Upload 2 files

モデルシェア＆マージ(V4).ipynb

@@ -0,0 +1,453 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "52e2911f-c3e0-46f0-b574-a356f5825a23",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "<bound method Module.modules of GPTNeoXForCausalLM(\n",
+      "  (gpt_neox): GPTNeoXModel(\n",
+      "    (embed_in): Embedding(51200, 3072)\n",
+      "    (emb_dropout): Dropout(p=0.0, inplace=False)\n",
+      "    (layers): ModuleList(\n",
+      "      (0-29): 30 x GPTNeoXLayer(\n",
+      "        (input_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "        (post_attention_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "        (post_attention_dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (post_mlp_dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (attention): GPTNeoXAttention(\n",
+      "          (rotary_emb): GPTNeoXRotaryEmbedding()\n",
+      "          (query_key_value): Linear(in_features=3072, out_features=9216, bias=True)\n",
+      "          (dense): Linear(in_features=3072, out_features=3072, bias=True)\n",
+      "          (attention_dropout): Dropout(p=0.0, inplace=False)\n",
+      "        )\n",
+      "        (mlp): GPTNeoXMLP(\n",
+      "          (dense_h_to_4h): Linear(in_features=3072, out_features=12288, bias=True)\n",
+      "          (dense_4h_to_h): Linear(in_features=12288, out_features=3072, bias=True)\n",
+      "          (act): GELUActivation()\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (final_layer_norm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "  )\n",
+      "  (embed_out): Linear(in_features=3072, out_features=51200, bias=False)\n",
+      ")>\n"
+     ]
+    }
+   ],
+   "source": [
+    "# 利用するGPUを制限\n",
+    "import os\n",
+    "import torch\n",
+    "from peft import PeftModel, PeftConfig\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
+    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"1\" # これは0, 1, 2のGPUから、2（3番目のGPU）だけ使うという宣言。\n",
+    "\n",
+    "# メインモデル\n",
+    "MODEL_NAME = 'line-corporation/japanese-large-lm-3.6b'\n",
+    "# モデルの準備\n",
+    "model = AutoModelForCausalLM.from_pretrained(\n",
+    "    MODEL_NAME,\n",
+    "    load_in_8bit=False, # 注意：8bitで読み込むと結合不可能\n",
+    "    device_map=\"auto\",\n",
+    ")\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=False, padding_size=\"left\", legacy=True)\n",
+    "tokenizer.do_lower_case = True  \n",
+    "\n",
+    "print(model.modules)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "0d93f697-9bf1-4214-ae31-47451edb3e7e",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "<bound method Module.modules of PeftModelForCausalLM(\n",
+      "  (base_model): LoraModel(\n",
+      "    (model): GPTNeoXForCausalLM(\n",
+      "      (gpt_neox): GPTNeoXModel(\n",
+      "        (embed_in): Embedding(51200, 3072)\n",
+      "        (emb_dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (layers): ModuleList(\n",
+      "          (0-29): 30 x GPTNeoXLayer(\n",
+      "            (input_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "            (post_attention_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "            (post_attention_dropout): Dropout(p=0.0, inplace=False)\n",
+      "            (post_mlp_dropout): Dropout(p=0.0, inplace=False)\n",
+      "            (attention): GPTNeoXAttention(\n",
+      "              (rotary_emb): GPTNeoXRotaryEmbedding()\n",
+      "              (query_key_value): Linear(\n",
+      "                in_features=3072, out_features=9216, bias=True\n",
+      "                (lora_dropout): ModuleDict(\n",
+      "                  (default): Dropout(p=0.05, inplace=False)\n",
