def build_graph(self, img_pl, train_stage, data_format="NCHW"):
# some graph info depending on what I will do with it
summary = self.TRAIN['summary']
train_lyr = self.NET['train_lyr']
n_k_lyr = self.NET['n_k_lyr']
n_b_lyr = self.NET['n_b_lyr']
if len(train_lyr) != 10:
print("Wrong length in train list for network. Exiting...")
quit()
self.num_classes = len(self.DATA['label_map'])
# build the graph
print("Building graph")
with tf.variable_scope('images'):
# resize input to desired size
img_resized = tf.image.resize_images(img_pl, [
self.DATA["img_prop"]["height"], self.DATA["img_prop"]["width"]
])
# if on GPU. transpose to NCHW
if data_format == "NCHW":
# convert from NHWC to NCHW (faster on GPU)
img_transposed = tf.transpose(img_resized, [0, 3, 1, 2])
else:
img_transposed = img_resized
# normalization of input
n_img = (img_transposed - 128) / 128
with tf.variable_scope("encoder"):
print("encoder")
with tf.variable_scope("downsample1"):
print("downsample1")
# input image 1024*512 - 960*720
down_lyr1 = lyr.uERF_downsample(n_img,
n_k_lyr[0],
3,
train_stage and train_lyr[0],
summary,
data_format=data_format)
with tf.variable_scope("incept-1"):
incept_1_lyr1 = lyr.dense_inception(
down_lyr1,
n_b_lyr[0],
train_stage and train_lyr[0],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
downsample1 = incept_1_lyr1
with tf.variable_scope("downsample2"):
print("downsample2")
# input image 512*256 - 480*360
down_lyr2 = lyr.uERF_downsample(downsample1,
n_k_lyr[1],
3,
train_stage and train_lyr[1],
summary,
data_format=data_format)
with tf.variable_scope("incept-1"):
print("incept-1")
incept_1_lyr2 = lyr.dense_inception(
down_lyr2,
n_b_lyr[1],
train_stage and train_lyr[1],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
downsample2 = incept_1_lyr2
with tf.variable_scope("downsample3"):
print("downsample3")
# input image 256*128 - 240*180
down_lyr3 = lyr.uERF_downsample(downsample2,
n_k_lyr[2],
3,
train_stage and train_lyr[2],
summary,
data_format=data_format)
downsample3 = down_lyr3
with tf.variable_scope("godeep"):
print("godeep")
# input image 64*32 - 60*45
with tf.variable_scope("dense-inception-1"):
print("dense-inception-1")
godeep_ly1 = lyr.dense_inception(
downsample3,
n_b_lyr[2],
train_stage and train_lyr[3],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("dense-inception-2"):
print("dense-inception-2")
godeep_ly2 = lyr.dense_inception(
godeep_ly1,
n_b_lyr[2],
train_stage and train_lyr[3],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("dense-inception-3"):
print("dense-inception-3")
godeep_ly3 = lyr.dense_inception(
godeep_ly2,
n_b_lyr[2],
train_stage and train_lyr[3],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("dense-inception-4"):
print("dense-inception-4")
godeep_ly4 = lyr.dense_inception(
godeep_ly3,
n_b_lyr[2],
train_stage and train_lyr[3],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
godeep = godeep_ly4
code = godeep
# end encoder, start decoder
print("============= End of encoder ===============")
print("size of code: ", code.get_shape().as_list())
print("=========== Beginning of decoder============")
with tf.variable_scope("decoder"):
print("decoder")
with tf.variable_scope("upsample"):
print("upsample")
with tf.variable_scope("unpool1"):
print("unpool1")
# input image 64*32 - 60*45
unpool_lyr1 = lyr.upsample_layer(code,
train_stage
and train_lyr[5],
kernels=n_k_lyr[3],
data_format=data_format)
with tf.variable_scope("incept-1"):
un_incept_1_lyr1 = lyr.dense_inception(
unpool_lyr1,
n_b_lyr[3],
train_stage and train_lyr[5],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
unpool1 = un_incept_1_lyr1
with tf.variable_scope("unpool2"):
print("unpool2")
# input image 128*64 - 120*90
unpool_lyr2 = lyr.upsample_layer(unpool1,
