the loss starts at 0.03 and the Jaccard coefficient does not increase and stays 0.018 what could possibly be wrong
The code:
model:
def conv_block(x, filter_size, size, dropout=0.6, batch_norm=True):
conv = layers.Conv2D(size, (filter_size, filter_size), padding="same")(x)
if batch_norm is True:
conv = layers.BatchNormalization(axis=3)(conv)
conv = layers.Activation("relu")(conv)
conv = layers.Conv2D(size, (filter_size, filter_size), padding="same")(conv)
if batch_norm is True:
conv = layers.BatchNormalization(axis=3)(conv)
conv = layers.Activation("relu")(conv)
if dropout > 0:
conv = layers.Dropout(dropout)(conv)
return conv
def repeat_elem(tensor, rep):
# lambda function to repeat Repeats the elements of a tensor along an axis
#by a factor of rep.
# If tensor has shape (None, 256,256,3), lambda will return a tensor of shape
#(None, 256,256,6), if specified axis=3 and rep=2.
return layers.Lambda(lambda x, repnum: K.repeat_elements(x, repnum, axis=3),
arguments={'repnum': rep})(tensor)
def res_conv_block(x, filter_size, size, dropout, batch_norm=True):
'''
Residual convolutional layer.
Two variants....
Either put activation function before the addition with shortcut
or after the addition (which would be as proposed in the original resNet).
1. conv - BN - Activation - conv - BN - Activation
- shortcut - BN - shortcut+BN
2. conv - BN - Activation - conv - BN
- shortcut - BN - shortcut+BN - Activation
Check fig 4 in https://arxiv.org/ftp/arxiv/papers/1802/1802.06955.pdf
'''
conv = layers.Conv2D(size, (filter_size, filter_size), padding='same')(x)
if batch_norm is True:
conv = layers.BatchNormalization(axis=3)(conv)
conv = layers.Activation('relu')(conv)
conv = layers.Conv2D(size, (filter_size, filter_size), padding='same')(conv)
if batch_norm is True:
conv = layers.BatchNormalization(axis=3)(conv)
#conv = layers.Activation('relu')(conv) #Activation before addition with shortcut
if dropout > 0:
conv = layers.Dropout(dropout)(conv)
shortcut = layers.Conv2D(size, kernel_size=(1, 1), padding='same')(x)
if batch_norm is True:
shortcut = layers.BatchNormalization(axis=3)(shortcut)
res_path = layers.add([shortcut, conv])
res_path = layers.Activation('relu')(res_path) #Activation after addition with shortcut (Original residual block)
return res_path
def gating_signal(input, out_size, batch_norm=True):
"""
resize the down layer feature map into the same dimension as the up layer feature map
using 1x1 conv
:return: the gating feature map with the same dimension of the up layer feature map
"""
x = layers.Conv2D(out_size, (1, 1), padding='same')(input)
if batch_norm:
x = layers.BatchNormalization()(x)
x = layers.Activation('relu')(x)
return x
def attention_block(x, gating, inter_shape):
shape_x = K.int_shape(x)
shape_g = K.int_shape(gating)
Getting the x signal to the same shape as the gating signal
theta_x = layers.Conv2D(inter_shape, (2, 2), strides=(2, 2), padding='same')(x) # 16
shape_theta_x = K.int_shape(theta_x)
Getting the gating signal to the same number of filters as the inter_shape
phi_g = layers.Conv2D(inter_shape, (1, 1), padding='same')(gating)
upsample_g = layers.Conv2DTranspose(inter_shape, (3, 3),
strides=(shape_theta_x[1] // shape_g[1], shape_theta_x[2] // shape_g[2]),
padding='same')(phi_g) # 16
concat_xg = layers.add([upsample_g, theta_x])
act_xg = layers.Activation('relu')(concat_xg)
psi = layers.Conv2D(1, (1, 1), padding='same')(act_xg)
sigmoid_xg = layers.Activation('sigmoid')(psi)
shape_sigmoid = K.int_shape(sigmoid_xg)
upsample_psi = layers.UpSampling2D(size=(shape_x[1] // shape_sigmoid[1], shape_x[2] // shape_sigmoid[2]))(sigmoid_xg) # 32
#upsample_psi = repeat_elem(upsample_psi, shape_x[3])
