class GradientRBM(Workflow):
"""This unit produces update weights using minibatch according to the
algorithm described in
http://deeplearning.net/tutorial/rbm.html (25.11.14).
Does Gibbs sampling
cd_k: number of iterations of Gibbs sampling
"""
def __init__(self, workflow, **kwargs):
super(GradientRBM, self).__init__(workflow, **kwargs)
self.stddev = kwargs["stddev"]
self.batch_size = -1
self.mem_cpy = MemCpy(self)
self.mem_cpy.link_from(self.start_point)
self.repeater = Repeater(self)
self.repeater.link_from(self.mem_cpy)
self.decision = IterationCounter(self, max_iterations=kwargs["cd_k"])
self.decision.link_from(self.repeater)
self.bino_h = BinarizationGradH(self,
rand=kwargs.get("rand_h", prng.get()))
self.bino_h.link_attrs(self.mem_cpy, ("input", "output"))
self.bino_h.link_from(self.decision)
self.bino_h.gate_block = self.decision.complete
self.make_v = All2AllSigmoidV(self,
weights_stddev=self.stddev,
weights_transposed=True,
output_sample_shape=kwargs["v_size"])
self.make_v.link_from(self.bino_h)
self.make_v.link_attrs(self.bino_h, ("input", "output"))
self.bino_v = BinarizationGradV(self,
rand=kwargs.get("rand_v", prng.get()))
self.bino_v.link_attrs(self.make_v, ("input", "output"))
self.bino_v.link_from(self.make_v)
self.make_h = All2AllSigmoidH(self,
weights_stddev=self.stddev,
output_sample_shape=kwargs["h_size"])
self.make_h.link_attrs(self.bino_v, ("input", "output"))
self.make_h.output = self.mem_cpy.output
self.make_h.link_from(self.bino_v)
self.repeater.link_from(self.make_h)
self.end_point.link_from(self.decision)
self.end_point.gate_block = ~self.decision.complete
self.bino_h.gate_block = self.decision.complete
self.mem_cpy.link_attrs(self, "input")
self.bino_h.link_attrs(self, "batch_size")
self.bino_v.link_attrs(self, "batch_size")
self.make_v.link_attrs(self, "weights")
self.make_v.link_attrs(self, ("bias", "vbias"))
self.make_h.link_attrs(self, "weights")
self.make_h.link_attrs(self, ("bias", "hbias"))
self.link_attrs(self.make_h, "output")
self.link_attrs(self.bino_v, ("v1", "output"))
self.link_attrs(self.make_h, ("h1", "output"))
self.demand("input", "weights", "hbias", "vbias", "batch_size")
def run(self):
self.decision.reset()
super(GradientRBM, self).run()
class GradientRBM(Workflow):
"""This unit produces update weights using minibatch according to the
algorithm described in
http://deeplearning.net/tutorial/rbm.html (25.11.14).
Does Gibbs sampling
cd_k: number of iterations of Gibbs sampling
"""
def __init__(self, workflow, **kwargs):
super(GradientRBM, self).__init__(workflow, **kwargs)
self.stddev = kwargs["stddev"]
self.batch_size = -1
self.mem_cpy = MemCpy(self)
self.mem_cpy.link_from(self.start_point)
self.repeater = Repeater(self)
self.repeater.link_from(self.mem_cpy)
self.decision = IterationCounter(
self, max_iterations=kwargs["cd_k"])
self.decision.link_from(self.repeater)
self.bino_h = BinarizationGradH(
self, rand=kwargs.get("rand_h", prng.get()))
self.bino_h.link_attrs(self.mem_cpy, ("input", "output"))
self.bino_h.link_from(self.decision)
self.bino_h.gate_block = self.decision.complete
self.make_v = All2AllSigmoidV(
self, weights_stddev=self.stddev, weights_transposed=True,
output_sample_shape=kwargs["v_size"])
self.make_v.link_from(self.bino_h)
self.make_v.link_attrs(self.bino_h, ("input", "output"))
self.bino_v = BinarizationGradV(
self, rand=kwargs.get("rand_v", prng.get()))
self.bino_v.link_attrs(self.make_v, ("input", "output"))
self.bino_v.link_from(self.make_v)
self.make_h = All2AllSigmoidH(
self, weights_stddev=self.stddev,
output_sample_shape=kwargs["h_size"])
self.make_h.link_attrs(self.bino_v, ("input", "output"))
self.make_h.output = self.mem_cpy.output
self.make_h.link_from(self.bino_v)
self.repeater.link_from(self.make_h)
self.end_point.link_from(self.decision)
self.end_point.gate_block = ~self.decision.complete
self.bino_h.gate_block = self.decision.complete
self.mem_cpy.link_attrs(self, "input")
self.bino_h.link_attrs(self, "batch_size")
self.bino_v.link_attrs(self, "batch_size")
self.make_v.link_attrs(self, "weights")
self.make_v.link_attrs(self, ("bias", "vbias"))
self.make_h.link_attrs(self, "weights")
self.make_h.link_attrs(self, ("bias", "hbias"))
self.link_attrs(self.make_h, "output")
self.link_attrs(self.bino_v, ("v1", "output"))
