  --output opt_model.pb \
  --input_names input_1 \
  --output_names predictions/Softmax

Text graph

Run the following script to make a text graph representation:

import tensorflow as tf

# Read the graph.
with tf.gfile.FastGFile('opt_model.pb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

# Remove Const nodes.
for i in reversed(range(len(graph_def.node))):
    if graph_def.node[i].op == 'Const':
        del graph_def.node[i]
    for attr in ['T', 'data_format', 'Tshape', 'N', 'Tidx', 'Tdim',
                 'use_cudnn_on_gpu', 'Index', 'Tperm', 'is_training',
                 'Tpaddings']:
        if attr in graph_def.node[i].attr:
            del graph_def.node[i].attr[attr]

# Save as text.
tf.train.write_graph(graph_def, "", "model.pbtxt", as_text=True)

Flatten

Open model.pbtxt and remove nodes with names flatten/Shape, flatten/strided_slice, flatten/Prod, flatten/stack. Replace a node

node {
  name: "flatten/Reshape"
  op: "Reshape"
  input: "block5_pool/MaxPool"
  input: "flatten/stack"
}

on

node {
  name: "flatten/Reshape"
  op: "Flatten"
  input: "block5_pool/MaxPool"
}

Launch

Now you can remove opt_model.pb and use both model.pb and model.pbtxt in OpenCV:

import cv2 as cv
import numpy as np
net = cv.dnn.readNetFromTensorflow('model.pb', 'model.pbtxt')

inp = np.random.standard_normal([1, 3, 224, 224]).astype(np.float32)
net.setInput(inp)
out = net.forward()