def data():
"""
Here we load the training set and the validation set with their labels without any data augmentation
In fact, we don't want data augmentation to influence the optimisation of hyperparameters
:return: training set, validation set, training labels, validation labels
"""
baseDir = r"D:\Arnaud\data_croutinet\ottawa\data"
trainDir = os.path.join(baseDir, "train/train.csv")
validationDir = os.path.join(baseDir, "validation/validation.csv")
trainLeft, trainRight, trainLabels = loadAsScalars(trainDir)
validationLeft, validationRight, validationLabels = loadAsScalars(
validationDir)
X_train = [trainLeft, trainRight]
y_train = trainLabels
X_test = [validationLeft, validationRight]
y_test = validationLabels
return X_train, X_test, y_train, y_test
#Define the img size
from keras.models import load_model
IMG_SIZE = 224
INPUT_DIM = (IMG_SIZE, IMG_SIZE, 3)
#Define directories
baseDir = r"D:\Arnaud\data_croutinet\ottawa\data"
trainDir = os.path.join(baseDir, "validation/validation.csv")
roads_loubna_dir = os.path.join(baseDir, "roads_loubna")
models_dir = os.path.join(baseDir, "models")
model = load_model(os.path.join(models_dir, "modelWithDataAugmentation5.h5"))
left, right, labels, namesLeft, namesRight = loadAsScalars(trainDir)
croutipointsLeft = [
CroutiPoint(namesLeft[i], left[i]) for i in range(len(namesLeft))
]
croutipointsRight = [
CroutiPoint(namesRight[i], right[i]) for i in range(len(namesRight))
]
# Training left points
for i in range(len(croutipointsLeft)):
print(i)
contenders = [rd.randint(0, len(croutipointsLeft) - 1) for k in range(30)]
predictions = model.predict([
np.array([croutipointsLeft[i].pixels for k in range(30)]),
np.array([croutipointsLeft[contenders[k]].pixels for k in range(30)])
def plotScoresAndHeatmap():
validationLeft, validationRight, validationLabels, namesLeft, namesRight = loadAsScalars(
validationDir)
validationLeft_score = model.predict(validationLeft)
validationLeft_score = validationLeft_score * (-1)
validationRight_score = model.predict(validationRight)
validationRight_score = validationRight_score * (-1)
prediction = []
correct = 0
for i in range(len(validationLeft_score)):
if validationLeft_score[i] > validationRight_score[i]:
if validationLabels[i] == 0:
prediction.append("prédiction vraie")
correct += 1
else:
prediction.append("prédiction fausse")
else:
if validationLabels[i] == 0:
prediction.append("prédiction fausse")
else:
prediction.append("prédiction vraie")
correct += 1
dict = {0: "left winner", 1: "right winner"}
for i in range(len(validationLabels)):
plt.figure()
leftImage = validationLeft[i]
rightImage = validationRight[i]
leftPath = os.path.join(roads_loubna_dir, namesLeft[i])
rightPath = os.path.join(roads_loubna_dir, namesRight[i])
heatmapLeft = heatmap(leftImage, leftPath)
heatmapRight = heatmap(rightImage, rightPath)
cv2.imwrite(
os.path.join(
activation_dir, "validationSetLeft" + str(i) + "score" +
str(validationLeft_score[i]) + ".jpg"), heatmapLeft)
cv2.imwrite(
os.path.join(
activation_dir, "validationSetRight" + str(i) + "score" +
str(validationRight_score[i]) + ".jpg"), heatmapRight)
plt.subplot(2, 2, 1)
plt.title("score : " + str(validationLeft_score[i]))
plt.imshow(loadImage(leftPath))
plt.suptitle(dict[validationLabels[i]] + " : " + prediction[i],
fontsize=16)
plt.subplot(2, 2, 2)
plt.title("score : " + str(validationRight_score[i]))
plt.imshow(loadImage(rightPath))
plt.subplot(2, 2, 3)
plt.imshow(
loadImage(
os.path.join(
activation_dir, "validationSetLeft" + str(i) + "score" +
str(validationLeft_score[i]) + ".jpg")))
