From d5fdd4acfd2935178112d41f54f19f626be4b55e Mon Sep 17 00:00:00 2001 From: Magnus0100 Date: Mon, 15 Dec 2025 01:32:31 +0800 Subject: [PATCH 1/6] [Feature] Add BERT SWAG Multiple Choice application (Ref: #ID9IPX) --- nlp/README.md | 6 +- .../swag_bert_run_result.ipynb | 909 ++++++++++++++++++ 2 files changed, 912 insertions(+), 3 deletions(-) create mode 100644 nlp/multiple_choice/swag_bert_run_result.ipynb diff --git a/nlp/README.md b/nlp/README.md index a2c9dae..a96be8f 100644 --- a/nlp/README.md +++ b/nlp/README.md @@ -4,9 +4,9 @@ This directory contains ready-to-use Natural Language Processing application not ## Application List -| No. | Model | Description | -| :-- | :---- | :------------------------------ | -| 1 | / | This section is empty for now — feel free to contribute your first application! | +| No. | Model | Description | +| :-- | :---- | :---------- | +| 1 | [BERT (SWAG)](./multiple_choice/swag_bert_run_result.ipynb) | Fine-tuning BERT on SWAG dataset for Multiple Choice tasks using MindSpore NLP. | ## Contributing New NLP Applications diff --git a/nlp/multiple_choice/swag_bert_run_result.ipynb b/nlp/multiple_choice/swag_bert_run_result.ipynb new file mode 100644 index 0000000..f2b7fd1 --- /dev/null +++ b/nlp/multiple_choice/swag_bert_run_result.ipynb @@ -0,0 +1,909 @@ +{ + "cells": [ + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "# SWAG Multiple Choice Finetune + Inference (MindSpore + MindHF)\n", + "\n", + "**特性：**\n", + "\n", + "- 流程：训练 -> 评估 -> 保存 -> 推理演示\n", + "\n", + "本案例基于 MindSpore 2.7.1 和 MindHF 0.6.0 实现 BERT 模型在 SWAG 数据集上的多选任务微调。\n", + "## 1. 导入依赖与环境配置\n", + "\n", + "导入必要的库，并设置 HuggingFace 镜像加速下载。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 1, + "metadata": { + "editable": true, + "execution": { + "iopub.execute_input": "2025-12-14T16:21:14.050863Z", + "iopub.status.busy": "2025-12-14T16:21:14.050528Z", + "iopub.status.idle": "2025-12-14T16:21:21.091146Z", + "shell.execute_reply": "2025-12-14T16:21:21.089643Z" + }, + "slideshow": { + "slide_type": "" + }, + "tags": [] + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for type is zero.\n", + " setattr(self, word, getattr(machar, word).flat[0])\n", + "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for type is zero.\n", + " return self._float_to_str(self.smallest_subnormal)\n", + "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for type is zero.\n", + " setattr(self, word, getattr(machar, word).flat[0])\n", + "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for type is zero.\n", + " return self._float_to_str(self.smallest_subnormal)\n" + ] + }, + { + "data": { + "text/plain": [ + "'false'" + ] + }, + "execution_count": 1, + "metadata": {}, + "output_type": "execute_result" + } + ], + "source": [ + "import os\n", + "from dataclasses import dataclass\n", + "from typing import Any, Dict, List, Optional, Union\n", + "\n", + "import numpy as np\n", + "import mindspore as ms\n", + "\n", + "# ----------------------------\n", + "# Environment (HF-Mirror)\n", + "# ----------------------------\n", + "os.environ.setdefault(\"HF_ENDPOINT\", \"https://hf-mirror.com\")\n", + "os.environ.setdefault(\"TOKENIZERS_PARALLELISM\", \"false\")\n" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## 2. 定义上下文设置函数\n", + "\n", + "定义设置 MindSpore 运行环境（Ascend/PYNATIVE）和导入 MindHF 组件的辅助函数。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 2, + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-14T16:21:21.094517Z", + "iopub.status.busy": "2025-12-14T16:21:21.094070Z", + "iopub.status.idle": "2025-12-14T16:21:21.102099Z", + "shell.execute_reply": "2025-12-14T16:21:21.100812Z" + } + }, + "outputs": [], + "source": [ + "# ----------------------------\n", + "# MindSpore context\n", + "# ----------------------------\n", + "def set_ms_context():\n", + " ms.set_seed(42)\n", + " ms.set_context(mode=ms.PYNATIVE_MODE)\n", + " try:\n", + " ms.set_device(\"Ascend\", 0)\n", + " except AttributeError:\n", + " ms.set_context(device_target=\"Ascend\", device_id=0)\n", + "\n", + "\n", + "def to_numpy(x) -> np.ndarray:\n", + " if isinstance(x, np.ndarray):\n", + " return x\n", + " if hasattr(x, \"asnumpy\"):\n", + " return x.asnumpy()\n", + " return np.asarray(x)\n", + "\n", + "\n", + "# ----------------------------\n", + "# Import mindhf.transformers\n", + "# ----------------------------\n", + "def import_mindhf_transformers():\n", + " from mindhf.transformers import (\n", + " AutoTokenizer,\n", + " AutoModelForMultipleChoice,\n", + " Trainer,\n", + " TrainingArguments,\n", + " )\n", + " return AutoTokenizer, AutoModelForMultipleChoice, Trainer, TrainingArguments\n" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## 3. 初始化与加载数据集\n", + "\n", + "初始化运行环境，加载 SWAG 数据集，并根据配置截取训练集和验证集。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 3, + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-14T16:21:21.104832Z", + "iopub.status.busy": "2025-12-14T16:21:21.104529Z", + "iopub.status.idle": "2025-12-14T16:21:31.986237Z", + "shell.execute_reply": "2025-12-14T16:21:31.984267Z" + } + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", + " from .autonotebook import tqdm as notebook_tqdm\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> MindSpore: 2.7.1\n", + ">>> Device: Ascend:0 | Mode: PYNATIVE\n", + ">>> Model: google-bert/bert-base-uncased\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> Data: Train=2000, Valid=1000\n" + ] + } + ], + "source": [ + "# 运行环境初始化\n", + "set_ms_context()\n", + "AutoTokenizer, AutoModelForMultipleChoice, Trainer, TrainingArguments = import_mindhf_transformers()\n", + "\n", + "# 配置\n", + "model_checkpoint = \"google-bert/bert-base-uncased\"\n", + "output_dir = \"./my_awesome_swag_model_ms\"\n", + "\n", + "# 样本量\n", + "max_train_samples = 2000\n", + "max_eval_samples = 1000\n", + "\n", + "print(f\">>> MindSpore: {ms.__version__}\")\n", + "print(f\">>> Device: Ascend:0 | Mode: PYNATIVE\")\n", + "print(f\">>> Model: {model_checkpoint}\")\n", + "\n", + "# 1. Dataset\n", + "from datasets import load_dataset\n", + "raw = load_dataset(\"swag\", \"regular\")\n", + "raw[\"train\"] = raw[\"train\"].select(range(min(max_train_samples, len(raw[\"train\"]))))\n", + "raw[\"validation\"] = raw[\"validation\"].select(range(min(max_eval_samples, len(raw[\"validation\"]))))\n", + "print(f\">>> Data: Train={len(raw['train'])}, Valid={len(raw['validation'])}\")\n" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## 4. 数据预处理\n", + "\n", + "加载 Tokenizer，将 SWAG 数据集的上下文与选项拼接，并进行编码（Tokenize）。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 4, + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-14T16:21:31.990076Z", + "iopub.status.busy": "2025-12-14T16:21:31.989390Z", + "iopub.status.idle": "2025-12-14T16:21:49.785728Z", + "shell.execute_reply": "2025-12-14T16:21:49.784341Z" + } + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "\r\n", + "Map: 0%| | 0/2000 [00:00 Dict[str, Any]:\n", + " label_key = \"labels\" if \"labels\" in features[0] else (\"label\" if \"label\" in features[0] else None)\n", + " labels = [f.pop(label_key) for f in features] if label_key else None\n", + "\n", + " batch_size = len(features)\n", + " num_choices = len(features[0][\"input_ids\"])\n", + "\n", + " flattened = []\n", + " for feat in features:\n", + " for i in range(num_choices):\n", + " flattened.append({k: v[i] for k, v in feat.items()})\n", + "\n", + " # Try returning MindSpore tensors directly\n", + " try:\n", + " batch = self.tokenizer.pad(\n", + " flattened,\n", + " padding=self.padding,\n", + " pad_to_multiple_of=self.pad_to_multiple_of,\n", + " return_tensors=\"ms\",\n", + " )\n", + " out = {k: v.reshape((batch_size, num_choices, -1)) for k, v in batch.items()}\n", + " if labels is not None:\n", + " out[\"labels\"] = ms.Tensor(np.asarray(labels, dtype=np.int32), dtype=self.label_dtype)\n", + " return out\n", + " except Exception:\n", + " # Fallback to numpy then convert\n", + " batch_np = self.tokenizer.pad(\n", + " flattened,\n", + " padding=self.padding,\n", + " pad_to_multiple_of=self.pad_to_multiple_of,\n", + " return_tensors=\"np\",\n", + " )\n", + " out = {}\n", + " for k, v in batch_np.items():\n", + " arr = np.asarray(v).reshape((batch_size, num_choices, -1))\n", + " if k in (\"input_ids\", \"attention_mask\", \"token_type_ids\"):\n", + " arr = arr.astype(np.int32, copy=False)\n", + " out[k] = ms.Tensor(arr)\n", + " if labels is not None:\n", + " out[\"labels\"] = ms.Tensor(np.asarray(labels, dtype=np.int32), dtype=self.label_dtype)\n", + " return out\n" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## 6. 