+      "                )\n",
+      "                (lora_A): ModuleDict(\n",
+      "                  (default): Linear(in_features=3072, out_features=128, bias=False)\n",
+      "                )\n",
+      "                (lora_B): ModuleDict(\n",
+      "                  (default): Linear(in_features=128, out_features=9216, bias=False)\n",
+      "                )\n",
+      "                (lora_embedding_A): ParameterDict()\n",
+      "                (lora_embedding_B): ParameterDict()\n",
+      "              )\n",
+      "              (dense): Linear(\n",
+      "                in_features=3072, out_features=3072, bias=True\n",
+      "                (lora_dropout): ModuleDict(\n",
+      "                  (default): Dropout(p=0.05, inplace=False)\n",
+      "                )\n",
+      "                (lora_A): ModuleDict(\n",
+      "                  (default): Linear(in_features=3072, out_features=128, bias=False)\n",
+      "                )\n",
+      "                (lora_B): ModuleDict(\n",
+      "                  (default): Linear(in_features=128, out_features=3072, bias=False)\n",
+      "                )\n",
+      "                (lora_embedding_A): ParameterDict()\n",
+      "                (lora_embedding_B): ParameterDict()\n",
+      "              )\n",
+      "              (attention_dropout): Dropout(p=0.0, inplace=False)\n",
+      "            )\n",
+      "            (mlp): GPTNeoXMLP(\n",
+      "              (dense_h_to_4h): Linear(in_features=3072, out_features=12288, bias=True)\n",
+      "              (dense_4h_to_h): Linear(in_features=12288, out_features=3072, bias=True)\n",
+      "              (act): GELUActivation()\n",
+      "            )\n",
+      "          )\n",
+      "        )\n",
+      "        (final_layer_norm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "      )\n",
+      "      (embed_out): Linear(in_features=3072, out_features=51200, bias=False)\n",
+      "    )\n",
+      "  )\n",
+      ")>\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Loraモジュールが入っているフォルダ\n",
+    "lora_name = \"lora-rinna-test_sample_V4\"\n",
+    "checkpoint = 500\n",
+    "\n",
+    "# LoRAモデルの準備\n",
+    "lora_model = PeftModel.from_pretrained(\n",
+    "    model, \n",
+    "    f'{lora_name}/checkpoint-{checkpoint}', \n",
+    "    device_map=\"auto\"\n",
+    ")\n",
+    "\n",
+    "print(lora_model.modules)\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "408c34f4-223d-4874-a610-94b41ddaa41e",
+   "metadata": {},
+   "source": [
+    "# 関数の読み込み"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "5eaf491b-da91-409f-87c2-7f731c569e70",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "['修正前:あなたは今どこにいますか? 修正後:あなたは今どこにいますか?']"
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# プロンプト変換\n",
+    "def prompt_infer(original):\n",
+    "  '''\n",
+    "  original: 修正前の文章\n",
+    "  revised: 修正後の文章\n",
+    "\n",
+    "  ---\n",
+    "  prompt_start: コンテクスト（前置き）\n",
+    "  original_token: 修正前の文章ののトークン、例：修正前\n",
+    "  revised_token: 修正後の文章のトークン、例：修正後\n",
+    "  '''\n",
+    "  prompt_start = '修正前の文章の誤字、脱字などを含むスペルチェックを行ってください。問題がない場合、そのままの文章を出力してください。'\n",
+    "  #original_token = '修正前'\n",
+    "  #revised_token = '修正後'\n",
+    "  return f'{prompt_start}修正前：{original} 修正後：'\n",
+    "prompt_infer('今日はいい転居ですね！')\n",
+    "\n",
+    "\n",
+    "import re\n",
+    "# 出力表示の簡略化\n",
+    "def converts(texts):\n",
+    "    return [re.findall('(修正前:.*)', text)[0]for text in texts]\n",
+    "\n",
