Story_Generation_T5.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "## Installing Transformers library"
      ],
      "metadata": {
        "id": "gMPsfSIMQTFs"
      }
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "id": "vpJtgoRkQf8x",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "c8e9da98-d57b-496c-e07a-35bf8ba2e50e"
      },
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/\n",
            "Collecting transformers\n",
            "  Downloading transformers-4.30.2-py3-none-any.whl (7.2 MB)\n",
            "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m7.2/7.2 MB\u001b[0m \u001b[31m85.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
            "\u001b[?25hRequirement already satisfied: filelock in /usr/local/lib/python3.10/dist-packages (from transformers) (3.12.0)\n",
            "Collecting huggingface-hub<1.0,>=0.14.1 (from transformers)\n",
            "  Downloading huggingface_hub-0.15.1-py3-none-any.whl (236 kB)\n",
            "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m236.8/236.8 kB\u001b[0m \u001b[31m28.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
            "\u001b[?25hRequirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.10/dist-packages (from transformers) (1.22.4)\n",
            "Requirement already satisfied: packaging>=20.0 in /usr/local/lib/python3.10/dist-packages (from transformers) (23.1)\n",
            "Requirement already satisfied: pyyaml>=5.1 in /usr/local/lib/python3.10/dist-packages (from transformers) (6.0)\n",
            "Requirement already satisfied: regex!=2019.12.17 in /usr/local/lib/python3.10/dist-packages (from transformers) (2022.10.31)\n",
            "Requirement already satisfied: requests in /usr/local/lib/python3.10/dist-packages (from transformers) (2.27.1)\n",
            "Collecting tokenizers!=0.11.3,<0.14,>=0.11.1 (from transformers)\n",
            "  Downloading tokenizers-0.13.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (7.8 MB)\n",
            "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m7.8/7.8 MB\u001b[0m \u001b[31m39.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
            "\u001b[?25hCollecting safetensors>=0.3.1 (from transformers)\n",
            "  Downloading safetensors-0.3.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (1.3 MB)\n",
            "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.3/1.3 MB\u001b[0m \u001b[31m24.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
            "\u001b[?25hRequirement already satisfied: tqdm>=4.27 in /usr/local/lib/python3.10/dist-packages (from transformers) (4.65.0)\n",
            "Requirement already satisfied: fsspec in /usr/local/lib/python3.10/dist-packages (from huggingface-hub<1.0,>=0.14.1->transformers) (2023.4.0)\n",
            "Requirement already satisfied: typing-extensions>=3.7.4.3 in /usr/local/lib/python3.10/dist-packages (from huggingface-hub<1.0,>=0.14.1->transformers) (4.5.0)\n",
            "Requirement already satisfied: urllib3<1.27,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests->transformers) (1.26.15)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests->transformers) (2022.12.7)\n",
            "Requirement already satisfied: charset-normalizer~=2.0.0 in /usr/local/lib/python3.10/dist-packages (from requests->transformers) (2.0.12)\n",
            "Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests->transformers) (3.4)\n",
            "Installing collected packages: tokenizers, safetensors, huggingface-hub, transformers\n",
            "Successfully installed huggingface-hub-0.15.1 safetensors-0.3.1 tokenizers-0.13.3 transformers-4.30.2\n"
          ]
        }
      ],
      "source": [
        "#Installing transformer package\n",
        "!pip install transformers"
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Rating Estimation by ChatGPT"
      ],
      "metadata": {
        "id": "pwf9LGiCzYmo"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!pip install openai"
      ],
      "metadata": {
        "id": "dzwU4eGtvhjh"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "import os\n",
        "import openai\n",
        "os.environ[\"OpenAI_API_Key\"] = \"\"\n",
        "api_key = os.environ.get(\"OpenAI_API_Key\")\n",
        "openai.api_key = api_key"
      ],
      "metadata": {
        "id": "s67L7e5yvjSX"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Loss to be added in the custom loss\n",
        "def get_chatgpt_rating(prompt, sample):\n",
        "  completion = openai.ChatCompletion.create(\n",
        "  model=\"gpt-3.5-turbo\",\n",
        "  messages=[\n",
        "    {\"role\": \"user\",\n",
        "     \"content\": f\"{prompt} {sample}\"}\n",
        "  ])\n",
        "  return 10 - int(completion.choices[0].message)"
      ],
      "metadata": {
        "id": "8tB5L6nFLhDS"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Different prompts tested in order to generate a sensible rating\n",