模型构建与训练配置\n", + "\n", + "加载 `BertForMultipleChoice` 模型，配置 `TrainingArguments`，并初始化 `Trainer`。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 6, + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-14T16:21:49.805891Z", + "iopub.status.busy": "2025-12-14T16:21:49.805554Z", + "iopub.status.idle": "2025-12-14T16:21:52.499460Z", + "shell.execute_reply": "2025-12-14T16:21:52.498081Z" + } + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "[MS_ALLOC_CONF] config: enable_vmm:True vmm_align_size:2MB\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "Some weights of BertForMultipleChoice were not initialized from the model checkpoint at google-bert/bert-base-uncased and are newly initialized: ['classifier.bias', 'classifier.weight']\n", + "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "/tmp/ipykernel_37813/2626949888.py:34: FutureWarning: `tokenizer` is deprecated and will be removed in version 5.0.0 for `Trainer.__init__`. Use `processing_class` instead.\n", + " trainer = Trainer(\n", + "The model is already on multiple devices. Skipping the move to device specified in `args`.\n" + ] + } + ], + "source": [ + "# 4. Model\n", + "model = AutoModelForMultipleChoice.from_pretrained(model_checkpoint)\n", + "\n", + "# 5. Metrics\n", + "def compute_metrics(eval_predictions):\n", + " if hasattr(eval_predictions, \"predictions\"):\n", + " logits = eval_predictions.predictions\n", + " labels = eval_predictions.label_ids\n", + " else:\n", + " logits, labels = eval_predictions\n", + " preds = np.argmax(to_numpy(logits), axis=1)\n", + " labels = to_numpy(labels)\n", + " return {\"accuracy\": float((preds == labels).mean())}\n", + "\n", + "# 6. TrainingArguments (Fixed for current environment)\n", + "\n", + "train_args = TrainingArguments(\n", + " output_dir=output_dir,\n", + " learning_rate=5e-5,\n", + " per_device_train_batch_size=8,\n", + " per_device_eval_batch_size=8,\n", + " num_train_epochs=3,\n", + " weight_decay=0.01,\n", + " eval_strategy=\"epoch\", \n", + " save_strategy=\"epoch\",\n", + " save_total_limit=1,\n", + " logging_steps=50,\n", + " push_to_hub=False,\n", + " remove_unused_columns=False,\n", + " report_to=[],\n", + ")\n", + "\n", + "# 7. Trainer\n", + "trainer = Trainer(\n", + " model=model,\n", + " args=train_args,\n", + " train_dataset=encoded[\"train\"],\n", + " eval_dataset=encoded[\"validation\"],\n", + " tokenizer=tokenizer,\n", + " data_collator=DataCollatorForMultipleChoice(tokenizer),\n", + " compute_metrics=compute_metrics,\n", + ")\n" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## 7. 执行训练与保存\n", + "\n", + "启动训练流程，完成后进行评估并保存模型权重。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 7, + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-14T16:21:52.502318Z", + "iopub.status.busy": "2025-12-14T16:21:52.501983Z", + "iopub.status.idle": "2025-12-14T16:25:25.956052Z", + "shell.execute_reply": "2025-12-14T16:25:25.953760Z" + } + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "You're using a BertTokenizerFast tokenizer. Please note that with a fast tokenizer, using the `__call__` method is faster than using a method to encode the text followed by a call to the `pad` method to get a padded encoding.\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + ">>> Starting training...\n" + ] + }, + { + "data": { + "text/html": [ + "\n", + "
\n", + " \n", + " \n", + " [750/750 03:19, Epoch 3/3]\n", + "
\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
EpochTraining LossValidation LossAccuracy
10.9603000.8831730.670000
20.3726000.9570890.679000
30.0976001.1632500.681000

" + ], + "text/plain": [ + "" + ] + }, + "metadata": {}, + "output_type": "display_data" + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + ">>> Starting evaluation...\n" + ] + }, + { + "data": { + "text/html": [ + "\n", + "

\n", + " \n", + " \n", + " [125/125 00:02]\n", + "
\n", + " " + ], + "text/plain": [ + "" + ] + }, + "metadata": {}, + "output_type": "display_data" + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> Eval metrics: {'eval_loss': 1.1632496118545532, 'eval_accuracy': 0.681, 'eval_runtime': 2.83, 'eval_samples_per_second': 353.355, 'eval_steps_per_second': 44.169, 'epoch': 3.0}\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> Model saved to: ./my_awesome_swag_model_ms\n" + ] + } + ], + "source": [ + "# 8. Run\n", + "print(\"\\n>>> Starting training...\")\n", + "trainer.train()\n", + "\n", + "print(\"\\n>>> Starting evaluation...\")\n", + "metrics = trainer.evaluate()\n", + "print(f\">>> Eval metrics: {metrics}\")\n", + "\n", + "# 9. Save\n", + "os.makedirs(output_dir, exist_ok=True)\n", + "trainer.save_model(output_dir)\n", + "try:\n", + " tokenizer.save_pretrained(output_dir)\n", + "except Exception:\n", + " pass\n", + "print(f\">>> Model saved to: {output_dir}\")\n" + ] + }, + { + "cell_type": "markdown", + "metadata": {}, + "source": [ + "## 8. 推理演示\n", + "\n", + "从验证集中抽取样本，进行端到端的推理预测演示。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 8, + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-14T16:25:25.962026Z", + "iopub.status.busy": "2025-12-14T16:25:25.961579Z", + "iopub.status.idle": "2025-12-14T16:25:26.000492Z", + "shell.execute_reply": "2025-12-14T16:25:25.999049Z" + } + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + ">>> Running Inference Demo...\n", + "--------------------------------------------------\n", + "Context: Students lower their eyes nervously.\n", + "Header : She\n", + " [ ] pats her shoulder, then saunters toward someone.\n", + " [ ] turns with two students.\n", + " [x] walks slowly towards someone.\n", + " [ ] wheels around as her dog thunders out.\n", + "--------------------------------------------------\n", + "Result: CORRECT (Pred: 2, Gold: 2)\n" + ] + } + ], + "source": [ + "# 10. Inference Demo\n", + "print(\"\\n>>> Running Inference Demo...\")\n", + "sample = raw[\"validation\"][0]\n", + "context = sample[\"sent1\"]\n", + "header = sample[\"sent2\"]\n", + "choices = [sample[e] for e in ending_names]\n", + "\n", + "first = [context] * 4\n", + "second = [f\"{header} {c}\" for c in choices]\n", + "\n", + "tok = tokenizer(first, second, truncation=True, padding=True, return_tensors=\"ms\")\n", + "inputs = {k: v.reshape((1, 4, -1)) for k, v in tok.items()}\n", + "\n", + "outputs = model(**inputs)\n", + "logits = outputs[\"logits\"] if isinstance(outputs, dict) else outputs.logits\n", + "pred = int(np.argmax(logits.asnumpy(), axis=1)[0])\n", + "\n", + "print(\"-\" * 50)\n", + "print(f\"Context: {context}\")\n", + "print(f\"Header : {header}\")\n", + "for i, c in enumerate(choices):\n", + " mark = \"[x]\" if i == pred else \"[ ]\"\n", + " print(f\" {mark} {c}\")\n", + "print(\"-\" * 50)\n", + "\n", + "gold = int(sample[\"label\"])\n", + "if pred == gold:\n", + " print(f\"Result: CORRECT (Pred: {pred}, Gold: {gold})\")\n", + "else:\n", + " print(f\"Result: INCORRECT (Pred: {pred}, Gold: {gold})\")\n" + ] + } + ], + "metadata": { + "kernelspec": { + "display_name": "ms_py311", + "language": "python", + "name": "ms_py311" + }, + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.11.14" + } + }, + "nbformat": 4, + "nbformat_minor": 4 +} From 92387420d72234400768e8cff8cc496f26e3eaab Mon Sep 17 00:00:00 2001 From: Magnus0100 Date: Sun, 21 Dec 2025 02:27:30 +0800 Subject: [PATCH 2/6] fixbug --- .DS_Store | Bin 0 -> 6148 bytes nlp/.DS_Store | Bin 0 -> 6148 bytes nlp/README.md | 2 +- nlp/multiple_choice/.DS_Store | Bin 0 -> 6148 bytes .../SWAG_Multiple_Choice.ipynb | 813 ++++++++++++++++ .../swag_bert_run_result.ipynb | 909 ------------------ 6 files changed, 814 insertions(+), 910 deletions(-) create mode 100644 .DS_Store create mode 100644 nlp/.DS_Store create mode 100644 nlp/multiple_choice/.DS_Store create mode 100644 nlp/multiple_choice/SWAG_Multiple_Choice.ipynb delete mode 100644 nlp/multiple_choice/swag_bert_run_result.ipynb diff --git a/.DS_Store b/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..225deb74a558d957ae8b21cc9eff906e52d07799 GIT binary patch literal 6148 zcmeHKU60a06ukqi6j6hTCi`O2#Mc#JU89LFEDMR7s1J+L#E)Q0VQnle(z2Kk5}x%x z_!oTjm-t_N(sO5qq%HWW#LOi#XF7B4?exqSLqwuI^O{68BC?PeOP7&tF>dFyVFg`t z9Vldsz^l$RnSk)JkXO zw;Z!T1G>m^#I0?dkrxl#C<-`!MTQ>H2rr5GB^2gDe946t~KtrYCetC)A~9 z)T3@TX5l=?E>eKsILne!d`(GFr7?OLP=bEmK{-O-9SU+@EKAy}2$2JtWLZ*}Ei88) zPxNN`-JJw}oHQCgM4?