import os
import sys
sys.path.append('../') # WSL lib denpendancy
import lib
import caffe
from lib.net_wrapper import net_wrapper
if __name__ == "__main__" :
# original network
deploy = '../models/vgg16_conv/test.prototxt'
model = '../models/vgg16_conv/pre.caffemodel'
net = net_wrapper(deploy, model)
# fully conv network
deploy = '../models/vgg16_conv/test_conv.prototxt'
net_full_conv = net_wrapper(deploy, model)
params = ['fc6', 'fc7', 'fc8']
fc_params = {pr: (net._net.params[pr][0].data, net._net.params[pr][1].data) for pr in params}
for fc in params:
print '{} weights are {} dimensional and biases are {} dimensional'.format(fc, fc_params[fc][0].shape, fc_params[fc][1].shape)
params_full_conv = ['fc6-conv', 'fc7-conv', 'fc8-conv']
conv_params = {pr: (net_full_conv._net.params[pr][0].data, net_full_conv._net.params[pr][1].data) for pr in params_full_conv}
for conv in params_full_conv:
print '{} weights are {} dimensional and biases are {} dimensional'.format(conv, conv_params[conv][0].shape, conv_params[conv][1].shape)
# converted network (fully conv. net)
for pr, pr_conv in zip(params, params_full_conv) :
from lib.pascal_db import pascal_db
from lib.net_wrapper import net_wrapper
from lib.analyzer import analyzer
import numpy as np
from PIL import Image
if __name__ == '__main__' :
pascal = pascal_db('val','2012','/data/PASCAL/VOCdevkit/', 'cls')
#arch = 'segmentation/stride_32_only_label'
arch = 'segmentation/cls_net'
deploy = '../models/' + arch + '/test.prototxt'
#model = '../models/' + arch + '/finetuned_models/alex_ift_iter_1500.caffemodel'
model = '../models/' + arch + '/finetuned_models/ftmodels_iter_3000.caffemodel'
net = net_wrapper(deploy, model,3)
#net = analyzer(deploy, model)
for i, ind in enumerate(pascal._image_index) :
image_path = os.path.join( pascal._data_path, 'JPEGImages', ind + pascal._image_ext )
net.run_forward(image_path)
res = net._output['prob']
#res = net._output['pool-global']
res = res.squeeze()
res = res[np.newaxis,:]
if i == 0 :
final_res = np.copy(res)
else :
final_res = np.vstack((final_res,res))
print '{:d} th image ... {:s}'.format(i+1, ind)
# wrote the results
pascal.write_voc_results_files(final_res)
import matplotlib.pyplot as plt
import scipy
if __name__ == '__main__' :
gpu_id = sys.argv[1]
arch = sys.argv[2]
setting = sys.argv[3]
# load pascal db
pascal = pascal_db('val','2012','/data/PASCAL/VOCdevkit/', 'seg')
deploy = './0__MODELS/' + arch + '/' + setting + '/test.prototxt'
model = './0__MODELS/' + arch + '/' + setting + '/ft_models/final.caffemodel'
net = net_wrapper(deploy, model, int(gpu_id))
comp_id = setting
save_path = os.path.join(pascal._devkit_path, 'results', 'VOC' + pascal._year, 'Segmentation', comp_id + '_' + pascal._image_set + '_cls' )
if not os.path.exists( save_path ) :
os.makedirs( save_path )
for i, ind in enumerate(pascal._image_index) :
image_path = os.path.join( pascal._data_path, 'JPEGImages', ind + pascal._image_ext )
net.run_forward(image_path)
res = net._output['prob']
label_map = np.argmax(res[0],axis=0).astype(np.uint8)
gt_path = os.path.join( pascal._data_path, 'SegmentationClass', ind + '.png' )
gt = Image.open(gt_path)
res_map = scipy.misc.toimage(label_map, pal=pascal._color_map, mode='P')
res_map = res_map.resize( gt.size, Image.NEAREST )
