Spaces:
Sleeping
Sleeping
Varun Wadhwa
commited on
Copying over
Browse files- app.py +252 -0
- requirements.txt +15 -0
app.py
ADDED
@@ -0,0 +1,252 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
|
3 |
+
from datasets import load_dataset
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import os
|
7 |
+
from sklearn.metrics import accuracy_score, precision_recall_fscore_support
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
import torch.optim as optim
|
12 |
+
from torch.utils.data import DataLoader
|
13 |
+
|
14 |
+
from transformers import AutoModelForTokenClassification, AutoTokenizer, DataCollatorForTokenClassification
|
15 |
+
from transformers import DebertaV2Config, DebertaV2ForTokenClassification
|
16 |
+
|
17 |
+
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
|
18 |
+
|
19 |
+
# print weights
|
20 |
+
def print_trainable_parameters(model):
|
21 |
+
pytorch_total_params = sum(p.numel() for p in model.parameters())
|
22 |
+
torch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
|
23 |
+
print(f'total params: {pytorch_total_params}. tunable params: {torch_total_params}')
|
24 |
+
|
25 |
+
device = torch.device('cpu')
|
26 |
+
print(f"Is CUDA available: {torch.cuda.is_available()}")
|
27 |
+
# True
|
28 |
+
if torch.cuda.is_available():
|
29 |
+
print(f"CUDA device: {torch.cuda.get_device_name(torch.cuda.current_device())}")
|
30 |
+
device = torch.device('cuda')
|
31 |
+
|
32 |
+
# Load models
|
33 |
+
st.write('Loading the pretrained model ...')
|
34 |
+
teacher_model_name = "iiiorg/piiranha-v1-detect-personal-information"
|
35 |
+
teacher_model = AutoModelForTokenClassification.from_pretrained(teacher_model_name)
|
36 |
+
tokenizer = AutoTokenizer.from_pretrained(teacher_model_name)
|
37 |
+
print(teacher_model)
|
38 |
+
print_trainable_parameters(teacher_model)
|
39 |
+
label2id = teacher_model.config.label2id
|
40 |
+
id2label = teacher_model.config.id2label
|
41 |
+
|
42 |
+
st.write("id2label: ", id2label)
|
43 |
+
st.write("label2id: ", label2id)
|
44 |
+
dimension = len(id2label)
|
45 |
+
st.write("dimension", dimension)
|
46 |
+
|
47 |
+
student_model_config = teacher_model.config
|
48 |
+
student_model_config.num_attention_heads = 8
|
49 |
+
student_model_config.num_hidden_layers = 4
|
50 |
+
student_model = DebertaV2ForTokenClassification.from_pretrained(
|
51 |
+
"microsoft/mdeberta-v3-base",
|
52 |
+
config=student_model_config)
|
53 |
+
# ignore_mismatched_sizes=True)
|
54 |
+
print(student_model)
|
55 |
+
print_trainable_parameters(student_model)
|
56 |
+
|
57 |
+
if torch.cuda.is_available():
|
58 |
+
teacher_model = teacher_model.to(device)
|
59 |
+
student_model = student_model.to(device)
|
60 |
+
|
61 |
+
# Load data.
|
62 |
+
raw_dataset = load_dataset("ai4privacy/pii-masking-400k", split='train')
|
63 |
+
raw_dataset = raw_dataset.filter(lambda example: example["language"].startswith("en"))
|
64 |
+
#raw_dataset = raw_dataset.select(range(2000))
|
65 |
+
raw_dataset = raw_dataset.filter(lambda example, idx: idx % 11 == 0, with_indices=True)
|
66 |
+
raw_dataset = raw_dataset.train_test_split(test_size=0.2)
|
67 |
+
print(raw_dataset)
|
68 |
+
print(raw_dataset.column_names)
|
69 |
+
|
70 |
+
# inputs = tokenizer(
|
71 |
+
# raw_dataset['train'][0]['mbert_tokens'],
|
72 |
+
# truncation=True,
|
73 |
+
# is_split_into_words=True)
|
74 |
+
# print(inputs)
|
75 |
+
# print(inputs.tokens())
|
76 |
+
# print(inputs.word_ids())
|
77 |
+
|
78 |
+
# function to align labels with tokens
|
79 |
+
# --> special tokens: -100 label id (ignored by cross entropy),
|
80 |
+
# --> if tokens are inside a word, replace 'B-' with 'I-'
|
81 |
+
def align_labels_with_tokens(labels):
|
82 |
+
aligned_label_ids = []
|
83 |
+
aligned_label_ids.append(-100)
|
84 |
+
for i, label in enumerate(labels):
|
85 |
+
if label.startswith("B-"):
|
86 |
+
label = label.replace("B-", "I-")
|
87 |
+
aligned_label_ids.append(label2id[label])
|
88 |
+
aligned_label_ids.append(-100)
|
89 |
+
return aligned_label_ids
|
90 |
+
|
91 |
+
# create tokenize function
|
92 |
+
def tokenize_function(examples):
