Spaces:
Running
Running
Update pages/21_NLP_Transformer.py
Browse files- pages/21_NLP_Transformer.py +56 -59
pages/21_NLP_Transformer.py
CHANGED
@@ -1,89 +1,86 @@
|
|
1 |
import torch
|
2 |
-
from torch.utils.data import DataLoader
|
3 |
from transformers import BertTokenizer, BertForSequenceClassification, AdamW, get_scheduler
|
4 |
from datasets import load_dataset
|
5 |
-
from
|
6 |
import streamlit as st
|
7 |
import matplotlib.pyplot as plt
|
|
|
8 |
|
9 |
-
# Load and
|
10 |
-
dataset = load_dataset("imdb")
|
11 |
-
train_dataset = dataset["train"]
|
12 |
-
test_dataset = dataset["test"]
|
13 |
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
|
14 |
-
|
15 |
-
def preprocess_function(examples):
|
16 |
-
return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=512)
|
17 |
-
|
18 |
-
encoded_train_dataset = train_dataset.map(preprocess_function, batched=True)
|
19 |
-
encoded_test_dataset = test_dataset.map(preprocess_function, batched=True)
|
20 |
-
train_dataloader = DataLoader(encoded_train_dataset, shuffle=True, batch_size=8)
|
21 |
-
test_dataloader = DataLoader(encoded_test_dataset, batch_size=8)
|
22 |
-
|
23 |
model = BertForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=2)
|
24 |
-
optimizer = AdamW(model.parameters(), lr=5e-5)
|
25 |
-
num_epochs = 3
|
26 |
-
num_training_steps = num_epochs * len(train_dataloader)
|
27 |
-
lr_scheduler = get_scheduler(name="linear", optimizer=optimizer, num_warmup_steps=0, num_training_steps=num_training_steps)
|
28 |
-
|
29 |
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
30 |
model.to(device)
|
31 |
|
32 |
-
#
|
33 |
-
|
34 |
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
loss = outputs.loss
|
41 |
-
loss.backward()
|
42 |
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
|
|
|
|
|
|
|
|
|
|
47 |
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
|
54 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
55 |
batch = {k: v.to(device) for k, v in batch.items()}
|
56 |
outputs = model(**batch)
|
57 |
-
|
58 |
-
|
59 |
-
total += batch["labels"].size(0)
|
60 |
-
return correct / total
|
61 |
|
62 |
-
|
63 |
-
|
|
|
|
|
|
|
64 |
|
65 |
-
|
66 |
-
st.title("Sentiment Analysis with BERT")
|
67 |
-
st.write(f"Test Accuracy: {accuracy * 100:.2f}%")
|
68 |
|
69 |
-
# Plot loss values
|
70 |
-
st.write("### Training Loss")
|
71 |
-
plt.figure(figsize=(10, 6))
|
72 |
-
plt.plot(loss_values, label="Training Loss")
|
73 |
-
plt.xlabel("Training Steps")
|
74 |
-
plt.ylabel("Loss")
|
75 |
-
plt.legend()
|
76 |
-
st.pyplot(plt)
|
77 |
|
78 |
# Text input for prediction
|
79 |
st.write("### Predict Sentiment")
|
80 |
user_input = st.text_area("Enter text:", "I loved this movie!")
