nazemi
commited on
Upload fine_ast.py
Browse files- fine_ast.py +83 -0
fine_ast.py
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
from datasets import load_dataset
|
3 |
+
dataset = load_dataset("audiofolder", data_dir="data")
|
4 |
+
#dataset= dataset["train"].train_test_split(seed=42, shuffle=True, test_size=0.1)
|
5 |
+
from transformers import ASTForAudioClassification
|
6 |
+
from transformers import ASTFeatureExtractor
|
7 |
+
from transformers import TrainingArguments
|
8 |
+
import numpy as np
|
9 |
+
from transformers import Trainer
|
10 |
+
import evaluate
|
11 |
+
batch_size = 8
|
12 |
+
gradient_accumulation_steps = 1
|
13 |
+
num_train_epochs = 10
|
14 |
+
labels=["noise","speech"]
|
15 |
+
num_labels = 2
|
16 |
+
max_duration = 5
|
17 |
+
model_id="bookbot/distil-ast-audioset"
|
18 |
+
model_name = "speechVSnoise"
|
19 |
+
|
20 |
+
label2id, id2label = dict(), dict()
|
21 |
+
for i, label in enumerate(labels):
|
22 |
+
label2id[label] = str(i)
|
23 |
+
id2label[str(i)] = label
|
24 |
+
|
25 |
+
model = ASTForAudioClassification.from_pretrained(
|
26 |
+
model_id,
|
27 |
+
num_labels=num_labels, label2id=label2id,
|
28 |
+
id2label=id2label,
|
29 |
+
ignore_mismatched_sizes=True
|
30 |
+
)
|
31 |
+
feature_extractor = ASTFeatureExtractor.from_pretrained(
|
32 |
+
model_id, do_normalize=True, return_attention_mask=False
|
33 |
+
)
|
34 |
+
|
35 |
+
|
36 |
+
def preprocess_function(examples):
|
37 |
+
audio_arrays = [x["array"] for x in examples["audio"]]
|
38 |
+
inputs = feature_extractor(
|
39 |
+
audio_arrays,
|
40 |
+
sampling_rate=feature_extractor.sampling_rate,
|
41 |
+
max_length=int(feature_extractor.sampling_rate * max_duration),
|
42 |
+
truncation=True,
|
43 |
+
|
44 |
+
)
|
45 |
+
return inputs
|
46 |
+
dataset_encoded = dataset.map(
|
47 |
+
preprocess_function,
|
48 |
+
batched=True,
|
49 |
+
batch_size=1674,
|
50 |
+
num_proc=1,
|
51 |
+
)
|
52 |
+
metric = evaluate.load("accuracy")
|
53 |
+
def compute_metrics(eval_pred):
|
54 |
+
predictions = np.argmax(eval_pred.predictions, axis=1)
|
55 |
+
return metric.compute(predictions=predictions, references=eval_pred.label_ids)
|
56 |
+
training_args = TrainingArguments(
|
57 |
+
f"{model_name}",
|
58 |
+
evaluation_strategy="epoch",
|
59 |
+
save_strategy="epoch",
|
60 |
+
learning_rate=5e-5,
|
61 |
+
per_device_train_batch_size=batch_size,
|
62 |
+
gradient_accumulation_steps=gradient_accumulation_steps,
|
63 |
+
per_device_eval_batch_size=batch_size,
|
64 |
+
num_train_epochs=num_train_epochs,
|
65 |
+
warmup_ratio=0.1,
|
66 |
+
logging_steps=5,
|
67 |
+
load_best_model_at_end=True,
|
68 |
+
# metric_for_best_model="accuracy",
|
69 |
+
# push_to_hub=True,
|
70 |
+
)
|
71 |
+
|
72 |
+
from transformers import Trainer
|
73 |
+
|
74 |
+
trainer = Trainer(
|
75 |
+
model,
|
76 |
+
training_args,
|
77 |
+
train_dataset=dataset_encoded["train"],
|
78 |
+
eval_dataset=dataset_encoded["train"],
|
79 |
+
tokenizer=feature_extractor,
|
80 |
+
# compute_metrics=compute_metrics,
|
81 |
+
)
|
82 |
+
trainer.train()
|
83 |
+
|