Spaces:
Sleeping
Sleeping
Update finetune.py
Browse files- finetune.py +28 -28
finetune.py
CHANGED
@@ -191,7 +191,34 @@ def finetune_model(lang, task, tempdir_path, log_every, max_epoch, scheduler, wa
|
|
191 |
gr.Info("Finished collect stats, starting training.")
|
192 |
log(tempdir_path, "Finished collect stats, starting training...")
|
193 |
trainer.train()
|
194 |
-
gr.Info("Finished Fine-tuning!
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
195 |
|
196 |
log(tempdir_path, "Finished fine-tuning.")
|
197 |
log(tempdir_path, "Start archiving experiment files...")
|
@@ -224,34 +251,7 @@ def finetune_model(lang, task, tempdir_path, log_every, max_epoch, scheduler, wa
|
|
224 |
|
225 |
gr.Info("Finished generating result file in zip!")
|
226 |
log(tempdir_path, "Finished generating result file in zip!")
|
227 |
-
|
228 |
-
gr.Info("Start generating output for test set!")
|
229 |
-
log(tempdir_path, "Start generating output for test set!")
|
230 |
-
|
231 |
-
del trainer
|
232 |
-
model = Speech2Text(
|
233 |
-
"assets/owsm_ebf_v3.1_base/config.yaml",
|
234 |
-
"assets/owsm_ebf_v3.1_base/owsm_v3.1_base.trained.pth",
|
235 |
-
device="cuda" if torch.cuda.is_available() else "cpu",
|
236 |
-
token_type="bpe",
|
237 |
-
bpemodel="assets/owsm_ebf_v3.1_base/bpe.model",
|
238 |
-
beam_size=5,
|
239 |
-
ctc_weight=0.0,
|
240 |
-
lang_sym=f"<{lang}>",
|
241 |
-
task_sym=f"<{task}>",
|
242 |
-
)
|
243 |
-
model.s2t_model.eval()
|
244 |
-
d = torch.load(f"{tempdir_path}/exp/finetune/valid.acc.ave.pth")
|
245 |
-
model.s2t_model.load_state_dict(d)
|
246 |
|
247 |
-
hyp = ""
|
248 |
-
with open(f"{tempdir_path}/hyp.txt", "w") as f_hyp:
|
249 |
-
for i in range(len(test_list)):
|
250 |
-
data = test_list[i]
|
251 |
-
out = model(librosa.load(data['audio_path'], sr=16000)[0])[0][3]
|
252 |
-
f_hyp.write(out + '\n')
|
253 |
-
hyp += out + '\n'
|
254 |
-
|
255 |
return [f"{tempdir_path}/finetune.zip", f"{tempdir_path}/ref.txt", f"{tempdir_path}/base.txt", f"{tempdir_path}/hyp.txt"], ref, base, hyp
|
256 |
|
257 |
|
|
|
191 |
gr.Info("Finished collect stats, starting training.")
|
192 |
log(tempdir_path, "Finished collect stats, starting training...")
|
193 |
trainer.train()
|
194 |
+
gr.Info("Finished Fine-tuning!")
|
195 |
+
|
196 |
+
gr.Info("Start generating output for test set!")
|
197 |
+
log(tempdir_path, "Start generating output for test set!")
|
198 |
+
|
199 |
+
del trainer
|
200 |
+
model = Speech2Text(
|
201 |
+
"assets/owsm_ebf_v3.1_base/config.yaml",
|
202 |
+
"assets/owsm_ebf_v3.1_base/owsm_v3.1_base.trained.pth",
|
203 |
+
device="cuda" if torch.cuda.is_available() else "cpu",
|
204 |
+
token_type="bpe",
|
205 |
+
bpemodel="assets/owsm_ebf_v3.1_base/bpe.model",
|
206 |
+
beam_size=5,
|
207 |
+
ctc_weight=0.0,
|
208 |
+
lang_sym=f"<{lang}>",
|
209 |
+
task_sym=f"<{task}>",
|
210 |
+
)
|
211 |
+
model.s2t_model.eval()
|
212 |
+
d = torch.load(f"{tempdir_path}/exp/finetune/valid.acc.ave.pth")
|
213 |
+
model.s2t_model.load_state_dict(d)
|
214 |
+
|
215 |
+
hyp = ""
|
216 |
+
with open(f"{tempdir_path}/hyp.txt", "w") as f_hyp:
|
217 |
+
for i in range(len(test_list)):
|
218 |
+
data = test_list[i]
|
219 |
+
out = model(librosa.load(data['audio_path'], sr=16000)[0])[0][3]
|
220 |
+
f_hyp.write(out + '\n')
|
221 |
+
hyp += out + '\n'
|
222 |
|
223 |
log(tempdir_path, "Finished fine-tuning.")
|
224 |
log(tempdir_path, "Start archiving experiment files...")
|
|
|
251 |
|
252 |
gr.Info("Finished generating result file in zip!")
|
253 |
log(tempdir_path, "Finished generating result file in zip!")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
254 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
255 |
return [f"{tempdir_path}/finetune.zip", f"{tempdir_path}/ref.txt", f"{tempdir_path}/base.txt", f"{tempdir_path}/hyp.txt"], ref, base, hyp
|
256 |
|
257 |
|