def eff_b0_01():
# 224
base = EfficientNetB0(
include_top=False,
input_shape=(config.img_size, config.img_size, config.channel),
weights=None
)
# Using BN to normalize images
# base.layers.pop(0)
# inp = Input(shape=(config.img_size, config.img_size, config.channel))
# x = BatchNormalization()(inp)
# base_out = base(inp)
x = GlobalAveragePooling2D()(base.output)
x = Dense(1280)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Dense(128)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)
x = Dense(config.n_labels, activation='softmax')(x)
model = Model(inputs=base.input, outputs=x)
return model
def getTrainModel(self):
input_shape = (self.cut_size["height"], self.cut_size["width"],
self.channel)
print("input_shape : {}".format(input_shape))
# EfficientNetのオプションについては随時調査（やりたいのは１から学習）
effnet_instance = EfficientNetB0(input_shape=input_shape,
weights='imagenet',
include_top=False)
x = effnet_instance.output
x = self.bottleneck(x)
model = Model(inputs=effnet_instance.input, outputs=x)
return model
def get_fe(fe, input_image):
if fe == 'effnetb0':
return EfficientNetB0(input_tensor=input_image, include_top=False), [
'swish_last', 'block5_i_MB_swish_1', 'block3_i_MB_swish_1'
]
elif fe == 'effnetb1':
return EfficientNetB1(input_tensor=input_image, include_top=False), [
'swish_last', 'block5_i_MB_swish_1', 'block3_i_MB_swish_1'
]
elif fe == 'effnetb2':
return EfficientNetB2(input_tensor=input_image, include_top=False), [
'swish_last', 'block5_i_MB_swish_1', 'block3_i_MB_swish_1'
]
elif fe == 'effnetb3':
return EfficientNetB3(input_tensor=input_image, include_top=False), [
'swish_last', 'block5_i_MB_swish_1', 'block3_i_MB_swish_1'
]
elif fe == 'effnetb4':
return EfficientNetB4(input_tensor=input_image, include_top=False), [
'swish_last', 'block5_i_MB_swish_1', 'block3_i_MB_swish_1'
]
elif fe == 'effnetb5':
return EfficientNetB5(input_tensor=input_image, include_top=False), [
'swish_last', 'block5_i_MB_swish_1', 'block3_i_MB_swish_1'
]
elif fe == 'd53':
return d53(input_image)
elif fe == 'mnetv2':
mnet = MobileNetV2(input_tensor=input_image, weights='imagenet')
return mnet, [
'out_relu', 'block_13_expand_relu', 'block_6_expand_relu'
]
elif fe == 'mnet':
mnet = MobileNet(input_tensor=input_image, weights='imagenet')
return mnet, ['conv_pw_13_relu', 'conv_pw_11_relu', 'conv_pw_5_relu']
elif fe == 'r50':
r50 = ResNet50(input_tensor=input_image, weights='imagenet')
return r50, ['activation_49', 'activation_40', 'activation_22']
raise ValueError('Pls put the correct fe')
def get_model(model_name="vgg16"):
tf.keras.backend.clear_session()
if (model_name == "vgg16"):
from tensorflow.keras.applications.vgg16 import preprocess_input
model = VGG16(weights='imagenet', include_top=False)
return (model, preprocess_input)
elif (model_name == "vgg19"):
return (tf.keras.applications.vgg19.VGG19(weights='imagenet',
include_top=False),
tf.keras.applications.vgg16.preprocess_input)
elif (model_name == "resnet50"):
return (tf.keras.applications.resnet50.ResNet50(weights="imagenet",
include_top=False),
tf.keras.applications.resnet50.preprocess_input)
elif (model_name == "mobilenet"):
return (tf.keras.applications.mobilenet.MobileNet(weights="imagenet",
include_top=False),
tf.keras.applications.mobilenet.preprocess_input)
# elif (model_name == "mobilenetv2"):
# return (tf.keras.applications.mobilenet_v2.MobileNetV2(weights="imagenet", include_top=False), tf.keras.applications.mobilenet_v2.preprocess_input)
elif (model_name == "xception"):
return (tf.keras.applications.xception.Xception(weights="imagenet",
include_top=False),
tf.keras.applications.xception.preprocess_input)
elif (model_name == "densenet121"):
return (tf.keras.applications.densenet.DenseNet121(weights="imagenet",
include_top=False),
tf.keras.applications.densenet.preprocess_input)
elif (model_name == "inceptionv3"):
return (tf.keras.applications.inception_v3.InceptionV3(
weights="imagenet", include_top=False),
