play_seq.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "play_seq.ipynb",
      "provenance": [],
      "collapsed_sections": [],
      "authorship_tag": "ABX9TyM/XRun6jod0zAcVZlaRW0s",
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU",
    "gpuClass": "standard"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "<a href=\"https://colab.research.google.com/github/myazdani/hacking-minGPT/blob/main/play_seq.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# Hacking minGPT with a toy sequential task\n",
        "\n",
        "François Fleuret came up with a toy sequential task to get better intuition with attention layers. Here I'll take the GPT architecture (minGPT by Andrej Karpathy specifically) to (sorta) solve the task. \n",
        "\n",
        "The task is summarized by this slide from François' deep learning course EE-559: \n",
        "\n",
        "<div>\n",
        "<img src=\"https://i.imgur.com/ez02xV9.png\" width=\"700\"/>\n",
        "</div>\n",
        "\n",
        "- the input: pairs of triangle and rectangle shapes at different heights. \n",
        "- the desired output: have the same shapes but each pair needs to have the same average height. \n",
        "\n",
        "\n",
        "<div>\n",
        "<img src=\"https://i.imgur.com/zDHnv7B.png\" width=\"700\"/>\n",
        "</div>\n",
        "\n",
        "For example, the output two triangle pairs both need to have the same height and this height needs to be the average height of the short and tall traingle pairs in the inputs."
      ],
      "metadata": {
        "id": "CeeogU4ByfWB"
      }
    },
    {
      "cell_type": "code",
      "execution_count": 1,
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "bfAysbgbIF1B",
        "outputId": "144d89df-d63e-4a9b-86b2-4c77a0c9e4e1"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Cloning into 'hacking-minGPT'...\n",
            "remote: Enumerating objects: 110, done.\u001b[K\n",
            "remote: Counting objects: 100% (110/110), done.\u001b[K\n",
            "remote: Compressing objects: 100% (77/77), done.\u001b[K\n",
            "remote: Total 110 (delta 69), reused 66 (delta 31), pack-reused 0\u001b[K\n",
            "Receiving objects: 100% (110/110), 1.10 MiB | 7.02 MiB/s, done.\n",
            "Resolving deltas: 100% (69/69), done.\n",
            "--2022-07-13 04:49:50--  https://raw.githubusercontent.com/myazdani/attention-layer-demo/master/utils.py\n",
            "Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 185.199.108.133, 185.199.109.133, 185.199.110.133, ...\n",
            "Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|185.199.108.133|:443... connected.\n",
            "HTTP request sent, awaiting response... 200 OK\n",
            "Length: 3583 (3.5K) [text/plain]\n",
            "Saving to: ‘utils.py’\n",
            "\n",
            "utils.py            100%[===================>]   3.50K  --.-KB/s    in 0s      \n",
            "\n",
            "2022-07-13 04:49:51 (53.5 MB/s) - ‘utils.py’ saved [3583/3583]\n",
            "\n"
          ]
        }
      ],
      "source": [
        "## only needed if running on google colab\n",
        "!git clone https://github.com/myazdani/hacking-minGPT.git\n",
        "\n",
        "import sys\n",
        "  \n",
        "# adding Folder_2 to the system path\n",
        "sys.path.insert(0, './hacking-minGPT/')\n",
        "\n",
        "\n",
        "!wget https://raw.githubusercontent.com/myazdani/attention-layer-demo/master/utils.py"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## get the data"
      ],
      "metadata": {
        "id": "Pct6_kwdyUjv"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# make deterministic\n",
        "import numpy as np\n",
        "import torch\n",
        "import torch.nn as nn\n",
        "from torch.nn import functional as F\n",
        "import matplotlib.pyplot as plt\n",
        "from mingpt.utils import set_seed\n",
        "set_seed(42)\n",
        "import utils as seq_utils\n",
        "\n",
        "data_gen = seq_utils.DataUtil(seq_height_min = 1.0, \n",
        "                          seq_height_max = 25.0,\n",
        "                          seq_width_min = 5.0, \n",
        "                          seq_width_max = 11.0,\n",
        "                          seq_length = 100,\n",
        "                          group_by_locations = False\n",
        "                         )"
      ],
      "metadata": {
        "id": "R3Gkb51pILS4"
      },
      "execution_count": 2,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "train_input, train_targets, train_tr, train_bx = data_gen.generate_sequences(25000)\n",
        "test_input, test_targets, test_tr, test_bx = data_gen.generate_sequences(1000)"
      ],
      "metadata": {
        "id": "sB0nzbcfIabR"
