In this example, we convert GoogleNet trained on Caffe to SINGA for image classification.
Download the parameter checkpoint file into this folder
$ wget https://s3-ap-southeast-1.amazonaws.com/dlfile/bvlc_googlenet.tar.gz $ tar xvf bvlc_googlenet.tar.gz
Run the program
# use cpu $ python serve.py -C & # use gpu $ python serve.py &
Submit images for classification
$ curl -i -F image=@image1.jpg http://localhost:9999/api $ curl -i -F image=@image2.jpg http://localhost:9999/api $ curl -i -F image=@image3.jpg http://localhost:9999/api
image1.jpg, image2.jpg and image3.jpg should be downloaded before executing the above commands.
We first extract the parameter values from Caffe's checkpoint file into a pickle version After downloading the checkpoint file into caffe_root/python folder, run the following script
# to be executed within caffe_root/python folder
import caffe
import numpy as np
import cPickle as pickle
model_def = '../models/bvlc_googlenet/deploy.prototxt'
weight = 'bvlc_googlenet.caffemodel' # must be downloaded at first
net = caffe.Net(model_def, weight, caffe.TEST)
params = {}
for layer_name in net.params.keys():
weights=np.copy(net.params[layer_name][0].data)
bias=np.copy(net.params[layer_name][1].data)
params[layer_name+'_weight']=weights
params[layer_name+'_bias']=bias
print layer_name, weights.shape, bias.shape
with open('bvlc_googlenet.pickle', 'wb') as fd:
pickle.dump(params, fd)
Then we construct the GoogleNet using SINGA's FeedForwardNet structure. Note that we added a EndPadding layer to resolve the issue from discrepancy of the rounding strategy of the pooling layer between Caffe (ceil) and cuDNN (floor). Only the MaxPooling layers outside inception blocks have this problem. Refer to this for more detials.