def trainModel(self, dataObj, save, plot):
#Define session
for i in range(self.displayPeriod):
#Get data from dataObj
data = dataObj.getData(self.batchSize)
feedDict = {self.input: data[0], self.gt: data[1]}
#Run optimizer
if (self.preTrain):
self.sess.run(self.optimizerPre, feed_dict=feedDict)
else:
self.sess.run(self.optimizer, feed_dict=feedDict)
if (i % self.writeStep == 0):
summary = self.sess.run(self.mergedSummary, feed_dict=feedDict)
self.train_writer.add_summary(summary, self.timestep)
if (i % self.progress == 0):
print "Timestep ", self.timestep
self.timestep += 1
if (save):
save_path = self.saver.save(self.sess,
self.saveFile,
global_step=self.timestep,
write_meta_graph=False)
print("Model saved in file: %s" % save_path)
if (plot):
filename = self.plotDir + "weights_" + str(self.timestep) + ".png"
np_w = self.sess.run(self.W_encode, feed_dict=feedDict)
plot_weights(np_w, filename, order=[3, 0, 1, 2])
def trainW(self):
feedDict = {self.inputImage: self.currImg}
#Visualization
if (self.plotTimestep % self.plotPeriod == 0):
np_V1_W = self.sess.run(self.weightImages)
np_V1_A = self.sess.run(self.V1_A)
rescaled_V1_W = np.exp(np.abs(np_V1_W * np.sqrt(self.patchSizeX * self.patchSizeY))) * np.sign(np_V1_W)
#plot_weights(rescaled_V1_W, self.plotDir+"dict_"+str(self.timestep), activity=np_V1_A)
plot_weights(rescaled_V1_W, self.plotDir+"dict_"+str(self.timestep))
#plot_1d_weights(rescaled_V1_W, self.plotDir+"dict_"+str(self.timestep), activity=np_V1_A)
np_inputImage = self.currImg
np_recon = self.sess.run(self.recon, feed_dict=feedDict)
#Draw recons
rescaled_inputImage = np.exp(np.abs(np_inputImage * np.sqrt(self.patchSizeX * self.patchSizeY))) * np.sign(np_inputImage)
rescaled_recon = np.exp(np.abs(np_recon * np.sqrt(self.patchSizeX * self.patchSizeY))) * np.sign(np_recon)
plotRecon(np_recon, np_inputImage, self.plotDir+"recon_"+str(self.timestep), r=range(4))
#plotRecon1d(np.squeeze(rescaled_recon), np.squeeze(rescaled_inputImage), self.plotDir+"recon_"+str(self.timestep), r=range(4))
#Update weights
self.sess.run(self.optimizerW, feed_dict=feedDict)
#New image
self.currImg = self.dataObj.getData(self.batchSize)
self.plotTimestep += 1
def trainW(self):
feedDict = {self.inputImage: self.currImg}
#Visualization
if (self.plotTimestep % self.plotPeriod == 0):
for l in range(self.numLayers):
np_V1_W = self.sess.run(self.visWeight[l])
plot_weights(
np_V1_W, self.plotDir + "dict_S" + str(l) + "_" +
str(self.timestep) + ".png")
#Draw recons
np_inputImage = self.currImg
np_recon = self.sess.run(self.visRecon[l], feed_dict=feedDict)
np_t_recon = self.sess.run(self.t_visRecon[l],
feed_dict=feedDict)
plotRecon(np_recon,
np_inputImage,
self.plotDir + "recon_S" + str(l) + "_" +
str(self.timestep) + ".png",
r=range(4))
plotRecon(np_t_recon,
np_inputImage,
self.plotDir + "t_recon_S" + str(l) + "_" +
str(self.timestep) + ".png",
r=range(4))
#Update weights
self.sess.run(self.optimizerW, feed_dict=feedDict)
#New image
self.currImg = self.dataObj.getData(self.batchSize)
self.plotTimestep += 1
def plot(self):
#Visualization
if (self.plotTimestep % self.plotPeriod == 0):
np_V1_W = self.sess.run(self.weightImages)
np_V1_A = self.sess.run(self.V1_A)
#plot_weights(rescaled_V1_W, self.plotDir+"dict_"+str(self.timestep), activity=np_V1_A)
plotStr = self.plotDir + "dict_"+str(self.timestep)
if(np_V1_W.ndim == 2):
rescaled_V1_W = np.exp(np.abs(np_V1_W * np.sqrt(self.patchSizeX * self.patchSizeY))) * np.sign(np_V1_W)
plot_1d_weights(rescaled_V1_W, plotStr, activity=np_V1_A)
else:
plot_weights(V1_W, plotStr)
np_inputImage = self.currImg
np_recon = self.sess.run(self.recon, feed_dict=feedDict)
#Draw recons
if(np.squeeze(np_recon).ndim == 2):
rescaled_inputImage = np_inputImage * np.sqrt(self.patchSizeX * self.patchSizeY)
rescaled_recon = np_recon * np.sqrt(self.patchSizeX * self.patchSizeY)