tf.keras.applications.inception_v3.preprocess_input)
elif (model_name == "efficientnetb0"):
return (EfficientNetB0(weights='imagenet',
include_top=False), effpreprocess)
elif (model_name == "efficientnetb5"):
return (EfficientNetB5(weights='imagenet',
include_top=False), effpreprocess)
def get_effnet_model(save_path,
model_res=1024,
image_size=256,
depth=1,
size=3,
activation='elu',
loss='logcosh',
optimizer='adam'):
if os.path.exists(save_path):
print('Loading model')
return load_model(save_path)
# Build model
print('Building model')
model_scale = int(2 *
(math.log(model_res, 2) - 1)) # For example, 1024 -> 18
if (size <= 0):
effnet = EfficientNetB0(include_top=False,
weights='imagenet',
input_shape=(image_size, image_size, 3))
if (size == 1):
effnet = EfficientNetB1(include_top=False,
weights='imagenet',
input_shape=(image_size, image_size, 3))
if (size == 2):
effnet = EfficientNetB2(include_top=False,
weights='imagenet',
input_shape=(image_size, image_size, 3))
if (size >= 3):
effnet = EfficientNetB3(include_top=False,
weights='imagenet',
input_shape=(image_size, image_size, 3))
layer_size = model_scale * 8 * 8 * 8
if is_square(layer_size): # work out layer dimensions
layer_l = int(math.sqrt(layer_size) + 0.5)
layer_r = layer_l
else:
layer_m = math.log(math.sqrt(layer_size), 2)
layer_l = 2**math.ceil(layer_m)
layer_r = layer_size // layer_l
layer_l = int(layer_l)
layer_r = int(layer_r)
x_init = None
inp = Input(shape=(image_size, image_size, 3))
x = effnet(inp)
if (size < 1):
x = Conv2D(model_scale * 8, 1, activation=activation)(x) # scale down
if (depth > 0):
x = Reshape((layer_r, layer_l))(
x
) # See https://github.com/OliverRichter/TreeConnect/blob/master/cifar.py - TreeConnect inspired layers instead of dense layers.
else:
if (depth < 1):
depth = 1
if (size <= 2):
x = Conv2D(model_scale * 8 * 4, 1,
activation=activation)(x) # scale down a bit
x = Reshape((layer_r * 2, layer_l * 2))(
x
) # See https://github.com/OliverRichter/TreeConnect/blob/master/cifar.py - TreeConnect inspired layers instead of dense layers.
else:
x = Reshape((384, 256))(x) # full size for B3
while (depth > 0):
x = LocallyConnected1D(layer_r, 1, activation=activation)(x)
x = Permute((2, 1))(x)
x = LocallyConnected1D(layer_l, 1, activation=activation)(x)
x = Permute((2, 1))(x)
if x_init is not None:
x = Add()([x, x_init]) # add skip connection
x_init = x
depth -= 1
if (
size >= 2
): # add unshared layers at end for different sections of the latent space
x_init = x
if layer_r % 3 == 0 and layer_l % 3 == 0:
a = LocallyConnected1D(layer_r, 1, activation=activation)(x)
b = LocallyConnected1D(layer_r, 1, activation=activation)(x)
c = LocallyConnected1D(layer_r, 1, activation=activation)(x)
a = Permute((2, 1))(a)
b = Permute((2, 1))(b)
c = Permute((2, 1))(c)
a = LocallyConnected1D(layer_l // 3, 1, activation=activation)(a)
b = LocallyConnected1D(layer_l // 3, 1, activation=activation)(b)
c = LocallyConnected1D(layer_l // 3, 1, activation=activation)(c)
x = Concatenate()([a, b, c])
else:
a = LocallyConnected1D(layer_r // 2, 1, activation=activation)(x)
b = LocallyConnected1D(layer_r // 2, 1, activation=activation)(x)
c = LocallyConnected1D(layer_r // 2, 1, activation=activation)(x)
d = LocallyConnected1D(layer_r // 2, 1, activation=activation)(x)
a = Permute((2, 1))(a)
b = Permute((2, 1))(b)
c = Permute((2, 1))(c)
d = Permute((2, 1))(d)
a = LocallyConnected1D(layer_l // 2, 1, activation=activation)(a)
b = LocallyConnected1D(layer_l // 2, 1, activation=activation)(b)
c = LocallyConnected1D(layer_l // 2, 1, activation=activation)(c)
d = LocallyConnected1D(layer_l // 2, 1, activation=activation)(d)
x = Concatenate()([a, b, c, d])
x = Add()([x, x_init]) # add skip connection
x = Reshape((model_scale, 512))(x) # train against all dlatent values
model = Model(inputs=inp, outputs=x)
model.compile(loss=loss, metrics=[], optimizer=optimizer
) # By default: adam optimizer, logcosh used for loss.
