datnguyentien204
commited on
Commit
•
db88be6
1
Parent(s):
b927a0f
Update finetuning.py
Browse files- finetuning.py +129 -129
finetuning.py
CHANGED
@@ -1,129 +1,129 @@
|
|
1 |
-
import os
|
2 |
-
|
3 |
-
import requests
|
4 |
-
from transformers import BlipProcessor, BlipForQuestionAnswering
|
5 |
-
from datasets import load_dataset
|
6 |
-
import torch
|
7 |
-
from PIL import Image
|
8 |
-
from torch.utils.data import DataLoader
|
9 |
-
from tqdm import tqdm
|
10 |
-
import pickle
|
11 |
-
|
12 |
-
model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")
|
13 |
-
processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
|
14 |
-
|
15 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
16 |
-
model.to(device)
|
17 |
-
|
18 |
-
torch.cuda.empty_cache()
|
19 |
-
torch.manual_seed(42)
|
20 |
-
|
21 |
-
class VQADataset(torch.utils.data.Dataset):
|
22 |
-
"""VQA (v2) dataset."""
|
23 |
-
|
24 |
-
def __init__(self, dataset, processor):
|
25 |
-
self.dataset = dataset
|
26 |
-
self.processor = processor
|
27 |
-
|
28 |
-
def __len__(self):
|
29 |
-
return len(self.dataset)
|
30 |
-
|
31 |
-
def __getitem__(self, idx):
|
32 |
-
# get image + text
|
33 |
-
question = self.dataset[idx]['question']
|
34 |
-
answer = self.dataset[idx]['answer']
|
35 |
-
image_id = self.dataset[idx]['pid']
|
36 |
-
image_path = f"Data/train_fill_in_blank/{image_id}/image.png"
|
37 |
-
image = Image.open(image_path).convert("RGB")
|
38 |
-
text = question
|
39 |
-
|
40 |
-
encoding = self.processor(image, text, padding="max_length", truncation=True, return_tensors="pt")
|
41 |
-
labels = self.processor.tokenizer.encode(
|
42 |
-
answer, max_length= 8, pad_to_max_length=True, return_tensors='pt'
|
43 |
-
)
|
44 |
-
encoding["labels"] = labels
|
45 |
-
# remove batch dimension
|
46 |
-
for k,v in encoding.items(): encoding[k] = v.squeeze()
|
47 |
-
return encoding
|
48 |
-
|
49 |
-
training_dataset = load_dataset("json", data_files="Data/train.jsonl", split="train[:90%]")
|
50 |
-
valid_dataset = load_dataset("json", data_files="Data/train.jsonl", split="train[90%:]")
|
51 |
-
print("Training sets: {} - Validating set: {}".format(len(training_dataset), len(valid_dataset)))
|
52 |
-
|
53 |
-
train_dataset = VQADataset(dataset=training_dataset,
|
54 |
-
processor=processor)
|
55 |
-
valid_dataset = VQADataset(dataset=valid_dataset,
|
56 |
-
processor=processor)
|
57 |
-
|
58 |
-
batch_size =
|
59 |
-
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=False, pin_memory=True)
|
60 |
-
valid_dataloader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=False, pin_memory=True)
|
61 |
-
|
62 |
-
|
63 |
-
optimizer = torch.optim.AdamW(model.parameters(), lr=4e-5)
|
64 |
-
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9, last_epoch=-1, verbose=False)
|
65 |
-
|
66 |
-
num_epochs = 100
|
67 |
-
patience = 10
|
68 |
-
min_eval_loss = float("inf")
|
69 |
-
early_stopping_hook = 0
|
70 |
-
tracking_information = []
|
71 |
-
scaler = torch.cuda.amp.GradScaler()
|
72 |
-
|
73 |
-
for epoch in range(num_epochs):
|
74 |
-
epoch_loss = 0
|
75 |
-
model.train()
|
76 |
-
for idx, batch in zip(tqdm(range(len(train_dataloader)), desc='Training batch: ...'), train_dataloader):
|
77 |
-
input_ids = batch.pop('input_ids').to(device)
|
78 |
-
pixel_values = batch.pop('pixel_values').to(device)
|
79 |
-
attention_masked = batch.pop('attention_mask').to(device)
|
80 |
-
labels = batch.pop('labels').to(device)
|
81 |
-
|
82 |
-
with torch.amp.autocast(device_type='cuda', dtype=torch.float16):
|
83 |
-
outputs = model(input_ids=input_ids,
|
84 |
-
pixel_values=pixel_values,
|
85 |
-
# attention_mask=attention_masked,
|
