Skip to content

Latest commit

 

History

History
203 lines (123 loc) · 4.77 KB

homework.md

File metadata and controls

203 lines (123 loc) · 4.77 KB
Raw

Homework

In this homework, we'll deploy the dino or dragon model we trained in the previous homework.

Download the model from here:

https://github.com/SVizor42/ML_Zoomcamp/releases/download/dino-dragon-model/dino_dragon_10_0.899.h5

Question 1

Now convert this model from Keras to TF-Lite format.

What's the size of the converted model?

  • 21 Mb
  • 43 Mb
  • 80 Mb
  • 164 Mb

Question 2

To be able to use this model, we need to know the index of the input and the index of the output.

What's the output index for this model?

  • 3
  • 7
  • 13
  • 24

Preparing the image

You'll need some code for downloading and resizing images. You can use this code:

from io import BytesIO
from urllib import request

from PIL import Image

def download_image(url):
    with request.urlopen(url) as resp:
        buffer = resp.read()
    stream = BytesIO(buffer)
    img = Image.open(stream)
    return img


def prepare_image(img, target_size):
    if img.mode != 'RGB':
        img = img.convert('RGB')
    img = img.resize(target_size, Image.NEAREST)
    return img

For that, you'll need to have pillow installed:

pip install pillow

Let's download and resize this image:

https://upload.wikimedia.org/wikipedia/commons/thumb/d/df/Smaug_par_David_Demaret.jpg/1280px-Smaug_par_David_Demaret.jpg

Based on the previous homework, what should be the target size for the image?

Question 3

Now we need to turn the image into numpy array and pre-process it.

Tip: Check the previous homework. What was the pre-processing we did there?

After the pre-processing, what's the value in the first pixel, the R channel?

  • 0.3353411
  • 0.5529412
  • 0.7458824
  • 0.9654902

Question 4

Now let's apply this model to this image. What's the output of the model?

  • 0.17049132
  • 0.39009996
  • 0.60146114
  • 0.82448614

Prepare the lambda code

Now you need to copy all the code into a separate python file. You will need to use this file for the next two questions.

Tip: you can test this file locally with ipython or Jupyter Notebook by importing the file and invoking the function from this file.

Docker

For the next two questions, we'll use a Docker image that we already prepared. This is the Dockerfile that we used for creating the image:

FROM public.ecr.aws/lambda/python:3.9
COPY dino-vs-dragon-v2.tflite .

And pushed it to svizor42/zoomcamp-dino-dragon-lambda:v2.

A few notes:

Question 5

Download the base image svizor42/zoomcamp-dino-dragon-lambda:v2. You can easily make it by using docker pull command.

So what's the size of this base image?

  • 139 Mb
  • 329 Mb
  • 639 Mb
  • 929 Mb

You can get this information when running docker images - it'll be in the "SIZE" column.

Question 6

Now let's extend this docker image, install all the required libraries and add the code for lambda.

You don't need to include the model in the image. It's already included. The name of the file with the model is dino-vs-dragon-v2.tflite and it's in the current workdir in the image (see the Dockerfile above for the reference).

Now run the container locally.

Score this image: https://upload.wikimedia.org/wikipedia/en/e/e9/GodzillaEncounterModel.jpg

What's the output from the model?

  • 0.12
  • 0.32
  • 0.52
  • 0.72

Publishing it to AWS

Now you can deploy your model to AWS!

  • Publish your image to ECR
  • Create a lambda function in AWS, use the ECR image
  • Give it more RAM and increase the timeout
  • Test it
  • Expose the lambda function using API Gateway

This is optional and not graded.

Publishing to Docker hub

This is just for reference, this is how we published our image to Docker hub:

docker build -t zoomcamp-dino-dragon-lambda .
docker tag zoomcamp-dino-dragon-lambda:latest svizor42/zoomcamp-dino-dragon-lambda:v2
docker push svizor42/zoomcamp-dino-dragon-lambda:v2

Submit the results

  • Submit your results here: https://forms.gle/Pnx563ELg9jgjxHX6
  • You can submit your solution multiple times. In this case, only the last submission will be used
  • If your answer doesn't match options exactly, select the closest one

Deadline

The deadline for submitting is 28 November 2022 (Monday), 23:00 CEST (Berlin time).

After that, the form will be closed.