def sample(self):
if hasattr(self, '_sample'):
return self._sample, self._sample_size
self._sample = {}
self._sample_size = 0
_, n = self.positive
positives = self._positive_expanded
n = int(n * self.ratio)
TT.debug("Collecting", n, "random samples.")
pixels_per_image = int(np.prod(self.image_size))
indices = xrange(len(self.files) * pixels_per_image)
ignored = 0
for index in random.sample(indices, n):
data_file, label_file = self.files[index / pixels_per_image]
if data_file not in self._sample:
self._sample[data_file] = []
pixel = index % pixels_per_image
p = 0.0
if data_file in positives and pixel in positives[data_file]:
p = 1.0
ignored += 1
col, row = pixel_at_index(pixel, self.image_size)
self._sample[data_file].append([col, row, p])
self._sample_size += 1
TT.debug(ignored, "samples out of", self._sample_size,
"random samples are positive.")
self.positive_in_sample = ignored
return self.sample
def on_dataset_epoch_end(self, epoch, logs={}):
numpy.savetxt(open(self.log_file, 'a'), [[self.old_epochs + self.epoch, self.loss]], fmt="%g")
if self.last_loss > self.loss:
filename = self.weights_file % (self.old_epochs + epoch)
TT.debug("Saving weights to", filename)
self.model.save_weights(filename)
self.last_loss = min(self.last_loss, self.loss)
def icpr2012():
"""
Filter dataset and labels.
:return:
"""
TT.debug("Loading configurations for ICPR 2012.")
def filename_filter(name):
"""
Filter dataset files.
Return True if file is in dataset.
:type name: str
:rtype: bool
"""
return re.compile(r'.+\.bmp').search(name)
def mapper(name):
"""
Map dataset with its labels.
:type name: str
:rtype: str
"""
return re.compile(r'\.[a-z]+$').sub('.csv', name)
return filename_filter, mapper
def dump(self):
_ = self.data # Create data if not already created.
TT.debug("Current dataset has", self._dataset_size, "images.",
self._positive_size, "positive and", self._sample_size, "negative.")
json.dump({'data': self._dataset, 'size': self._dataset_size,
'positive_size': self._positive_size + self.positive_in_sample,
'sample_size': self._sample_size - self.positive_in_sample},
open(self.dataset_store_path, 'w'))
def files(self):
if not hasattr(self, '_files'):
self._files = list_all_files(self.root_path,
filename_filter=self.filename_filter,
mapper=self.label_mapper)
TT.debug("Found", len(self._files), "matching files in",
self.root_path)
return self._files
def on_dataset_epoch_end(self, epoch, logs={}):
numpy.savetxt(open(self.log_file, 'a'),
[[self.old_epochs + self.epoch, self.loss]],
fmt="%g")
if self.last_loss > self.loss:
filename = self.weights_file % (self.old_epochs + epoch)
TT.debug("Saving weights to", filename)
self.model.save_weights(filename)
self.last_loss = min(self.last_loss, self.loss)
def __init__(self, root_path, patch_size=(101, 101), verbose=False, ratio=1.0, name='dataset', mapper=None,
filename_filter=None, rotation=True):
TT.debug("Dataset root path set to:", root_path)
self.name = name
self.patch_size = patch_size
self.ratio = ratio
self.root_path = os.path.abspath(root_path)
self.verbose = verbose
self.label_mapper = mapper
self.filename_filter = filename_filter
self.rotation = rotation
def task_train_filter(args):
ff, mapper = getattr(__import__('dataset'), args.dataset)()
dataset = Dataset(root_path=args.path,
verbose=args.verbose,
name='base-model',
mapper=mapper,
filename_filter=ff,
rotation=False)
dataset_batches = BatchGenerator(dataset, args.batch)
from mitosis import model_base
TT.debug("Compile base model.")
model = model_base(args.lr)
model_saved_weights_path = os.path.join(args.path,
'base-model.weights.npy')
if os.path.exists(model_saved_weights_path):
TT.info("Loading weights from %s" % model_saved_weights_path)
model.load_weights(model_saved_weights_path)
train_start = time.time()
log = LearnLog("filter", args.path)
for epoch in xrange(args.epoch):
TT.debug(epoch + 1, "of", args.epoch, "epochs")
log.on_dataset_epoch_begin(epoch + 1)
for x, y in dataset_batches:
model.fit(x,
y,
batch_size=args.mini_batch,
nb_epoch=1,
validation_split=.1,
callbacks=[log],
show_accuracy=True,
shuffle=True)
log.on_dataset_epoch_end(epoch + 1)
log.on_dataset_train_end()
TT.success("Training finished in %.2f hours." %
((time.time() - train_start) / 3600.))