+    "converts(['修正前の文章の誤字、脱字などを含むスペルチェックを行ってください。問題がない場合、そのままの文章を出力してください。修正前:あなたは今どこにいますか? 修正後:あなたは今どこにいますか?'])\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "id": "f8b6b4e1-3bb7-42ff-a102-5a3434b212e0",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "['修正前の文章の誤字、脱字などを含むスペルチェックを行ってください。問題がない場合、そのままの文章を出力してください。修正前：空も青く、今日はほとにいい天気ですね。 修正後：',\n",
+       " '修正前の文章の誤字、脱字などを含むスペルチェックを行ってください。問題がない場合、そのままの文章を出力してください。修正前：いつれ、私は仙台に戻ります。 修正後：',\n",
+       " '修正前の文章の誤字、脱字などを含むスペルチェックを行ってください。問題がない場合、そのままの文章を出力してください。修正前：私はとっくに修士課程を終了しています。 修正後：']"
+      ]
+     },
+     "execution_count": 5,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "texts = ['空も青く、今日はほとにいい天気ですね。', 'いつれ、私は仙台に戻ります。','私はとっくに修士課程を終了しています。']\n",
+    "\n",
+    "ctexts = [prompt_infer(t) for t in texts]\n",
+    "ctexts"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 7,
+   "id": "c39ba5eb-d2ad-408b-8b7f-8f7e2fa1caf5",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "修正前:空も青く、今日はほとにいい天気ですね。 修正後:空も青く、今日はとてもよい天気ですね。\n",
+      "修正前:いつれ、私は仙台に戻ります。 修正後:いつれ、私は仙台に戻ります。\n",
+      "修正前:私はとっくに修士課程を終了しています。 修正後:私はとっくに修士課程を終了しています。\n"
+     ]
+    }
+   ],
+   "source": [
+    "\n",
+    "# 評価モード\n",
+    "lora_model.eval()\n",
+    "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=False)\n",
+    "#ファインチューニングしたモデルをbest_modelと名付け、再度生成。モデル名がbest_modelに代えていることに注意。太宰的な文章が生成されている。\n",
+    "#best_model = trainer.model.cpu()\n",
+    "\n",
+    "for text in texts:\n",
+    "    prompt = prompt_infer(text)\n",
+    "    prompt_ids = tokenizer(prompt, return_tensors = 'pt',add_special_tokens=False).input_ids\n",
+    "    outputs = lora_model.generate(\n",
+    "        input_ids=prompt_ids.to(lora_model.device),\n",
+    "        #prompt_ids,\n",
+    "        do_sample=False,\n",
+    "        max_length=60,\n",
+    "        num_return_sequences=1,\n",
+    "        pad_token_id=tokenizer.eos_token_id\n",
+    "    )\n",
+    "    #tokenizer.batch_decode(outputs, skip_special_tokens = True)\n",
+    "    print('\\n'.join(converts(tokenizer.batch_decode(outputs, skip_special_tokens = True))))\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "e458339b-43bc-4c61-89d8-97e973c0ba3f",
+   "metadata": {},
+   "source": [
+    "# モデル結合\n",
+    "\n",
+    "loraパラメーターをもとのニューラルネットワークに加算し、結合する。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "a7415bdf-0af0-4596-bc44-4aed4b06e9d1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "lora_model2 = lora_model.merge_and_unload()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "5d627d97-dc39-4674-8881-be06ad0cee9a",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<bound method Module.modules of GPTNeoXForCausalLM(\n",
+       "  (gpt_neox): GPTNeoXModel(\n",
+       "    (embed_in): Embedding(51200, 3072)\n",
+       "    (emb_dropout): Dropout(p=0.0, inplace=False)\n",
+       "    (layers): ModuleList(\n",
+       "      (0-29): 30 x GPTNeoXLayer(\n",
+       "        (input_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+       "        (post_attention_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+       "        (post_attention_dropout): Dropout(p=0.0, inplace=False)\n",
+       "        (post_mlp_dropout): Dropout(p=0.0, inplace=False)\n",
+       "        (attention): GPTNeoXAttention(\n",
+       "          (rotary_emb): GPTNeoXRotaryEmbedding()\n",
+       "          (query_key_value): Linear(in_features=3072, out_features=9216, bias=True)\n",
+       "          (dense): Linear(in_features=3072, out_features=3072, bias=True)\n",
+       "          (attention_dropout): Dropout(p=0.0, inplace=False)\n",
+       "        )\n",
+       "        (mlp): GPTNeoXMLP(\n",
+       "          (dense_h_to_4h): Linear(in_features=3072, out_features=12288, bias=True)\n",