#PyHvKyR@HjjdmqhviIdFYwljUrPtPHxyDmHK3x7E8*KV9d zVdD7Va4H>qZwShpmwxC)^L8`~y;%D8RD)HqD*f92V$nQ2Jg|?Bn#%)w@$f+d<$L#+ z%ZhdL_T5Kk{SU!7j6UQ0U;^8}w5uA|z~iOAc|Fif)m0b0X&6Ky4}hwj@=VN$!y(e* zROk!?c}yL83amqVFY6%(H0^OE3VrgKICqrksVfcFO*iLs?|rBH4i$mA6OSVFZll=<%f``8*= z8dnO@0u#y=s9c4f1|4IQ->hwBoe3Cv}H$INfS{vyd5*zbY3Z)A&eH_aIAI1Af b($Hpe0oc;GQivLu`6D1@Fqu){uPX2x5fAnS literal 0 HcmV?d00001 diff --git a/nlp/.DS_Store b/nlp/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..fa701c2d7ccacf251c5f37c3558b0bc6fdb8145f GIT binary patch literal 6148 zcmeHKK~LK-6n-uR5+Sr6n8Y#Cu86e^G$bxrR}MSuLeO?#Q)!xPL<@1%r0b!ol=Jpu zcHJ-8-`T!r8w7My4@hW2?@7WR?*G%yC9AFkkU7~9>%$jk;$LQEBNkjV~Ccr z<7qi^?4R){&&sOZ{-qi#&DFJaZ{6GQK1LUM8dZa8Qg#R9Gp?O#oyS-4Abyvn(|&8~ zwJxebR-|J?$dVK(AKqt0qNiOwDU#A~BirHmp5Je6&t{#yy`5lxzcb$nW`{4^h+iDc z=f3y!+4G~5{+Il+(BI5FNZ}hcw(0N+z9U$o(Wj&w#&~NUVoB=PXk6sF_>4@>_UVL< zA!?7_0(Y?#$sNIeAnT^!sEUaKqJSuHKLy-==WX0?7-5S7qQE0mfcJ+0&KNqZEt;(Z zg}DL%n`kzMn12$P6F3YV))vtNQ_i_ib#f8YOalcXjJhyssF0oCk9y)LHY&eny=@mU+d&*5yG nS6jTMATUQUa``CUh8shBzyUCHSX)F3OnwB63{r>!530Z)BL#RA literal 0 HcmV?d00001 diff --git a/nlp/README.md b/nlp/README.md index a96be8f..8345145 100644 --- a/nlp/README.md +++ b/nlp/README.md @@ -6,7 +6,7 @@ This directory contains ready-to-use Natural Language Processing application not | No. | Model | Description | | :-- | :---- | :---------- | -| 1 | [BERT (SWAG)](./multiple_choice/swag_bert_run_result.ipynb) | Fine-tuning BERT on SWAG dataset for Multiple Choice tasks using MindSpore NLP. | +| 1 | [BERT (SWAG_Multiple_Choice)](./multiple_choice/SWAG_Multiple_Choice.ipynb) | Fine-tuning BERT on SWAG dataset for Multiple Choice tasks using Mind NLP. | ## Contributing New NLP Applications diff --git a/nlp/multiple_choice/.DS_Store b/nlp/multiple_choice/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..2a48537ade332bec09c484a35dcc1fa53beca53b GIT binary patch literal 6148 zcmeH~Jx;?w5QX1Dibxcsq}&rk;s#5U6x195Fd&2!kqC$oojZ;VZ+<{nhJuCyy^&_$ z?s&#)e}%^jpe%OR$G{B0lx~Vw8^&hW%}2JEkwT^O9_Q$Ag9kieKFa<+V6}&M#4|2& zJE*_I7B9F*ce>tOb?Z&XOnb}|`uxp}Gnoj8fCz|y2#CP%2w3l>O%Fmfihu}+z$XFw zK2*ACOYKAb)4{=y0MrS?Zk+pBf?6#=ZK-`IBeW8w)F>@^#3)hDc+I?)+J{Ctito(( z= 2.7.0 | >= 0.5.1 |\n", + "\n", + "首先导入必要的依赖库，并设置环境变量以使用 HF-Mirror 国内镜像加速下载。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 1, + "id": "49e9cf92", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:14:27.279544Z", + "iopub.status.busy": "2025-12-20T18:14:27.279122Z", + "iopub.status.idle": "2025-12-20T18:14:52.133383Z", + "shell.execute_reply": "2025-12-20T18:14:52.131939Z" + } + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "/root/.conda/envs/py311/lib/python3.11/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for type is zero.\n", + " setattr(self, word, getattr(machar, word).flat[0])\n", + "/root/.conda/envs/py311/lib/python3.11/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for type is zero.\n", + " return self._float_to_str(self.smallest_subnormal)\n", + "/root/.conda/envs/py311/lib/python3.11/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for type is zero.\n", + " setattr(self, word, getattr(machar, word).flat[0])\n", + "/root/.conda/envs/py311/lib/python3.11/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for type is zero.\n", + " return self._float_to_str(self.smallest_subnormal)\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "/root/.conda/envs/py311/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", + " from .autonotebook import tqdm as notebook_tqdm\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "Modular Diffusers is currently an experimental feature under active development. The API is subject to breaking changes in future releases.\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> MindSpore Version: 2.7.0\n" + ] + } + ], + "source": [ + "import os\n", + "import mindspore as ms\n", + "import numpy as np\n", + "from dataclasses import dataclass\n", + "from typing import Any, Dict, List, Optional, Union\n", + "\n", + "# ----------------------------\n", + "# Environment (HF-Mirror)\n", + "# ----------------------------\n", + "os.environ.setdefault(\"HF_ENDPOINT\", \"https://hf-mirror.com\")\n", + "os.environ.setdefault(\"TOKENIZERS_PARALLELISM\", \"false\")\n", + "\n", + "# ----------------------------\n", + "# Import mindnlp.transformers\n", + "# ----------------------------\n", + "import mindnlp # noqa: F401\n", + "\n", + "from mindnlp.transformers import (\n", + " AutoTokenizer,\n", + " AutoModelForMultipleChoice,\n", + " Trainer,\n", + " TrainingArguments,\n", + ")\n", + "\n", + "print(f\">>> MindSpore Version: {ms.__version__}\")" + ] + }, + { + "cell_type": "markdown", + "id": "4d355e84", + "metadata": {}, + "source": [ + "#### **定义辅助函数**\n", + "\n", + "为了确保代码的健壮性以及在不同硬件（Ascend/GPU/CPU）上的兼容性，我们定义以下工具函数：\n", + "- `set_ms_context`: 设置 MindSpore 运行模式（PYNATIVE）。\n", + "- `to_numpy`: 鲁棒的 Tensor 转 Numpy 函数，兼容 MindSpore Tensor 和 PyTorch Tensor。\n", + "- `move_inputs_to_device`: 确保推理时输入数据与模型在同一设备上。\n" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "36eb9465", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:14:52.136850Z", + "iopub.status.busy": "2025-12-20T18:14:52.136224Z", + "iopub.status.idle": "2025-12-20T18:14:52.153736Z", + "shell.execute_reply": "2025-12-20T18:14:52.152509Z" + } + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> Context set to PYNATIVE | Ascend:0\n" + ] + } + ], + "source": [ + "# ----------------------------\n", + "# MindSpore context\n", + "# ----------------------------\n", + "def set_ms_context():\n", + " ms.set_seed(42)\n", + " ms.set_context(mode=ms.PYNATIVE_MODE)\n", + " try:\n", + " ms.set_device(\"Ascend\", 0)\n", + " except AttributeError:\n", + " ms.set_context(device_target=\"Ascend\", device_id=0)\n", + "\n", + "def to_numpy(x) -> np.ndarray:\n", + " if x is None:\n", + " return None\n", + " if isinstance(x, np.ndarray):\n", + " return x\n", + " if isinstance(x, (list, tuple)):\n", + " return np.asarray(x)\n", + "\n", + " # MindSpore Tensor\n", + " if hasattr(x, \"asnumpy\"):\n", + " try:\n", + " return x.asnumpy()\n", + " except Exception:\n", + " pass\n", + "\n", + " # (mind)torch Tensor\n", + " if hasattr(x, \"detach\"):\n", + " try:\n", + " x = x.detach()\n", + " except Exception:\n", + " pass\n", + " if hasattr(x, \"cpu\"):\n", + " try:\n", + " x = x.cpu()\n", + " except Exception:\n", + " pass\n", + " if hasattr(x, \"numpy\"):\n", + " try:\n", + " return x.numpy()\n", + " except Exception:\n", + " pass\n", + "\n", + " return np.asarray(x)\n", + "\n", + "def get_model_device(model):\n", + " \"\"\"从参数上取 device（适配 mindtorch/torch 风格模型）。\"\"\"\n", + " try:\n", + " for p in model.parameters():\n", + " return p.device\n", + " except Exception:\n", + " return None\n", + " return None\n", + "\n", + "def move_inputs_to_device(inputs: Dict[str, Any], device):\n", + " \"\"\"把 Batch inputs 全部移动到同一 device（仅对有 .to 的张量生效）。\"\"\"\n", + " if device is None:\n", + " return inputs\n", + " out = {}\n", + " for k, v in inputs.items():\n", + " if hasattr(v, \"to\"):\n", + " out[k] = v.to(device)\n", + " else:\n", + " out[k] = v\n", + " return out\n", + "\n", + "# 初始化上下文\n", + "set_ms_context()\n", + "print(\">>> Context set to PYNATIVE | Ascend:0\")" + ] + }, + { + "cell_type": "markdown", + "id": "e6557024", + "metadata": {}, + "source": [ + "## 数据加载与预处理\n", + "\n", + "我们使用 HuggingFace `datasets` 库加载 SWAG 数据集。\n", + "为了演示效率，我们从原始训练集和验证集中截取部分数据进行训练。\n", + "\n", + "- **model_checkpoint**: 使用 `google-bert/bert-base-uncased`。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 3, + "id": "39916ecd", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:14:52.156477Z", + "iopub.status.busy": "2025-12-20T18:14:52.156148Z", + "iopub.status.idle": "2025-12-20T18:14:56.105448Z", + "shell.execute_reply": "2025-12-20T18:14:56.104088Z" + } + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> Model: google-bert/bert-base-uncased\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> Data: Train=2000, Valid=1000\n" + ] + } + ], + "source": [ + "# 配置\n", + "model_checkpoint = \"google-bert/bert-base-uncased\"\n", + "output_dir = \"./my_awesome_swag_model_ms\"\n", + "\n", + "# 样本量\n", + "max_train_samples = 2000\n", + "max_eval_samples = 1000\n", + "\n", + "print(f\">>> Model: {model_checkpoint}\")\n", + "\n", + "# 1. Dataset\n", + "from datasets import load_dataset\n", + "raw = load_dataset(\"swag\", \"regular\")\n", + "\n", + "# 截取子集\n", + "raw[\"train\"] = raw[\"train\"].select(range(min(max_train_samples, len(raw[\"train\"]))))\n", + "raw[\"validation\"] = raw[\"validation\"].select(range(min(max_eval_samples, len(raw[\"validation\"]))))\n", + "\n", + "print(f\">>> Data: Train={len(raw['train'])}, Valid={len(raw['validation'])}\")" + ] + }, + { + "cell_type": "markdown", + "id": "cd438d15", + "metadata": {}, + "source": [ + "#### **数据预处理 (Tokenization)**\n", + "\n", + "多选任务的数据预处理稍显特殊。