{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "7e66aa3d-a00c-491a-91ad-2a64feb2805d",
   "metadata": {},
   "source": [
    "# PHASE 1: EXPLAIN & BREAKDOWN (LEARNING PHASE)\n",
    "\n",
    "## 1. Simple Explanation of Transformers\n",
    "\n",
    "Transformers are a revolutionary neural network architecture that processes sequences (like sentences) by looking at all parts simultaneously rather than one word at a time. Think of it like reading a paragraph where you can instantly see how each word relates to every other word, rather than reading left-to-right. The key innovation is the \"attention mechanism\" - imagine highlighting the most important words in a sentence that help you understand the current word you're focusing on. This parallel processing makes Transformers much faster to train than older models like RNNs, and their ability to capture long-range dependencies makes them incredibly powerful for language tasks. They're the foundation behind ChatGPT, BERT, and most modern AI language models.\n",
    "\n",
    "## 2. Detailed Roadmap\n",
    "\n",
    "**Step 1: Foundation Concepts**\n",
    "- Sequence-to-sequence problems\n",
    "- Limitations of RNNs and CNNs for sequences\n",
    "- Parallel processing vs sequential processing\n",
    "\n",
    "**Step 2: Core Attention Mechanism**\n",
    "- Self-attention concept\n",
    "- Query, Key, Value matrices\n",
    "- Attention scores and weights\n",
    "- Scaled dot-product attention\n",
    "\n",
    "**Step 3: Multi-Head Attention**\n",
    "- Why multiple attention heads\n",
    "- Parallel attention computations\n",
    "- Concatenation and linear transformation\n",
    "\n",
    "**Step 4: Positional Encoding**\n",
    "- Why position matters in sequences\n",
    "- Sinusoidal encoding\n",
    "- Learned vs fixed positional embeddings\n",
    "\n",
    "**Step 5: Encoder Architecture**\n",
    "- Layer normalization\n",
    "- Residual connections\n",
    "- Feed-forward networks\n",
    "- Encoder block stacking\n",
    "\n",
    "**Step 6: Decoder Architecture**\n",
    "- Masked self-attention\n",
    "- Encoder-decoder attention\n",
    "- Autoregressive generation\n",
    "\n",
    "**Step 7: Training and Optimization**\n",
    "- Loss functions for different tasks\n",
    "- Teacher forcing\n",
    "- Beam search for inference\n",
    "\n",
    "## 3. Key Formulas\n",
    "\n",
    "**Attention Formula:**\n",
    "$$\\text{Attention}(Q, K, V) = \\text{softmax}\\left(\\frac{QK^T}{\\sqrt{d_k}}\\right)V$$\n",
    "\n",
    "- Q (Query): What we're looking for (d_model × d_k)\n",
    "- K (Key): What we're looking at (d_model × d_k) \n",
    "- V (Value): The actual information we extract (d_model × d_v)\n",
    "- $d_k$: Dimension of key vectors (typically d_model/num_heads)\n",
    "- $\\sqrt{d_k}$: Scaling factor to prevent softmax saturation\n",
    "\n",
    "**Multi-Head Attention:**\n",
    "$$\\text{MultiHead}(Q, K, V) = \\text{Concat}(\\text{head}_1, ..., \\text{head}_h)W^O$$\n",
    "$$\\text{head}_i = \\text{Attention}(QW_i^Q, KW_i^K, VW_i^V)$$\n",
    "\n",
    "- $W_i^Q, W_i^K, W_i^V$: Learned projection matrices for each head\n",
    "- $W^O$: Output projection matrix (d_model × d_model)\n",
    "- h: Number of attention heads\n",
    "\n",
    "**Positional Encoding:**\n",
    "$$PE_{(pos, 2i)} = \\sin(pos/10000^{2i/d_{model}})$$\n",
    "$$PE_{(pos, 2i+1)} = \\cos(pos/10000^{2i/d_{model}})$$\n",
    "\n",
    "- pos: Position in sequence\n",
    "- i: Dimension index\n",
    "- Creates unique encoding for each position\n",
    "\n",
    "## 4. Step-by-Step Numerical Example\n",
    "\n",
    "Let's work through a simple attention calculation with a 3-word sequence: \"I love AI\"\n",
    "\n",
    "**Given:**\n",
    "- Sequence length: 3\n",
    "- Embedding dimension: 4\n",
    "- Single attention head\n",
    "\n",
    "**Step 1: Input Embeddings**\n",
    "```\n",
    "X = [[1, 0, 1, 0],    # \"I\"\n",
    "     [0, 1, 0, 1],    # \"love\" \n",
    "     [1, 1, 0, 0]]    # \"AI\"\n",
    "```\n",
    "\n",
    "**Step 2: Create Q, K, V matrices**\n",
    "```\n",
    "W_Q = [[1, 0], [0, 1], [1, 1], [0, 0]]  # 4×2\n",
    "W_K = [[0, 1], [1, 0], [0, 1], [1, 1]]  # 4×2  \n",
    "W_V = [[1, 1], [0, 1], [1, 0], [0, 0]]  # 4×2\n",
    "```\n",
    "\n",
    "**Step 3: Compute Q, K, V**\n",
    "```\n",
    "Q = X × W_Q = [[2, 1], [1, 2], [1, 1]]  # 3×2\n",
    "K = X × W_K = [[1, 2], [2, 1], [1, 1]]  # 3×2\n",
    "V = X × W_V = [[2, 1], [1, 1], [1, 1]]  # 3×2\n",
    "```\n",
    "\n",
    "**Step 4: Attention Scores**\n",
    "```\n",
    "Scores = Q × K^T / √2 = [[2, 1], [1, 2], [1, 1]] × [[1, 2, 1], [2, 1, 1]] / 1.41\n",
    "       = [[4, 5, 3], [5, 4, 3], [3, 3, 2]] / 1.41\n",
    "       = [[2.83, 3.54, 2.12], [3.54, 2.83, 2.12], [2.12, 2.12, 1.41]]\n",
    "```\n",
    "\n",
    "**Step 5: Softmax**\n",
    "```\n",
    "Attention_weights = softmax(Scores) ≈ \n",
    "[[0.27, 0.49, 0.24], [0.49, 0.27, 0.24], [0.33, 0.33, 0.33]]\n",
    "```\n",
    "\n",
    "**Step 6: Final Output**\n",
    "```\n",
    "Output = Attention_weights × V = \n",
    "[[0.27×2 + 0.49×1 + 0.24×1, 0.27×1 + 0.49×1 + 0.24×1],\n",
    " [0.49×2 + 0.27×1 + 0.24×1, 0.49×1 + 0.27×1 + 0.24×1],\n",
    " [0.33×2 + 0.33×1 + 0.33×1, 0.33×1 + 0.33×1 + 0.33×1]]\n",
    "= [[1.27, 1.00], [1.51, 1.00], [1.33, 1.00]]\n",
    "```\n",
    "\n",
    "## 5. Real-World AI Use Case\n",
    "\n",
    "**Machine Translation (Google Translate)**\n",
    "Transformers revolutionized translation by allowing the model to simultaneously consider all words in a sentence when translating each word. For example, when translating \"The bank of the river is steep\" from English to French, the attention mechanism helps the model understand that \"bank\" refers to a riverbank (not a financial institution) by paying attention to \"river\" and \"steep\" simultaneously. This contextual understanding leads to more accurate translations: \"La rive de la rivière est escarpée\" rather than incorrectly using the financial term for bank.\n",
    "\n",
    "## 6. Tips for Mastering Transformers\n",
    "\n",
    "**Practice Sources:**\n",
    "- Implement attention mechanism from scratch using NumPy\n",
    "- Work through the \"Annotated Transformer\" tutorial\n",
    "- Practice with Hugging Face Transformers library\n",
    "- Build simple sequence-to-sequence tasks\n",
    "\n",
    "**Essential Resources:**\n",
    "- \"Attention Is All You Need\" paper (original Transformer paper)\n",
    "- Jay Alammar's \"Illustrated Transformer\" blog post\n",
    "- Harvard's \"Annotated Transformer\" \n",
    "- Stanford CS224N lectures on Transformers\n",
    "- 3Blue1Brown's attention mechanism visualization\n",
    "\n",
    "**Recommended Problems:**\n",
    "- Implement scaled dot-product attention\n",
    "- Build a simple sentiment classifier using pre-trained BERT\n",
    "- Create a text summarization model\n",
    "- Experiment with different positional encodings\n",
    "- Compare Transformer performance with RNN/LSTM on same task\n",
    "\n",
    "**Key Focus Areas:**\n",
    "- Understand why attention works better than recurrence\n",
    "- Master matrix dimensions in multi-head attention\n",
    "- Practice debugging attention weight visualizations\n",
    "- Learn to fine-tune pre-trained models\n",
    "- Understand the trade-offs between model size and performance\n",
    "\n",
    "Ready to move to Phase 2? Just say \"Understood\" and I'll provide the complete Python implementation with logging!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "f3688215-4c70-4334-bcf8-41d0e9e22dbc",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Requirement already satisfied: torch in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (2.7.1)\n",
      "Requirement already satisfied: torchvision in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (0.22.1)\n",
      "Requirement already satisfied: numpy in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (2.3.1)\n",
      "Requirement already satisfied: pandas in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (2.3.1)\n",
      "Requirement already satisfied: matplotlib in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (3.10.3)\n",
      "Requirement already satisfied: scikit-learn in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (1.7.0)\n",
      "Collecting transformers\n",
      "  Downloading transformers-4.53.2-py3-none-any.whl.metadata (40 kB)\n",
      "Collecting datasets\n",
      "  Downloading datasets-4.0.0-py3-none-any.whl.metadata (19 kB)\n",
      "Collecting accelerate\n",
      "  Downloading accelerate-1.8.1-py3-none-any.whl.metadata (19 kB)\n",
      "Requirement already satisfied: filelock in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from torch) (3.18.0)\n",
      "Requirement already satisfied: typing-extensions>=4.10.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from torch) (4.14.1)\n",
      "Requirement already satisfied: setuptools in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from torch) (80.9.0)\n",
      "Requirement already satisfied: sympy>=1.13.3 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from torch) (1.14.0)\n",
      "Requirement already satisfied: networkx in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from torch) (3.5)\n",
      "Requirement already satisfied: jinja2 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from torch) (3.1.6)\n",
      "Requirement already satisfied: fsspec in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from torch) (2025.5.1)\n",
      "Requirement already satisfied: pillow!=8.3.*,>=5.3.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from torchvision) (11.3.0)\n",
