This dataset was collected by Google from people drawing different objects. The dataset is a collection of 50 million drawings from 345 different objects that is available publically for everyone. We extract a simple subset of the dataset for simple machine learning tasks. The dataset is suggested as an alternative for MNIST.
MNIST has many disadvantages
- No great variablitiy in the data for each class. The number of strokes is limited for each drawn digit.
- Most MNIST pairs could be distinguised by just one pixel. See this.
- The data is overused in both universities and the literature.
| Name | Content | Examples | Size | Link |
|---|---|---|---|---|
train-ubyte.npz |
training set images and labels | 80,000 | 21 MBytes | Download |
test-ubyte.npz |
testing set images and labels | 20,000 | 6 MBytes | Download |
Alternatively, you can clone this GitHub repository; the dataset appears under dataset/.
Each training and test example is assigned to one of the following labels:
| Label | Description |
|---|---|
| 0 | Cloud |
| 1 | Sun |
| 2 | Pants |
| 3 | Umbrella |
| 4 | Table |
| 5 | Ladder |
| 6 | Eyeglasses |
| 7 | Clock |
| 8 | Scissors |
| 9 | Cup |
Loading data with Python (requires NumPy)
import numpy as np
train_data = np.load('dataset/train-ubyte.npz')
test_data = np.load('dataset/test-ubyte.npz')
x_train, y_train = train_data['a'], train_data['b']
x_test, y_test = test_data['a'], test_data['b']