app.py

import streamlit as st
import math
import tensorflow as tf
from transformers import GPT2Tokenizer, TFGPT2Model

st.title("GPT2 Sentiment Analysis")
st.write('Model detects if a specific tweet has positive or negative sentiment')
tweet = st.text_input("Enter your tweet", '')
#st.write(f"Hello {name}!")

PAD_TOKEN = "<|pad|>"
EOS_TOKEN = "<|endoftext|>"
MAX_LENGTH=20
# this will download and initialize the pre trained tokenizer
tokenizer = GPT2Tokenizer.from_pretrained("gpt2",pad_token=PAD_TOKEN,eos_token=EOS_TOKEN,max_length=MAX_LENGTH,is_split_into_words=True)

model = TFGPT2Model.from_pretrained("gpt2", use_cache=False,pad_token_id=tokenizer.pad_token_id,eos_token_id=tokenizer.eos_token_id)
model.training = True
model.resize_token_embeddings(len(tokenizer))

for layer in model.layers:
    layer.trainable = False
    
input = tf.keras.layers.Input(shape=(None,), dtype='int32')
mask = tf.keras.layers.Input(shape=(None,), dtype='int32')
x = model(input, attention_mask=mask)
#x = x.last_hidden_state[:, -1]
x = tf.reduce_mean(x.last_hidden_state, axis=1)
x = tf.keras.layers.Dense(16, activation='relu')(x)
x = tf.keras.layers.Dropout(0.3)(x)
output = tf.keras.layers.Dense(2, activation='softmax')(x)

clf = tf.keras.Model([input, mask], output)

clf.load_weights('./saved_weights/GPT2_sentiment')

#text="@newedge thanks for the follow, and the new icon looks great"

sample_text=[tweet]
EOS_TOKEN = "<|endoftext|>"
sample_text=[str(ex) + EOS_TOKEN for ex in sample_text]

sample_text_ = [tokenizer(str(x), return_tensors='tf', max_length=MAX_LENGTH, truncation=True, pad_to_max_length=True, add_special_tokens=True)['input_ids'] for x in sample_text]

sample_text_mask_ = [tokenizer(str(x), return_tensors='tf', max_length=MAX_LENGTH, truncation=True, pad_to_max_length=True, add_special_tokens=True)["attention_mask"] for x in sample_text]

pred = clf.predict([sample_text_, sample_text_mask_])

#pred_out = tf.math.argmax(pred, axis=1)
#pred_out=pred_out.numpy()

#st.write(f"Hello {sample_text[0]}!")
positive = round(pred[0][1],4)
negative = round(pred[0][0],4)

st.write(f"Positive Sentiment Prediction:  {positive}")
st.write(f"Negative Sentiment Prediction:  {negative}")

st.header('Below samples are outside of train or test data')
st.header('Sample Positive Sentiment Tweets')
st.write(f"watchin Espn's First Take! my favorite mornin show! lol Skip is great tv! fyi Im a Witness!")
st.write(f"I'm eating cheezits...with TWO flavors!  sharp cheddar &amp; parmesan. :-D")
st.write(f"Just drunk a coffe,but I'm still sleeping lol...now drink a fresh lemonade and eat some marshmallows mmm...then study guitar!")
st.write(f"On way home blasting mcfly in the back of the car in the sun good times ")
st.write(f"@AshenDestiny Just had a look at ur updates..quite thoughtful ones..")

st.header('Sample Negative Sentiment Tweets')
st.write(f"Man, I so desperately do NOT want to be doing this freelance work. Unfortunately, it looks like I'll be doing it the rest of the weekend.")
st.write(f"Is watching ripley's believe it or not. Totally bored.")
st.write(f"not been able to tweet today  at my dads and my sister had taken over the laptop, i was going to use my phone but it took all my credit :O")
st.write(f"Ughh I hate being broke  does anyone know of any jobs??")
st.write(f"Just realized I will miss th destroy build destroy premiere tonight.  I have failed @AndrewWK.")

st.header('Profile Links')
st.write("LinkedIn Profile: [https://www.linkedin.com/in/omkar-bhatkande/](https://www.linkedin.com/in/omkar-bhatkande/)")
st.write("Github repo: [https://github.com/omkarb09/gpt2-sentiment-analysis](https://github.com/omkarb09/gpt2-sentiment-analysis)")