      "Requirement already satisfied: python-dateutil>=2.8.2 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from pandas) (2.9.0.post0)\n",
      "Requirement already satisfied: pytz>=2020.1 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from pandas) (2025.2)\n",
      "Requirement already satisfied: tzdata>=2022.7 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from pandas) (2025.2)\n",
      "Requirement already satisfied: contourpy>=1.0.1 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from matplotlib) (1.3.2)\n",
      "Requirement already satisfied: cycler>=0.10 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from matplotlib) (0.12.1)\n",
      "Requirement already satisfied: fonttools>=4.22.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from matplotlib) (4.58.5)\n",
      "Requirement already satisfied: kiwisolver>=1.3.1 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from matplotlib) (1.4.8)\n",
      "Requirement already satisfied: packaging>=20.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from matplotlib) (25.0)\n",
      "Requirement already satisfied: pyparsing>=2.3.1 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from matplotlib) (3.2.3)\n",
      "Requirement already satisfied: scipy>=1.8.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from scikit-learn) (1.16.0)\n",
      "Requirement already satisfied: joblib>=1.2.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from scikit-learn) (1.5.1)\n",
      "Requirement already satisfied: threadpoolctl>=3.1.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from scikit-learn) (3.6.0)\n",
      "Collecting huggingface-hub<1.0,>=0.30.0 (from transformers)\n",
      "  Downloading huggingface_hub-0.33.4-py3-none-any.whl.metadata (14 kB)\n",
      "Requirement already satisfied: pyyaml>=5.1 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from transformers) (6.0.2)\n",
      "Collecting regex!=2019.12.17 (from transformers)\n",
      "  Downloading regex-2024.11.6-cp313-cp313-macosx_11_0_arm64.whl.metadata (40 kB)\n",
      "Requirement already satisfied: requests in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from transformers) (2.32.4)\n",
      "Collecting tokenizers<0.22,>=0.21 (from transformers)\n",
      "  Downloading tokenizers-0.21.2-cp39-abi3-macosx_11_0_arm64.whl.metadata (6.8 kB)\n",
      "Collecting safetensors>=0.4.3 (from transformers)\n",
      "  Downloading safetensors-0.5.3-cp38-abi3-macosx_11_0_arm64.whl.metadata (3.8 kB)\n",
      "Requirement already satisfied: tqdm>=4.27 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from transformers) (4.67.1)\n",
      "Collecting hf-xet<2.0.0,>=1.1.2 (from huggingface-hub<1.0,>=0.30.0->transformers)\n",
      "  Downloading hf_xet-1.1.5-cp37-abi3-macosx_11_0_arm64.whl.metadata (879 bytes)\n",
      "Collecting pyarrow>=15.0.0 (from datasets)\n",
      "  Downloading pyarrow-20.0.0-cp313-cp313-macosx_12_0_arm64.whl.metadata (3.3 kB)\n",
      "Collecting dill<0.3.9,>=0.3.0 (from datasets)\n",
      "  Downloading dill-0.3.8-py3-none-any.whl.metadata (10 kB)\n",
      "Collecting xxhash (from datasets)\n",
      "  Downloading xxhash-3.5.0-cp313-cp313-macosx_11_0_arm64.whl.metadata (12 kB)\n",
      "Collecting multiprocess<0.70.17 (from datasets)\n",
      "  Downloading multiprocess-0.70.16-py312-none-any.whl.metadata (7.2 kB)\n",
      "Collecting fsspec (from torch)\n",
      "  Downloading fsspec-2025.3.0-py3-none-any.whl.metadata (11 kB)\n",
      "Collecting aiohttp!=4.0.0a0,!=4.0.0a1 (from fsspec[http]<=2025.3.0,>=2023.1.0->datasets)\n",
      "  Downloading aiohttp-3.12.14-cp313-cp313-macosx_11_0_arm64.whl.metadata (7.6 kB)\n",
      "Requirement already satisfied: psutil in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from accelerate) (7.0.0)\n",
      "Collecting aiohappyeyeballs>=2.5.0 (from aiohttp!=4.0.0a0,!=4.0.0a1->fsspec[http]<=2025.3.0,>=2023.1.0->datasets)\n",
      "  Downloading aiohappyeyeballs-2.6.1-py3-none-any.whl.metadata (5.9 kB)\n",
      "Collecting aiosignal>=1.4.0 (from aiohttp!=4.0.0a0,!=4.0.0a1->fsspec[http]<=2025.3.0,>=2023.1.0->datasets)\n",
      "  Downloading aiosignal-1.4.0-py3-none-any.whl.metadata (3.7 kB)\n",
      "Requirement already satisfied: attrs>=17.3.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from aiohttp!=4.0.0a0,!=4.0.0a1->fsspec[http]<=2025.3.0,>=2023.1.0->datasets) (25.3.0)\n",
      "Collecting frozenlist>=1.1.1 (from aiohttp!=4.0.0a0,!=4.0.0a1->fsspec[http]<=2025.3.0,>=2023.1.0->datasets)\n",
      "  Downloading frozenlist-1.7.0-cp313-cp313-macosx_11_0_arm64.whl.metadata (18 kB)\n",
      "Collecting multidict<7.0,>=4.5 (from aiohttp!=4.0.0a0,!=4.0.0a1->fsspec[http]<=2025.3.0,>=2023.1.0->datasets)\n",
      "  Downloading multidict-6.6.3-cp313-cp313-macosx_11_0_arm64.whl.metadata (5.3 kB)\n",
      "Collecting propcache>=0.2.0 (from aiohttp!=4.0.0a0,!=4.0.0a1->fsspec[http]<=2025.3.0,>=2023.1.0->datasets)\n",
      "  Downloading propcache-0.3.2-cp313-cp313-macosx_11_0_arm64.whl.metadata (12 kB)\n",
      "Collecting yarl<2.0,>=1.17.0 (from aiohttp!=4.0.0a0,!=4.0.0a1->fsspec[http]<=2025.3.0,>=2023.1.0->datasets)\n",
      "  Downloading yarl-1.20.1-cp313-cp313-macosx_11_0_arm64.whl.metadata (73 kB)\n",
      "Requirement already satisfied: idna>=2.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from yarl<2.0,>=1.17.0->aiohttp!=4.0.0a0,!=4.0.0a1->fsspec[http]<=2025.3.0,>=2023.1.0->datasets) (3.10)\n",
      "Requirement already satisfied: six>=1.5 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from python-dateutil>=2.8.2->pandas) (1.17.0)\n",
      "Requirement already satisfied: charset_normalizer<4,>=2 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from requests->transformers) (3.4.2)\n",
      "Requirement already satisfied: urllib3<3,>=1.21.1 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from requests->transformers) (2.5.0)\n",
      "Requirement already satisfied: certifi>=2017.4.17 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from requests->transformers) (2025.7.9)\n",
      "Requirement already satisfied: mpmath<1.4,>=1.1.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from sympy>=1.13.3->torch) (1.3.0)\n",
      "Requirement already satisfied: MarkupSafe>=2.0 in /Users/karthik/Desktop/importants/venv/lib/python3.13/site-packages (from jinja2->torch) (3.0.2)\n",
      "Downloading transformers-4.53.2-py3-none-any.whl (10.8 MB)\n",
      "\u001b[2K   \u001b[38;2;114;156;31m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m10.8/10.8 MB\u001b[0m \u001b[31m18.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0mMB/s\u001b[0m eta \u001b[36m0:00:01\u001b[0m:01\u001b[0m\n",
      "\u001b[?25hDownloading huggingface_hub-0.33.4-py3-none-any.whl (515 kB)\n",
      "Downloading hf_xet-1.1.5-cp37-abi3-macosx_11_0_arm64.whl (2.6 MB)\n",
      "\u001b[2K   \u001b[38;2;114;156;31m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m2.6/2.6 MB\u001b[0m \u001b[31m17.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
      "\u001b[?25hDownloading tokenizers-0.21.2-cp39-abi3-macosx_11_0_arm64.whl (2.7 MB)\n",
      "\u001b[2K   \u001b[38;2;114;156;31m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m2.7/2.7 MB\u001b[0m \u001b[31m19.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
      "\u001b[?25hDownloading datasets-4.0.0-py3-none-any.whl (494 kB)\n",
      "Downloading dill-0.3.8-py3-none-any.whl (116 kB)\n",
      "Downloading fsspec-2025.3.0-py3-none-any.whl (193 kB)\n",
      "Downloading multiprocess-0.70.16-py312-none-any.whl (146 kB)\n",
      "Downloading accelerate-1.8.1-py3-none-any.whl (365 kB)\n",
      "Downloading aiohttp-3.12.14-cp313-cp313-macosx_11_0_arm64.whl (465 kB)\n",
      "Downloading multidict-6.6.3-cp313-cp313-macosx_11_0_arm64.whl (43 kB)\n",
      "Downloading yarl-1.20.1-cp313-cp313-macosx_11_0_arm64.whl (88 kB)\n",
      "Downloading aiohappyeyeballs-2.6.1-py3-none-any.whl (15 kB)\n",
      "Downloading aiosignal-1.4.0-py3-none-any.whl (7.5 kB)\n",
      "Downloading frozenlist-1.7.0-cp313-cp313-macosx_11_0_arm64.whl (45 kB)\n",
      "Downloading propcache-0.3.2-cp313-cp313-macosx_11_0_arm64.whl (41 kB)\n",
      "Downloading pyarrow-20.0.0-cp313-cp313-macosx_12_0_arm64.whl (30.8 MB)\n",
      "\u001b[2K   \u001b[38;2;114;156;31m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m30.8/30.8 MB\u001b[0m \u001b[31m21.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0mm eta \u001b[36m0:00:01\u001b[0m0:01\u001b[0m02\u001b[0m\n",
      "\u001b[?25hDownloading regex-2024.11.6-cp313-cp313-macosx_11_0_arm64.whl (284 kB)\n",
      "Downloading safetensors-0.5.3-cp38-abi3-macosx_11_0_arm64.whl (418 kB)\n",
      "Downloading xxhash-3.5.0-cp313-cp313-macosx_11_0_arm64.whl (30 kB)\n",
      "Installing collected packages: xxhash, safetensors, regex, pyarrow, propcache, multidict, hf-xet, fsspec, frozenlist, dill, aiohappyeyeballs, yarl, multiprocess, huggingface-hub, aiosignal, tokenizers, aiohttp, accelerate, transformers, datasets\n",
      "\u001b[2K  Attempting uninstall: fsspecm\u001b[38;5;237m╺\u001b[0m\u001b[38;5;237m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m 3/20\u001b[0m [pyarrow]\n",
      "\u001b[2K    Found existing installation: fsspec 2025.5.15;237m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m 3/20\u001b[0m [pyarrow]\n",
      "\u001b[2K    Uninstalling fsspec-2025.5.1:;237m╺\u001b[0m\u001b[38;5;237m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m 3/20\u001b[0m [pyarrow]\n",
      "\u001b[2K      Successfully uninstalled fsspec-2025.5.18;5;237m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m 3/20\u001b[0m [pyarrow]\n",
      "\u001b[2K   \u001b[38;2;114;156;31m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m20/20\u001b[0m [datasets]\u001b[0m \u001b[32m19/20\u001b[0m [datasets]ers]\n",
