notebooks/NN_transformer.ipynb

{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Transformer\n",
    "\n",
    "参考:\n",
    "\n",
    "* [Formal Algorithms for Transformers](https://arxiv.org/abs/2207.09238): シンプルなpseudocodeをまとめた論文。\n",
    "* [Attention Primer](https://github.com/greentfrapp/attention-primer)\n",
    "* [Transformerの最前線](https://www.slideshare.net/SSII_Slides/ssii2022-ts1-transformer)\n",
    "\n",
    "よく考えたらTransformerについて知らないので実装して確認します。Transformerは「任意の個数のベクトルを任意の個数のベクトルに変換するネットワーク」です。各アルゴリズムは論文を参照してください。\n",
    "\n",
    "## 用語など\n",
    "\n",
    "* Vocabulary: $[N_V]:= \\{1, \\dots, N_V\\}$のIDによって識別される単語や文字の集合のこと。この単語や文字をTokenと呼ぶ。何をTokenにするかは色々工夫があるよ。\n",
    "    * \"My grandma makes the best apple pie.\"を単語レベルでTokenizationすると、\"My\", \"grandma\", ... になり、それぞれをVocabularyのIDに対応させます。\n",
    "* Token embedding: Vocabularyの要素をベクトルに対応させます。例：\"grandma\" -> ID[grandma] -> $\\mathbb{R}^{d_e}$\n",
    "* Positional embedding: Tokenの場所をベクトルに対応させます。例：\"My grandma makes ...\" -> pos[grandma] はここでは２です -> $\\mathbb{R}^{d_e}$\n",
    "\n",
    "\n",
    "### Transformerの初歩\n",
    "\n",
    "まずは適当なデータセットを作ってAttentionの確認をしてみましょう (参考：[Attention Primer](https://github.com/greentfrapp/attention-primer))。文字のカウントをするAttentionを考えます。例えば次のような入力と出力を考えます。\n",
    "\n",
    "\n",
    "```python\n",
    "Input:\n",
    "['A'], ['B'], [' '], ['B'], ['C']\n",
    "Output:\n",
    "[[1], [2], [1]]\n",
    "```\n",
    "\n",
    "これは「文字列を文字の出現数に翻訳する」タスクなわけですね。\n",
    "この場合Positional embeddingは不要なので、tokenのembedはtoken embeddingのみで考えます。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([[[0., 1., 0., 0.],\n",
       "         [1., 0., 0., 0.],\n",
       "         [0., 1., 0., 0.],\n",
       "         [0., 0., 0., 1.],\n",
       "         [0., 1., 0., 0.]]]),\n",
       " array([[[0., 1., 0., 0., 0., 0.],\n",
       "         [0., 0., 0., 1., 0., 0.],\n",
       "         [1., 0., 0., 0., 0., 0.],\n",
       "         [0., 1., 0., 0., 0., 0.]]]))"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import numpy as np\n",
    "import string\n",
    "\n",
    "\n",
    "class Dataset:\n",
    "\tdef __init__(self, max_len=10, vocab_size=3):\n",
    "\t\tself.max_len = max_len\n",
    "\t\tself.vocab_size = vocab_size + 1\n",
    "\t\tself.vocab_embeds = np.eye(self.vocab_size)\n",
    "\t\tself.counts = np.eye(self.max_len + 1)\n",
    "\t\tassert self.vocab_size <= 26 + 1\n",
    "\n",
    "\tdef next_batch(self, batchsize=100):\n",
    "\t\t# A, B, C, ... をone-hotにしたやつです\n",
    "\t\tsampled_tokens = self.vocab_embeds[np.random.choice(np.arange(self.vocab_size), [batchsize, self.max_len])]\n",
    "\n",
    "\t\t# カウント 0, 1, 2, ..., max_len をone-hotにしたやつです\n",
    "\t\tcounts = np.sum(sampled_tokens, axis=1).astype(np.int32)\n",
    "\t\tsampled_counts = self.counts[counts]\n",
    "\t\treturn sampled_tokens, sampled_counts\n",
    "\n",
    "\tdef prettify(self, samples):\n",
    "\t\tsamples = samples.reshape(-1, self.max_len, self.vocab_size)\n",
    "\t\tidx = np.expand_dims(np.argmax(samples, axis=2), axis=2)\n",
    "\t\tdictionary = np.array(list(' ' + string.ascii_uppercase))\n",
    "\t\treturn dictionary[idx]\n",
    "\n",
    "\n",
    "data = Dataset(5, 3)\n",
    "data.next_batch(1)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Attention\n",
    "\n",
    "Attentionは全n個のベクトルから全n個のベクトルへ情報を伝える処理になります（多分）。\n",
    "CNNは近隣のみに情報を伝える処理、RNNは再帰的に情報を伝える処理であり、Attentionの方が広域に情報がロスなく伝達されます。\n",
    "self-attentionでは入力のSequenceが一つですが、encoder-decoder attentionのように、異なるベクトルを入力にすることもできます。\n",
    "今回はencoder-decoderを使います。\n",
    "\n",
    "例えば上の例では、Inputである```[' '], ['A'], ['B'], ['C'] ```がクエリを投げ、Memory (context)である```['A'], ['B'], [' '], ['B'], ['C']```のどれにそれぞれが注目するかをキーで指定します。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 169,