train_stage
and train_lyr[6],
kernels=n_k_lyr[4],
data_format=data_format)
with tf.variable_scope("incept-1"):
un_incept_1_lyr2 = lyr.dense_inception(
unpool_lyr2,
n_b_lyr[4],
train_stage and train_lyr[6],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
unpool2 = un_incept_1_lyr2
with tf.variable_scope("unpool3"):
print("unpool3")
# input image 256*128 - 240*180
unpool_lyr3 = lyr.upsample_layer(unpool2,
train_stage
and train_lyr[7],
kernels=n_k_lyr[5],
data_format=data_format)
with tf.variable_scope("incept-1"):
un_incept_1_lyr3 = lyr.dense_inception(
unpool_lyr3,
n_b_lyr[5],
train_stage and train_lyr[7],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
unpool3 = un_incept_1_lyr3
unpooled = unpool3
with tf.variable_scope("logits"):
# convert to logits with a linear layer
logits_linear = lyr.linear_layer(unpooled,
self.num_classes,
train_stage and train_lyr[9],
summary=summary,
data_format=data_format)
# transpose logits back to NHWC
if data_format == "NCHW":
logits = tf.transpose(logits_linear, [0, 2, 3, 1])
else:
logits = logits_linear
return logits, code, n_img
def build_graph(self, img_pl, train_stage, data_format="NCHW"):
# some graph info depending on what I will do with it
summary = self.TRAIN['summary']
train_lyr = self.NET['train_lyr']
n_k_lyr = self.NET['n_k_lyr']
n_lyr = self.NET['n_lyr']
if len(train_lyr) != 7:
print("Wrong length in train list for network. Exiting...")
quit()
self.num_classes = len(self.DATA['label_map'])
# build the graph
print("Building graph")
with tf.variable_scope('images'):
# resize input to desired size
img_resized = tf.image.resize_images(img_pl, [
self.DATA["img_prop"]["height"], self.DATA["img_prop"]["width"]
])
# if on GPU. transpose to NCHW
if data_format == "NCHW":
# convert from NHWC to NCHW (faster on GPU)
img_transposed = tf.transpose(img_resized, [0, 3, 1, 2])
else:
img_transposed = img_resized
# normalization of input
n_img = (img_transposed - 128) / 128
with tf.variable_scope("encoder"):
print("encoder")
with tf.variable_scope("downsample1"):
print("downsample1")
# input image 1024*512 - 960*720
down_lyr1 = lyr.uERF_downsample(n_img,
n_k_lyr[0],
3,
train_stage and train_lyr[0],
summary,
data_format=data_format)
inv_residual_lyr1 = [down_lyr1]
for n in range(n_lyr[0]):
with tf.variable_scope("inv-res-" + str(n)):
inv_residual_lyr1.append(
lyr.inv_residual(inv_residual_lyr1[-1],
n_k_lyr[1],
train_stage and train_lyr[0],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"]))
downsample1 = inv_residual_lyr1[-1]
with tf.variable_scope("downsample2"):
print("downsample2")
# input image 512*256 - 480*360
down_lyr2 = lyr.uERF_downsample(downsample1,
n_k_lyr[2],
3,
train_stage and train_lyr[1],
summary,
data_format=data_format)
inv_residual_lyr2 = [down_lyr2]
for n in range(n_lyr[1]):
with tf.variable_scope("inv-res-" + str(n)):
inv_residual_lyr2.append(
lyr.inv_residual(inv_residual_lyr2[-1],
n_k_lyr[3],
train_stage and train_lyr[1],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"]))
downsample2 = inv_residual_lyr2[-1]
with tf.variable_scope("downsample3"):
print("downsample3")
# input image 512*256 - 480*360
down_lyr3 = lyr.uERF_downsample(downsample2,
n_k_lyr[4],
3,
train_stage and train_lyr[2],
summary,
data_format=data_format)
inv_residual_lyr3 = [down_lyr3]
for n in range(n_lyr[2]):
with tf.variable_scope("inv-res-" + str(n)):
inv_residual_lyr3.append(
lyr.inv_residual(inv_residual_lyr3[-1],
n_k_lyr[5],
train_stage and train_lyr[2],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"]))
downsample3 = inv_residual_lyr3[-1]
code = downsample3
# end encoder, start decoder
print("============= End of encoder ===============")
print("size of code: ", code.get_shape().as_list())