      },
      "execution_count": 3,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "plt.plot(train_input[1,0,:], label = \"input\")\n",
        "plt.plot(train_targets[1,0,:], label = \"target\")\n",
        "plt.xlabel(\"Sequence id\")\n",
        "plt.legend();"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 279
        },
        "id": "WA8IUkJrth83",
        "outputId": "157008d8-c3d0-499e-e939-7610ee998388"
      },
      "execution_count": 4,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "### cluster the raw sequence values into discrete clusters\n",
        "\n",
        "Since the input of the minGPT needs to be discrete, I discretize with K-means. Karpathy does similar for applying minGPT to CIFAR images. "
      ],
      "metadata": {
        "id": "YwAnfJBOwYqt"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# get random 5 vals per seq and stack them all up \n",
        "pluck_vals = lambda x: x[0,torch.randperm(100)[:5]]\n",
        "random_vals = torch.cat([pluck_vals(x) for x in train_input]).numpy()\n",
        "print(random_vals.shape)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "7BIa0rAoImE7",
        "outputId": "f7e95609-6e99-4e6f-ffd2-5eed7f82b041"
      },
      "execution_count": 5,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "(125000,)\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "from sklearn.cluster import KMeans\n",
        "num_clusters = 256\n",
        "kmeans = KMeans(n_clusters=num_clusters, random_state=0).fit(random_vals[:,np.newaxis])\n",
        "kmeans.labels_"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "NTqsSl74IpaX",
        "outputId": "2e08f617-3f80-43b8-bf3b-f708d5df9f23"
      },
      "execution_count": 6,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "array([216, 209, 255, ...,  71, 228,  96], dtype=int32)"
            ]
          },
          "metadata": {},
          "execution_count": 6
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Wrap discretizes sequences into a dataloader"
      ],
      "metadata": {
        "id": "K9CaNHxnzssu"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "from torch.utils.data import Dataset\n",
        "\n",
        "class SeqDataset(Dataset):\n",
        "    \"\"\"\n",
        "    wrap up seq dataset into our own, which will convert floating points into sequences of integers\n",
        "    \"\"\"\n",
        "    \n",
        "    def __init__(self, inputs, targets, kmeans_clusters):\n",
        "        self.inputs = inputs\n",
        "        self.targets = targets\n",
        "        self.kmeans_clusters = kmeans_clusters\n",
        "        \n",
        "    def __len__(self):\n",
        "        return len(self.inputs)\n",
        "\n",
        "    def __getitem__(self, idx):\n",
        "        x, y = self.inputs[idx], self.targets[idx]\n",
        "        batch_size = x.shape[0]\n",
        "        x_np = x.view(-1,1).numpy()\n",
        "        y_np = y.view(-1,1).numpy()\n",
        "        x_clusters = torch.from_numpy(self.kmeans_clusters.predict(x_np)).view(batch_size,-1).long()\n",
        "        y_clusters = torch.from_numpy(self.kmeans_clusters.predict(y_np)).view(batch_size,-1).long()\n",
        "\n",
        "        return x_clusters.squeeze(), y_clusters.squeeze()"
      ],
      "metadata": {
        "id": "VWUv1riM1Sjc"
      },
      "execution_count": 7,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "train_dataset = SeqDataset(train_input, train_targets, kmeans)\n",
        "test_dataset = SeqDataset(test_input, test_targets, kmeans)\n",
        "train_dataset[0][0] # one example sequence flattened out into integers"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "VSysm7cG39E7",
        "outputId": "443403d6-ad3f-4852-8e72-9efff8219665"
      },
      "execution_count": 8,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "tensor([ 80, 117,  49, 246,  32, 124, 255,  80, 178,   0, 117, 216, 117,   0,\n",
              "          0, 228, 156,  95, 241,  95, 156, 241,  96, 246, 206,   5,  77, 191,\n",
              "        163, 106, 115, 170, 113, 232, 231, 176,  96, 178, 216,  80,  80, 110,\n",
              "         86,  50, 159, 137, 148, 144, 188,  75, 255,  32, 176,  32, 117, 180,\n",
              "         32, 216,  71,  49,  14,  80, 255, 178,   0,  80, 255, 178,  83, 209,\n",
              "         83, 223, 209, 133, 133, 124, 216, 246, 228, 167, 178, 176,  14, 216,\n",
              "        180, 180,  71, 178,  71,  49, 167, 176,  49,  14, 117,  96, 117,  49,\n",
              "         14, 246])"
            ]
          },
          "metadata": {},
          "execution_count": 8
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "# GPT training\n",
        "\n",
        "mostly using paramters used for CIFAR; might be a bit overkill?"
