def mitos12_cae_model_test():
model = {
'inputLayer' : ImageInputLayer(width=64,height=64,channels=3),
'hiddenLayers' : [
ConvolutionLayer(kernelsize=5, channels=3, features=12, stride=2, weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[64,64,3]),
ConvolutionLayer(kernelsize=5, channels=12, features=48, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[32,32,12]),
ConvolutionLayer(kernelsize=5, channels=48, features=192, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[16,16,48])
]
}
batch_size = 50
net = Network(Mitos12_CAE_Model, objective='reconstruction', batch_size=batch_size)
net.setupTraining("squared-diff", "Adam")
init = tf.initialize_all_variables()
mitos12 = MITOS12Data(train_dirs=["/media/sf_VirtualDropbox"])
saver = tf.train.Saver()
sess = tf.Session()
with sess.as_default():
assert tf.get_default_session() is sess
sess.run(init)
saver.restore(sess, "/home/adrien/workspace/DeepNet/mitos12ConvAutoEncoder3.ckpt")
xs = mitos12.next_batch(batch_size)
rs = net.predict(xs)
rs[0][rs[0]>1.] = 1.
plt.figure(1)
plt.subplot(2,3,1)
plt.title('Original')
plt.imshow(xs[0], interpolation='nearest')
plt.axis('off')
plt.subplot(2,3,2)
plt.title('Reconstruction')
plt.imshow(rs[0], interpolation='nearest')
plt.axis('off')
plt.gray()
plt.subplot(2,3,4)
plt.title('Diff - R')
plt.imshow(np.abs(rs[0]-xs[0])[:,:,0], interpolation='nearest')
plt.axis('off')
plt.subplot(2,3,5)
plt.title('Diff - G')
plt.imshow(np.abs(rs[0]-xs[0])[:,:,1], interpolation='nearest')
plt.axis('off')
plt.subplot(2,3,6)
plt.title('Diff - B')
plt.imshow(np.abs(rs[0]-xs[0])[:,:,2], interpolation='nearest')
plt.axis('off')
plt.show()
def mnistFullyConnectedClassif(train=False):
model = {
'inputLayer' : FlatInputLayer(inputsize=784),
'hiddenLayers' : [
FullyConnectedLayer(inputsize=784,outputsize=1000,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
FullyConnectedLayer(inputsize=1000,outputsize=200,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh)
],
'outputLayer' : FullyConnectedLayer(inputsize=200,outputsize=10,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.softmax)
}
batch_size = 50
net = Network(model, objective='classification', batch_size=batch_size)
net.setupTraining("cross-entropy", "Adam")
init = tf.initialize_all_variables()
mnist = MNISTData(train_dir='MNIST_data', one_hot=True)
saver = tf.train.Saver()
sess = tf.Session()
with sess.as_default():
assert tf.get_default_session() is sess
sess.run(init)
if train==False:
saver.restore(sess, "/home/adrien/workspace/DeepNet/mnistFullyConnectedClassif.ckpt")
print "Test accuracy %g"%net.evaluate(mnist.test['images'], mnist.test['labels'])
else:
for i in range(20000):
batch_xs, batch_ys = mnist.next_batch(batch_size, set=mnist.train)
if i%1000 == 0:
cost = net.cost.eval(feed_dict={net.x: batch_xs, net.target: batch_ys})
print "step %d, training cost %g"%(i,cost)
net.train(batch_xs, batch_ys)
print "Test accuracy %g"%net.evaluate(mnist.test['images'], mnist.test['labels'])
save_path = saver.save(sess, "/home/adrien/workspace/DeepNet/mnistFullyConnectedClassif.ckpt")