        "prompt1 = \"Provide me rating between 0 and 10 (without any explanation), where 0 is the best and 10 is the worst, for the following story summary: \"\n",
        "prompt2 = \"Assign a rating between 0 (best) and 10 (worst) to the given artificial story summary (only give rating as the response):\"\n",
        "prompt3 = \"Assign a rating between 0 (best) and 10 (worst) to the given artificial story summary (only give rating as the response). The rating should be based on writing style, coherence and capture strength. Summary:\"\n",
        "prompt4 = \"Assign a rating between 0 and 10 to the given artificial story summary. Only give rating as the response (no reasoning). The rating should be based on writing style, coherence, and capture strength. Summary:\" # Best\n",
        "prompt5 = \"Provide me rating between 0 and 10 (without any explanation),  for the following story summary:\""
      ],
      "metadata": {
        "id": "MQJRNPghzb7g"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Data Tokenization/Encoding"
      ],
      "metadata": {
        "id": "2dYlSUZNQYop"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "#Loading the standard T5 small model and tokenizer\n",
        "from transformers import AutoTokenizer, AutoModelForSeq2SeqLM\n",
        "# Can try different T5 models such as T5-large, T5-3B, T5-11B\n",
        "base_tokenizer = AutoTokenizer.from_pretrained('t5-small')\n",
        "base_model = AutoModelForSeq2SeqLM.from_pretrained(\"t5-small\")"
      ],
      "metadata": {
        "id": "bVkVGdHBYQat"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Encoding the sequences\n",
        "def encode_sequences(x, base_tokenizer = base_tokenizer):\n",
        "  # try:\n",
        "    # Input consists of different aspects of the story on which the output will be conditioned\n",
        "    input = x['Input']\n",
        "    # Label is the conditioned output - Story\n",
        "    label = x['Summary']\n",
        "    # Max length of the input sequence in T5 is 512 tokens (BART could be used for longer sequences - 1024 max length limit)\n",
        "    model_input = base_tokenizer(input, max_length = 512, truncation=True, padding='max_length')\n",
        "    model_input['labels'] = base_tokenizer(label, max_length = 512, truncation=True, padding='max_length')['input_ids']\n",
        "    return model_input"
      ],
      "metadata": {
        "id": "pz4X3H2rMDTE"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Loading the final dataset\n",
        "import pandas as pd\n",
        "df = pd.read_csv('/content/drive/MyDrive/Visual Story Telling/Dataset - Story Generation/Plot_Summary_Dataset')"
      ],
      "metadata": {
        "id": "9F_s4EA1YTER"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from sklearn.model_selection import train_test_split\n",
        "train_df, test_df = train_test_split(df, test_size=0.1, random_state=1)"
      ],
      "metadata": {
        "id": "q7Y0PUMjTqO_"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Tokenizing the dataset\n",
        "train_df = train_df.apply(encode_sequences, axis=1)\n",
        "test_df = test_df.apply(encode_sequences, axis = 1)"
      ],
      "metadata": {
        "id": "DTsB1rXyMp-x"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "train_df.dropna(inplace = True)\n",
        "test_df.dropna(inplace = True)"
      ],
      "metadata": {
        "id": "zqw4aSD8Agme"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "train_df.reset_index(drop=True, inplace=True)\n",
        "test_df.reset_index(drop=True, inplace=True)"
      ],
      "metadata": {
        "id": "fZr3FDXBoWgk"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Number of parameters"
      ],
      "metadata": {
        "id": "EOtQstapqEmC"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# p is accessing each layer's weight tensors of the model\n",
        "# p.requires_grad() checks whether the layer is frozen for training or not\n",
        "# p.numel returns total number of elements in the weight tensor - total number of parameters\n",
        "print(\"Parameters:\", sum(p.numel() for p in base_model.parameters() if p.requires_grad))"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "anOpPgTYqHKP",
        "outputId": "0bb95ec4-57ad-4390-8435-13755511b536"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Parameters: 60506624\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Accessing different layers of a Model"
      ],
      "metadata": {
        "id": "RBrX26Az6T0-"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "base_model"
      ],
      "metadata": {
        "id": "qlMuhybHs83r"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# We access each layer by the key. That is to access (x), we use base_model.x\n",
        "# Try and go layer by layer. First check the output of a.b, then go for a.b.c\n",
        "layer_id = 0\n",
