update

.gitignore

@@ -12,9 +12,10 @@
 
 # virtual machine crash logs, see http://www.java.com/en/download/help/error_hotspot.xml
 /model/*.hdf5
-/data/train/
-/data/validation/
+/data/
 /check/
+/log/
+/png/
 #/src
 #/build
 #/libs

README.md

@@ -21,7 +21,7 @@ to be sufficient to train a good fault segmentation network.**
 
 ### Training
 
-Run train3 to start training a new faultSeg model by using the 200 synthetic datasets
+Run train to start training a new faultSeg model by using the 200 synthetic datasets
 
 ## Publications
 

apply.py

@@ -0,0 +1,105 @@
+import math
+import skimage
+import numpy as np
+import os
+import matplotlib.pyplot as plt
+
+import tensorflow as tf
+#from keras import backend
+from keras.layers import *
+from keras.models import load_model
+from skimage.measure import compare_psnr
+from unet3 import cross_entropy_balanced
+import os
+pngDir = './png/'
+model = load_model('check/fseg-'+'25.hdf5',
+                 #'impd.hdf5',
+                 custom_objects={
+                     'cross_entropy_balanced': cross_entropy_balanced
+                 #} 
+                 )
+def main():
+  #goTrainTest()
+  goValidTest()
+  #goF3Test()
+
+def goTrainTest():
+  seismPath = "./data/train/seis/"
+  faultPath = "./data/train/fault/"
+  n1,n2,n3=128,128,128
+  dk = 100
+  gx = np.fromfile(seismPath+str(dk)+'.dat',dtype=np.single)
+  fx = np.fromfile(faultPath+str(dk)+'.dat',dtype=np.single)
+  gx = np.reshape(gx,(n1,n2,n3))
+  fx = np.reshape(fx,(n1,n2,n3))
+  gm = np.mean(gx)
+  gs = np.std(gx)
+  gx = gx-gm
+  gx = gx/gs
+  gx = np.transpose(gx)
+  fx = np.transpose(fx)
+  fp = model.predict(np.reshape(gx,(1,n1,n2,n3,1)),verbose=1)
+  fp = fp[0,:,:,:,0]
+  gx1 = gx[50,:,:]
+  fx1 = fx[50,:,:]
+  fp1 = fp[50,:,:]
+  plot2d(gx1,fx1,fp1,png='fp')
+
+def goValidTest():
+  seismPath = "./data/validation/seis/"
+  faultPath = "./data/validation/fault/"
+  n1,n2,n3=128,128,128
+  dk = 2
+  gx = np.fromfile(seismPath+str(dk)+'.dat',dtype=np.single)
+  fx = np.fromfile(faultPath+str(dk)+'.dat',dtype=np.single)
+  gx = np.reshape(gx,(n1,n2,n3))
+  fx = np.reshape(fx,(n1,n2,n3))
+  gm = np.mean(gx)
+  gs = np.std(gx)
+  gx = gx-gm
+  gx = gx/gs
+  gx = np.transpose(gx)
+  fx = np.transpose(fx)
+  fp = model.predict(np.reshape(gx,(1,n1,n2,n3,1)),verbose=1)
+  fp = fp[0,:,:,:,0]
+  gx1 = gx[50,:,:]
+  fx1 = fx[50,:,:]
+  fp1 = fp[50,:,:]
+  plot2d(gx1,fx1,fp1,png='fp')
+
+def goF3Test(): 
+  seismPath = "./data/prediction/f3d/"
+  n3,n2,n1=512,384,128
+  gx = np.fromfile(seismPath+'gxl.dat',dtype=np.single)
+  gx = np.reshape(gx,(n3,n2,n1))
+  gm = np.mean(gx)
+  gs = np.std(gx)
+  gx = gx-gm
+  gx = gx/gs
+  gx = np.transpose(gx)
+  fp = model.predict(np.reshape(gx,(1,n1,n2,n3,1)),verbose=1)
+  fp = fp[0,:,:,:,0]
+  gx1 = gx[99,:,:]
+  fp1 = fp[99,:,:]
+  plot2d(gx1,fp1,fp1,png='f3d/fp')
+
+def plot2d(gx,fx,fp,png=None):
+  fig = plt.figure(figsize=(15,5))
+  #fig = plt.figure()
+  ax = fig.add_subplot(131)
+  ax.imshow(gx,vmin=-2,vmax=2,cmap=plt.cm.bone,interpolation='bicubic',aspect=1)
+  ax = fig.add_subplot(132)
+  ax.imshow(fx,vmin=0,vmax=1,cmap=plt.cm.bone,interpolation='bicubic',aspect=1)
+  ax = fig.add_subplot(133)
+  ax.imshow(fp,vmin=0,vmax=1.0,cmap=plt.cm.bone,interpolation='bicubic',aspect=1)
+  if png:
+    plt.savefig(pngDir+png+'.png')
+  #cbar = plt.colorbar()
+  #cbar.set_label('Fault probability')
+  plt.tight_layout()
+  plt.show()
+
+if __name__ == '__main__':
+    main()
+
+

