def ssc_color_info_abstraction(input,
mlp_list,
is_training,
bn_decay,
scope,
kernel_size,
bn=True):
kernel_h = kernel_size[0]
kernel_w = kernel_size[1]
channel = input.get_shape()[-1]
with tf.variable_scope(scope) as sc:
each_layer = basic_tf.conv2d(input,
channel, [1, 1],
padding='SAME',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv_p0',
bn_decay=bn_decay)
for i, num_out_channel in enumerate(mlp_list):
print('conv', i)
input = basic_tf.conv2d(input,
mlp_list[i] // 2, [1, 1],
padding='SAME',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv_p1%d' % (i),
bn_decay=bn_decay)
input = basic_tf.conv2d(input,
mlp_list[i], [kernel_h, kernel_w],
padding='SAME',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='convp2_%d' % (i),
bn_decay=bn_decay)
each_layer = tf.concat(
[each_layer, input],
-1) #get a depth_concatened hypercolume (b,h1,w1,sum(mlp))
#pooling kernel-size[2,2]
each_layer = basic_tf.max_pool2d(each_layer,
kernel_size=[2, 2],
stride=[1, 1],
padding='SAME',
scope='maxpool1')
#(b,h2,w2,sum[mlp])
print('one turn finished')
return each_layer
def basic_detectModel(img, is_training, bn_decay, cell_size, num_class,
box_num): #512->4
with tf.variable_scope('conv_unit1'):
out = basic_tf.conv2d(img, 16, [3, 3], 'conv_11', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_11', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 32, [3, 3], 'conv_12', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_12', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 64, [3, 3], 'conv_13', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_13', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 128, [3, 3], 'conv_14', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_14', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 256, [3, 3], 'conv_15', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_15', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 512, [3, 3], 'conv_16', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_16', [2, 2], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_17', [2, 2], 'SAME')
with tf.variable_scope('conv_unit2'):
out1 = basic_tf.conv2d(out, 1024, [3, 3], 'conv_21', [1, 1], 'SAME')
out1 = basic_tf.conv2d(out1, 512, [1, 1], 'conv_22', [1, 1], 'SAME')
with tf.variable_scope('fully_connected_unit'):
out2 = tf.reshape(out1, (int(out1._shape[0]), -1)) #b,8192
out2 = basic_tf.fully_connected(out2, 4096, 'fc1')
out2 = basic_tf.dropout(out2, is_training, 'dp1', 0.5)
out2 = basic_tf.fully_connected(out2, 1024, 'fc2')
out2 = basic_tf.dropout(out2, is_training, 'dp2', 0.5)
out2 = basic_tf.fully_connected(
out2, cell_size[0] * cell_size[1] * (class_num + box_num * 5),
'fc3')
with tf.variable_scope('output_unit'):
n1 = cell_size[0] * cell_size[1] * num_class #4*4*2=32
n2 = n1 + cell_size[0] * cell_size[1] * box_num #32+4*4*2=64
class_pred = tf.reshape(
out2[:, 0:n1],
(-1, cell_size[0], cell_size[1], num_class)) #(b,4,4,2)
scales = tf.reshape(
out2[:,
n1:n2], (-1, cell_size[0], cell_size[1], box_num)) #(b,4,4,2)
boxes = tf.reshape(
out2[:, n2:],
(-1, cell_size[0], cell_size[1], box_num * 4)) #(b,4,4,8)
pred = tf.concat([class_pred, scales, boxes], 3) #(b,4,4,12)
return pred
def pointnet_fp_module(xyz1,
xyz2,
points1,
points2,
mlp,
is_training,
bn_decay,
scope,
bn=True,
weight_calc=True): #ok now
