myaug.py

# Augmentation routine, for both response and predictor
def myaug(X_train_org,y_train_org):
    x=[]
    y=[]
    halves=64
    for i in range(0,len(X_train_org)):
        r=random.randint(0,4)
        posneg=random.randint(0,1)
        if posneg==1:
            rpx=random.randint(5,20)
        else:
            rpx=random.randint(-20,-5)
        im=X_train_org[i].reshape(256,256)
        mask=y_train_org[i].reshape(256,256)
        mask[mask>0]=1
        avg=get_centroid(mask)
        if (r==0):
            extent=[int(avg[0]-halves),int(avg[0]+halves),int(avg[1]-halves),int(avg[1]+halves)]
        elif (r==1):
            extent=[int(avg[0]+rpx-halves),int(avg[0]+rpx+halves),int(avg[1]-halves),int(avg[1]+halves)]
        elif(r==2):
            extent=[int(avg[0]-halves),int(avg[0]+halves),int(avg[1]+rpx-halves),int(avg[1]+rpx+halves)]
        elif(r==3):
            extent=[int(avg[0]+rpx-halves),int(avg[0]+rpx+halves),int(avg[1]+rpx-halves),int(avg[1]+rpx+halves)]
        elif(r==4):
            ang=random.randint(5,15)
            im=sp.ndimage.interpolation.rotate(im,ang)
            mask=sp.ndimage.interpolation.rotate(mask,ang)
            avg=get_centroid(mask)
            extent=[int(avg[0]-halves),int(avg[0]+halves),int(avg[1]-halves),int(avg[1]+halves)]
        im=im[extent[0]:extent[1],extent[2]:extent[3]]
        im=(im-im.mean())/im.std()
        mask=mask[extent[0]:extent[1],extent[2]:extent[3]]    
        x.append(im)
        mask=imresize(mask,(32,32)).flatten()
        mask[mask>0]=1
        y.append(mask)
    X_batch=np.asarray(x).reshape(len(x),1,128,128)    
    y_batch=np.asarray(y).astype('int')
    return X_batch,y_batch