        "base_model.encoder.block[layer_id].layer[0].SelfAttention.o"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "lew66TwBunRw",
        "outputId": "9b75e849-d532-4dbb-89e3-c671f46c304c"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Linear(in_features=512, out_features=512, bias=False)"
            ]
          },
          "metadata": {},
          "execution_count": 52
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "base_model.encoder.block[layer_id].layer[1].DenseReluDense.wo"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "ODJLQz06zUdz",
        "outputId": "484965cb-0499-401f-bed3-7dd362261e05"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Linear(in_features=2048, out_features=512, bias=False)"
            ]
          },
          "metadata": {},
          "execution_count": 40
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "base_model.decoder.block[layer_id].layer[0].SelfAttention.o"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "smKahN3syi5e",
        "outputId": "71c8f93d-a0f4-4657-889b-0e07a5a9b804"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Linear(in_features=512, out_features=512, bias=False)"
            ]
          },
          "metadata": {},
          "execution_count": 37
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "base_model.decoder.block[layer_id].layer[2].DenseReluDense.wo"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "SnQ9hTzvziWN",
        "outputId": "03a8518f-7593-48c5-e6c9-741790155642"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "Linear(in_features=2048, out_features=512, bias=False)"
            ]
          },
          "metadata": {},
          "execution_count": 46
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## T5 w Adapters"
      ],
      "metadata": {
        "id": "Jj7Yr65aoj9I"
      }
    },
    {
      "cell_type": "markdown",
      "source": [
        "Could also utilize [adapter-transformers](https://github.com/adapter-hub/adapter-transformers) library - provides easy abstract way of incorporating adapters in the models"
      ],
      "metadata": {
        "id": "fhJizTKMw_Y9"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# Freeze all layers\n",
        "for param in base_model.parameters():\n",
        "    param.requires_grad = False"
      ],
      "metadata": {
        "id": "9Xc0E7qMonGV"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Function to count number of trainable paramters in a layer/block\n",
        "def count_parameters(model):\n",
        "  return sum(p.numel() for p in model.parameters() if p.requires_grad)"
      ],
      "metadata": {
        "id": "zv5ieus3zA2T"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "import torch\n",
        "# Creating a custom Adapter layer\n",
        "def make_adapter(in_dim, bottleneck_dim, out_dim):\n",
        "    adapter_layers = torch.nn.Sequential(\n",
        "        torch.nn.Linear(in_dim, bottleneck_dim),\n",
        "        torch.nn.GELU(),\n",
        "        torch.nn.Linear(bottleneck_dim, out_dim),\n",
        "    )\n",
        "    return adapter_layers"
      ],
      "metadata": {
        "id": "CtRwu_ikpZgI"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "Due to the forward pass code of DenseReluDense, we cannot add adpaters as a Sequential block in the Feed forward layer"
      ],
      "metadata": {
        "id": "mbQwZDl_tEgO"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "total_size = 0\n",
        "bottleneck_size = 32 # hyperparameter\n",
        "\n",
        "for layer_id in range(6):\n",
        "\n",
        "    #############################################################\n",
        "    # insert 1st adapter layer into transformer block in Encoder\n",
        "    #############################################################\n",
        "\n",
        "    orig_layer_1 = base_model.encoder.block[layer_id].layer[0].SelfAttention.o\n",
        "\n",
        "    adapter_layers_1 = make_adapter(\n",
        "        in_dim=orig_layer_1.out_features,\n",
        "        bottleneck_dim=bottleneck_size,\n",
        "        out_dim=orig_layer_1.out_features)\n",
        "\n",
        "    new_1 = torch.nn.Sequential(orig_layer_1, *adapter_layers_1)\n",
        "    base_model.encoder.block[layer_id].layer[0].SelfAttention.o = new_1\n",
        "\n",
        "    total_size += count_parameters(adapter_layers_1)\n",
        "\n",
        "    #############################################################\n",
        "    # insert 1st adapter layer into transformer block in Decoder\n",
        "    #############################################################\n",
        "\n",
        "    orig_layer_1 = base_model.decoder.block[layer_id].layer[0].SelfAttention.o\n",
        "\n",
        "    adapter_layers_1 = make_adapter(\n",
        "        in_dim=orig_layer_1.out_features,\n",
        "        bottleneck_dim=bottleneck_size,\n",
