Tensorflow：Inception v3批处理

Question

I'm extracting bottleneck tensors with Inception v3. My problem: I can feed only one image at a time:

sess.graph.get_tensor_by_name("pool_3:0").eval(session=sess, feed_dict={'DecodeJpeg:0':single_image})

Batched Processing of multiple images would speed-up things quite a bit, I guess. A solution is suggested here , but I cannot get it to work (tested with tensorflow v0.10.0 and 0.11.0rc0, inception model downloaded from http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz ).

import tensorflow as tf
from tensorflow.python.platform import gfile
import cv2
import numpy as np

def create_graph():
    with gfile.FastGFile('classify_image_graph_def.pb', 'rb') as f:
        graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    _ = tf.import_graph_def(graph_def, name='')

create_graph()

img = cv2.imread("some_img.jpg")
img = cv2.resize(img, (299,299), interpolation = cv2.INTER_CUBIC) 
# replicate image 10 times
img = np.array(10*[img]).astype('float')

with tf.Session() as sess:
    pooled_2 =  sess.graph.get_tensor_by_name("pool_3:0").eval(session=sess, feed_dict={'ResizeBilinear:0':img})

This gives me the (plausible) error message:

Traceback (most recent call last):
  File "extract_bottlenecks_minimal.py", line 26, in <module>
    pooled_2 = sess.graph.get_tensor_by_name("pool_3:0").eval(session=sess, feed_dict={'ResizeBilinear:0':img})
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 559, in eval
    return _eval_using_default_session(self, feed_dict, self.graph, session)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 3761, in _eval_using_default_session
    return session.run(tensors, feed_dict)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 717, in run
    run_metadata_ptr)
  File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 894, in _run
    % (np_val.shape, subfeed_t.name, str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (10, 299, 299, 3) for Tensor u'ResizeBilinear:0', which has shape '(1, 299, 299, 3)'

I also found this issue claiming that inception works with image batches, but this code needs to set up the whole network every time I feed it input (I would like to setup the the network only once).

Thanks for taking interest - any help is greatly appreciated ;-)

Answer 1

(1) Please try using the much more recent Inception v3 model, which you can find here:

http://download.tensorflow.org/models/image/imagenet/inception-v3-2016-03-01.tar.gz

This was the version released subsequently to the bug I opened in Feb, which should support a batch dimension on the evaluation path. The imagenet_eval script should work with this.

(2) You might need to decode the JPEG and resize to 299x299 before you feed into the network. I haven't looked at whether the pretrained one accepts a variable batch size to the decode, but it should at least accept it to the network itself.

For more information, see the image recognition tutorial.