return model
def _create_model():
print('Creating model')
# load the pretrained model, without the classification (top) layers
if self.transfer_model == 'Xception':
base_model = tf.keras.applications.Xception(
weights='imagenet',
include_top=False,
input_shape=(*self.target_size, 3))
based_model_last_block = 116 # last block 126, two blocks 116
elif self.transfer_model == 'Inception_Resnet':
base_model = tf.keras.applications.InceptionResNetV2(
weights='imagenet',
include_top=False,
input_shape=(*self.target_size, 3))
based_model_last_block = 287 # last block 630, two blocks 287
elif self.transfer_model == 'Resnet':
base_model = tf.keras.applications.ResNet50(
weights='imagenet',
include_top=False,
input_shape=(*self.target_size, 3))
based_model_last_block = 155 # last block 165, two blocks 155
elif self.transfer_model == 'B0':
base_model = EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=(*self.target_size, 3))
based_model_last_block = 213 # last block 229, two blocks 213
elif self.transfer_model == 'B3':
base_model = EfficientNetB3(weights='imagenet',
include_top=False,
input_shape=(*self.target_size, 3))
based_model_last_block = 354 # last block 370, two blocks 354
elif self.transfer_model == 'B5':
base_model = EfficientNetB5(weights='imagenet',
include_top=False,
input_shape=(*self.target_size, 3))
based_model_last_block = 417 # last block 559, two blocks 417
else:
base_model = tf.keras.applications.InceptionV3(
weights='imagenet',
include_top=False,
input_shape=(*self.target_size, 3))
based_model_last_block = 249 # last block 280, two blocks 249
# Set only the top layers as trainable (if we want to do fine-tuning,
# we can train the base layers as a second step)
base_model.trainable = False
# Target size infered from the base model
self.target_size = base_model.input_shape[1:3]
# Add the classification layers using Keras functional API
x = base_model.output
x = tf.keras.layers.GlobalAveragePooling2D()(x)
# Hidden layer for classification
if hidden_size == 0:
x = tf.keras.layers.Dropout(rate=dropout)(x)
elif bn_after_ac:
x = tf.keras.layers.Dense(
hidden_size,
activation=activation,
kernel_regularizer=tf.keras.regularizers.l2(
l=l2_lambda))(x)
x = tf.keras.layers.BatchNormalization()(x)
x = tf.keras.layers.Dropout(rate=dropout)(x)
else:
x = tf.keras.layers.Dense(
hidden_size,
use_bias=False,
kernel_regularizer=tf.keras.regularizers.l2(
l=l2_lambda))(x)
# scale: When the next layer is linear (also e.g. nn.relu), this can be disabled since the scaling can be done by the next layer.