+       "          (dense_4h_to_h): Linear(in_features=12288, out_features=3072, bias=True)\n",
+       "          (act): GELUActivation()\n",
+       "        )\n",
+       "      )\n",
+       "    )\n",
+       "    (final_layer_norm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+       "  )\n",
+       "  (embed_out): Linear(in_features=3072, out_features=51200, bias=False)\n",
+       ")>"
+      ]
+     },
+     "execution_count": 9,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "lora_model2.modules"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "id": "b588088b-1a6b-43d9-b023-3b6193291117",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 保存\n",
+    "\n",
+    "lora_model2.save_pretrained('merged/spellcheckGPT')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d22086c4-7b31-4573-a6f7-1c52c80ccb3d",
+   "metadata": {},
+   "source": [
+    "# 次から"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "faf07b2e-c374-466a-a5f5-fc417820cee6",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "732b45c8e9d049dc93ddf7bccc87bc41",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "import os\n",
+    "import torch\n",
+    "from peft import PeftModel, PeftConfig\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
+    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"1\" # これは0, 1, 2のGPUから、2（3番目のGPU）だけ使うという宣言。\n",
+    "\n",
+    "my_model = 'merged/spellcheckGPT'\n",
+    "spellcheckGPT = AutoModelForCausalLM.from_pretrained(my_model,\n",
+    "                                     load_in_8bit=True,\n",
+    "                                     device_map=\"auto\",)\n",
+    "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=False, padding_size=\"left\", legacy=True)\n",
+    "tokenizer.do_lower_case = True  "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "id": "32045d37-9a40-47e6-860e-0c521c9ae3b5",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "修正前:空も青く、今日はほとにいい天気ですね。 修正後:空も青く、今日はとても天気がいいですね。\n",
+      "修正前:いつれ、私は仙台に戻ります。 修正後:いつれ、私は仙台に戻ります。\n",
+      "修正前:私はとっくに修士課程を終了しています。 修正後:私はとっくに修士課程を終了しています。\n"
+     ]
+    }
+   ],
+   "source": [
+    "\n",
+    "# 評価モード\n",
+    "spellcheckGPT.eval()\n",
+    "\n",
+    "#ファインチューニングしたモデルをbest_modelと名付け、再度生成。モデル名がbest_modelに代えていることに注意。太宰的な文章が生成されている。\n",
+    "#best_model = trainer.model.cpu()\n",
+    "\n",
+    "for text in texts:\n",
+    "    prompt = prompt_infer(text)\n",
+    "    prompt_ids = tokenizer(prompt, return_tensors = 'pt',add_special_tokens=False).input_ids\n",
+    "    outputs = spellcheckGPT.generate(\n",
+    "        input_ids=prompt_ids.to(lora_model.device),\n",
+    "        #prompt_ids,\n",
+    "        do_sample=False,\n",
+    "        max_length=60,\n",
+    "        num_return_sequences=1,\n",
+    "        pad_token_id=tokenizer.eos_token_id\n",
+    "    )\n",
+    "    #tokenizer.batch_decode(outputs, skip_special_tokens = True)\n",
+    "    print('\\n'.join(converts(tokenizer.batch_decode(outputs, skip_special_tokens = True))))\n"
+   ]
+  }
+ ],
+ "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": "3.11.5"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

モデルシェア＆マージ(V5).ipynb

@@ -0,0 +1,470 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "52e2911f-c3e0-46f0-b574-a356f5825a23",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "<bound method Module.modules of GPTNeoXForCausalLM(\n",
+      "  (gpt_neox): GPTNeoXModel(\n",
+      "    (embed_in): Embedding(51200, 3072)\n",
+      "    (emb_dropout): Dropout(p=0.0, inplace=False)\n",
+      "    (layers): ModuleList(\n",
+      "      (0-29): 30 x GPTNeoXLayer(\n",
