Upload 11 files

app.py

@@ -0,0 +1,56 @@
+import os
+os.environ["USE_TF"] = "0"
+import torch
+import gradio as gr
+from transformers import DebertaV2Tokenizer
+from src.model import SentiNetTransformer
+from src.config import HPARAMS
+from src.ui import build_demo
+
+# CONFIGUARATION
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+hp = HPARAMS()
+backbone_config_path = "model/config.json"
+checkpint_path = "model/SentiNet_Transformer_params.pt"
+tokenizer_path = "model/"
+
+# LOAD MODEL & TOKENIZER
+model = SentiNetTransformer(model_path=backbone_config_path, fc_dropout=hp.transformer_fc_dropout).to(device)
+state_dict = torch.load(checkpint_path, map_location=device, weights_only=True)
+model.load_state_dict(state_dict)
+tokenizer = DebertaV2Tokenizer.from_pretrained(tokenizer_path)
+
+# INFERENCE FUNCTION
+@torch.no_grad()
+def sentiment_classifier(model, tokenizer, text, thresh=0.5, max_length_trun=256, device=None):
+    if device is None:
+        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    inputs = tokenizer(text, return_tensors="pt",
+                       add_special_tokens=True, max_length=max_length_trun,
+                       truncation=True, padding=True)
+    inputs = {k: v.to(device) for k, v in inputs.items()}
+    
+    model.eval()
+    logits = model(inputs)
+    prob = torch.sigmoid(logits).cpu().numpy()[0][0]
+
+    if prob >= thresh:
+        return "😀 Positive", round(float(prob), 3)
+    else:
+        return "😞 Negative", round(float(prob), 3)
+
+# GRADIO DEMO
+def generation_fn(text):
+    return sentiment_classifier(model, tokenizer, text, max_length_trun=256, device=None)
+    
+demo = build_demo(
+    generation_fn,
+    english_title = "# SentiNet: Transformer‑Based Sentiment Classifier",
+    persian_title = "# سنتی‌نت: تحلیل احساسات با ترنسفورمر",
+    assets_dir = "assets",
+    app_title = "SentiNet"
+)
+
+if __name__ == "__main__":
+    demo.launch()

assets/css/custom.css

@@ -0,0 +1,79 @@
+/* Background */
+.gradio-container {
+    background: linear-gradient(135deg, #fdfbfb, #ebedee) !important;
+    font-family: 'Inter', 'Segoe UI', sans-serif !important;
+}
+.dark .gradio-container {
+    background: linear-gradient(135deg, #1e1a5e, #2a0a3a) !important;
+}
+
+/* Buttons */
+button {
+    border-radius: 14px !important;
+    padding: 10px 18px !important;
+    font-weight: 600 !important;
+    background: linear-gradient(90deg, #6a11cb, #2575fc) !important;
+    color: white !important;
+    box-shadow: 0 4px 10px rgba(0,0,0,0.15) !important;
+    transition: transform 0.15s ease-in-out;
+}
+button:hover {
+    transform: translateY(-2px);
+    box-shadow: 0 6px 14px rgba(0,0,0,0.25) !important;
+}
+
+/* Title */
+#title {
+    font-size: 2.8em !important;
+    font-weight: 700 !important;
+    color: #1e3a8a;
+    text-align: center;
+    margin-top: 28px;
+    margin-bottom: 12px;
+    text-shadow: 1px 2px 6px rgba(0,0,0,0.1);
+}
+.dark #title {
+    color: #e0f7fa !important;
+    text-shadow: 1px 2px 6px rgba(0,0,0,0.4);
+}
+
+/* Summary / Description */
+#summary {
+    color: #374151;
+    background: rgba(255,255,255,0.7);
+    padding: 18px;
+    border-radius: 16px;
+    box-shadow: 0 4px 12px rgba(0,0,0,0.08);
+    margin-bottom: 16px;
+    text-align: justify !important;
+}
+.dark #summary {
+    color: #d1d5db !important;
+    background: rgba(30, 30, 46, 0.6) !important;
+}
+
+/* Help / Info Box */
+#help_text {
+    color: #1f2937;
+    background: rgba(240, 249, 255, 0.9);
+    padding: 16px;
+    border-left: 5px solid #3b82f6;
+    border-radius: 14px;
+    box-shadow: 0 4px 10px rgba(0,0,0,0.05);
+    margin-top: 12px;
+    text-align: justify !important;
+}
+.dark #help_text {
+    color: #d1d5db !important;
+    background: rgba(30, 30, 46, 0.7) !important;
+    border-left: 5px solid #60a5fa !important;
+}
+
+/* RTL Support */
+.persian {
+    direction: rtl;
+    text-align: right;
+}
+#summary.persian, #help_text.persian {
+    text-align: justify !important;
+}

