|
import streamlit as st |
|
from bert_model import BERTEmbedding |
|
from utils import build_vocab, get_embedding_tensor |
|
from sklearn.decomposition import PCA |
|
import plotly.express as px |
|
|
|
|
|
def plot_2d_embeddings(embeddings, sentence): |
|
pca = PCA(n_components=2) |
|
reduced = pca.fit_transform(embeddings[0].detach().numpy()) |
|
fig = px.scatter( |
|
x=reduced[:, 0], |
|
y=reduced[:, 1], |
|
text=sentence.split() |
|
) |
|
st.plotly_chart(fig) |
|
|
|
def plot_3d_embeddings(embeddings, sentence): |
|
pca = PCA(n_components=3) |
|
reduced = pca.fit_transform(embeddings[0].detach().numpy()) |
|
fig = px.scatter_3d( |
|
x=reduced[:, 0], |
|
y=reduced[:, 1], |
|
z=reduced[:, 2], |
|
text=sentence.split() |
|
) |
|
st.plotly_chart(fig) |
|
|
|
|
|
N_SEGMENTS = 2 |
|
MAX_LEN = 512 |
|
EMBED_DIM = 768 |
|
N_LAYERS = 12 |
|
ATTN_HEADS = 12 |
|
DROPOUT = 0.1 |
|
|
|
|
|
def main(): |
|
st.title("BERT Embeddings Visualization") |
|
|
|
uploaded_file = st.file_uploader("Upload a text file to build vocabulary", type=['txt']) |
|
|
|
if uploaded_file: |
|
uploaded_data = uploaded_file.read().decode('utf-8').splitlines() |
|
st.success("Vocabulary built successfully!") |
|
else: |
|
st.warning("Using default vocabulary.") |
|
with open('default_vocab.txt', 'r') as file: |
|
uploaded_data = file.read().splitlines() |
|
|
|
vocab = build_vocab(uploaded_data) |
|
|
|
VOCAB_SIZE = len(vocab) |
|
embedding_layer = BERTEmbedding(VOCAB_SIZE, N_SEGMENTS, MAX_LEN, EMBED_DIM, DROPOUT) |
|
|
|
user_sentence = st.text_input("Enter your sentence:", "AI in healthcare predicts patient outcomes and diagnoses.") |
|
|
|
viz_option = st.selectbox("Select Visualization Type", ["2D", "3D"]) |
|
|
|
if st.button('Visualize Embeddings'): |
|
embedding_tensor = get_embedding_tensor(user_sentence, vocab, embedding_layer) |
|
|
|
if viz_option == "2D": |
|
plot_2d_embeddings(embedding_tensor, user_sentence) |
|
elif viz_option == "3D": |
|
plot_3d_embeddings(embedding_tensor, user_sentence) |
|
|
|
if __name__ == "__main__": |
|
main() |