x = tf.keras.layers.BatchNormalization(
scale=activation != 'relu')(x)
x = tf.keras.layers.Activation(activation=activation)(x)
x = tf.keras.layers.Dropout(rate=dropout)(x)
predictions = tf.keras.layers.Dense(len(self.categories),
activation='softmax')(
x) # Output layer
# Define the optimizer and the loss and the optimizer
loss = 'sparse_categorical_crossentropy'
metrics = ['sparse_categorical_accuracy']
optimizer = tf.keras.optimizers.Adam(lr=learning_rate)
return tf.keras.Model(
inputs=base_model.input, outputs=predictions
), base_model, based_model_last_block, loss, metrics, optimizer
def build_model(self):
# tf.summary.image('input_image', image_input, 10)
# tf.summary.image('mask', mask, 10)
if (self.is_pb == False):
is_training_seg = tf.placeholder(tf.bool, name='is_training_seg')
is_training_dec = tf.placeholder(tf.bool, name='is_training_dec')
if IMAGE_MODE == 0:
image_input = tf.placeholder(tf.float32,
shape=(self.__batch_size,
IMAGE_SIZE[0],
IMAGE_SIZE[1], 1),
name='Image')
num_ch = 1
else:
image_input = tf.placeholder(tf.float32,
shape=(self.__batch_size,
IMAGE_SIZE[0],
IMAGE_SIZE[1], 3),
name='Image')
num_ch = 3
label = tf.placeholder(tf.float32,
shape=(self.__batch_size, CLASS_NUM),
name='Label')
mask = tf.placeholder(tf.float32,
shape=(self.__batch_size, IMAGE_SIZE[0],
IMAGE_SIZE[1], CLASS_NUM),
name='mask')
else:
is_training_seg = False
is_training_dec = False
if IMAGE_MODE == 0:
image_input = tf.placeholder(
tf.float32,
shape=(self.__batch_size_inference, IMAGE_SIZE[0],
IMAGE_SIZE[1], 1),
name='Image')
num_ch = 1
else:
image_input = tf.placeholder(
tf.float32,
shape=(self.__batch_size_inference, IMAGE_SIZE[0],
IMAGE_SIZE[1], 3),
name='Image')
num_ch = 3
label = tf.placeholder(tf.float32,
shape=(self.__batch_size_inference,
CLASS_NUM),
name='Label')
mask = tf.placeholder(tf.float32,
shape=(self.__batch_size_inference,
IMAGE_SIZE[0], IMAGE_SIZE[1],
CLASS_NUM),
name='mask')
# create backbone
if self.backbone == 'mixnet':
from MixNet import MixNetSmall
self.__checkPoint_dir = 'checkpoint/mixnet'
backbone_output = MixNetSmall(
image_input,
scope='segmentation',
include_top=False,
keep_dropout_backbone=self.keep_dropout_backbone,
training=is_training_seg)
elif self.backbone == 'mixnet_official':
from MixNet_official import build_model_base
self.__checkPoint_dir = 'checkpoint/mixnet_official'
backbone_output = build_model_base(image_input,
'mixnet-s',
training=is_training_seg,
override_params=None,
scope='segmentation')
elif self.backbone == 'efficientnet':
from efficientnet import EfficientNetB0
self.__checkPoint_dir = 'checkpoint/efficientnet'
backbone_output = EfficientNetB0(
image_input,
model_name='segmentation',
input_shape=(None, IMAGE_SIZE[0], IMAGE_SIZE[1], num_ch),
keep_dropout_backbone=self.keep_dropout_backbone,
training=is_training_seg,
include_top=True,
classes=CLASS_NUM)
elif self.backbone == 'fast_scnn':
from fast_scnn import build_fast_scnn
self.__checkPoint_dir = 'checkpoint/fast_scnn'
backbone_output = build_fast_scnn(
image_input,
'segmentation',
is_training=is_training_seg,
keep_dropout_backbone=self.keep_dropout_backbone)
elif self.backbone == 'ghostnet':
from GhostNet import ghostnet_base
self.__checkPoint_dir = 'checkpoint/ghostnet'
# Set depth_multiplier to change the depth of GhostNet
backbone_output = ghostnet_base(image_input,
mode=self.__mode,
data_format=DATA_FORMAT,
scope='segmentation',
dw_code=None,
ratio_code=None,
se=1,
min_depth=8,
depth=1,
depth_multiplier=0.5,
conv_defs=None,
is_training=is_training_seg,
momentum=self.__bn_momentum)
elif self.backbone == 'sinet':
from SiNet import SINet
self.__checkPoint_dir = 'checkpoint/sinet'
backbone_output = SINet(image_input,