      "\u001b[1A\u001b[2KSuccessfully installed accelerate-1.8.1 aiohappyeyeballs-2.6.1 aiohttp-3.12.14 aiosignal-1.4.0 datasets-4.0.0 dill-0.3.8 frozenlist-1.7.0 fsspec-2025.3.0 hf-xet-1.1.5 huggingface-hub-0.33.4 multidict-6.6.3 multiprocess-0.70.16 propcache-0.3.2 pyarrow-20.0.0 regex-2024.11.6 safetensors-0.5.3 tokenizers-0.21.2 transformers-4.53.2 xxhash-3.5.0 yarl-1.20.1\n"
     ]
    }
   ],
   "source": [
    "!pip install torch torchvision numpy pandas matplotlib scikit-learn transformers datasets accelerate"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "aca99530-0cea-4026-a9bb-cdfc9998e34f",
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "import torch.optim as optim\n",
    "from torch.utils.data import DataLoader, Dataset\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "import matplotlib.pyplot as plt\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.metrics import accuracy_score, classification_report\n",
    "import math\n",
    "import time\n",
    "\n",
    "# Custom Dataset class to handle text data and convert to tokens\n",
    "class TextDataset(Dataset):\n",
    "    def __init__(self, texts, labels, vocab_to_idx, max_length=128):\n",
    "        \"\"\"\n",
    "        Initialize the dataset with texts, labels, vocabulary mapping, and max sequence length\n",
    "        \"\"\"\n",
    "        self.texts = texts  # List of text strings\n",
    "        self.labels = labels  # List of corresponding labels (0 or 1 for binary classification)\n",
    "        self.vocab_to_idx = vocab_to_idx  # Dictionary mapping words to indices\n",
    "        self.max_length = max_length  # Maximum sequence length for padding/truncation\n",
    "        \n",
    "    def __len__(self):\n",
    "        \"\"\"Return the total number of samples in the dataset\"\"\"\n",
    "        return len(self.texts)\n",
    "    \n",
    "    def __getitem__(self, idx):\n",
    "        \"\"\"\n",
    "        Get a single sample from the dataset\n",
    "        Returns tokenized and padded text along with its label\n",
    "        \"\"\"\n",
    "        text = self.texts[idx]\n",
    "        label = self.labels[idx]\n",
    "        \n",
    "        # Tokenize text by splitting on whitespace and converting to lowercase\n",
    "        tokens = text.lower().split()\n",
    "        \n",
    "        # Convert tokens to indices using vocabulary, use <UNK> for unknown words\n",
    "        token_ids = [self.vocab_to_idx.get(token, self.vocab_to_idx['<UNK>']) for token in tokens]\n",
    "        \n",
    "        # Truncate if sequence is longer than max_length\n",
    "        if len(token_ids) > self.max_length:\n",
    "            token_ids = token_ids[:self.max_length]\n",
    "        else:\n",
    "            # Pad with <PAD> tokens if sequence is shorter than max_length\n",
    "            token_ids.extend([self.vocab_to_idx['<PAD>']] * (self.max_length - len(token_ids)))\n",
    "        \n",
    "        # Convert to PyTorch tensors\n",
    "        return torch.tensor(token_ids, dtype=torch.long), torch.tensor(label, dtype=torch.long)\n",
    "\n",
    "# Positional Encoding module to add position information to embeddings\n",
    "class PositionalEncoding(nn.Module):\n",
    "    def __init__(self, d_model, max_length=5000):\n",
    "        \"\"\"\n",
    "        Initialize positional encoding with sinusoidal patterns\n",
    "        d_model: dimension of the model embeddings\n",
    "        max_length: maximum sequence length to support\n",
    "        \"\"\"\n",
    "        super().__init__()\n",
    "        \n",
    "        # Create a matrix to hold positional encodings\n",
    "        pe = torch.zeros(max_length, d_model)\n",
    "        \n",
    "        # Create position indices (0, 1, 2, ..., max_length-1)\n",
    "        position = torch.arange(0, max_length, dtype=torch.float).unsqueeze(1)\n",
    "        \n",
    "        # Create the divisor term for the sinusoidal pattern\n",
    "        # This creates different frequencies for different dimensions\n",
    "        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))\n",
    "        \n",
    "        # Apply sine to even indices (0, 2, 4, ...)\n",
    "        pe[:, 0::2] = torch.sin(position * div_term)\n",
    "        \n",
    "        # Apply cosine to odd indices (1, 3, 5, ...)\n",
    "        pe[:, 1::2] = torch.cos(position * div_term)\n",
    "        \n",
    "        # Reshape to (max_length, 1, d_model) for broadcasting\n",
    "        pe = pe.unsqueeze(0).transpose(0, 1)\n",
    "        \n",
    "        # Register as buffer so it's saved with the model but not trained\n",
    "        self.register_buffer('pe', pe)\n",
    "    \n",
    "    def forward(self, x):\n",
    "        \"\"\"\n",
    "        Add positional encoding to input embeddings\n",
    "        x: input embeddings of shape (seq_len, batch_size, d_model)\n",
    "        \"\"\"\n",
    "        # Add positional encoding to input, only for the sequence length of x\n",
    "        return x + self.pe[:x.size(0), :]\n",
    "\n",
    "# Multi-Head Attention mechanism - the core of the Transformer\n",
    "class MultiHeadAttention(nn.Module):\n",
    "    def __init__(self, d_model, num_heads):\n",
    "        \"\"\"\n",
    "        Initialize multi-head attention\n",
    "        d_model: dimension of the model\n",
    "        num_heads: number of attention heads\n",
    "        \"\"\"\n",
    "        super().__init__()\n",
    "        self.d_model = d_model\n",
    "        self.num_heads = num_heads\n",
    "        \n",
    "        # Each head processes d_k dimensions\n",
    "        self.d_k = d_model // num_heads\n",
    "        \n",
    "        # Linear transformations for Query, Key, Value, and Output\n",
    "        self.W_q = nn.Linear(d_model, d_model)  # Query projection\n",
    "        self.W_k = nn.Linear(d_model, d_model)  # Key projection\n",
    "        self.W_v = nn.Linear(d_model, d_model)  # Value projection\n",
    "        self.W_o = nn.Linear(d_model, d_model)  # Output projection\n",
    "        \n",
    "    def scaled_dot_product_attention(self, Q, K, V, mask=None):\n",
    "        \"\"\"\n",
    "        Compute scaled dot-product attention\n",
    "        Q, K, V: Query, Key, Value matrices\n",
    "        mask: optional mask to prevent attention to certain positions\n",
    "        \"\"\"\n",
    "        # Compute attention scores: Q * K^T / sqrt(d_k)\n",
    "        scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.d_k)\n",
    "        \n",
    "        # Apply mask if provided (set masked positions to large negative value)\n",
    "        if mask is not None:\n",
    "            scores = scores.masked_fill(mask == 0, -1e9)\n",
    "        \n",
    "        # Apply softmax to get attention weights\n",
    "        attention_weights = F.softmax(scores, dim=-1)\n",
    "        \n",
    "        # Apply attention weights to values\n",
    "        output = torch.matmul(attention_weights, V)\n",
    "        \n",
    "        return output, attention_weights\n",
    "    \n",
    "    def forward(self, query, key, value, mask=None):\n",
    "        \"\"\"\n",
    "        Forward pass of multi-head attention\n",
    "        query, key, value: input tensors (can be the same for self-attention)\n",
    "        mask: optional attention mask\n",
    "        \"\"\"\n",
    "        batch_size = query.size(0)\n",
    "        \n",
    "        # Apply linear transformations and reshape for multi-head processing\n",
    "        # Shape: (batch_size, seq_len, d_model) -> (batch_size, num_heads, seq_len, d_k)\n",
    "        Q = self.W_q(query).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)\n",
    "        K = self.W_k(key).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)\n",
    "        V = self.W_v(value).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)\n",
    "        \n",
    "        # Apply scaled dot-product attention\n",
    "        attention_output, attention_weights = self.scaled_dot_product_attention(Q, K, V, mask)\n",
    "        \n",
    "        # Concatenate heads and reshape back to original dimensions\n",
    "        # Shape: (batch_size, num_heads, seq_len, d_k) -> (batch_size, seq_len, d_model)\n",
    "        attention_output = attention_output.transpose(1, 2).contiguous().view(\n",
    "            batch_size, -1, self.d_model)\n",
    "        \n",
    "        # Apply final linear transformation\n",
    "        output = self.W_o(attention_output)\n",
    "        return output\n",
    "\n",
    "# Single Transformer block containing attention and feed-forward layers\n",
    "class TransformerBlock(nn.Module):\n",
    "    def __init__(self, d_model, num_heads, d_ff, dropout=0.1):\n",
    "        \"\"\"\n",
    "        Initialize a Transformer block\n",
    "        d_model: model dimension\n",
    "        num_heads: number of attention heads\n",
    "        d_ff: dimension of feed-forward network\n",
    "        dropout: dropout probability\n",
    "        \"\"\"\n",
    "        super().__init__()\n",
    "        \n",
    "        # Multi-head attention layer\n",
    "        self.attention = MultiHeadAttention(d_model, num_heads)\n",
    "        \n",
    "        # Layer normalization layers\n",
    "        self.norm1 = nn.LayerNorm(d_model)\n",
    "        self.norm2 = nn.LayerNorm(d_model)\n",
    "        \n",
    "        # Feed-forward network: two linear layers with ReLU activation\n",
    "        self.feed_forward = nn.Sequential(\n",
    "            nn.Linear(d_model, d_ff),\n",
    "            nn.ReLU(),\n",
    "            nn.Linear(d_ff, d_model)\n",
    "        )\n",
    "        \n",
    "        # Dropout layer for regularization\n",
    "        self.dropout = nn.Dropout(dropout)\n",
