def LoadImages(X):
tic()
Images=[]
Labels=[]
for img_name in X:
if img_name.endswith(".tif"):
img = cv2.imread(CONST.PATH+'IMAGES/' + img_name,-1)
label = cv2.imread(CONST.PATH+'LABELS2/' + img_name,-1)
img= np.expand_dims(img, axis=0)
label= np.expand_dims(label, axis=0)
if Images == []:
Images=img
Labels=label
else:
Images=np.concatenate((Images, img), axis=0).astype(np.float32)
Labels = np.concatenate((Labels, label), axis=0).astype(np.float32)
#APPEND
else:
print('Wrong Image Type Found! Matrix is saved as all Zeros!')
toc("Images Loaded.")
return Images/255, Labels
def CoreSample(model, image_names, save, output_filename, test_flag, minmax_filename):
if model==[]:
model = LoadPretrainedModel(CONST.PRETRAINED_MODEL, weights=1)
if test_flag==0:
images,labels= LoadImages(image_names)
else:
images =LoadImages_Test(image_names,output_filename) #output_filename : path of images during testing
ctr=0 #Start of Image Number
#This loop is here to limit saving a maximum of 20 images on a file. Bigger than that has memory issues.
for i in range(int(np.ceil((images.shape[0])/CONST.MAXIMAGES_PER_FILE))):
last=min(len(images),ctr+CONST.MAXIMAGES_PER_FILE)
maps = GenerateCore(model, images[ctr:last, :, :])
maps = maps.reshape(len(maps) * CONST.NUM_PIXELS, CONST.FEATURES)
if test_flag==0:
tic()
targets = CreateTargets(labels[ctr:last,:,:])
targets = targets.reshape(len(maps))
UpdateNorm(find_minmax(maps), minmax_filename)
toc('Targets Created. MinMax updated .')
ctr=ctr+CONST.MAXIMAGES_PER_FILE
#Saving normal images before training (or even testing if it has labels
if save == 1:
SaveData(maps.astype(np.float32), targets.astype(np.float32), output_filename + '_' + str(ctr) + '.h5')
#Saving test images without labels
elif save==2:
print('Implementation Removed from this file!')
else:
print('Skipped saving.')
return maps,[]
return
def SaveData(X, Y, output_filename):
tic()
f = h5py.File(output_filename, 'w')
grp = f.create_group("my_data")
grp.create_dataset('X', data=X, compression="gzip")
grp.create_dataset('Y', data=Y, compression="gzip")
f.close()
toc('Data File saved to disk.')
return
def LoadPretrainedModel(filename, weights):
tic()
if weights == 0: # 1: load from weights file, 0: load from pickle file
model = pickle.load(open('kerasmodel', 'rb'))
else:
model = VGG_16(filename)
sgd = SGD(lr=0.1, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(optimizer=sgd, loss='categorical_crossentropy')
toc("Model Loaded. Compiled.")
# pickle.dump(model,open ('kerasmodel','wb'))
return model
def LoadImages_Test(X,path):
tic()
Images=[]
for img_name in X:
if img_name.endswith(".tif"):
img = cv2.imread(path +'/'+ img_name,-1)
img= np.expand_dims(img, axis=0)
if Images == []:
Images=img
else:
Images=np.concatenate((Images, img), axis=0).astype(np.float32)
#APPEND
else:
print('Wrong Image Type Found! Matrix is saved as all Zeros!')
toc("Images Loaded.")
return Images/255
def convert_dataset(dataset,c):
train_set, valid_set= dataset[0], dataset[1]
assert (train_set[0].shape)[1] == (valid_set[0].shape)[1], \
"Number of features for train,val do not match: {} and {}.".format(train_set.shape[1], valid_set.shape[1])
def shared_dataset(data_xy, borrow=True,classification=c):
data_x, data_y = data_xy
shared_x = theano.shared(numpy.asarray(data_x, dtype=theano.config.floatX), borrow=borrow)
shared_y = theano.shared(numpy.asarray(data_y, dtype=theano.config.floatX), borrow=borrow)
if classification:
shared_y = T.cast(shared_y, 'int32')
return shared_x,shared_y
tic()
train_set_x, train_set_y = shared_dataset(train_set)
valid_set_x, valid_set_y = shared_dataset(valid_set)
# test_set_x, test_set_y = shared_dataset(test_set)
toc('Dataset Converted to shared dataset.')