classes=CLASS_NUM,
p=2,
q=8,
chnn=1,
training=True,
bn_momentum=0.99,
scope='segmentation',
reuse=False)
elif self.backbone == 'lednet':
from LEDNet import lednet
self.__checkPoint_dir = 'checkpoint/lednet'
backbone_output = lednet(image_input,
training=is_training_seg,
scope='segmentation',
keep_dropout=False)
elif self.backbone == 'cspnet':
from CSPDenseNet import CSPPeleeNet
self.__checkPoint_dir = 'checkpoint/cspnet'
backbone_output = CSPPeleeNet(image_input,
data_format=DATA_FORMAT,
drop_rate=0.0,
training=self.is_training,
momentum=self.__bn_momentum,
name="segmentation",
mode=self.__mode,
activation=ACTIVATION)
elif self.backbone == 'FCNN':
self.__checkPoint_dir = 'checkpoint/fcnn'
from FCNN import conv_block, LearningToDownsample, GlobalFeatureExtractor, FeatureFusion, Classifier
def SegmentNet(input, scope, is_training, reuse=None):
with tf.variable_scope(scope, reuse=None):
lds2, lds1 = LearningToDownsample(input,
is_training,
scope='lds',
reuse=reuse)
gfe = GlobalFeatureExtractor(lds2,
is_training,
scope='gfe',
reuse=reuse)
ff = FeatureFusion(lds1,
lds2,
gfe,
is_training,
scope='ff',
reuse=reuse)
features, logits = Classifier(ff,
is_training,
scope='classify',
reuse=reuse)
mask = tf.nn.sigmoid(logits, name='softmax1')
return features, logits, mask
features, logits_pixel, mask = SegmentNet(image_input,
'segmentation',
self.is_training)
if self.is_pb == True:
self.mask_out_l = tf.nn.sigmoid(logits_pixel[0],
name='mask_out1')
self.mask_out_r = tf.nn.sigmoid(logits_pixel[1],
name='mask_out2')
if self.is_pb == False:
# Variable list
segmentation_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=logits_pixel, labels=mask))
train_segment_var_list = [
v for v in tf.trainable_variables()
if 'segmentation' in v.name
]
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
update_ops_segment = [
v for v in update_ops if 'segmentation' in v.name
]
optimizer_segmentation = tf.train.AdamOptimizer(
self.__learn_rate)
with tf.control_dependencies(update_ops_segment):
optimize_segment = optimizer_segmentation.minimize(
segmentation_loss, var_list=train_segment_var_list)
self.segmentation_loss = segmentation_loss
self.optimize_segment = optimize_segment
mask_out = tf.nn.sigmoid(logits_pixel, name='mask_out')
self.mask_out = mask_out
init_op = tf.global_variables_initializer()
self.Image = image_input
self.is_training_seg = is_training_seg
self.is_training_dec = is_training_dec
self.mask = mask
self.label = label
self.init_op = init_op
return
else:
raise ValueError("Unknown Backbone")
# create segmentation head and segmentation loss
if len(backbone_output) == 5:
if self.neck == 'bifpn':
from FPN import bifpn_neck
P3_out, P4_out, P5_out, P6_out, P7_out = bifpn_neck(
backbone_output,
64,
scope='segmentation',
is_training=is_training_seg,
momentum=self.__bn_momentum,
mode=self.__mode,
data_format=DATA_FORMAT)
with tf.variable_scope('segmentation'):
P3 = tf.layers.conv2d(P3_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P3',
data_format=DATA_FORMAT)
P4 = tf.layers.conv2d(P4_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P4',
data_format=DATA_FORMAT)
P5 = tf.layers.conv2d(P5_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P5',
data_format=DATA_FORMAT)
P6 = tf.layers.conv2d(P6_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P6',
data_format=DATA_FORMAT)
P7 = tf.layers.conv2d(P7_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P7',
data_format=DATA_FORMAT)
neck_output = [P3, P4, P5, P6, P7]
decision_in = [P3_out, P3]
elif self.neck == 'fpn':
from FPN import fpn_neck
neck_output = fpn_neck(
backbone_output,
CLASS_NUM,
drop_rate=0.2,
keep_dropout_head=self.keep_dropout_head,