cpurun

@@ -0,0 +1,5 @@
+#!/bin/sh
+
+#org.python.util.jython $*
+CUDA_VISIBLE_DEVICES='' python $*
+

train.py

@@ -36,7 +36,7 @@ def goTrain():
   valid_generator = DataGenerator(dpath=seismPathV,fpath=faultPathV,
                                   data_IDs=valid_ID,**params)
   model = unet(input_size=(None, None, None,1))
-  model.compile(optimizer=Adam(lr=1e-3), loss='binary_crossentropy', 
+  model.compile(optimizer=Adam(lr=1e-3), loss=cross_entropy_balanced, 
                 metrics=['accuracy'])
   model.summary()
 

unet3.py

@@ -31,8 +31,8 @@ def unet(pretrained_weights = None,input_size = (None,None,None,1)):
     conv3 = Conv3D(64, (3,3,3), activation='relu', padding='same')(conv3)
     pool3 = MaxPooling3D(pool_size=(2,2,2))(conv3)
 
-    conv4 = Conv3D(512, (3,3,3), activation='relu', padding='same')(pool3)
-    conv4 = Conv3D(512, (3,3,3), activation='relu', padding='same')(conv4)
+    conv4 = Conv3D(128, (3,3,3), activation='relu', padding='same')(pool3)
+    conv4 = Conv3D(128, (3,3,3), activation='relu', padding='same')(conv4)
 
     up5 = concatenate([UpSampling3D(size=(2,2,2))(conv4), conv3], axis=-1)
     conv5 = Conv3D(64, (3,3,3), activation='relu', padding='same')(up5)

utils.py

@@ -1,5 +1,6 @@
 import numpy as np
 import keras
+import random
 from keras.utils import to_categorical
 
 class DataGenerator(keras.utils.Sequence):
@@ -43,8 +44,6 @@ def on_epoch_end(self):
   def __data_generation(self, data_IDs_temp):
     'Generates data containing batch_size samples'
     # Initialization
-    X = np.zeros((4, *self.dim, self.n_channels),dtype=np.single)
-    Y = np.zeros((4, *self.dim, self.n_channels),dtype=np.single)
     gx  = np.fromfile(self.dpath+str(data_IDs_temp[0])+'.dat',dtype=np.single)
     fx  = np.fromfile(self.fpath+str(data_IDs_temp[0])+'.dat',dtype=np.single)
     gx = np.reshape(gx,self.dim)
@@ -59,6 +58,8 @@ def __data_generation(self, data_IDs_temp):
     gx = np.transpose(gx)
     fx = np.transpose(fx)
     # Generate data
+    X = np.zeros((4, *self.dim, self.n_channels),dtype=np.single)
+    Y = np.zeros((4, *self.dim, self.n_channels),dtype=np.single)
     for i in range(4):
       X[i,] = np.reshape(np.rot90(gx,i,(2,1)), (*self.dim,self.n_channels))
       Y[i,] = np.reshape(np.rot90(fx,i,(2,1)), (*self.dim,self.n_channels))