+      "        (input_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "        (post_attention_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "        (post_attention_dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (post_mlp_dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (attention): GPTNeoXAttention(\n",
+      "          (rotary_emb): GPTNeoXRotaryEmbedding()\n",
+      "          (query_key_value): Linear(in_features=3072, out_features=9216, bias=True)\n",
+      "          (dense): Linear(in_features=3072, out_features=3072, bias=True)\n",
+      "          (attention_dropout): Dropout(p=0.0, inplace=False)\n",
+      "        )\n",
+      "        (mlp): GPTNeoXMLP(\n",
+      "          (dense_h_to_4h): Linear(in_features=3072, out_features=12288, bias=True)\n",
+      "          (dense_4h_to_h): Linear(in_features=12288, out_features=3072, bias=True)\n",
+      "          (act): GELUActivation()\n",
+      "        )\n",
+      "      )\n",
+      "    )\n",
+      "    (final_layer_norm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "  )\n",
+      "  (embed_out): Linear(in_features=3072, out_features=51200, bias=False)\n",
+      ")>\n"
+     ]
+    }
+   ],
+   "source": [
+    "# 利用するGPUを制限\n",
+    "import os\n",
+    "import torch\n",
+    "from peft import PeftModel, PeftConfig\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
+    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"0\" # これは0, 1, 2のGPUから、2（3番目のGPU）だけ使うという宣言。\n",
+    "\n",
+    "# メインモデル\n",
+    "MODEL_NAME = 'line-corporation/japanese-large-lm-3.6b'\n",
+    "# モデルの準備\n",
+    "model = AutoModelForCausalLM.from_pretrained(\n",
+    "    MODEL_NAME,\n",
+    "    load_in_8bit=False, # 注意：8bitで読み込むと結合不可能\n",
+    "    device_map=\"auto\",\n",
+    ")\n",
+    "\n",
+    "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=False, padding_size=\"left\", legacy=True)\n",
+    "tokenizer.do_lower_case = True  \n",
+    "\n",
+    "print(model.modules)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "25768347-129c-4570-b9a1-7dff10422049",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import os\n",
+    "os.chdir('/data/shared_project/GPT')"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "0d93f697-9bf1-4214-ae31-47451edb3e7e",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "<bound method Module.modules of PeftModelForCausalLM(\n",
+      "  (base_model): LoraModel(\n",
+      "    (model): GPTNeoXForCausalLM(\n",
+      "      (gpt_neox): GPTNeoXModel(\n",
+      "        (embed_in): Embedding(51200, 3072)\n",
+      "        (emb_dropout): Dropout(p=0.0, inplace=False)\n",
+      "        (layers): ModuleList(\n",
+      "          (0-29): 30 x GPTNeoXLayer(\n",
+      "            (input_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "            (post_attention_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "            (post_attention_dropout): Dropout(p=0.0, inplace=False)\n",
+      "            (post_mlp_dropout): Dropout(p=0.0, inplace=False)\n",
+      "            (attention): GPTNeoXAttention(\n",
+      "              (rotary_emb): GPTNeoXRotaryEmbedding()\n",
+      "              (query_key_value): Linear(\n",
+      "                in_features=3072, out_features=9216, bias=True\n",
+      "                (lora_dropout): ModuleDict(\n",
+      "                  (default): Dropout(p=0.05, inplace=False)\n",
+      "                )\n",
+      "                (lora_A): ModuleDict(\n",
+      "                  (default): Linear(in_features=3072, out_features=24, bias=False)\n",
+      "                )\n",
+      "                (lora_B): ModuleDict(\n",
+      "                  (default): Linear(in_features=24, out_features=9216, bias=False)\n",