|
|
|
81 |
if user_input:
|
82 |
-
inputs = tokenizer(user_input, padding="max_length", truncation=True, max_length=
|
83 |
inputs = {k: v.to(device) for k, v in inputs.items()}
|
|
|
84 |
model.eval()
|
85 |
with torch.no_grad():
|
86 |
outputs = model(**inputs)
|
87 |
prediction = outputs.logits.argmax(dim=-1).item()
|
88 |
sentiment = "Positive" if prediction == 1 else "Negative"
|
89 |
-
|
|
|
|
1 |
import torch
|
|
|
2 |
from transformers import BertTokenizer, BertForSequenceClassification, AdamW, get_scheduler
|
3 |
from datasets import load_dataset
|
4 |
+
from torch.utils.data import DataLoader
|
5 |
import streamlit as st
|
6 |
import matplotlib.pyplot as plt
|
7 |
+
from tqdm.auto import tqdm
|
8 |
|
9 |
+
# Load pre-trained model and tokenizer from Hugging Face
|
|
|
|
|
|
|
10 |
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
model = BertForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=2)
|
|
|
|
|
|
|
|
|
|
|
12 |
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
13 |
model.to(device)
|
14 |
|
15 |
+
# Streamlit interface
|
16 |
+
st.title("Sentiment Analysis with BERT")
|
17 |
|
18 |
+
# Training setup
|
19 |
+
st.sidebar.title("Training Setup")
|
20 |
+
num_epochs = st.sidebar.slider("Number of Epochs", 1, 5, 3)
|
21 |
+
batch_size = st.sidebar.slider("Batch Size", 4, 32, 8)
|
22 |
+
learning_rate = st.sidebar.slider("Learning Rate", 1e-6, 1e-3, 5e-5, format="%.6f")
|
|
|
|
|
23 |
|
24 |
+
# Load and preprocess dataset
|
25 |
+
@st.cache(allow_output_mutation=True)
|
26 |
+
def load_and_preprocess_data():
|
27 |
+
dataset = load_dataset("imdb", split="train[:1%]")
|
28 |
+
def preprocess_function(examples):
|
29 |
+
return tokenizer(examples["text"], padding="max_length", truncation=True, max_length=128)
|
30 |
+
encoded_dataset = dataset.map(preprocess_function, batched=True)
|
31 |
+
encoded_dataset.set_format(type='torch', columns=['input_ids', 'attention_mask', 'label'])
|
32 |
+
return DataLoader(encoded_dataset, shuffle=True, batch_size=batch_size)
|
33 |
|
34 |
+
train_dataloader = load_and_preprocess_data()
|
35 |
+
|
36 |
+
# Training loop
|
37 |
+
if st.sidebar.button("Train"):
|
38 |
+
optimizer = AdamW(model.parameters(), lr=learning_rate)
|
39 |
+
num_training_steps = num_epochs * len(train_dataloader)
|
40 |
+
lr_scheduler = get_scheduler(
|
41 |
+
name="linear", optimizer=optimizer, num_warmup_steps=0, num_training_steps=num_training_steps
|
42 |
+
)
|
43 |
+
|
44 |
+
progress_bar = tqdm(range(num_training_steps))
|
45 |
+
loss_values = []
|
46 |
+
|
47 |
+
model.train()
|
48 |
+
for epoch in range(num_epochs):
|
49 |
+
for batch in train_dataloader:
|
50 |
batch = {k: v.to(device) for k, v in batch.items()}
|
51 |
outputs = model(**batch)
|
52 |
+
loss = outputs.loss
|
53 |
+
loss.backward()
|
|
|
|
|
54 |
|
55 |
+
optimizer.step()
|
56 |
+
lr_scheduler.step()
|
57 |
+
optimizer.zero_grad()
|
58 |
+
progress_bar.update(1)
|
59 |
+
loss_values.append(loss.item())
|
60 |
|
61 |
+
st.sidebar.success("Training completed")
|
|
|
|
|
62 |
|
63 |
+
# Plot loss values
|
64 |
+
st.write("### Training Loss")
|
65 |
+
plt.figure(figsize=(10, 6))
|
66 |
+
plt.plot(loss_values, label="Training Loss")
|
67 |
+
plt.xlabel("Training Steps")
|
68 |
+
plt.ylabel("Loss")
|
69 |
+
plt.legend()
|
70 |
+
st.pyplot(plt)
|
71 |
|
72 |
# Text input for prediction
|
73 |
st.write("### Predict Sentiment")
|
74 |
user_input = st.text_area("Enter text:", "I loved this movie!")
|
75 |
+
|
76 |
if user_input:
|
77 |
+
inputs = tokenizer(user_input, padding="max_length", truncation=True, max_length=128, return_tensors="pt")
|
78 |
inputs = {k: v.to(device) for k, v in inputs.items()}
|
79 |
+
|
80 |
model.eval()
|
81 |
with torch.no_grad():
|
82 |
outputs = model(**inputs)
|
83 |
prediction = outputs.logits.argmax(dim=-1).item()
|
84 |
sentiment = "Positive" if prediction == 1 else "Negative"
|
85 |
+
|
86 |
+
st.write(f"Sentiment: **{sentiment}**")
|