86 |
-
labels=labels)
|
87 |
-
|
88 |
-
loss = outputs.loss
|
89 |
-
epoch_loss += loss.item()
|
90 |
-
# loss.backward()
|
91 |
-
# optimizer.step()
|
92 |
-
optimizer.zero_grad()
|
93 |
-
|
94 |
-
scaler.scale(loss).backward()
|
95 |
-
scaler.step(optimizer)
|
96 |
-
scaler.update()
|
97 |
-
|
98 |
-
model.eval()
|
99 |
-
eval_loss = 0
|
100 |
-
for idx, batch in zip(tqdm(range(len(valid_dataloader)), desc='Validating batch: ...'), valid_dataloader):
|
101 |
-
input_ids = batch.pop('input_ids').to(device)
|
102 |
-
pixel_values = batch.pop('pixel_values').to(device)
|
103 |
-
attention_masked = batch.pop('attention_mask').to(device)
|
104 |
-
labels = batch.pop('labels').to(device)
|
105 |
-
|
106 |
-
with torch.amp.autocast(device_type='cuda', dtype=torch.float16):
|
107 |
-
outputs = model(input_ids=input_ids,
|
108 |
-
pixel_values=pixel_values,
|
109 |
-
attention_mask=attention_masked,
|
110 |
-
labels=labels)
|
111 |
-
|
112 |
-
loss = outputs.loss
|
113 |
-
eval_loss += loss.item()
|
114 |
-
|
115 |
-
tracking_information.append((epoch_loss/len(train_dataloader), eval_loss/len(valid_dataloader), optimizer.param_groups[0]["lr"]))
|
116 |
-
print("Epoch: {} - Training loss: {} - Eval Loss: {} - LR: {}".format(epoch+1, epoch_loss/len(train_dataloader), eval_loss/len(valid_dataloader), optimizer.param_groups[0]["lr"]))
|
117 |
-
scheduler.step()
|
118 |
-
if eval_loss < min_eval_loss:
|
119 |
-
model.save_pretrained("Model/blip-saved-model", from_pt=True)
|
120 |
-
print("Saved model to Model/blip-saved-model")
|
121 |
-
min_eval_loss = eval_loss
|
122 |
-
early_stopping_hook = 0
|
123 |
-
else:
|
124 |
-
early_stopping_hook += 1
|
125 |
-
if early_stopping_hook > patience:
|
126 |
-
break
|
127 |
-
|
128 |
-
pickle.dump(tracking_information, open("tracking_information.pkl", "wb"))
|
129 |
-
print("The finetuning process has done!")
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
import requests
|
4 |
+
from transformers import BlipProcessor, BlipForQuestionAnswering
|
5 |
+
from datasets import load_dataset
|
6 |
+
import torch
|
7 |
+
from PIL import Image
|
8 |
+
from torch.utils.data import DataLoader
|
9 |
+
from tqdm import tqdm
|
10 |
+
import pickle
|
11 |
+
|
12 |
+
model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base")
|
13 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")
|
14 |
+
|
15 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
16 |
+
model.to(device)
|
17 |
+
|
18 |
+
torch.cuda.empty_cache()
|
19 |
+
torch.manual_seed(42)
|
20 |
+
|
21 |
+
class VQADataset(torch.utils.data.Dataset):
|
22 |
+
"""VQA (v2) dataset."""
|
23 |
+
|
24 |
+
def __init__(self, dataset, processor):
|
25 |
+
self.dataset = dataset
|
26 |
+
self.processor = processor
|
27 |
+
|
28 |
+
def __len__(self):
|
29 |
+
return len(self.dataset)
|
30 |
+
|
31 |
+
def __getitem__(self, idx):
|
32 |
+
# get image + text
|
33 |
+
question = self.dataset[idx]['question']
|
34 |
+
answer = self.dataset[idx]['answer']
|
35 |
+
image_id = self.dataset[idx]['pid']
|
36 |
+
image_path = f"Data/train_fill_in_blank/{image_id}/image.png"
|
37 |
+
image = Image.open(image_path).convert("RGB")
|
38 |
+
text = question
|
39 |
+
|
40 |
+
encoding = self.processor(image, text, padding="max_length", truncation=True, return_tensors="pt")
|
41 |
+
labels = self.processor.tokenizer.encode(
|
42 |
+
answer, max_length= 8, pad_to_max_length=True, return_tensors='pt'
|
43 |
+
)
|
44 |
+
encoding["labels"] = labels
|
45 |
+
# remove batch dimension
|
46 |
+
for k,v in encoding.items(): encoding[k] = v.squeeze()
|
47 |
+
return encoding
|
48 |
+
|
49 |
+
