MycroftAI
/
mycroft-precise
mirror of https://github.com/MycroftAI/mycroft-precise.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

Fix partial ModelParams documentation

pull/147/head
Matthew D. Scholefield 2020-04-20 09:17:33 -05:00
parent 13db6f01a2
commit ba2e34c9f6
1 changed files with 11 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
precise/model.py
View File

@ -29,17 +29,20 @@ if TYPE_CHECKING:
class ModelParams: class ModelParams:
""" """
Attributes: Attributes:
recurrent_units: recurrent_units: Number of GRU units. Higher values increase computation
dropout: but allow more complex learning. Too high of a value causes overfitting
extra_metrics: Whether to include false positive and false negative metrics dropout: Reduces overfitting but can potentially decrease accuracy if too high
extra_metrics: Whether to include false positive and false negative metrics while training
skip_acc: Whether to skip accuracy calculation while training skip_acc: Whether to skip accuracy calculation while training
loss_bias: Near 1.0 reduces false positives. See <set_loss_bias>
freeze_till: Layer number from start to freeze after loading (allows for partial training)
""" """
recurrent_units = attr.ib(20) # type: int recurrent_units = attr.ib(20) # type: int
dropout = attr.ib(0.2) # type: float dropout = attr.ib(0.2) # type: float
extra_metrics = attr.ib(False) # type: bool extra_metrics = attr.ib(False) # type: bool
skip_acc = attr.ib(False) # type: bool skip_acc = attr.ib(False) # type: bool
loss_bias = attr.ib(0.7) # type: float loss_bias = attr.ib(0.7) # type: float
freeze_till = attr.ib(0) # type: bool freeze_till = attr.ib(0) # type: int
def load_precise_model(model_name: str) -> Any: def load_precise_model(model_name: str) -> Any:
@ -73,7 +76,8 @@ def create_model(model_name: Optional[str], params: ModelParams) -> 'Sequential'
model = Sequential() model = Sequential()
model.add(GRU( model.add(GRU(
params.recurrent_units, activation='linear', params.recurrent_units, activation='linear',
input_shape=(pr.n_features, pr.feature_size), dropout=params.dropout, name='net' input_shape=(
pr.n_features, pr.feature_size), dropout=params.dropout, name='net'
)) ))
model.add(Dense(1, activation='sigmoid')) model.add(Dense(1, activation='sigmoid'))
@ -82,5 +86,6 @@ def create_model(model_name: Optional[str], params: ModelParams) -> 'Sequential'
set_loss_bias(params.loss_bias) set_loss_bias(params.loss_bias)
for i in model.layers[:params.freeze_till]: for i in model.layers[:params.freeze_till]:
i.trainable = False i.trainable = False
model.compile('rmsprop', weighted_log_loss, metrics=(not params.skip_acc) * metrics) model.compile('rmsprop', weighted_log_loss,
metrics=(not params.skip_acc) * metrics)
return model return model