scope='segmentation',
training=is_training_seg)
elif self.neck == 'bfp':
from BFP import bfp_segmentation_head
P3_out, P4_out, P5_out, P6_out, P7_out = bfp_segmentation_head(
backbone_output,
64,
scope='segmentation',
is_training=is_training_seg)
with tf.variable_scope('segmentation'):
P3 = tf.layers.conv2d(P3_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P3')
P4 = tf.layers.conv2d(P4_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P4')
P5 = tf.layers.conv2d(P5_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P5')
P6 = tf.layers.conv2d(P6_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P6')
P7 = tf.layers.conv2d(P7_out,
CLASS_NUM, (1, 1), (1, 1),
use_bias=False,
name='P7')
neck_output = [P3, P4, P5, P6, P7]
decision_in = [P3_out, P3]
elif self.neck == 'pan':
from FPN import PAN
with tf.variable_scope('segmentation'):
backbone_output = PAN(backbone_output,
128,
training=self.is_training_seg)
seg_fea = backbone_output[-1]
seg_fea = tf.keras.layers.UpSampling2D((4, 4))(seg_fea)
seg_fea = tf.keras.layers.DepthwiseConv2D(
(3, 3), (1, 1), padding='same')(seg_fea)
seg_fea = tf.layers.batch_normalization(
seg_fea, training=is_training_seg)
seg_fea = tf.nn.relu(seg_fea)
seg_fea = tf.layers.conv2d(seg_fea, 128, (1, 1))
seg_fea = tf.keras.layers.SeparableConv2D(
128, (3, 3), padding='same')(seg_fea)
seg_fea = tf.layers.batch_normalization(
seg_fea, training=is_training_seg)
seg_fea = tf.nn.relu(seg_fea)
seg_fea = tf.keras.layers.SeparableConv2D(
128, (3, 3), padding='same')(seg_fea)
seg_fea = tf.layers.batch_normalization(
seg_fea, training=is_training_seg)
seg_fea = tf.nn.relu(seg_fea)
seg_fea1 = tf.layers.conv2d(seg_fea, 1, (1, 1))
neck_output = [seg_fea, seg_fea1]
else:
raise ValueError(" Unknown neck ")
if len(neck_output) == 5:
if DATA_FORMAT == 'channels_first':
for nec_index in range(len(neck_output)):
neck_output[nec_index] = tf.transpose(
neck_output[nec_index], [0, 2, 3, 1])
logits_pixel = tf.image.resize_images(
neck_output[0], (IMAGE_SIZE[0], IMAGE_SIZE[1]),
align_corners=True,
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
segmentation_loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=logits_pixel, labels=mask)) + \
tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neck_output[1], labels=tf.image.resize_images(mask, (IMAGE_SIZE[0] // 4, IMAGE_SIZE[1] // 4)))) + \
tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neck_output[2], labels=tf.image.resize_images(mask, (IMAGE_SIZE[0] // 8, IMAGE_SIZE[1] // 8)))) + \
tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neck_output[3], labels=tf.image.resize_images(mask, (IMAGE_SIZE[0] // 16, IMAGE_SIZE[1] // 16)))) + \
tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neck_output[4], labels=tf.image.resize_images(mask, (IMAGE_SIZE[0] // 32, IMAGE_SIZE[1] // 32))))
elif len(neck_output) == 2:
decision_in = neck_output
if DATA_FORMAT == 'channels_first':
for nec_index in range(len(neck_output)):
neck_output[nec_index] = tf.transpose(
neck_output[nec_index], [0, 2, 3, 1])
logits_pixel = tf.image.resize_images(
neck_output[1], (IMAGE_SIZE[0], IMAGE_SIZE[1]),
align_corners=True)
segmentation_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(
logits=logits_pixel, labels=mask))
elif len(backbone_output) == 2:
decision_in = [backbone_output[0], backbone_output[1]]
if DATA_FORMAT == 'channels_first':
for nec_index in range(len(backbone_output)):
backbone_output[nec_index] = tf.transpose(
backbone_output[nec_index], [0, 2, 3, 1])
logits_pixel = tf.image.resize_images(
backbone_output[1], (IMAGE_SIZE[0], IMAGE_SIZE[1]),
align_corners=True)
segmentation_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=logits_pixel,
