Tensorflow slow in extracting pool3 features from a retrained model

Question

I used the retrain code to train the final layer of the Inception model on my own images. But on my CPU machine, it is taking almost 0.5-0.8 second to extract the pool3 features.

pool3 = sess.graph.get_tensor_by_name('pool_3:0')
pf = sess.run(pool3, {png_data: imgBuffer.getvalue()})

The above lines of code take all the time. Is it because of my CPU memory? Any way to make it faster?

Was it faster with the unmodified code (i.e. without retraining)? — Olivier Moindrot, Jun 21 '16 at 12:39

score 0 · Answer 1 · edited May 23 '17 at 10:29

0

Tensorflow uses lazy evaluation; basically all nodes in graph are evaluated on a need basis at sess.run().

This is why you noticed most of time is spent by sess.run and it is not that pool3 in particular is more expensive

You can use Timeline to find detailed time breakdown.

edited May 23 '17 at 10:29

Community

1
1

answered Jun 22 '16 at 18:10

Yao Zhang

5,591
1
16
22

Tensorflow slow in extracting pool3 features from a retrained model

1 Answers1