print("=========== Beginning of decoder============")
with tf.variable_scope("decoder"):
print("decoder")
with tf.variable_scope("upsample"):
print("upsample")
with tf.variable_scope("unpool1"):
print("unpool1")
unpool_lyr1 = lyr.upsample_layer(
code,
train_stage and train_lyr[3],
kernels=n_k_lyr[6],
data_format=data_format) + downsample2
inv_residual_unpool_1 = [unpool_lyr1]
for n in range(n_lyr[4]):
with tf.variable_scope("inv-res-" + str(n)):
inv_residual_unpool_1.append(
lyr.inv_residual(
inv_residual_unpool_1[-1],
n_k_lyr[7],
train_stage and train_lyr[3],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"]))
unpool1 = inv_residual_unpool_1[-1]
with tf.variable_scope("unpool2"):
print("unpool2")
unpool_lyr2 = lyr.upsample_layer(
unpool1,
train_stage and train_lyr[4],
kernels=n_k_lyr[8],
data_format=data_format) + downsample1
inv_residual_unpool_2 = [unpool_lyr2]
for n in range(n_lyr[4]):
with tf.variable_scope("inv-res-" + str(n)):
inv_residual_unpool_2.append(
lyr.inv_residual(
inv_residual_unpool_2[-1],
n_k_lyr[9],
train_stage and train_lyr[4],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"]))
unpool2 = inv_residual_unpool_2[-1]
with tf.variable_scope("unpool3"):
print("unpool3")
# input image 64*32 - 60*45
unpool_lyr3 = lyr.upsample_layer(
unpool2 + inv_residual_lyr1[-1],
train_stage and train_lyr[5],
kernels=n_k_lyr[10],
data_format=data_format)
inv_residual_unpool_3 = [unpool_lyr3]
for n in range(n_lyr[4]):
with tf.variable_scope("inv-res-" + str(n)):
inv_residual_unpool_3.append(
lyr.inv_residual(
inv_residual_unpool_3[-1],
n_k_lyr[11],
train_stage and train_lyr[5],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"]))
unpool3 = inv_residual_unpool_3[-1]
unpooled = unpool3
with tf.variable_scope("logits"):
# convert to logits with a linear layer
logits_linear = lyr.linear_layer(unpooled,
self.num_classes,
train_stage and train_lyr[6],
summary=summary,
data_format=data_format)
# transpose logits back to NHWC
if data_format == "NCHW":
logits = tf.transpose(logits_linear, [0, 2, 3, 1])
else:
logits = logits_linear
return logits, code, n_img
def build_graph(self, img_pl, train_stage, data_format="NCHW"):
# some graph info depending on what I will do with it
summary = self.TRAIN['summary']
train_lyr = self.NET['train_lyr']
n_k_lyr = self.NET['n_k_lyr']
if len(train_lyr) != 25:
print("Wrong length in train list for network. Exiting...")
quit()
self.num_classes = len(self.DATA['label_map'])
# build the graph
print("Building graph")
with tf.variable_scope('images'):
# resize input to desired size
img_resized = tf.image.resize_images(img_pl, [
self.DATA["img_prop"]["height"], self.DATA["img_prop"]["width"]
])
# if on GPU. transpose to NCHW
if data_format == "NCHW":
# convert from NHWC to NCHW (faster on GPU)
img_transposed = tf.transpose(img_resized, [0, 3, 1, 2])
else:
img_transposed = img_resized
# normalization of input
n_img = (img_transposed - 128) / 128
with tf.variable_scope("encoder"):
print("encoder")
with tf.variable_scope("Block1"):
print("Block 1")
block1_lyr1 = lyr.uERF_downsample(n_img,
n_k_lyr[0],
3,
train_stage and train_lyr[0],
summary,
data_format=data_format)
block_1 = block1_lyr1
with tf.variable_scope("Block2"):
print("Block 2")
block2_lyr1 = lyr.uERF_downsample(block_1,
n_k_lyr[1],
3,
train_stage and train_lyr[1],
summary,
data_format=data_format)
with tf.variable_scope("non-bt-1"):
block2_lyr2 = lyr.uERF_non_bt(
block2_lyr1,
3,
train_stage and train_lyr[2],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-2"):
block2_lyr3 = lyr.uERF_non_bt(
block2_lyr2,
3,
train_stage and train_lyr[3],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-3"):
block2_lyr4 = lyr.uERF_non_bt(
block2_lyr3,
