mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-07 08:04:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

keep model_avg on cpu

Browse Source
This commit is contained in:
yaozengwei 2022-05-06 14:45:00 +08:00
parent 70634d58a1
commit c3bb03253f
2 changed files with 3 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
egs/librispeech/ASR/pruned_transducer_stateless4/train.py
View File

@ -867,10 +867,6 @@ def run(rank, world_size, args):
model = DDP(model, device_ids=[rank]) model = DDP(model, device_ids=[rank])
model.device = device model.device = device
if rank == 0:
model_avg.to(device)
model_avg.device = device
optimizer = Eve(model.parameters(), lr=params.initial_lr) optimizer = Eve(model.parameters(), lr=params.initial_lr)
scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs)

4
icefall/checkpoint.py
View File

@ -467,5 +467,7 @@ def average_state_dict(
uniqued_names = list(uniqued.values()) uniqued_names = list(uniqued.values())
for k in uniqued_names: for k in uniqued_names:
state_dict_1[k] *= weight_1 state_dict_1[k] *= weight_1
state_dict_1[k] += state_dict_2[k] * weight_2 state_dict_1[k] += (
state_dict_2[k].to(device=state_dict_1[k].device) * weight_2
)
state_dict_1[k] *= scaling_factor state_dict_1[k] *= scaling_factor