y = layers.multiply([upsample_psi, x])
result = layers.Conv2D(shape_x[3], (1, 1), padding='same')(y)
result_bn = layers.BatchNormalization()(result)
return result_bn
def Attention_ResUNet(input_shape, NUM_CLASSES=1, dropout_rate=0.8, batch_norm=True):
'''
Rsidual UNet, with attention
'''
# network structure
FILTER_NUM = 64 # number of basic filters for the first layer
FILTER_SIZE = 3 # size of the convolutional filter
UP_SAMP_SIZE = 2 # size of upsampling filters
# input data
# dimension of the image depth
inputs = layers.Input(input_shape, dtype=tf.float32)
axis = 3
# Downsampling layers
# DownRes 1, double residual convolution + pooling
conv_128 = res_conv_block(inputs, FILTER_SIZE, FILTER_NUM, dropout_rate, batch_norm)
pool_64 = layers.MaxPooling2D(pool_size=(2,2))(conv_128)
# DownRes 2
conv_64 = res_conv_block(pool_64, FILTER_SIZE, 2*FILTER_NUM, dropout_rate, batch_norm)
pool_32 = layers.MaxPooling2D(pool_size=(2,2))(conv_64)
# DownRes 3
conv_32 = res_conv_block(pool_32, FILTER_SIZE, 4*FILTER_NUM, dropout_rate, batch_norm)
pool_16 = layers.MaxPooling2D(pool_size=(2,2))(conv_32)
# DownRes 4
conv_16 = res_conv_block(pool_16, FILTER_SIZE, 8*FILTER_NUM, dropout_rate, batch_norm)
pool_8 = layers.MaxPooling2D(pool_size=(2,2))(conv_16)
# DownRes 5, convolution only
conv_8 = res_conv_block(pool_8, FILTER_SIZE, 16*FILTER_NUM, dropout_rate, batch_norm)
# Upsampling layers
# UpRes 6, attention gated concatenation + upsampling + double residual convolution
gating_16 = gating_signal(conv_8, 8*FILTER_NUM, batch_norm)
att_16 = attention_block(conv_16, gating_16, 8*FILTER_NUM)
up_16 = layers.UpSampling2D(size=(UP_SAMP_SIZE, UP_SAMP_SIZE), data_format="channels_last")(conv_8)
up_16 = layers.concatenate([up_16, att_16], axis=axis)
up_conv_16 = res_conv_block(up_16, FILTER_SIZE, 8*FILTER_NUM, dropout_rate, batch_norm)
# UpRes 7
gating_32 = gating_signal(up_conv_16, 4*FILTER_NUM, batch_norm)
att_32 = attention_block(conv_32, gating_32, 4*FILTER_NUM)
up_32 = layers.UpSampling2D(size=(UP_SAMP_SIZE, UP_SAMP_SIZE), data_format="channels_last")(up_conv_16)
up_32 = layers.concatenate([up_32, att_32], axis=axis)
up_conv_32 = res_conv_block(up_32, FILTER_SIZE, 4*FILTER_NUM, dropout_rate, batch_norm)
# UpRes 8
gating_64 = gating_signal(up_conv_32, 2*FILTER_NUM, batch_norm)
att_64 = attention_block(conv_64, gating_64, 2*FILTER_NUM)
up_64 = layers.UpSampling2D(size=(UP_SAMP_SIZE, UP_SAMP_SIZE), data_format="channels_last")(up_conv_32)
up_64 = layers.concatenate([up_64, att_64], axis=axis)
up_conv_64 = res_conv_block(up_64, FILTER_SIZE, 2*FILTER_NUM, dropout_rate, batch_norm)
# UpRes 9
gating_128 = gating_signal(up_conv_64, FILTER_NUM, batch_norm)
att_128 = attention_block(conv_128, gating_128, FILTER_NUM)
up_128 = layers.UpSampling2D(size=(UP_SAMP_SIZE, UP_SAMP_SIZE), data_format="channels_last")(up_conv_64)
up_128 = layers.concatenate([up_128, att_128], axis=axis)
up_conv_128 = res_conv_block(up_128, FILTER_SIZE, FILTER_NUM, dropout_rate, batch_norm)
# 1*1 convolutional layers
conv_final = layers.Conv2D(NUM_CLASSES, kernel_size=(1,1))(up_conv_128)
conv_final = layers.BatchNormalization(axis=axis)(conv_final)
conv_final = layers.Activation('sigmoid')(conv_final) #Change to softmax for multichannel
# Model integration
model = models.Model(inputs, conv_final, name="AttentionResUNet")
return model
input_shape = (256,256,1)
model=Attention_ResUNet(input_shape, NUM_CLASSES=1, dropout_rate=0.6, batch_norm=True)
#model.summary()
DATA:
image_directory = '/content/drive/MyDrive/MRA/new/train/images/'
mask_directory = '/content/drive/MyDrive/MRA/new/train/masks/'
SIZE = 256
image_dataset = [] #Many ways to handle data, you can use pandas. Here, we are using a list format.