        "        out_dim=orig_layer_1.out_features)\n",
        "\n",
        "    new_1 = torch.nn.Sequential(orig_layer_1, *adapter_layers_1)\n",
        "    base_model.decoder.block[layer_id].layer[0].SelfAttention.o = new_1\n",
        "\n",
        "    total_size += count_parameters(adapter_layers_1)\n",
        "\n",
        "print(\"Number of adapter parameters added:\", total_size)"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "71_MYe2qptFJ",
        "outputId": "cda08659-7347-4f2d-c9b4-9a3268bbb6ec"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Number of adapter parameters added: 399744\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## T5 w LoRA"
      ],
      "metadata": {
        "id": "8XA-3Jngonjc"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# Installing PEFT Library\n",
        "!pip install peft"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "M6tJK6CZosIn",
        "outputId": "e9592691-46cf-4e2a-d283-a9fbd379f3e3"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "Looking in indexes: https://pypi.org/simple, https://us-python.pkg.dev/colab-wheels/public/simple/\n",
            "Collecting peft\n",
            "  Downloading peft-0.3.0-py3-none-any.whl (56 kB)\n",
            "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m56.8/56.8 kB\u001b[0m \u001b[31m7.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
            "\u001b[?25hRequirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.10/dist-packages (from peft) (1.22.4)\n",
            "Requirement already satisfied: packaging>=20.0 in /usr/local/lib/python3.10/dist-packages (from peft) (23.1)\n",
            "Requirement already satisfied: psutil in /usr/local/lib/python3.10/dist-packages (from peft) (5.9.5)\n",
            "Requirement already satisfied: pyyaml in /usr/local/lib/python3.10/dist-packages (from peft) (6.0)\n",
            "Requirement already satisfied: torch>=1.13.0 in /usr/local/lib/python3.10/dist-packages (from peft) (2.0.1+cu118)\n",
            "Requirement already satisfied: transformers in /usr/local/lib/python3.10/dist-packages (from peft) (4.30.1)\n",
            "Collecting accelerate (from peft)\n",
            "  Downloading accelerate-0.20.3-py3-none-any.whl (227 kB)\n",
            "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m227.6/227.6 kB\u001b[0m \u001b[31m20.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
            "\u001b[?25hRequirement already satisfied: filelock in /usr/local/lib/python3.10/dist-packages (from torch>=1.13.0->peft) (3.12.0)\n",
            "Requirement already satisfied: typing-extensions in /usr/local/lib/python3.10/dist-packages (from torch>=1.13.0->peft) (4.5.0)\n",
            "Requirement already satisfied: sympy in /usr/local/lib/python3.10/dist-packages (from torch>=1.13.0->peft) (1.11.1)\n",
            "Requirement already satisfied: networkx in /usr/local/lib/python3.10/dist-packages (from torch>=1.13.0->peft) (3.1)\n",
            "Requirement already satisfied: jinja2 in /usr/local/lib/python3.10/dist-packages (from torch>=1.13.0->peft) (3.1.2)\n",
            "Requirement already satisfied: triton==2.0.0 in /usr/local/lib/python3.10/dist-packages (from torch>=1.13.0->peft) (2.0.0)\n",
            "Requirement already satisfied: cmake in /usr/local/lib/python3.10/dist-packages (from triton==2.0.0->torch>=1.13.0->peft) (3.25.2)\n",
            "Requirement already satisfied: lit in /usr/local/lib/python3.10/dist-packages (from triton==2.0.0->torch>=1.13.0->peft) (16.0.5)\n",
            "Requirement already satisfied: huggingface-hub<1.0,>=0.14.1 in /usr/local/lib/python3.10/dist-packages (from transformers->peft) (0.15.1)\n",
            "Requirement already satisfied: regex!=2019.12.17 in /usr/local/lib/python3.10/dist-packages (from transformers->peft) (2022.10.31)\n",
            "Requirement already satisfied: requests in /usr/local/lib/python3.10/dist-packages (from transformers->peft) (2.27.1)\n",
            "Requirement already satisfied: tokenizers!=0.11.3,<0.14,>=0.11.1 in /usr/local/lib/python3.10/dist-packages (from transformers->peft) (0.13.3)\n",
            "Requirement already satisfied: safetensors>=0.3.1 in /usr/local/lib/python3.10/dist-packages (from transformers->peft) (0.3.1)\n",
            "Requirement already satisfied: tqdm>=4.27 in /usr/local/lib/python3.10/dist-packages (from transformers->peft) (4.65.0)\n",
            "Requirement already satisfied: fsspec in /usr/local/lib/python3.10/dist-packages (from huggingface-hub<1.0,>=0.14.1->transformers->peft) (2023.4.0)\n",
            "Requirement already satisfied: MarkupSafe>=2.0 in /usr/local/lib/python3.10/dist-packages (from jinja2->torch>=1.13.0->peft) (2.1.2)\n",
            "Requirement already satisfied: urllib3<1.27,>=1.21.1 in /usr/local/lib/python3.10/dist-packages (from requests->transformers->peft) (1.26.15)\n",
            "Requirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.10/dist-packages (from requests->transformers->peft) (2022.12.7)\n",
