def taste():
logDir = "../log"
resetLog = False
if resetLog:
if tf.gfile.Exists(logDir):
tf.gfile.DeleteRecursively(logDir)
model = Model_C(28, 28, 1, nbCategories=2, favoritism=(1, 5))
model.verbose = True
now = time.time()
summary_writer = tf.summary.FileWriter(logDir + "/" + str(now),
model.sess.graph)
lossTrain = []
lossValid = []
accuracyTrain = []
accuracyValide = []
try:
nbStep = 2000
for itr in range(nbStep):
print("itr:", itr)
X, Y = batchOneVignette(50)
model.fit(X, Y)
lossTrain.append(model.loss)
accuracyTrain.append(model.accuracy)
summary_writer.add_summary(model.summary, global_step=itr)
if itr % 20 == 0:
print("\nVALIDATION")
X, Y = batchOneVignette(50)
model.validate(X, Y)
lossValid.append(model.loss)
accuracyValide.append(model.accuracy)
print("\n")
else:
lossValid.append(np.nan)
accuracyValide.append(np.nan)
except KeyboardInterrupt:
print("on a stoppé")
X, Y = batchOneVignette(16)
hat_Y, hat_Y_proba = model.predict(X)
draw16imgs(Y, "Y")
draw16imgs(hat_Y, "hat Y")
plt.show()
model.close()
def tasteClassif():
logDir = "../log"
reset = True
if reset and tf.gfile.Exists(logDir):
tf.gfile.DeleteRecursively(logDir)
img_size = 28
multi_label = False
if multi_label:
dataDealer = lambda batchSize: batchSeveralVignette(batchSize)
nbCat = 3
else:
dataDealer = lambda batchSize: batchOneVignette(batchSize)
nbCat = 3
#Xs, Ys_cat,Ys_background=dataDealer(1)
#print("Ys_cat.shape:",Ys_cat.shape)
model = Model_A(multi_label, img_size, img_size, 1, nbCat)
model.verbose = True
batchSize = 50
model.learning_rate = 1e-3
now = time.time()
summary_writer = tf.summary.FileWriter(logDir + "/" + str(now),
model.sess.graph)
nbStep = 2000
lossTrain = []
lossValid = []
accuracyTrain = []
accuracyValide = []
try:
for itr in range(nbStep):
print("itr:", itr)
X, Y = dataDealer(batchSize)
model.fit(X, Y)
lossTrain.append(model.loss)
accuracyTrain.append(model.accuracy)
summary_writer.add_summary(model.summary, global_step=itr)
if itr % 20 == 0:
print("\nVALIDATION")
X, Y = dataDealer(batchSize)
model.validate(X, Y)
lossValid.append(model.loss)
accuracyValide.append(model.accuracy)
print("\n")
else:
lossValid.append(np.nan)
accuracyValide.append(np.nan)
except KeyboardInterrupt:
print("on a stoppé")
plt.subplot(1, 2, 1)
plt.plot(accuracyTrain, label="train")
plt.plot(accuracyValide, '.', label="valid")
plt.title("accuracy")
plt.legend()
plt.subplot(1, 2, 2)
plt.plot(lossTrain, label="train")
plt.plot(lossValid, '.', label="valid")
plt.title("loss")
plt.legend()
""" TEST"""
X, Y = dataDealer(16)
hat_Y = model.predict(X)
if multi_label:
subTitles = (hat_Y > 0.5).astype(np.int32)
else:
subTitles = list(zip(np.argmax(Y, axis=1), np.argmax(hat_Y, axis=1)))
print(subTitles)
draw16imgs(X, "X", 0, 1, cmap="gray", subTitles=subTitles)
plt.show()
model.close()
plt.title(str(label))
plt.show()
if __name__=="__main__":
#X,Y=batchOneVignette(4)
X,Y=batchSeveralVignette(16)
print("X.shape",X.shape)
print("Y.shape",Y.shape)
print(Y[0,:])
print(Y[1, :])
print(Y[2, :])
print(Y[3, :])
draw16imgs(X,subTitles=Y)
plt.show()
Y = Conv2D(1, 1)(Y)
return Y
import tensorflow as tf
if __name__ == '__main__':
patch_size = 32
X, Y = None, None
for X, Y in dataGenerator_croix_sur_ile(batch_size=16,
patch_size=patch_size):
break
print("X.shape:", X.shape)
print("Y.shape:", Y.shape)
_X = tf.constant(X)
_out = Model_interpolation_2D(patch_size=patch_size, dropout=1)(_X)
print("out.get_shape():", _out.get_shape())