   "metadata": {},
   "outputs": [],
   "source": [
    "import haiku as hk\n",
    "import jax\n",
    "import jax.numpy as jnp\n",
    "from chex import Array\n",
    "import chex\n",
    "\n",
    "\n",
    "def build_pred(\n",
    "    dim_out: int,\n",
    "    dim_attn: int,\n",
    "    dataset: Dataset,\n",
    ") -> hk.Transformed:\n",
    "    \"\"\"単層のAttentionを作ります。また、どの文字が何個出力されたのか？も推論します\"\"\"\n",
    "\n",
    "    vocab_seq = dataset.vocab_embeds\n",
    "    length_vocab, dim_vocab = vocab_seq.shape\n",
    "    length_count = dataset.max_len + 1\n",
    "\n",
    "    @jax.vmap\n",
    "    def forward(memory_seq: Array) -> Array:\n",
    "        chex.assert_rank(memory_seq, 2)\n",
    "        length_memory, dim_memory = memory_seq.shape\n",
    "\n",
    "        # ----- attentionです -----\n",
    "\n",
    "        # queryの計算\n",
    "        W_q = hk.get_parameter(\"W_q\", [dim_vocab, dim_attn], init=hk.initializers.RandomNormal())\n",
    "        b_q = hk.get_parameter(\"b_q\", [1, dim_attn], init=hk.initializers.RandomNormal())\n",
    "        query = vocab_seq @ W_q + b_q\n",
    "        chex.assert_shape(query, (length_vocab, dim_attn))\n",
    "\n",
    "        # keyの計算\n",
    "        W_k = hk.get_parameter(\"W_k\", [dim_memory, dim_attn], init=hk.initializers.RandomNormal())\n",
    "        b_k = hk.get_parameter(\"b_k\", [1, dim_attn], init=hk.initializers.RandomNormal())\n",
    "        key = memory_seq @ W_k + b_k\n",
    "        chex.assert_shape(key, (length_memory, dim_attn))\n",
    "\n",
    "        # valueの計算\n",
    "        W_v = hk.get_parameter(\"W_v\", [dim_memory, dim_out], init=hk.initializers.RandomNormal())\n",
    "        b_v = hk.get_parameter(\"b_v\", [1, dim_out], init=hk.initializers.RandomNormal())\n",
    "        val = memory_seq @ W_v + b_v\n",
    "        chex.assert_shape(val, (length_memory, dim_out))\n",
    "\n",
    "        # scoreと重み付き和の計算\n",
    "        score = (query @ key.T) / jnp.sqrt(dim_attn)\n",
    "        weights = jax.nn.softmax(score, axis=-1)\n",
    "        chex.assert_shape(weights, (length_vocab, length_memory))\n",
    "        weighted_val = weights @ val\n",
    "        chex.assert_shape(weighted_val, (length_vocab, dim_out))\n",
    "        weighted_val = weighted_val\n",
    "        weighted_val = hk.LayerNorm(axis=-1, create_scale=False, create_offset=False)(weighted_val)\n",
    "\n",
    "        # ----- 文字カウントの推論 -----\n",
    "        W_pred = hk.get_parameter(\"W_pred\", [dim_out, length_count], init=hk.initializers.RandomNormal())\n",
    "        pred = weighted_val @ W_pred\n",
    "        chex.assert_shape(pred, (length_vocab, length_count))\n",
    "        return weights, pred\n",
    "\n",
    "    return hk.without_apply_rng(hk.transform(forward))\n",
    "\n",
    "\n",
    "max_len = 5  # \"全n個のベクトル\"のn\n",
    "vocab_size = 3\n",
    "dataset = Dataset(max_len, vocab_size)\n",
    "\n",
    "# attentionの準備\n",
    "dim_out = dim_attn = 32\n",
    "pred_count = build_pred(dim_out, dim_attn, dataset)\n",
    "rng = jax.random.PRNGKey(0)\n",
    "dummy_samples, dummy_labels = dataset.next_batch()\n",
    "_params = pred_count.init(rng, dummy_samples)\n",
    "\n",
    "# テスト\n",
    "batchsize = 16\n",
    "_samples, _labels = dataset.next_batch(batchsize)\n",
    "attention, pred = pred_count.apply(_params, _samples)\n",
    "chex.assert_shape(attention, (batchsize, dataset.vocab_size, dataset.max_len))\n",
    "chex.assert_shape(pred, (batchsize, dataset.vocab_size, dataset.max_len + 1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 178,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[<matplotlib.lines.Line2D at 0x7f43e263d220>]"