''' PointNet Feature Propogation (FP) Module
Input:
xyz1: (b, n1, 3) TF tensor
xyz2: (b, n2, 3) TF tensor, sparser than xyz1
points1: (b, n1, n1) TF tensor
points2: (b, n2, n2) TF tensor
mlp: list of int32 -- output size for MLP on each point
Return:
new_points: (b, n1, mlp[-1]) TF tensor
'''
with tf.variable_scope(scope) as sc:
dist, idx = three_nn(xyz1, xyz2)
dist = tf.maximum(dist, 1e-10)
if weight_calc:
norm = tf.reduce_sum(tf.truediv(1.0, dist), axis=2,
keep_dims=True) #(b,n,3)->(b,n,1)
norm = tf.tile(norm, [1, 1, 3]) #(b,n,1)->(b,n,3)
weight = tf.truediv(tf.truediv(1.0, dist), norm) #(b,n,3)
else:
weight = get_weight(True, dist)
interpolated_points = three_interpolate(points2, idx, weight)
if points1 is not None:
new_points1 = tf.concat(
axis=2, values=[interpolated_points,
points1]) # B,ndataset1,nchannel1+nchannel2
else:
new_points1 = interpolated_points
new_points1 = tf.expand_dims(new_points1, 2)
for i, num_out_channel in enumerate(mlp):
new_points1 = basic_tf.conv2d(new_points1,
num_out_channel, [1, 1],
padding='VALID',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv_%d' % (i),
bn_decay=bn_decay)
new_points1 = tf.squeeze(new_points1, [2]) # B,n1,mlp[-1]
return new_points1
def input_down_size_unit(img,
c_out1,
cout_2,
is_training,
bn_decay,
scope,
bn=True):
img1 = basic_tf.conv2d(img,
c_out1, [3, 3],
padding='SAME',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv_input0',
bn_decay=bn_decay) #down_size
img1 = basic_tf.max_pool2d(img1, [3, 3],
scope='max_pool_input',
stride=[2, 2],
padding='SAME')
img1 = basic_tf.conv2d(img1,
cout_2, [3, 3],
padding='SAME',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv_input1',
bn_decay=bn_decay) #down_size
img1 = basic_tf.max_pool2d(img1, [3, 3],
scope='max_pool_input',
stride=[2, 2],
padding='SAME')
return img1
def basic_detectModel(img, is_training, bn_decay, num_class): #512->4
with tf.variable_scope('conv_unit1_G'):
out = basic_tf.conv2d(img, 16, [3, 3], 'conv_11', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_11', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 32, [3, 3], 'conv_12', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_12', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 64, [3, 3], 'conv_13', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_13', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 128, [3, 3], 'conv_14', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_14', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 256, [3, 3], 'conv_15', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_15', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 512, [3, 3], 'conv_16', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_16', [2, 2], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_17', [2, 2], 'SAME')
with tf.variable_scope('conv_unit2'):
out1 = basic_tf.conv2d(out, 1024, [3, 3], 'conv_21', [1, 1], 'SAME')
out1 = basic_tf.conv2d(out1, 512, [1, 1], 'conv_22', [1, 1], 'SAME')
out1 = basic_tf.avg_pool2d(out1, [2, 2], 'pre_avepool', [2, 2], 'SAME')
with tf.variable_scope('fully_connected_unit_G'):
out2 = tf.reshape(out1, (int(out1._shape[0]), -1)) #b,8192
#out2 = basic_tf.fully_connected(out2,4096,'fc1')
#out2 = basic_tf.dropout(out2,is_training,'dp1',0.5)
out2 = basic_tf.fully_connected(out2, 1024, 'fc2')
