mrq/vall-e
5
9
Fork 0
Code Issues 8 Pull Requests Packages Projects Releases Wiki Activity

do not like that

Browse Source
This commit is contained in:
mrq 2025-02-27 23:56:02 -06:00
parent f4f435d7f5
commit fc25a9a7dc
2 changed files with 4 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
vall_e/models/ar_nar_v2.py
View File

@ -1037,7 +1037,8 @@ def example_usage():
texts, proms, resps, tasks = sample_data()
stats = {"step": i}
stats |= engine.traverse(phns_list=texts, proms_list=proms, resps_list=resps, task_list=tasks, training=True)
with torch.autograd.set_detect_anomaly(cfg.trainer.detect_grad_anomaly):
stats |= engine.traverse(phns_list=texts, proms_list=proms, resps_list=resps, task_list=tasks, training=True)
stats |= {"grad_norm": engine.get_global_grad_norm()}
tqdm.write(f"{stats}")

9
vall_e/models/base_v2.py
View File

@ -944,13 +944,8 @@ class Base_V2(nn.Module):
sequence = token.t()
nll, metrics = _calc_loss( logits[batch_index][:, start:end], sequence.long(), causal )
for level in enumerate(self.n_resp_levels):
loss_key = f'{name}[{level}].nll'
if loss_key not in loss:
loss[loss_key] = []
loss[loss_key].append( nll[level] * loss_factor )
nll = None
if nll is not None:
nll = nll.sum()
loss_key = f'{name}.nll'
acc_key = f'{name}.acc'