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
out = sess.run(_out)
print("out.shape:", out.shape)
draw16imgs(out[:, :, :, 0])
#get_keras_model(32, dropout= 0.2)
def taste():
logDir = "../log"
reset = True
if reset and tf.gfile.Exists(logDir):
tf.gfile.DeleteRecursively(logDir)
img_size = 28
nbStrates = 6
"""attention Y est d'ordre 4 : chaque strate représentant une catégorie = une instance ou le background"""
dataDealer = lambda batchSize: oneBatch_of_rect_stratify(
either__any_disjoint_separated=0,
img_size=img_size,
batchSize=16,
nbRectPerImg=2,
deltaRange=(8, 12),
nbStrates=nbStrates)
""" img """
nbChannel = 1
model = Model_G(img_size, img_size, nbChannel, nbStrates, 2, 32)
#model=Model_fullyConv_regCat(img_size,img_size,nbChannel,nbStrates,None,30,60)
model.verbose = True
model.nbConsecutiveOptForOneFit = 1
batchSize = 50
model.learning_rate = 1e-2
now = time.time()
summary_writer_train = tf.summary.FileWriter(logDir + "/" + str(now),
model.sess.graph)
nbStep = 2000
lossTrain = []
lossValid = []
try:
for itr in range(nbStep):
print("itr:", itr)
X, Y_cat, Y_background = dataDealer(batchSize)
model.fit(X, Y_cat, Y_background, itr)
lossTrain.append(model.loss)
summary_writer_train.add_summary(model.summary, global_step=itr)
if itr % 20 == 0:
print("\nVALIDATION")
X, Y_cat, Y_background = dataDealer(batchSize)
model.validate(X, Y_cat, Y_background, itr)
lossValid.append(model.loss)
print("\n")
else:
lossValid.append(model.loss)
except KeyboardInterrupt:
print("on a stoppé")
plt.subplot(1, 2, 2)
plt.plot(lossTrain, label="train")
plt.plot(lossValid, '.', label="valid class")
plt.title("loss")
plt.legend()
""" TEST"""
X, Y_cat, Y_background = dataDealer(16)
hat_Y_cat, hat_Y_cat_sum = model.predict(X)
print("hat_Y_cat.shape", hat_Y_cat.shape)
draw16imgs(X, "X", vmin=0, vmax=1, cmap="gray")
#draw16imgs(hat_Y_background)
draw16imgs(hat_Y_cat_sum, "hat cat", 0, nbStrates, "jet")
draw16imgs(np.sum(Y_cat, axis=3), "Y_cat", 0, nbStrates, "jet")
# draw_proba_Vs_cat=1
# if draw_proba_Vs_cat==0:
# for i in range(0,nbCat):
# drawOne(hat_Y_test_proba[:, :, :,i],"proba cat:"+str(i-1),1)
# else:
# drawOne(hat_Y_test_cat[:, :, :], "hat_cat", nbCat)
plt.show()
model.close()
noise_gauss = np.random.normal(loc=0,
scale=0.2,
size=[batchSize, img_size, img_size, 1])
#noise_salt= (np.random.random(size=[batchSize, img_size, img_size,1])>0.2)
Xs += noise_gauss
#Xs[noise_salt]=0
return Xs, Ys_strate, Ys_background
if __name__ == "__main__":
nbStrates = 10
Xs, Ys_strate, Ys_background = oneBatch_of_rect_stratify(
either__any_disjoint_separated=0,
img_size=28,
batchSize=16,
nbRectPerImg=1,
deltaRange=(7, 18),
nbStrates=nbStrates)
draw16imgs(Xs, cmap="gray")
draw16imgs(np.sum(Ys_strate, axis=3))
draw16imgs(np.argmax(Ys_background, axis=3))
plt.show()
def tastetRot():
X = np.arange(100).reshape((10, 10))
X_rot = ndi.interpolation.rotate(X, angle=45, reshape=False)
X_zoom = ndi.zoom(X, zoom=2)
print(X_rot.shape)
print(X_zoom.shape)
plt.imshow(X_rot)
if __name__ == "__main__":
xs = []
ys = []
for x, y in oneEpoch(50, False):
xs.append(x)
ys.append(y)
xs = np.array(xs)
ys = np.array(ys)
print(xs.shape)
print(ys.shape)
decal = 60
draw16imgs(xs[decal:], cmap="gray")
draw16imgs(np.expand_dims(ys[decal:], 3), vmin=0, vmax=3)
plt.show()
for i in range(batch_size):
img,gt=createImgOneVignette()
imgs[i,:,:,0]=img
gts[i,:,:]=gt
return imgs,gts
if __name__=="__main__":
X,Y=batchOneVignette(16)
draw16imgs(X,"X",cmap="gray")
draw16imgs(Y,"Y")
plt.show()