我们需要将 **Context (sent1)** 与 **Header (sent2)** 结合，分别与 4 个 **Ending** 选项拼接，形成 4 个独立的输入序列。\n", + "\n", + "1. **Flatten**: 将 `(Batch, 4)` 的结构展平为 `(Batch * 4)` 进行 Tokenize。\n", + "2. **Tokenize**: 使用 BERT Tokenizer 进行编码。\n", + "3. **Un-flatten**: 将编码后的结果重新 reshape 回 `(Batch, 4, Seq_Len)`。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 4, + "id": "f14c6fcc", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:14:56.108448Z", + "iopub.status.busy": "2025-12-20T18:14:56.108110Z", + "iopub.status.idle": "2025-12-20T18:14:56.945719Z", + "shell.execute_reply": "2025-12-20T18:14:56.944331Z" + } + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Data preprocessing completed.\n" + ] + } + ], + "source": [ + "# 2. Tokenizer\n", + "tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, use_fast=True)\n", + "\n", + "# 3. Preprocess\n", + "ending_names = [\"ending0\", \"ending1\", \"ending2\", \"ending3\"]\n", + "\n", + "def preprocess_function(examples):\n", + " first_sentences = [[c] * 4 for c in examples[\"sent1\"]]\n", + " question_headers = examples[\"sent2\"]\n", + " second_sentences = [\n", + " [f\"{h} {examples[end][i]}\" for end in ending_names]\n", + " for i, h in enumerate(question_headers)\n", + " ]\n", + "\n", + " # Flatten\n", + " first_sentences = sum(first_sentences, [])\n", + " second_sentences = sum(second_sentences, [])\n", + "\n", + " tokenized = tokenizer(first_sentences, second_sentences, truncation=True)\n", + "\n", + " # Un-flatten\n", + " result = {k: [v[i:i + 4] for i in range(0, len(v), 4)] for k, v in tokenized.items()}\n", + " result[\"labels\"] = examples[\"label\"]\n", + " return result\n", + "\n", + "# 执行 Map 操作\n", + "encoded = raw.map(preprocess_function, batched=True, remove_columns=raw[\"train\"].column_names)\n", + "print(\"Data preprocessing completed.\")" + ] + }, + { + "cell_type": "markdown", + "id": "8f739f70", + "metadata": {}, + "source": [ + "#### **定义 DataCollator**\n", + "\n", + "定义数据整理器，负责在 Batch 层面进行 **Dynamic Padding**，并将数据转换为 MindSpore Tensor。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 5, + "id": "f6808579", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:14:56.948719Z", + "iopub.status.busy": "2025-12-20T18:14:56.948398Z", + "iopub.status.idle": "2025-12-20T18:14:56.963209Z", + "shell.execute_reply": "2025-12-20T18:14:56.962011Z" + } + }, + "outputs": [], + "source": [ + "# ----------------------------\n", + "# DataCollator\n", + "# ----------------------------\n", + "@dataclass\n", + "class DataCollatorForMultipleChoice:\n", + " tokenizer: Any\n", + " padding: Union[bool, str] = \"longest\"\n", + " pad_to_multiple_of: Optional[int] = None\n", + " label_dtype: ms.dtype = ms.int32\n", + "\n", + " def __call__(self, features: List[Dict[str, Any]]) -> Dict[str, Any]:\n", + " label_key = \"labels\" if \"labels\" in features[0] else (\"label\" if \"label\" in features[0] else None)\n", + " labels = [f.pop(label_key) for f in features] if label_key else None\n", + "\n", + " batch_size = len(features)\n", + " num_choices = len(features[0][\"input_ids\"])\n", + "\n", + " flattened = []\n", + " for feat in features:\n", + " for i in range(num_choices):\n", + " flattened.append({k: v[i] for k, v in feat.items()})\n", + "\n", + " # 训练阶段：优先返回 MindSpore 张量\n", + " try:\n", + " batch = self.tokenizer.pad(\n", + " flattened,\n", + " padding=self.padding,\n", + " pad_to_multiple_of=self.pad_to_multiple_of,\n", + " return_tensors=\"ms\",\n", + " )\n", + " out = {k: v.reshape((batch_size, num_choices, -1)) for k, v in batch.items()}\n", + " if labels is not None:\n", + " out[\"labels\"] = ms.Tensor(np.asarray(labels, dtype=np.int32), dtype=self.label_dtype)\n", + " return out\n", + " except Exception:\n", + " # Fallback\n", + " batch_np = self.tokenizer.pad(\n", + " flattened,\n", + " padding=self.padding,\n", + " pad_to_multiple_of=self.pad_to_multiple_of,\n", + " return_tensors=\"np\",\n", + " )\n", + " out = {}\n", + " for k, v in batch_np.items():\n", + " arr = np.asarray(v).reshape((batch_size, num_choices, -1))\n", + " if k in (\"input_ids\", \"attention_mask\", \"token_type_ids\"):\n", + " arr = arr.astype(np.int32, copy=False)\n", + " out[k] = ms.Tensor(arr)\n", + " if labels is not None:\n", + " out[\"labels\"] = ms.Tensor(np.asarray(labels, dtype=np.int32), dtype=self.label_dtype)\n", + " return out" + ] + }, + { + "cell_type": "markdown", + "id": "368fbebe", + "metadata": {}, + "source": [ + "## 模型构建\n", + "\n", + "使用 `AutoModelForMultipleChoice` 加载预训练模型。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 6, + "id": "41a6399a", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:14:56.966036Z", + "iopub.status.busy": "2025-12-20T18:14:56.965545Z", + "iopub.status.idle": "2025-12-20T18:14:58.261180Z", + "shell.execute_reply": "2025-12-20T18:14:58.260039Z" + } + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "[MS_ALLOC_CONF]Runtime config: enable_vmm:True vmm_align_size:2MB\n" + ] + }, + { + "name": "stderr", + "output_type": "stream", + "text": [ + "Some weights of BertForMultipleChoice were not initialized from the model checkpoint at google-bert/bert-base-uncased and are newly initialized: ['classifier.bias', 'classifier.weight']\n", + "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Model loaded.\n" + ] + } + ], + "source": [ + "# 4. Model\n", + "model = AutoModelForMultipleChoice.from_pretrained(model_checkpoint)\n", + "print(\"Model loaded.\")" + ] + }, + { + "cell_type": "markdown", + "id": "16ac1dcf", + "metadata": {}, + "source": [ + "## 模型训练\n", + "\n", + "配置 `TrainingArguments` 并初始化 `Trainer`。\n", + "我们定义 `compute_metrics` 函数来计算准确率 (Accuracy)。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 7, + "id": "f6188891", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:14:58.264194Z", + "iopub.status.busy": "2025-12-20T18:14:58.263795Z", + "iopub.status.idle": "2025-12-20T18:15:00.066353Z", + "shell.execute_reply": "2025-12-20T18:15:00.064796Z" + } + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "/tmp/ipykernel_54703/2519269147.py:30: FutureWarning: `tokenizer` is deprecated and will be removed in version 5.0.0 for `Trainer.__init__`. Use `processing_class` instead.\n", + " trainer = Trainer(\n" + ] + } + ], + "source": [ + "# 5. Metrics\n", + "def compute_metrics(eval_predictions):\n", + " if hasattr(eval_predictions, \"predictions\"):\n", + " logits = eval_predictions.predictions\n", + " labels = eval_predictions.label_ids\n", + " else:\n", + " logits, labels = eval_predictions\n", + " preds = np.argmax(to_numpy(logits), axis=1)\n", + " labels = to_numpy(labels)\n", + " return {\"accuracy\": float((preds == labels).mean())}\n", + "\n", + "# 6. TrainingArguments\n", + "train_args = TrainingArguments(\n", + " output_dir=output_dir,\n", + " learning_rate=5e-5,\n", + " per_device_train_batch_size=8,\n", + " per_device_eval_batch_size=8,\n", + " num_train_epochs=3,\n", + " weight_decay=0.01,\n", + " eval_strategy=\"epoch\",\n", + " save_strategy=\"epoch\",\n", + " save_total_limit=1,\n", + " logging_steps=50,\n", + " push_to_hub=False,\n", + " remove_unused_columns=False,\n", + " report_to=[],\n", + ")\n", + "\n", + "# 7. Trainer\n", + "trainer = Trainer(\n", + " model=model,\n", + " args=train_args,\n", + " train_dataset=encoded[\"train\"],\n", + " eval_dataset=encoded[\"validation\"],\n", + " tokenizer=tokenizer,\n", + " data_collator=DataCollatorForMultipleChoice(tokenizer),\n", + " compute_metrics=compute_metrics,\n", + ")" + ] + }, + { + "cell_type": "markdown", + "id": "4bf8bd26", + "metadata": {}, + "source": [ + "#### **执行训练与保存**\n", + "\n", + "调用 `trainer.train()` 开始训练，训练结束后保存模型和 Tokenizer。\n" + ] + }, + { + "cell_type": "code", + "execution_count": 8, + "id": "9611e086", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:15:00.069802Z", + "iopub.status.busy": "2025-12-20T18:15:00.069470Z", + "iopub.status.idle": "2025-12-20T18:22:24.765471Z", + "shell.execute_reply": "2025-12-20T18:22:24.763076Z" + } + }, + "outputs": [ + { + "name": "stderr", + "output_type": "stream", + "text": [ + "You're using a BertTokenizerFast tokenizer. Please note that with a fast tokenizer, using the `__call__` method is faster than using a method to encode the text followed by a call to the `pad` method to get a padded encoding.\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + ">>> Starting training...\n" + ] + }, + { + "data": { + "text/html": [ + "\n", + "
