How to Build and Deploy a Deep-Learning Multi-Classifier Web App (Part 1)

Create a simple neural network model
machinelearning
ai
Author

Tony Phung

Published

January 24, 2024

I’ve just created my first multi-category classifier using Jeremy Howard’s popular fast ai which is an astraction layer library built on top of the most world’s used deep-learning library PyTorch. I’ve documented the process including the issues I faced (i.e. bugs)

I found it more easier to digest and understand this process by splitting the steps into 3 parts:

Part 1: Create Learner (.pkl file)
Part 2: Create Gradio application file (app.py)
Part 3: Host on HuggingFace

Part 1: Create Learner (.pkl file)

1.1 Install and import libraries

!pip install timm
!pip install fastai 
from fastai.vision.all import *
import timm

1.2 Download Pets Breed Data

path = untar_data(URLs.PETS)/'images'

1.3 Create Data Loader

1.3.1 (A different) Labelling Function

Here a different method to label our data was used:

  • In ‘noodles vs rice’ model: There were two parent folders separating two categories of data: get_y=parent_label

  • In ‘saving a fast ai’ model: There was a custom labelling function that looked for capital letters for cat breeds def is_cat(x): return x[0].isupper()

  • In this model, I used Regex to find breed names before the last ’_’ in the file name: label_func=RegexLabeller(pat=r'^([^/]+)_\d). See show_batch() output to see the file names examples.

Did you notice this is the same dataset as the is_cat model? So changing our label resulted in a different model!

1.3.2 Data Loader Code

pets_dataloaders =  ImageDataLoaders.from_name_func(
    '.',
    get_image_files(path),
    valid_pct=0.2,
    seed=42,
    label_func=RegexLabeller(pat=r'^([^/]+)_\d+'),
    item_tfms=Resize(224))
c:\Users\tonyp\miniconda3\envs\fastai\Lib\site-packages\fastai\torch_core.py:263: UserWarning: 'has_mps' is deprecated, please use 'torch.backends.mps.is_built()'
  return getattr(torch, 'has_mps', False)

1.4 Batch Examples, Create Learner, Fine-Tune and Export

I grouped these steps as the code are exactly the same in previous posts.

1.4.1 Batch Examples

This function is also a good way to find out what is the file name structure if we were not sure.

pets_dataloaders.show_batch(max_n=8)

1.4.2 Create Learner

I am still using resnet model architecture for starters for reasons mentioned previously by Jeremy Howard

pets_learner = vision_learner(pets_dataloaders, resnet34, metrics=error_rate)

1.4.3 Fine-Tune

pets_learner.fine_tune(3)

1.4.4 Export

pets_learner.export('pets_learner.pkl')

To be Continued…

Part 1: Create Learner (.pkl file)
Part 2: Create Gradio application file (app.py)
Part 3: Host on HuggingFace