2martens/masterthesis
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fix issue with non-checkpointable custom class object

Browse Source 
Signed-off-by: Jim Martens <github@2martens.de>
This commit is contained in:
Jim Martens
2019-06-05 12:06:05 +02:00
parent 2ce6687efc
commit a586dd8297
1 changed files with 17 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
src/twomartens/masterthesis/ssd.py
View File

@ -73,7 +73,7 @@ class SSD:
""" """
def __init__(self, mode: str, weights_path: Optional[str] = None) -> None: def __init__(self, mode: str, weights_path: Optional[str] = None) -> None:
self._model, self.predictor_sizes = \ self.model, self.predictor_sizes = \
keras_ssd300.ssd_300(image_size=IMAGE_SIZE, n_classes=N_CLASSES, keras_ssd300.ssd_300(image_size=IMAGE_SIZE, n_classes=N_CLASSES,
mode=mode, iou_threshold=IOU_THRESHOLD, top_k=TOP_K, mode=mode, iou_threshold=IOU_THRESHOLD, top_k=TOP_K,
scales=[0.07, 0.15, 0.33, 0.51, 0.69, 0.87, 1.05], scales=[0.07, 0.15, 0.33, 0.51, 0.69, 0.87, 1.05],
@ -82,7 +82,7 @@ class SSD:
# load existing weights # load existing weights
if weights_path is not None: if weights_path is not None:
self._model.load_weights(weights_path, by_name=True) self.model.load_weights(weights_path, by_name=True)
if mode == "training": if mode == "training":
# set non-classifier layers to non-trainable # set non-classifier layers to non-trainable
@ -92,12 +92,12 @@ class SSD:
'conv7_2_mbox_conf', 'conv7_2_mbox_conf',
'conv8_2_mbox_conf', 'conv8_2_mbox_conf',
'conv9_2_mbox_conf'] 'conv9_2_mbox_conf']
for layer in self._model.layers: for layer in self.model.layers:
if layer.name not in classifier_names: if layer.name not in classifier_names:
layer.trainable = False layer.trainable = False
def __call__(self, inputs: tf.Tensor, *args, **kwargs) -> tf.Tensor: def __call__(self, inputs: tf.Tensor, *args, **kwargs) -> tf.Tensor:
return self._model(inputs) return self.model(inputs)
class DropoutSSD: class DropoutSSD:
@ -110,7 +110,7 @@ class DropoutSSD:
""" """
def __init__(self, mode: str, weights_path: Optional[str] = None) -> None: def __init__(self, mode: str, weights_path: Optional[str] = None) -> None:
self._model, self.predictor_sizes = \ self.model, self.predictor_sizes = \
keras_ssd300_dropout.ssd_300_dropout(image_size=IMAGE_SIZE, keras_ssd300_dropout.ssd_300_dropout(image_size=IMAGE_SIZE,
n_classes=N_CLASSES, n_classes=N_CLASSES,
dropout_rate=DROPOUT_RATE, mode=mode, dropout_rate=DROPOUT_RATE, mode=mode,
@ -122,7 +122,7 @@ class DropoutSSD:
# load existing weights # load existing weights
if weights_path is not None: if weights_path is not None:
self._model.load_weights(weights_path, by_name=True) self.model.load_weights(weights_path, by_name=True)
if mode == "training": if mode == "training":
# set non-classifier layers to non-trainable # set non-classifier layers to non-trainable
@ -132,12 +132,12 @@ class DropoutSSD:
'conv7_2_mbox_conf', 'conv7_2_mbox_conf',
'conv8_2_mbox_conf', 'conv8_2_mbox_conf',
'conv9_2_mbox_conf'] 'conv9_2_mbox_conf']
for layer in self._model.layers: for layer in self.model.layers:
if layer.name not in classifier_names: if layer.name not in classifier_names:
layer.trainable = False layer.trainable = False
def __call__(self, inputs: tf.Tensor, *args, **kwargs) -> tf.Tensor: def __call__(self, inputs: tf.Tensor, *args, **kwargs) -> tf.Tensor:
return self._model(inputs) return self.model(inputs)
def predict(dataset: tf.data.Dataset, def predict(dataset: tf.data.Dataset,
@ -374,13 +374,13 @@ def train(dataset: tf.data.Dataset,
} }
# model # model
if use_dropout: if use_dropout:
checkpointables.update({ ssd = DropoutSSD(mode='training', weights_path=weights_path)
'ssd': DropoutSSD(mode='training', weights_path=weights_path)
})
else: else:
checkpointables.update({ ssd = SSD(mode='training', weights_path=weights_path)
'ssd': SSD(mode='training', weights_path=weights_path)
}) checkpointables.update({
'ssd': ssd.model
})
checkpointables.update({ checkpointables.update({
# optimizer # optimizer
@ -398,10 +398,12 @@ def train(dataset: tf.data.Dataset,
latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir) latest_checkpoint = tf.train.latest_checkpoint(checkpoint_dir)
checkpoint = tf.train.Checkpoint(**checkpointables) checkpoint = tf.train.Checkpoint(**checkpointables)
checkpoint.restore(latest_checkpoint) checkpoint.restore(latest_checkpoint)
# update model inside SSD object with version from checkpoint
ssd.model = checkpointables['ssd']
# input encoder # input encoder
input_encoder = ssd_input_encoder.SSDInputEncoder(IMAGE_SIZE[0], IMAGE_SIZE[1], input_encoder = ssd_input_encoder.SSDInputEncoder(IMAGE_SIZE[0], IMAGE_SIZE[1],
N_CLASSES, checkpointables['ssd'].predictor_sizes) N_CLASSES, ssd.predictor_sizes)
def _get_last_epoch(epoch_var: tf.Variable, **kwargs) -> int: def _get_last_epoch(epoch_var: tf.Variable, **kwargs) -> int:
return int(epoch_var) return int(epoch_var)