            "Requirement already satisfied: charset-normalizer~=2.0.0 in /usr/local/lib/python3.10/dist-packages (from requests->transformers->peft) (2.0.12)\n",
            "Requirement already satisfied: idna<4,>=2.5 in /usr/local/lib/python3.10/dist-packages (from requests->transformers->peft) (3.4)\n",
            "Requirement already satisfied: mpmath>=0.19 in /usr/local/lib/python3.10/dist-packages (from sympy->torch>=1.13.0->peft) (1.3.0)\n",
            "Installing collected packages: accelerate, peft\n",
            "Successfully installed accelerate-0.20.3 peft-0.3.0\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "from peft import get_peft_model, LoraConfig, TaskType\n",
        "\n",
        "# We are usig LoRA for parameter efficient fine tuning of a Seq2SeqLM. Arguments' default values are taken from the demonstrations\n",
        "peft_config = LoraConfig(\n",
        "    task_type=TaskType.SEQ_2_SEQ_LM, inference_mode=False, r=8, lora_alpha=32, lora_dropout=0.1\n",
        ")"
      ],
      "metadata": {
        "id": "MBIVz6M12bdX"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Customizing the model foe peft\n",
        "base_model = get_peft_model(base_model, peft_config)\n",
        "# Number of trainable parameters in the model in PEFT model\n",
        "base_model.print_trainable_parameters()"
      ],
      "metadata": {
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "id": "dgyK0Z7l3O4W",
        "outputId": "97ad5f6f-cb28-410e-f85d-218674b58576"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stdout",
          "text": [
            "trainable params: 294912 || all params: 60801536 || trainable%: 0.4850403779272945\n"
          ]
        }
      ]
    },
    {
      "cell_type": "markdown",
      "source": [
        "## T5 Training Setup (Includes Custom Loss Function)"
      ],
      "metadata": {
        "id": "_ASa0fbnQe4u"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# Initializing the Data Collator for batching of the dataset\n",
        "from transformers import DataCollatorForSeq2Seq\n",
        "data_collator = DataCollatorForSeq2Seq(\n",
        "        tokenizer=base_tokenizer,\n",
        "        return_tensors=\"pt\",\n",
        "        max_length = 512,\n",
        "        padding = 'max_length',\n",
        "        model = base_model\n",
        "    )"
      ],
      "metadata": {
        "id": "k_oMtcLHNMyo"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "!pip install accelerate"
      ],
      "metadata": {
        "id": "jXxRUyW39Fr6"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from transformers import Seq2SeqTrainingArguments, Seq2SeqTrainer\n",
        "\n",
        "# Path where model training loss and intermediate weights will be stored\n",
        "\"\"\"\n",
        "Change model path to\n",
        "1) /content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/Basic - when using no pEDT methods\n",
        "2) /content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/Adapters - when using Adapters\n",
        "3) /content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/LoRA - when using LoRA\n",
        "\"\"\"\n",
        "model_path = f'/content/drive/MyDrive/Visual Story Telling/Story_Gen_Model'\n",
        "\n",
        "\n",
        "#Specifying the training argument\n",
        "training_args = Seq2SeqTrainingArguments(\n",
        "    output_dir=model_path,\n",
        "    per_device_train_batch_size=2,\n",
        "    overwrite_output_dir = True,\n",
        "    evaluation_strategy=\"no\",\n",
        "    gradient_accumulation_steps=8,\n",
        "    num_train_epochs=30,\n",
        "    weight_decay=0.01,\n",
        "    lr_scheduler_type=\"cosine\",\n",
        "    learning_rate=5e-4,\n",
        "    fp16=True\n",
        ")"
      ],
      "metadata": {
        "id": "95wIk1YxNOTb"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Overwrite the Trainer API for utilizing custom loss function\n",
        "import torch\n",
        "import torch.nn.functional as F\n",
        "class CustomSeq2SeqTrainer(Seq2SeqTrainer):\n",
        "  def compute_loss(self, model, inputs, return_outputs=False):\n",
        "      labels = inputs.pop(\"labels\")\n",
        "      input_ids = inputs.pop(\"input_ids\")\n",
        "      attention_mask = inputs.pop(\"attention_mask\")\n",
        "      outputs = model(input_ids = input_ids, labels = labels, attention_mask = attention_mask)\n",
        "      loss_cross_entropy = outputs.loss\n",
        "      logits = outputs.logits\n",
        "      loss_GPT = custom_loss_function(logits)\n",
        "      loss = (loss_cross_entropy + torch.tensor(loss_GPT))\n",
        "      return (loss, outputs) if return_outputs else loss\n",
        "\n",
        "# Check whether the tokenizer is being passed as an argument\n",