plt.figure(1)
plt.gray()
plt.imshow(mnist.test['images'][0].reshape((28,28)), interpolation='nearest')
plt.title(str(np.argmax(net.predict([mnist.test['images'][0]]))))
plt.show()
def mitos12ClassifierFromConvAutoEncoder3(train=False, resumeTraining=False, iterations=20000):
autoEncoderModel = {
'inputLayer' : ImageInputLayer(width=64,height=64,channels=3),
'hiddenLayers' : [
ConvolutionLayer(kernelsize=5, channels=3, features=12, stride=2, weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[64,64,3]),
ConvolutionLayer(kernelsize=5, channels=12, features=48, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[32,32,12]),
ConvolutionLayer(kernelsize=5, channels=48, features=192, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[16,16,48])
]
}
classifierModel = {
'inputLayer': ImageInputLayer(width=8,height=8,channels=192),
'hiddenLayers': [
ImageToVectorLayer(imagesize=(8,8,192)),
FullyConnectedLayer(inputsize=8*8*192,outputsize=200,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
FullyConnectedLayer(inputsize=200,outputsize=100,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh)
],
'outputLayer' : FullyConnectedLayer(inputsize=100,outputsize=2,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.softmax)
}
batch_size = 50
autoEncoder = Network(autoEncoderModel, objective='reconstruction', batch_size=batch_size)
clf = Network(classifierModel, objective='classification', batch_size=batch_size)
clf.setupTraining("cross-entropy", "Adam", a=0.995)
init = tf.initialize_all_variables()
basedir = "/media/sf_E_DRIVE/Dropbox/ULB/Doctorat/ImageSet/MITOS12/"
mitos12 = MITOS12Data(train_dirs=[os.path.join(basedir,d) for d in ["A00_v2", "A01_v2", "A02_v2", "A03_v2", "A04_v2"]])
aesaver = tf.train.Saver(autoEncoder.getVariables())
clfsaver = tf.train.Saver(clf.getVariables())
sess = tf.Session()
with sess.as_default():
assert tf.get_default_session() is sess
sess.run(init)
aesaver.restore(sess, "/home/adrien/workspace/DeepNet/mitos12ConvAutoEncoder3.ckpt")
if train==False or resumeTraining==True:
clfsaver.restore(sess, "/home/adrien/workspace/DeepNet/mitos12ClfFromConvAutoEncoder3.ckpt")
if train==True:
for i in range(iterations):
batch = mitos12.next_supervised_batch(batch_size)
input_images = [b[0] for b in batch]
batch_xs = autoEncoder.encode(input_images)
batch_ys = [b[1] for b in batch]
if i%1000==0:
cost = clf.cost.eval(feed_dict={clf.x: batch_xs, clf.target: batch_ys})
loss = clf.loss.eval(feed_dict={clf.x: batch_xs, clf.target: batch_ys})
l2loss = clf.l2loss.eval()
print "step %d, training cost %g, loss %g, l2loss %g"%(i,cost,loss,l2loss)
# cost = clf.cost.eval(feed_dict={clf.x: batch_xs, clf.target: batch_ys})
# print "step %d, training cost %g"%(i,cost)
save_path = clfsaver.save(sess, "/home/adrien/workspace/DeepNet/mitos12ClfFromConvAutoEncoder3.ckpt")
clf.train(batch_xs, batch_ys)
save_path = clfsaver.save(sess, "/home/adrien/workspace/DeepNet/mitos12ClfFromConvAutoEncoder3.ckpt")
# Eval :
Cmat = np.zeros((2,2))
for i in range(50):
batch = mitos12.next_supervised_batch(batch_size)
input_images = [b[0] for b in batch]