\n", + " \n", + " \n", + " [750/750 07:10, Epoch 3/3]\n", + "
\n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + " \n", + "
EpochTraining LossValidation LossAccuracy
10.9305000.8971920.650000
20.4527000.9427330.666000
30.1942001.1431440.681000

" + ], + "text/plain": [ + "" + ] + }, + "metadata": {}, + "output_type": "display_data" + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + ">>> Starting evaluation...\n" + ] + }, + { + "data": { + "text/html": [ + "\n", + "

\n", + " \n", + " \n", + " [125/125 00:05]\n", + "
\n", + " " + ], + "text/plain": [ + "" + ] + }, + "metadata": {}, + "output_type": "display_data" + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> Eval metrics: {'eval_loss': 1.1431440114974976, 'eval_accuracy': 0.681, 'eval_runtime': 5.1806, 'eval_samples_per_second': 193.028, 'eval_steps_per_second': 24.129, 'epoch': 3.0}\n" + ] + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + ">>> Model saved to: ./my_awesome_swag_model_ms\n" + ] + } + ], + "source": [ + "# 8. Run\n", + "print(\"\\n>>> Starting training...\")\n", + "trainer.train()\n", + "\n", + "print(\"\\n>>> Starting evaluation...\")\n", + "metrics = trainer.evaluate()\n", + "print(f\">>> Eval metrics: {metrics}\")\n", + "\n", + "# 9. Save\n", + "os.makedirs(output_dir, exist_ok=True)\n", + "trainer.save_model(output_dir)\n", + "try:\n", + " tokenizer.save_pretrained(output_dir)\n", + "except Exception:\n", + " pass\n", + "print(f\">>> Model saved to: {output_dir}\")" + ] + }, + { + "cell_type": "markdown", + "id": "e52e851f", + "metadata": {}, + "source": [ + "## 模型推理\n", + "\n", + "为了验证模型效果，我们进行一次端到端的推理演示。\n", + "\n", + "**注意（关键修复）：**\n", + "在 MindNLP 环境下，为了确保推理的稳定性和跨后端兼容性，我们采取以下策略：\n", + "1. **`return_tensors=\"pt\"`**: 使用 PyTorch 兼容的 Tensor 格式（MindNLP 会自动代理到 MindTorch）。\n", + "2. **`move_inputs_to_device`**: 显式将输入数据移动到模型参数所在的设备，避免 \"All tensor arguments must be on the same device\" 错误。\n" + ] + }, + { + "cell_type": "code", + "execution_count": null, + "id": "01860b80", + "metadata": { + "execution": { + "iopub.execute_input": "2025-12-20T18:22:24.770484Z", + "iopub.status.busy": "2025-12-20T18:22:24.770015Z", + "iopub.status.idle": "2025-12-20T18:22:24.834243Z", + "shell.execute_reply": "2025-12-20T18:22:24.832738Z" + } + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "\n", + ">>> Running Inference Demo...\n", + ">>> Inference model device: device(type=npu, index=0)\n", + "--------------------------------------------------\n", + "Context: Students lower their eyes nervously.\n", + "Header : She\n", + " [ ] pats her shoulder, then saunters toward someone.\n", + " [ ] turns with two students.\n", + " [x] walks slowly towards someone.\n", + " [ ] wheels around as her dog thunders out.\n", + "--------------------------------------------------\n", + "Result: CORRECT (Pred: 2, Gold: 2)\n" + ] + } + ], + "source": [ + "# 10. Inference Demo (FIXED)\n", + "print(\"\\n>>> Running Inference Demo...\")\n", + "\n", + "# 推理阶段需要 torch/no_grad；在 mindnlp 环境下 torch 会被代理到 mindtorch\n", + "import torch\n", + "\n", + "model.eval()\n", + "device = get_model_device(model)\n", + "print(f\">>> Inference model device: {device}\")\n", + "\n", + "# 1. 准备单条样本\n", + "sample = raw[\"validation\"][0]\n", + "context = sample[\"sent1\"]\n", + "header = sample[\"sent2\"]\n", + "choices = [sample[e] for e in ending_names]\n", + "\n", + "first = [context] * 4\n", + "second = [f\"{header} {c}\" for c in choices]\n", + "\n", + "# 2. Tokenize\n", + "tok = tokenizer(first, second, truncation=True, padding=True, return_tensors=\"pt\")\n", + "inputs = {k: v.reshape((1, 4, -1)) for k, v in tok.items()}\n", + "inputs = move_inputs_to_device(inputs, device)\n", + "\n", + "# 3. 执行前向计算 (No Grad)\n", + "with torch.no_grad():\n", + " outputs = model(**inputs)\n", + "\n", + "logits = outputs[\"logits\"] if isinstance(outputs, dict) else outputs.logits\n", + "pred = int(np.argmax(to_numpy(logits), axis=1)[0])\n", + "\n", + "# 4. 打印结果\n", + "print(\"-\" * 50)\n", + "print(f\"Context: {context}\")\n", + "print(f\"Header : {header}\")\n", + "for i, c in enumerate(choices):\n", + " mark = \"[x]\" if i == pred else \"[ ]\"\n", + " print(f\" {mark} {c}\")\n", + "print(\"-\" * 50)\n", + "\n", + "gold = int(sample[\"label\"])\n", + "if pred == gold:\n", + " print(f\"Result: CORRECT (Pred: {pred}, Gold: {gold})\")\n", + "else:\n", + " print(f\"Result: INCORRECT (Pred: {pred}, Gold: {gold})\")" + ] + } + ], + "metadata": { + "language_info": { + "codemirror_mode": { + "name": "ipython", + "version": 3 + }, + "file_extension": ".py", + "mimetype": "text/x-python", + "name": "python", + "nbconvert_exporter": "python", + "pygments_lexer": "ipython3", + "version": "3.11.14" + } + }, + "nbformat": 4, + "nbformat_minor": 5 +} diff --git a/nlp/multiple_choice/swag_bert_run_result.ipynb b/nlp/multiple_choice/swag_bert_run_result.ipynb deleted file mode 100644 index f2b7fd1..0000000 --- a/nlp/multiple_choice/swag_bert_run_result.ipynb +++ /dev/null @@ -1,909 +0,0 @@ -{ - "cells": [ - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "# SWAG Multiple Choice Finetune + Inference (MindSpore + MindHF)\n", - "\n", - "**特性：**\n", - "\n", - "- 流程：训练 -> 评估 -> 保存 -> 推理演示\n", - "\n", - "本案例基于 MindSpore 2.7.1 和 MindHF 0.6.0 实现 BERT 模型在 SWAG 数据集上的多选任务微调。\n", - "## 1. 导入依赖与环境配置\n", - "\n", - "导入必要的库，并设置 HuggingFace 镜像加速下载。\n" - ] - }, - { - "cell_type": "code", - "execution_count": 1, - "metadata": { - "editable": true, - "execution": { - "iopub.execute_input": "2025-12-14T16:21:14.050863Z", - "iopub.status.busy": "2025-12-14T16:21:14.050528Z", - "iopub.status.idle": "2025-12-14T16:21:21.091146Z", - "shell.execute_reply": "2025-12-14T16:21:21.089643Z" - }, - "slideshow": { - "slide_type": "" - }, - "tags": [] - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for type is zero.\n", - " setattr(self, word, getattr(machar, word).flat[0])\n", - "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for type is zero.\n", - " return self._float_to_str(self.smallest_subnormal)\n", - "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for type is zero.\n", - " setattr(self, word, getattr(machar, word).flat[0])\n", - "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for type is zero.\n", - " return self._float_to_str(self.smallest_subnormal)\n" - ] - }, - { - "data": { - "text/plain": [ - "'false'" - ] - }, - "execution_count": 1, - "metadata": {}, - "output_type": "execute_result" - } - ], - "source": [ - "import os\n", - "from dataclasses import dataclass\n", - "from typing import Any, Dict, List, Optional, Union\n", - "\n", - "import numpy as np\n", - "import mindspore as ms\n", - "\n", - "# ----------------------------\n", - "# Environment (HF-Mirror)\n", - "# ----------------------------\n", - "os.environ.setdefault(\"HF_ENDPOINT\", \"https://hf-mirror.com\")\n", - "os.environ.setdefault(\"TOKENIZERS_PARALLELISM\", \"false\")\n" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "## 2. 定义上下文设置函数\n", - "\n", - "定义设置 MindSpore 运行环境（Ascend/PYNATIVE）和导入 MindHF 组件的辅助函数。\n" - ] - }, - { - "cell_type": "code", - "execution_count": 2, - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-14T16:21:21.094517Z", - "iopub.status.busy": "2025-12-14T16:21:21.094070Z", - "iopub.status.idle": "2025-12-14T16:21:21.102099Z", - "shell.execute_reply": "2025-12-14T16:21:21.100812Z" - } - }, - "outputs": [], - "source": [ - "# ----------------------------\n", - "# MindSpore context\n", - "# ----------------------------\n", - "def set_ms_context():\n", - " ms.set_seed(42)\n", - " ms.set_context(mode=ms.PYNATIVE_MODE)\n", - " try:\n", - " ms.set_device(\"Ascend\", 0)\n", - " except AttributeError:\n", - " ms.set_context(device_target=\"Ascend\", device_id=0)\n", - "\n", - "\n", - "def to_numpy(x) -> np.ndarray:\n", - " if isinstance(x, np.ndarray):\n", - " return x\n", - " if hasattr(x, \"asnumpy\"):\n", - " return x.asnumpy()\n", - " return np.asarray(x)\n", - "\n", - "\n", - "# ----------------------------\n", - "# Import mindhf.transformers\n", - "# ----------------------------\n", - "def import_mindhf_transformers():\n", - " from mindhf.transformers import (\n", - " AutoTokenizer,\n", - " AutoModelForMultipleChoice,\n", - " Trainer,\n", - " TrainingArguments,\n", - " )\n", - " return AutoTokenizer, AutoModelForMultipleChoice, Trainer, TrainingArguments\n" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "## 3. 