Update all files: Fix Parler-TTS imports, PyTorch version, and model loading
Browse files- download_and_finetune_sst.py +48 -0
- download_and_finetune_tts.py +44 -0
- requirements.txt +1 -2
download_and_finetune_sst.py
CHANGED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor, Trainer, TrainingArguments
|
2 |
+
from datasets import load_dataset
|
3 |
+
|
4 |
+
# Download model
|
5 |
+
model_name = "facebook/wav2vec2-base-960h"
|
6 |
+
model = Wav2Vec2ForCTC.from_pretrained(model_name)
|
7 |
+
processor = Wav2Vec2Processor.from_pretrained(model_name)
|
8 |
+
|
9 |
+
# Load dataset (replace with your dataset)
|
10 |
+
dataset = load_dataset("librispeech_asr", "clean", split="train.100") # Example dataset
|
11 |
+
|
12 |
+
# Preprocess function
|
13 |
+
def preprocess_function(examples):
|
14 |
+
audio = examples["audio"]
|
15 |
+
inputs = processor(audio["array"], sampling_rate=audio["sampling_rate"], return_tensors="pt", padding=True)
|
16 |
+
with processor.as_target_processor():
|
17 |
+
labels = processor(examples["text"], return_tensors="pt", padding=True)
|
18 |
+
return {
|
19 |
+
"input_values": inputs["input_values"][0],
|
20 |
+
"labels": labels["input_ids"][0]
|
21 |
+
}
|
22 |
+
|
23 |
+
train_dataset = dataset.map(preprocess_function, remove_columns=dataset.column_names)
|
24 |
+
|
25 |
+
# Training arguments
|
26 |
+
training_args = TrainingArguments(
|
27 |
+
output_dir="./sst_finetuned",
|
28 |
+
per_device_train_batch_size=8,
|
29 |
+
num_train_epochs=3,
|
30 |
+
save_steps=500,
|
31 |
+
logging_steps=10,
|
32 |
+
)
|
33 |
+
|
34 |
+
# Initialize Trainer
|
35 |
+
trainer = Trainer(
|
36 |
+
model=model,
|
37 |
+
args=training_args,
|
38 |
+
train_dataset=train_dataset,
|
39 |
+
)
|
40 |
+
|
41 |
+
# Fine-tune
|
42 |
+
trainer.train()
|
43 |
+
|
44 |
+
# Save fine-tuned model
|
45 |
+
trainer.save_model("./sst_finetuned")
|
46 |
+
processor.save_pretrained("./sst_finetuned")
|
47 |
+
|
48 |
+
print("SST model fine-tuned and saved to './sst_finetuned'. Upload to models/sst_model in your Space.")
|
download_and_finetune_tts.py
CHANGED
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import ParlerTTSForConditionalGeneration, AutoTokenizer, Trainer, TrainingArguments
|
2 |
+
from datasets import load_dataset
|
3 |
+
|
4 |
+
# Download model
|
5 |
+
model_name = "parler-tts/parler-tts-mini-v1"
|
6 |
+
model = ParlerTTSForConditionalGeneration.from_pretrained(model_name)
|
7 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
8 |
+
|
9 |
+
# Load dataset (replace with your dataset)
|
10 |
+
dataset = load_dataset("lj_speech") # Example dataset; adjust as needed
|
11 |
+
|
12 |
+
# Preprocess function (customize based on your dataset)
|
13 |
+
def preprocess_function(examples):
|
14 |
+
# Tokenize text and prepare audio (example; adjust for your data)
|
15 |
+
inputs = tokenizer(examples["text"], return_tensors="pt", padding=True, truncation=True)
|
16 |
+
# Add audio processing if needed
|
17 |
+
return {"input_ids": inputs["input_ids"], "attention_mask": inputs["attention_mask"]}
|
18 |
+
|
19 |
+
train_dataset = dataset["train"].map(preprocess_function, batched=True)
|
20 |
+
|
21 |
+
# Training arguments
|
22 |
+
training_args = TrainingArguments(
|
23 |
+
output_dir="./tts_finetuned",
|
24 |
+
per_device_train_batch_size=8,
|
25 |
+
num_train_epochs=3,
|
26 |
+
save_steps=500,
|
27 |
+
logging_steps=10,
|
28 |
+
)
|
29 |
+
|
30 |
+
# Initialize Trainer
|
31 |
+
trainer = Trainer(
|
32 |
+
model=model,
|
33 |
+
args=training_args,
|
34 |
+
train_dataset=train_dataset,
|
35 |
+
)
|
36 |
+
|
37 |
+
# Fine-tune
|
38 |
+
trainer.train()
|
39 |
+
|
40 |
+
# Save fine-tuned model
|
41 |
+
trainer.save_model("./tts_finetuned")
|
42 |
+
tokenizer.save_pretrained("./tts_finetuned")
|
43 |
+
|
44 |
+
print("TTS model fine-tuned and saved to './tts_finetuned'. Upload to models/tts_model in your Space.")
|
requirements.txt
CHANGED
@@ -1,9 +1,8 @@
|
|
1 |
fastapi==0.103.2
|
2 |
uvicorn==0.23.2
|
3 |
-
|
4 |
torch==2.1.2
|
5 |
soundfile==0.12.1
|
6 |
numpy==1.26.4
|
7 |
-
llama-cpp-python==0.2.28
|
8 |
pydantic==2.5.3
|
9 |
datasets==2.16.1
|
|
|
1 |
fastapi==0.103.2
|
2 |
uvicorn==0.23.2
|
3 |
+
transformers==4.41.0
|
4 |
torch==2.1.2
|
5 |
soundfile==0.12.1
|
6 |
numpy==1.26.4
|
|
|
7 |
pydantic==2.5.3
|
8 |
datasets==2.16.1
|