assets/markdown/english_summary.md

@@ -0,0 +1,4 @@
+**SentiNet🤖** is an experimental project exploring different approaches to sentiment classification, with a focus on handling nuanced language phenomena such as sarcasm, shifting tones, and negation. 
+By fine‑tuning the Microsoft DeBERTa‑v3 encoder and comparing it against classic machine learning baselines and recurrent models, SentiNet demonstrates how modern Transformers capture contextual meaning beyond word‑level cues.
+The system highlights the strengths and weaknesses of each approach while providing an interactive demo that outputs clear sentiment labels (😀 Positive / 😞 Negative) alongside confidence scores, making evaluation both rigorous and accessible.  
+Project GitHub: [https://github.com/HooMAN/SentiNet](https://github.com/HooMAN/SentiNet)

assets/markdown/persian_summary.md

@@ -0,0 +1,3 @@
+**سنتی‌نت🤖** یک پروژه‌ی آزمایشی برای بررسی رویکردهای مختلف در دسته‌بندی احساسات است؛ تمرکز اصلی آن بر چالش‌های زبانی مانند طعنه، تغییر لحن در طول جمله و نفی می‌باشد. 
+این مدل با استفاده از ترنسفورمر DeBERTa‑v3 مایکروسافت و آموزش روی داده‌های نقد فیلم، توانسته فراتر از روش‌های کلاسیک عمل کند و معنای متنی را در سطح جمله درک کند. سنتی‌نت علاوه بر مقایسه‌ی رویکردهای سنتی و بازگشتی با ترنسفورمر، یک دموی تعاملی ارائه می‌دهد که نتیجه را همراه با برچسب احساسی (😀 مثبت / 😞 منفی) و میزان اطمینان نمایش می‌دهد تا ارزیابی برای کاربر ساده و شفاف باشد.  
+پروژه در گیت‌هاب: [https://github.com/HooMAN/SentiNet](https://github.com/HooMAN/SentiNet)

model/config.json

@@ -0,0 +1,22 @@
+{
+		"model_type": "deberta-v2",
+		"attention_probs_dropout_prob": 0.1,
+		"hidden_act": "gelu",
+		"hidden_dropout_prob": 0.1,
+		"hidden_size": 768,
+		"initializer_range": 0.02,
+		"intermediate_size": 3072,
+		"max_position_embeddings": 512,
+		"relative_attention": true,
+		"position_buckets": 256,
+		"norm_rel_ebd": "layer_norm",
+		"share_att_key": true,
+		"pos_att_type": "p2c|c2p",
+		"layer_norm_eps": 1e-7,
+		"max_relative_positions": -1,
+		"position_biased_input": false,
+		"num_attention_heads": 12,
+		"num_hidden_layers": 12,
+		"type_vocab_size": 0,
+		"vocab_size": 128100
+}

model/spm.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c679fbf93643d19aab7ee10c0b99e460bdbc02fedf34b92b05af343b4af586fd
+size 2464616

model/tokenizer_config.json

@@ -0,0 +1,4 @@
+{
+  "do_lower_case": false,
+  "vocab_type": "spm"
+}

requirements.txt

@@ -0,0 +1,4 @@
+torch
+transformers
+gradio
+numpy

src/config.py

@@ -0,0 +1,66 @@
+from dataclasses import dataclass, field
+
+# ---- Hyperparameter configuration ----
+
+@dataclass
+class HPARAMS:
+    # common
+    seed: int = 42
+    url: str = "https://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz"
+
+    scheduler_hparams: dict = field(default_factory=lambda: {
+        "factor": 0.5,
+        "patience": 2,
+        "mode": "min"
+    })
+    
+    # BiGRU hparams
+    max_seq_len_gru: int = 256
+    batch_size_gru: int = 128
+    vocab_size: int = 10000
+    glove_txt_path: str = ("/mnt/e/ML_Files/PreTrained_Models/GloVe_Embeddings/glove.2024.wikigiga"
+                           ".200d/wiki_giga_2024_200_MFT20_vectors_seed_2024_alpha_0.75_eta_0.05_combined.txt")
+    
+    model_hparams_gru: dict = field(default_factory=lambda: {
+        "embedding_dim": 128,
+        "hidden_size": 128,
+        "dropout": 0.12,
+        "num_gru_layers": 2,
+        "use_dense": False,
+        "dense_dropout_prob": 0.1
+    })
+
+    optimizer_hparams_gru: dict = field(default_factory=lambda: {
+        "lr": 1e-3,
+        "weight_decay": 5e-4
+    })
+
+    trainer_hparams_gru: dict = field(default_factory=lambda: {
+        "n_epochs": 20,
+        "use_early_stopping" : True,
+        "early_stopping_patience" : 3,
+        "scheduler_monitor" : "val_loss",
+        "restore_best_model": True,
+    })
+    
+    # Transformer hparams
+    max_seq_len_transformer:int = 288
+    #transformer_path: str = "/mnt/e/ML_Files/PreTrained_Models/HuggingFace/deberta-v3-base/"
+    transformer_path: str = "/mnt/d/ML-Files/PreTrained-Models/HuggingFace/Transformer-Encoder/microsoft_deberta-v3-base/"
+    batch_size_transformer: int = 32
+    transformer_fc_dropout: float = 0.1
+    
+    optimizer_hparams_transformer: dict = field(default_factory=lambda: {
+        "lr": 3e-5,
+        "weight_decay": 5e-4
+    })
+
+    trainer_hparams_transformer: dict = field(default_factory=lambda: {
+        "n_epochs": 5,
+        "use_early_stopping" : True,
+        "early_stopping_patience" : 2,
+        "scheduler_monitor" : "val_loss",
+        "restore_best_model": False,
+    })
+    
+hp = HPARAMS()