初始化与加载数据集\n", - "\n", - "初始化运行环境，加载 SWAG 数据集，并根据配置截取训练集和验证集。\n" - ] - }, - { - "cell_type": "code", - "execution_count": 3, - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-14T16:21:21.104832Z", - "iopub.status.busy": "2025-12-14T16:21:21.104529Z", - "iopub.status.idle": "2025-12-14T16:21:31.986237Z", - "shell.execute_reply": "2025-12-14T16:21:31.984267Z" - } - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/root/.conda/envs/ms_py311/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", - " from .autonotebook import tqdm as notebook_tqdm\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> MindSpore: 2.7.1\n", - ">>> Device: Ascend:0 | Mode: PYNATIVE\n", - ">>> Model: google-bert/bert-base-uncased\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> Data: Train=2000, Valid=1000\n" - ] - } - ], - "source": [ - "# 运行环境初始化\n", - "set_ms_context()\n", - "AutoTokenizer, AutoModelForMultipleChoice, Trainer, TrainingArguments = import_mindhf_transformers()\n", - "\n", - "# 配置\n", - "model_checkpoint = \"google-bert/bert-base-uncased\"\n", - "output_dir = \"./my_awesome_swag_model_ms\"\n", - "\n", - "# 样本量\n", - "max_train_samples = 2000\n", - "max_eval_samples = 1000\n", - "\n", - "print(f\">>> MindSpore: {ms.__version__}\")\n", - "print(f\">>> Device: Ascend:0 | Mode: PYNATIVE\")\n", - "print(f\">>> Model: {model_checkpoint}\")\n", - "\n", - "# 1. Dataset\n", - "from datasets import load_dataset\n", - "raw = load_dataset(\"swag\", \"regular\")\n", - "raw[\"train\"] = raw[\"train\"].select(range(min(max_train_samples, len(raw[\"train\"]))))\n", - "raw[\"validation\"] = raw[\"validation\"].select(range(min(max_eval_samples, len(raw[\"validation\"]))))\n", - "print(f\">>> Data: Train={len(raw['train'])}, Valid={len(raw['validation'])}\")\n" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "## 4. 数据预处理\n", - "\n", - "加载 Tokenizer，将 SWAG 数据集的上下文与选项拼接，并进行编码（Tokenize）。\n" - ] - }, - { - "cell_type": "code", - "execution_count": 4, - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-14T16:21:31.990076Z", - "iopub.status.busy": "2025-12-14T16:21:31.989390Z", - "iopub.status.idle": "2025-12-14T16:21:49.785728Z", - "shell.execute_reply": "2025-12-14T16:21:49.784341Z" - } - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "\r\n", - "Map: 0%| | 0/2000 [00:00 Dict[str, Any]:\n", - " label_key = \"labels\" if \"labels\" in features[0] else (\"label\" if \"label\" in features[0] else None)\n", - " labels = [f.pop(label_key) for f in features] if label_key else None\n", - "\n", - " batch_size = len(features)\n", - " num_choices = len(features[0][\"input_ids\"])\n", - "\n", - " flattened = []\n", - " for feat in features:\n", - " for i in range(num_choices):\n", - " flattened.append({k: v[i] for k, v in feat.items()})\n", - "\n", - " # Try returning MindSpore tensors directly\n", - " try:\n", - " batch = self.tokenizer.pad(\n", - " flattened,\n", - " padding=self.padding,\n", - " pad_to_multiple_of=self.pad_to_multiple_of,\n", - " return_tensors=\"ms\",\n", - " )\n", - " out = {k: v.reshape((batch_size, num_choices, -1)) for k, v in batch.items()}\n", - " if labels is not None:\n", - " out[\"labels\"] = ms.Tensor(np.asarray(labels, dtype=np.int32), dtype=self.label_dtype)\n", - " return out\n", - " except Exception:\n", - " # Fallback to numpy then convert\n", - " batch_np = self.tokenizer.pad(\n", - " flattened,\n", - " padding=self.padding,\n", - " pad_to_multiple_of=self.pad_to_multiple_of,\n", - " return_tensors=\"np\",\n", - " )\n", - " out = {}\n", - " for k, v in batch_np.items():\n", - " arr = np.asarray(v).reshape((batch_size, num_choices, -1))\n", - " if k in (\"input_ids\", \"attention_mask\", \"token_type_ids\"):\n", - " arr = arr.astype(np.int32, copy=False)\n", - " out[k] = ms.Tensor(arr)\n", - " if labels is not None:\n", - " out[\"labels\"] = ms.Tensor(np.asarray(labels, dtype=np.int32), dtype=self.label_dtype)\n", - " return out\n" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "## 6. 模型构建与训练配置\n", - "\n", - "加载 `BertForMultipleChoice` 模型，配置 `TrainingArguments`，并初始化 `Trainer`。\n" - ] - }, - { - "cell_type": "code", - "execution_count": 6, - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-14T16:21:49.805891Z", - "iopub.status.busy": "2025-12-14T16:21:49.805554Z", - "iopub.status.idle": "2025-12-14T16:21:52.499460Z", - "shell.execute_reply": "2025-12-14T16:21:52.498081Z" - } - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "[MS_ALLOC_CONF] config: enable_vmm:True vmm_align_size:2MB\n" - ] - }, - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Some weights of BertForMultipleChoice were not initialized from the model checkpoint at google-bert/bert-base-uncased and are newly initialized: ['classifier.bias', 'classifier.weight']\n", - "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n" - ] - }, - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/tmp/ipykernel_37813/2626949888.py:34: FutureWarning: `tokenizer` is deprecated and will be removed in version 5.0.0 for `Trainer.__init__`. Use `processing_class` instead.\n", - " trainer = Trainer(\n", - "The model is already on multiple devices. Skipping the move to device specified in `args`.\n" - ] - } - ], - "source": [ - "# 4. Model\n", - "model = AutoModelForMultipleChoice.from_pretrained(model_checkpoint)\n", - "\n", - "# 5. Metrics\n", - "def compute_metrics(eval_predictions):\n", - " if hasattr(eval_predictions, \"predictions\"):\n", - " logits = eval_predictions.predictions\n", - " labels = eval_predictions.label_ids\n", - " else:\n", - " logits, labels = eval_predictions\n", - " preds = np.argmax(to_numpy(logits), axis=1)\n", - " labels = to_numpy(labels)\n", - " return {\"accuracy\": float((preds == labels).mean())}\n", - "\n", - "# 6. TrainingArguments (Fixed for current environment)\n", - "\n", - "train_args = TrainingArguments(\n", - " output_dir=output_dir,\n", - " learning_rate=5e-5,\n", - " per_device_train_batch_size=8,\n", - " per_device_eval_batch_size=8,\n", - " num_train_epochs=3,\n", - " weight_decay=0.01,\n", - " eval_strategy=\"epoch\", \n", - " save_strategy=\"epoch\",\n", - " save_total_limit=1,\n", - " logging_steps=50,\n", - " push_to_hub=False,\n", - " remove_unused_columns=False,\n", - " report_to=[],\n", - ")\n", - "\n", - "# 7. Trainer\n", - "trainer = Trainer(\n", - " model=model,\n", - " args=train_args,\n", - " train_dataset=encoded[\"train\"],\n", - " eval_dataset=encoded[\"validation\"],\n", - " tokenizer=tokenizer,\n", - " data_collator=DataCollatorForMultipleChoice(tokenizer),\n", - " compute_metrics=compute_metrics,\n", - ")\n" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "## 7. 执行训练与保存\n", - "\n", - "启动训练流程，完成后进行评估并保存模型权重。\n" - ] - }, - { - "cell_type": "code", - "execution_count": 7, - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-14T16:21:52.502318Z", - "iopub.status.busy": "2025-12-14T16:21:52.501983Z", - "iopub.status.idle": "2025-12-14T16:25:25.956052Z", - "shell.execute_reply": "2025-12-14T16:25:25.953760Z" - } - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "You're using a BertTokenizerFast tokenizer. Please note that with a fast tokenizer, using the `__call__` method is faster than using a method to encode the text followed by a call to the `pad` method to get a padded encoding.\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - "\n", - ">>> Starting training...\n" - ] - }, - { - "data": { - "text/html": [ - "\n", - "
\n", - " \n", - " \n", - " [750/750 03:19, Epoch 3/3]\n", - "
\n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - "
EpochTraining LossValidation LossAccuracy
10.9603000.8831730.670000
20.3726000.9570890.679000
30.0976001.1632500.681000

" - ], - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - "\n", - ">>> Starting evaluation...\n" - ] - }, - { - "data": { - "text/html": [ - "\n", - "

\n", - " \n", - " \n", - " [125/125 00:02]\n", - "