      ]
     },
     "execution_count": 178,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 640x480 with 1 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "# 学習させてみます\n",
    "\n",
    "import optax\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "max_len = 10  # \"全n個のベクトル\"のn\n",
    "vocab_size = 3\n",
    "dataset = Dataset(max_len, vocab_size)\n",
    "\n",
    "# 推論器の準備\n",
    "dim_out = dim_attn = 64\n",
    "pred_count = build_pred(dim_out, dim_attn, dataset)\n",
    "\n",
    "\n",
    "# optimizerの準備\n",
    "optimizer = optax.adam(1e-2)\n",
    "\n",
    "\n",
    "def compute_loss(params, input_seqs, labels):\n",
    "    _, preds = pred_count.apply(params, input_seqs)\n",
    "    return jnp.mean(optax.softmax_cross_entropy(preds, labels))\n",
    "\n",
    "\n",
    "@jax.jit\n",
    "def compute_params(params, opt_state, input_seqs, labels):\n",
    "    loss, grad = jax.value_and_grad(compute_loss)(params, input_seqs, labels)\n",
    "    update, opt_state = optimizer.update(grad, opt_state, params)\n",
    "    params = optax.apply_updates(params, update)\n",
    "    return loss, params\n",
    "\n",
    "\n",
    "rng = jax.random.PRNGKey(0)\n",
    "dummy_samples, dummy_labels = dataset.next_batch()\n",
    "params = pred_count.init(rng, dummy_samples)\n",
    "opt_state = optimizer.init(params)\n",
    "\n",
    "\n",
    "# 学習\n",
    "batch_size = 100\n",
    "losses = []\n",
    "for i in range(10000):\n",
    "    input_seqs, labels = dataset.next_batch(batchsize)\n",
    "    loss, params = compute_params(params, opt_state, input_seqs, labels)\n",
    "    if i % 100 == 0:\n",
    "        losses.append(loss)\n",
    "\n",
    "plt.plot(losses)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "いい感じに学習できた気がします (意外とLayer Normが大事っぽい)。Attentionが何を見ているのか可視化してみましょう。"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 184,
   "metadata": {},
   "outputs": [],
   "source": [
    "input_seq, label = dataset.next_batch(1)\n",
    "attention, pred = pred_count.apply(params, input_seq)\n",
    "\n",
    "label = label[0, :, :]\n",
    "input_seq = input_seq[0, :, :]\n",
    "attention = np.array(attention[0, :, :])\n",
    "_pred = np.array(pred[0, :, :])\n",
    "pred = np.zeros_like(_pred)\n",
    "for i in range(pred.shape[0]):\n",
    "    pred[i, _pred[i].argmax(-1)] = 1.0"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 185,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 2400x400 with 8 Axes>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "import seaborn as sns\n",
    "\n",
    "plt.figure(figsize=(24, 4))\n",
    "\n",
    "plt.subplot(1, 4, 1)\n",
    "sns.heatmap(input_seq.T)\n",
    "plt.ylabel(\"Character ID\")\n",
    "plt.xlabel(\"Sequence\")\n",
    "plt.title(\"Input Sequence\")\n",
    "\n",
    "plt.subplot(1, 4, 2)\n",
    "sns.heatmap(label)\n",
    "plt.ylabel(\"Character ID\")\n",
    "plt.xlabel(\"Counts\")\n",
    "plt.title(\"Label\")\n",
    "\n",
    "plt.subplot(1, 4, 3)\n",
    "sns.heatmap(pred)\n",
    "plt.ylabel(\"Character ID\")\n",
    "plt.xlabel(\"Counts\")\n",
    "plt.title(\"Prediction\")\n",
    "\n",
    "\n",
    "plt.subplot(1, 4, 4)\n",
    "sns.heatmap(attention)\n",
    "plt.ylabel(\"Character ID\")\n",
    "plt.xlabel(\"Counts\")\n",
    "plt.title(f\"Attention\")\n",
    "\n",
    "plt.tight_layout()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3.9.12 ('shumi-VTLwuKSy-py3.9')",
   "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": "3.9.12"
  },
  "orig_nbformat": 4,
  "vscode": {
   "interpreter": {
    "hash": "d6b7cac5e0d2ff733f340da4d53ae5ecfef7f7ad39623f5982b029a09306b36b"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}