|
93 |
+
# tokenize and truncate text. The examples argument would have already stripped
|
94 |
+
# the train or test label.
|
95 |
+
new_labels = []
|
96 |
+
inputs = tokenizer(
|
97 |
+
examples['mbert_tokens'],
|
98 |
+
is_split_into_words=True,
|
99 |
+
padding=True,
|
100 |
+
truncation=True,
|
101 |
+
max_length=512)
|
102 |
+
for _, labels in enumerate(examples['mbert_token_classes']):
|
103 |
+
new_labels.append(align_labels_with_tokens(labels))
|
104 |
+
|
105 |
+
inputs["labels"] = new_labels
|
106 |
+
return inputs
|
107 |
+
|
108 |
+
# tokenize training and validation datasets
|
109 |
+
tokenized_data = raw_dataset.map(
|
110 |
+
tokenize_function,
|
111 |
+
batched=True)
|
112 |
+
tokenized_data.set_format(type='torch', columns=['input_ids', 'attention_mask', 'labels'])
|
113 |
+
# data collator
|
114 |
+
data_collator = DataCollatorForTokenClassification(tokenizer)
|
115 |
+
|
116 |
+
st.write(tokenized_data["train"][:2]["labels"])
|
117 |
+
|
118 |
+
# Function to evaluate model performance
|
119 |
+
def evaluate_model(model, dataloader, device):
|
120 |
+
model.eval() # Set model to evaluation mode
|
121 |
+
all_preds = []
|
122 |
+
all_labels = []
|
123 |
+
|
124 |
+
# Disable gradient calculations
|
125 |
+
with torch.no_grad():
|
126 |
+
for batch in dataloader:
|
127 |
+
input_ids = batch['input_ids'].to(device)
|
128 |
+
attention_mask = batch['attention_mask'].to(device)
|
129 |
+
labels = batch['labels'].to(device)
|
130 |
+
|
131 |
+
# Forward pass to get logits
|
132 |
+
outputs = model(input_ids, attention_mask=attention_mask)
|
133 |
+
logits = outputs.logits
|
134 |
+
|
135 |
+
# Get predictions
|
136 |
+
preds = torch.argmax(logits, dim=-1).cpu().numpy()
|
137 |
+
all_preds.extend(preds)
|
138 |
+
all_labels.extend(labels.cpu().numpy())
|
139 |
+
|
140 |
+
# Calculate evaluation metrics
|
141 |
+
print("evaluate_model sizes")
|
142 |
+
print(len(all_preds[0]))
|
143 |
+
print(len(all_labels[0]))
|
144 |
+
all_preds = np.asarray(all_preds, dtype=np.float32)
|
145 |
+
all_labels = np.asarray(all_labels, dtype=np.float32)
|
146 |
+
print("Flattened sizes")
|
147 |
+
print(all_preds.size)
|
148 |
+
print(all_labels.size)
|
149 |
+
all_preds = all_preds.flatten()
|
150 |
+
all_labels = all_labels.flatten()
|
151 |
+
accuracy = accuracy_score(all_labels, all_preds)
|
152 |
+
precision, recall, f1, _ = precision_recall_fscore_support(all_labels, all_preds, average='micro')
|
153 |
+
|
154 |
+
return accuracy, precision, recall, f1
|
155 |
+
|
156 |
+
# Function to compute distillation and hard-label loss
|
157 |
+
def distillation_loss(student_logits, teacher_logits, true_labels, temperature, alpha):
|
158 |
+
# print("Distillation loss sizes")
|
159 |
+
# print(teacher_logits.size())
|
160 |
+
# print(student_logits.size())
|
161 |
+
# print(true_labels.size())
|
162 |
+
# Compute soft targets from teacher logits
|
163 |
+
soft_targets = nn.functional.softmax(teacher_logits / temperature, dim=-1)
|
164 |
+
student_soft = nn.functional.log_softmax(student_logits / temperature, dim=-1)
|
165 |
+
|
166 |
+
# KL Divergence loss for distillation
|
167 |
+
distill_loss = nn.functional.kl_div(student_soft, soft_targets, reduction='batchmean') * (temperature ** 2)
|
168 |
+
|
169 |
+
# Cross-entropy loss for hard labels
|
170 |
+
student_logit_reshape = torch.transpose(student_logits, 1, 2) # transpose to match the labels dimension
|
171 |
+
hard_loss = nn.CrossEntropyLoss()(student_logit_reshape, true_labels)
|
172 |
+
|
173 |
+
# Combine losses
|
174 |
+
loss = alpha * distill_loss + (1.0 - alpha) * hard_loss
|
175 |
+
|
176 |
+
return loss
|
177 |
+
|
178 |
+
# hyperparameters
|
179 |
+
batch_size = 32
|
180 |
+
lr = 1e-4
|
181 |
+
num_epochs = 30
|
182 |
+
temperature = 2.0
|
183 |
+
alpha = 0.5
|
184 |
+
|
185 |
+
# define optimizer
|
186 |
+
optimizer = optim.Adam(student_model.parameters(), lr=lr)
