5. Transformation

[1]:

%matplotlib inline
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt
from torchvision.transforms import *

np.random.seed(37)

image = Image.open('./shapes/train/circle/0001.jpg', 'r')
print('width ', image.width, 'height ', image.height)

plt.imshow(image)

width  144 height  144

[1]:

<matplotlib.image.AxesImage at 0x7f2e5bd240f0>

5.1. PIL transforms

5.1.1. Crop

[2]:

t1 = CenterCrop(100)
t2 = CenterCrop(144)
t3 = CenterCrop(175)

fig, axes = plt.subplots(1, 3, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))
ax[2].imshow(t3(image))

ax[0].set_title('Zoom In')
ax[1].set_title('Normal')
ax[2].set_title('Zoom Out')

[2]:

Text(0.5, 1.0, 'Zoom Out')

[3]:

t1 = RandomCrop(size=(144, 144), padding=1)
t2 = RandomCrop(size=(144, 144), padding=5)

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[3]:

<matplotlib.image.AxesImage at 0x7f2e5bb15a90>

5.1.2. ColorJitter

[4]:

t1 = ColorJitter(brightness=10, contrast=0, saturation=0, hue=0.5)
t2 = ColorJitter(brightness=10, contrast=0, saturation=0.1, hue=0.5)
t3 = ColorJitter(brightness=10, contrast=0.5, saturation=0.1, hue=0.5)

fig, axes = plt.subplots(1, 3, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))
ax[2].imshow(t3(image))

[4]:

<matplotlib.image.AxesImage at 0x7f2e5ba1bef0>

5.1.3. Grayscale

[5]:

t1 = Grayscale(num_output_channels=1)
t2 = Grayscale(num_output_channels=3)

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[5]:

<matplotlib.image.AxesImage at 0x7f2e5b965d30>

[6]:

t1 = RandomGrayscale(p=0.1)
t2 = RandomGrayscale(p=0.6)

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[6]:

<matplotlib.image.AxesImage at 0x7f2e5b990cc0>

5.1.4. Pad

[7]:

t1 = Pad(5)
t2 = Pad(10)

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[7]:

<matplotlib.image.AxesImage at 0x7f2e5b8d3588>

5.1.5. RandomAffine

[8]:

t1 = RandomAffine(degrees=(0, 30))
t2 = RandomAffine(degrees=(30, 50), shear=3)

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[8]:

<matplotlib.image.AxesImage at 0x7f2e5b919710>

5.1.6. RandomApply

[9]:

t1 = RandomApply([CenterCrop(175),
                  RandomAffine(degrees=(0, 30)),
                  Pad(3)], p=0.5)
t1 = RandomApply([CenterCrop(175),
                  RandomAffine(degrees=(0, 30)),
                  Pad(3)], p=0.5)

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[9]:

<matplotlib.image.AxesImage at 0x7f2e5b653f28>

5.1.7. RandomChoice

[10]:

t1 = RandomChoice([CenterCrop(175),
                  RandomAffine(degrees=(0, 30)),
                  Pad(3)])
t1 = RandomChoice([CenterCrop(175),
                  RandomAffine(degrees=(0, 30)),
                  Pad(3)])

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[10]:

<matplotlib.image.AxesImage at 0x7f2e5b5ac9e8>

5.1.8. RandomFlip

[11]:

t1 = RandomHorizontalFlip(p=0.99)
t2 = RandomHorizontalFlip(p=0.99)

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[11]:

<matplotlib.image.AxesImage at 0x7f2e5b4b2d68>

5.1.9. RandomOrder

[12]:

t1 = RandomOrder([CenterCrop(175),
                  RandomAffine(degrees=(0, 30)),
                  Pad(3)])
t1 = RandomOrder([CenterCrop(175),
                  RandomAffine(degrees=(0, 30)),
                  Pad(3)])

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[12]:

<matplotlib.image.AxesImage at 0x7f2e5b380da0>

5.1.10. RandomPerspective

[13]:

t1 = RandomPerspective(distortion_scale=0.8, p=0.99)
t2 = RandomPerspective(distortion_scale=0.8, p=0.99)

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[13]:

<matplotlib.image.AxesImage at 0x7f2e5b4c6390>

5.1.11. RandomRotation

[14]:

t1 = RandomRotation(degrees=(45, 90))
t2 = RandomRotation(degrees=(45, 90))

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[14]:

<matplotlib.image.AxesImage at 0x7f2e5b415ac8>

5.1.12. Resize

[15]:

t1 = Resize(size=(200, 200))
t2 = Resize(size=(130, 130))

fig, axes = plt.subplots(1, 2, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))

[15]:

<matplotlib.image.AxesImage at 0x7f2e5b16eb38>

5.2. Tensor transforms

[16]:

import torch

data = ToTensor()(image)
c, h, w = data.shape[0], data.shape[1], data.shape[2]

5.2.1. Normalize

[17]:

t = Normalize(mean=[0.5, 0.5, 0.5], std=[1., 1., 1.])
o = t(data)
o.shape

[17]:

torch.Size([3, 144, 144])

5.2.2. RandomErasing

[18]:

t = RandomErasing(p=0.5)
o = t(data)
o.shape

[18]:

torch.Size([3, 144, 144])

5.3. Conversion transforms

5.3.1. PIL image

[19]:

t1 = transforms.Compose([
    ToTensor(),
    ToPILImage()
])
t2 = transforms.Compose([
    ToTensor(),
    Normalize(mean=[0.5, 0.5, 0.5], std=[1., 1., 1.]),
    ToPILImage()
])
t3 = transforms.Compose([
    ToTensor(),
    RandomErasing(p=0.9),
    ToPILImage()
])

fig, axes = plt.subplots(1, 3, figsize=(10, 5))
ax = axes.ravel()

ax[0].imshow(t1(image))
ax[1].imshow(t2(image))
ax[2].imshow(t3(image))

ax[0].set_title('Normal')
ax[1].set_title('Normalize')
ax[2].set_title('Erasing')

[19]:

Text(0.5, 1.0, 'Erasing')