Skip to content

Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.

fixbug for cif

游雁93f9a424f2bc
diff --git a/funasr/models/paraformer/cif_predictor.py b/funasr/models/paraformer/cif_predictor.py
index 83ca464da..535131f39 100644
--- a/funasr/models/paraformer/cif_predictor.py
+++ b/funasr/models/paraformer/cif_predictor.py
@@ -506,7 +506,10 @@ def cif_v1_export(hidden, alphas, threshold: float):
frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)
fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)
- prefix_sum = torch.cumsum(alphas, dim=1)
+ # prefix_sum = torch.cumsum(alphas, dim=1)
+ prefix_sum = torch.cumsum(alphas, dim=1, dtype=torch.float64).to(
+ torch.float32
+ ) # cumsum precision degradation cause wrong result in extreme
prefix_sum_floor = torch.floor(prefix_sum)
dislocation_prefix_sum = torch.roll(prefix_sum, 1, dims=1)
dislocation_prefix_sum_floor = torch.floor(dislocation_prefix_sum)
@@ -518,8 +521,8 @@ def cif_v1_export(hidden, alphas, threshold: float):
fires[fire_idxs] = 1
fires = fires + prefix_sum - prefix_sum_floor
+ # prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)
prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)
-
frames = prefix_sum_hidden[fire_idxs]
shift_frames = torch.roll(frames, 1, dims=0)
@@ -530,6 +533,7 @@ def cif_v1_export(hidden, alphas, threshold: float):
shift_frames[shift_batch_idxs] = 0
remains = fires - torch.floor(fires)
+ # remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
shift_remain_frames = torch.roll(remain_frames, 1, dims=0)
@@ -537,8 +541,11 @@ def cif_v1_export(hidden, alphas, threshold: float):
frames = frames - shift_frames + shift_remain_frames - remain_frames
- max_label_len = batch_len.max()
+ # max_label_len = batch_len.max()
+ max_label_len = alphas.sum(dim=-1)
+ max_label_len = torch.floor(max_label_len).max().to(dtype=torch.int64)
+ # frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)
frame_fires_idxs = indices < batch_len.unsqueeze(1)
@@ -667,7 +674,10 @@ def cif_wo_hidden_v1(alphas, threshold, return_fire_idxs=False):
fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)
- prefix_sum = torch.cumsum(alphas, dim=1)
+ # prefix_sum = torch.cumsum(alphas, dim=1)
+ prefix_sum = torch.cumsum(alphas, dim=1, dtype=torch.float64).to(
+ torch.float32
+ ) # cumsum precision degradation cause wrong result in extreme
prefix_sum_floor = torch.floor(prefix_sum)
dislocation_prefix_sum = torch.roll(prefix_sum, 1, dims=1)
dislocation_prefix_sum_floor = torch.floor(dislocation_prefix_sum)
@@ -689,6 +699,8 @@ def cif_v1(hidden, alphas, threshold):
device = hidden.device
dtype = hidden.dtype
batch_size, len_time, hidden_size = hidden.size()
+ # frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)
+ # prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)
frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)
prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)
@@ -702,6 +714,7 @@ def cif_v1(hidden, alphas, threshold):
shift_frames[shift_batch_idxs] = 0
remains = fires - torch.floor(fires)
+ # remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]
shift_remain_frames = torch.roll(remain_frames, 1, dims=0)
@@ -709,8 +722,12 @@ def cif_v1(hidden, alphas, threshold):
frames = frames - shift_frames + shift_remain_frames - remain_frames
- max_label_len = batch_len.max()
+ # max_label_len = batch_len.max()
+ max_label_len = (
+ torch.round(alphas.sum(-1)).int().max()
+ ) # torch.round to calculate the max length
+ # frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)
indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)
frame_fires_idxs = indices < batch_len.unsqueeze(1)