num_features = (train_set[0].shape)[1]
rval = [(train_set_x, train_set_y), (valid_set_x, valid_set_y)]#, (test_set_x, test_set_y)]
return rval, num_features
def load_saved_data(folder, train, num_imgs):
if train == 1:
tic()
data_list = listdir(folder)
X_all = []
Y_all = []
# random.shuffle(data_list)
ctr = epoch_ctr % len(data_list)
file_name = data_list[ctr]
train_file = h5py.File(folder + file_name, 'r')
X = train_file['/my_data/X'][:]
Y = train_file['/my_data/Y'][:]
start = random.randint(0,
int(X.shape[0] / CONST.NUM_PIXELS) - num_imgs)
if X_all == []:
X_all = X[start * CONST.NUM_PIXELS:(start + num_imgs) *
CONST.NUM_PIXELS, :]
Y_all = Y[start * CONST.NUM_PIXELS:(start + num_imgs) *
CONST.NUM_PIXELS]
else:
Y_all = np.concatenate((Y_all, Y), axis=0)
#Uncomment Line below
X_all = Normalize(X_all)
# Y_all=np.expand_dims(Y_all,axis=1)
elif train == 0:
tic()
data_list = listdir(folder)
X_all = []
Y_all = []
# random.shuffle(data_list)
# for i in range(len(data_list)):
file_name = data_list[0] #file_ctr]
test_file = h5py.File(folder + file_name, 'r')
X = test_file['/my_data/X'][:]
Y = test_file['/my_data/Y'][:]
if X_all == []:
X_all = X
Y_all = Y
else:
X_all = np.concatenate((X_all, X), axis=0)
Y_all = np.concatenate((Y_all, Y), axis=0)
X_all = Normalize(X_all)
elif train == 2:
tic()
data_list = listdir(folder)
data_list.sort()
X_all = []
Y_all = []
# for i in range(len(data_list)):
# random.shuffle(data_list)
file_name = data_list[file_ctr]
test_file = h5py.File(folder + file_name, 'r')
X = test_file['/my_data/X'][:]
if X_all == []:
X_all = X
else:
X_all = np.concatenate((X_all, X), axis=0)
X_all = Normalize(X_all)
X = X_all
Y = (Y_all)
# tic()
# f = h5py.File('/home/exx/PycharmProjects/SAR/Norm_'+file_name, 'w')
# grp = f.create_group("my_data")
# grp.create_dataset('X', data=X, compression="gzip")
# grp.create_dataset('Y', data=Y, compression="gzip")
# f.close()
# toc('Data File saved to disk.')
toc("Data loaded and Normalized from " + folder + file_name)
# import os
# os.remove(folder+file_name)
return X, Y
def GenerateCore(model, Images):
# from keras import backend as K
tic()
num_samples = len(Images)
layers_extract = CONST.LAYER_NUMS
all_hc = np.zeros((num_samples, CONST.NUM_PIXELS, CONST.FEATURES))
layers = [model.layers[li].output for li in layers_extract]
get_feature = theano.function([model.layers[0].input], layers,
allow_input_downcast=False)
def extract_hypercolumn(instance):
# fc6_output = get_fc6([instance])
# fc7_output = get_fc7([instance, 0])
feature_maps = get_feature(instance)
hypercolumns = np.zeros((CONST.NUM_PIXELS, CONST.FEATURES))
# fc6_map= np.reshape(fc6_output, (64,64))
# fc7_map= np.reshape(fc7_output, (64,64))
# hypercolumns[:, 1473] = np.reshape(sp.misc.imresize(fc6_map, size=(224, 224), mode="F", interp='bilinear'), (50176))
# hypercolumns[:, 1474] = np.reshape(sp.misc.imresize(fc7_map, size=(224, 224), mode="F", interp='bilinear'), (50176))
original= instance[:,0, :, :] + .407
hypercolumns[:, 0] = np.reshape(original, (CONST.NUM_PIXELS))
ctr = 1
for convmap in feature_maps:
for fmap in convmap[0]:
upscaled = sp.misc.imresize(fmap, size=(CONST.IMAGE_DIM),
mode="F", interp='bilinear')
hypercolumns[:, ctr] = np.reshape(upscaled, (CONST.NUM_PIXELS))
ctr += 1
return np.asarray(hypercolumns)
print("Starting Loop")
counter = 0
for i in range(len(Images)):
Y_Channel = Images[i, :, :]
R = Y_Channel - .407
G = Y_Channel - .458
B = Y_Channel - .485
Y_Image = np.stack((R, G, B), axis=0)
Y_Image = np.expand_dims(Y_Image, axis=0).astype(np.float32)
hc = extract_hypercolumn(Y_Image)
hc = np.expand_dims(hc, axis=0)
all_hc[counter] = hc
counter += 1
if not counter % np.ceil(num_samples/10):
print(counter)
if not counter % num_samples:
break
toc("Hypercolumns Extracted.")
return all_hc