self.link_attrs(self.make_h, ("h1", "output"))
self.demand("input", "weights", "hbias", "vbias", "batch_size")
def run(self):
self.decision.reset()
super(GradientRBM, self).run()
def __init__(self, workflow, **kwargs):
super(GradientRBM, self).__init__(workflow, **kwargs)
self.stddev = kwargs["stddev"]
self.batch_size = -1
self.mem_cpy = MemCpy(self)
self.mem_cpy.link_from(self.start_point)
self.repeater = Repeater(self)
self.repeater.link_from(self.mem_cpy)
self.decision = IterationCounter(self, max_iterations=kwargs["cd_k"])
self.decision.link_from(self.repeater)
self.bino_h = BinarizationGradH(self,
rand=kwargs.get("rand_h", prng.get()))
self.bino_h.link_attrs(self.mem_cpy, ("input", "output"))
self.bino_h.link_from(self.decision)
self.bino_h.gate_block = self.decision.complete
self.make_v = All2AllSigmoidV(self,
weights_stddev=self.stddev,
weights_transposed=True,
output_sample_shape=kwargs["v_size"])
self.make_v.link_from(self.bino_h)
self.make_v.link_attrs(self.bino_h, ("input", "output"))
self.bino_v = BinarizationGradV(self,
rand=kwargs.get("rand_v", prng.get()))
self.bino_v.link_attrs(self.make_v, ("input", "output"))
self.bino_v.link_from(self.make_v)
self.make_h = All2AllSigmoidH(self,
weights_stddev=self.stddev,
output_sample_shape=kwargs["h_size"])
self.make_h.link_attrs(self.bino_v, ("input", "output"))
self.make_h.output = self.mem_cpy.output
self.make_h.link_from(self.bino_v)
self.repeater.link_from(self.make_h)
self.end_point.link_from(self.decision)
self.end_point.gate_block = ~self.decision.complete
self.bino_h.gate_block = self.decision.complete
self.mem_cpy.link_attrs(self, "input")
self.bino_h.link_attrs(self, "batch_size")
self.bino_v.link_attrs(self, "batch_size")
self.make_v.link_attrs(self, "weights")
self.make_v.link_attrs(self, ("bias", "vbias"))
self.make_h.link_attrs(self, "weights")
self.make_h.link_attrs(self, ("bias", "hbias"))
self.link_attrs(self.make_h, "output")
self.link_attrs(self.bino_v, ("v1", "output"))
self.link_attrs(self.make_h, ("h1", "output"))
self.demand("input", "weights", "hbias", "vbias", "batch_size")
def __init__(self, workflow, **kwargs):
super(NNWorkflow, self).__init__(workflow, **kwargs)
self._repeater = Repeater(self)
self._loader = None
self._forwards = []
self._evaluator = None
self._decision = None
self._gds = []
def __init__(self, workflow, **kwargs):
super(GradientRBM, self).__init__(workflow, **kwargs)
self.stddev = kwargs["stddev"]
self.batch_size = -1
self.mem_cpy = MemCpy(self)
self.mem_cpy.link_from(self.start_point)
self.repeater = Repeater(self)
self.repeater.link_from(self.mem_cpy)
self.decision = IterationCounter(
self, max_iterations=kwargs["cd_k"])
self.decision.link_from(self.repeater)
self.bino_h = BinarizationGradH(
self, rand=kwargs.get("rand_h", prng.get()))
self.bino_h.link_attrs(self.mem_cpy, ("input", "output"))
self.bino_h.link_from(self.decision)
self.bino_h.gate_block = self.decision.complete
self.make_v = All2AllSigmoidV(
self, weights_stddev=self.stddev, weights_transposed=True,
output_sample_shape=kwargs["v_size"])
self.make_v.link_from(self.bino_h)
self.make_v.link_attrs(self.bino_h, ("input", "output"))
self.bino_v = BinarizationGradV(
self, rand=kwargs.get("rand_v", prng.get()))
self.bino_v.link_attrs(self.make_v, ("input", "output"))
self.bino_v.link_from(self.make_v)
self.make_h = All2AllSigmoidH(
self, weights_stddev=self.stddev,
output_sample_shape=kwargs["h_size"])
self.make_h.link_attrs(self.bino_v, ("input", "output"))
self.make_h.output = self.mem_cpy.output
self.make_h.link_from(self.bino_v)
self.repeater.link_from(self.make_h)
self.end_point.link_from(self.decision)
self.end_point.gate_block = ~self.decision.complete
self.bino_h.gate_block = self.decision.complete
self.mem_cpy.link_attrs(self, "input")
self.bino_h.link_attrs(self, "batch_size")
self.bino_v.link_attrs(self, "batch_size")
self.make_v.link_attrs(self, "weights")
self.make_v.link_attrs(self, ("bias", "vbias"))
self.make_h.link_attrs(self, "weights")
self.make_h.link_attrs(self, ("bias", "hbias"))
self.link_attrs(self.make_h, "output")
self.link_attrs(self.bino_v, ("v1", "output"))
self.link_attrs(self.make_h, ("h1", "output"))
self.demand("input", "weights", "hbias", "vbias", "batch_size")
