def hourglass(self, inputs, n, numOut, name = 'hourglass'): with tf.name_scope(name): up_1 = residual(inputs, numOut, name = 'up1') low_ = tf.contrib.layers.max_pool2d(inputs, [2,2],[2,2], 'VALID') low_1 = residual(low_, numOut, name = 'low1') if n > 0: low_2 = self.hourglass(low_1, n-1, numOut, name='low2') else: low_2 = residual(low_1, numOut, name='low2') low_3 = residual(low_2, numOut, name = 'low3') up_2 = tf.image.resize_nearest_neighbor(low_3, tf.shape(low_3)[1:3]*2, name= 'upsampling') return tf.add([up_1, up_2])
def hourglass(self, inputs, n, numOut, name = 'hourglass'): with tf.name_scope(name): up_1 = residual(inputs, numOut, name = 'up1') low_ = tf.contrib.layers.max_pool2d(inputs, [2,2],[2,2], 'VALID') low_1 = residual(low_, numOut, name = 'low1') if n > 0: low_2 = self.hourglass(low_1, n-1, numOut, name='low2') else: low_2 = residual(low_1, numOut, name='low2') low_3 = residual(low_2, numOut, name = 'low3') up_2 = tf.image.resize_nearest_neighbor(low_3, tf.shape(low_3)[1:3]*2, name= 'upsampling') return tf.add([up_1, up_2])
def __call__(self, inputs):
with tf.name_scope(self.name):
with tf.name_scope('preprocessing'):
pad_1 = tf.pad(inputs, np.array([[0,0],[2,2],[2,2],[0,0]]))
conv_1 = conv2d(pad_1, 64, kernel_size=6, strides = 2, name = '256to128')
res_1 = residual(conv_1, 128)
pool_1 = tf.contrib.layers.max_pool2d(res_1, [2,2], [2,2], padding= 'VALID')
res_2 = residual(pool_1, 128)
res_3 = residual(res_2, self.nFeat)
# Supervision Table
hg = [None] * self.nbStack
ll = [None] * self.nbStack
ll_ = [None] * self.nbStack
drop = [None] * self.nbStack
out = [None] * self.nbStack
out_ = [None] * self.nbStack
sum_ = [None] * self.nbStack
with tf.name_scope('stacks'):
with tf.name_scope('hourglass.1'):
hg[0] = self.hourglass(res_3, self.nLow, self.nFeat, 'hourglass')
ll[0] = convBnrelu(hg[0], self.nFeat, name= 'conv_1')
ll_[0] = conv2d(ll[0],self.nFeat,1,1,'VALID','ll')
drop[0] = tf.layers.dropout(ll_[0], rate = 0.1, training = self.train)
out[0] = conv2d(ll[0],self.outDim,1,1,'VALID','out')
out_[0] = conv2d(out[0],self.nFeat,1,1,'VALID','out_')
sum_[0] = tf.add_n([drop[0], out_[0], res_3])
for i in range(1, self.nbStack-1):
with tf.name_scope('hourglass.' + str(i+1)):
hg[i] = self.hourglass(sum_[i-1], self.nLow, self.nFeat, 'hourglass')
ll[i] = convBnrelu(hg[i], self.nFeat, name='conv_1')
ll_[i] = conv2d(ll[i],self.nFeat,1,1,'VALID','ll')
drop[i] = tf.layers.dropout(ll_[i],rate=0.1, training = self.train)
out[i] = conv2d(ll[i],self.outDim,1,1,'VALID','out')
out_[i] = conv2d(out[i],self.nFeat,1,1,'VALID','out_')
sum_[i] = tf.add_n([drop[i], out_[i], sum_[i-1]])
with tf.name_scope('hourglass.' + str(self.nbStack)):
hg[self.nbStack-1] = self.hourglass(sum_[self.nbStack - 2], self.nLow, self.nFeat, 'hourglass')
ll[self.nbStack-1] = convBnrelu(hg[self.nbStack - 1], self.nFeat, name='conv_1')
drop[self.nbStack-1] = tf.layers.dropout(ll[self.nbStack-1], rate=0.1, training = self.train)
