Revision history [back]

UPDATE

@Paul Kuo, the following are steps to create a graph without training-testing switches. An extra steps are required to import it into OpenCV.

Step 1: Create a graph definition

Make a script with the following code at the root folder of Age-Gender-Estimate-TF.

import tensorflow as tf import inception_resnet_v1

inp = tf.placeholder(tf.float32, shape=[None, 160, 160, 3], name='input_image')
age_logits, gender_logits, _ = inception_resnet_v1.inference(inp, keep_probability=0.8,
                                                             phase_train=False,
                                                             weight_decay=1e-5)
print age_logits  # logits/age/BiasAdd
print gender_logits # logits/gender/BiasAdd

with tf.Session() as sess:
    graph_def = sess.graph.as_graph_def()
    tf.train.write_graph(graph_def, "", 'inception_resnet_v1.pb', as_text=False)

Here we create a graph definition for testing mode only (phase_train=False).

Step 2: Freeze resulting graph with checkpoint.

python ~/tensorflow/tensorflow/python/tools/freeze_graph.py \
    --input_graph=inception_resnet_v1.pb \
    --input_checkpoint=savedmodel.ckpt \
    --output_graph=frozen_inception_resnet_v1.pb \
    --output_node_names="logits/age/BiasAdd,logits/gender/BiasAdd" \
    --input_binary

Step 3: Checking

Using TensorBoard, check that our graph has no training subgraphs (compare with images above). C:\fakepath\Screenshot from 2018-03-30 10-41-03.png

Step 4: Help OpenCV to import graph.

Unfortunately, current version of OpenCV cannot interpret this graph correctly because of single Reshape layer that takes dynamically estimated target shape: C:\fakepath\Screenshot from 2018-03-30 10-30-43.png

Actually, it's just a flattening that means reshaping from 4-dimensional blob to 2-dimensional keeping the same batch size. We replace manage it during graph definition because it's out of user's code:

# inception_resnet_v1.py, line 262:
net = slim.fully_connected(net, bottleneck_layer_size, activation_fn=None,
                           scope='Bottleneck', reuse=False)

But we can help OpenCV to manage it by modifying a text graph.

Step 5: Create a text graph

import tensorflow as tf

# Read the graph.
with tf.gfile.FastGFile('frozen_inception_resnet_v1.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, "", "frozen_inception_resnet_v1.pbtxt", as_text=True)

Step 6: Modify a text graph

Remove the Shape node

node {
  name: "InceptionResnetV1/Bottleneck/BatchNorm/Shape"
  op: "Shape"
  input: "InceptionResnetV1/Bottleneck/MatMul"
  attr {
    key: "out_type"
    value {
      type: DT_INT32
    }
  }
}

Replace Reshape to Identity:

from

node {
  name: "InceptionResnetV1/Bottleneck/BatchNorm/Reshape_1"
  op: "Reshape"
  input: "InceptionResnetV1/Bottleneck/BatchNorm/FusedBatchNorm"
  input: "InceptionResnetV1/Bottleneck/BatchNorm/Shape"
}

to

node {
  name: "InceptionResnetV1/Bottleneck/BatchNorm/Reshape_1"
  op: "Identity"
  input: "InceptionResnetV1/Bottleneck/BatchNorm/FusedBatchNorm"
}

Also replace node

from

node {
  name: "InceptionResnetV1/Bottleneck/BatchNorm/Reshape"
  op: "Reshape"
  input: "InceptionResnetV1/Bottleneck/MatMul"
  input: "InceptionResnetV1/Bottleneck/BatchNorm/Reshape/shape"
}

to

node {
  name: "InceptionResnetV1/Bottleneck/BatchNorm/Reshape"
  op: "Identity"
  input: "InceptionResnetV1/Bottleneck/MatMul"
}

Additionally, remove nodes with names InceptionResnetV1/Logits/Flatten/flatten/Shape InceptionResnetV1/Logits/Flatten/flatten/strided_slice InceptionResnetV1/Logits/Flatten/flatten/Reshape/shape

Replace node with a name InceptionResnetV1/Logits/Flatten/flatten/Reshape

from

node {
  name: "InceptionResnetV1/Logits/Flatten/flatten/Reshape"
  op: "Reshape"
  input: "InceptionResnetV1/Logits/AvgPool_1a_8x8/AvgPool"
  input: "InceptionResnetV1/Logits/Flatten/flatten/Reshape/shape"
}

to

node {
  name: "InceptionResnetV1/Logits/Flatten/flatten/Reshape"
  op: "Flatten"
  input: "InceptionResnetV1/Logits/AvgPool_1a_8x8/AvgPool"
}

Step 7: check an accuracy

graph = 'frozen_inception_resnet_v1.pb'
config = 'frozen_inception_resnet_v1.pbtxt'
cvNet = cv.dnn.readNetFromTensorflow(graph, config)

with tf.gfile.FastGFile(graph) 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='')

    np.random.seed(324)
    inp = np.random.standard_normal([1, 160, 160, 3]).astype(np.float32)

    out = sess.run([sess.graph.get_tensor_by_name('logits/age/BiasAdd:0'),
                     sess.graph.get_tensor_by_name('logits/gender/BiasAdd:0')],
                   feed_dict={'input_image:0': inp})

    cvNet.setInput(inp.transpose(0, 3, 1, 2))
    cvOut = cvNet.forward(['logits/age/MatMul', 'logits/gender/MatMul'])

print np.max(np.abs(cvOut[0] - out[0]))
print np.max(np.abs(cvOut[1] - out[1]))

Note that this is a kind of rude solution but at least that works with the latest master branch. You may download resulting .pbtxt here. BTW you also can skip all the steps above and use frozenGraph_ageGender.pb with this text file (cvNet = cv.dnn.readNetFromTensorflow('frozenGraph_ageGender.pb', 'frozen_inception_resnet_v1.pbtxt')).