def _build_pos_and_neg_contribs(self):
if (self.nonlinear_mxts_mode ==
NonlinearMxtsMode.DeepLIFT_GenomicsDefault):
preceding_linear_layer = self.get_inputs()
while (type(preceding_linear_layer).__name__
in ["BatchNormalization", "NoOp"]):
preceding_linear_layer = preceding_linear_layer.get_inputs()
if (self.verbose):
print("For layer " + str(self.get_name()) +
" the preceding linear"
" layer is " + str(preceding_linear_layer.get_name()) +
" of"
" type " + type(preceding_linear_layer).__name__ + ";")
if ("Conv" in type(preceding_linear_layer).__name__):
if (self.verbose):
print("In accordance with nonlinear_mxts_mode=" +
self.nonlinear_mxts_mode +
" we are setting the NonlinearMxtsMode to " +
NonlinearMxtsMode.Rescale)
self.nonlinear_mxts_mode = NonlinearMxtsMode.Rescale
elif (type(preceding_linear_layer).__name__ == "Dense"):
if (self.verbose):
print("In accordance with nonlinear_mxts_mode" +
self.nonlinear_mxts_mode +
" we are setting the NonlinearMxtsMode to " +
NonlinearMxtsMode.RevealCancel)
self.nonlinear_mxts_mode = NonlinearMxtsMode.RevealCancel
else:
raise RuntimeError("Unsure how to resolve " +
self.nonlinear_mxts_mode + " when the"
" preceding linear layer is of type"
" " + type(preceding_linear_layer).__name__)
if (self.nonlinear_mxts_mode == NonlinearMxtsMode.RevealCancel):
input_pos_contribs, input_neg_contribs =\
self._get_input_pos_and_neg_contribs()
input_pos_contribs = hf.pseudocount_near_zero(input_pos_contribs)
input_neg_contribs = hf.pseudocount_near_zero(input_neg_contribs)
input_ref = self._get_input_reference_vars()
pos_contribs = 0.5 * (
(self._build_activation_vars(input_ref + input_pos_contribs) -
self._build_activation_vars(input_ref)) +
(self._build_activation_vars(input_ref + input_neg_contribs +
input_pos_contribs) -
self._build_activation_vars(input_ref + input_neg_contribs)))
neg_contribs = 0.5 * (
(self._build_activation_vars(input_ref + input_neg_contribs) -
self._build_activation_vars(input_ref)) +
(self._build_activation_vars(input_ref + input_pos_contribs +
input_neg_contribs) -
self._build_activation_vars(input_ref + input_pos_contribs)))
return (pos_contribs, neg_contribs)
else:
scale_factor = self._get_naive_rescale_factor()
input_pos_contribs, input_neg_contribs =\
self._get_input_pos_and_neg_contribs()
return (input_pos_contribs * scale_factor,
input_neg_contribs * scale_factor)
def _build_pos_and_neg_contribs(self):
if (self.nonlinear_mxts_mode==
NonlinearMxtsMode.DeepLIFT_GenomicsDefault):
preceding_linear_layer = self.get_inputs()
while (type(preceding_linear_layer).__name__ in [
"BatchNormalization", "NoOp"]):
preceding_linear_layer = preceding_linear_layer.get_inputs()
if (self.verbose):
print("For layer "+str(self.get_name())+" the preceding linear"
" layer is "+str(preceding_linear_layer.get_name())+" of"
" type "+type(preceding_linear_layer).__name__+";")
if ("Conv" in type(preceding_linear_layer).__name__):
if (self.verbose):
print("In accordance with nonlinear_mxts_mode="
+self.nonlinear_mxts_mode+
" we are setting the NonlinearMxtsMode to "+
NonlinearMxtsMode.Rescale)
self.nonlinear_mxts_mode=NonlinearMxtsMode.Rescale
elif (type(preceding_linear_layer).__name__=="Dense"):
if (self.verbose):
print("In accordance with nonlinear_mxts_mode"
+self.nonlinear_mxts_mode+
" we are setting the NonlinearMxtsMode to "+
NonlinearMxtsMode.RevealCancel)
self.nonlinear_mxts_mode=NonlinearMxtsMode.RevealCancel
else:
raise RuntimeError("Unsure how to resolve "
+self.nonlinear_mxts_mode+" when the"
" preceding linear layer is of type"
" "+type(preceding_linear_layer).__name__)
if (self.nonlinear_mxts_mode == NonlinearMxtsMode.RevealCancel):
input_pos_contribs, input_neg_contribs =\
self._get_input_pos_and_neg_contribs()
input_pos_contribs = hf.pseudocount_near_zero(input_pos_contribs)
input_neg_contribs = hf.pseudocount_near_zero(input_neg_contribs)
input_ref = self._get_input_reference_vars()
pos_contribs = 0.5*(
(self._build_activation_vars(input_ref+input_pos_contribs)
- self._build_activation_vars(input_ref))
+(self._build_activation_vars(
input_ref+input_neg_contribs+input_pos_contribs)
- self._build_activation_vars(input_ref+input_neg_contribs)))
neg_contribs = 0.5*(
(self._build_activation_vars(input_ref+input_neg_contribs)
- self._build_activation_vars(input_ref))
+(self._build_activation_vars(
input_ref+input_pos_contribs+input_neg_contribs)
- self._build_activation_vars(input_ref+input_pos_contribs)))
