Skip to content

Instantly share code, notes, and snippets.

@FranciscoCanas
Last active May 26, 2020 04:05
Show Gist options
  • Save FranciscoCanas/844aab52fcff6b1c652a to your computer and use it in GitHub Desktop.
Save FranciscoCanas/844aab52fcff6b1c652a to your computer and use it in GitHub Desktop.
Combining Pre-trained Left and Right nets into a single joint model
import numpy as np
import sys, os
# Edit the paths as needed:
caffe_root = '../caffe/'
sys.path.insert(0, caffe_root + 'python')
import caffe
# Path to your combined net prototxt files:
combined_model_root_path = './models/combined_net/'
l_model_root_path = './models/left/2015-07-20/'
r_model_root_path= './models/right/2015-07-20/'
# The pre-trained Caffemodel files:
lnet_file = '_iter_50000.caffemodel'
rnet_file = '_iter_60000.caffemodel'
# Their respective prototxt files:
lnet_proto = 'net.prototxt'
rnet_proto = 'net.prototxt'
# Chdir if your prototxt files specify your training and testing files
# in relative paths:
os.chdir(l_model_root_path)
lnet = caffe.Net(lnet_proto, lnet_file, caffe.TRAIN)
os.chdir(r_model_root_path)
rnet = caffe.Net(rnet_proto, rnet_file, caffe.TRAIN)
os.chdir(combined_model_root_path)
comb_net = caffe.Net('net.prototxt', caffe.TRAIN)
# The layers you want to combine into the new caffe net:
layer_names = ['ff1', 'ff2']
# The two nets we have already loaded:
nets = {
'l': lnet,
'r': rnet,
}
# For each of the pretrained net sides, copy the params to
# the corresponding layer of the combined net:
for side, net in nets.items():
for layer in layer_names:
W = net.params[layer][0].data[...] # Grab the pretrained weights
b = net.params[layer][1].data[...] # Grab the pretrained bias
comb_net.params['{}_{}'.format(side, layer)][0].data[...] = W # Insert into new combined net
comb_net.params['{}_{}'.format(side, layer)][1].data[...] = b
# Save the combined model with pretrained weights to a caffemodel file:
comb_net.save('pretrained.caffemodel')
@Engineering-Course
Copy link

I wonder the two pre-trained model are the same networks or different?

@priyapaul
Copy link

priyapaul commented Apr 21, 2017

I am not able to load two nets, my terminal becomes unresponsive. Is there anything to be taken care of? The following codes prints loaded rgb net and then tries to load the second net . After printing out the following, it hangs

I0421 18:26:05.947526 10280 net.cpp:408] data -> label
I0421 18:26:05.947546 10280 cpm_data_transformer.cpp:242] CPMDataTransformer constructor done.

os.chdir(combined_model_root_path)
net_rgb = caffe.Net(net_proto, net_file, caffe.TRAIN)
print("loaded rgb net")
comb_net = caffe.Net('pose_train_test_rgb.prototxt', caffe.TRAIN)
print("loaded comb net")

@balajib363
Copy link

W0523 12:27:00.884069 6835 _caffe.cpp:139] DEPRECATION WARNING - deprecated use of Python interface
I am facing this error, when running the script.
I have installed caffe from source and version is 1.0.0.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment