def reduce_model(modin):
rsm = reduce_model_list(modin, [
Dropout, nn.ReLU, nn.Softmax, PositionalEmb, TokenDropout, Sparsemax,
CoordinateEmb, Swish
], [
lambda m: (
m.p,
m.inplace,
), lambda m: (m.inplace, ), lambda m: (m.dim, ), lambda m: (
m.num_pos,
m.num_dim,
m.poff,
m.doff,
m.alpha,
), lambda m: (
m.p,
m.keep_magnitude,
), lambda m: (m.dim, ), lambda m: (
m.num_pos,
m.num_dim,
m.poff,
m.doff,
m.alpha,
m.num_steps,
), lambda m: (
m.reset_beta,
m.beta,
)
])
return reduce_model_list(rsm, [GeLU_GPT, GeLU_BERT, nn.Tanh, nn.Sigmoid])
def reduce_model(modin):
rsm = reduce_model_list(modin, [nn.ReLU, nn.Softmax, Sparsemax, Swish], [
lambda m: (m.inplace, ), lambda m: (m.dim, ), lambda m:
(m.dim, ), lambda m: (m.reset_beta, m.beta, m.dim, m.eps)
])
return reduce_model_list(
rsm, [GELU, GeLU_GPT, GeLU_BERT, Mish, nn.Tanh, nn.Sigmoid])
def reduce_model(modin):
return reduce_model_list(modin, [Dropout, TokenDropout, NGramDropout], [
lambda m: (
m.p,
m.inplace,
), lambda m: (
m.p,
m.inplace,
m.keep_magnitude,
), lambda m: (
m.p,
m.inplace,
m.seqdim,
m.keep_magnitude,
m.sample_p,
m.max_n,
)
])
def reduce_model(modin):
rsm = reduce_model_list(modin, [PositionalEmb, CoordinateEmb], [
lambda m: (
m.num_pos,
m.num_dim,
m.poff,
m.doff,
m.alpha,
),
lambda m: (
m.num_pos,
m.num_dim,
m.poff,
m.doff,
m.alpha,
m.num_steps,
),
])
return reduce_model_drop(reduce_model_act(rsm))