batch_xs = autoEncoder.encode(input_images)
batch_ys = [b[1] for b in batch]
pred = clf.predict(batch_xs)
Cmat += C(pred,batch_ys)
print Cmat
im_, path, basename = mitos12.images[18]
im = np.array(im_)
stride = 15
rangex = np.arange(0,im.shape[0]-64,stride)
rangey = np.arange(0,im.shape[1]-64,stride)
ts = [(t/len(rangey), t%len(rangey)) for t in range(len(rangex)*len(rangey))]
chunks = [im[tx*stride:tx*stride+64,ty*stride:ty*stride+64,:] for tx,ty in ts]
chunksPos = [(tx*stride,ty*stride) for tx,ty in ts]
pMitosis = np.zeros((im.shape[0], im.shape[1], 3))
print len(chunks)
for t in range(len(chunks)/50):
batch = chunks[t*50:t*50+50]
batch_xs = autoEncoder.encode(batch)
is_mitosis = clf.predict(batch_xs)
for i,p in enumerate(is_mitosis):
cp = chunksPos[t*50+i]
pMitosis[cp[0]:cp[0]+64, cp[1]:cp[1]+64, 0] += p[0]
pMitosis[cp[0]:cp[0]+64, cp[1]:cp[1]+64, 1] += p[1]
pMitosis[cp[0]:cp[0]+64, cp[1]:cp[1]+64, 2] += 1
plt.figure()
plt.gray()
plt.imshow(pMitosis[:,:,0], interpolation=None)
plt.figure()
plt.imshow(pMitosis[:,:,1], interpolation=None)
plt.figure()
plt.imshow(pMitosis[:,:,2], interpolation=None)
plt.figure()
plt.imshow(pMitosis[:,:,0]/pMitosis[:,:,2], interpolation=None)
plt.figure()
plt.imshow(plt.imread(basename+".jpg"))
plt.show()
def mitos12ConvAutoEncoder3(train=False, resumeTraining=False, iterations=20000):
model = {
'inputLayer' : ImageInputLayer(width=64,height=64,channels=3),
'hiddenLayers' : [
ConvolutionLayer(kernelsize=5, channels=3, features=12, stride=2, weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[64,64,3]),
ConvolutionLayer(kernelsize=5, channels=12, features=48, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[32,32,12]),
ConvolutionLayer(kernelsize=5, channels=48, features=192, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[16,16,48])
]
}
batch_size = 50
net = Network(model, objective='reconstruction', batch_size=batch_size)
net.setupTraining("squared-diff", "Adam")
init = tf.initialize_all_variables()
basedir = "/media/sf_E_DRIVE/Dropbox/ULB/Doctorat/ImageSet/MITOS12/"
mitos12 = MITOS12Data(train_dirs=[os.path.join(basedir,d) for d in ["A00_v2", "A01_v2", "A02_v2", "A03_v2", "A04_v2"]])
saver = tf.train.Saver()
sess = tf.Session()
with sess.as_default():
assert tf.get_default_session() is sess
sess.run(init)
if train==False or resumeTraining==True:
saver.restore(sess, "/home/adrien/workspace/DeepNet/mitos12ConvAutoEncoder3.ckpt")
if train==True:
for i in range(iterations):
batch_xs = mitos12.next_batch(batch_size)
batch_ys = batch_xs
if i%1000 == 0:
cost = net.cost.eval(feed_dict={net.x: batch_xs, net.target: batch_ys})
loss = net.loss.eval(feed_dict={net.x: batch_xs, net.target: batch_ys})
l2loss = net.l2loss.eval()
print "step %d, training cost %g, loss %g, l2loss %g"%(i,cost,loss,l2loss)
save_path = saver.save(sess, "/home/adrien/workspace/DeepNet/mitos12ConvAutoEncoder3.ckpt")
net.train(batch_xs, batch_ys)
save_path = saver.save(sess, "/home/adrien/workspace/DeepNet/mitos12ConvAutoEncoder3.ckpt")
xs = mitos12.next_batch(batch_size)
#es = net.encode(xs)