\n", - " " - ], - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> Eval metrics: {'eval_loss': 1.1632496118545532, 'eval_accuracy': 0.681, 'eval_runtime': 2.83, 'eval_samples_per_second': 353.355, 'eval_steps_per_second': 44.169, 'epoch': 3.0}\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> Model saved to: ./my_awesome_swag_model_ms\n" - ] - } - ], - "source": [ - "# 8. Run\n", - "print(\"\\n>>> Starting training...\")\n", - "trainer.train()\n", - "\n", - "print(\"\\n>>> Starting evaluation...\")\n", - "metrics = trainer.evaluate()\n", - "print(f\">>> Eval metrics: {metrics}\")\n", - "\n", - "# 9. Save\n", - "os.makedirs(output_dir, exist_ok=True)\n", - "trainer.save_model(output_dir)\n", - "try:\n", - " tokenizer.save_pretrained(output_dir)\n", - "except Exception:\n", - " pass\n", - "print(f\">>> Model saved to: {output_dir}\")\n" - ] - }, - { - "cell_type": "markdown", - "metadata": {}, - "source": [ - "## 8. 推理演示\n", - "\n", - "从验证集中抽取样本，进行端到端的推理预测演示。\n" - ] - }, - { - "cell_type": "code", - "execution_count": 8, - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-14T16:25:25.962026Z", - "iopub.status.busy": "2025-12-14T16:25:25.961579Z", - "iopub.status.idle": "2025-12-14T16:25:26.000492Z", - "shell.execute_reply": "2025-12-14T16:25:25.999049Z" - } - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "\n", - ">>> Running Inference Demo...\n", - "--------------------------------------------------\n", - "Context: Students lower their eyes nervously.\n", - "Header : She\n", - " [ ] pats her shoulder, then saunters toward someone.\n", - " [ ] turns with two students.\n", - " [x] walks slowly towards someone.\n", - " [ ] wheels around as her dog thunders out.\n", - "--------------------------------------------------\n", - "Result: CORRECT (Pred: 2, Gold: 2)\n" - ] - } - ], - "source": [ - "# 10. Inference Demo\n", - "print(\"\\n>>> Running Inference Demo...\")\n", - "sample = raw[\"validation\"][0]\n", - "context = sample[\"sent1\"]\n", - "header = sample[\"sent2\"]\n", - "choices = [sample[e] for e in ending_names]\n", - "\n", - "first = [context] * 4\n", - "second = [f\"{header} {c}\" for c in choices]\n", - "\n", - "tok = tokenizer(first, second, truncation=True, padding=True, return_tensors=\"ms\")\n", - "inputs = {k: v.reshape((1, 4, -1)) for k, v in tok.items()}\n", - "\n", - "outputs = model(**inputs)\n", - "logits = outputs[\"logits\"] if isinstance(outputs, dict) else outputs.logits\n", - "pred = int(np.argmax(logits.asnumpy(), axis=1)[0])\n", - "\n", - "print(\"-\" * 50)\n", - "print(f\"Context: {context}\")\n", - "print(f\"Header : {header}\")\n", - "for i, c in enumerate(choices):\n", - " mark = \"[x]\" if i == pred else \"[ ]\"\n", - " print(f\" {mark} {c}\")\n", - "print(\"-\" * 50)\n", - "\n", - "gold = int(sample[\"label\"])\n", - "if pred == gold:\n", - " print(f\"Result: CORRECT (Pred: {pred}, Gold: {gold})\")\n", - "else:\n", - " print(f\"Result: INCORRECT (Pred: {pred}, Gold: {gold})\")\n" - ] - } - ], - "metadata": { - "kernelspec": { - "display_name": "ms_py311", - "language": "python", - "name": "ms_py311" - }, - "language_info": { - "codemirror_mode": { - "name": "ipython", - "version": 3 - }, - "file_extension": ".py", - "mimetype": "text/x-python", - "name": "python", - "nbconvert_exporter": "python", - "pygments_lexer": "ipython3", - "version": "3.11.14" - } - }, - "nbformat": 4, - "nbformat_minor": 4 -} From 0ef39c43c7ffff3e284d1948f3cf91ca931e3dff Mon Sep 17 00:00:00 2001 From: Magnus0100 Date: Sun, 21 Dec 2025 02:31:51 +0800 Subject: [PATCH 3/6] 1 --- nlp/multiple_choice/SWAG_Multiple_Choice.ipynb | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/nlp/multiple_choice/SWAG_Multiple_Choice.ipynb b/nlp/multiple_choice/SWAG_Multiple_Choice.ipynb index 19c4fde..9beed3c 100644 --- a/nlp/multiple_choice/SWAG_Multiple_Choice.ipynb +++ b/nlp/multiple_choice/SWAG_Multiple_Choice.ipynb @@ -18,7 +18,7 @@ "2. **数据处理**：加载 SWAG 数据集，进行 Tokenization、Flatten 处理及动态 Padding。\n", "3. **模型构建**：加载 BERT 预训练权重，构建多选分类网络。\n", "4. **模型训练**：定义损失函数与优化器，执行微调。\n", - "5. **模型推理**：加载微调后的模型，演示端到端推理（包含针对 MindNLP 跨框架特性的适配）。\n" + "5. **模型推理**：加载微调后的模型，演示端到端推理。\n" ] }, { From 1595147a65d32f76853b080b777f42149193555d Mon Sep 17 00:00:00 2001 From: Magnus0100 Date: Tue, 30 Dec 2025 15:31:03 +0800 Subject: [PATCH 4/6] [Docs] clean and rename SWAG multiple choice notebook --- nlp/README.md | 2 +- ...nb => finetune_bert_multiple_choice.ipynb} | 371 ++---------------- 2 files changed, 44 insertions(+), 329 deletions(-) rename nlp/multiple_choice/{SWAG_Multiple_Choice.ipynb => finetune_bert_multiple_choice.ipynb} (62%) diff --git a/nlp/README.md b/nlp/README.md index 8345145..6c5b393 100644 --- a/nlp/README.md +++ b/nlp/README.md @@ -6,7 +6,7 @@ This directory contains ready-to-use Natural Language Processing application not | No. | Model | Description | | :-- | :---- | :---------- | -| 1 | [BERT (SWAG_Multiple_Choice)](./multiple_choice/SWAG_Multiple_Choice.ipynb) | Fine-tuning BERT on SWAG dataset for Multiple Choice tasks using Mind NLP. | +| 1 | [BERT (SWAG Multiple Choice)](./multiple_choice/finetune_bert_multiple_choice.ipynb) | Fine-tuning BERT on SWAG dataset for Multiple Choice tasks using Mind NLP. | ## Contributing New NLP Applications diff --git a/nlp/multiple_choice/SWAG_Multiple_Choice.ipynb b/nlp/multiple_choice/finetune_bert_multiple_choice.ipynb similarity index 62% rename from nlp/multiple_choice/SWAG_Multiple_Choice.ipynb rename to nlp/multiple_choice/finetune_bert_multiple_choice.ipynb index 9beed3c..1f72324 100644 --- a/nlp/multiple_choice/SWAG_Multiple_Choice.ipynb +++ b/nlp/multiple_choice/finetune_bert_multiple_choice.ipynb @@ -2,7 +2,7 @@ "cells": [ { "cell_type": "markdown", - "id": "bc899378", + "id": "0", "metadata": {}, "source": [ "# 基于 MindSpore 的 BERT 模型 SWAG 多选阅读理解任务\n", @@ -23,7 +23,7 @@ }, { "cell_type": "markdown", - "id": "fffcc076", + "id": "1", "metadata": {}, "source": [ "## 环境准备\n", @@ -39,54 +39,10 @@ }, { "cell_type": "code", - "execution_count": 1, - "id": "49e9cf92", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:14:27.279544Z", - "iopub.status.busy": "2025-12-20T18:14:27.279122Z", - "iopub.status.idle": "2025-12-20T18:14:52.133383Z", - "shell.execute_reply": "2025-12-20T18:14:52.131939Z" - } - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/root/.conda/envs/py311/lib/python3.11/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for type is zero.\n", - " setattr(self, word, getattr(machar, word).flat[0])\n", - "/root/.conda/envs/py311/lib/python3.11/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for type is zero.\n", - " return self._float_to_str(self.smallest_subnormal)\n", - "/root/.conda/envs/py311/lib/python3.11/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for type is zero.\n", - " setattr(self, word, getattr(machar, word).flat[0])\n", - "/root/.conda/envs/py311/lib/python3.11/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for type is zero.\n", - " return self._float_to_str(self.smallest_subnormal)\n" - ] - }, - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/root/.conda/envs/py311/lib/python3.11/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n", - " from .autonotebook import tqdm as notebook_tqdm\n" - ] - }, - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Modular Diffusers is currently an experimental feature under active development. The API is subject to breaking changes in future releases.