Image Classification with Convolutional Neural Networks¶

This tutorial explains how image classification with CNNs can be handled in safeds with the NeuralNetworkClassifier. The data used contains of multiple drawn shapes and can be found on kaggle.

Note

All operations on an NeuralNetworkClassifier return a new NeuralNetworkClassifier. The original NeuralNetworkClassifier will not be changed.

Load data into an ImageDataset¶

1. Load images via files in an ImageList. The data is available under docs/tutorials/data/shapes. If the return_filenames parameter is set to True, a list of all filepaths will be returned as well in the same order as the images in the returned ImageList.

In [1]:

from safeds.data.image.containers import ImageList

images, filepaths = ImageList.from_files("data/shapes", return_filenames=True)

from safeds.data.image.containers import ImageList images, filepaths = ImageList.from_files("data/shapes", return_filenames=True)

2. Create a Column with the labels of the images:

In [2]:

import re

from safeds.data.tabular.containers import Column

labels = Column(
    "label",
    [re.search(r"(.*)[\\/](.*)[\\/](.*)\.", filepath).group(2) for filepath in filepaths],
)

import re from safeds.data.tabular.containers import Column labels = Column( "label", [re.search(r"(.*)[\\/](.*)[\\/](.*)\.", filepath).group(2) for filepath in filepaths], )

3. Create an ImageDataset from the ImageList and the Column of labels. If the shuffle parameter is set to True, the ImageDataset will be shuffled after each epoch while training a neural network.

In [3]:

from safeds.data.labeled.containers import ImageDataset

dataset = ImageDataset[Column](images, labels, shuffle=True)

from safeds.data.labeled.containers import ImageDataset dataset = ImageDataset[Column](images, labels, shuffle=True)

Create the neural network with a NeuralNetworkClassifier¶

1. Create a list of Layer instances for your neural network:

In [4]:

from safeds.ml.nn.layers import Convolutional2DLayer, FlattenLayer, ForwardLayer, MaxPooling2DLayer

layers = [
    Convolutional2DLayer(output_channel=16, kernel_size=3, padding=1),
    MaxPooling2DLayer(kernel_size=2, stride=2),
    Convolutional2DLayer(output_channel=32, kernel_size=3, padding=1),
    MaxPooling2DLayer(kernel_size=2, stride=2),
    FlattenLayer(),
    ForwardLayer(neuron_count=128),
    ForwardLayer(neuron_count=3),
]

from safeds.ml.nn.layers import Convolutional2DLayer, FlattenLayer, ForwardLayer, MaxPooling2DLayer layers = [ Convolutional2DLayer(output_channel=16, kernel_size=3, padding=1), MaxPooling2DLayer(kernel_size=2, stride=2), Convolutional2DLayer(output_channel=32, kernel_size=3, padding=1), MaxPooling2DLayer(kernel_size=2, stride=2), FlattenLayer(), ForwardLayer(neuron_count=128), ForwardLayer(neuron_count=3), ]

2. Create a NeuralNetworkClassifier from an InputConversion, the list of Layer instances:

In [5]:

from safeds.ml.nn import NeuralNetworkClassifier
from safeds.ml.nn.converters import InputConversionImageToColumn

cnn = NeuralNetworkClassifier[ImageDataset[Column], ImageList](
    InputConversionImageToColumn(dataset.input_size),
    layers,
)

from safeds.ml.nn import NeuralNetworkClassifier from safeds.ml.nn.converters import InputConversionImageToColumn cnn = NeuralNetworkClassifier[ImageDataset[Column], ImageList]( InputConversionImageToColumn(dataset.input_size), layers, )

Fit and predict the NeuralNetworkClassifier¶

1. Fit the NeuralNetworkClassifier:

In [6]:

cnn_fitted = cnn.fit(dataset, epoch_count=8, batch_size=16)

cnn_fitted = cnn.fit(dataset, epoch_count=8, batch_size=16)

2. Predict values from the NeuralNetworkClassifier:

In [7]:

prediction = cnn_fitted.predict(dataset.get_input())

prediction = cnn_fitted.predict(dataset.get_input())

3. Shuffle the prediction to get a random order:

In [8]:

shuffled_prediction = prediction.shuffle()

shuffled_prediction = prediction.shuffle()

4. Display a subset of the input data:

In [9]:

shuffled_prediction.get_input().remove_image_by_index(list(range(9, len(prediction))))

shuffled_prediction.get_input().remove_image_by_index(list(range(9, len(prediction))))

Out[9]:

5. Display the corresponding predicted labels:

In [10]:

shuffled_prediction.get_output().to_list()[0:9]

shuffled_prediction.get_output().to_list()[0:9]

Out[10]:

['triangles',
 'triangles',
 'triangles',
 'triangles',
 'triangles',
 'triangles',
 'triangles',
 'triangles',
 'triangles']