vision.data

Computer vision data¶

This module contains the classes that define datasets handling Image objects and their transformations. As usual, we'll start with a quick overview, before we get in to the detailed API docs.

Before any work can be done a dataset needs to be converted into a DataBunch object, and in the case of the computer vision data - specifically into an ImageDataBunch subclass.

This is done with the help of data block API and the ImageList class and its subclasses.

However, there is also a group of shortcut methods provided by ImageDataBunch which reduce the multiple stages of the data block API, into a single wrapper method. These shortcuts methods work really well for:

• Imagenet-style of datasets (ImageDataBunch.from_folder)
• A pandas DataFrame with a column of filenames and a column of labels which can be strings for classification, strings separated by a label_delim for multi-classification or floats for a regression problem (ImageDataBunch.from_df)
• A csv file with the same format as above (ImageDataBunch.from_csv)
• A list of filenames and a list of targets (ImageDataBunch.from_lists)
• A list of filenames and a function to get the target from the filename (ImageDataBunch.from_name_func)
• A list of filenames and a regex pattern to get the target from the filename (ImageDataBunch.from_name_re)

In the last five factory methods, a random split is performed between train and validation, in the first one it can be a random split or a separation from a training and a validation folder.

If you're just starting out you may choose to experiment with these shortcut methods, as they are also used in the first lessons of the fastai deep learning course. However, you can completely skip them and start building your code using the data block API from the very beginning. Internally, these shortcuts use this API anyway.

The first part of this document is dedicated to the shortcut ImageDataBunch factory methods. Then all the other computer vision data-specific methods that are used with the data block API are presented.

Quickly get your data ready for training¶

To get you started as easily as possible, the fastai provides two helper functions to create a DataBunch object that you can directly use for training a classifier. To demonstrate them you'll first need to download and untar the file by executing the following cell. This will create a data folder containing an MNIST subset in data/mnist_sample.

There are a number of ways to create an ImageDataBunch. One common approach is to use Imagenet-style folders (see a ways down the page below for details) with ImageDataBunch.from_folder:

Here the datasets will be automatically created in the structure of Imagenet-style folders. The parameters specified:

• the transforms to apply to the images in ds_tfms (here with do_flip=False because we don't want to flip numbers),
• the target size of our pictures (here 24).

As with all DataBunch usage, a train_dl and a valid_dl are created that are of the type PyTorch DataLoader.

If you want to have a look at a few images inside a batch, you can use DataBunch.show_batch. The rows argument is the number of rows and columns to display.

The second way to define the data for a classifier requires a structure like this:

path\
  train\
  test\
  labels.csv

where the labels.csv file defines the label(s) of each image in the training set. This is the format you will need to use when each image can have multiple labels. It also works with single labels:

You can then use ImageDataBunch.from_csv:

An example of multiclassification can be downloaded with the following cell. It's a sample of the planet dataset.

If we open the labels files, we seach that each image has one or more tags, separated by a space.

The show_batchmethod will then print all the labels that correspond to each image.

You can find more ways to build an ImageDataBunch without the factory methods in data_block.

This is the same initialization as a regular DataBunch so you probably don't want to use this directly, but one of the factory methods instead.

Factory methods¶

If you quickly want to get a ImageDataBunch and train a model, you should process your data to have it in one of the formats the following functions handle.

Refer to create_from_ll to see all the **kwargs arguments.

"Imagenet-style" datasets look something like this (note that the test folder is optional):

path\
  train\
    class1\
    class2\
    ...
  valid\
    class1\
    class2\
    ...
  test\

For example:

Note that this (and all factory methods in this section) pass any kwargs to DataBunch.create.

Refer to create_from_ll to see all the **kwargs arguments.

Create an ImageDataBunch from path by splitting the data in folder and labelled in a file csv_labels between a training and validation set. Use valid_pct to indicate the percentage of the total images to use as the validation set. An optional test folder contains unlabelled data and suffix contains an optional suffix to add to the filenames in csv_labels (such as '.jpg'). fn_col is the index (or the name) of the the column containing the filenames and label_col is the index (indices) (or the name(s)) of the column(s) containing the labels. Use header to specify the format of the csv header, and delimiter to specify a non-standard csv-field separator. In case your csv has no header, column parameters can only be specified as indices. If label_delim is passed, split what's in the label column according to that separator.

For example:

Refer to create_from_ll to see all the **kwargs arguments.

Same as ImageDataBunch.from_csv, but passing in a DataFrame instead of a csv file. e.g

Different datasets are labeled in many different ways. The following methods can help extract the labels from the dataset in a wide variety of situations. The way they are built in fastai is constructive: there are methods which do a lot for you but apply in specific circumstances and there are methods which do less for you but give you more flexibility.

In this case the hierarchy is:

1. ImageDataBunch.from_name_re: Gets the labels from the filenames using a regular expression
2. ImageDataBunch.from_name_func: Gets the labels from the filenames using any function
3. ImageDataBunch.from_lists: Labels need to be provided as an input in a list

Refer to create_from_ll to see all the **kwargs arguments.