+      "                )\n",
+      "                (lora_embedding_A): ParameterDict()\n",
+      "                (lora_embedding_B): ParameterDict()\n",
+      "              )\n",
+      "              (dense): Linear(\n",
+      "                in_features=3072, out_features=3072, bias=True\n",
+      "                (lora_dropout): ModuleDict(\n",
+      "                  (default): Dropout(p=0.05, inplace=False)\n",
+      "                )\n",
+      "                (lora_A): ModuleDict(\n",
+      "                  (default): Linear(in_features=3072, out_features=24, bias=False)\n",
+      "                )\n",
+      "                (lora_B): ModuleDict(\n",
+      "                  (default): Linear(in_features=24, out_features=3072, bias=False)\n",
+      "                )\n",
+      "                (lora_embedding_A): ParameterDict()\n",
+      "                (lora_embedding_B): ParameterDict()\n",
+      "              )\n",
+      "              (attention_dropout): Dropout(p=0.0, inplace=False)\n",
+      "            )\n",
+      "            (mlp): GPTNeoXMLP(\n",
+      "              (dense_h_to_4h): Linear(in_features=3072, out_features=12288, bias=True)\n",
+      "              (dense_4h_to_h): Linear(in_features=12288, out_features=3072, bias=True)\n",
+      "              (act): GELUActivation()\n",
+      "            )\n",
+      "          )\n",
+      "        )\n",
+      "        (final_layer_norm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+      "      )\n",
+      "      (embed_out): Linear(in_features=3072, out_features=51200, bias=False)\n",
+      "    )\n",
+      "  )\n",
+      ")>\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Loraモジュールが入っているフォルダ\n",
+    "lora_name = \"lora-rinna-test_sample_V5\"\n",
+    "checkpoint = 300\n",
+    "\n",
+    "# LoRAモデルの準備\n",
+    "lora_model = PeftModel.from_pretrained(\n",
+    "    model, \n",
+    "    f'{lora_name}/checkpoint-{checkpoint}', \n",
+    "    device_map=\"auto\"\n",
+    ")\n",
+    "\n",
+    "print(lora_model.modules)\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "408c34f4-223d-4874-a610-94b41ddaa41e",
+   "metadata": {},
+   "source": [
+    "# 関数の読み込み"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "5eaf491b-da91-409f-87c2-7f731c569e70",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "本文の文章の誤字、脱字などを含むスペルチェックを行ってください、「チェック：」の後に出力してください。例：修正前：仮名の「ん」で表される音は、後続の音によって などとにる。 チェック：\"にる\"->\"なる\", 。修正前：今日はいい転居ですね！ チェック：\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "['修正前:あなたは今どこにいますか? 修正後:あなたは今どこにいますか?']"
+      ]
+     },
+     "execution_count": 9,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# プロンプト変換\n",
+    "def prompt_infer(original):\n",
+    "  '''\n",
+    "  original: 修正前の文章\n",
+    "\n",
+    "  ---\n",
+    "  prompt_start: コンテクスト（前置き）\n",
+    "  original_token: 修正前の文章ののトークン、例：修正前\n",
+    "  revised_token: 修正後の文章のトークン、例：修正後\n",
+    "  '''\n",
+    "  prompt_start = '本文の文章の誤字、脱字などを含むスペルチェックを行ってください、「チェック：」の後に出力してください。例：修正前：仮名の「ん」で表される音は、後続の音によって などとにる。 チェック：\"にる\"->\"なる\", 。'\n",
+    "  #original_token = '修正前'\n",
+    "  #revised_token = '修正後'\n",
+    "  return f'{prompt_start}修正前：{original} チェック：'\n",
+    "print(prompt_infer('今日はいい転居ですね！'))\n",
+    "\n",
+    "\n",
+    "import re\n",
+    "# 出力表示の簡略化\n",
+    "def converts(texts):\n",
+    "    return [re.findall('(修正前:.*)', text)[0]for text in texts]\n",
+    "\n",
