a-One-Fan/DL-Art-School
1
0
Fork 0
forked from mrq/DL-Art-School
Code Pull Requests Activity

Attempt to fix nan

Browse Source
This commit is contained in:
James Betker 2021-06-07 11:43:42 -06:00
parent eda796985b
commit 1c574c5bd1
1 changed files with 8 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
codes/models/classifiers/cifar_resnet_branched.py
View File

@ -158,6 +158,7 @@ class DropoutNorm(SwitchNorm):
# Compute the dropout probabilities. This module is a no-op before the accumulator is initialized. # Compute the dropout probabilities. This module is a no-op before the accumulator is initialized.
if self.accumulator_filled > 0: if self.accumulator_filled > 0:
with torch.no_grad():
probs = torch.mean(self.accumulator, dim=0) * self.dropout_rate probs = torch.mean(self.accumulator, dim=0) * self.dropout_rate
bs, br = x.shape[:2] bs, br = x.shape[:2]
drop = torch.rand((bs, br), device=x.device) > probs.unsqueeze(0) drop = torch.rand((bs, br), device=x.device) > probs.unsqueeze(0)
@ -172,7 +173,7 @@ class DropoutNorm(SwitchNorm):
class HardRoutingGate(nn.Module): class HardRoutingGate(nn.Module):
def __init__(self, breadth, dropout_rate=.8): def __init__(self, breadth, dropout_rate=.8):
super().__init__() super().__init__()
self.norm = DropoutNorm(breadth, dropout_rate, accumulator_size=2) self.norm = DropoutNorm(breadth, dropout_rate, accumulator_size=128)
def forward(self, x): def forward(self, x):
soft = self.norm(nn.functional.softmax(x, dim=1)) soft = self.norm(nn.functional.softmax(x, dim=1))