app.py

import os
import torch
import spacy
import spaces
import numpy as np
import pandas as pd
from transformers import AutoModelForSequenceClassification
from transformers import AutoTokenizer
import gradio as gr
import matplotlib.pyplot as plt
from matplotlib.colors import LinearSegmentedColormap
import plotly.express as px
import seaborn as sns

PATH = '/data/' # at least 150GB storage needs to be attached
os.environ['TRANSFORMERS_CACHE'] = PATH
os.environ['HF_HOME'] = PATH
os.environ['HF_DATASETS_CACHE'] = PATH
os.environ['TORCH_HOME'] = PATH

HF_TOKEN = os.environ["hf_read"]

SENTIMENT_LABEL_NAMES = {0: "Negative", 1: "No sentiment or Neutral sentiment", 2: "Positive"}
LANGUAGES = ["Czech", "English", "French", "German", "Hungarian", "Polish", "Slovakian"]

id2label = {
    0: "Anger",
    1: "Fear",
    2: "Disgust",
    3: "Sadness",
    4: "Joy",
    5: "None of Them"
}
def load_spacy_model(model_name="xx_sent_ud_sm"):
    try:
        model = spacy.load(model_name)
    except OSError:
        spacy.cli.download(model_name)
        model = spacy.load(model_name)
    return model

def split_sentences(text, model):
    # disable pipeline components not necessary for splitting
    model.disable_pipes(model.pipe_names)  # first disable all the pipes
    model.enable_pipe("senter") # then enable the sentence splitter only

    doc = model(text)
    sentences = [sent.text for sent in doc.sents]

    return sentences

def build_huggingface_path(language: str):
    if language == "Czech" or language == "Slovakian":
        return "visegradmedia-emotion/Emotion_RoBERTa_pooled_V4"
    return "poltextlab/xlm-roberta-large-pooled-MORES"

@spaces.GPU
def predict(text, model_id, tokenizer_id):
    model = AutoModelForSequenceClassification.from_pretrained(model_id, low_cpu_mem_usage=True, device_map="auto", offload_folder="offload", token=HF_TOKEN)
    tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)

    inputs = tokenizer(text,
                       max_length=64,
                       truncation=True,
                       padding="do_not_pad",
                       return_tensors="pt")
    model.eval()

    with torch.no_grad():
        logits = model(**inputs).logits

    probs = torch.nn.functional.softmax(logits, dim=1).cpu().numpy().flatten()
    return probs

def get_most_probable_label(probs):
    label = id2label[probs.argmax()]
    probability = f"{round(100 * probs.max(), 2)}%"
    return label, probability


def prepare_heatmap_data(data):
    heatmap_data = pd.DataFrame(0.0, index=id2label.values(), columns=range(len(data)))

    for idx, row in enumerate(data):
        confidences = row["emotions"].tolist()
        for idy, confidence in enumerate(confidences):
            emotion = id2label[idy]
            heatmap_data.at[emotion, idx] = round(confidence, 4)

    heatmap_data.columns = [item["sentence"][:18]+"..." for item in data]
    return heatmap_data


def plot_emotion_heatmap(heatmap_data):
    heatmap_data = heatmap_data.T

    fig = plt.figure(figsize=(len(heatmap_data.columns) * 0.5 + 4, len(heatmap_data.index) * 0.5 + 2))

    cmap = LinearSegmentedColormap.from_list("white_to_grey", ["#ffffff", "#aaaaaa"])

    sns.heatmap(
        heatmap_data,
        annot=False,
        cmap=cmap,
        cbar=True,
        linewidths=0.5,
        linecolor='gray',
        vmin=0,
        vmax=1
    )

    plt.xlabel("Emotions")
    plt.ylabel("Sentences")
    plt.xticks(rotation=0, ha='center')
    plt.yticks(rotation=0)
    plt.tight_layout()

    return fig

def plot_sunburst_chart(heatmap_data):
    data = []
    for item in heatmap_data:
        sentence = item['sentence']
        emotions = item['emotions']

        sentence_wrapped = "\n".join([sentence[i:i + 50] for i in range(0, len(sentence), 50)])

