Tensorflow：将图形中的张量保存到文件（或图像）中

Question

I currently have a complex set of operations that iteratively create images and feed them to processes to be used. This is part of a very huge and complex graph. I would like to make sure that those images are being created correctly.

Normally for debugging, we have tf.Print , which creates a no-op with a side-effect of printing to the screen.

Is there either
a) Some way I can store a very large intermediate tensor to a file?
b) Some way to specifically store intermediate tensor images to file (or the screen)?

Obviously, if I have an evaluated numeric tensor this is no problem as I can visualize it using matplotlib's imshow , but as a symbolic tensor it's not so obvious how to do this.

I could save a bunch of intermediate tensors to evaluate with sess.run , but the way my code is architected, it will be very hard to manually access and gather those all.

Answer 1

The best way to save tensors with spatial correlations (image like tensors) in tensorflow is via tf.summary.image . Check out the tensorboard tutorial for how it all fits together, but the crux is like the following:

tf.summary.image("img", img)
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('logs', sess.graph)
for i in range(num_iters):
    summary = sess.run(merged)
    train_writer.add_summary(summary, i)

Then start up tensorboard:

tensorboard --logdir logs

Then go to localhost:6006 in a browser to see the summaries.

The downside to this approach is that you don't have fine grain control over what slice you see unless you explicitly specify (eg tf.summary.image("img", img[...,10]) )