# for idx in range(len(es)):
# max_act = 0
# arg_max_act = 0
# es0 = es[idx]
# for t in np.arange(0,192):
# if( es0[:,:,t].sum() > max_act ):
# max_act = es0[:,:,t].sum()
# arg_max_act = t
# print arg_max_act
# return
# featt = np.zeros(es0.shape)
# featt[:,:,arg_max_act] = es0[:,:,arg_max_act]
# rs = net.decode([featt])
# print es0[:,:,arg_max_act]
# plt.figure()
# plt.gray()
# plt.imshow(es0[:,:,arg_max_act], interpolation='nearest')
# plt.figure()
# plt.imshow(rs[0]/rs[0].max(), interpolation='nearest')
# plt.show()
# return
# rs = net.predict(xs)
# rs[0][rs[0]>1.] = 1.
# plt.figure(1)
# plt.subplot(2,3,1)
# plt.title('Original')
# plt.imshow(xs[0], interpolation='nearest')
# plt.axis('off')
# plt.subplot(2,3,2)
# plt.title('Reconstruction')
# plt.imshow(rs[0], interpolation='nearest')
# plt.axis('off')
# plt.gray()
# plt.subplot(2,3,4)
# plt.title('Diff - R')
# plt.imshow(np.abs(rs[0]-xs[0])[:,:,0], interpolation='nearest')
# plt.axis('off')
# plt.subplot(2,3,5)
# plt.title('Diff - G')
# plt.imshow(np.abs(rs[0]-xs[0])[:,:,1], interpolation='nearest')
# plt.axis('off')
# plt.subplot(2,3,6)
# plt.title('Diff - B')
# plt.imshow(np.abs(rs[0]-xs[0])[:,:,2], interpolation='nearest')
# plt.axis('off')
# plt.show()
W_1 = net.layers[1].W.eval()
W_1n = (W_1-W_1.min())/(W_1.max()-W_1.min())
plt.figure()
for i in range(W_1.shape[3]):
plt.subplot(4,3,i+1)
plt.imshow(W_1n[:,:,:,i], interpolation='none')
plt.show()
def mitos12ClassifierFromConvAutoEncoder3(train=False, resumeTraining=False, iterations=20000):
autoEncoderModel = {
'inputLayer' : ImageInputLayer(width=64,height=64,channels=3),
'hiddenLayers' : [
ConvolutionLayer(kernelsize=5, channels=3, features=12, stride=2, weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[64,64,3]),
ConvolutionLayer(kernelsize=5, channels=12, features=48, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[32,32,12]),
ConvolutionLayer(kernelsize=5, channels=48, features=192, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[16,16,48])
]
}
classifierModel = {
'inputLayer': ImageInputLayer(width=8,height=8,channels=192),
'hiddenLayers': [
ImageToVectorLayer(imagesize=(8,8,192)),
FullyConnectedLayer(inputsize=8*8*192,outputsize=200,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
FullyConnectedLayer(inputsize=200,outputsize=100,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh)
],
'outputLayer' : FullyConnectedLayer(inputsize=100,outputsize=2,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.softmax)
}
batch_size = 50
autoEncoder = Network(autoEncoderModel, objective='reconstruction', batch_size=batch_size)
clf = Network(classifierModel, objective='classification', batch_size=batch_size)
clf.setupTraining("cross-entropy", "Adam")
init = tf.initialize_all_variables()
basedir = "/media/sf_E_DRIVE/Dropbox/ULB/Doctorat/ImageSet/MITOS12/"
mitos12 = MITOS12Data(train_dirs=[os.path.join(basedir,d) for d in ["A00_v2", "A01_v2", "A02_v2", "A03_v2", "A04_v2"]])
aesaver = tf.train.Saver(autoEncoder.getVariables())
clfsaver = tf.train.Saver(clf.getVariables())
sess = tf.Session()
with sess.as_default():
assert tf.get_default_session() is sess
sess.run(init)
aesaver.restore(sess, "/home/adrien/workspace/DeepNet/mitos12ConvAutoEncoder3.ckpt")
if train==False or resumeTraining==True:
clfsaver.restore(sess, "/home/adrien/workspace/DeepNet/mitos12ClfFromConvAutoEncoder3.ckpt")
if train==True:
for i in range(iterations):
batch = mitos12.next_supervised_batch(batch_size)
input_images = [b[0] for b in batch]
batch_xs = autoEncoder.encode(input_images)
batch_ys = [b[1] for b in batch]
if i%1000==0:
cost = clf.cost.eval(feed_dict={clf.x: batch_xs, clf.target: batch_ys})
print "step %d, training cost %g"%(i,cost)