out2 = basic_tf.dropout(out2, is_training, 'dp2', 0.5)
out2 = basic_tf.fully_connected(out2, 128, 'fc3')
out2 = basic_tf.dropout(out2, is_training, 'dp3', 0.5)
with tf.variable_scope('output_unit_G'):
pred = basic_tf.fully_connected(out2, (num_class + 4),
'fc4',
activation_fn=None)
return pred
def feature_transform_net(inputs, mlp, is_training, bn_decay=None, K=64):
""" Feature Transform Net, input is BxNx1xK
mlp:list of output_channels
Return:
transformed_inputs (b,n,k)"""
b = xyz.get_shape()[0].value
n = xyz.get_shape()[1].value
net = inputs
for i, num_out_channel in enumerate(mlp):
net = basic_tf.conv2d(net,
num_out_channel, [1, 1],
padding='VALID',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='tconv%d' % (i + 1),
bn_decay=bn_decay)
net = basic_tf.max_pool2d(net, [n, 1], padding='VALID', scope='tmaxpool')
net = tf.reshape(net, [b, -1])
net = basic_tf.fully_connected(net,
512,
bn=True,
is_training=is_training,
scope='tfc1',
bn_decay=bn_decay)
net = basic_tf.fully_connected(net,
256,
bn=True,
is_training=is_training,
scope='tfc2',
bn_decay=bn_decay)
with tf.variable_scope('transform_feat') as sc:
weights = tf.get_variable('weights', [256, K * K],
initializer=tf.constant_initializer(0.0),
dtype=tf.float32)
biases = tf.get_variable('biases', [K * K],
initializer=tf.constant_initializer(0.0),
dtype=tf.float32)
biases += tf.constant(np.eye(K).flatten(), dtype=tf.float32)
transform = tf.matmul(net, weights)
transform = tf.nn.bias_add(transform, biases)
transform = tf.reshape(transform, [b, K, K])
transformed_inputs = tf.matmul(input, transform)
return transformed_inputs
def brute_classify(img, num_class, is_training, bn_decay): #512->4
with tf.variable_scope('conv_unit1'):
out = basic_tf.conv2d(img, 16, [3, 3], 'conv_11', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_11', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 32, [3, 3], 'conv_12', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_12', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 64, [3, 3], 'conv_13', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_13', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 128, [3, 3], 'conv_14', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_14', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 256, [3, 3], 'conv_15', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_15', [2, 2], 'SAME')
out = basic_tf.conv2d(out, 512, [3, 3], 'conv_16', [1, 1], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_16', [2, 2], 'SAME')
out = basic_tf.max_pool2d(out, [2, 2], 'maxpool_17', [2, 2], 'SAME')
with tf.variable_scope('conv_unit2'):
out1 = basic_tf.conv2d(out, 1024, [3, 3], 'conv_21', [1, 1], 'SAME')
out1 = basic_tf.conv2d(out1, 512, [1, 1], 'conv_22', [1, 1], 'SAME')
out1 = basic_tf.avg_pool2d(out1, [2, 2], 'pre_avepool', [2, 2], 'SAME')
with tf.variable_scope('fully_connected_unit'):
out2 = tf.reshape(out1, (int(out1._shape[0]), -1)) #b,4096
out2 = basic_tf.fully_connected(out2, 1024, 'fc1')
out2 = basic_tf.dropout(out2, is_training, 'dp1', 0.5)
out2 = basic_tf.fully_connected(out2, 128, 'fc2')
out2 = basic_tf.dropout(out2, is_training, 'dp2', 0.5)
pred = basic_tf.fully_connected(out2, num_class, 'fc3')
print(pred)
return pred
def output_down_size_unit(input, cout, is_training, bn_decay, scope, bn=True):
