vision.transform

Image transforms¶

fastai provides a complete image transformation library written from scratch in PyTorch. Although the main purpose of the library is data augmentation for use when training computer vision models, you can also use it for more general image transformation purposes. Before we get in to the detail of the full API, we'll look at a quick overview of the data augmentation pieces that you'll almost certainly need to use.

Data augmentation¶

Data augmentation is perhaps the most important regularization technique when training a model for Computer Vision: instead of feeding the model with the same pictures every time, we do small random transformations (a bit of rotation, zoom, translation, etc...) that don't change what's inside the image (to the human eye) but do change its pixel values. Models trained with data augmentation will then generalize better.

To get a set of transforms with default values that work pretty well in a wide range of tasks, it's often easiest to use get_transforms. Depending on the nature of the images in your data, you may want to adjust a few arguments, the most important being:

• do_flip: if True the image is randomly flipped (default behavior)
• flip_vert: limit the flips to horizontal flips (when False) or to horizontal and vertical flips as well as 90-degrees rotations (when True)

get_transforms returns a tuple of two lists of transforms: one for the training set and one for the validation set (we don't want to modify the pictures in the validation set, so the second list of transforms is limited to resizing the pictures). This can be passed directly to define a DataBunch object (see below) which is then associated with a model to begin training.

Note that the defaults for get_transforms are generally pretty good for regular photos - although here we'll add a bit of extra rotation so it's easier to see the differences.

We first define a function to return a new image, since transformation functions modify their inputs. We also define a little helper function plots_f to let us output a grid of transformed images based on a function - the details of this function aren't important here.

If we want to have a look at what these transforms actually do, we need to use the apply_tfms function. It will be in charge of picking the values of the random parameters and doing the transformations to the Image object. This function has multiple arguments you can customize (see its documentation for details); here we will highlight the most useful. The first one we'll need to set, especially if our images are of different shapes, is the target size. It will ensure that all the images are cropped or padded to the same size so we can then collate them into batches.

Note that the target size can be a rectangle if you specify a tuple of int.

The second argument that can be customized is how we treat missing pixels: when applying transforms (like a rotation), some of the pixels inside the square won't have values from the image. We can set missing pixels to one of the following:

• black (padding_mode='zeros')
• the value of the pixel at the nearest border (padding_mode='border')
• the value of the pixel symmetric to the nearest border (padding_mode='reflection')

padding_mode='reflection' is the default. Here is what padding_mode='zeros' looks like:

And padding_mode='border' looks like this:

The third argument that might be useful to change is resize_method. Images are often rectangles of different ratios, so to get them to the target size, we may need to crop, squish, or pad them to get the ratio right.

By default, the library resizes the image while keeping its original ratio so that the smaller size corresponds to the given size, then takes a crop (ResizeMethod.CROP). You can choose to resize the image while keeping its original ratio so that the bigger size corresponds to the given size, then take a pad (ResizeMethod.PAD). Another way is to just squish the image to the given size (ResizeMethod.SQUISH).

Data augmentation details¶

If you want to quickly get a set of random transforms that have worked well in a wide range of tasks, you should use the get_transforms function. The most important parameters to adjust are do_flip and flip_vert, depending on the type of images you have.

This function returns a tuple of two lists of transforms, one for the training set and the other for the validation set (which is limited to a center crop by default).

Let's see how get_transforms changes this little kitten now.

Another useful function that gives basic transforms is zoom_crop:

scale should be a given float if do_rand is False, otherwise it can be a range of floats (and the zoom will have a random value in between). Again, here is a sense of what this can give us:

This transform is an implementation of the main approach used for nearly all winning Imagenet entries since 2013, based on Andrew Howard's Some Improvements on Deep Convolutional Neural Network Based Image Classification. It determines a new width and height of the image after the random scale and squish to the new ratio are applied. Those are switched with probability 0.5. Then we return the part of the image with the width and height computed, centered in row_pct, col_pct if width and height are both less than the corresponding size of the image. Otherwise we try again with new random parameters.