3,
train_stage and train_lyr[4],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-4"):
block2_lyr5 = lyr.uERF_non_bt(
block2_lyr4,
3,
train_stage and train_lyr[5],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
block_2 = block2_lyr5
with tf.variable_scope("Block3"):
print("Block 3")
block3_lyr1 = lyr.uERF_downsample(block_2,
n_k_lyr[2],
3,
train_stage and train_lyr[6],
summary,
data_format=data_format)
with tf.variable_scope("non-bt-1"):
block3_lyr2 = lyr.uERF_non_bt(
block3_lyr1,
3,
train_stage and train_lyr[7],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-2"):
block3_lyr3 = lyr.uERF_non_bt(
block3_lyr2,
5,
train_stage and train_lyr[8],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-3"):
block3_lyr4 = lyr.uERF_non_bt(
block3_lyr3,
7,
train_stage and train_lyr[9],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-4"):
block3_lyr5 = lyr.uERF_non_bt(
block3_lyr4,
9,
train_stage and train_lyr[10],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-5"):
block3_lyr6 = lyr.uERF_non_bt(
block3_lyr5,
3,
train_stage and train_lyr[11],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-6"):
block3_lyr7 = lyr.uERF_non_bt(
block3_lyr6,
5,
train_stage and train_lyr[12],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-7"):
block3_lyr8 = lyr.uERF_non_bt(
block3_lyr7,
7,
train_stage and train_lyr[13],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-8"):
block3_lyr9 = lyr.uERF_non_bt(
block3_lyr8,
9,
train_stage and train_lyr[14],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
code = block_3 = block3_lyr9
# end encoder, start decoder
print("============= End of encoder ===============")
print("size of code: ", code.get_shape().as_list())
print("=========== Beginning of decoder============")
with tf.variable_scope("decoder"):
print("decoder")
with tf.variable_scope("Block4"):
print("Block 4")
with tf.variable_scope("unpool"):
block4_unpool = lyr.upsample_layer(code,
train_stage
and train_lyr[15],
kernels=n_k_lyr[3],
data_format=data_format)
with tf.variable_scope("non-bt-1"):
block4_lyr1 = lyr.uERF_non_bt(
block4_unpool,
3,
train_stage and train_lyr[16],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-2"):
block4_lyr2 = lyr.uERF_non_bt(
block4_lyr1,
3,
train_stage and train_lyr[17],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
block_4 = block4_lyr2
with tf.variable_scope("Block5"):
print("Block 5")
with tf.variable_scope("unpool"):
block5_unpool = lyr.upsample_layer(block_4,
train_stage
and train_lyr[18],
kernels=n_k_lyr[4],
data_format=data_format)
with tf.variable_scope("non-bt-1"):
block5_lyr1 = lyr.uERF_non_bt(
block5_unpool,
3,
train_stage and train_lyr[19],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
with tf.variable_scope("non-bt-2"):
block5_lyr2 = lyr.uERF_non_bt(
block5_lyr1,
3,
train_stage and train_lyr[20],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
block_5 = block5_lyr2
with tf.variable_scope("Block6"):
print("Block 6")
with tf.variable_scope("unpool"):
block6_unpool = lyr.upsample_layer(block_5,
train_stage
and train_lyr[21],
kernels=n_k_lyr[5],
data_format=data_format)
with tf.variable_scope("non-bt-1"):
block6_lyr1 = lyr.uERF_non_bt(
block6_unpool,
3,
train_stage and train_lyr[22],
summary,
data_format=data_format,
dropout=self.NET["dropout"],
bn_decay=self.NET["bn_decay"])
block_6 = block6_lyr1
with tf.variable_scope("logits"):
# convert to logits with a linear layer
logits_linear = lyr.linear_layer(block_6,
self.num_classes,
train_stage and train_lyr[24],
summary=summary,
data_format=data_format)
# transpose logits back to NHWC
if data_format == "NCHW":
logits = tf.transpose(logits_linear, [0, 2, 3, 1])
else:
logits = logits_linear
return logits, code, n_img