b = input.get_shape()[0].value
out = basic_tf.avg_pool2d(input, [3, 3],
scope='ave_pool_output',
stride=[1, 1],
padding='SAME')
out = basic_tf.conv2d(out,
cout, [1, 1],
padding='SAME',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='conv_output1',
bn_decay=bn_decay) #down_size
out = basic_tf.max_pool2d(out, [3, 3],
scope='max_pool_output',
stride=[2, 2],
padding='SAME')
out = tf.reshape(out, (b, -1)) #(-1,2048)
return out
def wnet_model_unit(input,
mlp_list,
kernel_size_list,
is_training,
bn_decay,
scope,
bn=True,
maxpool=True):
bf = input._shape[-1]
with tf.variable_scope(scope) as sc:
output = basic_tf.max_pool2d(input, [3, 3], 'maxpool_input', [1, 1],
'SAME')
output = basic_tf.conv2d(output,
bf // 3, [1, 1],
padding='SAME',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv_input',
bn_decay=bn_decay) #filter(output1)
for i, ks in enumerate(kernel_size_list):
kernel_h = ks[0]
kernel_w = ks[1]
out = input #conv from down_sized input
c_out = mlp_list[i]
if kernel_h == 1 and kernel_w == 1:
out = basic_tf.conv2d(
out,
c_out, [1, 1],
padding='SAME',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv_p1%d' % (i),
bn_decay=bn_decay) #conv from down_sized input
else:
out = basic_tf.conv2d(
out,
c_out // 2, [1, 1],
padding='SAME',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv_p0%d' % (i),
bn_decay=bn_decay) #conv from down_sized input
out = basic_tf.conv2d(
out,
c_out, [kernel_h, kernel_w],
padding='SAME',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv_p1%d' % (i),
bn_decay=bn_decay) #conv from down_sized input
#print(output.shape)
output = tf.concat([output, out], -1)
if maxpool:
output = basic_tf.max_pool2d(output, [3, 3], 'max_pool', [2, 2],
'SAME')
#(b,h2,w2,sum[mlp])
print(output.shape, 'one turn finished')
return output
def pointnet_AB_module_msg(xyz,
points,
m,
r_list,
ns_list,
mlp_list,
is_training,
bn_decay,
scope,
bn=True,
use_xyz=True):
''' PointNet Set Abstraction (SA) module with Multi-Scale Grouping (MSG)
Input:
xyz: (b, n, 3) TF tensor
points: (b, n, c) TF tensor
m: int32 -- #points sampled in farthest point sampling
r: list of float32 -- search radius in local region
ns: list of int32 -- how many points in each local region
mlp: list of list of int32 -- output size for MLP on each point
use_xyz: bool, if True concat XYZ with local point features, otherwise just use point features
Return:
new_xyz: (b, m, 3) TF tensor
new_points: (b, m, \sum_k{mlp[k][-1]}) TF tensor
'''
with tf.variable_scope(scope) as sc:
# sampling
new_xyz = gather_point(xyz, farthest_point_sample(m,
xyz)) #(b,npoint,3)
new_points_list = []
print('sample ok')
#grouping & convolution mlp
for i in range(len(r_list)):
print('grouping', i)
#grouping
r = r_list[i]
ns = ns_list[i]
idx, pts_cnt = query_ball_point(r, ns, xyz, new_xyz)
grouped_xyz = group_point(xyz, idx) #b,m,ns[i]
grouped_xyz -= tf.tile(tf.expand_dims(new_xyz, 2),
[1, 1, ns, 1]) #(b,m,ns[i],3)
if points is not None:
grouped_points = group_point(points, idx) #(b,m,ns[i],c)
if use_xyz:
grouped_points = tf.concat([grouped_points, grouped_xyz],
axis=-1)
else:
grouped_points = grouped_xyz
print('convolutioning')
#convolutional layers
for j, num_out_channel in enumerate(mlp_list[i]):
print('conv', j)
grouped_points = basic_tf.conv2d(
grouped_points,
