Mod: Refactor model forward.
Browse files
app.py
CHANGED
@@ -1,6 +1,8 @@
|
|
1 |
#!/usr/bin/env python
|
2 |
# -*- coding: utf-8 -*-
|
3 |
|
|
|
|
|
4 |
from tempfile import NamedTemporaryFile
|
5 |
from typing import Any
|
6 |
|
@@ -20,12 +22,56 @@ def format_cand(cand: str) -> str:
|
|
20 |
return f"{cand[0].title()}{cand[1:]}."
|
21 |
|
22 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
23 |
def main() -> None:
|
24 |
st.header("Describe audio content with CoNeTTE")
|
25 |
|
26 |
model = load_conette(model_kwds=dict(device="cpu"))
|
27 |
|
28 |
-
st.warning(
|
|
|
|
|
29 |
audios = st.file_uploader(
|
30 |
"**Upload audio files here:**",
|
31 |
type=["wav", "flac", "mp3", "ogg", "avi"],
|
@@ -78,7 +124,7 @@ def main() -> None:
|
|
78 |
)
|
79 |
del allow_rep_mode
|
80 |
|
81 |
-
|
82 |
task=task,
|
83 |
beam_size=beam_size,
|
84 |
min_pred_size=min_pred_size,
|
@@ -87,39 +133,7 @@ def main() -> None:
|
|
87 |
)
|
88 |
|
89 |
if audios is not None and len(audios) > 0:
|
90 |
-
|
91 |
-
cands = [""] * len(audios)
|
92 |
-
tmp_files = []
|
93 |
-
tmp_fpaths = []
|
94 |
-
audio_fnames = []
|
95 |
-
|
96 |
-
for i, audio in enumerate(audios):
|
97 |
-
audio_fname = audio.name
|
98 |
-
audio_fnames.append(audio_fname)
|
99 |
-
cand_key = f"{audio_fname}-{kwargs}"
|
100 |
-
|
101 |
-
if cand_key in st.session_state:
|
102 |
-
cand = st.session_state[cand_key]
|
103 |
-
cands[i] = cand
|
104 |
-
else:
|
105 |
-
tmp_file = NamedTemporaryFile()
|
106 |
-
tmp_file.write(audio.getvalue())
|
107 |
-
tmp_files.append(tmp_file)
|
108 |
-
audio_to_predict.append((i, cand_key, tmp_file))
|
109 |
-
|
110 |
-
tmp_fpath = tmp_file.name
|
111 |
-
tmp_fpaths.append(tmp_fpath)
|
112 |
-
|
113 |
-
if len(tmp_fpaths) > 0:
|
114 |
-
outputs = model(
|
115 |
-
tmp_fpaths,
|
116 |
-
**kwargs,
|
117 |
-
)
|
118 |
-
for i, (j, cand_key, tmp_file) in enumerate(audio_to_predict):
|
119 |
-
cand = outputs["cands"][i]
|
120 |
-
cands[j] = cand
|
121 |
-
st.session_state[cand_key] = cand
|
122 |
-
tmp_file.close()
|
123 |
|
124 |
for audio_fname, cand in zip(audio_fnames, cands):
|
125 |
st.success(f"**Output for {audio_fname}:**\n- {format_cand(cand)}")
|
@@ -127,10 +141,12 @@ def main() -> None:
|
|
127 |
if len(record) > 0:
|
128 |
outputs = model(
|
129 |
record_fpath,
|
130 |
-
**
|
131 |
)
|
132 |
cand = outputs["cands"][0]
|
133 |
-
st.success(
|
|
|
|
|
134 |
|
135 |
|
136 |
if __name__ == "__main__":
|
|
|
1 |
#!/usr/bin/env python
|
2 |
# -*- coding: utf-8 -*-
|
3 |
|
4 |
+
import os.path as osp
|
5 |
+
|
6 |
from tempfile import NamedTemporaryFile
|
7 |
from typing import Any
|
8 |
|
|
|
22 |
return f"{cand[0].title()}{cand[1:]}."
