def _create_siamese(net_path, net_x, net_z):
# read mat file from net_path and start TF Siamese graph from placeholders X and Z
params_names_list, params_values_list = _import_from_matconvnet(net_path)
# loop through the flag arrays and re-construct network, reading parameters of conv and bnorm layers
for i in range(_num_layers):
print('> Layer ' + str(i + 1))
# conv
conv_W_name = _find_params('conv' + str(i + 1) + 'f',
params_names_list)[0]
conv_b_name = _find_params('conv' + str(i + 1) + 'b',
params_names_list)[0]
print('Conv:setting %s %s' % (conv_W_name, conv_b_name))
print('Conv : stride %s filter-group %s' %
(str(_conv_stride[i]), str(_filtergroup_yn[i])))
conv_W = params_values_list[params_names_list.index(conv_W_name)]
conv_b = params_values_list[params_names_list.index(conv_b_name)]
# batchnorm
if _bnorm_yn[i]:
bn_beta_name = _find_params('bn' + str(i + 1) + 'b',
params_names_list)[0]
bn_gamma_name = _find_params('bn' + str(i + 1) + 'm',
params_names_list)[0]
bn_moments_name = _find_params('bn' + str(i + 1) + 'x',
params_names_list)[0]
#print ('\t\tBNORM: setting '+bn_beta_name+' '+bn_gamma_name+' '+bn_moments_name
bn_beta = params_values_list[params_names_list.index(bn_beta_name)]
bn_gamma = params_values_list[params_names_list.index(
bn_gamma_name)]
bn_moments = params_values_list[params_names_list.index(
bn_moments_name)]
bn_moving_mean = bn_moments[:, 0]
bn_moving_variance = bn_moments[:,
1]**2 # saved as std in matconvnet
else:
bn_beta = bn_gamma = bn_moving_mean = bn_moving_variance = []
# set up conv "block" with bnorm and activation
net_x = set_convolutional(net_x, conv_W, np.swapaxes(conv_b,0,1), _conv_stride[i], \
bn_beta, bn_gamma, bn_moving_mean, bn_moving_variance, \
filtergroup=_filtergroup_yn[i], batchnorm=_bnorm_yn[i], activation=_relu_yn[i], \
scope='conv'+str(i+1), reuse=False)
# notice reuse=True for Siamese parameters sharing
net_z = set_convolutional(net_z, conv_W, np.swapaxes(conv_b,0,1), _conv_stride[i], \
bn_beta, bn_gamma, bn_moving_mean, bn_moving_variance, \
filtergroup=_filtergroup_yn[i], batchnorm=_bnorm_yn[i], activation=_relu_yn[i], \
scope='conv'+str(i+1), reuse=True)
# add max pool if required
if _pool_stride[i] > 0:
print("_pool_stride")
return net_z, net_x, params_names_list, params_values_list
def _create_siamese(net_path, net_x, net_z):
# read mat file from net_path and start TF Siamese graph from placeholders X and Z
params_names_list, params_values_list = _import_from_matconvnet(net_path)
# loop through the flag arrays and re-construct network, reading parameters of conv and bnorm layers
for i in xrange(_num_layers):
print '> Layer '+str(i+1)
# conv
conv_W_name = _find_params('conv'+str(i+1)+'f', params_names_list)[0]
conv_b_name = _find_params('conv'+str(i+1)+'b', params_names_list)[0]
print '\t\tCONV: setting '+conv_W_name+' '+conv_b_name
print '\t\tCONV: stride '+str(_conv_stride[i])+', filter-group '+str(_filtergroup_yn[i])
conv_W = params_values_list[params_names_list.index(conv_W_name)]
conv_b = params_values_list[params_names_list.index(conv_b_name)]
# batchnorm
if _bnorm_yn[i]:
bn_beta_name = _find_params('bn'+str(i+1)+'b', params_names_list)[0]
bn_gamma_name = _find_params('bn'+str(i+1)+'m', params_names_list)[0]
bn_moments_name = _find_params('bn'+str(i+1)+'x', params_names_list)[0]
print '\t\tBNORM: setting '+bn_beta_name+' '+bn_gamma_name+' '+bn_moments_name
bn_beta = params_values_list[params_names_list.index(bn_beta_name)]
bn_gamma = params_values_list[params_names_list.index(bn_gamma_name)]
bn_moments = params_values_list[params_names_list.index(bn_moments_name)]
bn_moving_mean = bn_moments[:,0]
bn_moving_variance = bn_moments[:,1]**2 # saved as std in matconvnet
else:
bn_beta = bn_gamma = bn_moving_mean = bn_moving_variance = []
# set up conv "block" with bnorm and activation
net_x = set_convolutional(net_x, conv_W, np.swapaxes(conv_b,0,1), _conv_stride[i], \
bn_beta, bn_gamma, bn_moving_mean, bn_moving_variance, \
filtergroup=_filtergroup_yn[i], batchnorm=_bnorm_yn[i], activation=_relu_yn[i], \
scope='conv'+str(i+1), reuse=False)
# notice reuse=True for Siamese parameters sharing
net_z = set_convolutional(net_z, conv_W, np.swapaxes(conv_b,0,1), _conv_stride[i], \
bn_beta, bn_gamma, bn_moving_mean, bn_moving_variance, \
filtergroup=_filtergroup_yn[i], batchnorm=_bnorm_yn[i], activation=_relu_yn[i], \
scope='conv'+str(i+1), reuse=True)