|
187 |
+
|
188 |
+
# create training data loader
|
189 |
+
dataloader = DataLoader(tokenized_data['train'], batch_size=batch_size, collate_fn=data_collator)
|
190 |
+
# create testing data loader
|
191 |
+
test_dataloader = DataLoader(tokenized_data['test'], batch_size=batch_size, collate_fn=data_collator)
|
192 |
+
|
193 |
+
# put student model in train mode
|
194 |
+
student_model.train()
|
195 |
+
|
196 |
+
# train model
|
197 |
+
for epoch in range(num_epochs):
|
198 |
+
for batch in dataloader:
|
199 |
+
# Prepare inputs
|
200 |
+
input_ids = batch['input_ids'].to(device)
|
201 |
+
attention_mask = batch['attention_mask'].to(device)
|
202 |
+
labels = batch['labels'].to(device)
|
203 |
+
|
204 |
+
# Disable gradient calculation for teacher model
|
205 |
+
with torch.no_grad():
|
206 |
+
teacher_outputs = teacher_model(input_ids, attention_mask=attention_mask)
|
207 |
+
teacher_logits = teacher_outputs.logits
|
208 |
+
|
209 |
+
# Forward pass through the student model
|
210 |
+
student_outputs = student_model(input_ids, attention_mask=attention_mask)
|
211 |
+
student_logits = student_outputs.logits
|
212 |
+
|
213 |
+
# Compute the distillation loss
|
214 |
+
loss = distillation_loss(student_logits, teacher_logits, labels, temperature, alpha)
|
215 |
+
|
216 |
+
# Backpropagation
|
217 |
+
optimizer.zero_grad()
|
218 |
+
loss.backward()
|
219 |
+
optimizer.step()
|
220 |
+
|
221 |
+
print(f"Epoch {epoch + 1} completed with loss: {loss.item()}")
|
222 |
+
|
223 |
+
# Evaluate the teacher model
|
224 |
+
teacher_accuracy, teacher_precision, teacher_recall, teacher_f1 = evaluate_model(teacher_model, test_dataloader, device)
|
225 |
+
print(f"Teacher (test) - Accuracy: {teacher_accuracy:.4f}, Precision: {teacher_precision:.4f}, Recall: {teacher_recall:.4f}, F1 Score: {teacher_f1:.4f}")
|
226 |
+
|
227 |
+
# Evaluate the student model
|
228 |
+
student_accuracy, student_precision, student_recall, student_f1 = evaluate_model(student_model, test_dataloader, device)
|
229 |
+
print(f"Student (test) - Accuracy: {student_accuracy:.4f}, Precision: {student_precision:.4f}, Recall: {student_recall:.4f}, F1 Score: {student_f1:.4f}")
|
230 |
+
print("\n")
|
231 |
+
|
232 |
+
# put student model back into train mode
|
233 |
+
student_model.train()
|
234 |
+
|
235 |
+
#Compare the models
|
236 |
+
# create testing data loader
|
237 |
+
validation_dataloader = DataLoader(tokenized_data['test'], batch_size=8, collate_fn=data_collator)
|
238 |
+
# Evaluate the teacher model
|
239 |
+
teacher_accuracy, teacher_precision, teacher_recall, teacher_f1 = evaluate_model(teacher_model, validation_dataloader, device)
|
240 |
+
print(f"Teacher (validation) - Accuracy: {teacher_accuracy:.4f}, Precision: {teacher_precision:.4f}, Recall: {teacher_recall:.4f}, F1 Score: {teacher_f1:.4f}")
|
241 |
+
# Evaluate the student model
|
242 |
+
student_accuracy, student_precision, student_recall, student_f1 = evaluate_model(student_model, validation_dataloader, device)
|
243 |
+
print(f"Student (validation) - Accuracy: {student_accuracy:.4f}, Precision: {student_precision:.4f}, Recall: {student_recall:.4f}, F1 Score: {student_f1:.4f}")
|
244 |
+
|
245 |
+
|
246 |
+
st.write('Pushing model to huggingface')
|
247 |
+
|
248 |
+
# Push model to huggingface
|
249 |
+
hf_name = 'CarolXia' # your hf username or org name
|
250 |
+
mode_name = "pii-kd-deberta-v2"
|
251 |
+
model_id = hf_name + "/" + mode_name
|
252 |
+
student_model.push_to_hub(model_id, token=st.secrets["HUGGINGFACE_TOKEN"])
|
requirements.txt
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
auto-gptq
|
2 |
+
bitsandbytes
|
3 |
+
datasets
|
4 |
+
evaluate
|
5 |
+
seqeval
|
6 |
+
gliner
|
7 |
+
torch>=2.0.0
|
8 |
+
transformers>=4.38.2
|
9 |
+
huggingface_hub>=0.21.4
|
10 |
+
onnxruntime
|
11 |
+
optimum
|
12 |
+
peft
|
13 |
+
sentencepiece
|
14 |
+
tqdm
|
15 |
+
|