+    "converts(['修正前の文章の誤字、脱字などを含むスペルチェックを行ってください。問題がない場合、そのままの文章を出力してください。修正前:あなたは今どこにいますか? 修正後:あなたは今どこにいますか?'])\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "id": "f8b6b4e1-3bb7-42ff-a102-5a3434b212e0",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "['本文の文章の誤字、脱字などを含むスペルチェックを行ってください、「チェック：」の後に出力してください。例：修正前：仮名の「ん」で表される音は、後続の音によって などとにる。 チェック：\"にる\"->\"なる\", 。修正前：空も青く、今日はほとにいい天気ですね。 チェック：',\n",
+       " '本文の文章の誤字、脱字などを含むスペルチェックを行ってください、「チェック：」の後に出力してください。例：修正前：仮名の「ん」で表される音は、後続の音によって などと��る。 チェック：\"にる\"->\"なる\", 。修正前：いつれ、私は仙台に戻ります。 チェック：',\n",
+       " '本文の文章の誤字、脱字などを含むスペルチェックを行ってください、「チェック：」の後に出力してください。例：修正前：仮名の「ん」で表される音は、後続の音によって などとにる。 チェック：\"にる\"->\"なる\", 。修正前：私はとっくに修士課程を終了しています。 チェック：']"
+      ]
+     },
+     "execution_count": 10,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "texts = ['空も青く、今日はほとにいい天気ですね。', 'いつれ、私は仙台に戻ります。','私はとっくに修士課程を終了しています。']\n",
+    "\n",
+    "ctexts = [prompt_infer(t) for t in texts]\n",
+    "ctexts"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "c39ba5eb-d2ad-408b-8b7f-8f7e2fa1caf5",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "修正前:仮名の「ん」で表される音は、後続の音によって などとにる。 チェック:\"にる\"->\"なる\", 。修正前:空も青く、今日はほとにいい天気ですね。 チェック:\"ほとに\"->\"とても\",\n",
+      "修正前:仮名の「ん」で表される音は、後続の音によって などとにる。 チェック:\"にる\"->\"なる\", 。修正前:いつれ、私は仙台に戻ります。 チェック:\"いつれ\"->\"いずれ\",\n",
+      "修正前:仮名の「ん」で表される音は、後続の音によって などとにる。 チェック:\"にる\"->\"なる\", 。修正前:私はとっくに修士課程を終了しています。 チェック:\"を\"->\"の\",\n"
+     ]
+    }
+   ],
+   "source": [
+    "\n",
+    "# 評価モード\n",
+    "lora_model.eval()\n",
+    "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=False, legacy=True)\n",
+    "#ファインチューニングしたモデルをbest_modelと名付け、再度生成。モデル名がbest_modelに代えていることに注意。太宰的な文章が生成されている。\n",
+    "#best_model = trainer.model.cpu()\n",
+    "\n",
+    "for text in texts:\n",
+    "    prompt = prompt_infer(text)\n",
+    "    prompt_ids = tokenizer(prompt, return_tensors = 'pt',add_special_tokens=False).input_ids\n",
+    "    outputs = lora_model.generate(\n",
+    "        input_ids=prompt_ids.to(lora_model.device),\n",
+    "        #prompt_ids,\n",
+    "        do_sample=False,\n",
+    "        max_length=256,\n",
+    "        num_return_sequences=1,\n",
+    "        pad_token_id=tokenizer.eos_token_id\n",
+    "    )\n",
+    "    #tokenizer.batch_decode(outputs, skip_special_tokens = True)\n",
+    "    print('\\n'.join(converts(tokenizer.batch_decode(outputs, skip_special_tokens = True))))\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "e458339b-43bc-4c61-89d8-97e973c0ba3f",
+   "metadata": {},
+   "source": [
+    "# モデル結合\n",
+    "\n",
+    "loraパラメーターをもとのニューラルネットワークに加算し、結合する。"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "a7415bdf-0af0-4596-bc44-4aed4b06e9d1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "lora_model2 = lora_model.merge_and_unload()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "5d627d97-dc39-4674-8881-be06ad0cee9a",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<bound method Module.modules of GPTNeoXForCausalLM(\n",
+       "  (gpt_neox): GPTNeoXModel(\n",
+       "    (embed_in): Embedding(51200, 3072)\n",
+       "    (emb_dropout): Dropout(p=0.0, inplace=False)\n",
+       "    (layers): ModuleList(\n",
+       "      (0-29): 30 x GPTNeoXLayer(\n",
+       "        (input_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+       "        (post_attention_layernorm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+       "        (post_attention_dropout): Dropout(p=0.0, inplace=False)\n",
+       "        (post_mlp_dropout): Dropout(p=0.0, inplace=False)\n",
+       "        (attention): GPTNeoXAttention(\n",
+       "          (rotary_emb): GPTNeoXRotaryEmbedding()\n",
+       "          (query_key_value): Linear(in_features=3072, out_features=9216, bias=True)\n",
+       "          (dense): Linear(in_features=3072, out_features=3072, bias=True)\n",