Creates an ImageDataBunch from fnames, calling a regular expression (containing one re group) on the file names to get the labels, putting aside valid_pct for the validation. In the same way as ImageDataBunch.from_csv, an optional test folder contains unlabelled data.

Our previously created dataframe contains the labels in the filenames so we can leverage it to test this new method. ImageDataBunch.from_name_re needs the exact path of each file so we will append the data path to each filename before creating our ImageDataBunch object.

Refer to create_from_ll to see all the **kwargs arguments.

Works in the same way as ImageDataBunch.from_name_re, but instead of a regular expression it expects a function that will determine how to extract the labels from the filenames. (Note that from_name_re uses this function in its implementation).

To test it we could build a function with our previous regex. Let's try another, similar approach to show that the labels can be obtained in a different way.

Refer to create_from_ll to see all the **kwargs arguments.

The most flexible factory function; pass in a list of labels that correspond to each of the filenames in fnames.

To show an example we have to build the labels list outside our ImageDataBunch object and give it as an argument when we call from_lists. Let's use our previously created function to create our labels list.

Use bs, num_workers, collate_fn and a potential test folder. ds_tfms is a tuple of two lists of transforms to be applied to the training and the validation (plus test optionally) set. tfms are the transforms to apply to the DataLoader. The size and the kwargs are passed to the transforms for data augmentation.

Other methods¶

In the next few methods we will use another dataset, CIFAR. This is because the second method will get the statistics for our dataset and we want to be able to show different statistics per channel. If we were to use MNIST, these statistics would be the same for every channel. White pixels are [255,255,255] and black pixels are [0,0,0] (or in normalized form [1,1,1] and [0,0,0]) so there is no variance between channels.

This function takes a tensor and flattens all dimensions except the channels, which it keeps as the first axis. This function is used to feed ImageDataBunch.batch_stats so that it can get the pixel statistics of a whole batch.

Let's take as an example the dimensions our MNIST batches: 128, 3, 24, 24.

In the fast.ai library we have imagenet_stats, cifar_stats and mnist_stats so we can add normalization easily with any of these datasets. Let's see an example with our dataset of choice: MNIST.

Data normalization¶

You may also want to normalize your data, which can be done by using the following functions.

On MNIST the mean and std are 0.1307 and 0.3081 respectively (looked on Google). If you're using a pretrained model, you'll need to use the normalization that was used to train the model. The imagenet norm and denorm functions are stored as constants inside the library named imagenet_norm and imagenet_denorm. If you're training a model on CIFAR-10, you can also use cifar_norm and cifar_denorm.

You may sometimes see warnings about clipping input data when plotting normalized data. That's because even although it's denormalized when plotting automatically, sometimes floating point errors may make some values slightly out or the correct range. You can safely ignore these warnings in this case.

To use this dataset and collate samples into batches, you'll need to following function:

Finally, to apply transformations to Image in a Dataset, we use this last class.

ItemList specific to vision¶

The vision application adds a few subclasses of ItemList specific to images.

It inherits from ItemList and overwrite ItemList.get to call open_image in order to turn an image file in Path object into an Image object. label_cls can be specified for the labels, xtra contains any extra information (usually in the form of a dataframe) and processor is applied to the ItemList after splitting and labelling.

How ImageList.__init__ overwrites on ItemList.__init__?

ImageList.__init__ creates additional attributes like convert_mode, after_open, c, size upon ItemList.__init__; and convert_mode and sizes in particular are necessary to make use of ImageList.get (which also overwrites on ItemList.get) and ImageList.open.

How ImageList.from_folder overwrites on ItemList.from_folder?

ImageList.from_folder adds some constraints on extensions upon ItemList.from_folder, to work with image files specifically; and can take additional input arguments like convert_mode and after_open which are not available to ItemList.

Let's get a feel of how open is used with the following example.

open takes only one input fn as filename in the type of Path or String.

The reason why imagelistRGB[10] print out an image, is because behind the scene we have ImageList.get calls ImageList.open which calls open_image which uses PIL.Image.open(fn).convert(convert_mode) to open an image file (how we print the image), and finally turns it into an Image object with shape (3, 128, 128)

Internally, ImageList.open passes ImageList.convert_mode and ImageList.after_open to open_image to adjust the appearance of the Image object. For example, setting convert_mode to L can make images black and white.

Building your own dataset¶

This module also contains a few helper functions to allow you to build you own dataset for image classification.

It will try if every image in this folder can be opened and has n_channels. If n_channels is 3 – it'll try to convert image to RGB. If delete=True, it'll be removed it this fails. If resume – it will skip already existent images in dest. If max_size is specified, image is resized to the same ratio so that both sizes are less than max_size, using interp. Result is stored in dest, ext forces an extension type, img_format and kwargs are passed to PIL.Image.save. Use max_workers CPUs.