    "        \n",
    "    def forward(self, x, mask=None):\n",
    "        \"\"\"\n",
    "        Forward pass through the Transformer block\n",
    "        x: input tensor\n",
    "        mask: optional attention mask\n",
    "        \"\"\"\n",
    "        # Self-attention with residual connection and layer normalization\n",
    "        attention_output = self.attention(x, x, x, mask)\n",
    "        x = self.norm1(x + self.dropout(attention_output))\n",
    "        \n",
    "        # Feed-forward with residual connection and layer normalization\n",
    "        ff_output = self.feed_forward(x)\n",
    "        x = self.norm2(x + self.dropout(ff_output))\n",
    "        \n",
    "        return x\n",
    "\n",
    "# Complete Transformer model for text classification\n",
    "class TransformerClassifier(nn.Module):\n",
    "    def __init__(self, vocab_size, d_model, num_heads, num_layers, d_ff, max_length, num_classes, dropout=0.1):\n",
    "        \"\"\"\n",
    "        Initialize the Transformer classifier\n",
    "        vocab_size: size of vocabulary\n",
    "        d_model: model dimension\n",
    "        num_heads: number of attention heads\n",
    "        num_layers: number of Transformer blocks\n",
    "        d_ff: feed-forward dimension\n",
    "        max_length: maximum sequence length\n",
    "        num_classes: number of output classes\n",
    "        dropout: dropout probability\n",
    "        \"\"\"\n",
    "        super().__init__()\n",
    "        self.d_model = d_model\n",
    "        \n",
    "        # Embedding layer to convert token indices to dense vectors\n",
    "        self.embedding = nn.Embedding(vocab_size, d_model)\n",
    "        \n",
    "        # Positional encoding to add position information\n",
    "        self.positional_encoding = PositionalEncoding(d_model, max_length)\n",
    "        \n",
    "        # Stack of Transformer blocks\n",
    "        self.transformer_blocks = nn.ModuleList([\n",
    "            TransformerBlock(d_model, num_heads, d_ff, dropout)\n",
    "            for _ in range(num_layers)\n",
    "        ])\n",
    "        \n",
    "        # Final layer normalization\n",
    "        self.norm = nn.LayerNorm(d_model)\n",
    "        \n",
    "        # Classification head\n",
    "        self.classifier = nn.Linear(d_model, num_classes)\n",
    "        \n",
    "        # Dropout for regularization\n",
    "        self.dropout = nn.Dropout(dropout)\n",
    "        \n",
    "    def forward(self, x, mask=None):\n",
    "        \"\"\"\n",
    "        Forward pass through the entire model\n",
    "        x: input token indices\n",
    "        mask: optional attention mask\n",
    "        \"\"\"\n",
    "        # Convert token indices to embeddings and scale by sqrt(d_model)\n",
    "        x = self.embedding(x) * math.sqrt(self.d_model)\n",
    "        \n",
    "        # Add positional encoding\n",
    "        x = self.positional_encoding(x)\n",
    "        \n",
    "        # Apply dropout\n",
    "        x = self.dropout(x)\n",
    "        \n",
    "        # Pass through each Transformer block\n",
    "        for transformer in self.transformer_blocks:\n",
    "            x = transformer(x, mask)\n",
    "        \n",
    "        # Apply final layer normalization\n",
    "        x = self.norm(x)\n",
    "        \n",
    "        # Global average pooling to get sequence representation\n",
    "        x = torch.mean(x, dim=1)\n",
    "        \n",
    "        # Apply classification head\n",
    "        x = self.classifier(x)\n",
    "        \n",
    "        return x\n",
    "\n",
    "def create_sample_dataset():\n",
    "    \"\"\"\n",
    "    Create a sample dataset of movie reviews for sentiment analysis\n",
    "    Returns lists of texts and corresponding labels (1=positive, 0=negative)\n",
    "    \"\"\"\n",
    "    # Positive movie reviews\n",
    "    positive_reviews = [\n",
    "        \"this movie is absolutely fantastic and amazing\",\n",
    "        \"i loved every minute of this incredible film\",\n",
    "        \"outstanding performance by all actors brilliant\",\n",
    "        \"best movie i have seen in years\",\n",
    "        \"wonderful story and excellent cinematography\",\n",
    "        \"this film exceeded all my expectations\",\n",
    "        \"amazing plot and great character development\",\n",
    "        \"absolutely loved the soundtrack and visuals\",\n",
    "        \"this is a masterpiece of modern cinema\",\n",
    "        \"fantastic acting and brilliant direction\"\n",
    "    ]\n",
    "    \n",
    "    # Negative movie reviews\n",
    "    negative_reviews = [\n",
    "        \"this movie was terrible and boring\",\n",
    "        \"worst film i have ever watched\",\n",
    "        \"awful acting and poor storyline\",\n",
    "        \"complete waste of time and money\",\n",
    "        \"terrible plot and bad character development\",\n",
    "        \"boring and uninteresting throughout\",\n",
    "        \"poor directing and weak performances\",\n",
    "        \"disappointing and predictable story\",\n",
    "        \"bad cinematography and awful soundtrack\",\n",
    "        \"terrible movie with no redeeming qualities\"\n",
    "    ]\n",
    "    \n",
    "    # Combine texts and create labels\n",
    "    texts = positive_reviews + negative_reviews\n",
    "    labels = [1] * len(positive_reviews) + [0] * len(negative_reviews)\n",
    "    \n",
    "    return texts, labels\n",
    "\n",
    "def build_vocabulary(texts, min_freq=1):\n",
    "    \"\"\"\n",
    "    Build vocabulary from a list of texts\n",
    "    texts: list of text strings\n",
    "    min_freq: minimum frequency for a word to be included in vocabulary\n",
    "    Returns: dictionary mapping words to indices\n",
    "    \"\"\"\n",
    "    # Count word frequencies\n",
    "    word_counts = {}\n",
    "    for text in texts:\n",
    "        words = text.lower().split()\n",
    "        for word in words:\n",
    "            word_counts[word] = word_counts.get(word, 0) + 1\n",
    "    \n",
    "    # Initialize vocabulary with special tokens\n",
    "    vocab_to_idx = {'<PAD>': 0, '<UNK>': 1}\n",
    "    idx = 2\n",
    "    \n",
    "    # Add words that meet minimum frequency requirement\n",
    "    for word, count in word_counts.items():\n",
    "        if count >= min_freq:\n",
    "            vocab_to_idx[word] = idx\n",
    "            idx += 1\n",
    "    \n",
    "    return vocab_to_idx\n",
    "\n",
    "def create_padding_mask(x, pad_idx=0):\n",
    "    \"\"\"\n",
    "    Create a mask to prevent attention to padding tokens\n",
    "    x: input tensor with token indices\n",
    "    pad_idx: index of padding token\n",
    "    Returns: mask tensor\n",
    "    \"\"\"\n",
    "    return (x != pad_idx).unsqueeze(1).unsqueeze(2)\n",
    "\n",
    "def train_model(model, train_loader, val_loader, num_epochs, device):\n",
    "    \"\"\"\n",
    "    Train the Transformer model\n",
    "    model: the model to train\n",
    "    train_loader: DataLoader for training data\n",
    "    val_loader: DataLoader for validation data\n",
    "    num_epochs: number of training epochs\n",
    "    device: device to train on (CPU/GPU/MPS)\n",
    "    \"\"\"\n",
    "    # Define loss function and optimizer\n",
    "    criterion = nn.CrossEntropyLoss()\n",
    "    optimizer = optim.Adam(model.parameters(), lr=0.001)\n",
    "    \n",
    "    # Lists to store training progress\n",
    "    train_losses = []\n",
    "    val_accuracies = []\n",
    "    \n",
    "    # Training loop\n",
    "    for epoch in range(num_epochs):\n",
    "        # Training phase\n",
    "        model.train()\n",
    "        total_loss = 0\n",
    "        num_batches = 0\n",
    "        \n",
    "        # Iterate through training batches\n",
    "        for batch_idx, (data, target) in enumerate(train_loader):\n",
    "            # Move data to device\n",
    "            data, target = data.to(device), target.to(device)\n",
    "            \n",
    "            # Create padding mask\n",
    "            mask = create_padding_mask(data).to(device)\n",
    "            \n",
    "            # Zero gradients\n",
    "            optimizer.zero_grad()\n",
    "            \n",
    "            # Forward pass\n",
    "            output = model(data, mask)\n",
    "            \n",
    "            # Calculate loss\n",
    "            loss = criterion(output, target)\n",
    "            \n",
    "            # Backward pass\n",
    "            loss.backward()\n",
    "            \n",
    "            # Update weights\n",
    "            optimizer.step()\n",
    "            \n",
    "            # Accumulate loss\n",
    "            total_loss += loss.item()\n",
    "            num_batches += 1\n",
    "        \n",
    "        # Calculate average training loss\n",
    "        avg_loss = total_loss / num_batches\n",
    "        train_losses.append(avg_loss)\n",
    "        \n",
    "        # Validation phase\n",
    "        model.eval()\n",
    "        correct = 0\n",
    "        total = 0\n",
    "        \n",
    "        # Evaluate on validation set\n",
    "        with torch.no_grad():\n",
    "            for data, target in val_loader:\n",
    "                data, target = data.to(device), target.to(device)\n",
    "                mask = create_padding_mask(data).to(device)\n",
    "                \n",
    "                # Forward pass\n",
    "                output = model(data, mask)\n",
    "                \n",
    "                # Get predictions\n",
    "                _, predicted = torch.max(output.data, 1)\n",
    "                total += target.size(0)\n",
