Some notes from my experiments so far:
Speed(tested on 0.4MP image, 4 octaves, 10 iterations):
Use "caffe.set_mode_gpu()" to enable GPU (if you have a CUDA compatible setup)
CPU (Intel i7-4710HQ CPU @ 2.50GHz): 154.99 s
GPU (Nvidia 850M): 7.03 s (22x faster)
Using NVIDIA cuDNN may improve GPU speed by a additional factor of 2x (haven't tried)
MemoryOn my 2GB GPU, I run out of memory when running on images larger than ~0.5 MP.
I have made the following function to automatically resize:
def loadWithMaxSize(imagename, maxSize = 400000):
ig = PIL.Image.open(imagename)
dim = ig.size[0]*ig.size[1]
print "Initial size: ", dim, " ", ig.size[0], "x", ig.size[1]
factor = math.sqrt(float(maxSize)/dim)
ig = ig.resize((int(factor*ig.size[0]),int(factor*ig.size[1])), PIL.Image.ANTIALIAS)
dim = ig.size[0]*ig.size[1]
print "Final size: ", dim, " ", ig.size[0], "x", ig.size[1]
return ig
imagename = '/home/mikael/Downloads/cheese1.jpg'
img = np.float32(loadWithMaxSize(imagename))
showarray(img)
It is also interesting to see what the net actually predicts to be in the image.
To do this you need the label definitions, which for instance can be downloaded from here:
https://github.com/HoldenCaulfieldRye/caffe/blob/master/data/ilsvrc12/synset_words.txtwith open("/home/mikael/Downloads/synset_words.txt") as f:
labels = f.readlines()
input_image = caffe.io.load_image(imagename)
prediction = net.predict([input_image], oversample=False)
top5predictions = prediction[0].argsort()[-5:][::-1]
for p in top5predictions:
print "Predicted class:", labels[p].strip('\n').split(' ', 1)[1], " (", "{:.3f}".format(100*prediction[0][p]), "%)"which will output something like:
Predicted class: cheeseburger ( 99.991 %)
Predicted class: hotdog, hot dog, red hot ( 0.004 %)
Predicted class: bagel, beigel ( 0.003 %)
Predicted class: guacamole ( 0.001 %)
Predicted class: plate ( 0.000 %)
In order for this to work, you need to added a 'raw_scale' parameter to the classifier initalization, e.g.:
net = caffe.Classifier('tmp.prototxt', param_fn,
mean = np.float32([104.0, 116.0, 122.0]), # ImageNet mean, training set dependent
channel_swap = (2,1,0), # the reference model has channels in BGR order instead of RGB
raw_scale=255) 
September 06, 2011, 04:07:55 PM
Dilber MC Theme by HarzeM