mirrors/pytorch-image-models
1
0
Fork 0
mirror of https://github.com/huggingface/pytorch-image-models.git synced 2025-06-03 15:01:08 +08:00
Code Issues Projects Releases Wiki Activity

Update metaformers.py

Browse Source
This commit is contained in:
Fredo Guan 2023-01-08 11:19:28 -08:00
parent 1b3318bf19
commit 4cfecf8acb
1 changed files with 18 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
timm/models/metaformers.py
View File

@ -647,6 +647,9 @@ class MetaFormer(nn.Module):
if not isinstance(res_scale_init_values, (list, tuple)): if not isinstance(res_scale_init_values, (list, tuple)):
res_scale_init_values = [res_scale_init_values] * num_stage res_scale_init_values = [res_scale_init_values] * num_stage
self.grad_checkpointing = False
self.feature_info = []
stages = nn.ModuleList() # each stage consists of multiple metaformer blocks stages = nn.ModuleList() # each stage consists of multiple metaformer blocks
cur = 0 cur = 0
for i in range(num_stage): for i in range(num_stage):
@ -665,6 +668,7 @@ class MetaFormer(nn.Module):
) )
stages.append(stage) stages.append(stage)
cur += depths[i] cur += depths[i]
self.feature_info += [dict(num_chs=dims[stage_id], reduction=2, module=f'stages.{stage_id}')]
self.stages = nn.Sequential(*stages) self.stages = nn.Sequential(*stages)
self.norm = output_norm(dims[-1]) self.norm = output_norm(dims[-1])
@ -688,16 +692,25 @@ class MetaFormer(nn.Module):
def no_weight_decay(self): def no_weight_decay(self):
return {'norm'} return {'norm'}
def forward_features(self, x): def forward_head(self, x, pre_logits: bool = False):
x = self.stages(x) if pre_logits:
return x
x = x.mean([1,2]) # TODO use adaptive pool instead of mean x = x.mean([1,2]) # TODO use adaptive pool instead of mean
x = self.norm(x) x = self.norm(x)
# (B, H, W, C) -> (B, C) # (B, H, W, C) -> (B, C)
x = self.head(x)
return x
def forward_features(self, x):
x = self.stages(x)
return x return x
def forward(self, x): def forward(self, x):
x = self.forward_features(x) x = self.forward_features(x)
x = self.head(x) x = forward_head(x)
return x return x
def checkpoint_filter_fn(state_dict, model): def checkpoint_filter_fn(state_dict, model):