# TODO: Are we using max pooling? Remove if not.
# add max pool if required
if _pool_stride[i]>0:
print '\t\tMAX-POOL: size '+str(_pool_sz)+ ' and stride '+str(_pool_stride[i])
net_x = tf.nn.max_pool(net_x, [1,_pool_sz,_pool_sz,1], strides=[1,_pool_stride[i],_pool_stride[i],1], padding='VALID', name='pool'+str(i+1))
net_z = tf.nn.max_pool(net_z, [1,_pool_sz,_pool_sz,1], strides=[1,_pool_stride[i],_pool_stride[i],1], padding='VALID', name='pool'+str(i+1))
print
return net_z, net_x, params_names_list, params_values_list
def _create_siamese(net_path, net_x, net_z, istrain):
# loop through the flag arrays and re-construct network, reading parameters of conv and bnorm layers
for i in range(_num_layers):
print( '> Layer '+str(i+1))
# set up conv "block" with bnorm and activation
net_x = set_convolutional(net_x, conv_W, np.swapaxes(conv_b,0,1), _conv_stride[i], istrain=istrain, \
filtergroup=_filtergroup_yn[i], batchnorm=_bnorm_yn[i], activation=_relu_yn[i], \
scope='conv'+str(i+1), reuse=False)
# notice reuse=True for Siamese parameters sharing
net_z = set_convolutional(net_z, conv_W, np.swapaxes(conv_b,0,1), _conv_stride[i], istrain=istrain, \
filtergroup=_filtergroup_yn[i], batchnorm=_bnorm_yn[i], activation=_relu_yn[i], \
scope='conv'+str(i+1), reuse=True)
# add max pool if required
if _pool_stride[i]>0:
print( '\t\tMAX-POOL: size '+str(_pool_sz)+ ' and stride '+str(_pool_stride[i]))
net_x = tf.nn.max_pool(net_x, [1,_pool_sz,_pool_sz,1], strides=[1,_pool_stride[i],_pool_stride[i],1], padding='VALID', name='pool'+str(i+1))
net_z = tf.nn.max_pool(net_z, [1,_pool_sz,_pool_sz,1], strides=[1,_pool_stride[i],_pool_stride[i],1], padding='VALID', name='pool'+str(i+1))
print()
return net_z, net_x, params_names_list, params_values_list
def create_net(net_x, net_z):
#-------------------------------------------------------------------------
#function//net_x:instance frame ;net_z:template frame
#-------------------------------------------------------------------------
#not sure
#W_param_list=[n for n in range(0,_nums_layers)]
#b_param_list=[n for n in range(0,_nums_layers)]
for i in range(_nums_layers):
print('Layer ' + str(i + 1))
#set up the conv bolck
#set_convolutional(X,stride,bn_beta,bn_gamma,bn_init_mean,bn_init_var,batchnorm=True,activation=True,reuse=False,scope=None):
print(net_z)
print(net_x)
net_x = set_convolutional(
net_x,
[_conv_w_sz[i], _conv_w_sz[i], _conv_w_in_c[i], _conv_w_out[i]],
_conv_w_out[i], #the shape of W and b
_conv_stride[i],
0,
0,
0,
0,
batchnorm=False,
activation=_if_relu[i],
reuse=True,
scope='conv' + str(i + 1))
net_z = set_convolutional(
net_z,
[_conv_w_sz[i], _conv_w_sz[i], _conv_w_in_c[i], _conv_w_out[i]],
_conv_w_out[i], #the shape of W and b
_conv_stride[i],
0,
0,
0,
0,
batchnorm=False,
activation=_if_relu[i],
reuse=True,
scope='conv' + str(i + 1))
print(net_z)
print(net_x)
print('Layer ' + str(i + 1) + ' conv end')
#if having the pooling
if _pool_stride[i] > 0:
print("_pool_stride")
net_x = tf.nn.max_pool(
net_x, [1, _pool_sz[i], _pool_sz[i], 1],
strides=[1, _pool_stride[i], _pool_stride[i], 1],
padding='VALID',
name='pool' + str(i + 1))
net_z = tf.nn.max_pool(
net_z, [1, _pool_sz[i], _pool_sz[i], 1],
strides=[1, _pool_stride[i], _pool_stride[i], 1],
padding='VALID',
name='pool' + str(i + 1))
print(net_z)
print(net_x)
print('Layer ' + str(i + 1) + ' end')
return net_z, net_x
