Kevin Fink
commited on
Commit
·
a479880
1
Parent(s):
e1dcc24
dev
Browse files- app.py +18 -8
- requirements.txt +1 -0
app.py
CHANGED
@@ -8,6 +8,7 @@ from sklearn.metrics import accuracy_score
|
|
8 |
import numpy as np
|
9 |
import torch
|
10 |
import os
|
|
|
11 |
from huggingface_hub import login
|
12 |
from peft import get_peft_model, LoraConfig
|
13 |
|
@@ -28,14 +29,23 @@ model.save_pretrained(model_save_path)
|
|
28 |
|
29 |
def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size, lr, grad):
|
30 |
try:
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
39 |
login(api_key.strip())
|
40 |
|
41 |
|
|
|
8 |
import numpy as np
|
9 |
import torch
|
10 |
import os
|
11 |
+
import evaluate
|
12 |
from huggingface_hub import login
|
13 |
from peft import get_peft_model, LoraConfig
|
14 |
|
|
|
29 |
|
30 |
def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size, lr, grad):
|
31 |
try:
|
32 |
+
metric = evaluate.load("rouge", cache_dir='/cache')
|
33 |
+
def compute_metrics(eval_preds):
|
34 |
+
preds, labels = eval_preds
|
35 |
+
if isinstance(preds, tuple):
|
36 |
+
preds = preds[0]
|
37 |
+
# Replace -100s used for padding as we can't decode them
|
38 |
+
preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
|
39 |
+
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
|
40 |
+
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
|
41 |
+
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
|
42 |
+
|
43 |
+
result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
|
44 |
+
result = {k: round(v * 100, 4) for k, v in result.items()}
|
45 |
+
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
|
46 |
+
result["gen_len"] = np.mean(prediction_lens)
|
47 |
+
return result
|
48 |
+
|
49 |
login(api_key.strip())
|
50 |
|
51 |
|
requirements.txt
CHANGED
@@ -6,3 +6,4 @@ huggingface_hub
|
|
6 |
scikit-learn
|
7 |
numpy
|
8 |
torch
|
|
|
|
6 |
scikit-learn
|
7 |
numpy
|
8 |
torch
|
9 |
+
evaluate
|