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, set plotting parameters, and run ./scripts/download_model_binary.py models/bvlc_reference_caffenet to get the pretrained CaffeNet model if it hasn't already been fetched.
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'
import os
if not os.path.isfile(caffe_root + 'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel'):
print("Downloading pre-trained CaffeNet model...")
!../scripts/download_model_binary.py ../models/bvlc_reference_caffenet
Set Caffe to CPU mode, load the net in the test phase for inference, and configure input preprocessing.
In [2]:
caffe.set_mode_cpu()
net = caffe.Net(caffe_root + 'models/bvlc_reference_caffenet/deploy.prototxt',
caffe_root + 'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel',
caffe.TEST)
# input preprocessing: 'data' is the name of the input blob == net.inputs[0]
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2,0,1))
transformer.set_mean('data', np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy').mean(1).mean(1)) # mean pixel
transformer.set_raw_scale('data', 255) # the reference model operates on images in [0,255] range instead of [0,1]
transformer.set_channel_swap('data', (2,1,0)) # the reference model has channels in BGR order instead of RGB
Classify the image by reshaping the net for the single input then doing the forward pass.
In [3]:
net.blobs['data'].reshape(1,3,227,227)
net.blobs['data'].data[...] = transformer.preprocess('data', caffe.io.load_image(caffe_root + 'examples/images/cat.jpg'))
out = net.forward()
print("Predicted class is #{}.".format(out['prob'].argmax()))
The layer features and their shapes (1 is the batch size, corresponding to the single input image in this example).
In [4]:
[(k, v.data.shape) for k, v in net.blobs.items()]
Out[4]:
The parameters and their shapes. The parameters are net.params['name'][0] while biases are net.params['name'][1].
In [5]:
[(k, v[0].data.shape) for k, v in net.params.items()]
Out[5]:
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]:
plt.imshow(transformer.deprocess('data', net.blobs['data'].data[0]))
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[0, :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[0, :36]
vis_square(feat, padval=1)
The third layer output, conv3 (rectified, all 384 channels)
In [12]:
feat = net.blobs['conv3'].data[0]
vis_square(feat, padval=0.5)
The fourth layer output, conv4 (rectified, all 384 channels)
In [13]:
feat = net.blobs['conv4'].data[0]
vis_square(feat, padval=0.5)
The fifth layer output, conv5 (rectified, all 256 channels)
In [14]:
feat = net.blobs['conv5'].data[0]
vis_square(feat, padval=0.5)
The fifth layer after pooling, pool5
In [15]:
feat = net.blobs['pool5'].data[0]
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[0]
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[0]
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[0]
plt.plot(feat.flat)
Out[18]:
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[0].flatten().argsort()[-1:-6:-1]
print labels[top_k]