%reload_ext autoreload
%autoreload 2
%matplotlib inline

Convolutions and basics of CNNs¶

How convolutions work

from exp.nb_07 import *
from PIL import Image
import numpy as np

Data¶

Let's grab out MNIST dataset.

x_train, y_train, x_valid, y_valid = get_data()

x_train.shape

torch.Size([50000, 784])

Each row of the dataset is an image,

x_train[0].shape

torch.Size([784])

But in order to properly view them we'll reshape them for now into rank 3 tensors:

[channels, height, width]

x_train = x_train.view(-1, 28,28)

five = x_train[0]

plt.imshow(five)

<matplotlib.image.AxesImage at 0x2990cc70d30>

Let's try to create a top edge detection kernel from scratch and convolve it over the image.

k = tensor([
    [1.,1.,1.],
    [-1.,-1.,-1.,],
    [0.,0.,0.]
]); k.shape

torch.Size([3, 3])

Pytorch F.conv2d requires [batch size, channels, height, width] so we'll reshape using view

Then we can pass it into our nn.Conv2d with our kernel and since the stride=1 the result will be a rank 3 tensor

`[filters, channels, height, width]'

top = F.conv2d(five.view(1,1,28,28), k[None, None])

top.shape

torch.Size([1, 1, 26, 26])

plt.imshow(top.squeeze())

<matplotlib.image.AxesImage at 0x2990d41e6a0>

And if we transpose our kernel it looks like it will detect edges on the leftside of an object.

k.t()

tensor([[ 1., -1.,  0.],
        [ 1., -1.,  0.],
        [ 1., -1.,  0.]])

left = F.conv2d(five.view(1,1,28,28), k.t()[None, None])

plt.imshow(left.squeeze())

<matplotlib.image.AxesImage at 0x2990e0458e0>

So how do this basic

F.adaptive_avg_pool2d

Applies a 2D adaptive average pooling over an input signal composed of several input planes.

avg_pool_1d = F.adaptive_avg_pool2d(feature_map, 1); avg_pool_1d.shape

feature_map.squeeze().view(-1).mean()

avg_pool_1d.squeeze()

F.adaptive_max_pool2d

Applies a 2D adaptive max pooling over an input signal composed of several input planes.

Max pooling will return a tensor of the max of some specified shape.

max_pool = F.adaptive_max_pool2d(feature_map, 1); max_pool.shape

feature_map.squeeze().view(-1).max()

max_pool.squeeze()

Let's look at the activations of a pretrained model.

from fastai.vision import *

model = models.resnet34(pretrained=True)

To pass an image we need to: normalize, turn into a mini-batch, and put onto GPU

Resources