Kevin Fink
commited on
Commit
·
6fdec3f
1
Parent(s):
41f77cb
deve
Browse files
app.py
CHANGED
@@ -38,8 +38,10 @@ def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size
|
|
38 |
preds = preds[0]
|
39 |
# Replace -100s used for padding as we can't decode them
|
40 |
preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
|
|
|
41 |
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
|
42 |
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
|
|
|
43 |
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
|
44 |
|
45 |
result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
|
@@ -47,7 +49,7 @@ def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size
|
|
47 |
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
|
48 |
result["gen_len"] = np.mean(prediction_lens)
|
49 |
accuracy = accuracy_score(decoded_labels, decoded_preds)
|
50 |
-
result["
|
51 |
return result
|
52 |
|
53 |
login(api_key.strip())
|
@@ -70,9 +72,9 @@ def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size
|
|
70 |
weight_decay=0.01,
|
71 |
#gradient_accumulation_steps=int(grad),
|
72 |
#max_grad_norm = 3.0,
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
logging_dir='/data/logs',
|
77 |
logging_steps=200,
|
78 |
#push_to_hub=True,
|
@@ -207,8 +209,8 @@ def fine_tune_model(model, dataset_name, hub_id, api_key, num_epochs, batch_size
|
|
207 |
#return 'RUN AGAIN TO LOAD REST OF DATA'
|
208 |
dataset = load_dataset(dataset_name.strip())
|
209 |
#dataset['train'] = dataset['train'].select(range(8000))
|
210 |
-
dataset['train'] = dataset['train'].select(range(
|
211 |
-
dataset['validation'] = dataset['validation'].select(range(
|
212 |
train_set = dataset.map(tokenize_function, batched=True)
|
213 |
#valid_set = dataset['validation'].map(tokenize_function, batched=True)
|
214 |
|
|
|
38 |
preds = preds[0]
|
39 |
# Replace -100s used for padding as we can't decode them
|
40 |
preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
|
41 |
+
preds = np.array(preds)
|
42 |
decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
|
43 |
labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
|
44 |
+
labels = np.array(labels)
|
45 |
decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)
|
46 |
|
47 |
result = metric.compute(predictions=decoded_preds, references=decoded_labels, use_stemmer=True)
|
|
|
49 |
prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
|
50 |
result["gen_len"] = np.mean(prediction_lens)
|
51 |
accuracy = accuracy_score(decoded_labels, decoded_preds)
|
52 |
+
result["eval_accuracy"] = round(accuracy * 100, 4)
|
53 |
return result
|
54 |
|
55 |
login(api_key.strip())
|
|
|
72 |
weight_decay=0.01,
|
73 |
#gradient_accumulation_steps=int(grad),
|
74 |
#max_grad_norm = 3.0,
|
75 |
+
load_best_model_at_end=True,
|
76 |
+
metric_for_best_model="accuracy",
|
77 |
+
greater_is_better=True,
|
78 |
logging_dir='/data/logs',
|
79 |
logging_steps=200,
|
80 |
#push_to_hub=True,
|
|
|
209 |
#return 'RUN AGAIN TO LOAD REST OF DATA'
|
210 |
dataset = load_dataset(dataset_name.strip())
|
211 |
#dataset['train'] = dataset['train'].select(range(8000))
|
212 |
+
dataset['train'] = dataset['train'].select(range(4000))
|
213 |
+
dataset['validation'] = dataset['validation'].select(range(200))
|
214 |
train_set = dataset.map(tokenize_function, batched=True)
|
215 |
#valid_set = dataset['validation'].map(tokenize_function, batched=True)
|
216 |
|