Fast.ai Image Classification

Example Dataset
We're going to use udt-dataset-cats-and-dogs, a dataset of labeled images of cats and dogs created from COCO. For this guide you don't need to download it directly, because we'll load it in right from our notebook.
Don't like cats and dogs? You can also use any classification from the Common Objects in Context with the Import COCO button! Maybe try to classify bears vs cats!
coco_dogs_and_cats.udt.csv
Import CSV Into Pandas Dataframe
We can begin by importing the fastai library, pandas, and our udt.csv file.
1
from fastai.vision import *
2
import pandas as pd
3
​
4
url_to_csv = "https://raw.githubusercontent.com/UniversalDataTool/udt-dataset-cats-and-dogs/master/coco_dogs_and_cats.udt.csv"
5
udt_csv = pd.read_csv(url_to_csv)
Copied!
You can use the udt.json format too, tables are just a nice way to visualize the data!
Download Images
UDT Datasets just have links to images, so we'll need to download the actual images. Let's do that using the fast.ai download_images function.
1
# Get the lines of our CSV that have sample data
2
samples = udt_csv[udt_csv["path"].str.contains("samples.")]
3
​
4
# Create two csvs that just have
5
# our cat image urls and dog image curls
6
​
7
dog_samples = samples[samples["annotation"] == "dog"]
8
cat_samples = samples[samples["annotation"] == "cat"]
9
​
10
dog_samples.to_csv("dog_urls.csv", columns=["imageUrl"], header=False, index=False)
11
cat_samples.to_csv("cat_urls.csv", columns=["imageUrl"], header=False, index=False)
Copied!
dog_urls.csv
1
# Now we can download the images!
2
download_images("dog_urls.csv", "images/dog" , max_pics=500)
3
download_images("cat_urls.csv", "images/cat" , max_pics=500)
4
​
5
# Let's make sure all the images are readable
6
verify_images("images/dog", delete=True, max_size=500)
7
verify_images("images/cat", delete=True, max_size=500)
Copied!
Create an ImageDataBunch
From here, everything should should seem pretty normal. We can create an ImageDataBunch from our images directory.
1
data = ImageDataBunch.from_folder("images", train=".", valid_pct=0.2, ds_tfms=get_transforms(), size=224, num_workers=4).normalize(imagenet_stats)
2
​
3
# Let's take a look at the data
4
data.show_batch(rows=3, figsize=(7,8))
Copied!
show_batch output
Train a Model
We can now train a model! This is just a simple one, don't forget to fine tune!
1
learn = cnn_learner(data, models.resnet34, metrics=error_rate)
2
learn.fit_one_cycle(4)
Copied!
fit_one_cycle output

Machine Learning - Previous

Fast.ai

Fast.ai Image Segmentation

Last modified 1yr ago

Copy link

Contents

Example Dataset

Import CSV Into Pandas Dataframe

Download Images

Create an ImageDataBunch

Train a Model