    "                correct += (predicted == target).sum().item()\n",
    "        \n",
    "        # Calculate validation accuracy\n",
    "        val_accuracy = 100 * correct / total\n",
    "        val_accuracies.append(val_accuracy)\n",
    "    \n",
    "    return train_losses, val_accuracies\n",
    "\n",
    "def evaluate_model(model, test_loader, device):\n",
    "    \"\"\"\n",
    "    Evaluate the trained model on test data\n",
    "    model: trained model\n",
    "    test_loader: DataLoader for test data\n",
    "    device: device to run evaluation on\n",
    "    \"\"\"\n",
    "    # Set model to evaluation mode\n",
    "    model.eval()\n",
    "    all_predictions = []\n",
    "    all_targets = []\n",
    "    \n",
    "    # Evaluate without gradient computation\n",
    "    with torch.no_grad():\n",
    "        for data, target in test_loader:\n",
    "            data, target = data.to(device), target.to(device)\n",
    "            mask = create_padding_mask(data).to(device)\n",
    "            \n",
    "            # Forward pass\n",
    "            output = model(data, mask)\n",
    "            \n",
    "            # Get predictions\n",
    "            _, predicted = torch.max(output, 1)\n",
    "            \n",
    "            # Store predictions and targets\n",
    "            all_predictions.extend(predicted.cpu().numpy())\n",
    "            all_targets.extend(target.cpu().numpy())\n",
    "    \n",
    "    # Calculate metrics\n",
    "    accuracy = accuracy_score(all_targets, all_predictions)\n",
    "    report = classification_report(all_targets, all_predictions, target_names=['Negative', 'Positive'])\n",
    "    \n",
    "    return accuracy, all_predictions, all_targets\n",
    "\n",
    "def main():\n",
    "    \"\"\"\n",
    "    Main function to orchestrate the entire training and evaluation process\n",
    "    \"\"\"\n",
    "    # Determine the best available device\n",
    "    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
    "    if device.type == 'cuda':\n",
    "        print(f\"Using GPU: {torch.cuda.get_device_name(0)}\")\n",
    "    elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():\n",
    "        device = torch.device('mps')\n",
    "        print(\"Using Apple Silicon MPS acceleration\")\n",
    "    else:\n",
    "        print(\"Using CPU\")\n",
    "    \n",
    "    # Create sample dataset\n",
    "    texts, labels = create_sample_dataset()\n",
    "    \n",
    "    # Build vocabulary from the texts\n",
    "    vocab_to_idx = build_vocabulary(texts)\n",
    "    \n",
    "    # Split data into train, validation, and test sets\n",
    "    X_train, X_temp, y_train, y_temp = train_test_split(texts, labels, test_size=0.4, random_state=42)\n",
    "    X_val, X_test, y_val, y_test = train_test_split(X_temp, y_temp, test_size=0.5, random_state=42)\n",
    "    \n",
    "    # Create datasets and data loaders\n",
    "    max_length = 32\n",
    "    train_dataset = TextDataset(X_train, y_train, vocab_to_idx, max_length)\n",
    "    val_dataset = TextDataset(X_val, y_val, vocab_to_idx, max_length)\n",
    "    test_dataset = TextDataset(X_test, y_test, vocab_to_idx, max_length)\n",
    "    \n",
    "    train_loader = DataLoader(train_dataset, batch_size=4, shuffle=True)\n",
    "    val_loader = DataLoader(val_dataset, batch_size=4, shuffle=False)\n",
    "    test_loader = DataLoader(test_dataset, batch_size=4, shuffle=False)\n",
    "    \n",
    "    # Define model configuration\n",
    "    model_config = {\n",
    "        'vocab_size': len(vocab_to_idx),\n",
    "        'd_model': 64,\n",
    "        'num_heads': 4,\n",
    "        'num_layers': 2,\n",
    "        'd_ff': 128,\n",
    "        'max_length': max_length,\n",
    "        'num_classes': 2,\n",
    "        'dropout': 0.1\n",
    "    }\n",
    "    \n",
    "    # Create and move model to device\n",
    "    model = TransformerClassifier(**model_config).to(device)\n",
    "    \n",
    "    # Count model parameters\n",
    "    total_params = sum(p.numel() for p in model.parameters())\n",
    "    trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)\n",
    "    \n",
    "    # Train the model\n",
    "    train_losses, val_accuracies = train_model(model, train_loader, val_loader, num_epochs=10, device=device)\n",
    "    \n",
    "    # Evaluate on test set\n",
    "    test_accuracy, predictions, targets = evaluate_model(model, test_loader, device)\n",
    "    \n",
    "    # Create visualization of training progress\n",
    "    plt.figure(figsize=(12, 4))\n",
    "    \n",
    "    # Plot training loss\n",
    "    plt.subplot(1, 2, 1)\n",
    "    plt.plot(train_losses)\n",
    "    plt.title('Training Loss')\n",
    "    plt.xlabel('Epoch')\n",
    "    plt.ylabel('Loss')\n",
    "    \n",
    "    # Plot validation accuracy\n",
    "    plt.subplot(1, 2, 2)\n",
    "    plt.plot(val_accuracies)\n",
    "    plt.title('Validation Accuracy')\n",
    "    plt.xlabel('Epoch')\n",
    "    plt.ylabel('Accuracy (%)')\n",
    "    \n",
    "    plt.tight_layout()\n",
    "    plt.savefig('transformer_training_progress.png')\n",
    "    plt.show()\n",
    "    \n",
    "    # Function to make predictions on new text\n",
    "    def predict_sentiment(text, model, vocab_to_idx, device, max_length=32):\n",
    "        \"\"\"\n",
    "        Predict sentiment for a given text\n",
    "        text: input text string\n",
    "        model: trained model\n",
    "        vocab_to_idx: vocabulary mapping\n",
    "        device: device to run prediction on\n",
    "        max_length: maximum sequence length\n",
    "        \"\"\"\n",
    "        model.eval()\n",
    "        \n",
    "        # Tokenize and convert to indices\n",
    "        tokens = text.lower().split()\n",
    "        token_ids = [vocab_to_idx.get(token, vocab_to_idx['<UNK>']) for token in tokens]\n",
    "        \n",
    "        # Pad or truncate to max_length\n",
    "        if len(token_ids) > max_length:\n",
    "            token_ids = token_ids[:max_length]\n",
    "        else:\n",
    "            token_ids.extend([vocab_to_idx['<PAD>']] * (max_length - len(token_ids)))\n",
    "        \n",
    "        # Convert to tensor and add batch dimension\n",
    "        input_tensor = torch.tensor([token_ids], dtype=torch.long).to(device)\n",
    "        mask = create_padding_mask(input_tensor).to(device)\n",
    "        \n",
    "        # Make prediction\n",
    "        with torch.no_grad():\n",
    "            output = model(input_tensor, mask)\n",
    "            probabilities = F.softmax(output, dim=1)\n",
    "            _, predicted = torch.max(output, 1)\n",
    "        \n",
    "        # Convert prediction to sentiment label\n",
    "        sentiment = \"Positive\" if predicted.item() == 1 else \"Negative\"\n",
    "        confidence = probabilities[0][predicted.item()].item()\n",
    "        \n",
    "        return sentiment, confidence\n",
    "    \n",
    "    # Test the model with sample sentences\n",
    "    test_sentences = [\n",
    "        \"this movie is absolutely amazing\",\n",
    "        \"terrible film with bad acting\",\n",
    "        \"good story but poor execution\"\n",
    "    ]\n",
    "    \n",
    "    print(\"\\nSample predictions:\")\n",
    "    for sentence in test_sentences:\n",
    "        sentiment, confidence = predict_sentiment(sentence, model, vocab_to_idx, device)\n",
    "        print(f\"Text: '{sentence}' -> Sentiment: {sentiment} (Confidence: {confidence:.3f})\")\n",
    "\n",
    "# Run the main function if this script is executed directly\n",
    "if __name__ == \"__main__\":\n",
    "    main()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "be43abe2-06cf-4d29-8590-f72d3cbecc7b",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2025-07-14 13:11:54,911 - INFO - Starting M4-optimized Transformer implementation\n",
      "2025-07-14 13:11:54,912 - INFO - Using Apple Silicon M4 MPS acceleration\n",
      "2025-07-14 13:11:54,912 - INFO - Creating extensive movie review dataset for M4 chip\n",
      "2025-07-14 13:11:54,912 - INFO - Created dataset with 200 samples\n",
      "2025-07-14 13:11:54,912 - INFO - Positive: 100, Negative: 100\n",
      "2025-07-14 13:11:54,913 - INFO - Building vocabulary from text data\n",
      "2025-07-14 13:11:54,913 - INFO - Vocabulary size: 106\n",
      "2025-07-14 13:11:54,915 - INFO - Dataset splits - Train: 140, Val: 30, Test: 30\n",
      "2025-07-14 13:11:54,934 - INFO - Total parameters: 544,002\n",
      "2025-07-14 13:11:54,935 - INFO - Starting model training\n",
      "2025-07-14 13:11:54,935 - INFO - Epoch 1/30\n",
      "2025-07-14 13:11:54,937 - INFO - Input data shape: torch.Size([16, 24])\n",
      "2025-07-14 13:11:54,937 - INFO - Target shape: torch.Size([16])\n",
      "2025-07-14 13:11:55,233 - INFO - Model output shape: torch.Size([16, 2])\n",
      "2025-07-14 13:11:55,535 - INFO - Batch 0/9, Loss: 0.8154\n",
      "2025-07-14 13:11:56,072 - INFO - Epoch 1 - Loss: 0.6878, Val Acc: 80.00%\n",
      "2025-07-14 13:11:56,073 - INFO - Epoch 2/30\n",
      "2025-07-14 13:11:56,095 - INFO - Batch 0/9, Loss: 0.6519\n",
      "2025-07-14 13:11:56,305 - INFO - Epoch 2 - Loss: 0.5778, Val Acc: 100.00%\n",
      "2025-07-14 13:11:56,305 - INFO - Epoch 3/30\n",
      "2025-07-14 13:11:56,329 - INFO - Batch 0/9, Loss: 0.5265\n",
      "2025-07-14 13:11:56,517 - INFO - Epoch 3 - Loss: 0.4763, Val Acc: 96.67%\n",
      "2025-07-14 13:11:56,517 - INFO - Epoch 4/30\n",
      "2025-07-14 13:11:56,541 - INFO - Batch 0/9, Loss: 0.3375\n",
      "2025-07-14 13:11:56,731 - INFO - Epoch 4 - Loss: 0.3694, Val Acc: 96.67%\n",
      "2025-07-14 13:11:56,732 - INFO - Epoch 5/30\n",