def load(self):
if os.path.exists(self.dataset_store_path):
TT.debug("Loading dataset from", self.dataset_store_path)
data = json.load(open(self.dataset_store_path))
self._dataset = data['data']
self._dataset_size = data['size']
self._positive = {}
self._positive_size = data['positive_size']
self._sample = {}
self._sample_size = data['sample_size']
self.positive_in_sample = 0
TT.debug("Current dataset has", self._dataset_size, "images.",
self._positive_size, "positive and", self._sample_size, "negative.")
return True
return False
def __iter__(self):
data = Queue(self.MAX_NUM)
def append(dst, pool, item):
if item is not None:
pool.append(item)
if len(pool) < min(self.pool_size, self.batch_size):
return dst, pool
if dst is None:
return np.asarray(pool, dtype=np.float64), []
if len(pool):
return np.concatenate((dst, pool)), []
return dst, []
def produce():
i = 1
count = 0
data_x = data_y = None
pool_x = []
pool_y = []
for x, y in self.dataset:
data_x, pool_x = append(data_x, pool_x, x)
data_y, pool_y = append(data_y, pool_y, (y, 1 - y))
count += 1
if count >= self.batch_size:
data_x, pool_x = append(data_x, pool_x, None)
data_y, pool_y = append(data_y, pool_y, None)
data.put([data_x, data_y])
i += 1
count = 0
data_x = data_y = None
if count > 0:
data_x, pool_x = append(data_x, pool_x, None)
data_y, pool_y = append(data_y, pool_y, None)
data.put([data_x, data_y])
start_new_thread(produce, ())
i = 1
while i <= self.n:
start = time.clock()
X, Y = data.get()
if self.verbose:
TT.debug("batch", i, "of", self.n, "completed in",
time.clock() - start, "seconds. This batch has",
int(np.sum(Y[:, 0])), "positive pixels and",
int(np.sum(Y[:, 1])), "negative pixels.")
yield X, Y
i += 1
def positive(self):
if hasattr(self, '_positive'):
return self._positive, self._positive_size
TT.debug("Collecting positive samples.")
self._positive = {}
self._positive_size = 0
self._positive_expanded = {}
for data_file, label_file in self.files:
labels = load_csv(os.path.join(self.root_path, label_file))
self._positive[data_file] = labels
self._positive_size += len(labels)
self._positive_expanded[data_file] = {}
for col, row, p in labels:
self._positive_expanded[data_file][index_at_pixel(col=col, row=row, size=self.image_size)] = p
TT.debug("Found", self._positive_size, "positive samples.")
return self.positive
def data(self):
if hasattr(self, '_dataset'):
return self._dataset, self._dataset_size
if self.load():
return self.data
TT.debug("Creating new dataset.")
pos, pos_c = self.positive
sam, sam_c = self.sample
for filename in pos:
if filename not in sam:
sam[filename] = pos[filename]
else:
sam[filename] += pos[filename]
self._dataset = sam
self._dataset_size = sam_c + pos_c
self.dump()
return self.data
def main():
parser, args = parse_args()
TT.verbose = args.verbose
if args.task == 'train-filter':
TT.debug("Running: Task Train Filter")
task_train_filter(args)
if args.task == 'train-cnn':
TT.debug("Running: Task Train CNN")
task_train_cnn(args)
elif args.task == 'test-filter':
TT.debug("Running: Task Test Filter")
task_test_filter(args)
elif args.task == 'test-cnn':
TT.debug("Running: Task Test CNN")
task_test_cnn(args)
else:
parser.print_help()
exit(0)
def main():
parser, args = parse_args()
TT.verbose = args.verbose
if args.task == 'train-filter':
TT.debug("Running: Task Train Filter")
task_train_filter(args)
if args.task == 'train-cnn':
TT.debug("Running: Task Train CNN")
task_train_cnn(args)
elif args.task == 'test-filter':
TT.debug("Running: Task Test Filter")
task_test_filter(args)
elif args.task == 'test-cnn':
TT.debug("Running: Task Test CNN")
task_test_cnn(args)
else:
parser.print_help()
exit(0)
def task_train_filter(args):
ff, mapper = getattr(__import__('dataset'), args.dataset)()
dataset = Dataset(root_path=args.path, verbose=args.verbose, name='base-model',
mapper=mapper, filename_filter=ff, rotation=False)
dataset_batches = BatchGenerator(dataset, args.batch)
from mitosis import model_base
TT.debug("Compile base model.")