|
23 |
|
24 |
|
25 |
+
def get_results(
|
26 |
+
model: CoNeTTEModel,
|
27 |
+
audios: list,
|
28 |
+
generate_kwds: dict[str, Any],
|
29 |
+
) -> tuple[list[str], list[str]]:
|
30 |
+
audio_to_predict = []
|
31 |
+
cands = [""] * len(audios)
|
32 |
+
tmp_files = []
|
33 |
+
tmp_fpaths = []
|
34 |
+
audio_fnames = []
|
35 |
+
|
36 |
+
for i, audio in enumerate(audios):
|
37 |
+
audio_fname = audio.name
|
38 |
+
audio_fnames.append(audio_fname)
|
39 |
+
cand_key = f"{audio_fname}-{generate_kwds}"
|
40 |
+
|
41 |
+
if cand_key in st.session_state:
|
42 |
+
cand = st.session_state[cand_key]
|
43 |
+
cands[i] = cand
|
44 |
+
else:
|
45 |
+
tmp_file = NamedTemporaryFile()
|
46 |
+
tmp_file.write(audio.getvalue())
|
47 |
+
tmp_files.append(tmp_file)
|
48 |
+
audio_to_predict.append((i, cand_key, tmp_file))
|
49 |
+
|
50 |
+
tmp_fpath = tmp_file.name
|
51 |
+
tmp_fpaths.append(tmp_fpath)
|
52 |
+
|
53 |
+
if len(tmp_fpaths) > 0:
|
54 |
+
outputs = model(
|
55 |
+
tmp_fpaths,
|
56 |
+
**generate_kwds,
|
57 |
+
)
|
58 |
+
for i, (j, cand_key, tmp_file) in enumerate(audio_to_predict):
|
59 |
+
cand = outputs["cands"][i]
|
60 |
+
cands[j] = cand
|
61 |
+
st.session_state[cand_key] = cand
|
62 |
+
tmp_file.close()
|
63 |
+
|
64 |
+
return audio_fnames, cands
|
65 |
+
|
66 |
+
|
67 |
def main() -> None:
|
68 |
st.header("Describe audio content with CoNeTTE")
|
69 |
|
70 |
model = load_conette(model_kwds=dict(device="cpu"))
|
71 |
|
72 |
+
st.warning(
|
73 |
+
"Recommanded audio: lasting from **1 to 30s**, sampled at **32 kHz** minimum."
|
74 |
+
)
|
75 |
audios = st.file_uploader(
|
76 |
"**Upload audio files here:**",
|
77 |
type=["wav", "flac", "mp3", "ogg", "avi"],
|
|
|
124 |
)
|
125 |
del allow_rep_mode
|
126 |
|
127 |
+
generate_kwds: dict[str, Any] = dict(
|
128 |
task=task,
|
129 |
beam_size=beam_size,
|
130 |
min_pred_size=min_pred_size,
|
|
|
133 |
)
|
134 |
|
135 |
if audios is not None and len(audios) > 0:
|
136 |
+
audio_fnames, cands = get_results(model, audios, generate_kwds)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
137 |
|
138 |
for audio_fname, cand in zip(audio_fnames, cands):
|
139 |
st.success(f"**Output for {audio_fname}:**\n- {format_cand(cand)}")
|
|
|
141 |
if len(record) > 0:
|
142 |
outputs = model(
|
143 |
record_fpath,
|
144 |
+
**generate_kwds,
|
145 |
)
|
146 |
cand = outputs["cands"][0]
|
147 |
+
st.success(
|
148 |
+
f"**Output for {osp.basename(record_fpath)}:**\n- {format_cand(cand)}"
|
149 |
+
)
|
150 |
|
151 |
|
152 |
if __name__ == "__main__":
|