      "2025-07-14 13:11:56,755 - INFO - Batch 0/9, Loss: 0.4018\n",
      "2025-07-14 13:11:56,938 - INFO - Epoch 5 - Loss: 0.2218, Val Acc: 100.00%\n",
      "2025-07-14 13:11:56,938 - INFO - Epoch 6/30\n",
      "2025-07-14 13:11:56,960 - INFO - Batch 0/9, Loss: 0.1139\n",
      "2025-07-14 13:11:57,145 - INFO - Epoch 6 - Loss: 0.0939, Val Acc: 96.67%\n",
      "2025-07-14 13:11:57,145 - INFO - Epoch 7/30\n",
      "2025-07-14 13:11:57,172 - INFO - Batch 0/9, Loss: 0.0501\n",
      "2025-07-14 13:11:57,344 - INFO - Epoch 7 - Loss: 0.0513, Val Acc: 96.67%\n",
      "2025-07-14 13:11:57,345 - INFO - Early stopping at epoch 7\n",
      "2025-07-14 13:11:57,362 - INFO - Best validation accuracy: 100.00%\n",
      "2025-07-14 13:11:57,362 - INFO - Evaluating model on test set\n",
      "2025-07-14 13:11:57,371 - INFO - Test Accuracy: 0.8667\n",
      "2025-07-14 13:11:57,373 - INFO - Classification Report:\n",
      "              precision    recall  f1-score   support\n",
      "\n",
      "    Negative       0.87      0.87      0.87        15\n",
      "    Positive       0.87      0.87      0.87        15\n",
      "\n",
      "    accuracy                           0.87        30\n",
      "   macro avg       0.87      0.87      0.87        30\n",
      "weighted avg       0.87      0.87      0.87        30\n",
      "\n"
     ]
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 1500x500 with 3 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "2025-07-14 13:11:57,628 - INFO - Testing model predictions on M4 chip\n",
      "2025-07-14 13:11:57,742 - INFO - 'this movie is absolutely amazing and fantastic wit...' -> Positive (0.629)\n",
      "2025-07-14 13:11:57,746 - INFO - 'terrible film with awful acting and poor boring st...' -> Negative (0.691)\n",
      "2025-07-14 13:11:57,750 - INFO - 'good story but disappointing execution and weak ch...' -> Positive (0.542)\n",
      "2025-07-14 13:11:57,760 - INFO - 'excellent film with brilliant performances and won...' -> Positive (0.626)\n",
      "2025-07-14 13:11:57,764 - INFO - 'boring movie with terrible plot and awful disappoi...' -> Negative (0.713)\n",
      "2025-07-14 13:11:57,767 - INFO - 'fantastic entertainment with amazing visuals and g...' -> Positive (0.706)\n",
      "2025-07-14 13:11:57,771 - INFO - 'poor acting with boring dialogue and terrible char...' -> Negative (0.649)\n",
      "2025-07-14 13:11:57,771 - INFO - Final test accuracy: 0.8667\n",
      "2025-07-14 13:11:57,772 - INFO - M4 Transformer implementation completed successfully!\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "import torch.optim as optim\n",
    "from torch.utils.data import DataLoader, Dataset\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.metrics import accuracy_score, classification_report\n",
    "import logging\n",
    "import math\n",
    "import random\n",
    "\n",
    "# Configure logging\n",
    "logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')\n",
    "logger = logging.getLogger(__name__)\n",
    "\n",
    "# Set seeds for reproducibility\n",
    "torch.manual_seed(42)\n",
    "np.random.seed(42)\n",
    "random.seed(42)\n",
    "\n",
    "class TextDataset(Dataset):\n",
    "    def __init__(self, texts, labels, vocab_to_idx, max_length=32):\n",
    "        self.texts = texts\n",
    "        self.labels = labels\n",
    "        self.vocab_to_idx = vocab_to_idx\n",
    "        self.max_length = max_length\n",
    "        \n",
    "    def __len__(self):\n",
    "        return len(self.texts)\n",
    "    \n",
    "    def __getitem__(self, idx):\n",
    "        text = self.texts[idx]\n",
    "        label = self.labels[idx]\n",
    "        \n",
    "        tokens = text.lower().split()\n",
    "        token_ids = [self.vocab_to_idx.get(token, self.vocab_to_idx['<UNK>']) for token in tokens]\n",
    "        \n",
    "        if len(token_ids) > self.max_length:\n",
    "            token_ids = token_ids[:self.max_length]\n",
    "        else:\n",
    "            token_ids.extend([self.vocab_to_idx['<PAD>']] * (self.max_length - len(token_ids)))\n",
    "        \n",
    "        return torch.tensor(token_ids, dtype=torch.long), torch.tensor(label, dtype=torch.long)\n",
    "\n",
    "class PositionalEncoding(nn.Module):\n",
    "    def __init__(self, d_model, max_len=512):\n",
    "        super().__init__()\n",
    "        self.dropout = nn.Dropout(0.1)\n",
    "        \n",
    "        pe = torch.zeros(max_len, d_model)\n",
    "        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)\n",
    "        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))\n",
    "        \n",
    "        pe[:, 0::2] = torch.sin(position * div_term)\n",
    "        pe[:, 1::2] = torch.cos(position * div_term)\n",
    "        \n",
    "        self.register_buffer('pe', pe)\n",
    "        \n",
    "    def forward(self, x):\n",
    "        # x shape: (batch_size, seq_len, d_model)\n",
    "        seq_len = x.size(1)\n",
    "        x = x + self.pe[:seq_len, :].unsqueeze(0)\n",
    "        return self.dropout(x)\n",
    "\n",
    "class MultiHeadAttention(nn.Module):\n",
    "    def __init__(self, d_model, num_heads, dropout=0.1):\n",
    "        super().__init__()\n",
    "        assert d_model % num_heads == 0\n",
    "        \n",
    "        self.d_model = d_model\n",
    "        self.num_heads = num_heads\n",
    "        self.d_k = d_model // num_heads\n",
    "        \n",
    "        self.W_q = nn.Linear(d_model, d_model)\n",
    "        self.W_k = nn.Linear(d_model, d_model)\n",
    "        self.W_v = nn.Linear(d_model, d_model)\n",
    "        self.W_o = nn.Linear(d_model, d_model)\n",
    "        \n",
    "        self.dropout = nn.Dropout(dropout)\n",
    "        \n",
    "    def scaled_dot_product_attention(self, Q, K, V, mask=None):\n",
    "        d_k = Q.size(-1)\n",
    "        scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(d_k)\n",
    "        \n",
    "        if mask is not None:\n",
    "            scores = scores.masked_fill(mask == 0, -1e9)\n",
    "        \n",
    "        attention_weights = F.softmax(scores, dim=-1)\n",
    "        attention_weights = self.dropout(attention_weights)\n",
    "        \n",
    "        return torch.matmul(attention_weights, V), attention_weights\n",
    "    \n",
    "    def forward(self, query, key, value, mask=None):\n",
    "        batch_size = query.size(0)\n",
    "        \n",
    "        # Linear transformations\n",
    "        Q = self.W_q(query).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)\n",
    "        K = self.W_k(key).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)\n",
    "        V = self.W_v(value).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)\n",
    "        \n",
    "        # Attention\n",
    "        attention_output, attention_weights = self.scaled_dot_product_attention(Q, K, V, mask)\n",
    "        \n",
    "        # Concatenate heads\n",
    "        attention_output = attention_output.transpose(1, 2).contiguous().view(\n",
    "            batch_size, -1, self.d_model)\n",
    "        \n",
    "        return self.W_o(attention_output)\n",
    "\n",
    "class FeedForward(nn.Module):\n",
    "    def __init__(self, d_model, d_ff, dropout=0.1):\n",
    "        super().__init__()\n",
    "        self.linear1 = nn.Linear(d_model, d_ff)\n",
    "        self.linear2 = nn.Linear(d_ff, d_model)\n",
    "        self.dropout = nn.Dropout(dropout)\n",
    "        \n",
    "    def forward(self, x):\n",
    "        return self.linear2(self.dropout(F.relu(self.linear1(x))))\n",
    "\n",
    "class TransformerBlock(nn.Module):\n",
    "    def __init__(self, d_model, num_heads, d_ff, dropout=0.1):\n",
    "        super().__init__()\n",
    "        self.attention = MultiHeadAttention(d_model, num_heads, dropout)\n",
    "        self.feed_forward = FeedForward(d_model, d_ff, dropout)\n",
    "        self.norm1 = nn.LayerNorm(d_model)\n",
    "        self.norm2 = nn.LayerNorm(d_model)\n",
    "        self.dropout = nn.Dropout(dropout)\n",
    "        \n",
    "    def forward(self, x, mask=None):\n",
    "        # Self-attention with residual connection\n",
    "        attn_output = self.attention(x, x, x, mask)\n",
    "        x = self.norm1(x + self.dropout(attn_output))\n",
    "        \n",
    "        # Feed-forward with residual connection\n",
    "        ff_output = self.feed_forward(x)\n",
    "        x = self.norm2(x + self.dropout(ff_output))\n",
    "        \n",
    "        return x\n",
    "\n",
    "class TransformerClassifier(nn.Module):\n",
    "    def __init__(self, vocab_size, d_model, num_heads, num_layers, d_ff, max_len, num_classes, dropout=0.1):\n",
    "        super().__init__()\n",
    "        self.d_model = d_model\n",
    "        self.embedding = nn.Embedding(vocab_size, d_model)\n",
    "        self.pos_encoding = PositionalEncoding(d_model, max_len)\n",
    "        \n",
    "        self.transformer_blocks = nn.ModuleList([\n",
    "            TransformerBlock(d_model, num_heads, d_ff, dropout)\n",
    "            for _ in range(num_layers)\n",
    "        ])\n",
    "        \n",
    "        self.norm = nn.LayerNorm(d_model)\n",
    "        self.classifier = nn.Linear(d_model, num_classes)\n",
    "        self.dropout = nn.Dropout(dropout)\n",
    "        \n",
    "    def forward(self, x, mask=None):\n",
    "        # Input shape: (batch_size, seq_len)\n",
    "        batch_size, seq_len = x.size()\n",
    "        \n",
    "        # Embedding: (batch_size, seq_len, d_model)\n",
    "        x = self.embedding(x) * math.sqrt(self.d_model)\n",
    "        \n",
    "        # Positional encoding\n",
    "        x = self.pos_encoding(x)\n",
    "        \n",
