Fast.ai Image Classification

Quickly import *.udt.csv files into fast.ai for 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.

from fastai.vision import *
import pandas as pd

url_to_csv = "https://raw.githubusercontent.com/UniversalDataTool/udt-dataset-cats-and-dogs/master/coco_dogs_and_cats.udt.csv"
udt_csv = pd.read_csv(url_to_csv)

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.

# Get the lines of our CSV that have sample data
samples = udt_csv[udt_csv["path"].str.contains("samples.")]

# Create two csvs that just have
# our cat image urls and dog image curls

dog_samples = samples[samples["annotation"] == "dog"]
cat_samples = samples[samples["annotation"] == "cat"]

dog_samples.to_csv("dog_urls.csv", columns=["imageUrl"], header=False, index=False)
cat_samples.to_csv("cat_urls.csv", columns=["imageUrl"], header=False, index=False)

dog_urls.csv

# Now we can download the images!
download_images("dog_urls.csv", "images/dog" , max_pics=500)
download_images("cat_urls.csv", "images/cat" , max_pics=500)

# Let's make sure all the images are readable
verify_images("images/dog", delete=True, max_size=500)
verify_images("images/cat", delete=True, max_size=500)

Create an ImageDataBunch

From here, everything should should seem pretty normal. We can create an ImageDataBunch from our images directory.

data = ImageDataBunch.from_folder("images", train=".", valid_pct=0.2, ds_tfms=get_transforms(), size=224, num_workers=4).normalize(imagenet_stats)

# Let's take a look at the data
data.show_batch(rows=3, figsize=(7,8))

show_batch output

Train a Model

We can now train a model! This is just a simple one, don't forget to fine tune!

learn = cnn_learner(data, models.resnet34, metrics=error_rate)
learn.fit_one_cycle(4)

fit_one_cycle output