save_path = clfsaver.save(sess, "/home/adrien/workspace/DeepNet/mitos12ClfFromConvAutoEncoder3.ckpt")
clf.train(batch_xs, batch_ys)
save_path = clfsaver.save(sess, "/home/adrien/workspace/DeepNet/mitos12ClfFromConvAutoEncoder3.ckpt")
# Eval :
Cmat = np.zeros((2,2))
for i in range(50):
batch = mitos12.next_supervised_batch(batch_size)
input_images = [b[0] for b in batch]
batch_xs = autoEncoder.encode(input_images)
batch_ys = [b[1] for b in batch]
pred = clf.predict(batch_xs)
Cmat += C(pred,batch_ys)
print Cmat
def mitos12ConvAutoEncoder4(train=False, resumeTraining=False, iterations=20000):
model = {
'inputLayer' : ImageInputLayer(width=128,height=128,channels=3),
'hiddenLayers' : [
ConvolutionLayer(kernelsize=15, channels=3, features=12, stride=4, weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[128,128,3]),
ConvolutionLayer(kernelsize=7, channels=12, features=40, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[32,32,12]),
ConvolutionLayer(kernelsize=5, channels=40, features=80, stride=2,weightInitFunc=WeightInit.truncatedNormal, biasInitFunc=WeightInit.positive, activationFunc=tf.nn.relu, inputshape=[16,16,40])
]
}
batch_size = 50
net = Network(model, objective='reconstruction', batch_size=batch_size)
net.setupTraining("squared-diff", "Adam", a=0.998)
autoEncoderName = "mitos12ConvAutoEncoder4WithDistortion"
init = tf.initialize_all_variables()
basedir = "/media/sf_E_DRIVE/Dropbox/ULB/Doctorat/ImageSet/MITOS12/"
mitos12 = MITOS12Data(train_dirs=[os.path.join(basedir,d) for d in ["A00_v2", "A01_v2", "A02_v2", "A03_v2", "A04_v2"]],chunksize=(128,128))
saver = tf.train.Saver()
sess = tf.Session()
with sess.as_default():
assert tf.get_default_session() is sess
sess.run(init)
if train==False or resumeTraining==True:
saver.restore(sess, "/home/adrien/workspace/DeepNet/%s.ckpt"%autoEncoderName)
if train==True:
for i in range(iterations):
batch_xs = mitos12.next_batch(batch_size, noise=True, nc=0.02)
batch_ys = batch_xs
if i%1000 == 0:
cost = net.cost.eval(feed_dict={net.x: batch_xs, net.target: batch_ys})
loss = net.loss.eval(feed_dict={net.x: batch_xs, net.target: batch_ys})
l2loss = net.l2loss.eval()
print "step %d, training cost %g, loss %g, l2loss %g"%(i,cost,loss,l2loss)
save_path = saver.save(sess, "/home/adrien/workspace/DeepNet/%s.ckpt"%autoEncoderName)
with open("/home/adrien/workspace/DeepNet/%s_results.txt"%autoEncoderName, "a") as resFile:
resFile.write("step %d, training cost %g, loss %g, l2loss %g\n"%(i,cost,loss,l2loss))
net.train(batch_xs, batch_ys)
save_path = saver.save(sess, "/home/adrien/workspace/DeepNet/%s.ckpt"%autoEncoderName)
xs = mitos12.next_batch(batch_size)
cost = net.cost.eval(feed_dict={net.x: xs, net.target: xs})
loss = net.loss.eval(feed_dict={net.x: xs, net.target: xs})
l2loss = net.l2loss.eval()
print "test cost %g, loss %g, l2loss %g"%(cost,loss,l2loss)
es = net.encode(xs)
# for idx in range(40):
W = net.layers[1].W.eval()
W = (W-W.min())/(W.max()-W.min())
plt.figure()
for idx in range(12):
plt.subplot(4,3,idx+1)
plt.axis('off')
plt.imshow(W[:,:,:,idx], interpolation='nearest', vmin=-1, vmax=1)
# plt.show()
# return
# for idx in range(len(es)):
# max_act = 0
# arg_max_act = 0
# es0 = es[idx]
# for t in np.arange(0,160):
# if( es0[:,:,t].sum() > max_act ):
# max_act = es0[:,:,t].sum()
# arg_max_act = t
# print arg_max_act
# return
# featt = np.zeros(es0.shape)
# featt[:,:,arg_max_act] = es0[:,:,arg_max_act]
# rs = net.decode([featt])
# print es0[:,:,arg_max_act]
# plt.figure()
# plt.gray()
# plt.imshow(es0[:,:,arg_max_act], interpolation='nearest')
# plt.figure()
# plt.imshow(rs[0]/rs[0].max(), interpolation='nearest')
# plt.show()
# return
rs = net.predict(xs)
rs[0][rs[0]>1.] = 1.
