A first draft of "loading data: images"

[MarkDaoust]

It still needs an introduction and conclusion paragraph.

But this won't be visible on the site unless we add it to the "_toc.yaml"

Staging:

https://github.com/MarkDaoust/docs/blob/load_data/site/en/tutorials/load_data/images.ipynb

https://colab.sandbox.google.com/github/MarkDaoust/docs/blob/load_data/site/en/tutorials/load_data/images.ipynb

[random-forests]

Thank you for writing this! Good timing, too. Together with Kiran's notebook on tf records, this would make a great start to a "Loading data" collection.

Suggestions:

Under "Load and format the images"

• Instead of normalizing the image by "image = (image/128.0) - 1", instead, how about "image = image / 255"?

• This would also simplify the code to display the images (instead of "plt.imshow((image+1)/2)", we can write "plt.imshow(image)")

• How about demo'ing the "load_and_format_image" function, just to demonstrate it works like regular Python code when eager is enabled? To do so, you could add a short block under it, that calls the method using something like "load_and_format_image('/root/.keras/datasets/flower_photos/tulips/7166644048_b00a14f01b.jpg', 0)" (I know there's a bunch of output, maybe catch it in a object, show how to get the .numpy() value, and print out the shape

Under "Pipe to a model for training"

• Explain the prefetch parameter.

Under "Quick Transfer learning with keras.Applications"

• Leave the depth_multiplier at the default value.

• Finally, split the dataset into train / test, so we can show a quick evaluation after the transfer learning section, which is really cool.

Overall this is great!

[MarkDaoust]

Instead of normalizing the image by "image = (image/128.0) - 1", instead, how about "image = image / 255"?

Done.
The tf.keras.applications.mobilenet expects data to be normalized to [-1,1].
I've switched it to use [0,1] for most of it, with a transform to [-1,1] in the transfer learning section.

demo'ing the "load_and_format_image"

Done.

Explain the prefetch parameter.

Done

Leave the depth_multiplier at the default value.

Done

Finally, split the dataset into train / test, so we can show a quick evaluation after the transfer learning section, which is really cool.

I'd rather avoid going into too much detail here. I just want to get to "see how easy that was!" and leave the rest for a real tutorial.

[MarkDaoust]

+tf.data

Hey Derek, Jiri,

Does this seem like a reasonable approach to loading image directories with tf.data?

[jsimsa]

Thanks @MarkDaoust did a pass and left some comments.

[mrry]

Thanks for getting this started, Mark!

[mrry]

I think @rohan100jain has some changes in flight that will help this no longer be the case.

[MarkDaoust]

(not really merging, just testing copybara)