      ],
      "metadata": {
        "id": "Lz8etkO7z0TB"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "from mingpt.model import GPT, GPTConfig\n",
        "block_size = 100\n",
        "mconf = GPTConfig(num_clusters, block_size,\n",
        "                  embd_pdrop=0.0, resid_pdrop=0.0, attn_pdrop=0.0,\n",
        "                  n_layer=12, n_head=8, n_embd=256)\n",
        "model = GPT(mconf)\n"
      ],
      "metadata": {
        "id": "bmLdj1Ym4aHc"
      },
      "execution_count": 9,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from mingpt.trainer import Trainer, TrainerConfig\n",
        "\n",
        "# initialize a trainer instance and kick off training\n",
        "# kinda guessing these\n",
        "tconf = TrainerConfig(max_epochs=10, batch_size=256, learning_rate=1e-4,\n",
        "                      lr_decay=True, warmup_tokens=1024, \n",
        "                      num_workers=2)\n",
        "\n",
        "trainer = Trainer(model, train_dataset, None, tconf)\n",
        "trainer.train()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "op7rQ4uf4kbD",
        "outputId": "aec05b62-647a-437f-fad2-e46db1149ebb"
      },
      "execution_count": 10,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "epoch 1 iter 97: train loss 3.76646. lr 1.000000e-04: 100%|██████████| 98/98 [00:44<00:00,  2.18it/s]\n",
            "epoch 2 iter 97: train loss 3.52670. lr 1.000000e-04: 100%|██████████| 98/98 [00:42<00:00,  2.32it/s]\n",
            "epoch 3 iter 97: train loss 3.46497. lr 1.000000e-04: 100%|██████████| 98/98 [00:41<00:00,  2.34it/s]\n",
            "epoch 4 iter 97: train loss 3.38630. lr 1.000000e-04: 100%|██████████| 98/98 [00:42<00:00,  2.33it/s]\n",
            "epoch 5 iter 97: train loss 3.32625. lr 1.000000e-04: 100%|██████████| 98/98 [00:42<00:00,  2.33it/s]\n",
            "epoch 6 iter 97: train loss 3.27124. lr 1.000000e-04: 100%|██████████| 98/98 [00:42<00:00,  2.33it/s]\n",
            "epoch 7 iter 97: train loss 3.21525. lr 1.000000e-04: 100%|██████████| 98/98 [00:41<00:00,  2.33it/s]\n",
            "epoch 8 iter 97: train loss 3.17643. lr 1.000000e-04: 100%|██████████| 98/98 [00:42<00:00,  2.32it/s]\n",
            "epoch 9 iter 97: train loss 3.14479. lr 1.000000e-04: 100%|██████████| 98/98 [00:42<00:00,  2.33it/s]\n",
            "epoch 10 iter 97: train loss 3.11946. lr 1.000000e-04: 100%|██████████| 98/98 [00:41<00:00,  2.34it/s]\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "def compare_predictions(example_pair):\n",
        "    device = 'cuda' if torch.cuda.is_available() else 'cpu'\n",
        "    pred = model(example_pair[0].view(1,-1).to(device))\n",
        "\n",
        "    # inverse kmeans transform the inputs, predictions, and outputs\n",
        "    raw_input = kmeans.cluster_centers_[:,0][example_pair[0].cpu().numpy()]\n",
        "    raw_pred = kmeans.cluster_centers_[:,0][pred[0].argmax(axis=2).cpu().numpy().squeeze()]\n",
        "    raw_output = kmeans.cluster_centers_[:,0][example_pair[1].cpu().numpy()]\n",
        "    plt.plot(raw_input, label = 'input')\n",
        "    plt.plot(raw_pred, label = 'pred target')\n",
        "    plt.plot(raw_output, label = 'actual target')\n",
        "    plt.legend();"
      ],
      "metadata": {
        "id": "mOblsisCs0LS"
      },
      "execution_count": 11,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "compare_predictions(test_dataset[0])"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 265
        },
        "id": "gxGw_Ba20o_x",
        "outputId": "81420a88-26cb-48e6-9086-0083eac3bfa2"
      },
      "execution_count": 12,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 432x288 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {
            "needs_background": "light"
          }
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "# looks like we need some more training iterations.."
      ],
      "metadata": {
        "id": "V30sIs70xX61"
      },
      "execution_count": 25,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        ""
      ],
      "metadata": {
        "id": "KQMgQpsr4q0Y"
      },
      "execution_count": null,
      "outputs": []
    }
  ]
}