Add more instructions
Browse files
README.md
CHANGED
@@ -34,7 +34,7 @@ This model is a fine-tuned version of [openai/whisper-large-v2](https://huggingf
|
|
34 |
To use the model, use the following code. It should be able to inference with less than 16GB VRAM.
|
35 |
```
|
36 |
from peft import PeftModel, PeftConfig
|
37 |
-
from transformers import WhisperForConditionalGeneration, Seq2SeqTrainer
|
38 |
|
39 |
peft_model_id = "alvanlii/whisper-largev2-cantonese-peft-lora"
|
40 |
peft_config = PeftConfig.from_pretrained(peft_model_id)
|
@@ -42,6 +42,16 @@ model = WhisperForConditionalGeneration.from_pretrained(
|
|
42 |
peft_config.base_model_name_or_path, load_in_8bit=True, device_map="auto"
|
43 |
)
|
44 |
model = PeftModel.from_pretrained(model, peft_model_id)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
```
|
46 |
|
47 |
## Training and evaluation data
|
@@ -64,12 +74,11 @@ For training, three datasets were used:
|
|
64 |
|
65 |
## Training Results
|
66 |
|
67 |
-
| Training Loss | Epoch | Step | Validation Loss |
|
68 |
-
|
69 |
-
| <TBA> | 0.55 | 2000 | <TBA> |
|
70 |
-
| <TBA> | 1.11 | 4000 | <TBA> |
|
71 |
-
| <TBA> | 1.66 | 6000 | <TBA> |
|
72 |
-
| <TBA> | 2.22 | 8000 | <TBA> |
|
73 |
-
| <TBA> | 2.77 | 10000 | <TBA> |
|
74 |
-
| <TBA> | 3.32 | 12000 | <TBA> |
|
75 |
-
| <TBA> | 3.88 | 14000 | <TBA> | <TBA> |
|
|
|
34 |
To use the model, use the following code. It should be able to inference with less than 16GB VRAM.
|
35 |
```
|
36 |
from peft import PeftModel, PeftConfig
|
37 |
+
from transformers import WhisperForConditionalGeneration, Seq2SeqTrainer, WhisperTokenizer, WhisperProcessor
|
38 |
|
39 |
peft_model_id = "alvanlii/whisper-largev2-cantonese-peft-lora"
|
40 |
peft_config = PeftConfig.from_pretrained(peft_model_id)
|
|
|
42 |
peft_config.base_model_name_or_path, load_in_8bit=True, device_map="auto"
|
43 |
)
|
44 |
model = PeftModel.from_pretrained(model, peft_model_id)
|
45 |
+
|
46 |
+
task = "transcribe"
|
47 |
+
tokenizer = WhisperTokenizer.from_pretrained(peft_config.base_model_name_or_path, task=task)
|
48 |
+
processor = WhisperProcessor.from_pretrained(peft_config.base_model_name_or_path, task=task)
|
49 |
+
feature_extractor = processor.feature_extractor
|
50 |
+
forced_decoder_ids = processor.get_decoder_prompt_ids(language=language, task=task)
|
51 |
+
pipe = AutomaticSpeechRecognitionPipeline(model=model, tokenizer=tokenizer, feature_extractor=feature_extractor)
|
52 |
+
|
53 |
+
audio = # load audio here
|
54 |
+
text = pipe(audio, generate_kwargs={"forced_decoder_ids": forced_decoder_ids}, max_new_tokens=255)["text"]
|
55 |
```
|
56 |
|
57 |
## Training and evaluation data
|
|
|
74 |
|
75 |
## Training Results
|
76 |
|
77 |
+
| Training Loss | Epoch | Step | Validation Loss |
|
78 |
+
|:-------------:|:-----:|:----:|:---------------:|
|
79 |
+
| <TBA> | 0.55 | 2000 | <TBA> |
|
80 |
+
| <TBA> | 1.11 | 4000 | <TBA> |
|
81 |
+
| <TBA> | 1.66 | 6000 | <TBA> |
|
82 |
+
| <TBA> | 2.22 | 8000 | <TBA> |
|
83 |
+
| <TBA> | 2.77 | 10000 | <TBA> |
|
84 |
+
| <TBA> | 3.32 | 12000 | <TBA> |
|
|