training_dataset = load_dataset("json", data_files="Data/train.jsonl", split="train[:90%]")
|
50 |
+
valid_dataset = load_dataset("json", data_files="Data/train.jsonl", split="train[90%:]")
|
51 |
+
print("Training sets: {} - Validating set: {}".format(len(training_dataset), len(valid_dataset)))
|
52 |
+
|
53 |
+
train_dataset = VQADataset(dataset=training_dataset,
|
54 |
+
processor=processor)
|
55 |
+
valid_dataset = VQADataset(dataset=valid_dataset,
|
56 |
+
processor=processor)
|
57 |
+
|
58 |
+
batch_size = 8
|
59 |
+
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=False, pin_memory=True)
|
60 |
+
valid_dataloader = DataLoader(valid_dataset, batch_size=batch_size, shuffle=False, pin_memory=True)
|
61 |
+
|
62 |
+
|
63 |
+
optimizer = torch.optim.AdamW(model.parameters(), lr=4e-5)
|
64 |
+
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9, last_epoch=-1, verbose=False)
|
65 |
+
|
66 |
+
num_epochs = 100
|
67 |
+
patience = 10
|
68 |
+
min_eval_loss = float("inf")
|
69 |
+
early_stopping_hook = 0
|
70 |
+
tracking_information = []
|
71 |
+
scaler = torch.cuda.amp.GradScaler()
|
72 |
+
|
73 |
+
for epoch in range(num_epochs):
|
74 |
+
epoch_loss = 0
|
75 |
+
model.train()
|
76 |
+
for idx, batch in zip(tqdm(range(len(train_dataloader)), desc='Training batch: ...'), train_dataloader):
|
77 |
+
input_ids = batch.pop('input_ids').to(device)
|
78 |
+
pixel_values = batch.pop('pixel_values').to(device)
|
79 |
+
attention_masked = batch.pop('attention_mask').to(device)
|
80 |
+
labels = batch.pop('labels').to(device)
|
81 |
+
|
82 |
+
with torch.amp.autocast(device_type='cuda', dtype=torch.float16):
|
83 |
+
outputs = model(input_ids=input_ids,
|
84 |
+
pixel_values=pixel_values,
|
85 |
+
# attention_mask=attention_masked,
|
86 |
+
labels=labels)
|
87 |
+
|
88 |
+
loss = outputs.loss
|
89 |
+
epoch_loss += loss.item()
|
90 |
+
# loss.backward()
|
91 |
+
# optimizer.step()
|
92 |
+
optimizer.zero_grad()
|
93 |
+
|
94 |
+
scaler.scale(loss).backward()
|
95 |
+
scaler.step(optimizer)
|
96 |
+
scaler.update()
|
97 |
+
|
98 |
+
model.eval()
|
99 |
+
eval_loss = 0
|
100 |
+
for idx, batch in zip(tqdm(range(len(valid_dataloader)), desc='Validating batch: ...'), valid_dataloader):
|
101 |
+
input_ids = batch.pop('input_ids').to(device)
|
102 |
+
pixel_values = batch.pop('pixel_values').to(device)
|
103 |
+
attention_masked = batch.pop('attention_mask').to(device)
|
104 |
+
labels = batch.pop('labels').to(device)
|
105 |
+
|
106 |
+
with torch.amp.autocast(device_type='cuda', dtype=torch.float16):
|
107 |
+
outputs = model(input_ids=input_ids,
|
108 |
+
pixel_values=pixel_values,
|
109 |
+
attention_mask=attention_masked,
|
110 |
+
labels=labels)
|
111 |
+
|
112 |
+
loss = outputs.loss
|
113 |
+
eval_loss += loss.item()
|
114 |
+
|
115 |
+
tracking_information.append((epoch_loss/len(train_dataloader), eval_loss/len(valid_dataloader), optimizer.param_groups[0]["lr"]))
|
116 |
+
print("Epoch: {} - Training loss: {} - Eval Loss: {} - LR: {}".format(epoch+1, epoch_loss/len(train_dataloader), eval_loss/len(valid_dataloader), optimizer.param_groups[0]["lr"]))
|
117 |
+
scheduler.step()
|
118 |
+
if eval_loss < min_eval_loss:
|
119 |
+
model.save_pretrained("Model/blip-saved-model", from_pt=True)
|
120 |
+
print("Saved model to Model/blip-saved-model")
|
121 |
+
min_eval_loss = eval_loss
|
122 |
+
early_stopping_hook = 0
|
123 |
+
else:
|
124 |
+
early_stopping_hook += 1
|
125 |
+
if early_stopping_hook > patience:
|
126 |
+
break
|
127 |
+
|
128 |
+
pickle.dump(tracking_information, open("tracking_information.pkl", "wb"))
|
129 |
+
print("The finetuning process has done!")
|