labels=mask))
else:
raise ValueError(
" Incorrect backbone output number, must be 3 or 5")
mask_out = tf.nn.sigmoid(logits_pixel, name='mask_out')
if self.is_pb == True:
self.mask_out_l = tf.nn.sigmoid(logits_pixel[0], name='mask_out1')
self.mask_out_r = tf.nn.sigmoid(logits_pixel[1], name='mask_out2')
# Create decision head
dec_out = decision_head(decision_in[0],
decision_in[1],
class_num=CLASS_NUM,
scope='decision',
keep_dropout_head=self.keep_dropout_head,
training=is_training_dec,
data_format=DATA_FORMAT,
momentum=self.__bn_momentum,
mode=self.__mode,
activation=ACTIVATION)
decision_out = tf.nn.sigmoid(dec_out, name='decision_out')
decision_loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=dec_out,
labels=label)
decision_loss = tf.reduce_mean(decision_loss)
total_loss = segmentation_loss + decision_loss
if self.is_pb == False:
# Variable list
train_segment_var_list = [
v for v in tf.trainable_variables() if 'segmentation' in v.name
]
train_decision_var_list = [
v for v in tf.trainable_variables() if 'decision' in v.name
]
train_var_list = train_segment_var_list + train_decision_var_list
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
update_ops_segment = [
v for v in update_ops if 'segmentation' in v.name
]
update_ops_decision = [
v for v in update_ops if 'decision' in v.name
]
# optimizer_segmentation = tf.train.GradientDescentOptimizer(self.__learn_rate)
optimizer_segmentation = tf.train.AdamOptimizer(self.__learn_rate)
optimizer_decision = tf.train.AdamOptimizer(self.__learn_rate)
# optimizer_decision = tf.train.GradientDescentOptimizer(self.__learn_rate)
optimizer_total = tf.train.AdamOptimizer(self.__learn_rate)
with tf.control_dependencies(update_ops_segment):
optimize_segment = optimizer_segmentation.minimize(
segmentation_loss, var_list=train_segment_var_list)
with tf.control_dependencies(update_ops_decision):
optimize_decision = optimizer_decision.minimize(
decision_loss, var_list=train_decision_var_list)
with tf.control_dependencies(update_ops):
optimize_total = optimizer_total.minimize(
total_loss, var_list=train_var_list)
self.segmentation_loss = segmentation_loss
self.decision_loss = decision_loss
self.optimize_segment = optimize_segment
self.optimize_decision = optimize_decision
self.optimize_total = optimize_total
if not os.path.exists(self.tensorboard_logdir):
os.makedirs(self.tensorboard_logdir)
# merged = tf.summary.merge_all()
# train_writer = tf.summary.FileWriter(self.tensorboard_logdir, self.session.graph)
init_op = tf.global_variables_initializer()
# self.update_ops_segment = update_ops_segment
# self.train_segment_var_list = [v for v in tf.trainable_variables() if 'segmentation' in v.name]
# self.backbone_output = backbone_output
self.Image = image_input
self.is_training_seg = is_training_seg
self.is_training_dec = is_training_dec
self.mask = mask
self.label = label
self.decison_out = decision_out
self.mask_out = mask_out
self.init_op = init_op
from efficientnet import EfficientNetB0, preprocess_input
from sklearn.metrics import f1_score
from keras.preprocessing.image import ImageDataGenerator
import numpy as np
model = EfficientNetB0(weights=None, classes=80)
model.load_weights('/home/palm/PycharmProjects/ptt/weights/b0.h5')
d = ImageDataGenerator(preprocessing_function=preprocess_input)
val_generator = d.flow_from_directory(
'/media/palm/data/scene_validation_20170908/ai_challenger_scene_validation_20170908/images',
batch_size=32,
target_size=(224, 224),
shuffle=False)
y = model.predict_generator(val_generator, steps=len(val_generator), verbose=1)
y_pred = np.argmax(y, axis=1)
print(f1_score(val_generator.classes, y_pred, average='macro'))