        for i, score in enumerate(emotions):
            data.append({
                'sentence': sentence_wrapped,
                'emotion': id2label[i],
                'score': float(score)
            })

    df = pd.DataFrame(data)

    fig = px.sunburst(
        df,
        path=['sentence', 'emotion'],
        values='score',
        color='emotion',
        hover_data={'score': ':.3f'},
        title='Sentence-Level Emotion Confidence'
    )

    fig.update_layout(
        width=800,
        height=800,
        margin=dict(t=50, l=0, r=0, b=0)
    )

    return fig


def plot_average_emotion_pie(heatmap_data):
    all_emotion_scores = np.array([item['emotions'] for item in heatmap_data])
    mean_scores = all_emotion_scores.mean(axis=0)

    labels = [id2label[i] for i in range(len(mean_scores))]
    sizes = mean_scores

    # optional: remove emotions with near-zero average
    labels_filtered = []
    sizes_filtered = []
    for l, s in zip(labels, sizes):
        if s > 0.01:  # You can change this threshold
            labels_filtered.append(l)
            sizes_filtered.append(s)

    fig, ax = plt.subplots(figsize=(6, 6))
    wedges, texts, autotexts = ax.pie(
        sizes_filtered,
        labels=labels_filtered,
        autopct='%1.1f%%',
        startangle=140,
        textprops={'fontsize': 12}
    )

    ax.axis('equal')  # Equal aspect ratio ensures a circle
    plt.title("Average Emotion Confidence Across Sentences", fontsize=14)

    return fig

def plot_emotion_barplot(heatmap_data):
    most_probable_emotions = heatmap_data.idxmax(axis=0)
    emotion_counts = most_probable_emotions.value_counts()
    all_emotions = heatmap_data.index
    emotion_frequencies = (emotion_counts.reindex(all_emotions, fill_value=0) / emotion_counts.sum()).sort_values(ascending=False)
    fig, ax = plt.subplots(figsize=(8, 6))
    sns.barplot(x=emotion_frequencies.values, y=emotion_frequencies.index, palette="coolwarm", ax=ax)
    ax.set_title("Relative Frequencies of Predicted Emotions")
    ax.set_xlabel("Relative Frequency")
    ax.set_ylabel("Emotions")
    plt.tight_layout()
    return fig

def predict_wrapper(text, language):
    model_id = build_huggingface_path(language)
    tokenizer_id = "xlm-roberta-large"

    spacy_model = load_spacy_model()
    sentences = split_sentences(text, spacy_model)

    results = []
    results_heatmap = []
    for sentence in sentences:
        probs = predict(sentence, model_id, tokenizer_id)
        label, probability = get_most_probable_label(probs)
        results.append([sentence, label, probability])
        results_heatmap.append({"sentence":sentence, "emotions":probs})

    # let's see...
    print(results)
    print(results_heatmap)

    figure = plot_emotion_barplot(prepare_heatmap_data(results_heatmap))
    heatmap = plot_emotion_heatmap(prepare_heatmap_data(results_heatmap))
    piechart = plot_average_emotion_pie(prepare_heatmap_data(results_heatmap))
    output_info = f'Prediction was made using the <a href="https://huggingface.co/{model_id}">{model_id}</a> model.'
    return results, figure, piechart, heatmap, output_info


with gr.Blocks() as demo:
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(lines=6, label="Input", placeholder="Enter your text here...")
        with gr.Column():
            with gr.Row():
                language_choice = gr.Dropdown(choices=LANGUAGES, label="Language", value="English")
            with gr.Row():
                predict_button = gr.Button("Submit")

    with gr.Row():
        result_table = gr.Dataframe(
            headers=["Sentence", "Prediction", "Confidence"],
            column_widths=["65%", "25%", "10%"],
            wrap=True # important
        )

    with gr.Row():
        plot = gr.Plot()

    with gr.Row():
        piechart = gr.Plot()

    with gr.Row():
        heatmap = gr.Plot()

    with gr.Row():
        model_info = gr.Markdown()

    predict_button.click(
        fn=predict_wrapper,
        inputs=[input_text, language_choice],
        outputs=[result_table, plot, piechart, heatmap, model_info]
    )

if __name__ == "__main__":
    demo.launch()