def __init__(self, hyperParams=None):
self.params = hyperParams
if(self.checkSavedModel()):
return
p = self.params["model"]
model = Sequential()
embed_matrix = self.params["embedding"].getWord2VecMatrix()
emb = Embedding(
embed_matrix.shape[0],
embed_matrix.shape[1],
# weights=[embed_matrix],
mask_zero=True,
learn=(int(self.params["learn_embedding"]) == 1)
)
emb.W.set_value(floatX(embed_matrix))
model.add(emb)
print "Initialized Embeddings"
srnn = SimpleRNN(
input_dim=embed_matrix.shape[1],
output_dim=int(p.get("hidden_nodes", 100)),
activation='tanh',
init='uniform',
inner_init='uniform',
# inner_activation='hard_sigmoid',
return_sequences=False,
truncate_gradient=int(p.get("depth", 3)),
)
model.add(srnn)
print "Initialized Recurrent Layer"
denseL = Dense(
input_dim=int(p.get("hidden_nodes", 100)),
output_dim=int(p.get("output_nodes", 4)),
activation='softmax',
init='uniform',
)
model.add(denseL)
print "Initialized Dense Layer"
self.model = model
self.compile()
self.model.layers[0].params = [self.model.layers[0].W]
self.saveModel()
def setEmbeddingWeights(self, embed_matrix):
# emb = Embedding(
# embed_matrix.shape[0],
# embed_matrix.shape[1],
# weights=[embed_matrix],
# mask_zero=True,
# # learn=(self.params["learn_embedding"] == 1)
# )
#
# self.model.layers[0] = emb
self.model.layers[0].W.set_value(floatX(embed_matrix))
# self.model.layers[0].set_weights([embed_matrix])
print "Changed embed layer weights!"