return (pos_contribs, neg_contribs)
else:
scale_factor = self._get_naive_rescale_factor()
input_pos_contribs, input_neg_contribs =\
self._get_input_pos_and_neg_contribs()
return (input_pos_contribs*scale_factor,
input_neg_contribs*scale_factor)
def _get_mxts_increments_for_inputs(self):
if (self.nonlinear_mxts_mode == NonlinearMxtsMode.DeconvNet):
#apply the given nonlinearity in reverse
pos_mxts = self._build_activation_vars(self.get_pos_mxts())
neg_mxts = self._build_activation_vars(self.get_neg_mxts())
else:
#all the other ones here are of the form:
# scale_factor*self.get_mxts()
# recall that for all modes except RevealCancel, the treatment
# of positive and negative terms is the same (and if RevealCancel
# occurs nowhere, then pos_mxts=neg_mxts
if (self.nonlinear_mxts_mode == NonlinearMxtsMode.Rescale):
scale_factor = self._get_naive_rescale_factor()
pos_scale_factor = scale_factor
neg_scale_factor = scale_factor
elif (self.nonlinear_mxts_mode ==
NonlinearMxtsMode.GuidedBackpropRescale):
naive_scale_factor = self._get_naive_rescale_factor()
pos_scale_factor = (naive_scale_factor *
hf.gt_mask(self.get_pos_mxts(), 0.0))
neg_scale_factor = (naive_scale_factor *
hf.gt_mask(self.get_neg_mxts(), 0.0))
elif (self.nonlinear_mxts_mode == NonlinearMxtsMode.Gradient):
scale_factor = self._gradients_get_scale_factor()
pos_scale_factor = scale_factor
neg_scale_factor = scale_factor
elif (self.nonlinear_mxts_mode == NonlinearMxtsMode.GuidedBackprop
):
scale_factor = self._gradients_get_scale_factor()\
*hf.gt_mask(self.get_pos_mxts(),0.0)
pos_scale_factor = scale_factor
neg_scale_factor = scale_factor
elif (self.nonlinear_mxts_mode == NonlinearMxtsMode.RevealCancel):
pos_contribs, neg_contribs = self.get_pos_and_neg_contribs()
input_pos_contribs, input_neg_contribs =\
self._get_input_pos_and_neg_contribs()
pos_scale_factor = (
pos_contribs /
hf.pseudocount_near_zero(input_pos_contribs))
neg_scale_factor = (
neg_contribs /
hf.pseudocount_near_zero(input_neg_contribs))
elif (self.nonlinear_mxts_mode == NonlinearMxtsMode.PassThrough):
pos_scale_factor = 1.0
neg_scale_factor = 1.0
else:
raise RuntimeError("Unsupported nonlinear_mxts_mode: " +
str(self.nonlinear_mxts_mode))
pos_mxts = pos_scale_factor * self.get_pos_mxts()
neg_mxts = neg_scale_factor * self.get_neg_mxts()
return pos_mxts, neg_mxts
def _get_mxts_increments_for_inputs(self):
if (self.nonlinear_mxts_mode==NonlinearMxtsMode.DeconvNet):
#apply the given nonlinearity in reverse
pos_mxts = self._build_activation_vars(self.get_pos_mxts())
neg_mxts = self._build_activation_vars(self.get_neg_mxts())
else:
#all the other ones here are of the form:
# scale_factor*self.get_mxts()
# recall that for all modes except RevealCancel, the treatment
# of positive and negative terms is the same (and if RevealCancel
# occurs nowhere, then pos_mxts=neg_mxts
if (self.nonlinear_mxts_mode==NonlinearMxtsMode.Rescale):
scale_factor = self._get_naive_rescale_factor()
pos_scale_factor = scale_factor
neg_scale_factor = scale_factor
elif (self.nonlinear_mxts_mode==
NonlinearMxtsMode.GuidedBackpropRescale):
naive_scale_factor = self._get_naive_rescale_factor()
pos_scale_factor = (naive_scale_factor*
hf.gt_mask(self.get_pos_mxts(),0.0))
neg_scale_factor = (naive_scale_factor*
hf.gt_mask(self.get_neg_mxts(),0.0))
elif (self.nonlinear_mxts_mode==NonlinearMxtsMode.Gradient):
scale_factor = self._gradients_get_scale_factor()
pos_scale_factor = scale_factor
neg_scale_factor = scale_factor
elif (self.nonlinear_mxts_mode==NonlinearMxtsMode.GuidedBackprop):
scale_factor = self._gradients_get_scale_factor()\
*hf.gt_mask(self.get_pos_mxts(),0.0)
pos_scale_factor = scale_factor
neg_scale_factor = scale_factor
elif (self.nonlinear_mxts_mode==NonlinearMxtsMode.RevealCancel):
pos_contribs, neg_contribs = self.get_pos_and_neg_contribs()
input_pos_contribs, input_neg_contribs =\
self._get_input_pos_and_neg_contribs()
pos_scale_factor = (
pos_contribs/hf.pseudocount_near_zero(input_pos_contribs))
neg_scale_factor = (
neg_contribs/hf.pseudocount_near_zero(input_neg_contribs))
elif (self.nonlinear_mxts_mode==NonlinearMxtsMode.PassThrough):
pos_scale_factor = 1.0
neg_scale_factor = 1.0
else:
raise RuntimeError("Unsupported nonlinear_mxts_mode: "
+str(self.nonlinear_mxts_mode))
pos_mxts = pos_scale_factor*self.get_pos_mxts()
neg_mxts = neg_scale_factor*self.get_neg_mxts()
return pos_mxts, neg_mxts