Datasets in salad

This is a test notebook that will later hold demos for a subpackage of salad

Introduction

Lorem ipsum dolor sit amet, consetetur sadipscing elitr, sed diam nonumy eirmod tempor invidunt ut labore et dolore magna aliquyam erat, sed diam voluptua. At vero eos et accusam et justo duo dolores et ea rebum. Stet clita kasd gubergren, no sea takimata sanctus est Lorem ipsum dolor sit amet. Lorem ipsum dolor sit amet, consetetur sadipscing elitr, sed diam nonumy eirmod tempor invidunt ut labore et dolore magna aliquyam erat, sed diam voluptua. At vero eos et accusam et justo duo dolores et ea rebum. Stet clita kasd gubergren, no sea takimata sanctus est Lorem ipsum dolor sit amet.

import torch
from torch import nn

import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns

%matplotlib inline
#%config InlineBackend.figure_format = 'svg'

Digits Datasets

A standard benchmark for domain adaptation is digit classification. In salad.datasets, it is easy to get access to this standard benchmark.

from salad.datasets import DigitsLoader
from salad.utils import panelize

We currently implemented unified loading functions for four different benchmarks.

from salad.datasets import MNIST, USPS, SVHN, Synth

mnist = MNIST('/tmp/data')
usps  = USPS('/tmp/data')
svhn  = SVHN('/tmp/data')
synth = Synth('/tmp/data')

Extracting /tmp/data/zip.train.gz
Extracting /tmp/data/synth_train_32x32.mat?raw=true

Accessing them is also possible directy by DigitsLoader, which subclasses the usual torch.utils.data.DataLoader:

dataset_names = ['mnist', 'usps', 'synth', 'svhn']
data = DigitsLoader('/tmp/data', dataset_names, shuffle=True, batch_size = 64, normalize=False)

Extracting /tmp/data/zip.train.gz
Extracting /tmp/data/synth_train_32x32.mat?raw=true
Using downloaded and verified file: /tmp/data/train_32x32.mat

for batch in data:

    for (x,y), name in zip(batch, dataset_names):

        print(name, x.size(), y.dtype, np.unique(y.numpy()))

    break

mnist torch.Size([64, 3, 32, 32]) torch.int64 [0 1 2 3 4 5 6 7 8 9]
usps torch.Size([64, 3, 32, 32]) torch.int64 [0 1 2 3 4 5 6 7 8 9]
synth torch.Size([64, 3, 32, 32]) torch.int64 [0 1 2 3 4 5 6 7 8 9]
svhn torch.Size([64, 3, 32, 32]) torch.int64 [0 1 2 3 4 5 6 7 8 9]

fig, axes = plt.subplots(2,2,figsize=(10,10))
axes = axes.flatten()

for batch in data:

    for (x,y), ax, name in zip(batch, axes, dataset_names):

        ax.imshow(panelize(x.numpy()))
        ax.set_title(name.upper())
        ax.axis('off')

    break

plt.show()

For training adaptation models, it is generally a good idea to normalize the images in some way prior to feeding them into a neural network. salad provide some means of doing this in a standardized way. In particular, it is easy to check statistics of the dataset:

default_normalization = {}

data_normalized = DigitsLoader('/tmp/data', dataset_names, shuffle=True, batch_size = 200, normalize = True)

for dataset in data_normalized.datasets:

    samples = np.concatenate([x.numpy() for x,_ in dataset])
    print(samples.mean(axis=(0,2,3)), samples.std(axis=(0,2,3)))

Normalize data
Normalize data
Extracting /tmp/data/zip.train.gz
Normalize data
Downloading https://github.com/domainadaptation/datasets/blob/master/synth/synth_train_32x32_small.mat?raw=true to /tmp/data/synth_train_32x32_small.mat?raw=true
Downloading https://github.com/domainadaptation/datasets/blob/master/synth/synth_test_32x32_small.mat?raw=true to /tmp/data/synth_test_32x32_small.mat?raw=true
Done!
Extracting /tmp/data/synth_train_32x32_small.mat?raw=true
Normalize data
Using downloaded and verified file: /tmp/data/train_32x32.mat
[9.4225504e-07 9.4225504e-07 9.4225504e-07] [1.0009977 1.0009977 1.0009977]
[-3.891881e-05 -3.891881e-05 -3.891881e-05] [1.0014582 1.0014582 1.0014582]
[1.4131116e-05 7.6292954e-06 4.8976064e-05] [1.0168213 1.0188137 1.0175495]
[-9.3011549e-05 -1.5506116e-04  2.3364362e-05] [1.0297976 1.0292693 1.0348536]

The normalization values can be found in salad.datasets.transforms.

Toy Datasets

The salad.datasets package also provides a loader for toy datasets.

from salad.datasets import ToyDatasetLoader

loader_stacked = ToyDatasetLoader(augment = False, collate='stack', batch_size = 2048, seed=1306)
loader_concat  = ToyDatasetLoader(augment = False, collate='cat', batch_size = 2048, seed=1306)

A stacked loader returns a tuples containing the return values of the individual datasets:

for (xs,ys), (xt,yt) in loader_stacked:

    plt.figure(figsize=(12,3))
    plt.subplot(1,3,1)
    plt.scatter(*xs.transpose(1,0), c = ys, s= 1)
    plt.axis('off')
    plt.subplot(1,3,2)
    plt.scatter(*xt.transpose(1,0), c = yt, s = 1)
    plt.axis('off')
    plt.subplot(1,3,3)
    plt.scatter(*xs.transpose(1,0), s=1, alpha = .75)
    plt.scatter(*xt.transpose(1,0), s=1, alpha = .75)
    plt.axis('off')

    break

plt.show()