model = model_base(args.lr)
model_saved_weights_path = os.path.join(args.path, 'base-model.weights.npy')
if os.path.exists(model_saved_weights_path):
TT.info("Loading weights from %s" % model_saved_weights_path)
model.load_weights(model_saved_weights_path)
train_start = time.time()
log = LearnLog("filter", args.path)
for epoch in xrange(args.epoch):
TT.debug(epoch + 1, "of", args.epoch, "epochs")
log.on_dataset_epoch_begin(epoch + 1)
for x, y in dataset_batches:
model.fit(x, y, batch_size=args.mini_batch, nb_epoch=1, validation_split=.1,
callbacks=[log], show_accuracy=True, shuffle=True)
log.on_dataset_epoch_end(epoch + 1)
log.on_dataset_train_end()
TT.success("Training finished in %.2f hours." % ((time.time() - train_start) / 3600.))
def task_test_filter(args):
dataset = ImageIterator(args.input, args.output)
dataset_batches = BatchGenerator(dataset, args.batch)
from mitosis import model_base
TT.debug("Compile base model.")
model = model_base(args.lr)
model_saved_weights_path = os.path.join(args.path, 'base-model.weights.npy')
TT.info("Loading weights from %s" % model_saved_weights_path)
model.load_weights(model_saved_weights_path)
test_start = time.time()
out = None
for x, y in dataset_batches:
tmp = model.predict(x, args.mini_batch, args.verbose)
out = np_append(out, tmp)
width, height = dataset.image_size
out = numpy.reshape(out[:, 0], (height, width))
numpy.save(change_ext(args.input, 'predicted.npy'), out)
numpy.save(change_ext(args.input, 'expected.npy'), dataset.output)
TT.success("Testing finished in %.2f minutes." % ((time.time() - test_start) / 60.))
def task_test_filter(args):
dataset = ImageIterator(args.input, args.output)
dataset_batches = BatchGenerator(dataset, args.batch)
from mitosis import model_base
TT.debug("Compile base model.")
model = model_base(args.lr)
model_saved_weights_path = os.path.join(args.path, 'base-model.weights.npy')
TT.info("Loading weights from %s" % model_saved_weights_path)
model.load_weights(model_saved_weights_path)
test_start = time.time()
out = None
for x, y in dataset_batches:
tmp = model.predict(x, args.mini_batch, args.verbose)
out = np_append(out, tmp)
width, height = dataset.image_size
out = numpy.reshape(out[:, 0], (height, width))
numpy.save(change_ext(args.input, 'predicted.npy'), out)
numpy.save(change_ext(args.input, 'expected.npy'), dataset.output)
TT.success("Testing finished in %.2f minutes." % ((time.time() - test_start) / 60.))
def on_dataset_train_end(self, logs={}):
filename = self.weights_file % 0
TT.debug("Saving weights to", filename)
self.model.save_weights(filename)
def task_test_cnn(args):
dataset = ImageIterator(args.input, args.output)
dataset_batches = BatchGenerator(dataset, args.batch)
from mitosis import model_base, model_1, model_2
TT.debug("Compile base model.")
model = model_base(0)
TT.debug("Compile model 1.")
model1 = model_1(0)
TT.debug("Compile model 2.")
model2 = model_2(0)
model_saved_weights_path = os.path.join(args.path,
'base-model.weights.npy')
model1_saved_weights_path = os.path.join(args.path, 'model1.weights.npy')
model2_saved_weights_path = os.path.join(args.path, 'model2.weights.npy')
TT.info("Loading weights from %s" % model_saved_weights_path)
model.load_weights(model_saved_weights_path)
TT.info("Loading weights from %s" % model1_saved_weights_path)
model1.load_weights(model1_saved_weights_path)
TT.info("Loading weights from %s" % model2_saved_weights_path)
model2.load_weights(model2_saved_weights_path)
test_start = time.time()
out = out1 = out2 = None
for x, y in dataset_batches:
tmp = model.predict(x, args.mini_batch, args.verbose)
local1 = numpy.zeros(tmp.shape)
local2 = numpy.zeros(tmp.shape)
out = np_append(out, tmp)
x = 1. - x
x_new = []
indices = []
for i in range(len(tmp)):
if tmp[i][0] > .6:
x_new.append(x[i])
indices.append(i)
x_new = numpy.asarray(x_new)
if len(x_new):
tmp1 = model1.predict(x_new, args.mini_batch, args.verbose)
local1[indices] = tmp1
out1 = np_append(out1, local1)
if len(x_new):
tmp2 = model2.predict(x_new, args.mini_batch, args.verbose)
local2[indices] = tmp2
out2 = np_append(out2, local2)
width, height = dataset.image_size
out = numpy.reshape(out[:, 0], (height, width))
out1 = numpy.reshape(out1[:, 0], (height, width))
out2 = numpy.reshape(out2[:, 0], (height, width))
numpy.save(change_ext(args.input, 'predicted.npy'), out)
numpy.save(change_ext(args.input, 'model1.predicted.npy'), out1)
numpy.save(change_ext(args.input, 'model2.predicted.npy'), out2)
numpy.save(change_ext(args.input, 'expected.npy'), dataset.output)
TT.success("Testing finished in %.2f minutes." %
((time.time() - test_start) / 60.))