out[self.nbStack-1] = conv2d(drop[self.nbStack-1],self.outDim,1,1,'VALID', 'out')
return tf.stack(out, name = 'output')
def __call__(self, inputs):
with tf.name_scope(self.name):
with tf.name_scope('preprocessing'):
pad_1 = tf.pad(inputs, np.array([[0,0],[2,2],[2,2],[0,0]]))
conv_1 = conv2d(pad_1, 64, kernel_size=6, strides = 2, name = '256to128')
res_1 = residual(conv_1, 128)
pool_1 = tf.contrib.layers.max_pool2d(res_1, [2,2], [2,2], padding= 'VALID')
res_2 = residual(pool_1, 128)
res_3 = residual(res_2, self.nFeat)
# Supervision Table
hg = [None] * self.nbStack
ll = [None] * self.nbStack
ll_ = [None] * self.nbStack
drop = [None] * self.nbStack
out = [None] * self.nbStack
out_ = [None] * self.nbStack
sum_ = [None] * self.nbStack
with tf.name_scope('stacks'):
with tf.name_scope('hourglass.1'):
hg[0] = self.hourglass(res_3, self.nLow, self.nFeat, 'hourglass')
ll[0] = convBnrelu(hg[0], self.nFeat, name= 'conv_1')
ll_[0] = conv2d(ll[0],self.nFeat,1,1,'VALID','ll')
drop[0] = tf.layers.dropout(ll_[0], rate = 0.1, training = self.train)
out[0] = conv2d(ll[0],self.outDim,1,1,'VALID','out')
out_[0] = conv2d(out[0],self.nFeat,1,1,'VALID','out_')
sum_[0] = tf.add_n([drop[0], out_[0], res_3])
for i in range(1, self.nbStack-1):
with tf.name_scope('hourglass.' + str(i+1)):
hg[i] = self.hourglass(sum_[i-1], self.nLow, self.nFeat, 'hourglass')
ll[i] = convBnrelu(hg[i], self.nFeat, name='conv_1')
ll_[i] = conv2d(ll[i],self.nFeat,1,1,'VALID','ll')
drop[i] = tf.layers.dropout(ll_[i],rate=0.1, training = self.train)
out[i] = conv2d(ll[i],self.outDim,1,1,'VALID','out')
out_[i] = conv2d(out[i],self.nFeat,1,1,'VALID','out_')
sum_[i] = tf.add_n([drop[i], out_[i], sum_[i-1]])
with tf.name_scope('hourglass.' + str(self.nbStack)):
hg[self.nbStack-1] = self.hourglass(sum_[self.nbStack - 2], self.nLow, self.nFeat, 'hourglass')
ll[self.nbStack-1] = convBnrelu(hg[self.nbStack - 1], self.nFeat, name='conv_1')
drop[self.nbStack-1] = tf.layers.dropout(ll[self.nbStack-1], rate=0.1, training = self.train)
out[self.nbStack-1] = conv2d(drop[self.nbStack-1],self.outDim,1,1,'VALID', 'out')
return tf.stack(out, name = 'output')
def network_arm(inputs):
with tf.variable_scope("conv1"):
o_conv1 = layer(inputs, 16, [5, 5])
with tf.variable_scope("res1"):
o_res1 = residual(o_conv1, 16, 1)
with tf.variable_scope("res2"):
o_res2 = residual(o_res1, 16, 1)
with tf.variable_scope("res3"):
o_res3 = residual(o_res2, 32, 2)
with tf.variable_scope("res4"):
o_res4 = residual(o_res3, 32, 1)
with tf.variable_scope("res5"):
o_res5 = residual(o_res4, 32, 1)
with tf.variable_scope("res6"):
o_res6 = residual(o_res5, 64, 2)
with tf.variable_scope("res7"):
o_res7 = residual(o_res6, 64, 1)
with tf.variable_scope("res8"):
o_res8 = residual(o_res7, 64, 1)
with tf.variable_scope("fc"):
fc = layer(o_res8, 1, [1,1])
with tf.variable_scope("flat"):
return flatten_multi(fc)