Vertex AI Image Classifier with Explainable Predictions (solution notebook)

[kylesteckler]

Solution Notebook for XAI Image Lab with Vertex. #165

• Explore flower dataset
• Build/package tensorflow model to train on Vertex AI
• Train on Vertex AI
• Upload/deploy model to Vertex AI
• Serve predictions and visualize feature attributions from IG

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:07Z
----------------------------------------------------------------

• "with Explanations" -> "with explanations"

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:08Z
----------------------------------------------------------------

Consider grouping all the imports at the beginning as in a standard Python file, so that all the notebook dependencies are immediately clear.

New comment: Did you run pre-commit? The import ordering within groups does not seem to be sorted by alphabetical order.

_sanjanalreddy commented on 2022-03-04T23:51:45Z_ ----------------------------------------------------------------

+1

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:08Z
----------------------------------------------------------------

Maybe shorten that cell into the following one-liner:

!gsutil ls -d | grep -w gs://${BUCKET}/ || gsutil mb -l ${REGION} gs://${BUCKET}

kylesteckler commented on 2022-03-08T20:45:09Z
----------------------------------------------------------------

When I run this one-liner with BUCKET='kylesteckler-demo' I get an error:

InvalidUrlError: Invalid bucket name in URL "-demo".

BenoitDherin commented on 2022-03-09T18:46:20Z
----------------------------------------------------------------

Have you tried quoting the bucket path as follows:

!gsutil ls -d | grep -w "gs://${BUCKET}/" || gsutil mb -l ${REGION} "gs://${BUCKET}"

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:09Z
----------------------------------------------------------------

• Have all the imports at the beginning to the notebook (see comment above).

• Instead of creating an alias, why not importing directly aiplatform, which is more readable than the acronym aip:

from google.cloud import aiplatform

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:10Z
----------------------------------------------------------------

* "shuffle, split and copy" -> "shuffle, split, and copy"

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:10Z
----------------------------------------------------------------

IMAGE_SIZE and BATCH_SIZE seem not to be used in this cell, maybe move them where we start using them?

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:11Z
----------------------------------------------------------------

• "task.py" -> "task.py"

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:12Z
----------------------------------------------------------------

Suggestion: What about using fire in order to parse the command line argument? Doing that would essentially get rid of the task.py.

takumiohym commented on 2022-03-07T03:10:34Z
----------------------------------------------------------------

+1

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:13Z
----------------------------------------------------------------

• The code of this cell does not seem pep8 compliant (lines too long, only one space between functions, for instance). Have you passed it in a pep8 formater?(I notice that sometimes pre-commit fails to clean up code in notebook cells.)

• What about importing the layers you need at the beginning of the file to avoid all the verbosity of having tf.keras.layers repeated at each layer? This way the code would be shorter and the model clearer to comprehend at a glance.

• Same remark for tf.keras.Sequential

sanjanalreddy commented on 2022-03-04T23:55:32Z
----------------------------------------------------------------

Just a brief comment on what the architecture of the model is would be helpful in the markdown above. Since we aren't focusing on model building here, it might be useful to mention what kind of model we are using.

takumiohym commented on 2022-03-07T03:12:20Z
----------------------------------------------------------------

In terms of the model, how about just doing transfer learning?

That can shorten the training time, and get better result (and better explainability).

BenoitDherin commented on 2022-03-09T18:50:17Z
----------------------------------------------------------------

New comment: Did you run pre-commit? The code still doesn't look fully pep8 compliant (e.g., only one line between functions, lines looking too long)

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:13Z
----------------------------------------------------------------

Is is possible to escape the quote as follows so that the yaml files is more readable:

echo > ./config.yaml "\
workerPoolSpecs:
  machineSpec:
    machineType: n1-standard-8
  replicaCount: 1
  pythonPackageSpec:

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

BenoitDherin commented on 2022-03-04T23:09:14Z
----------------------------------------------------------------

Maybe adding something about what is going on in the following code could be useful since it is not obvious to understand:

m_call = tf.function(local_model.call).get_concrete_function(
    [
        tf.TensorSpec(
            shape=[None, 192, 192, 3], dtype=tf.float32, name=CONCRETE_INPUT
        )
    ]
)

and also why the xai_model needs it.

sanjanalreddy commented on 2022-03-04T23:59:21Z
----------------------------------------------------------------

Agree. It's not clear what the purpose is.

[sanjanalreddy]

+1

---

View entire conversation on ReviewNB

[sanjanalreddy]

Just a brief comment on what the architecture of the model is would be helpful in the markdown above. Since we aren't focusing on model building here, it might be useful to mention what kind of model we are using.

---

View entire conversation on ReviewNB

[sanjanalreddy]

Agree. It's not clear what the purpose is.

---

View entire conversation on ReviewNB

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

sanjanalreddy commented on 2022-03-05T00:04:17Z
----------------------------------------------------------------

Add a note that restart of the kernel is needed

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

sanjanalreddy commented on 2022-03-05T00:04:18Z
----------------------------------------------------------------

Would be better to set the PROJECT ID using this to have consistency across notebooks in the repo

PROJECT = !(gcloud config get-value core/project)

PROJECT = PROJECT[0]

BUCKET = PROJECT # defaults to PROJECT

REGION = "us-central1" # Replace with your REGION

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

sanjanalreddy commented on 2022-03-05T00:04:19Z
----------------------------------------------------------------

Add a line on why we have to denormalize the image here

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

sanjanalreddy commented on 2022-03-05T00:04:19Z
----------------------------------------------------------------

Add a note above this cell that we are using

endpoint.explian() here to get the explanations and what steps should be followed when we need to get explanations.

[sanjanalreddy]

Thanks for updating this. Just added minor suggestions to improve the instructions!

[BenoitDherin]

Added a few minor comments. Thanks for updating the lab.

[takumiohym]

+1

---

View entire conversation on ReviewNB

[takumiohym]

In terms of the model, how about just doing transfer learning?

That can shorten the training time, and get better result (and better explainability).

---

View entire conversation on ReviewNB

[kylesteckler]

When I run this one-liner with BUCKET='kylesteckler-demo' I get an error:

InvalidUrlError: Invalid bucket name in URL "-demo".

---

View entire conversation on ReviewNB

[BenoitDherin]

Have you tried quoting the bucket path as follows:

!gsutil ls -d | grep -w "gs://${BUCKET}/" || gsutil mb -l ${REGION} "gs://${BUCKET}"

---

View entire conversation on ReviewNB

[BenoitDherin]

New comment: Did you run pre-commit? The code still doesn't look fully pep8 compliant (e.g., only one line between functions, lines looking too long)

---

View entire conversation on ReviewNB

[BenoitDherin]

LGTM. nit: It seems at time the code is not fully pep8 compliant.

[review-notebook-app]

View / edit / reply to this conversation on ReviewNB

sanjanalreddy commented on 2022-03-09T19:02:27Z
----------------------------------------------------------------

Looks like the confidence fell drastically using the tfhub model but nothing to worry about

[sanjanalreddy]

LGTM

[takumiohym]

LGTM! Thanks for updating this.