Peter Rupnik
commited on
Commit
·
8941911
1
Parent(s):
1cdf067
Add frames_to_intervals function with filtering
Browse files
README.md
CHANGED
@@ -22,13 +22,13 @@ te test split of the same dataset.
|
|
22 |
|
23 |
# Evaluation
|
24 |
|
25 |
-
Although the output of the model is a series 0 or 1, describing their 20ms frames, the evaluation was done on
|
26 |
-
event level; spans of consecutive outputs 1 were bundled together into one event. When the true and predicted
|
27 |
events partially overlap, this is counted as a true positive.
|
28 |
|
29 |
## Evaluation on ROG corpus
|
30 |
|
31 |
-
In evaluation, we only evaluate positive events, i.e.
|
32 |
```
|
33 |
precision recall f1-score support
|
34 |
|
@@ -41,18 +41,18 @@ Evaluation on 800 human-annotated instances ParlaSpeech-HR and ParlaSpeech-RS p
|
|
41 |
|
42 |
```
|
43 |
Performance on RS:
|
44 |
-
Classification report for human vs model on event level:
|
45 |
precision recall f1-score support
|
46 |
|
47 |
1 0.95 0.99 0.97 542
|
48 |
Performance on HR:
|
49 |
-
Classification report for human vs model on event level:
|
50 |
precision recall f1-score support
|
51 |
|
52 |
1 0.93 0.98 0.95 531
|
53 |
```
|
54 |
-
The metrics reported are on event level, which means that if true and
|
55 |
-
predicted filled pauses at least partially overlap, we count them as a
|
56 |
True Positive event.
|
57 |
|
58 |
|
@@ -80,6 +80,51 @@ ds = Dataset.from_dict(
|
|
80 |
).cast_column("audio", Audio(sampling_rate=16_000, mono=True))
|
81 |
|
82 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
83 |
def evaluator(chunks):
|
84 |
sampling_rate = chunks["audio"][0]["sampling_rate"]
|
85 |
with torch.no_grad():
|
@@ -90,13 +135,18 @@ def evaluator(chunks):
|
|
90 |
).to(device)
|
91 |
logits = model(**inputs).logits
|
92 |
y_pred = np.array(logits.cpu()).argmax(axis=-1)
|
93 |
-
|
|
|
94 |
|
95 |
|
96 |
ds = ds.map(evaluator, batched=True)
|
97 |
print(ds["y_pred"][0])
|
98 |
-
#
|
99 |
# with 0 indicating no filled pause detected in that frame
|
|
|
|
|
|
|
|
|
100 |
```
|
101 |
|
102 |
|
|
|
22 |
|
23 |
# Evaluation
|
24 |
|
25 |
+
Although the output of the model is a series 0 or 1, describing their 20ms frames, the evaluation was done on
|
26 |
+
event level; spans of consecutive outputs 1 were bundled together into one event. When the true and predicted
|
27 |
events partially overlap, this is counted as a true positive.
|
28 |
|
29 |
## Evaluation on ROG corpus
|
30 |
|
31 |
+
In evaluation, we only evaluate positive events, i.e.
|
32 |
```
|
33 |
precision recall f1-score support
|
34 |
|
|
|
41 |
|
42 |
```
|
43 |
Performance on RS:
|
44 |
+
Classification report for human vs model on event level:
|
45 |
precision recall f1-score support
|
46 |
|
47 |
1 0.95 0.99 0.97 542
|
48 |
Performance on HR:
|
49 |
+
Classification report for human vs model on event level:
|
50 |
precision recall f1-score support
|
51 |
|
52 |
1 0.93 0.98 0.95 531
|
53 |
```
|
54 |
+
The metrics reported are on event level, which means that if true and
|
55 |
+
predicted filled pauses at least partially overlap, we count them as a
|
56 |
True Positive event.
|
57 |
|
58 |
|
|
|
80 |
).cast_column("audio", Audio(sampling_rate=16_000, mono=True))
|
81 |
|
82 |
|
83 |
+
def frames_to_intervals(
|
84 |
+
frames: list[int], drop_short=True, drop_initial=True, short_cutoff_s=0.08
|
85 |
+
) -> list[tuple[float]]:
|
86 |
+
"""Transforms a list of ones or zeros, corresponding to annotations on frame
|
87 |
+
levels, to a list of intervals ([start second, end second]).
|
88 |
+
|
89 |
+
Allows for additional filtering on duration (false positives are often short)
|
90 |
+
and start times (false positives starting at 0.0 are often an artifact of
|
91 |
+
poor segmentation).
|
92 |
+
|
93 |
+
:param list[int] frames: Input frame labels
|
94 |
+
:param bool drop_short: Drop everything shorter than short_cutoff_s, defaults to True
|
95 |
+
:param bool drop_initial: Drop predictions starting at 0.0, defaults to True
|
96 |
+
:param float short_cutoff_s: Duration in seconds of shortest allowable prediction, defaults to 0.08
|
97 |
+
:return list[tuple[float]]: List of intervals [start_s, end_s]
|
98 |
+
"""
|
99 |
+
from itertools import pairwise
|
100 |
+
import pandas as pd
|
101 |
+
|
102 |
+
results = []
|
103 |
+
ndf = pd.DataFrame(
|
104 |
+
data={
|
105 |
+
"time_s": [0.020 * i for i in range(len(frames))],
|
106 |
+
"frames": frames,
|
107 |
+
}
|
108 |
+
)
|
109 |
+
ndf = ndf.dropna()
|
110 |
+
indices_of_change = ndf.frames.diff()[ndf.frames.diff() != 0].index.values
|
111 |
+
for si, ei in pairwise(indices_of_change):
|
112 |
+
if ndf.loc[si : ei - 1, "frames"].mode()[0] == 0:
|
113 |
+
pass
|
114 |
+
else:
|
115 |
+
results.append(
|
116 |
+
(
|
117 |
+
round(ndf.loc[si, "time_s"], 3),
|
118 |
+
round(ndf.loc[ei - 1, "time_s"], 3),
|
119 |
+
)
|
120 |
+
)
|
121 |
+
if drop_short and (len(results) > 0):
|
122 |
+
results = [i for i in results if (i[1] - i[0] >= short_cutoff_s)]
|
123 |
+
if drop_initial and (len(results) > 0):
|
124 |
+
results = [i for i in results if i[0] != 0.0]
|
125 |
+
return results
|
126 |
+
|
127 |
+
|
128 |
def evaluator(chunks):
|
129 |
sampling_rate = chunks["audio"][0]["sampling_rate"]
|
130 |
with torch.no_grad():
|
|
|
135 |
).to(device)
|
136 |
logits = model(**inputs).logits
|
137 |
y_pred = np.array(logits.cpu()).argmax(axis=-1)
|
138 |
+
intervals = [frames_to_intervals(i) for i in y_pred]
|
139 |
+
return {"y_pred": y_pred.tolist(), "intervals": intervals}
|
140 |
|
141 |
|
142 |
ds = ds.map(evaluator, batched=True)
|
143 |
print(ds["y_pred"][0])
|
144 |
+
# Prints a list of 20ms frames: [0,0,0,0,1,1,1,1,1,1,1,1,1,1,1,0....]
|
145 |
# with 0 indicating no filled pause detected in that frame
|
146 |
+
|
147 |
+
print(ds["intervals"][0])
|
148 |
+
# Prints the identified intervals as a list of [start_s, ends_s]:
|
149 |
+
# [[0.08, 0.28 ], ...]
|
150 |
```
|
151 |
|
152 |
|