exp_inputImage = np.squeeze(np.exp(np.abs(rescaled_inputImage) - 1e-10) * np.sign(np_inputImage))
exp_recon = np.squeeze(np.exp(np.abs(rescaled_recon) - 1e-10) * np.sign(np_recon))
plotRecon1d(exp_recon, exp_inputImage, self.plotDir+"recon_"+str(self.timestep), r=range(4))
else:
plotRecon(np_recon, np_inputImage, self.plotDir+"recon_"+str(self.timestep), r=range(4))
self.plotTimestep += 1
def trainW(self):
feedDict = {self.inputImage: self.currImg}
#Visualization
if (self.plotTimestep % self.plotPeriod == 0):
np_V1_W = self.sess.run(self.weightImages)
plot_weights(np_V1_W, self.plotDir+"dict_"+str(self.timestep)+".png")
#Draw recons
np_inputImage = self.currImg
np_recon = self.sess.run(self.recon, feed_dict=feedDict)
plotRecon(np_recon, np_inputImage, self.plotDir+"recon_"+str(self.timestep), r=range(4))
#Update weights
self.sess.run(self.optimizerW, feed_dict=feedDict)
#New image
self.currImg = self.dataObj.getData(self.batchSize)
self.plotTimestep += 1
def trainW(self):
feedDict = {self.inputImage: self.currImg}
#Visualization
if (self.plotTimestep % self.plotPeriod == 0):
np_V1_W = self.sess.run(self.weightImages)
plot_weights(np_V1_W, self.plotDir+"dict_"+str(self.timestep)+".png")
#Draw recons
np_inputImage = self.currImg
np_recon = self.sess.run(self.recon, feed_dict=feedDict)
plotRecon(np_recon, np_inputImage, self.plotDir+"recon_"+str(self.timestep), r=range(4))
#Update weights
self.sess.run(self.optimizerW, feed_dict=feedDict)
#New image
self.currImg = self.dataObj.getData(self.batchSize)
self.plotTimestep += 1
def plotWeight(self):
#Make directory for timestep
outPlotDir = self.plotDir + "/" + str(self.timestep) + "/"
if not os.path.exists(outPlotDir):
os.makedirs(outPlotDir)
np_V1_W = self.sess.run(self.weightImages)
np_V1_A = self.sess.run(self.V1_A)
#plot_weights(rescaled_V1_W, self.plotDir+"dict_"+str(self.timestep), activity=np_V1_A)
plotStr = outPlotDir + "dict_"
if (np_V1_W.ndim == 3):
plot_1d_weights(np_V1_W,
plotStr,
activity=np_V1_A,
sepFeatures=True)
else:
plot_weights(V1_W, plotStr)
def trainW(self):
feedDict = {self.inputImage: self.currImg}
#Visualization
if (self.plotTimestep % self.plotPeriod == 0):
for l in range(self.numLayers):
np_V1_W = self.sess.run(self.visWeight[l])
plot_weights(np_V1_W, self.plotDir+"dict_S" + str(l) + "_" +str(self.timestep)+".png")
#Draw recons
np_inputImage = self.currImg
np_recon = self.sess.run(self.visRecon[l], feed_dict=feedDict)
np_t_recon = self.sess.run(self.t_visRecon[l], feed_dict=feedDict)
plotRecon(np_recon, np_inputImage, self.plotDir+"recon_S"+str(l)+"_"+str(self.timestep)+".png", r=range(4))
plotRecon(np_t_recon, np_inputImage, self.plotDir+"t_recon_S"+str(l)+"_"+str(self.timestep)+".png", r=range(4))
#Update weights
self.sess.run(self.optimizerW, feed_dict=feedDict)
#New image
self.currImg = self.dataObj.getData(self.batchSize)
self.plotTimestep += 1
def trainModel(self, dataObj, save, plot):
#Define session
for i in range(self.displayPeriod):
#Get data from dataObj
data = dataObj.getData(self.batchSize)
feedDict = {self.input: data[0], self.gt: data[1]}
#Run optimizer
if(self.preTrain):
self.sess.run(self.optimizerPre, feed_dict=feedDict)
else:
self.sess.run(self.optimizer, feed_dict=feedDict)
if(i%self.writeStep == 0):
summary = self.sess.run(self.mergedSummary, feed_dict=feedDict)
self.train_writer.add_summary(summary, self.timestep)
if(i%self.progress == 0):
print "Timestep ", self.timestep
self.timestep+=1
if(save):
save_path = self.saver.save(self.sess, self.saveFile, global_step=self.timestep, write_meta_graph=False)
print("Model saved in file: %s" % save_path)
if(plot):
filename = self.plotDir + "weights_" + str(self.timestep) + ".png"
np_w = self.sess.run(self.W_encode, feed_dict=feedDict)
plot_weights(np_w, filename, order=[3, 0, 1, 2])