num_out_channel, [1, 1],
padding='VALID',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv%d_%d' % (i, j),
bn_decay=bn_decay) #(b,m,ns[i],mlp[i][-1])
new_points = tf.reduce_max(grouped_points,
axis=[2]) #(b,m,mlp[i][-1])
new_points_list.append(new_points)
new_points_concat = tf.concat(new_points_list,
axis=-1) #size(b,m,sum_k{mlp[k][-1])
print('one turn')
return new_xyz, new_points_concat
def pointnet_AB_module(xyz,
points,
m,
r,
ns,
mlp,
mlp2,
group_all,
is_training,
bn_decay,
scope,
bn=True,
pooling='max',
tnet_spec=None,
knn=False,
use_xyz=True):
''' PointNet Set Abstraction (SA) Module
Input:
xyz: (b, n, 3) TF tensor
points: (b, n, c) TF tensor
m: int32 -- #points sampled in farthest point sampling
r: float32 -- search radius in local region
ns: int32 -- how many points in each local region
mlp: list of int32 -- output size for MLP on each point
mlp2: list of int32 -- output size for MLP on each region
group_all: bool -- group all points into one PC if set true, OVERRIDE
npoint, radius and nsample settings
use_xyz: bool, if True concat XYZ with local point features, otherwise just use point features
Return:
new_xyz: (b,m, 3) TF tensor
new_points: (b,m, mlp[-1] or mlp2[-1]) TF tensor
idx: (b,m, ns) int32 -- indices for local regions
'''
with tf.variable_scope(scope) as sc:
#Sampling&Grouping
if group_all:
ns = xyz.get_shape()[1].value
new_xyz, new_points, idx, grouped_xyz = sample_and_group_all(
xyz, points, use_xyz)
else:
new_xyz, new_points, idx, grouped_xyz = sample_and_group(
m, r, ns, xyz, points, tnet_spec, knn,
use_xyz) #here we got the idx from sampling&grouping
print('convolution')
#convolutional layer mlp(handling the new_points we got)
for i, num_out_channel in enumerate(mlp):
print('conv', i)
new_points = basic_tf.conv2d(new_points,
num_out_channel, [1, 1],
padding='VALID',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv%d' % (i),
bn_decay=bn_decay)
#pooling
print('pooling')
if pooling == 'avg':
new_points = basic_tf.avg_pool2d(new_points, [1, ns],
stride=[1, 1],
padding='VALID',
scope='avgpool1')
elif pooling == 'weighted_avg':
with tf.variable_scope('weighted_avg1'):
dists = tf.norm(grouped_xyz, axis=-1, ord=2, keep_dims=True)
exp_dists = tf.exp(-dists * 5)
weights = exp_dists / tf.reduce_sum(
exp_dists, axis=2, keep_dims=True) # (b, m, ns, 1)
new_points *= weights # (b, m, ns, mlp[-1])
new_points = tf.reduce_sum(new_points, axis=2, keep_dims=True)
elif pooling == 'max':
new_points = tf.reduce_max(new_points, axis=[2], keep_dims=True)
elif pooling == 'min':
new_points = basic_tf.max_pool2d(-1 * new_points, [1, ns],
stride=[1, 1],
padding='VALID',
scope='minpool1')
elif pooling == 'max_and_avg':
avg_points = basic_tf.max_pool2d(new_points, [1, ns],
stride=[1, 1],
padding='VALID',
scope='maxpool1')
max_points = basic_tf.avg_pool2d(new_points, [1, ns],
stride=[1, 1],
padding='VALID',
scope='avgpool1')
new_points = tf.concat([avg_points, max_points], axis=-1)
#convolutional layer mlp2
if mlp2 is None: mlp2 = []
for i, num_out_channel in enumerate(mlp2):
new_points = basic_tf.conv2d(new_points,
num_out_channel, [1, 1],
padding='VALID',
stride=[1, 1],
bn=bn,
is_training=is_training,
scope='conv_post_%d' % (i),
bn_decay=bn_decay)
#prepare the result
new_points = tf.squeeze(new_points, [2]) # (b,m,mlp2[-1])
print('1 turn')
