Reading in jpeg files and displaying them in Tensorboard

[Durham Smith]

Hi all I am wondering if anyone can help. 

My goal is to read in a batch of jpg files and convert them to a tensor of shape [height, width, channels] where channels =3. 

I then would like to display use tensorboard to check if these images have loaded correctly.

As far as I understand I should do the following:

import tensorflow as tf
import os

def read_image_data():
 dirname, filename = os.path.split(os.path.abspath(__file__))
 #Create a list of filenames (for whaledata they are jpegs)
 filenames = ['/data_dir/w_7489.jpg']
 #Create a queue that produces the filenames to read
 filename_queue = tf.train.string_input_producer(filenames)
 #Create a reader for the filequeue
 reader = tf.WholeFileReader()
 #Read in the files
 key, value = reader.read(filename_queue)
 #Convert the Tensor(of type string) to representing the Tensor of type uint8
 # and shape [height, width, channels] representing the images
 images = tf.image.decode_jpeg(value, channels=3)
 #convert images to floats and attach image summary
 float_images = tf.cast(images, tf.float32)
 tf.image_summary('images', float_images)
 #Create session
 sess = tf.Session()
 summary_op = tf.merge_all_summaries()
 tf.initialize_all_variables()
 #Write summary
 summary_writer = tf.train.SummaryWriter(dirname+'/log/', graph_def=sess.graph_def)
 summary_str = sess.run(summary_op)
 summary_writer.add_summary(summary_str)
 #Close session
 sess.close()


def main(argv=None):
 read_image_data()
 return 0

if __name__ == '__main__':
 tf.app.run()

When I execute this code the following happens:
1) The /log/ directory and an event file called events.out.tfevents.1447627799.Durhams-MacBook-Pro.local is created
2) The program never finishes execution (I need to kill -9 <pid> it)

Any insight into solving this would be great.

Additionally it would be great if someone can tell me how to handle cases where the jpegs I need to read in do not have the same dimensions.

[Vijay Vasudevan]

I suspect the section called "Creating threads to prefetch using QueueRunner objects" in http://tensorflow.org/how_tos/reading_data/index.md#QueueRunners may be useful.  My guess is that you need to spawn a few threads / queue runners to decouple the running of the queue from the processing of the results.

For your second question, http://tensorflow.org/api_docs/python/image.md#AUTOGENERATED-resizing will probably help.

--
You received this message because you are subscribed to the Google Groups "Discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to discuss+u...@tensorflow.org.
To post to this group, send email to dis...@tensorflow.org.
To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/126e2ffe-e6d8-4e55-b2d9-09717d6569f9%40tensorflow.org.

[martin...@gmail.com]

Hey, I got here from a stackoverflow thread and while this helped, I wanted you to know that the links in your reply 404 - this should fix the first link: http://tensorflow.org/how_tos/reading_data/index.html#QueueRunner, but I have no idea about the second!

[durham...@gmail.com]

Thanks for the help Vijay and Martin!

For anyone that has the same issue in future two things are needed to be changed from the code posted above. They are:

1. Add a batchsize to float_images in my code above, if you want to test it with only one image you can use the expand_dims function.
2. Start the queue runner [ tf.train.start_queue_runners(sess=sess)]

Martin as for the 2nd link not working you can find it going to the api page ->python->images->resizing (I've notice some issues when linking to the pages containing the api quite often now).

[Michele Campus]

Noticed that this question refer to here, I'd like to know how is possibile that the answer are consider correct.
I tried and the code does not properly work if you don't modify the "eval" line like this: image = my_img.eval(session=sess).

Is there something that I miss ?

Thanks

[Martin Tutek]

Hi,

it is not mentioned in the post, however the problem was posted for tensorflow version 0.5 (current stable vesion is 0.8). So probably there will be some syntax differences - so image = my_img.eval(session=sess)   is the only change required?

I'll edit this ASAP,

Cheers,

--
You received this message because you are subscribed to a topic in the Google Groups "Discuss" group.
To unsubscribe from this topic, visit https://groups.google.com/a/tensorflow.org/d/topic/discuss/ik7hSLIKEBM/unsubscribe.
To unsubscribe from this group and all its topics, send an email to discuss+u...@tensorflow.org.

To post to this group, send email to dis...@tensorflow.org.

To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/discuss/71ce2932-dfe6-458a-8c29-d3d989b66bd8%40tensorflow.org.

[Michele Campus]

Glad to help you.
I also have problem with Image.show().

Image.show(Image.fromarray(np.asarray(image)))

AttributeError: 'module' object has no attribute 'show'

--

Michele Campus