3DUnet_tensorflow2.0

This Repo is for implementation of 3D unet in Tensorflow 2.0v

Files:

• i) unet_config.py -|--> All the Netword and Training configuration
• ii) Unet3D |--> Network architecture
• iii) Train_Unet3D |--> Training Script. it has tfrecord decoder, tfdataset reading pipeline and training loop,Losses and Matrics function. Binary Dice Coefficent and Dice Loss

How to run

To run the model all is to need to configure the unet_config.py based on your requiremnet.

###---Number-of-GPU
NUM_OF_GPU=4
DISTRIIBUTED_STRATEGY_GPUS=["gpu:0","gpu:1","gpu:2","gpu:3"]
'''
if want to resume training from the weights Set
RESUME_TRAINING=1
'''
###----Resume-Training
RESUME_TRAINING=1
RESUME_TRAIING_MODEL='/Path/of/the/model/weight/Model.h5'
TRAINING_INITIAL_EPOCH=1381
NUMBER_OF_CLASSES=1
INPUT_PATCH_SIZE=(384,192,192, 1)
##Training Hyper-Parameter
TRAIN_CLASSIFY_LEARNING_RATE =1e-4
#TRAIN_CLASSIFY_LOSS=tf.keras.losses.binary_crossentropy()
OPTIMIZER=tf.keras.optimizers.Adam(lr=TRAIN_CLASSIFY_LEARNING_RATE,epsilon=1e-5)
#TRAIN_CLASSIFY_METRICS=tf.keras.metrics.binary_accuracy()
BATCH_SIZE=4
TRAINING_STEP_PER_EPOCH=math.ceil((76)/BATCH_SIZE)
VALIDATION_STEP=math.ceil((6)/BATCH_SIZE)
TRAING_EPOCH=1600
NUMBER_OF_PARALLEL_CALL=4
PARSHING=2*BATCH_SIZE
#--Callbacks-----
ModelCheckpoint_MOTITOR='LUNGSegVal_loss'
TRAINING_SAVE_MODEL_PATH=''/Path/to/save/model/weight/Model.h5''
TRAINING_CSV='LungSEG_Model_March30_2020.csv'
####
TRAINING_TF_RECORDS='/Training/tfrecords/path/'
VALIDATION_TF_RECORDS='/Val/tfrecords/path/'

Dice Loss

def dice_coe(y_true,y_pred, loss_type='jaccard', smooth=1.):

    y_true_f = tf.reshape(y_true,[-1])
    y_pred_f = tf.reshape(y_pred,[-1])

    intersection = tf.reduce_sum(y_true_f * y_pred_f)

    if loss_type == 'jaccard':
        union = tf.reduce_sum(tf.square(y_pred_f)) + tf.reduce_sum(tf.square(y_true_f))

    elif loss_type == 'sorensen':
        union = tf.reduce_sum(y_pred_f) + tf.reduce_sum(y_true_f)

    else:
        raise ValueError("Unknown `loss_type`: %s" % loss_type)

    return (2. * intersection + smooth) / (union + smooth)

def dice_loss(y_true,y_pred, loss_type='jaccard', smooth=1.):

    y_true_f = tf.cast(tf.reshape(y_true,[-1]),tf.float32)
    y_pred_f =tf.cast(tf.reshape(y_pred,[-1]),tf.float32)

    intersection = tf.reduce_sum(y_true_f * y_pred_f)

    if loss_type == 'jaccard':
        union = tf.reduce_sum(tf.square(y_pred_f)) + tf.reduce_sum(tf.square(y_true_f))

    elif loss_type == 'sorensen':
        union = tf.reduce_sum(y_pred_f) + tf.reduce_sum(y_true_f)

    else:
        raise ValueError("Unknown `loss_type`: %s" % loss_type)

    return (1-(2. * intersection + smooth) / (union + smooth))