Train_Skin_Lesion_Segmentation.py

# -*- coding: utf-8 -*-
"""
Created on Sat Jun  8 18:15:43 2019
@author: Reza Azad
"""
from __future__ import division
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "9"
import models as M
import numpy as np
from keras.callbacks import ModelCheckpoint, TensorBoard,ReduceLROnPlateau
from keras import callbacks
import pickle

# ===== normalize over the dataset 
def dataset_normalized(imgs):
    imgs_normalized = np.empty(imgs.shape)
    imgs_std = np.std(imgs)
    imgs_mean = np.mean(imgs)
    imgs_normalized = (imgs-imgs_mean)/imgs_std
    for i in range(imgs.shape[0]):
        imgs_normalized[i] = ((imgs_normalized[i] - np.min(imgs_normalized[i])) / (np.max(imgs_normalized[i])-np.min(imgs_normalized[i])))*255
    return imgs_normalized
       
    
####################################  Load Data #####################################
base_add = '../'
tr_data    = np.load(base_add+'data_train.npy')
val_data   = np.load(base_add+'data_val.npy')

tr_mask    = np.load(base_add+'mask_train.npy')
val_mask   = np.load(base_add+'mask_val.npy')

tr_mask    = np.expand_dims(tr_mask, axis=3)
val_mask   = np.expand_dims(val_mask, axis=3)

print('ISIC18 Dataset loaded')

tr_data   = dataset_normalized(tr_data)
val_data  = dataset_normalized(val_data)

tr_mask   = tr_mask /255.
val_mask  = val_mask /255.

print('dataset Normalized')

# Build model
model = M.Deeplabv3pa(input_shape = (256, 256, 3))
model.summary()

print('Training')
batch_size = 7
nb_epoch   = 100


mcp_save = ModelCheckpoint('weight_isic18_deeplab_v3pa', save_best_only=True, monitor='val_loss', mode='min')
reduce_lr_loss = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=7, verbose=1, epsilon=1e-4, mode='min')
history = model.fit(tr_data, tr_mask,
              batch_size=batch_size,
              epochs=nb_epoch,
              shuffle=True,
              verbose=1,
              validation_data=(val_data, val_mask), callbacks=[mcp_save, reduce_lr_loss] )