    "        # Transformer blocks\n",
    "        for transformer in self.transformer_blocks:\n",
    "            x = transformer(x, mask)\n",
    "        \n",
    "        # Final normalization\n",
    "        x = self.norm(x)\n",
    "        \n",
    "        # Global average pooling: (batch_size, d_model)\n",
    "        x = x.mean(dim=1)\n",
    "        \n",
    "        # Classification: (batch_size, num_classes)\n",
    "        x = self.classifier(x)\n",
    "        \n",
    "        return x\n",
    "\n",
    "def create_extensive_dataset():\n",
    "    logger.info(\"Creating extensive movie review dataset for M4 chip\")\n",
    "    \n",
    "    # 100 positive reviews\n",
    "    positive_reviews = [\n",
    "        \"this movie is absolutely fantastic and amazing\",\n",
    "        \"i loved every minute of this incredible film\",\n",
    "        \"outstanding performance by all actors brilliant\",\n",
    "        \"best movie i have seen in years wonderful\",\n",
    "        \"excellent story and great cinematography\",\n",
    "        \"this film exceeded all my expectations perfectly\",\n",
    "        \"amazing plot and fantastic character development\",\n",
    "        \"absolutely loved the soundtrack and visuals\",\n",
    "        \"this is a masterpiece of modern cinema\",\n",
    "        \"fantastic acting and brilliant direction\",\n",
    "        \"incredible storyline with amazing twists\",\n",
    "        \"wonderful performances throughout the film\",\n",
    "        \"excellent directing and great production values\",\n",
    "        \"loved the characters and their development\",\n",
    "        \"fantastic visual effects and sound design\",\n",
    "        \"amazing chemistry between the actors\",\n",
    "        \"brilliant writing and excellent execution\",\n",
    "        \"outstanding cinematography and direction\",\n",
    "        \"wonderful movie with great emotional depth\",\n",
    "        \"excellent film with fantastic performances\",\n",
    "        \"great story with amazing character arcs\",\n",
    "        \"loved the plot and character interactions\",\n",
    "        \"fantastic movie with excellent pacing\",\n",
    "        \"wonderful acting and brilliant storytelling\",\n",
    "        \"amazing film with great visual appeal\",\n",
    "        \"excellent direction and outstanding cast\",\n",
    "        \"loved the creativity and original story\",\n",
    "        \"fantastic entertainment with great humor\",\n",
    "        \"wonderful experience with amazing visuals\",\n",
    "        \"excellent movie with brilliant performances\",\n",
    "        \"superb acting and incredible storytelling\",\n",
    "        \"amazing cinematography and great direction\",\n",
    "        \"fantastic film with wonderful characters\",\n",
    "        \"excellent plot with brilliant execution\",\n",
    "        \"loved the soundtrack and visual effects\",\n",
    "        \"wonderful movie with great performances\",\n",
    "        \"fantastic story with amazing development\",\n",
    "        \"excellent film with brilliant acting\",\n",
    "        \"amazing direction and wonderful cast\",\n",
    "        \"great movie with fantastic visuals\",\n",
    "        \"excellent entertainment with brilliant story\",\n",
    "        \"wonderful film with amazing performances\",\n",
    "        \"fantastic movie with great character work\",\n",
    "        \"amazing story with excellent direction\",\n",
    "        \"brilliant film with wonderful acting\",\n",
    "        \"excellent movie with fantastic plot\",\n",
    "        \"great film with amazing performances\",\n",
    "        \"wonderful story with brilliant execution\",\n",
    "        \"fantastic acting and excellent direction\",\n",
    "        \"amazing movie with great entertainment value\"\n",
    "    ] * 2  # Duplicate to get 100\n",
    "    \n",
    "    # 100 negative reviews\n",
    "    negative_reviews = [\n",
    "        \"this movie was terrible and boring\",\n",
    "        \"worst film i have ever watched\",\n",
    "        \"awful acting and poor storyline\",\n",
    "        \"complete waste of time and money\",\n",
    "        \"terrible plot and bad character development\",\n",
    "        \"boring and uninteresting throughout\",\n",
    "        \"poor directing and weak performances\",\n",
    "        \"disappointing and predictable story\",\n",
    "        \"bad cinematography and awful soundtrack\",\n",
    "        \"terrible movie with no redeeming qualities\",\n",
    "        \"awful script with poor execution\",\n",
    "        \"boring storyline with terrible acting\",\n",
    "        \"disappointing film with weak characters\",\n",
    "        \"poor production and bad direction\",\n",
    "        \"terrible dialogue and awful performances\",\n",
    "        \"boring plot with predictable ending\",\n",
    "        \"awful movie with poor character development\",\n",
    "        \"disappointing story with weak execution\",\n",
    "        \"terrible film with boring characters\",\n",
    "        \"poor acting and awful screenplay\",\n",
    "        \"boring movie with terrible pacing\",\n",
    "        \"awful direction and poor cinematography\",\n",
    "        \"disappointing film with weak storyline\",\n",
    "        \"terrible performances and boring plot\",\n",
    "        \"poor script with awful character development\",\n",
    "        \"boring film with disappointing ending\",\n",
    "        \"awful movie with terrible acting\",\n",
    "        \"disappointing story with poor execution\",\n",
    "        \"terrible film with boring dialogue\",\n",
    "        \"poor movie with awful performances\",\n",
    "        \"boring story with terrible direction\",\n",
    "        \"awful film with poor character work\",\n",
    "        \"disappointing movie with weak plot\",\n",
    "        \"terrible acting and boring story\",\n",
    "        \"poor film with awful direction\",\n",
    "        \"boring movie with terrible performances\",\n",
    "        \"awful story with disappointing execution\",\n",
    "        \"terrible film with poor acting\",\n",
    "        \"boring plot with awful characters\",\n",
    "        \"disappointing movie with terrible story\",\n",
    "        \"poor acting and boring direction\",\n",
    "        \"awful film with terrible plot\",\n",
    "        \"boring movie with disappointing performances\",\n",
    "        \"terrible story with poor execution\",\n",
    "        \"awful acting and boring film\",\n",
    "        \"disappointing plot with terrible direction\",\n",
    "        \"poor movie with awful story\",\n",
    "        \"boring film with terrible acting\",\n",
    "        \"awful direction and disappointing story\",\n",
    "        \"terrible movie with boring performances\"\n",
    "    ] * 2  # Duplicate to get 100\n",
    "    \n",
    "    texts = positive_reviews + negative_reviews\n",
    "    labels = [1] * len(positive_reviews) + [0] * len(negative_reviews)\n",
    "    \n",
    "    # Shuffle\n",
    "    combined = list(zip(texts, labels))\n",
    "    random.shuffle(combined)\n",
    "    texts, labels = zip(*combined)\n",
    "    \n",
    "    logger.info(f\"Created dataset with {len(texts)} samples\")\n",
    "    logger.info(f\"Positive: {len(positive_reviews)}, Negative: {len(negative_reviews)}\")\n",
    "    \n",
    "    return list(texts), list(labels)\n",
    "\n",
    "def build_vocabulary(texts, min_freq=2):\n",
    "    logger.info(\"Building vocabulary from text data\")\n",
    "    \n",
    "    word_counts = {}\n",
    "    for text in texts:\n",
    "        words = text.lower().split()\n",
    "        for word in words:\n",
    "            word_counts[word] = word_counts.get(word, 0) + 1\n",
    "    \n",
    "    vocab_to_idx = {'<PAD>': 0, '<UNK>': 1}\n",
    "    idx = 2\n",
    "    \n",
    "    for word, count in word_counts.items():\n",
    "        if count >= min_freq:\n",
    "            vocab_to_idx[word] = idx\n",
    "            idx += 1\n",
    "    \n",
    "    logger.info(f\"Vocabulary size: {len(vocab_to_idx)}\")\n",
    "    return vocab_to_idx\n",
    "\n",
    "def create_padding_mask(x, pad_idx=0):\n",
    "    # Create mask for padding tokens\n",
    "    mask = (x != pad_idx).unsqueeze(1).unsqueeze(2)\n",
    "    return mask\n",
    "\n",
    "def train_model(model, train_loader, val_loader, num_epochs, device):\n",
    "    logger.info(\"Starting model training\")\n",
    "    \n",
    "    criterion = nn.CrossEntropyLoss()\n",
    "    optimizer = optim.AdamW(model.parameters(), lr=0.0001, weight_decay=0.01)\n",
    "    scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=num_epochs)\n",
    "    \n",
    "    train_losses = []\n",
    "    val_accuracies = []\n",
    "    best_val_acc = 0\n",
    "    patience = 5\n",
    "    patience_counter = 0\n",
    "    \n",
    "    for epoch in range(num_epochs):\n",
    "        logger.info(f\"Epoch {epoch+1}/{num_epochs}\")\n",
    "        \n",
    "        # Training\n",
    "        model.train()\n",
    "        total_loss = 0\n",
    "        num_batches = 0\n",
    "        \n",
    "        for batch_idx, (data, target) in enumerate(train_loader):\n",
    "            data, target = data.to(device), target.to(device)\n",
    "            \n",
    "            # Debug shapes on first batch\n",
    "            if batch_idx == 0 and epoch == 0:\n",
    "                logger.info(f\"Input data shape: {data.shape}\")\n",
    "                logger.info(f\"Target shape: {target.shape}\")\n",
    "            \n",
    "            optimizer.zero_grad()\n",
    "            output = model(data)\n",