plt.figure()
plt.subplot(2,3,1)
plt.title('Original')
plt.imshow(xs[0], interpolation='nearest')
plt.axis('off')
plt.subplot(2,3,2)
plt.title('Reconstruction')
plt.imshow(rs[0], interpolation='nearest')
plt.axis('off')
plt.gray()
plt.subplot(2,3,4)
plt.title('Diff - R')
plt.imshow(np.abs(rs[0]-xs[0])[:,:,0], interpolation='nearest')
plt.axis('off')
plt.subplot(2,3,5)
plt.title('Diff - G')
plt.imshow(np.abs(rs[0]-xs[0])[:,:,1], interpolation='nearest')
plt.axis('off')
plt.subplot(2,3,6)
plt.title('Diff - B')
plt.imshow(np.abs(rs[0]-xs[0])[:,:,2], interpolation='nearest')
plt.axis('off')
plt.show()
def visuNetwork(n=0):
model = {
'inputLayer' : FlatInputLayer(inputsize=64*64*3),
'hiddenLayers' : [
FullyConnectedLayer(inputsize=64*64*3,outputsize=32*32,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
FullyConnectedLayer(inputsize=32*32,outputsize=16*16,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
FullyConnectedLayer(inputsize=16*16,outputsize=8*8,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
]
}
savedModelPath = "/home/adrien/workspace/DeepNet/mitos12FullyConnectedAE.ckpt"
batch_size = 200
net = Network(model, objective='reconstruction', batch_size=batch_size)
net.setupTraining("squared-diff", "Adam", False)
mitos12 = MITOS12Data(train_dirs=["/media/sf_VirtualDropbox"])
varsToRestore = []
for l in net.layers:
varsToRestore += l.trainables
saver = tf.train.Saver(varsToRestore)
sess = tf.Session()
init = tf.initialize_all_variables()
with sess.as_default():
assert tf.get_default_session() is sess
sess.run(init)
saver.restore(sess, savedModelPath)
batch_xs = mitos12.next_batch(batch_size, flat=True)
ys = net.encode(batch_xs)
print ys.shape
plt.figure(1)
plt.imshow(ys, vmin=ys.min(), vmax=ys.max(), interpolation='nearest')
plt.gray()
plt.show()
return
#x = tf.Variable(tf.truncated_normal([1,64*64*3], stddev=0.1))