return new_xyz, new_points, idx
def input_transform_net(xyz, mlp, is_training, bn_decay=None, K=3):
""" Input (XYZ) Transform Net, input is BxNx3 gray image
mlp:list of output_channels
Return:
transformed_xyz (b,n,3)"""
b = xyz.get_shape()[0].value
n = xyz.get_shape()[1].value
print('transing')
input = tf.expand_dims(xyz, -1) #(b,n,3,1)
for i, num_out_channel in enumerate(mlp):
if i == 0:
net = basic_tf.conv2d(input,
num_out_channel, [1, 3],
padding='VALID',
stride=[1, 1],
bn=True,
is_training=is_training,
scope='tconv%d' % (i + 1),
bn_decay=bn_decay)
else:
net = basic_tf.conv2d(net,
num_out_channel, [1, 1],
padding='VALID',
stride=[1, 1],
bn=bn_decay,
is_training=is_training,
scope='tconv%d' % (i + 1),
bn_decay=bn_decay) #(b,n,mlp[-1])
net = basic_tf.max_pool2d(net, [n, 1], padding='VALID',
scope='tmaxpool') #(b,1,mlp[-1])
net = tf.reshape(net, [b, -1]) #(b * mlp[-1])
net = basic_tf.fully_connected(net,
512,
bn=True,
is_training=is_training,
scope='tfc1',
bn_decay=bn_decay) #(b,512)
net = basic_tf.fully_connected(net,
256,
bn=True,
is_training=is_training,
scope='tfc2',
bn_decay=bn_decay) #(b,256)
print('transing2')
with tf.variable_scope('transform_XYZ') as sc:
assert (K == 3)
weights = tf.get_variable('weights', [256, 3 * K],
initializer=tf.constant_initializer(0.0),
dtype=tf.float32)
biases = tf.get_variable('biases', [3 * K],
initializer=tf.constant_initializer(0.0),
dtype=tf.float32)
biases += tf.constant([1, 0, 0, 0, 1, 0, 0, 0, 1], dtype=tf.float32)
transform = tf.matmul(net, weights)
transform = tf.nn.bias_add(transform, biases) #now,net_shape:(b,3*K)
transform = tf.reshape(transform, [b, 3, K])
transformed_xyz = tf.matmul(xyz, transform)
return transformed_xyz
def color_net(rgb, is_training, bn_decay=None):
with tf.variable_scope('input_layer'):
h = rgb.get_shape()[1].value
w = rgb.get_shape()[2].value
b = rgb.get_shape()[0].value
og = [h, w]
end_data = {}
end_data['rgb_data'] = rgb
#end_data['srgb_data']=srgb #both of them are been normalized
out1 = basic_tf.conv2d(rgb, 96, [1, 1], 'input_conv', [1, 1], 'SAME')
out1 = basic_tf.max_pool2d(out1, [2, 2], 'input_pool', [1, 1], 'SAME')
with tf.variable_scope('intermidate_layer'):
for i, kernels in enumerate(list_of_kernel):
mlps = list_of_mlplist[i]
out1 = bm.ssc_color_info_abstraction(out1,
mlps,
is_training=is_training,
bn_decay=bn_decay,
scope='ssc_section_%d' % (i),
kernel_size=kernels,
bn=True)
if i == 0:
hyper_colume = out1
else:
hyper_colume = tf.concat([hyper_colume, out1], -1)
hyper_colume = basic_tf.avg_pool2d(hyper_colume, [2, 2],
'medium_avepool', [1, 1], 'SAME')
c = hyper_colume.get_shape()[-1].value
print(hyper_colume.shape)
hyper_colume = tf.reshape(hyper_colume, (b * h * w, c))
with tf.variable_scope('output_layer'):
out = basic_tf.fully_connected(hyper_colume,
256,
bn=True,
is_training=is_training,
scope='fc2',
bn_decay=bn_decay)
out = basic_tf.dropout(out,
keep_prob=0.5,
is_training=is_training,
scope='dp2')
out = basic_tf.fully_connected(out,
64,
bn=True,
is_training=is_training,
scope='fc3',
bn_decay=bn_decay)
out = basic_tf.dropout(out,
keep_prob=0.5,
is_training=is_training,
scope='dp3')
out = basic_tf.fully_connected(out,
3,
bn=True,
is_training=is_training,
scope='fc4',
bn_decay=bn_decay)
pred = tf.reshape(out, (b, h, w, 3))
return pred, end_data
