import gradio as gr
import re

# ---------- Dictionaries of words for each "layer" ----------
# Layer 1: basic positive/negative detection
positive_words = {"love", "great", "good", "amazing", "awesome", "happy"}
negative_words = {"hate", "terrible", "awful", "bad", "horrible", "sad"}

# Layer 2: sarcasm or contradictory cues
sarcasm_cues = {"not!", "jk", "/s"}  # e.g. "The food was great, not!"

# Layer 3: intensifiers
intensifiers = {"very", "extremely", "so", "super", "really"}

# ----------------------------------------------------------------
# 1) Classification Layers
# ----------------------------------------------------------------

def layer_1_basic(text):
    """
    Layer 1:
    - Count how many known positive words vs. negative words.
    - If pos >= neg, label = "Positive", else "Negative"
    """
    words = re.findall(r"[a-zA-Z]+", text.lower())
    pos_count = sum(1 for w in words if w in positive_words)
    neg_count = sum(1 for w in words if w in negative_words)

    label = "Positive" if pos_count >= neg_count else "Negative"
    return label, pos_count, neg_count

def layer_2_sarcasm_or_flip(text, prev_label):
    """
    Layer 2:
    - Check for sarcasm cues (like "not!", "jk", "/s").
    - If found, flip the classification from Layer 1.
    """
    text_lower = text.lower()
    if any(cue in text_lower for cue in sarcasm_cues):
        flipped = "Positive" if prev_label == "Negative" else "Negative"
        return flipped + " (sarcasm-corrected)"
    else:
        return prev_label

def layer_3_intensifiers(text, prev_label):
    """
    Layer 3:
    - If there's an intensifier ("very", "extremely", "so", "super", "really"),
      we add "(HIGH CONFIDENCE)" to the label.
    """
    text_lower = text.lower()
    if any(i in text_lower for i in intensifiers):
        return prev_label + " (HIGH CONFIDENCE)"
    else:
        return prev_label

def deep_classify(text):
    """
    Pass text through 3 layers, returning a final classification
    plus details about each layer.
    """
    # LAYER 1
    layer1_label, pos_count, neg_count = layer_1_basic(text)

    # LAYER 2
    layer2_label = layer_2_sarcasm_or_flip(text, layer1_label)

    # LAYER 3
    final_label = layer_3_intensifiers(text, layer2_label)

    return layer1_label, layer2_label, final_label, pos_count, neg_count

# ----------------------------------------------------------------
# 2) Gradio Demo
# ----------------------------------------------------------------

# Some sample statements from your original dropdown
sample_statements = [
    "I love this movie!",
    "This place is awful.",
    "The food was amazing, not!",
    "I'm so happy right now.",
    "Everything was terrible... JK!",
    "I really hate slow service.",
    "This is so good, I'm super impressed.",
]

def simulate_layers(sample_choice, custom_text):
    """
    1) If the user typed something in 'custom_text', use that.
    2) Otherwise, use the selected 'sample_choice'.
    3) Run the text through the 3-layer classification.
    4) Return an HTML block describing each layer's result.
    """
    text_to_classify = ""
    if custom_text.strip():
        text_to_classify = custom_text.strip()
    elif sample_choice:
        text_to_classify = sample_choice
    else:
        return "<p style='color:red;'>Please provide a sentence.</p>"

    layer1_label, layer2_label, final_label, pos_count, neg_count = deep_classify(text_to_classify)

    # Build HTML explanation
    html_output = f"""
    <h3>Deep Learning Simulation</h3>
    <p><b>Input Text:</b> {text_to_classify}</p>

    <ol>
      <li><b>Layer 1</b> checks positive vs. negative words (simple count).
          <br>Positive words found: {pos_count}, Negative words found: {neg_count}
          <br>Layer 1 Classification: <i>{layer1_label}</i>
      </li>
      <li><b>Layer 2</b> checks for sarcasm cues (like "not!", "jk", "/s").
          <br>Result after sarcasm check: <i>{layer2_label}</i>
      </li>
      <li><b>Layer 3</b> looks for intensifiers ("very", "extremely", "so", "super", "really").
          <br>Final Classification: <span style="color:blue; font-weight:bold;">{final_label}</span>
      </li>
    </ol>

    <h4>How this relates to Deep Learning</h4>
    <ul>
      <li>In actual deep learning, each layer learns more complex
          <i>features</i> automatically from data, rather than
          using manually defined checks.</li>
      <li>However, the <b>concept</b> of passing through multiple
          layers to refine the result <i>is</i> how neural networks work.</li>
    </ul>

    <p><b>Observation:</b> This layered approach can handle nuances
    like sarcasm or intensifiers better than a single rule or single-layer model.</p>
    """
    return html_output

with gr.Blocks(css="footer{display:none !important}") as demo:
    gr.Markdown("Deep Learning (Layered Approach) Simulation")
    gr.Markdown("Pick a sample statement **or** type your own, then click **Simulate**.")

    with gr.Row():
        sample_choice = gr.Dropdown(
            label="Sample Statement",
            choices=["(None)"] + sample_statements,
            value="(None)",
        )
        custom_text = gr.Textbox(
            label="Custom Text",
            placeholder="Type your own statement here..."
        )
    simulate_button = gr.Button("Simulate Layers")
    output = gr.HTML()

    def simulate_wrapper(smpl, cstm):
        # If the user picks "(None)", interpret that as an empty choice
        if smpl == "(None)":
            smpl = ""
        return simulate_layers(smpl, cstm)

    simulate_button.click(
        fn=simulate_wrapper,
        inputs=[sample_choice, custom_text],
        outputs=[output]
    )

demo.launch()