trained model error about mobilenet HOT 7 CLOSED

I assume that you have make the tfrecord correctly. (Both test set and train set)

And you said that you used 'eval_image_classifier.py' to do the evaluation and got a low accuracy.

Then you write a new evaluation script??? And it come up with 'MobileNet/Conv1/Weight2 is not initialized'??? Right?

from mobilenet.

To figure out whether there is any problem with the evaluation procedure, I think you could use the training set to do the evaluation. If the accuracy is high (as the training loss is low enough), it means there is no problem with the evaluation program.

from mobilenet.

Sorry，I have check the test img , the bbox is wrong，
now the test Accuracy is normal
2017-05-26 20:05:00.298230: I tensorflow/core/kernels/logging_ops.cc:79] eval/Accuracy[0.99077183]
2017-05-26 20:05:00.298274: I tensorflow/core/kernels/logging_ops.cc:79] eval/Recall_5[0.99731543624161079]

but how can i extract the avg_pool_15 feature? Thank you

from mobilenet.

Glad to hear that~

To extract the feature, you can use

end_point['avg_pool_15']

from mobilenet.

Are you applying MobileNet to face recognition task? The result seems quite well.

from mobilenet.

I am on person re-id task
evaluate step as
slim.evaluation.evaluate_once(
master=FLAGS.master,
checkpoint_path=checkpoint_path,
logdir=FLAGS.eval_dir,
num_evals=num_batches,
eval_op=list(names_to_updates.values()),
variables_to_restore=variables_to_restore)
how to get end_point by this function?

from mobilenet.

Cool~
It is a high level API. You have to hack into it or write a new script to extract the feature. It is not so difficult to write one.

from mobilenet.