def process( (i, image, label) ):
image1_caffe = image.reshape(1, *image.shape)
data = np.asarray([ image1_caffe ])
out = net.forward_all(data=data, blobs=[ args.layer ])
pt1 = utils.flat_shape(out[args.layer][0])
diffs = []
dists = []
for tr in trs:
f, name = tr['f'], tr['name']
image2 = f(image)
image2_caffe = image2.reshape(1, *image.shape)
data = np.asarray([ image2_caffe ])
out = net.forward_all(data=data, blobs=[ args.layer ])
pt2 = utils.flat_shape(out[args.layer][0])
# compute angle diff and distance
norm = np.linalg.norm(pt1) * np.linalg.norm(pt2)
diff = pt2 - pt1
dist = np.linalg.norm(diff)
diffs.append( (label, name, diff) )
dists.append( (label, name, dist) )
return (diffs, dists)
def init_params(self):
#展平纬度, 初始化参数值
std = 0.01
shape = (utils.flat_shape(self.inshape),
utils.flat_shape(self.outshape))
wval = np.random.randn(shape[0], shape[1]) * std
bval = np.zeros(shape[1])
self.__W = LayerParam(self.name, 'W', wval)
self.__b = LayerParam(self.name, 'b', bval)
def weight_initializers(self):
initializers = {}
initializers['uniform'] = lambda shape: np.random.uniform(
-1, 1, shape) * 0.1
initializers['normal'] = lambda shape: np.random.randn(
utils.flat_shape(shape)).reshape(shape) * 0.1
return initializers
def process( (i, image, label) ):
res = []
for tr in trs:
f, name = tr['f'], tr['name']
image2 = f(image)
image2_caffe = image2.reshape(1, *image.shape)
data = np.asarray([ image2_caffe/255. ]) # normalize
out = net.forward_all(data=data, blobs=[ args.layer ])
pt = utils.flat_shape(out[args.layer][0])
res.append((label, name, pt))
return res
def init_params(self):
inshape = self.__inshape
outshape = self.__outshape
in_chnls = inshape[0]
out_chnls = outshape[0]
#pdb.set_trace()
#展平形状(c*kh*kw, c_), 把卷积运算转换成矩阵运算, 优化性能
shape = (in_chnls * utils.flat_shape(self.__ks), out_chnls)
wval = self.__W(shape)
bval = self.__b((shape[1], ))
W = LayerParam(self.name, 'weight', wval)
b = LayerParam(self.name, 'bias', bval)
self.__W = W
self.__b = b
relu_conv1x1_inter_res4_rgb = transf_block(inter_res4_rgb, dt, conv_kernel_constraint, training_phase, 'redux_rgb_inter_res4')
relu_conv1x1_res4_rgb = transf_block(res4_rgb, dt, conv_kernel_constraint, training_phase, 'redux_rgb_res4')
relu_conv1x1_inter_res5_rgb = transf_block(inter_res5_rgb, dt, conv_kernel_constraint, training_phase, 'redux_rgb_inter_res5')
relu_conv1x1_res5_rgb = transf_block(res5_rgb, dt, conv_kernel_constraint, training_phase, 'redux_rgb_res5')
relu_conv1x1_res1_depth = transformation_block(res1_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_res1')
relu_conv1x1_inter_res2_depth = transf_block(inter_res2_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_inter_res2')
relu_conv1x1_res2_depth = transf_block(res2_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_res2')
relu_conv1x1_inter_res3_depth = transf_block(inter_res3_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_inter_res3')
relu_conv1x1_res3_depth = transf_block(res3_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_res3')
relu_conv1x1_inter_res4_depth = transf_block(inter_res4_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_inter_res4')
relu_conv1x1_res4_depth = transf_block(res4_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_res4')
relu_conv1x1_inter_res5_depth = transf_block(inter_res5_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_inter_res5')
relu_conv1x1_res5_depth = transf_block(res5_depth, dt, conv_kernel_constraint, training_phase, 'redux_depth_res5')
relu_conv1x1_res1_rgb = tf.reshape(relu_conv1x1_res1_rgb, [-1, flat_shape(relu_conv1x1_res1_rgb)])
relu_conv1x1_inter_res2_rgb = tf.reshape(relu_conv1x1_inter_res2_rgb, [-1, flat_shape(relu_conv1x1_inter_res2_rgb)])
relu_conv1x1_res2_rgb = tf.reshape(relu_conv1x1_res2_rgb, [-1, flat_shape(relu_conv1x1_res2_rgb)])
relu_conv1x1_inter_res3_rgb = tf.reshape(relu_conv1x1_inter_res3_rgb, [-1, flat_shape(relu_conv1x1_inter_res3_rgb)])
relu_conv1x1_res3_rgb = tf.reshape(relu_conv1x1_res3_rgb, [-1, flat_shape(relu_conv1x1_res3_rgb)])
relu_conv1x1_inter_res4_rgb = tf.reshape(relu_conv1x1_inter_res4_rgb, [-1, flat_shape(relu_conv1x1_inter_res4_rgb)])
relu_conv1x1_res4_rgb = tf.reshape(relu_conv1x1_res4_rgb, [-1, flat_shape(relu_conv1x1_res4_rgb)])
relu_conv1x1_inter_res5_rgb = tf.reshape(relu_conv1x1_inter_res5_rgb, [-1, flat_shape(relu_conv1x1_inter_res5_rgb)])
relu_conv1x1_res5_rgb = tf.reshape(relu_conv1x1_res5_rgb, [-1, flat_shape(relu_conv1x1_res5_rgb)])
relu_conv1x1_res1_depth = tf.reshape(relu_conv1x1_res1_depth, [-1, flat_shape(relu_conv1x1_res1_depth)])
relu_conv1x1_inter_res2_depth = tf.reshape(relu_conv1x1_inter_res2_depth, [-1, flat_shape(relu_conv1x1_inter_res2_depth)])
relu_conv1x1_res2_depth = tf.reshape(relu_conv1x1_res2_depth, [-1, flat_shape(relu_conv1x1_res2_depth)])
relu_conv1x1_inter_res3_depth = tf.reshape(relu_conv1x1_inter_res3_depth, [-1, flat_shape(relu_conv1x1_inter_res3_depth)])
relu_conv1x1_res3_depth = tf.reshape(relu_conv1x1_res3_depth, [-1, flat_shape(relu_conv1x1_res3_depth)])
relu_conv1x1_inter_res4_depth = tf.reshape(relu_conv1x1_inter_res4_depth, [-1, flat_shape(relu_conv1x1_inter_res4_depth)])
def set_prev(self, prev_layer):
inshape = prev_layer.outshape
self.__inshape = inshape
self.__outshape = (utils.flat_shape(inshape), )