    "            \n",
    "            # Debug output shape\n",
    "            if batch_idx == 0 and epoch == 0:\n",
    "                logger.info(f\"Model output shape: {output.shape}\")\n",
    "            \n",
    "            loss = criterion(output, target)\n",
    "            loss.backward()\n",
    "            \n",
    "            # Gradient clipping\n",
    "            torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)\n",
    "            \n",
    "            optimizer.step()\n",
    "            \n",
    "            total_loss += loss.item()\n",
    "            num_batches += 1\n",
    "            \n",
    "            if batch_idx % 20 == 0:\n",
    "                logger.info(f\"Batch {batch_idx}/{len(train_loader)}, Loss: {loss.item():.4f}\")\n",
    "        \n",
    "        avg_loss = total_loss / num_batches\n",
    "        train_losses.append(avg_loss)\n",
    "        \n",
    "        # Validation\n",
    "        model.eval()\n",
    "        correct = 0\n",
    "        total = 0\n",
    "        \n",
    "        with torch.no_grad():\n",
    "            for data, target in val_loader:\n",
    "                data, target = data.to(device), target.to(device)\n",
    "                output = model(data)\n",
    "                _, predicted = torch.max(output, 1)\n",
    "                total += target.size(0)\n",
    "                correct += (predicted == target).sum().item()\n",
    "        \n",
    "        val_accuracy = 100 * correct / total\n",
    "        val_accuracies.append(val_accuracy)\n",
    "        \n",
    "        # Early stopping and model saving\n",
    "        if val_accuracy > best_val_acc:\n",
    "            best_val_acc = val_accuracy\n",
    "            torch.save(model.state_dict(), 'best_transformer_model.pth')\n",
    "            patience_counter = 0\n",
    "        else:\n",
    "            patience_counter += 1\n",
    "        \n",
    "        logger.info(f\"Epoch {epoch+1} - Loss: {avg_loss:.4f}, Val Acc: {val_accuracy:.2f}%\")\n",
    "        \n",
    "        # Early stopping\n",
    "        if patience_counter >= patience:\n",
    "            logger.info(f\"Early stopping at epoch {epoch+1}\")\n",
    "            break\n",
    "            \n",
    "        scheduler.step()\n",
    "    \n",
    "    # Load best model\n",
    "    model.load_state_dict(torch.load('best_transformer_model.pth'))\n",
    "    logger.info(f\"Best validation accuracy: {best_val_acc:.2f}%\")\n",
    "    \n",
    "    return train_losses, val_accuracies\n",
    "\n",
    "def evaluate_model(model, test_loader, device):\n",
    "    logger.info(\"Evaluating model on test set\")\n",
    "    \n",
    "    model.eval()\n",
    "    all_predictions = []\n",
    "    all_targets = []\n",
    "    \n",
    "    with torch.no_grad():\n",
    "        for data, target in test_loader:\n",
    "            data, target = data.to(device), target.to(device)\n",
    "            output = model(data)\n",
    "            _, predicted = torch.max(output, 1)\n",
    "            \n",
    "            all_predictions.extend(predicted.cpu().numpy())\n",
    "            all_targets.extend(target.cpu().numpy())\n",
    "    \n",
    "    accuracy = accuracy_score(all_targets, all_predictions)\n",
    "    logger.info(f\"Test Accuracy: {accuracy:.4f}\")\n",
    "    \n",
    "    report = classification_report(all_targets, all_predictions, target_names=['Negative', 'Positive'])\n",
    "    logger.info(f\"Classification Report:\\n{report}\")\n",
    "    \n",
    "    return accuracy, all_predictions, all_targets\n",
    "\n",
    "def main():\n",
    "    logger.info(\"Starting M4-optimized Transformer implementation\")\n",
    "    \n",
    "    # M4 chip setup\n",
    "    if torch.backends.mps.is_available():\n",
    "        device = torch.device('mps')\n",
    "        logger.info(\"Using Apple Silicon M4 MPS acceleration\")\n",
    "    else:\n",
    "        device = torch.device('cpu')\n",
    "        logger.info(\"Using CPU\")\n",
    "    \n",
    "    # Create extensive dataset\n",
    "    texts, labels = create_extensive_dataset()\n",
    "    vocab_to_idx = build_vocabulary(texts)\n",
    "    \n",
    "    # Proper data splitting\n",
    "    X_train, X_temp, y_train, y_temp = train_test_split(\n",
    "        texts, labels, test_size=0.3, random_state=42, stratify=labels\n",
    "    )\n",
    "    X_val, X_test, y_val, y_test = train_test_split(\n",
    "        X_temp, y_temp, test_size=0.5, random_state=42, stratify=y_temp\n",
    "    )\n",
    "    \n",
    "    logger.info(f\"Dataset splits - Train: {len(X_train)}, Val: {len(X_val)}, Test: {len(X_test)}\")\n",
    "    \n",
    "    # Create datasets\n",
    "    max_length = 24\n",
    "    train_dataset = TextDataset(X_train, y_train, vocab_to_idx, max_length)\n",
    "    val_dataset = TextDataset(X_val, y_val, vocab_to_idx, max_length)\n",
    "    test_dataset = TextDataset(X_test, y_test, vocab_to_idx, max_length)\n",
    "    \n",
    "    # Data loaders optimized for M4\n",
    "    train_loader = DataLoader(train_dataset, batch_size=16, shuffle=True, num_workers=0)\n",
    "    val_loader = DataLoader(val_dataset, batch_size=16, shuffle=False, num_workers=0)\n",
    "    test_loader = DataLoader(test_dataset, batch_size=16, shuffle=False, num_workers=0)\n",
    "    \n",
    "    # Model configuration optimized for M4\n",
    "    model = TransformerClassifier(\n",
    "        vocab_size=len(vocab_to_idx),\n",
    "        d_model=128,\n",
    "        num_heads=8,\n",
    "        num_layers=4,\n",
    "        d_ff=256,\n",
    "        max_len=max_length,\n",
    "        num_classes=2,\n",
    "        dropout=0.1\n",
    "    ).to(device)\n",
    "    \n",
    "    total_params = sum(p.numel() for p in model.parameters())\n",
    "    logger.info(f\"Total parameters: {total_params:,}\")\n",
    "    \n",
    "    # Train model\n",
    "    train_losses, val_accuracies = train_model(model, train_loader, val_loader, num_epochs=30, device=device)\n",
    "    \n",
    "    # Evaluate\n",
    "    test_accuracy, predictions, targets = evaluate_model(model, test_loader, device)\n",
    "    \n",
    "    # Visualization\n",
    "    plt.figure(figsize=(15, 5))\n",
    "    \n",
    "    plt.subplot(1, 3, 1)\n",
    "    plt.plot(train_losses, label='Training Loss')\n",
    "    plt.title('Training Loss')\n",
    "    plt.xlabel('Epoch')\n",
    "    plt.ylabel('Loss')\n",
    "    plt.legend()\n",
    "    \n",
    "    plt.subplot(1, 3, 2)\n",
    "    plt.plot(val_accuracies, label='Validation Accuracy')\n",
    "    plt.title('Validation Accuracy')\n",
    "    plt.xlabel('Epoch')\n",
    "    plt.ylabel('Accuracy (%)')\n",
    "    plt.legend()\n",
    "    \n",
    "    plt.subplot(1, 3, 3)\n",
    "    plt.hist(predictions, bins=2, alpha=0.7, label='Predictions')\n",
    "    plt.hist(targets, bins=2, alpha=0.7, label='Targets')\n",
    "    plt.title('Prediction Distribution')\n",
    "    plt.xlabel('Class')\n",
    "    plt.ylabel('Count')\n",
    "    plt.legend()\n",
    "    \n",
    "    plt.tight_layout()\n",
    "    plt.savefig('m4_transformer_results.png', dpi=300)\n",
    "    plt.show()\n",
    "    \n",
    "    # Test predictions\n",
    "    def predict_sentiment(text, model, vocab_to_idx, device, max_length=24):\n",
    "        model.eval()\n",
    "        tokens = text.lower().split()\n",
    "        token_ids = [vocab_to_idx.get(token, vocab_to_idx['<UNK>']) for token in tokens]\n",
    "        \n",
    "        if len(token_ids) > max_length:\n",
    "            token_ids = token_ids[:max_length]\n",
    "        else:\n",
    "            token_ids.extend([vocab_to_idx['<PAD>']] * (max_length - len(token_ids)))\n",
    "        \n",
    "        input_tensor = torch.tensor([token_ids], dtype=torch.long).to(device)\n",
    "        \n",
    "        with torch.no_grad():\n",
    "            output = model(input_tensor)\n",
    "            probabilities = F.softmax(output, dim=1)\n",
    "            _, predicted = torch.max(output, 1)\n",
    "        \n",
    "        sentiment = \"Positive\" if predicted.item() == 1 else \"Negative\"\n",
    "        confidence = probabilities[0][predicted.item()].item()\n",
    "        \n",
    "        return sentiment, confidence\n",
    "    \n",
    "    logger.info(\"Testing model predictions on M4 chip\")\n",
    "    \n",
    "    test_sentences = [\n",
    "        \"this movie is absolutely amazing and fantastic with great acting\",\n",
    "        \"terrible film with awful acting and poor boring story\",\n",
    "        \"good story but disappointing execution and weak characters\",\n",
    "        \"excellent film with brilliant performances and wonderful direction\",\n",
    "        \"boring movie with terrible plot and awful disappointing ending\",\n",
    "        \"fantastic entertainment with amazing visuals and great soundtrack\",\n",
    "        \"poor acting with boring dialogue and terrible character development\"\n",
    "    ]\n",
    "    \n",
    "    for sentence in test_sentences:\n",
    "        sentiment, confidence = predict_sentiment(sentence, model, vocab_to_idx, device)\n",
    "        logger.info(f\"'{sentence[:50]}...' -> {sentiment} ({confidence:.3f})\")\n",
    "    \n",
    "    logger.info(f\"Final test accuracy: {test_accuracy:.4f}\")\n",
    "    logger.info(\"M4 Transformer implementation completed successfully!\")\n",
    "\n",
    "if __name__ == \"__main__\":\n",
    "    main()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "37c4844e-d5a5-4cc7-8dc7-9750c318bfcd",
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.13.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}