mask_dataset = [] #Place holders to define add labels. We will add 0 to all parasitized images and 1 to uninfected.
images = sorted(os.listdir(image_directory))
for i, image_name in enumerate(images): #Remember enumerate method adds a counter and returns the enumerate object
if (image_name.split('.')[1] == 'png'):
#print(image_directory+image_name)
image = cv2.imread(image_directory+image_name, 1)
image = Image.fromarray(image)
image = image.resize((SIZE, SIZE))
image_dataset.append(np.array(image))
#Iterate through all images in Uninfected folder, resize to 64 x 64
#Then save into the same numpy array 'dataset' but with label 1
masks = sorted(os.listdir(mask_directory))
for i, image_name in enumerate(masks):
if (image_name.split('.')[1] == 'png'):
image = cv2.imread(mask_directory+image_name, 0)
image = Image.fromarray(image)
image = image.resize((SIZE, SIZE))
mask_dataset.append(np.array(image))
#Normalize images
image_dataset = np.array(image_dataset)/255.
#D not normalize masks, just rescale to 0 to 1.
mask_dataset = np.expand_dims((np.array(mask_dataset)),3) /255.
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(image_dataset, mask_dataset, test_size = 0.05, random_state = 0)
#Sanity check, view few mages
import random
import numpy as np
image_number = random.randint(0, len(X_train))
plt.figure(figsize=(12, 6))
plt.subplot(121)
plt.imshow(np.reshape(X_train[image_number], (256, 256, 3)), cmap='gray')
plt.subplot(122)
plt.imshow(np.reshape(y_train[image_number], (256, 256)), cmap='gray')
plt.show()
TRAINING:
IMG_HEIGHT = X_train.shape[1]
IMG_WIDTH = X_train.shape[2]
IMG_CHANNELS = X_train.shape[3]
num_labels = 1 #Binary
input_shape = (IMG_HEIGHT,IMG_WIDTH,IMG_CHANNELS)
batch_size = 16
#from focal_loss import BinaryFocalLoss
'''
Attention Residual Unet
'''
att_res_unet_model = Attention_ResUNet(input_shape)
att_res_unet_model.compile(optimizer=Adam(lr = 1e-2), loss=BinaryFocalLoss(gamma=2),
metrics=['accuracy', jacard_coef])
#att_res_unet_model.compile(optimizer=Adam(lr = 1e-3), loss='binary_crossentropy',
metrics=['accuracy', jacard_coef])
#print(att_res_unet_model.summary())
start3 = datetime.now()
att_res_unet_history = att_res_unet_model.fit(X_train, y_train,
verbose=1,
batch_size = batch_size,
validation_data=(X_test, y_test ),
shuffle=False,
epochs=50)
stop3 = datetime.now()
#Execution time of the model
execution_time_AttResUnet = stop3-start3
#print("Attention ResUnet execution time is: ", execution_time_AttResUnet)
att_res_unet_model.save('AttResUnet.hdf5')