        "def custom_loss_function(logits, tokenizer=base_tokenizer):\n",
        "    loss_GPT = 10\n",
        "    summaries = base_tokenizer.batch_decode(F.softmax(logits, dim=-1).argmax(dim=-1), skip_special_tokens=True)\n",
        "    for summary in summaries:\n",
        "      loss_GPT = loss_GPT + (10 - int(get_chatgpt_rating(prompt4, summary)))\n",
        "    return loss_GPT\n"
      ],
      "metadata": {
        "id": "iSFkn4YDOXl1"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Initializing the trainer\n",
        "\n",
        "trainer = Seq2SeqTrainer(\n",
        "    model=base_model,                         # the instantiated  Transformers model to be trained\n",
        "    args=training_args,                  # training arguments, defined above\n",
        "    data_collator=data_collator,\n",
        "    train_dataset=train_df,       # training dataset\n",
        "    eval_dataset = test_df,\n",
        ")"
      ],
      "metadata": {
        "id": "DQsfkoziNn2O"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#Custom Trainer function using Customized Loss function\n",
        "\n",
        "trainer = CustomSeq2SeqTrainer(\n",
        "    model=base_model,                         # the instantiated  Transformers model to be trained\n",
        "    args=training_args,                  # training arguments, defined above\n",
        "    data_collator=data_collator,\n",
        "    train_dataset=train_df,       # training dataset\n",
        "    eval_dataset = test_df,\n",
        ")"
      ],
      "metadata": {
        "id": "Mhswv40DCJef"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Model Training"
      ],
      "metadata": {
        "id": "gpPf0iO7QiYn"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "# Starting the training\n",
        "trainer.train()"
      ],
      "metadata": {
        "id": "RXKGeFzINqA5"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# Saving the final model\n",
        "trainer.save_model()"
      ],
      "metadata": {
        "id": "u4ugp2eQNs_u"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Model Evaluation (Perplexity)"
      ],
      "metadata": {
        "id": "ZsCbDbfbwSBd"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "eval_results = trainer.evaluate()"
      ],
      "metadata": {
        "id": "AM3GvSURwP3-"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "print(f\"Perplexity: {math.exp(eval_results['eval_loss']):.2f}\")"
      ],
      "metadata": {
        "id": "2vRsQ7NKsVN7"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Loading and Testing of the FT model"
      ],
      "metadata": {
        "id": "sdHZ53sg35wB"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "from transformers import AutoTokenizer, AutoModelForSeq2SeqLM\n",
        "\"\"\"\n",
        "Change model path to\n",
        "1) /content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/Basic - when using no PEFT methods\n",
        "2) /content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/Adapters - when using Adapters\n",
        "3) /content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/LoRA - when using LoRA -----> Have to save the PEFT file on Hugging Face and then load it if want to perform inference\n",
        "\"\"\"\n",
        "model_path = f'/content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/Basic'\n",
        "\n",
        "base_tokenizer = AutoTokenizer.from_pretrained(\"t5-small\")\n",
        "base_model = AutoModelForSeq2SeqLM.from_pretrained(model_path)"
      ],
      "metadata": {
        "id": "GdiHx03234_l",
        "colab": {
          "base_uri": "https://localhost:8080/"
        },
        "outputId": "837ea88f-9e82-4879-f4ab-e30ae5f3b620"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "stream",
          "name": "stderr",
          "text": [
            "Some weights of the model checkpoint at /content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/Adapters were not used when initializing T5ForConditionalGeneration: ['encoder.block.5.layer.0.SelfAttention.o.1.weight', 'decoder.block.0.layer.0.SelfAttention.o.0.weight', 'decoder.block.3.layer.0.SelfAttention.o.0.weight', 'encoder.block.4.layer.0.SelfAttention.o.1.bias', 'decoder.block.0.layer.0.SelfAttention.o.3.bias', 'decoder.block.4.layer.0.SelfAttention.o.3.bias', 'encoder.block.2.layer.0.SelfAttention.o.1.bias', 'decoder.block.4.layer.0.SelfAttention.o.0.weight', 'decoder.block.2.layer.0.SelfAttention.o.1.weight', 'decoder.block.3.layer.0.SelfAttention.o.1.bias', 'encoder.block.5.layer.0.SelfAttention.o.3.weight', 'decoder.block.1.layer.0.SelfAttention.o.1.bias', 'decoder.block.3.layer.0.SelfAttention.o.3.weight', 'encoder.block.5.layer.0.SelfAttention.o.0.weight', 'decoder.block.0.layer.0.SelfAttention.o.3.weight', 'decoder.block.4.layer.0.SelfAttention.o.3.weight', 'decoder.block.1.layer.0.SelfAttention.o.3.weight', 'encoder.block.4.layer.0.SelfAttention.o.3.weight', 'encoder.block.3.layer.0.SelfAttention.o.3.weight', 