plt.subplot(2, 2, 4)
plt.imshow(
loadImage(
os.path.join(
activation_dir, "validationSetRight" + str(i) + "score" +
str(validationRight_score[i]) + ".jpg")))
plt.savefig(os.path.join(check_dir, "validationSet" + str(i) + ".jpg"))
INPUT_DIM = (IMG_SIZE, IMG_SIZE, 3)
#Define directories
baseDir = r"D:\Arnaud\data_croutinet\ottawa\data"
trainDir = os.path.join(baseDir, "train/train.csv")
validationDir = os.path.join(baseDir, "validation/validation.csv")
testDir = os.path.join(baseDir, "test/test.csv")
roads_loubna_dir = os.path.join(baseDir, "roads_loubna")
ranking_dir = os.path.join(baseDir, "rankingNoSigmoid")
activation_dir = os.path.join(ranking_dir, "activation")
check_dir = os.path.join(ranking_dir, "checkdata")
models_dir = os.path.join(baseDir, "models")
model = load_model(os.path.join(models_dir, "scoreNetworkNoSigmoid.h5"))
validationLeft, validationRight, validationLabels, namesLeft, namesRight = loadAsScalars(
validationDir)
def mutliplePredict(data, nb_predict):
"""
Compute multipes predict on same dataset
:param data: the array of puictures
:param nb_predict: numbre of predictions that you wannt
:return: the mean of predictions for each value
"""
scores_array = []
for i in range(nb_predict):
scores_array.append(model.predict(data))
scores_array = np.array(scores_array)
#Define the img size
IMG_SIZE = 224
INPUT_DIM = (IMG_SIZE, IMG_SIZE, 3)
#Define directories
baseDir = r"D:\Arnaud\data_croutinet\ottawa\data"
trainDir = os.path.join(baseDir, "train/train.csv")
validationDir = os.path.join(baseDir, "validation/validation.csv")
base_network_save = os.path.join(baseDir, "scoreNetworkRetrain2.h5")
ranking_network_save = os.path.join(baseDir, "rankingNetworkRetrain.h5")
base_network_save2 = os.path.join(baseDir, "scoreNetworkRetrain3.h5")
#load training and validation set with labels as scalars between 0 and 1
trainLeft, trainRight, trainLabels = loadAsScalars(trainDir)
validationLeft, validationRight, validationLabels = loadAsScalars(
validationDir)
#Here is the architecture of ScoreCroutinet that we create below
base_network = load_model(base_network_save)
model = create_meta_network(INPUT_DIM, base_network)
#We fit the model to the training set
history = model.fit([trainLeft, trainRight],
trainLabels,
batch_size=16,
epochs=30,
validation_data=([validationLeft,
validationRight], validationLabels))
duelsDF['left_id'] = roads_loubna_dir + "/" + duelsDF['left_id']
duelsDF['right_id'] = roads_loubna_dir + "/" + duelsDF['right_id']
#print(duelsDF)
mask_yes = duelsDF['winner'] == '1'
yes = duelsDF[mask_yes]
mask_no = duelsDF['winner'] == '0'
no = duelsDF[mask_no]
# Here we load the hyperas optimiszed version of ScoreCroutinet
base_network = create_base_network(INPUT_DIM)
# On top of it we add the Ranking one who teach ScoreCroutinet how to score
model = create_meta_network(INPUT_DIM, base_network)
validationLeft, validationRight, validationLabels = loadAsScalars(
validationDir)
# For batch training, the number of iterations of training model
n_iter = 150
for iteration in range(n_iter):
print(iteration / n_iter)
# sample positive and negative cases for current iteration. It is faster to use fit on batch of n yesno and augment
# that batch using datagen_class_aug_test than to use fit_generator with the datagen_class_aug_test and small batch
# sizes.
yesno = yes.sample(500).append(no.sample(500))
print('yesno created')
labels = dict(
zip([str(x) for x in yesno.index.tolist()],
[1 if x == '1' else 0 for x in yesno.winner.tolist()]))
print('labels created')