src/model.py

@@ -0,0 +1,25 @@
+import torch
+import torch.nn as nn
+from transformers import AutoModel, DebertaV2Tokenizer, AutoConfig
+
+class SentiNetTransformer(nn.Module):
+    """Sentiment classifier built on top of a pretrained Transformer backbone."""
+    
+    def __init__(self, model_path: str, fc_dropout: float = 0.1):
+        super().__init__()
+        config = AutoConfig.from_pretrained(model_path)
+        self.transformer = AutoModel.from_config(config)
+        hidden_dim = self.transformer.config.hidden_size
+
+        self.fc = nn.Sequential(
+            nn.Linear(hidden_dim, hidden_dim),
+            nn.ReLU(),
+            nn.Dropout(fc_dropout)
+        )
+        self.output = nn.Linear(hidden_dim, 1)
+
+    def forward(self, encodings: dict):
+        transformer_outputs = self.transformer(**encodings)  # last_hidden_state: (N, L, H)
+        cls_embedding = transformer_outputs.last_hidden_state[:, 0, :]  # CLS token: (N, H)
+        x = self.fc(cls_embedding)  # (N, H)
+        return self.output(x)  # (N, 1)

src/ui.py

@@ -0,0 +1,77 @@
+import os
+import gradio as gr
+
+def read_file(path: str, default_content: str = "") -> str:
+    """
+    Ensure file exists (with default_content if missing) and return its contents.
+    """
+    os.makedirs(os.path.dirname(path), exist_ok=True)
+    if not os.path.exists(path):
+        with open(path, "w", encoding="utf-8") as f:
+            f.write(default_content)
+    with open(path, "r", encoding="utf-8") as f:
+        return f.read()
+        
+
+def build_demo(
+    generation_fn,
+    english_title: str,
+    persian_title: str,
+    assets_dir: str = "assets",
+    app_title: str = "SentiNet Demo"
+):
+    md_dir = os.path.join(assets_dir, "markdown")
+    css_dir = os.path.join(assets_dir, "css")
+    english_md = os.path.join(md_dir, "english_summary.md")
+    persian_md = os.path.join(md_dir, "persian_summary.md")
+    english_summary = read_file(english_md)
+    persian_summary = read_file(persian_md)
+
+    css_file = os.path.join(css_dir, "custom.css")
+    css = read_file(css_file, "/* Custom CSS overrides */\n")
+
+    with gr.Blocks(css=css, title=app_title) as demo:
+        title_md = gr.Markdown(english_title, elem_id="title")
+
+        with gr.Row():
+            english_btn = gr.Button("English")
+            persian_btn = gr.Button("فارسی (Persian)")
+
+        summary_md = gr.Markdown(english_summary, elem_id="summary")
+
+        # generation panel
+        with gr.Row(variant="panel"):
+            with gr.Column(scale=1, variant="panel"):
+                text_inp = gr.Textbox(
+                    label="Enter a sentence",
+                    placeholder="Type your movie review here..."
+                )
+                generate_btn = gr.Button("🔍 Classify Sentiment", variant="primary")
+
+            with gr.Column(scale=1, variant="panel"):
+                sentiment_label = gr.Label(label="Prediction")
+                confidence_out = gr.Number(label="Confidence Score")
+
+        # events
+        generate_btn.click(
+            generation_fn,
+            inputs=[text_inp],
+            outputs=[sentiment_label, confidence_out]
+        )
+
+        def set_english():
+            return (
+                gr.update(value=english_title, elem_classes=[]),
+                gr.update(value=english_summary, elem_classes=[]),
+            )
+
+        def set_persian():
+            return (
+                gr.update(value=persian_title, elem_classes=["persian"]),
+                gr.update(value=persian_summary, elem_classes=["persian"]),
+            )
+
+        english_btn.click(set_english, outputs=[title_md, summary_md])
+        persian_btn.click(set_persian, outputs=[title_md, summary_md])
+
+    return demo