'encoder.block.4.layer.0.SelfAttention.o.3.bias', 'encoder.block.3.layer.0.SelfAttention.o.0.weight', 'encoder.block.0.layer.0.SelfAttention.o.3.bias', 'decoder.block.1.layer.0.SelfAttention.o.3.bias', 'decoder.block.5.layer.0.SelfAttention.o.1.bias', 'encoder.block.0.layer.0.SelfAttention.o.1.weight', 'encoder.block.1.layer.0.SelfAttention.o.3.weight', 'encoder.block.0.layer.0.SelfAttention.o.1.bias', 'decoder.block.2.layer.0.SelfAttention.o.0.weight', 'encoder.block.2.layer.0.SelfAttention.o.3.bias', 'encoder.block.5.layer.0.SelfAttention.o.3.bias', 'decoder.block.0.layer.0.SelfAttention.o.1.bias', 'decoder.block.5.layer.0.SelfAttention.o.0.weight', 'encoder.block.3.layer.0.SelfAttention.o.1.weight', 'decoder.block.2.layer.0.SelfAttention.o.3.weight', 'decoder.block.5.layer.0.SelfAttention.o.3.bias', 'decoder.block.1.layer.0.SelfAttention.o.1.weight', 'encoder.block.2.layer.0.SelfAttention.o.3.weight', 'encoder.block.0.layer.0.SelfAttention.o.0.weight', 'encoder.block.3.layer.0.SelfAttention.o.1.bias', 'decoder.block.2.layer.0.SelfAttention.o.1.bias', 'encoder.block.4.layer.0.SelfAttention.o.1.weight', 'decoder.block.4.layer.0.SelfAttention.o.1.weight', 'encoder.block.2.layer.0.SelfAttention.o.0.weight', 'decoder.block.5.layer.0.SelfAttention.o.1.weight', 'encoder.block.0.layer.0.SelfAttention.o.3.weight', 'encoder.block.1.layer.0.SelfAttention.o.3.bias', 'encoder.block.1.layer.0.SelfAttention.o.1.bias', 'decoder.block.3.layer.0.SelfAttention.o.3.bias', 'decoder.block.5.layer.0.SelfAttention.o.3.weight', 'decoder.block.2.layer.0.SelfAttention.o.3.bias', 'encoder.block.4.layer.0.SelfAttention.o.0.weight', 'encoder.block.1.layer.0.SelfAttention.o.0.weight', 'decoder.block.4.layer.0.SelfAttention.o.1.bias', 'decoder.block.1.layer.0.SelfAttention.o.0.weight', 'decoder.block.3.layer.0.SelfAttention.o.1.weight', 'encoder.block.5.layer.0.SelfAttention.o.1.bias', 'decoder.block.0.layer.0.SelfAttention.o.1.weight', 'encoder.block.2.layer.0.SelfAttention.o.1.weight', 'encoder.block.1.layer.0.SelfAttention.o.1.weight', 'encoder.block.3.layer.0.SelfAttention.o.3.bias']\n",
            "- This IS expected if you are initializing T5ForConditionalGeneration from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n",
            "- This IS NOT expected if you are initializing T5ForConditionalGeneration from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n",
            "Some weights of T5ForConditionalGeneration were not initialized from the model checkpoint at /content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/Adapters and are newly initialized: ['decoder.block.0.layer.0.SelfAttention.o.weight', 'decoder.block.2.layer.0.SelfAttention.o.weight', 'decoder.block.3.layer.0.SelfAttention.o.weight', 'decoder.block.1.layer.0.SelfAttention.o.weight', 'decoder.block.4.layer.0.SelfAttention.o.weight', 'encoder.block.0.layer.0.SelfAttention.o.weight', 'encoder.block.4.layer.0.SelfAttention.o.weight', 'decoder.block.5.layer.0.SelfAttention.o.weight', 'encoder.block.2.layer.0.SelfAttention.o.weight', 'encoder.block.1.layer.0.SelfAttention.o.weight', 'encoder.block.3.layer.0.SelfAttention.o.weight', 'encoder.block.5.layer.0.SelfAttention.o.weight']\n",
            "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n"
          ]
        }
      ]
    },
    {
      "cell_type": "code",
      "source": [
        "\"\"\"\n",
        "Format of the input\n",
        "Title: ADD_TITLE. Genre: ADD_GENRES. Characters: ADD_CHARS. Relations: Neutral: <>. Positive: <>. Negative: <>.  Plot: ADD_PLOT\n",
        "\"\"\"\n",
        "\n",
        "input = \"Title: Saiyan Warrior. Genre: Action, Adventure, Drama. Characters: Goku, Vegeta, Beerus. Relations: Neutral: <Beerus, Vegeta>. Positive: <Goku, Vegeta>. Negative: <Goku, Beerus>. Plot: Beerus is a god of destruction, who came to Earth to destroy it. Goku and Vegeta in order to protect their family fight and defeat him.\"\n",
        "input = base_tokenizer(input, return_tensors=\"pt\").input_ids\n",
        "\n",
        "#For different decoding strategies\n",
        "do_sample = False\n",
        "num_beams = 1\n",
        "top_p = 0.9\n",
        "\n",
        "output = base_model.generate(input, num_beams=num_beams, do_sample=do_sample, top_p = top_p, min_length=200, max_length=250)\n",
        "print(base_tokenizer.batch_decode(output, skip_special_tokens=True)[0])"
      ],
      "metadata": {
        "id": "eFrtEoUOPCr-"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Model Eval (BLEU)"
      ],
      "metadata": {
        "id": "ooAGQEE-EBn0"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "!pip install evaluate"
      ],
      "metadata": {
        "id": "oQ5o39Y0DcXR"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "from evaluate import load\n",
        "#Loading the BLEU score metric\n",
        "bleu = load(\"bleu\")"
      ],
      "metadata": {