def task_train_cnn(args):
ff, mapper = getattr(__import__('dataset'), args.dataset)()
dataset = Dataset(root_path=args.path, verbose=args.verbose, name='cnn',
mapper=mapper, filename_filter=ff, ratio=9)
dataset_batches = BatchGenerator(dataset, args.batch)
from mitosis import model_base, model_1, model_2
TT.debug("Compile base model.")
model = model_base(lr=0)
TT.debug("Compile model 1.")
model1 = model_1(args.lr)
TT.debug("Compile model 2.")
model2 = model_2(args.lr)
model_saved_weights_path = os.path.join(args.path, 'base-model.weights.npy')
model1_saved_weights_path = os.path.join(args.path, 'model1.weights.npy')
model2_saved_weights_path = os.path.join(args.path, 'model2.weights.npy')
if os.path.exists(model_saved_weights_path):
TT.info("Loading weights from %s" % model_saved_weights_path)
model.load_weights(model_saved_weights_path)
if os.path.exists(model1_saved_weights_path):
TT.info("Loading weights from %s" % model1_saved_weights_path)
model1.load_weights(model1_saved_weights_path)
if os.path.exists(model2_saved_weights_path):
TT.info("Loading weights from %s" % model2_saved_weights_path)
model2.load_weights(model2_saved_weights_path)
train_start = time.time()
log1 = LearnLog("model1", args.path)
log2 = LearnLog("model2", args.path)
for epoch in xrange(args.epoch):
TT.debug(epoch + 1, "of", args.epoch, "epochs")
log1.on_dataset_epoch_begin(epoch + 1)
log2.on_dataset_epoch_begin(epoch + 1)
for x, y in dataset_batches:
outputs = model.predict(x, batch_size=args.mini_batch, verbose=args.verbose)
# Multiply each window with it's prediction and then pass it to the next layer
# x_new = []
# y_new = []
x = 1. - x
for i in range(len(outputs)):
if y[i][0] < 1.:
# x_new.append(x[i])
# y_new.append(y[i])
x[i] *= outputs[i][0]
TT.debug("Model 1 on epoch %d" % (epoch + 1))
model1.fit(numpy.asarray(x), numpy.asarray(y), batch_size=args.mini_batch, nb_epoch=1, validation_split=.1,
callbacks=[log1], show_accuracy=True, shuffle=True)
TT.debug("Model 2 on epoch %d" % (epoch + 1))
model2.fit(numpy.asarray(x), numpy.asarray(y), batch_size=args.mini_batch, nb_epoch=1, validation_split=.1,
callbacks=[log2], show_accuracy=True, shuffle=True)
log1.on_dataset_epoch_end(epoch + 1)
log2.on_dataset_epoch_end(epoch + 1)
log1.on_dataset_train_end()
log2.on_dataset_train_end()
TT.success("Training finished in %.2f hours." % ((time.time() - train_start) / 3600.))
def task_train_cnn(args):
ff, mapper = getattr(__import__('dataset'), args.dataset)()
dataset = Dataset(root_path=args.path,
verbose=args.verbose,
name='cnn',
mapper=mapper,
filename_filter=ff,
ratio=9)
dataset_batches = BatchGenerator(dataset, args.batch)
from mitosis import model_base, model_1, model_2
TT.debug("Compile base model.")
model = model_base(lr=0)
TT.debug("Compile model 1.")
model1 = model_1(args.lr)
TT.debug("Compile model 2.")
