def weight(self, ext, module, g_inp, g_out, backproped):
X, dE_dY = convUtils.get_weight_gradient_factors(
module.input0, g_out[0], module)
if module.has_bias():
X = module.append_ones(X)
return einsum('bml,bkl,bmi,bki->b', (dE_dY, X, dE_dY, X))
def weight(self, ext, module, g_inp, g_out, backproped):
X, dE_dY = convUtils.get_weight_gradient_factors(
module.input0, g_out[0], module)
if module.has_bias():
X = module.append_ones(X)
d1 = einsum('bml,bkl->bmk', (dE_dY, X))
return (d1**2).sum(0).view_as(module.weight)
def weight(self, ext, module, g_inp, g_out, backproped):
"""batch_l2 for weight.
Args:
ext: extension
module: module
g_inp: input gradients
g_out: output gradients
backproped: backpropagation quantities
Returns:
batch_l2 for weight
"""
X, dE_dY = convUtils.get_weight_gradient_factors(
module.input0, g_out[0], module)
return einsum("nmi,nki,nmj,nkj->n", dE_dY, X, dE_dY, X)
def weight(self, ext, module, g_inp, g_out, backproped):
X, dE_dY = convUtils.get_weight_gradient_factors(
module.input0, g_out[0], module)
return einsum("nmi,nki,nmj,nkj->n", (dE_dY, X, dE_dY, X))
def weight(self, ext, module, g_inp, g_out, backproped):
N_axis = 0
X, dE_dY = convUtils.get_weight_gradient_factors(
module.input0, g_out[0], module)
d1 = einsum("nml,nkl->nmk", (dE_dY, X))
return (d1**2).sum(N_axis).view_as(module.weight)