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, 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)
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
class MultitaskBD(Layer):
"""Text to image and back."""
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.ToVis = Dense(self.size, self.size_out)
self.FromVis = Dense(self.size_out, self.size)
self.Decode = 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.params = params(self.Embed, self.Encode, self.ToVis, self.FromVis, self.Decode, self.PredictT)
def __call__(self, inp, out_prev, img):
img_out = self.visual_activation(self.ToVis(last(self.Encode(self.Embed(inp)))))
txt_out = softmax3d(self.Embed.unembed(self.PredictT(self.Decode(self.visual_activation(self.FromVis(img)),
self.Embed(out_prev)))))
return (img_out, txt_out)
def predictor_v(self):
"""Return function to predict image vector from input."""
input = T.imatrix()
return theano.function([input], self.visual_activation(self.ToVis(last(self.Encode(self.Embed(input))))))
class MultitaskMM(Layer):
"""Shared recurrent encoder with visual decoder + textual decoder."""
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 __call__(self, inp, out_prev, _img):
rep = last(self.Encode(self.Embed(inp)))
img = self.visual_activation(self.DecodeV(rep))
txt = softmax3d(self.Embed.unembed(self.PredictT(self.DecodeT(rep, self.Embed(out_prev)))))
return (img, txt)
def predictor_v(self):
"""Return function to predict image vector from input."""
input = T.imatrix()
return theano.function([input], self.visual_activation(self.DecodeV(last(self.Encode(self.Embed(input))))))
class MultitaskY(Layer):
"""Joint Encode + separate pathways."""
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.Joint = 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, out_depth=self.out_depth,
gru_activation=self.gru_activation, dropout_prob=self.dropout_prob)
self.Textual = textual(self.size, self.size, self.depth,
gru_activation=self.gru_activation, dropout_prob=self.dropout_prob)
self.params = params(self.Embed, self.Visual, self.Textual)
def __call__(self, inp, output_prev, _img):
inp_e = self.Joint(self.Embed(inp))
output_prev_e = self.Embed(output_prev)
img = self.visual_activation(self.Visual(inp_e))
txt = softmax3d(self.Embed.unembed(self.Textual(inp_e, output_prev_e, _img)))
return (img, txt)
def predictor_v(self):
"""Return function to predict image vector from input."""
input = T.imatrix()
return theano.function([input], self.visual_activation(self.Visual(self.Embed(input))))
class Multitask(Layer):
"""Visual encoder combined with a textual task."""
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)
def __call__(self, inp, output_prev, img):
inp_e = self.Embed(inp)
output_prev_e = self.Embed(output_prev)
img_pred = self.visual_activation(self.Visual(inp_e))
txt_pred = softmax3d(self.Embed.unembed(self.Textual(inp_e, output_prev_e, img)))
return (img_pred, txt_pred)
def predictor_v(self):
"""Return function to predict image vector from input."""
input = T.imatrix()
return theano.function([input], self.visual_activation(self.Visual(self.Embed(input))))
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)
class MultitaskLMY(Layer):
"""Alternative visual encoder combined with a textual decoder.
Textual decoder starts from final state of encoder instead of from
image. Shared hidden layer plus specialized layers.
"""
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 params(self):
return params(self.Embed, self.Shared, self.Visual, self.LM, self.ToTxt)
def __call__(self, inp, output_prev, _img):
shared = self.Shared(self.Embed(inp))
img_pred = self.Visual(shared)
txt_pred = softmax3d(self.Embed.unembed(self.ToTxt(self.LM(last(shared), self.Embed(output_prev)))))
return (img_pred, txt_pred)
def predictor_v(self):
"""Return function to predict image vector from input."""
input = T.imatrix()
return theano.function([input], self.Visual(self.Shared(self.Embed(input))))
def predictor_r(self):
"""Return function to predict representation from input."""
input = T.imatrix()
return theano.function([input], last(self.Shared(self.Embed(input))))
class MultitaskLMD(Layer):
"""Alternative visual encoder combined with a textual decoder.
Textual decoder starts from final state of encoder instead of from image.
"""
def __init__(self,
size_vocab,
size_embed,
size,
size_out,
depth,
gru_activation=clipped_rectify,
visual_activation=linear,
visual_encoder=StackedGRUH0,
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 = StackedGRU(self.size_embed,
self.size,
self.depth,
activation=self.gru_activation,
dropout_prob=self.dropout_prob)
self.ToTxt = Dense(self.size, self.size_embed) # try direct softmax
def params(self):
return params(self.Embed, self.Visual, self.LM, self.ToTxt)
def __call__(self, inp, output_prev, _img):
rep = self.Visual.encode(self.Embed(inp))
img_pred = self.Visual.visual_activation(self.Visual.ToImg(rep))
txt_pred = softmax3d(
self.Embed.unembed(
self.ToTxt(self.LM(rep, self.Embed(output_prev)))))
return (img_pred, txt_pred)
def predictor_v(self):
"""Return function to predict image vector from input."""
input = T.imatrix()
return theano.function([input], self.Visual(self.Embed(input)))
def predictor_r(self):
"""Return function to predict representation from input."""
input = T.imatrix()
return theano.function([input], self.Visual.encode(self.Embed(input)))
def __init__(self, size_vocab, size_embed, size, size_out, depth,
gru_activation=clipped_rectify,
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,
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_embed) # map to embeddings
class MultitaskLMC(Layer):
"""Visual encoder combined with a textual decoder."""
def __init__(
self,
size_vocab,
size_embed,
size,
size_out,
depth,
gru_activation=clipped_rectify,
visual_activation=linear,
visual_encoder=StackedGRUH0,
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 = StackedGRU(
self.size_embed, self.size, self.depth, activation=self.gru_activation, dropout_prob=self.dropout_prob
)
self.FromImg = Dense(self.size_out, self.size)
self.ToTxt = Dense(self.size, self.size_embed) # try direct softmax
def params(self):
return params(self.Embed, self.Visual, self.LM, self.FromImg, self.ToTxt)
def __call__(self, inp, output_prev, img):
img_pred = self.Visual(self.Embed(inp))
txt_pred = softmax3d(self.Embed.unembed(self.ToTxt(self.LM(self.FromImg(img), self.Embed(output_prev)))))
return (img_pred, txt_pred)
def predictor_v(self):
"""Return function to predict image vector from input."""
input = T.imatrix()
return theano.function([input], self.Visual(self.Embed(input)))
def predictor_r(self):
"""Return function to predict representation from input."""
input = T.imatrix()
return theano.function([input], self.Visual.encode(self.Embed(input)))