# x = np.ones((1,64*64*3))*0.5
# batch_size = 1
# net = Network(model, objective='reconstruction', batch_size=batch_size)
# net.setupTraining("squared-diff", "Adam", False)
# varsToRestore = []
# for l in net.layers:
# varsToRestore += l.trainables
# mitos12 = MITOS12Data(train_dirs=["/media/sf_VirtualDropbox"])
# saver = tf.train.Saver(varsToRestore)
# sess = tf.Session()
# init = tf.initialize_all_variables()
# with sess.as_default():
# assert tf.get_default_session() is sess
# sess.run(init)
# saver.restore(sess, savedModelPath)
# W = net.layers[3].W.eval()
# print W.shape, W.min(), W.max()
# plt.figure(1)
# plt.imshow(W, vmin=W.min(), vmax=W.max())
# plt.gray()
# plt.show()
# return
# target = [0. for i in range(64)]
# target[n] = 1.
# print target
# y = net.encoded
# loss = y-target
# grad = tf.gradients(loss, net.x)
# for i in range(100):
# gradx = grad[0].eval(feed_dict={net.x: x})
# if i%100 == 0:
# print gradx
# x += gradx*0.1
# fig = plt.figure(1)
# plt.imshow(x.reshape([64,64,3]))
# fig.savefig('neuron_%d_max_input.png'%(n), bbox_inches='tight')
# print "Saved figure"
def mitos12FullyConnectedAE(train=False, resumeTraining=False, iterations=20000):
model = {
'inputLayer' : FlatInputLayer(inputsize=64*64*3),
'hiddenLayers' : [
FullyConnectedLayer(inputsize=64*64*3,outputsize=64*64,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
FullyConnectedLayer(inputsize=64*64,outputsize=32*32,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
FullyConnectedLayer(inputsize=32*32,outputsize=16*16,weightInitFunc=WeightInit.truncatedNormal,biasInitFunc=WeightInit.positive,activationFunc=tf.nn.tanh),
]
}
savedModelPath = "/home/adrien/workspace/DeepNet/mitos12FullyConnectedAE.ckpt"
batch_size = 50
net = Network(model, objective='reconstruction', batch_size=batch_size)
net.setupTraining("squared-diff", "Adam", True)
init = tf.initialize_all_variables()
mitos12 = MITOS12Data(train_dirs=["/media/sf_VirtualDropbox"])
saver = tf.train.Saver()
sess = tf.Session()
with sess.as_default():
assert tf.get_default_session() is sess
sess.run(init)
if train==False or resumeTraining==True:
print "Restoring from "+savedModelPath
saver.restore(sess, savedModelPath)
if train==True:
for i in range(iterations):
batch_xs = mitos12.next_batch(batch_size, flat=True)
batch_ys = batch_xs
if i%1000 == 0:
cost = net.cost.eval(feed_dict={net.x: batch_xs, net.target: batch_ys})
print "step %d, training cost %g"%(i,cost)
save_path = saver.save(sess,savedModelPath)
net.train(batch_xs, batch_ys)
save_path = saver.save(sess,savedModelPath)
im = np.array(mitos12.getRandomImage())
print im.shape
cols = im.shape[0]/64
rows = im.shape[1]/64
xs_1 = [im[i*64:i*64+64,j*64:j*64+64,:].flatten() for i in xrange(cols) for j in xrange(rows)]
xs_2 = [im[32+i*64:32+i*64+64,32+j*64:32+j*64+64,:].flatten() for i in xrange(cols) for j in xrange(rows)]
rs_1 = net.predict(xs_1)
rs_2 = net.predict(xs_2)
rim = np.zeros(im.shape)
im_k = np.zeros((im.shape[0],im.shape[1]))
for k,r in enumerate(rs_1):
i = (k/cols)*64
j = (k%cols)*64
rim[i:i+64,j:j+64,:] += r.reshape((64,64,3))
im_k[i:i+64,j:j+64] += 1
for k,r in enumerate(rs_2):
i = (k/cols)*64
j = (k%cols)*64
rim[32+i:32+i+64,32+j:32+j+64,:] += r.reshape((64,64,3))
im_k[32+i:32+i+64,32+j:32+j+64] += 1
rim[im_k>0,0] /= im_k[im_k>0]
rim[im_k>0,1] /= im_k[im_k>0]
rim[im_k>0,2] /= im_k[im_k>0]
rim[rim>1.] = 1.
rim[rim<0.] = 0.
plt.figure(0)
plt.subplot(1,2,1)
plt.imshow(im, interpolation='nearest')
plt.axis('off')
plt.subplot(1,2,2)
plt.imshow(rim, interpolation='nearest')
plt.axis('off')
plt.show()