Randomness¶

The functions that define each transform, such as rotateor flip_lr are deterministic. The fastai library will then randomize them in two different ways:

• each transform can be defined with an argument named p representing the probability for it to be applied
• each argument that is type-annotated with a random function (like uniform or rand_bool) can be replaced by a tuple of arguments accepted by this function, and on each call of the transform, the argument that is passed inside the function will be picked randomly using that random function.

If we look at the function rotate for instance, we see it has an argument degrees that is type-annotated as uniform.

First level of randomness: We can define a transform using rotate with degrees fixed to a value, but by passing an argument p. The rotation will then be executed with a probability of p but always with the same value of degrees.

Second level of randomness: We can define a transform using rotate with degrees defined as a range, without an argument p. The rotation will then always be executed with a random value picked uniformly between the two floats we put in degrees.

All combined: We can define a transform using rotate with degrees defined as a range, and an argument p. The rotation will then be executed with a probability p and with a random value picked uniformly between the two floats we put in degrees.

List of transforms¶

Here is the list of all the deterministic functions on which the transforms are built. As explained before, each of them can have a probability p of being executed, and any time an argument is type-annotated with a random function, it's possible to randomize it via that function.

This transform adjusts the brightness of the image depending on the value of change. A change of 0 will transform the image to black, and a change of 1 will transform the image to white. change=0.5 doesn't adjust the brightness.

scale adjusts the contrast. A scale of 0 will transform the image to grey, and a scale over 1 will transform the picture to super-contrast. scale = 1. doesn't adjust the contrast.

This transform takes a crop of the image to return one of the given size. The position is given by (col_pct, row_pct), with col_pct and row_pct being normalized between 0. and 1.

This works like crop but if the target size is bigger than the size of the image (on either dimension), padding is applied according to padding_mode (see pad for an example of all the options) and the position of center is ignored on that dimension.

This transform applies combines a flip (horizontal or vertical) and a rotation of a multiple of 90 degrees.

This is an affine implementation of dihedral that should be used if the target is an ImagePoints or an ImageBBox.

This transform horizontally flips the image. flip_lr mirrors the image.

This is an affine implementation of flip_lr that should be used if the target is an ImagePoints or an ImageBBox.

This transform changes the pixels of the image by randomly replacing them with pixels from the neighborhood (how far the neighborhood extends is controlled by the value of magnitude).

Pad the image by adding padding pixel on each side of the picture according to mode:

• mode='zeros': pads with zeros,
• mode='border': repeats the pixels at the border.
• mode='reflection': pads by taking the pixels symmetric to the border.

Perspective warping is a deformation of the image as seen in a different plane of the 3D-plane. The new plane is determined by telling where we want each of the four corners of the image (from -1 to 1, -1 being left/top, 1 being right/bottom).

resize¶

Pytorch's transforms.Resize(size) equivalent is implemented without an explicit transform function in fastai. It's done via the arguments size and resize_method.

The size argument can be either a single int 224, or a tuple of ints (224,400). The default behavior is to crop the image to a square when a single int is passed and to squish it in the case of a tuple, so that:

• if size=224 is passed, it will resize and then crop to (224,224)
• if size=(224,400) is passed, it will squish it to (224,400)
• if size=(224,224) is passed, it will squish (not crop!) it to (224,224)

You can override the default resize_method.

Note:

If you receive an error similar to the one below:

RuntimeError: Argument #4: Padding size should be less than the corresponding input dimension, but got: padding (46, 46) at dimension 3 of input [1, 3, 128, 36]

this is caused by an issue with PyTorch's reflection padding, which the library uses by default. Adding an extra keyword argument padding_mode = 'zeros' should be able to serve as a workaround for now.