+       "          (attention_dropout): Dropout(p=0.0, inplace=False)\n",
+       "        )\n",
+       "        (mlp): GPTNeoXMLP(\n",
+       "          (dense_h_to_4h): Linear(in_features=3072, out_features=12288, bias=True)\n",
+       "          (dense_4h_to_h): Linear(in_features=12288, out_features=3072, bias=True)\n",
+       "          (act): GELUActivation()\n",
+       "        )\n",
+       "      )\n",
+       "    )\n",
+       "    (final_layer_norm): LayerNorm((3072,), eps=1e-05, elementwise_affine=True)\n",
+       "  )\n",
+       "  (embed_out): Linear(in_features=3072, out_features=51200, bias=False)\n",
+       ")>"
+      ]
+     },
+     "execution_count": 9,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "lora_model2.modules"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 10,
+   "id": "b588088b-1a6b-43d9-b023-3b6193291117",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# 保存\n",
+    "\n",
+    "lora_model2.save_pretrained('merged/spellcheckGPT')"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d22086c4-7b31-4573-a6f7-1c52c80ccb3d",
+   "metadata": {},
+   "source": [
+    "# 次から"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "faf07b2e-c374-466a-a5f5-fc417820cee6",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "application/vnd.jupyter.widget-view+json": {
+       "model_id": "732b45c8e9d049dc93ddf7bccc87bc41",
+       "version_major": 2,
+       "version_minor": 0
+      },
+      "text/plain": [
+       "Loading checkpoint shards:   0%|          | 0/2 [00:00<?, ?it/s]"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "import os\n",
+    "import torch\n",
+    "from peft import PeftModel, PeftConfig\n",
+    "from transformers import AutoModelForCausalLM, AutoTokenizer\n",
+    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"1\" # これは0, 1, 2のGPUから、2（3番目のGPU）だけ使うという宣言。\n",
+    "\n",
+    "my_model = 'merged/spellcheckGPT'\n",
+    "spellcheckGPT = AutoModelForCausalLM.from_pretrained(my_model,\n",
+    "                                     load_in_8bit=True,\n",
+    "                                     device_map=\"auto\",)\n",
+    "tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=False, padding_size=\"left\", legacy=True)\n",
+    "tokenizer.do_lower_case = True  "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "id": "32045d37-9a40-47e6-860e-0c521c9ae3b5",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "修正前:空も青く、今日はほとにいい天気ですね。 修正後:空も青く、今日はとても天気がいいですね。\n",
+      "修正前:いつれ、私は仙台に戻ります。 修正後:いつれ、私は仙台に戻ります。\n",
+      "修正前:私はとっくに修士課程を終了しています。 修正後:私はとっくに修士課程を終了しています。\n"
+     ]
+    }
+   ],
+   "source": [
+    "\n",
+    "# 評価モード\n",
+    "spellcheckGPT.eval()\n",
+    "\n",
+    "#ファインチューニングしたモデルをbest_modelと名付け、再度生成。モデル名がbest_modelに代えていることに注意。太宰的な文章が生成されている。\n",
+    "#best_model = trainer.model.cpu()\n",
+    "\n",
+    "for text in texts:\n",
+    "    prompt = prompt_infer(text)\n",
+    "    prompt_ids = tokenizer(prompt, return_tensors = 'pt',add_special_tokens=False).input_ids\n",
+    "    outputs = spellcheckGPT.generate(\n",
+    "        input_ids=prompt_ids.to(lora_model.device),\n",
+    "        #prompt_ids,\n",
+    "        do_sample=False,\n",
+    "        max_length=60,\n",
+    "        num_return_sequences=1,\n",
+    "        pad_token_id=tokenizer.eos_token_id\n",
+    "    )\n",
+    "    #tokenizer.batch_decode(outputs, skip_special_tokens = True)\n",
+    "    print('\\n'.join(converts(tokenizer.batch_decode(outputs, skip_special_tokens = True))))\n"
+   ]
+  }
+ ],
+ "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": "3.11.5"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}