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 matplotlib.colors as mcolors
import plotly.express as px
import seaborn as sns
from tqdm import tqdm

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

css = '''
.info {font-size: 3em; !important}
'''

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"
}

emotion_colors = {
    "Anger": "#D96459",
    "Fear": "#6A8EAE",
    "Disgust": "#A4C639",
    "Sadness": "#9DBCD4",
    "Joy": "#F3E9A8",
    "None of Them": "#C0C0C0"
}
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-emotions6"

@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):
    # Transpose: now rows = sentences, columns = emotions
    heatmap_data = heatmap_data.T

    # Normalize each row (sentence-wise)
    normalized_data = heatmap_data.copy()
    for row in normalized_data.index:
        max_val = normalized_data.loc[row].max()
        normalized_data.loc[row] = normalized_data.loc[row] / max_val if max_val > 0 else 0

    # Create color matrix
    color_matrix = np.empty((len(normalized_data.index), len(normalized_data.columns), 3))
    for i, sentence in enumerate(normalized_data.index):
        for j, emotion in enumerate(normalized_data.columns):
            val = normalized_data.loc[sentence, emotion]
            base_rgb = mcolors.to_rgb(emotion_colors[emotion])
            # Blend from white to base color
            blended = tuple(1 - val * (1 - c) for c in base_rgb)
            color_matrix[i, j] = blended

    fig, ax = plt.subplots(figsize=(len(normalized_data.columns) * 0.8 + 2, len(normalized_data.index) * 0.5 + 2))
    ax.imshow(color_matrix, aspect='auto')

    # Set ticks and labels
    ax.set_xticks(np.arange(len(normalized_data.columns)))
    ax.set_xticklabels(normalized_data.columns, rotation=45, ha='right', fontsize=10)

    ax.set_yticks(np.arange(len(normalized_data.index)))
    ax.set_yticklabels(normalized_data.index, rotation=0, fontsize=10)

    ax.set_xlabel("Emotions")
    ax.set_ylabel("Sentences")

    plt.tight_layout()
    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},
        colors=[emotion_colors[l] for l in labels_filtered]
    )

    ax.axis('equal')  # Equal aspect ratio ensures a circle
    plt.title("Average Emotion Confidence Across Sentences", fontsize=14, pad=15)

    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)
    palette = [emotion_colors[emotion] for emotion in emotion_frequencies.index]
    fig, ax = plt.subplots(figsize=(8, 6))
    sns.barplot(x=emotion_frequencies.values, y=emotion_frequencies.index, palette=palette, 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 tqdm(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(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(css=css) as demo:
    placeholder = "Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua."
    introduction = """This platform is designed to detect and visualize emotions in text. The model behind it operates using a 6-label codebook, including the following labels: ‘Anger’, ‘Fear’, ‘Disgust’, ‘Sadness’, ‘Joy’, and ‘None of Them’.
    The model is optimized for sentence-level analysis, and make predictions in the following languages: Czech, English, French, German, Hungarian, Polish, and Slovak.
    The text you enter in the input box is automatically divided into sentences, and the analysis is performed on each sentence. Depending on the length of the text, this process may take a few seconds, but for longer texts, it can take up to 2-3 minutes.
    """
    gr.Markdown(introduction, elem_classes="info")
    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():
        with gr.Column(scale=7):
            piechart = gr.Plot()
        with gr.Column(scale=3):
            gr.Markdown("The following pie chart shows the average probabilities of all emotions associated with the entire text.", elem_classes="info")

    with gr.Row():
        with gr.Column(scale=7):
            plot = gr.Plot()
        with gr.Column(scale=3):
            gr.Markdown("The bar plot represents the relative frequency of emotion predictions at the sentence level.", elem_classes="info")

    with gr.Row():
        with gr.Column(scale=7):
            result_table = gr.Dataframe(
                headers=["Sentence", "Prediction", "Confidence"],
                column_widths=["65%", "25%", "10%"],
                wrap=True  # important
            )
        with gr.Column(scale=3):
            gr.Markdown("The table displays the predicted emotion and its confidence value for each sentence in the text.", elem_classes="info")

    with gr.Row():
        with gr.Column(scale=7):
            heatmap = gr.Plot()
        with gr.Column(scale=3):
            gr.Markdown("The heatmap visualizes the probability of emotions within each sentence. The darker the color, the stronger the presence of the given emotion.", elem_classes="info")

    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()