Here we visualize filters and outputs using the network architecture proposed by Krizhevsky et al. for ImageNet and implemented in caffe.

(This page follows DeCAF visualizations originally by Yangqing Jia.)

First, import required modules and set plotting parameters

In [1]:
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline

# Make sure that caffe is on the python path:
caffe_root = '../'  # this file is expected to be in {caffe_root}/examples
import sys
sys.path.insert(0, caffe_root + 'python')

import caffe

plt.rcParams['figure.figsize'] = (10, 10)
plt.rcParams['image.interpolation'] = 'nearest'
plt.rcParams['image.cmap'] = 'gray'

Follow the instructions for getting the pretrained models, load the net, specify test phase and CPU mode, and configure input preprocessing.

In [2]:
net = caffe.Classifier(caffe_root + 'examples/imagenet/imagenet_deploy.prototxt',
                       caffe_root + 'examples/imagenet/caffe_reference_imagenet_model')
net.set_phase_test()
net.set_mode_cpu()
# input preprocessing: 'data' is the name of the input blob == net.inputs[0]
net.set_mean('data', np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy'))  # ImageNet mean
net.set_raw_scale('data', 255)  # the reference model operates on images in [0,255] range instead of [0,1]
net.set_channel_swap('data', (2,1,0))  # the reference model has channels in BGR order instead of RGB

Run a classification pass

In [3]:
scores = net.predict([caffe.io.load_image(caffe_root + 'examples/images/cat.jpg')])

The layer features and their shapes (10 is the batch size, corresponding to the the ten subcrops used by Krizhevsky et al.)

In [4]:
[(k, v.data.shape) for k, v in net.blobs.items()]
Out[4]:
[('data', (10, 3, 227, 227)),
 ('conv1', (10, 96, 55, 55)),
 ('pool1', (10, 96, 27, 27)),
 ('norm1', (10, 96, 27, 27)),
 ('conv2', (10, 256, 27, 27)),
 ('pool2', (10, 256, 13, 13)),
 ('norm2', (10, 256, 13, 13)),
 ('conv3', (10, 384, 13, 13)),
 ('conv4', (10, 384, 13, 13)),
 ('conv5', (10, 256, 13, 13)),
 ('pool5', (10, 256, 6, 6)),
 ('fc6', (10, 4096, 1, 1)),
 ('fc7', (10, 4096, 1, 1)),
 ('fc8', (10, 1000, 1, 1)),
 ('prob', (10, 1000, 1, 1))]

The parameters and their shapes (each of these layers also has biases which are omitted here)

In [5]:
[(k, v[0].data.shape) for k, v in net.params.items()]
Out[5]:
[('conv1', (96, 3, 11, 11)),
 ('conv2', (256, 48, 5, 5)),
 ('conv3', (384, 256, 3, 3)),
 ('conv4', (384, 192, 3, 3)),
 ('conv5', (256, 192, 3, 3)),
 ('fc6', (1, 1, 4096, 9216)),
 ('fc7', (1, 1, 4096, 4096)),
 ('fc8', (1, 1, 1000, 4096))]

Helper functions for visualization

In [6]:
# take an array of shape (n, height, width) or (n, height, width, channels)
#  and visualize each (height, width) thing in a grid of size approx. sqrt(n) by sqrt(n)
def vis_square(data, padsize=1, padval=0):
    data -= data.min()
    data /= data.max()
    
    # force the number of filters to be square
    n = int(np.ceil(np.sqrt(data.shape[0])))
    padding = ((0, n ** 2 - data.shape[0]), (0, padsize), (0, padsize)) + ((0, 0),) * (data.ndim - 3)
    data = np.pad(data, padding, mode='constant', constant_values=(padval, padval))
    
    # tile the filters into an image
    data = data.reshape((n, n) + data.shape[1:]).transpose((0, 2, 1, 3) + tuple(range(4, data.ndim + 1)))
    data = data.reshape((n * data.shape[1], n * data.shape[3]) + data.shape[4:])
    
    plt.imshow(data)

The input image

In [7]:
# index four is the center crop
plt.imshow(net.deprocess('data', net.blobs['data'].data[4]))

The first layer filters, conv1

In [8]:
# the parameters are a list of [weights, biases]
filters = net.params['conv1'][0].data
vis_square(filters.transpose(0, 2, 3, 1))

The first layer output, conv1 (rectified responses of the filters above, first 36 only)

In [9]:
feat = net.blobs['conv1'].data[4, :36]
vis_square(feat, padval=1)

The second layer filters, conv2

There are 256 filters, each of which has dimension 5 x 5 x 48. We show only the first 48 filters, with each channel shown separately, so that each filter is a row.

In [10]:
filters = net.params['conv2'][0].data
vis_square(filters[:48].reshape(48**2, 5, 5))

The second layer output, conv2 (rectified, only the first 36 of 256 channels)

In [11]:
feat = net.blobs['conv2'].data[4, :36]
vis_square(feat, padval=1)

The third layer output, conv3 (rectified, all 384 channels)

In [12]:
feat = net.blobs['conv3'].data[4]
vis_square(feat, padval=0.5)

The fourth layer output, conv4 (rectified, all 384 channels)

In [13]:
feat = net.blobs['conv4'].data[4]
vis_square(feat, padval=0.5)

The fifth layer output, conv5 (rectified, all 256 channels)

In [14]:
feat = net.blobs['conv5'].data[4]
vis_square(feat, padval=0.5)

The fifth layer after pooling, pool5

In [15]:
feat = net.blobs['pool5'].data[4]
vis_square(feat, padval=1)

The first fully connected layer, fc6 (rectified)

We show the output values and the histogram of the positive values

In [16]:
feat = net.blobs['fc6'].data[4]
plt.subplot(2, 1, 1)
plt.plot(feat.flat)
plt.subplot(2, 1, 2)
_ = plt.hist(feat.flat[feat.flat > 0], bins=100)

The second fully connected layer, fc7 (rectified)

In [17]:
feat = net.blobs['fc7'].data[4]
plt.subplot(2, 1, 1)
plt.plot(feat.flat)
plt.subplot(2, 1, 2)
_ = plt.hist(feat.flat[feat.flat > 0], bins=100)

The final probability output, prob

In [18]:
feat = net.blobs['prob'].data[4]
plt.plot(feat.flat)
Out[18]:
[<matplotlib.lines.Line2D at 0x7fd63060d990>]

Let's see the top 5 predicted labels.

In [19]:
# load labels
imagenet_labels_filename = caffe_root + 'data/ilsvrc12/synset_words.txt'
try:
    labels = np.loadtxt(imagenet_labels_filename, str, delimiter='\t')
except:
    !../data/ilsvrc12/get_ilsvrc_aux.sh
    labels = np.loadtxt(imagenet_labels_filename, str, delimiter='\t')

# sort top k predictions from softmax output
top_k = net.blobs['prob'].data[4].flatten().argsort()[-1:-6:-1]
print labels[top_k]
['n02123045 tabby, tabby cat' 'n02123159 tiger cat'
 'n02124075 Egyptian cat' 'n02119022 red fox, Vulpes vulpes'
 'n02127052 lynx, catamount']