        "id": "hGayRQ7_Dind"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "model = AutoModelForSeq2SeqLM.from_pretrained('/content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/Basic')\n",
        "tokenizer = AutoTokenizer.from_pretrained('t5-small')"
      ],
      "metadata": {
        "id": "y6vjP1z4HnC0"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def encode_input(x, base_tokenizer = tokenizer):\n",
        "  x = str(x)\n",
        "  return base_tokenizer(x, max_length = 512, truncation=True, padding='max_length')"
      ],
      "metadata": {
        "id": "Tf9FrB-UH7qd"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "import pandas as pd\n",
        "df = pd.read_csv('/content/drive/MyDrive/Visual Story Telling/Dataset - Story Generation/Plot_Summary_Dataset')\n",
        "train_df, test_df = train_test_split(df, test_size=0.1. random_state=1)\n",
        "test_df['Input'] = test_df['Input'].apply(encode_input)"
      ],
      "metadata": {
        "id": "ZwqKW4CQHOwM"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "def apply_len(x):\n",
        "    return len(x)\n",
        "test_df['len'] = test_df['Summary'].apply(apply_len)"
      ],
      "metadata": {
        "id": "EQCQziBaNvn0"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "test_df = test_df[test_df['len']<350]\n",
        "test_df.reset_index(drop=True, inplace=True)"
      ],
      "metadata": {
        "id": "iXChy4h_Nyg2"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "num_beams = 1\n",
        "top_p = 1.0\n",
        "do_sample = False\n",
        "#Greedy, Ancestral and Nucleus\n",
        "decoding_strats = [[1,1.0, False],[1,1.0, True],[1,0.9, True]]"
      ],
      "metadata": {
        "id": "iAbAIxPMJLQ5"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#BLEU score eval of FT model\n",
        "results = []\n",
        "for x in decoding_strats:\n",
        "  predictions = []\n",
        "  references = []\n",
        "  num_beams, top_p, do_sample = x\n",
        "  for idx in test_df.iterrows():\n",
        "    references.append(test_df['Summary'].iloc[idx])\n",
        "    predictions.append(tokenizer.batch_decode(model.generate(test_df['Input'].iloc[idx], num_beams = num_beams, top_p = top_p, do_sample = do_sample, max_length = len(test_df['Summary'].iloc[idx])), skip_special_tokens=True))\n",
        "  base_results = bleu.compute(predictions=predictions, references=references)\n",
        "  results.append(base_results)\n",
        "print(results)"
      ],
      "metadata": {
        "id": "pfvKq0aYJN1F"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## Loss curve"
      ],
      "metadata": {
        "id": "hfwwzwOOJWnN"
      }
    },
    {
      "cell_type": "code",
      "source": [
        "import matplotlib.pyplot as plt\n",
        "%matplotlib inline"
      ],
      "metadata": {
        "id": "baSelc20JYB8"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "import pandas as pd"
      ],
      "metadata": {
        "id": "w2pWkpbXGPI-"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "# List of different configurations of T5 model\n",
        "models = ['Basic','Adapters','LoRA']"
      ],
      "metadata": {
        "id": "kXeK8boEJunm"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "#Abstraction of the loss and epochs from the log history\n",
        "def get_epoch(x):\n",
        "  return x['epoch']\n",
        "\n",
        "def get_loss(x):\n",
        "  return x['loss']"
      ],
      "metadata": {
        "id": "B94hciBHJwWd"
      },
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "code",
      "source": [
        "for i in models:\n",
        "  #Loading the trainer_state json file, which consists of the losses at different steps/epochs\n",
        "  df = pd.read_json(f'/content/drive/MyDrive/Visual Story Telling/Story_Gen_Model/{i}/checkpoint-8000/trainer_state.json')\n",
        "  #Organize in a way to extract losses and steps\n",
        "  df['epoch_number'] = df['log_history'].apply(get_epoch)\n",
        "  df ['loss'] = df['log_history'].apply(get_loss)\n",
        "  #Plotting of the training loss curve for the FT Model\n",
        "  plt.plot(df['epoch_number'].values, df['loss'].values, label = f\"{i} T5\")\n",
        "plt.xlabel('#Epochs')\n",
        "plt.ylabel('Loss value')\n",
        "plt.title('Training curve T5 models')\n",
        "plt.legend()\n",
        "plt.show()"
      ],
      "metadata": {
        "id": "ou9bZY2CJzbp",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 472
        },
        "outputId": "61e35de3-044e-45eb-a38a-98286769b71d"
      },
      "execution_count": null,
      "outputs": [
        {
          "output_type": "display_data",
          "data": {
            "text/plain": [
              "<Figure size 640x480 with 1 Axes>"
            ],
            "image/png": "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\n"
          },
          "metadata": {}
        }
      ]
    }
  ]
}