def cache_train(bands):
train_shapes = shapes[~shapes['image_id'].isin(train_wkt['ImageId'].unique(
))]
print('num_train_images = ', train_shapes.shape[0])
min_train_height = train_shapes['height'].min()
min_train_width = train_shapes['width'].min()
# num_train = train_shapes.shape[0]
num_train = 50
image_rows = min_train_height
image_cols = min_train_width
num_channels = bands
num_mask_channels = 11
f = h5py.File(os.path.join(data_path, 'unlabeled_train_3.h5'), 'w')
imgs = f.create_dataset('train',
(num_train, num_channels, image_rows, image_cols),
dtype=np.float16)
# imgs_mask = f.create_dataset('train_mask', (num_train, num_mask_channels, image_rows, image_cols), dtype=np.uint8)
ids = []
i = 0
for image_id in tqdm(sorted(train_shapes['image_id'].unique())):
# drop the test image
if image_id == '6110_3_1':
continue
if image_id in informative_set:
image = extra_functions.read_image(image_id, bands)
_, height, width = image.shape
imgs[i] = image[:, :min_train_height, :min_train_width]
# imgs_mask[i] = read_mask(image_id, height, width)[:, :min_train_height, :min_train_width] / 255
ids += [image_id]
i += 1
if i == 36:
break
# fix from there: https://github.com/h5py/h5py/issues/441
f['train_ids'] = np.array(ids).astype('|S9')
f.close()
def cache_train():
fp_files = []
for (dirpath, dirnames, filenames) in walk(data_path):
fp_files.extend(filenames)
break
print("Number of Training images: ", len(fp_files))
print("Processing...")
num_channels = 3
num_mask_channels = 1
image_rows = 3705
image_cols = 4800
num_train = len(fp_files)
f = h5py.File(os.path.join(data_path, 'train_16.h5'), 'w')
imgs = f.create_dataset('train',
(num_train, image_rows, image_cols, num_channels),
dtype=np.float16)
imgs_mask = f.create_dataset(
'train_mask', (num_train, image_rows, image_cols, num_mask_channels),
dtype=np.uint8)
ids = []
i = 0
for image_id in sorted(fp_files):
print(image_id)
image = extra_functions.read_image(image_id)
mask = extra_functions.read_mask(image_id)
height, width, _ = image.shape
imgs[i] = image
imgs_mask[i] = mask
ids += [image_id]
i += 1
# fix from there: https://github.com/h5py/h5py/issues/441
f['train_ids'] = np.array(ids).astype('|S9')
f.close()
print("Training Images Cached Successfully.")
t[t > b] = b
out[:, :, i] = t
return out.astype(np.uint8)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.gpu_options.visible_device_list = gpu_id
sess = tf.Session(config=config)
for image_id in test_ids:
# Print image name
print(image_id)
# Read in the testing image
image = read_image(image_id, bands=input_channels)
c, h, w = image.shape
h = int(int(h / img_rows) * img_rows)
w = int(int(w / img_cols) * img_cols)
# Make prediction ( Also for flip and swap_axis)
predicted_mask_p1 = make_prediction(model, model_path, image, input_size=(img_rows, img_cols),
crop=cropping, num_masks=output_classes, num_channels=input_channels, sess=sess)
predicted_mask_p2 = make_prediction(model, model_path, image[:, int(img_rows/2):h-int(img_rows/2), int(img_cols/2):w-int(img_cols/2)], input_size=(img_rows, img_cols), crop=cropping, num_masks=output_classes, num_channels=input_channels, sess=sess)
p2_mask = np.zeros((11, img_rows, img_cols))
for i in range(img_rows):
if i < img_rows / 2 - 1:
j = int(img_cols / 2) - 2 - i
def __init__(self,
image_ids,
img_rows,
img_cols,
num_channels,
cropping,
horizontal_flip=False,
vertical_flip=False,
swap_axis=False,
mask=-1):
self.CLASSES = {
1: 'Building',
2: 'Structure',
3: 'Road',
4: 'Track',
5: 'Trees',
6: 'Crops',
7: 'Fast_H20',
8: 'Slow_H20',
9: 'Truck',
10: 'Car',
11: 'Background'
}
self.image_ids = image_ids
train_wkt = pd.read_csv(os.path.join('../data', 'train_wkt_v4.csv'))
shapes = pd.read_csv(os.path.join('../data', '3_shapes.csv'))
train_shapes = shapes[shapes['image_id'].isin(
train_wkt['ImageId'].unique())]
min_train_height = train_shapes['height'].min()
min_train_width = train_shapes['width'].min()
self.y_train = np.zeros((24, 1, min_train_height, min_train_width))
self.X_train = np.zeros((24, 3, min_train_height, min_train_width))
for i in range(len(image_ids)):
image_id = image_ids[i]
# X_train
image = extra_functions.read_image(image_id, 3)
self.X_train[:] = image[:, :min_train_height, :min_train_width]
# y_train
_, height, width = image.shape
self.y_train[:] = self.read_mask(
image_id, height, width,
mask)[:, :min_train_height, :min_train_width] / 255
print("Loaded image {}".format(str(i)))
self.y_train = np.reshape(self.y_train,
[24, 1, min_train_height, min_train_width])
self.horizontal_flip = horizontal_flip
self.vertical_flip = vertical_flip
self.swap_axis = swap_axis
self.num_channels = num_channels
self.img_rows = img_rows
self.img_cols = img_cols
self.cropping = cropping
if mask == -1:
# set this value as variable later
self.num_mask_channels = 11
else:
self.num_mask_channels = 1
num_channels=input_channels,
cropping=cropping,
horizontal_flip=horizontal_flip,
vertical_flip=vertical_flip,
swap_axis=swap_axis,
mask=mask,
unlabeled=True)
# initialize trainer
trainer = Trainer(model,
batch_size=batch_size,
verification_batch_size=veri_batch_size,
norm_grads=False,
optimizer=optimizer,
mask=mask,
test_x=read_image(veri_image_id, input_channels),
test_y=read_GT_Masks(veri_image_id, output_classes,
mask),
save_image_dims=save_image_dims,
opt_kwargs=op_param)
# start training
trainer.train(
data,
unlabeled_data,
model_save_path,
training_iters=batch_pass,
epochs=nb_epoch,
dropout=dropout, # probability to keep units
display_step=display_step,
restore=restore,