Adding an example on handwriting recognition #594
Conversation
| [IAM Dataset](https://fki.tic.heia-fr.ch/databases/iam-handwriting-database) | ||
| that has variable length ground-truths. IAM Dataset is widely used across many OCR | ||
| benchmarks so we hope this example serves as a good starting point. | ||
| """ |
There was a problem hiding this comment.
Can we add a little more description of the dataset, like each sample in the dataset is an image of hand-written sentences, the prediction target of a sample is a string?
| """ | ||
| ## Introduction | ||
|
|
||
| This example shows how the [Captcha OCR](https://keras.io/examples/vision/captcha_ocr/) |
There was a problem hiding this comment.
May add one sentence introduction to what is OCR.
There was a problem hiding this comment.
I don't think this is required since it's a sequel example.
| """ | ||
|
|
||
|
|
||
| class CTCLayer(keras.layers.Layer): |
There was a problem hiding this comment.
Do we have to make the loss a Layer instead of a loss function or Loss subclass?
There was a problem hiding this comment.
Our idea with this example is to keep it as close to the Captcha OCR example as possible while showing the bits that need to be changed. IMO, that helps to ensure a good reading experience. Ccing @AakashKumarNain if he has other points of view.
There was a problem hiding this comment.
When defining a normal loss function, we generally assume that the shape of y_true and y_pred are same which isn't the case here. Defining it as an endpoint layer makes it easy to calculate the length of inputs and the labels on the fly during training, and then pass the same to the ctc_batch_cost(...)
|
@haifeng-jin, |
|
@sayakpaul Sure, just ping me when it's ready. Thank you! |
|
@haifeng-jin now, it's good to go for another round of review. We incorporated Edit Distance as an evaluation metric. |
There was a problem hiding this comment.
Awesome example! Very strong follow-up to the original OCR example.
I added a few minor copyedits. Please add the generated files. Thanks @haifeng-jin for the review.
|
@fchollet added the generated files. The only major change is the reduced number of epochs because my system was getting stalled after 10 epochs. Tried on a commodity GCP Notebook instance too but didn't help much. Cc: @AakashKumarNain |
|
Thanks for the review @haifeng-jin @fchollet . @sayakpaul should we run this on a bigger VM (only if it is required)? |
|
I don't think that's required given that we have explicitly noted the number of epochs should be at least 50. However, if you want to do it go right ahead. |
|
Yes, I think the restriction should be fine. Thank you! |
|
@sayakpaul @AakashKumarNain there were only 2 image files included in the PR, but the example is intended to have 3 figures (it's missing the last one). Please add the missing figure (in a new PR) |
A Colab Notebook is available here.
Cc: @AakashKumarNain