The resize is performed slightly differently depending on how ImageDataBunch is created:

1. When the shortcut ImageDataBunch from_* methods are used, the size and resize_method arguments are passed with the rest of the arguments. For example, to resize images on the fly to 224x224 with from_name_re method, do:

   data = ImageDataBunch.from_name_re(path_img, fnames, pat, size=224, bs=bs)

   and to override the resize_method:

   data = ImageDataBunch.from_name_re(path_img, fnames, pat, size=224, resize_method=ResizeMethod.SQUISH, bs=bs)

2. When data block API is used, the size and resize_method are passed via the transform method. For example:

   src = ImageList.from_folder(path).split_none().label_from_folder()
   tfms = get_transforms() # or tfms=None if none are needed
   size=224 # size=(224,224) or (400,224)
   data = src.transform(tfms=tfms, size=size, resize_method=ResizeMethod.SQUISH).databunch(bs=bs, num_workers=4).normalize()

Resizing before training

Do note that if you just want to resize the input images, doing it on the fly via transform is inefficient, since it will have to be done on every notebook re-run. Chances are that you will want to resize the images on the filesystem and use the resized datasets as needed.

For example, you could use fastai code to do that, to create low-resolution images under 'small-96', and mid-resolution images under small-256:

from fastai.vision import *

path = untar_data(URLs.PETS)
path_hr = path/'images'
path_lr = path/'small-96'
path_mr = path/'small-256'

il = ImageList.from_folder(path_hr)

def resize_one(fn, i, path, size):
    dest = path/fn.relative_to(path_hr)
    dest.parent.mkdir(parents=True, exist_ok=True)
    img = PIL.Image.open(fn)
    targ_sz = resize_to(img, size, use_min=True)
    img = img.resize(targ_sz, resample=PIL.Image.BILINEAR).convert('RGB')
    img.save(dest, quality=75)

# create smaller image sets the first time this nb is run
sets = [(path_lr, 96), (path_mr, 256)]
for p,size in sets:
    if not p.exists(): 
        print(f"resizing to {size} into {p}")
        parallel(partial(resize_one, path=p, size=size), il.items)

Of course, adjust quality, resample, and other arguments to suit your needs. You will also need to tweak it for custom directories (train, test, etc.)

imagemagick's mogrify and convert are commonly-used tools to resize images via your shell. For example, if you're in a data directory, containing a test directory:

ls -1 *
test

and you want to create a new directory 300x224/test with images resized to 300x224:

SRC=train; DEST=300x224; mkdir -p $DEST/$SRC; find $SRC -name "*.jpg" -exec convert -resize 300x224 -gravity center -extent 300x224 {} $DEST/{} \;

Check the imagemagick documentation for the many various options.

If you already have a directory which is a copy of original images, mogrify is usually applied directly to the files with the same result:

mkdir 300x224
cp -r train 300x224
cd 300x224/train
mogrify -resize 300x224 -gravity center -extent 300x224 *jpg

• channel: Which channel (RGB) to randomise
• thresh: After randomising, scale the values to not exceed the thresh value

By randomizing one of the three channels, the learner essentially sees the same image, but with different colors. Usually, every RGB image has one channel that is dominant, and randomizing this channel is the riskiest; thus, a low thresh (threshold) value must be applied. In this example, the Green channel is the dominant one.

Apply the four tilts at the same time, each with a strength given in the vector magnitude. See tilt just below for the effect of each individual tilt.

direction is a number (0: left, 1: right, 2: top, 3: bottom). A positive magnitude is a tilt forward (toward the person looking at the picture), a negative magnitude a tilt backward.

The normalization technique described in this paper: Improved Regularization of Convolutional Neural Networks with Cutout

By default, it will apply a single cutout (n_holes=1) of length=40) with probability p=1. The cutout position is always random. If you choose to do multiple cutouts, they may overlap.

The paper above used cutouts of size 16x16 for CIFAR-10 (10 categiries classification) and cutouts of size 8x8 for CIFAR-100 (100 categories). Generally, the more categories, the less cutout you want.

You can add some randomness to the cutouts like this:

Convenience functions¶

These functions simplify creating random versions of crop_pad and zoom.

The args are for internal purposes and shouldn't be touched.