train_generator = train_datagen.flow_from_directory(
# This is the target directory
train_dir,
# All images will be resized to target height and width.
target_size=(height, width),
batch_size=batch_size,
# Since we use categorical_crossentropy loss, we need categorical labels
class_mode='categorical')
validation_generator = validation_datagen.flow_from_directory(
validation_dir,
target_size=(height, width),
batch_size=batch_size,
class_mode='categorical')
conv_base = Net(weights='imagenet', include_top=False, input_shape=input_shape)
# model = add_new_last_layer(conv_base, nb_classes)
model = models.Sequential()
# model.add(layers.Flatten(name="flatten"))
model.add(conv_base)
model.add(layers.GlobalMaxPooling2D(name="gap"))
if dropout_rate > 0:
model.add(layers.Dropout(dropout_rate, name="dropout_out"))
# model.add(layers.Dense(256, activation='relu', name="fc1"))
model.add(layers.Dense(2, activation='softmax', name="fc_out")) # 这里的2代表要分类的数目
# model.summary()
# 冻结卷积层不参与训练
conv_base.trainable = False
# In[14]:
HEIGHT = 299
WIDTH = 299
input_shape = (HEIGHT, WIDTH, 3)
FC_LAYERS = [1024]
dropout = 0.7
epochs = 150
swa = SWA('./keras_swa.model', epochs - 3)
base_model = EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=(HEIGHT, WIDTH, 3))
finetune_model = build_finetune_model(base_model,
dropout=dropout,
num_classes=196)
finetune_model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
snapshot = SnapshotCallbackBuilder(nb_epochs=epochs,
nb_snapshots=1,
init_lr=1e-3)
history = finetune_model.fit_generator(generator=train_generator,
# 添加最后的层
# 输入
# base_model和分类数量
# 输出
# 新的keras的model
# """
# x = base_model.output
# x = GlobalAveragePooling2D()(x)
# predictions = Dense(nb_classes, activation='softmax')(x) # new softmax layer
# model = Model(input=base_model.input, output=predictions)
# return model
# loading pretrained conv base model
# 使用在ImageNet ILSVRC比赛中已经训练好的权重，权重训练自ImageNet，ImageNet预训练权重
# 对输入类型等实例化 参数weights为指定模型的初始化权重检查点
conv_base = Net(weights='imagenet', include_top=False, input_shape=input_shape)
# model = add_new_last_layer(conv_base, nb_classes)
model = models.Sequential()
# model.add(layers.Flatten(name="flatten"))
model.add(conv_base)
model.add(layers.GlobalMaxPooling2D(name="gap"))
if dropout_rate > 0:
model.add(layers.Dropout(dropout_rate, name="dropout_out"))
# model.add(layers.Dense(256, activation='relu', name="fc1"))
model.add(layers.Dense(2, activation='softmax', name="fc_out")) # 这里的2代表要分类的数目
# 输出网络模型参数 查看一下实例化后卷积基模型
# model.summary()
# 冻结卷积层不参与训练
conv_base.trainable = False
batch_size = 50 #จำนวน data ที่ส่งไป Train ในแต่ละครั้ง
width = 150
height = 150
input_shape = (height, width, 3) #ขนาด image enter
epochs = 2000 #จำนวนรอบในการ Train
NUM_TRAIN = 2914# จำนวนภาพ Train
NUM_TEST = 125 #จำนวนภาพ Test
dropout_rate = 0.2
import sys
sys.path.append('/media/tohn/SSD/Sub_Efficient_USAI/content/efficientnet_keras_transfer_learning')
from efficientnet import EfficientNetB0 as Net
from efficientnet import center_crop_and_resize, preprocess_input
# loading pretrained conv base model
conv_base = Net(weights='imagenet', include_top=False, input_shape=input_shape)
# data augmentation เพื่อลดโอกาสการเกิด overfitting
from tensorflow.keras.preprocessing.image import ImageDataGenerator
train_datagen = ImageDataGenerator(
rescale=1./255, # image input 0-255 --> 0-1 เปลี่ยนค่าสี
rotation_range=40, # หมุนภาพในองศา
width_shift_range=0.2, #เปลี่ยนความกว้าง
height_shift_range=0.2, #ปลี่ยนความสูง
shear_range=0.2, #ทำให้ภาพเบี้ยว
zoom_range=0.2, #ซุม image มากสุด 20%
horizontal_flip=True, #พลิกภาพแบบสุ่มตามแนวนอน
fill_mode='nearest')
# Note that the validation data should not be augmented!
shear_range=0.2,
zoom_range=0.2,
horizontal_flip=True,
vertical_flip=True,
# validation_split=0.2,
preprocessing_function=randaug)
d = ImageDataGenerator(preprocessing_function=randaug)
train_generator = g.flow_from_directory(impath,
batch_size=32,
target_size=(224, 224))
val_generator = d.flow_from_directory(
'/media/palm/data/scene_validation_20170908/ai_challenger_scene_validation_20170908/images',
batch_size=32,
target_size=(224, 224))
model = EfficientNetB0(include_top=False, pooling='avg')
x = model.output
x = layers.Dense(80, activation='softmax')(x)
model = models.Model(model.input, x)
model.compile(optimizer=optimizers.SGD(0.01, momentum=0.9),
loss='categorical_crossentropy',
metrics=['acc', f1_m])
tb = CustomTensorBoard('B0, preprocess, 224*224, batch_size 32',
log_dir='logs/B0-1',
write_graph=False)
cp = ModelCheckpoint('weights/b0.h5',
save_weights_only=True,
save_best_only=True,
monitor='val_acc',
mode='max')