def _create_callbacks():
os.makedirs(ccp.output('weights'), exist_ok=True)
return [
CSVLogger(ccp.output('keras_history.csv'), append=True),
ModelCheckpoint(
ccp.output("weights/weights.{epoch:02d}-{val_loss:.2f}.hdf5")),
TensorBoard(log_dir=ccp.output('tensorboard')),
DensityCheckpoint("data/shakecam/shakeshack-1504543773.jpg")
]
def handle(self, *args, **kwargs):
now = int(time())
source = "https://cdn.shakeshack.com/camera.jpg?{}".format(now)
destination = ccp.output("shakeshack-{}.jpg".format(now))
with _fetch_image(source) as data:
if len(data) > 0:
_upload(data, "shakeshack/{}".format(basename(destination)))
def _predict_images(self, image_keys, save=False):
for image_key in image_keys:
prediction = self.predictor.predict_line(ml.load_img(image_key))
truth = self._get_truth(image_key)
if save:
dest = ccp.output("predictions/{}".format(os.path.basename(image_key)))
self.previewer.save(dest, image_key, prediction, truth)
else:
self.previewer.show(image_key, prediction, truth)
if input("Continue? [y]/n: ") == 'n':
break
class DensityCheckpoint(Callback):
image_key = attr.ib()
output_dir = attr.ib(default=ccp.output("prediction_checkpoint/{}".format(
int(time.time()))))
def on_train_begin(self, logs=None):
self._save_prediction("begin")
def on_epoch_end(self, epoch, logs=None):
self._save_prediction("epoch_{:03}".format(epoch))
def _save_prediction(self, label):
x = ml.image_to_batch(ml.load_img(self.image_key))
y = self.model.predict(x, batch_size=1)
self._save_image(label, y)
def _save_image(self, label, y):
os.makedirs(self.output_dir, exist_ok=True)
destination = "{}.jpg".format(os.path.join(self.output_dir, label))
previewer.save(destination, self.image_key, Prediction(y, y.sum()))
class LineCountCheckpoint(Callback):
image_key = attr.ib()
output_dir = attr.ib(default=ccp.output("prediction_checkpoint/{}".format(
int(time.time()))))
def on_train_begin(self, logs=None):
self._save_prediction("begin")
def on_epoch_end(self, epoch, logs=None):
self._save_prediction("epoch_{:03}".format(epoch))
def _save_prediction(self, name):
truth = dm.generate_truth_batch(self.image_key, True)
inline = float(self.model.predict(truth, batch_size=1))
print("checkpoint linecount prediction: {}".format(inline))
self._save_result(name, truth, inline)
def _save_result(self, name, y, inline):
os.makedirs(self.output_dir, exist_ok=True)
destination = "{}.jpg".format(os.path.join(self.output_dir, name))
previewer.save(destination, self.image_key, Prediction(y, line=inline))
def _write_images_to_tmp(self, anns):
os.makedirs(ccp.output("previews/"), exist_ok=True)
for key in anns.keys():
previewer.save(ccp.output("previews/{}.jpg".format(_index_from_key(key))), key)
def _create_callbacks():
os.makedirs(ccp.output('weights/linecount'), exist_ok=True)
return [CSVLogger(ccp.output('keras_history.csv'), append=True),
ModelCheckpoint(ccp.output("weights/linecount/weights.{epoch:03d}-{val_loss:.2f}.hdf5")),
TensorBoard(log_dir=ccp.output('tensorboard')),
callbacks.LineCountCheckpoint("data/shakecam/shakeshack-1500859164.jpg")]