deit/sparse_linear.py
2023-06-20 02:44:21 +08:00

97 lines
3.5 KiB
Python

import torch.nn as nn
import torch.nn.functional as F
import torch
def compute_mask(t, N, M):
out_channel, in_channel = t.shape
percentile = N / M
t_reshaped = t.reshape(out_channel, -1, M)
mask = torch.ones_like(t)
mask_reshaped = mask.reshape(out_channel, -1, M)
nparams_topprune = int(M * (1-percentile))
if nparams_topprune != 0:
topk = torch.topk(torch.abs(t_reshaped), k=nparams_topprune, largest=False, dim = -1)
mask_reshaped = mask_reshaped.scatter(dim = -1, index = topk.indices, value = 0)
return mask_reshaped.reshape(out_channel, in_channel)
class SparseLinearSuper(nn.Module):
def __init__(self, in_features, out_features, bias=True):
super().__init__()
self.in_features = in_features
self.out_features = out_features
self.weight = nn.Parameter(torch.ones(out_features, in_features))
if bias:
self.bias = nn.Parameter(torch.ones(out_features))
else:
self.bias = None
self.sparsity_idx = 0
self.sparsity_config = (4, 4)
self.nas_config_list = ['no_separate']
self.mask = torch.ones_like(self.weight)
self.set_sample_config(self.sparsity_config)
def set_seperate_config(self, seperate_configs):
# supernet training: used after loading pre-trained weights; ea: used before loading nas weights
if seperate_configs:
self.nas_config_list = seperate_configs
# Repeat
self.weight = nn.Parameter(self.weight.repeat(len(self.nas_config_list), 1, 1))
if self.bias != None:
self.bias = nn.Parameter(self.bias.repeat(len(self.nas_config_list), 1))
def set_sample_config(self, sample_config):
self.sparsity_config = sample_config
self._set_mask()
def _set_mask(self):
n, m = self.sparsity_config
# Find the corresponding index
if len(self.nas_config_list) == 1: # No separate
self.mask = compute_mask(self.weight, n, m)
else:
for config in self.nas_config_list:
if [n, m] in config:
self.sparsity_idx = self.nas_config_list.index(config)
self.mask = compute_mask(self.weight[self.sparsity_idx], n, m)
def __repr__(self):
return f"SparseLinearSuper(in_features={self.in_features}, out_features={self.out_features}, sparse_config:{self.sparsity_config})"
def forward(self, x):
if len(self.nas_config_list) == 1: # No seperate
weight = self.weight * self.mask
else: # separate
weight = self.weight[self.sparsity_idx] * self.mask
if self.bias is not None:
if len(self.nas_config_list) == 1: # No seperate
x = F.linear(x, weight, self.bias)
else: # seperate
x = F.linear(x, weight, self.bias[self.sparsity_idx])
else:
x = F.linear(x, weight)
return x
def num_pruned_params(self):
if self.mask.size() == self.weight.size():
return int(torch.sum(self.mask==0).item())
else:
return int(torch.sum(self.mask==0).item()) + self.weight[0].numel() * (len(self.nas_config_list) - 1)
if __name__ == '__main__':
m = SparseLinearSuper(12, 12)
input = torch.randn(12)
print(m(input))
m.set_seperate_config(seperate_configs=[[[1, 4], [2, 4]], [[4, 4]]])
m.set_sample_config((1, 4))
print(m(input))
print(m.num_pruned_params())
#print(sum(p.numel() for p in m.parameters() if p.requires_grad))