def write_activations(model_file, x, write_dir):
if not os.path.exists(write_dir):
os.mkdir(write_dir)
xn = normalize(x, np.max(x), np.max(x) - np.min(x))
tiff.imsave(os.path.join(write_dir, 'input.tif'),
(xn * 255).astype('uint8'))
# transform data to cuda tensor
x = torch.FloatTensor(x[np.newaxis, np.newaxis, ...]).cuda()
# load network
net = load_net(model_file=model_file)
net.eval()
net.cuda()
# apply forward prop and extract network activations
encoder_outputs, encoded_outputs = net.encoder(x)
decoder_outputs, final_outputs = net.decoder(encoded_outputs,
encoder_outputs)
# write random activations
for i, encoder_output in enumerate(encoder_outputs):
c = np.random.randint(encoder_output.size(1))
act = encoder_output[0, c, :, :].data.cpu().numpy()
act = normalize(act, np.max(act), np.max(act) - np.min(act))
tiff.imsave(
os.path.join(
write_dir,
'enc_act_' + str(len(encoder_outputs) - i - 1) + '.tif'),
(act * 255).astype('uint8'))
c = np.random.randint(encoded_outputs.size(1))
act = encoded_outputs[0, c, :, :].data.cpu().numpy()
act = normalize(act, np.max(act), np.max(act) - np.min(act))
tiff.imsave(
os.path.join(write_dir,
'enc_act_' + str(len(encoder_outputs)) + '.tif'),
(act * 255).astype('uint8'))
for i, decoder_output in enumerate(decoder_outputs):
c = np.random.randint(decoder_output.size(1))
act = decoder_output[0, c, :, :].data.cpu().numpy()
act = normalize(act, np.max(act), np.max(act) - np.min(act))
tiff.imsave(
os.path.join(
write_dir,
'dec_act_' + str(len(decoder_outputs) - i - 1) + '.tif'),
(act * 255).astype('uint8'))
def __init__(self, data_path, input_shapes, len_epoch=1000, preprocess='z', transform=None, dtype='uint8'):
self.data_path = data_path
self.input_shapes = input_shapes
self.len_epoch = len_epoch
self.transform = transform
self.data = read_tif(data_path, dtype=dtype)
mu, std = self.get_stats()
self.mu = mu
self.std = std
self.preprocess = preprocess
if preprocess == 'z':
self.data = normalize(self.data, mu, std)
elif preprocess == 'unit':
self.data = normalize(self.data, 0, 255)
def __init__(self,
data_path,
label_path,
input_shape,
mode='image-level',
split=None,
train=None,
len_epoch=1000,
preprocess='z',
transform=None,
target_transform=None,
dtypes=('uint8', 'uint8')):
self.data_path = data_path
self.label_path = label_path
self.input_shape = input_shape
self.mode = mode
self.len_epoch = len_epoch
self.transform = transform
self.target_transform = target_transform
self.data = read_tif(data_path, dtype=dtypes[0])
self.labels = read_tif(label_path, dtype=dtypes[1])
mu, std = self.get_stats()
self.mu = mu
self.std = std
self.preprocess = preprocess
if preprocess == 'z':
self.data = normalize(self.data, mu, std)
elif preprocess == 'unit':
self.data = normalize(self.data, 0, 255)
self.labels = normalize(self.labels, 0, 255)
if split is not None:
if train:
s = int(split * self.data.shape[2])
self.data = self.data[:, :, :s]
self.labels = self.labels[:, :, :s]
else:
s = int(split * self.data.shape[2])
self.data = self.data[:, :, s:]
self.labels = self.labels[:, :, s:]
def __init__(self, data_path, label_path, input_shape, len_epoch=1000, preprocess='z', transform=None, target_transform=None, dtypes=('uint8','uint8')):
self.data_path = data_path
self.label_path = label_path
self.input_shape = input_shape
self.len_epoch = len_epoch
self.transform = transform
self.target_transform = target_transform
self.data = read_tif(data_path, dtype=dtypes[0])
self.labels = read_tif(label_path, dtype=dtypes[1])
mu, std = self.get_stats()
self.mu = mu
self.std = std
self.preprocess = preprocess
if preprocess == 'z':
self.data = normalize(self.data, mu, std)
elif preprocess == 'unit':
self.data = normalize(self.data, 0, 255)
self.labels = normalize(self.labels, 0, 255)
args.input_size = [int(item) for item in args.input_size.split(',')]
"""
Setup writing directory
"""
print('[%s] Setting up write directories' % (datetime.datetime.now()))
if args.write_dir is not None:
if not os.path.exists(args.write_dir):
os.mkdir(args.write_dir)
"""
Load and normalize the data
"""
print('[%s] Loading and normalizing the data' % (datetime.datetime.now()))
test_data = read_tif(args.data, dtype='uint8')
mu = np.mean(test_data)
std = np.std(test_data)
test_data = normalize(test_data, mu, std)
if len(test_data.shape) < 3:
test_data = test_data[np.newaxis, ...]
"""
Load the network
"""
print('[%s] Loading network' % (datetime.datetime.now()))
net = load_net(args.net)
"""
Segmentation
"""
print('[%s] Starting segmentation' % (datetime.datetime.now()))
segmentation = segment_pixels(test_data,
net,
args.input_size,
batch_size=args.batch_size,