model2 = model_2(args.lr)
model_saved_weights_path = os.path.join(args.path,
'base-model.weights.npy')
model1_saved_weights_path = os.path.join(args.path, 'model1.weights.npy')
model2_saved_weights_path = os.path.join(args.path, 'model2.weights.npy')
if os.path.exists(model_saved_weights_path):
TT.info("Loading weights from %s" % model_saved_weights_path)
model.load_weights(model_saved_weights_path)
if os.path.exists(model1_saved_weights_path):
TT.info("Loading weights from %s" % model1_saved_weights_path)
model1.load_weights(model1_saved_weights_path)
if os.path.exists(model2_saved_weights_path):
TT.info("Loading weights from %s" % model2_saved_weights_path)
model2.load_weights(model2_saved_weights_path)
train_start = time.time()
log1 = LearnLog("model1", args.path)
log2 = LearnLog("model2", args.path)
for epoch in xrange(args.epoch):
TT.debug(epoch + 1, "of", args.epoch, "epochs")
log1.on_dataset_epoch_begin(epoch + 1)
log2.on_dataset_epoch_begin(epoch + 1)
for x, y in dataset_batches:
outputs = model.predict(x,
batch_size=args.mini_batch,
verbose=args.verbose)
# Multiply each window with it's prediction and then pass it to the next layer
# x_new = []
# y_new = []
x = 1. - x
for i in range(len(outputs)):
if y[i][0] < 1.:
# x_new.append(x[i])
# y_new.append(y[i])
x[i] *= outputs[i][0]
TT.debug("Model 1 on epoch %d" % (epoch + 1))
model1.fit(numpy.asarray(x),
numpy.asarray(y),
batch_size=args.mini_batch,
nb_epoch=1,
validation_split=.1,
callbacks=[log1],
show_accuracy=True,
shuffle=True)
TT.debug("Model 2 on epoch %d" % (epoch + 1))
model2.fit(numpy.asarray(x),
numpy.asarray(y),
batch_size=args.mini_batch,
nb_epoch=1,
validation_split=.1,
callbacks=[log2],
show_accuracy=True,
shuffle=True)
log1.on_dataset_epoch_end(epoch + 1)
log2.on_dataset_epoch_end(epoch + 1)
log1.on_dataset_train_end()
log2.on_dataset_train_end()
TT.success("Training finished in %.2f hours." %
((time.time() - train_start) / 3600.))
def task_test_cnn(args):
dataset = ImageIterator(args.input, args.output)
dataset_batches = BatchGenerator(dataset, args.batch)
from mitosis import model_base, model_1, model_2
TT.debug("Compile base model.")
model = model_base(0)
TT.debug("Compile model 1.")
model1 = model_1(0)
TT.debug("Compile model 2.")
model2 = model_2(0)
model_saved_weights_path = os.path.join(args.path, 'base-model.weights.npy')
model1_saved_weights_path = os.path.join(args.path, 'model1.weights.npy')
model2_saved_weights_path = os.path.join(args.path, 'model2.weights.npy')
TT.info("Loading weights from %s" % model_saved_weights_path)
model.load_weights(model_saved_weights_path)
TT.info("Loading weights from %s" % model1_saved_weights_path)
model1.load_weights(model1_saved_weights_path)
TT.info("Loading weights from %s" % model2_saved_weights_path)
model2.load_weights(model2_saved_weights_path)
test_start = time.time()
out = out1 = out2 = None
for x, y in dataset_batches:
tmp = model.predict(x, args.mini_batch, args.verbose)
local1 = numpy.zeros(tmp.shape)
local2 = numpy.zeros(tmp.shape)
out = np_append(out, tmp)
x = 1. - x
x_new = []
indices = []
for i in range(len(tmp)):
if tmp[i][0] > .6:
x_new.append(x[i])
indices.append(i)
x_new = numpy.asarray(x_new)
if len(x_new):
tmp1 = model1.predict(x_new, args.mini_batch, args.verbose)
local1[indices] = tmp1
out1 = np_append(out1, local1)
if len(x_new):
tmp2 = model2.predict(x_new, args.mini_batch, args.verbose)
local2[indices] = tmp2
out2 = np_append(out2, local2)
width, height = dataset.image_size
out = numpy.reshape(out[:, 0], (height, width))
out1 = numpy.reshape(out1[:, 0], (height, width))
out2 = numpy.reshape(out2[:, 0], (height, width))
numpy.save(change_ext(args.input, 'predicted.npy'), out)
numpy.save(change_ext(args.input, 'model1.predicted.npy'), out1)
numpy.save(change_ext(args.input, 'model2.predicted.npy'), out2)
numpy.save(change_ext(args.input, 'expected.npy'), dataset.output)
TT.success("Testing finished in %.2f minutes." % ((time.time() - test_start) / 60.))
def on_dataset_train_end(self, logs={}):
filename = self.weights_file % 0
TT.debug("Saving weights to", filename)
self.model.save_weights(filename)