def __init__(self,
size_vocab,
size_embed,
size,
size_out,
depth,
gru_activation=clipped_rectify,
visual_encoder=StackedGRUH0,
visual_activation=linear,
dropout_prob=0.0):
autoassign(locals())
self.Embed = Embedding(self.size_vocab, self.size_embed)
self.Visual = Visual(self.size_embed,
self.size,
self.size_out,
self.depth,
encoder=self.visual_encoder,
gru_activation=self.gru_activation,
visual_activation=self.visual_activation,
dropout_prob=self.dropout_prob)
self.LM = StackedGRUH0(self.size_embed,
self.size,
self.depth,
activation=self.gru_activation,
dropout_prob=self.dropout_prob)
self.ToTxt = Dense(self.size, self.size_vocab) # map to vocabulary
def __init__(self,
size_vocab,
size_embed,
size,
size_out,
depth,
depth_spec=1,
visual_encoder=StackedGRUH0,
gru_activation=clipped_rectify,
visual_activation=linear,
dropout_prob=0.0):
autoassign(locals())
self.Embed = Embedding(self.size_vocab, self.size_embed)
self.Shared = StackedGRUH0(self.size_embed,
self.size,
self.depth,
activation=self.gru_activation,
dropout_prob=self.dropout_prob)
self.Visual = Visual(self.size,
self.size,
self.size_out,
self.depth_spec,
encoder=self.visual_encoder,
gru_activation=self.gru_activation,
visual_activation=self.visual_activation,
dropout_prob=self.dropout_prob)
self.LM = StackedGRU(self.size,
self.size,
self.depth_spec,
activation=self.gru_activation,
dropout_prob=self.dropout_prob)
self.ToTxt = Dense(self.size, self.size_embed) # try direct softmax
def __init__(self, size_vocab, size_embed, size, depth):
autoassign(locals())
self.Embed = Embedding(self.size_vocab, self.size_embed)
self.GRU = StackedGRUH0(self.size_embed,
self.size,
self.depth,
activation=clipped_rectify)
def __init__(self, size_vocab, size_embed, size, size_out, depth, out_depth=1, # FIXME USE THIS PARAM
gru_activation=tanh, visual_activation=linear,
dropout_prob=0.0):
autoassign(locals())
self.Embed = Embedding(self.size_vocab, self.size_embed)
self.Encode = StackedGRUH0(self.size_embed, self.size, self.depth,
activation=self.gru_activation, dropout_prob=self.dropout_prob)
self.DecodeT = StackedGRU(self.size_embed, self.size, self.depth,
activation=self.gru_activation, dropout_prob=self.dropout_prob)
self.PredictT = Dense(size_in=self.size, size_out=self.size_embed)
self.DecodeV = Dense(self.size, self.size_out)
self.params = params(self.Embed, self.DecodeT, self.PredictT, self.DecodeV)
def __init__(self, size_vocab, size_embed, size, size_out, depth, textual,
out_depth=1,
gru_activation=tanh,
visual_activation=linear,
dropout_prob=0.0):
autoassign(locals())
self.Embed = Embedding(self.size_vocab, self.size_embed)
self.Visual = Visual(self.size_embed, self.size, self.size_out, self.depth, out_depth=self.out_depth,
gru_activation=self.gru_activation, dropout_prob=self.dropout_prob)
self.Textual = textual(self.size_embed, self.size, self.size_out, self.depth,
gru_activation=self.gru_activation, dropout_prob=self.dropout_prob)
self.params = params(self.Embed, self.Visual, self.Textual)