\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> MindSpore Version: 2.7.0\n" - ] - } - ], + "execution_count": null, + "id": "2", + "metadata": {}, + "outputs": [], "source": [ "import os\n", "import mindspore as ms\n", @@ -117,7 +73,7 @@ }, { "cell_type": "markdown", - "id": "4d355e84", + "id": "3", "metadata": {}, "source": [ "#### **定义辅助函数**\n", @@ -131,24 +87,9 @@ { "cell_type": "code", "execution_count": null, - "id": "36eb9465", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:14:52.136850Z", - "iopub.status.busy": "2025-12-20T18:14:52.136224Z", - "iopub.status.idle": "2025-12-20T18:14:52.153736Z", - "shell.execute_reply": "2025-12-20T18:14:52.152509Z" - } - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> Context set to PYNATIVE | Ascend:0\n" - ] - } - ], + "id": "4", + "metadata": {}, + "outputs": [], "source": [ "# ----------------------------\n", "# MindSpore context\n", @@ -223,7 +164,7 @@ }, { "cell_type": "markdown", - "id": "e6557024", + "id": "5", "metadata": {}, "source": [ "## 数据加载与预处理\n", @@ -236,32 +177,10 @@ }, { "cell_type": "code", - "execution_count": 3, - "id": "39916ecd", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:14:52.156477Z", - "iopub.status.busy": "2025-12-20T18:14:52.156148Z", - "iopub.status.idle": "2025-12-20T18:14:56.105448Z", - "shell.execute_reply": "2025-12-20T18:14:56.104088Z" - } - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> Model: google-bert/bert-base-uncased\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> Data: Train=2000, Valid=1000\n" - ] - } - ], + "execution_count": null, + "id": "6", + "metadata": {}, + "outputs": [], "source": [ "# 配置\n", "model_checkpoint = \"google-bert/bert-base-uncased\"\n", @@ -286,7 +205,7 @@ }, { "cell_type": "markdown", - "id": "cd438d15", + "id": "7", "metadata": {}, "source": [ "#### **数据预处理 (Tokenization)**\n", @@ -300,25 +219,10 @@ }, { "cell_type": "code", - "execution_count": 4, - "id": "f14c6fcc", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:14:56.108448Z", - "iopub.status.busy": "2025-12-20T18:14:56.108110Z", - "iopub.status.idle": "2025-12-20T18:14:56.945719Z", - "shell.execute_reply": "2025-12-20T18:14:56.944331Z" - } - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Data preprocessing completed.\n" - ] - } - ], + "execution_count": null, + "id": "8", + "metadata": {}, + "outputs": [], "source": [ "# 2. Tokenizer\n", "tokenizer = AutoTokenizer.from_pretrained(model_checkpoint, use_fast=True)\n", @@ -352,7 +256,7 @@ }, { "cell_type": "markdown", - "id": "8f739f70", + "id": "9", "metadata": {}, "source": [ "#### **定义 DataCollator**\n", @@ -362,16 +266,9 @@ }, { "cell_type": "code", - "execution_count": 5, - "id": "f6808579", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:14:56.948719Z", - "iopub.status.busy": "2025-12-20T18:14:56.948398Z", - "iopub.status.idle": "2025-12-20T18:14:56.963209Z", - "shell.execute_reply": "2025-12-20T18:14:56.962011Z" - } - }, + "execution_count": null, + "id": "10", + "metadata": {}, "outputs": [], "source": [ "# ----------------------------\n", @@ -429,7 +326,7 @@ }, { "cell_type": "markdown", - "id": "368fbebe", + "id": "11", "metadata": {}, "source": [ "## 模型构建\n", @@ -439,40 +336,10 @@ }, { "cell_type": "code", - "execution_count": 6, - "id": "41a6399a", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:14:56.966036Z", - "iopub.status.busy": "2025-12-20T18:14:56.965545Z", - "iopub.status.idle": "2025-12-20T18:14:58.261180Z", - "shell.execute_reply": "2025-12-20T18:14:58.260039Z" - } - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "[MS_ALLOC_CONF]Runtime config: enable_vmm:True vmm_align_size:2MB\n" - ] - }, - { - "name": "stderr", - "output_type": "stream", - "text": [ - "Some weights of BertForMultipleChoice were not initialized from the model checkpoint at google-bert/bert-base-uncased and are newly initialized: ['classifier.bias', 'classifier.weight']\n", - "You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Model loaded.\n" - ] - } - ], + "execution_count": null, + "id": "12", + "metadata": {}, + "outputs": [], "source": [ "# 4. Model\n", "model = AutoModelForMultipleChoice.from_pretrained(model_checkpoint)\n", @@ -481,7 +348,7 @@ }, { "cell_type": "markdown", - "id": "16ac1dcf", + "id": "13", "metadata": {}, "source": [ "## 模型训练\n", @@ -492,26 +359,10 @@ }, { "cell_type": "code", - "execution_count": 7, - "id": "f6188891", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:14:58.264194Z", - "iopub.status.busy": "2025-12-20T18:14:58.263795Z", - "iopub.status.idle": "2025-12-20T18:15:00.066353Z", - "shell.execute_reply": "2025-12-20T18:15:00.064796Z" - } - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "/tmp/ipykernel_54703/2519269147.py:30: FutureWarning: `tokenizer` is deprecated and will be removed in version 5.0.0 for `Trainer.__init__`. Use `processing_class` instead.\n", - " trainer = Trainer(\n" - ] - } - ], + "execution_count": null, + "id": "14", + "metadata": {}, + "outputs": [], "source": [ "# 5. Metrics\n", "def compute_metrics(eval_predictions):\n", @@ -555,7 +406,7 @@ }, { "cell_type": "markdown", - "id": "4bf8bd26", + "id": "15", "metadata": {}, "source": [ "#### **执行训练与保存**\n", @@ -565,120 +416,10 @@ }, { "cell_type": "code", - "execution_count": 8, - "id": "9611e086", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:15:00.069802Z", - "iopub.status.busy": "2025-12-20T18:15:00.069470Z", - "iopub.status.idle": "2025-12-20T18:22:24.765471Z", - "shell.execute_reply": "2025-12-20T18:22:24.763076Z" - } - }, - "outputs": [ - { - "name": "stderr", - "output_type": "stream", - "text": [ - "You're using a BertTokenizerFast tokenizer. Please note that with a fast tokenizer, using the `__call__` method is faster than using a method to encode the text followed by a call to the `pad` method to get a padded encoding.\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - "\n", - ">>> Starting training...\n" - ] - }, - { - "data": { - "text/html": [ - "\n", - "
\n", - " \n", - " \n", - " [750/750 07:10, Epoch 3/3]\n", - "
\n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - " \n", - "
EpochTraining LossValidation LossAccuracy
10.9305000.8971920.650000
20.4527000.9427330.666000
30.1942001.1431440.681000

" - ], - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - "\n", - ">>> Starting evaluation...\n" - ] - }, - { - "data": { - "text/html": [ - "\n", - "

\n", - " \n", - " \n", - " [125/125 00:05]\n", - "
\n", - " " - ], - "text/plain": [ - "" - ] - }, - "metadata": {}, - "output_type": "display_data" - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> Eval metrics: {'eval_loss': 1.1431440114974976, 'eval_accuracy': 0.681, 'eval_runtime': 5.1806, 'eval_samples_per_second': 193.028, 'eval_steps_per_second': 24.129, 'epoch': 3.0}\n" - ] - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - ">>> Model saved to: ./my_awesome_swag_model_ms\n" - ] - } - ], + "execution_count": null, + "id": "16", + "metadata": {}, + "outputs": [], "source": [ "# 8. Run\n", "print(\"\\n>>> Starting training...\")\n", @@ -700,7 +441,7 @@ }, { "cell_type": "markdown", - "id": "e52e851f", + "id": "17", "metadata": {}, "source": [ "## 模型推理\n", @@ -716,35 +457,9 @@ { "cell_type": "code", "execution_count": null, - "id": "01860b80", - "metadata": { - "execution": { - "iopub.execute_input": "2025-12-20T18:22:24.770484Z", - "iopub.status.busy": "2025-12-20T18:22:24.770015Z", - "iopub.status.idle": "2025-12-20T18:22:24.834243Z", - "shell.execute_reply": "2025-12-20T18:22:24.832738Z" - } - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "\n", - ">>> Running Inference Demo...\n", - ">>> Inference model device: device(type=npu, index=0)\n", - "--------------------------------------------------\n", - "Context: Students lower their eyes nervously.\n", - "Header : She\n", - " [ ] pats her shoulder, then saunters toward someone.\n", - " [ ] turns with two students.\n", - " [x] walks slowly towards someone.\n", - " [ ] wheels around as her dog thunders out.\n", - "--------------------------------------------------\n", - "Result: CORRECT (Pred: 2, Gold: 2)\n" - ] - } - ], + "id": "18", + "metadata": {}, + "outputs": [], "source": [ "# 10. Inference Demo (FIXED)\n", "print(\"\\n>>> Running Inference Demo...\")\n", From b75c3a390b4419f9f11b3c7f21198d94686df77f Mon Sep 17 00:00:00 2001 From: Magnus0100 Date: Wed, 31 Dec 2025 12:57:08 +0800 Subject: [PATCH 5/6] Move BERT SWAG notebook to llm --- .DS_Store | Bin 6148 -> 0 bytes llm/README.md | 1 + .../bert}/finetune_bert_multiple_choice.ipynb | 12 ++++++------ nlp/.DS_Store | Bin 6148 -> 0 bytes nlp/README.md | 4 +--- nlp/multiple_choice/.DS_Store | Bin 6148 -> 0 bytes 6 files changed, 8 insertions(+), 9 deletions(-) delete mode 100644 .DS_Store rename {nlp/multiple_choice => llm/bert}/finetune_bert_multiple_choice.ipynb (96%) delete mode 100644 nlp/.DS_Store delete mode 100644 nlp/multiple_choice/.DS_Store diff --git a/.DS_Store b/.DS_Store deleted file mode 100644 index 225deb74a558d957ae8b21cc9eff906e52d07799..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 6148 zcmeHKU60a06ukqi6j6hTCi`O2#Mc#JU89LFEDMR7s1J+L#E)Q0VQnle(z2Kk5}x%x z_!oTjm-t_N(sO5qq%HWW#LOi#XF7B4?exqSLqwuI^O{68BC?PeOP7&tF>dFyVFg`t z9Vldsz^l$RnSk)JkXO zw;Z!T1G>m^#I0?dkrxl#C<-`!MTQ>H2rr5GB^2gDe946t~KtrYCetC)A~9 z)T3@TX5l=?E>eKsILne!d`(GFr7?OLP=bEmK{-O-9SU+@EKAy}2$2JtWLZ*}Ei88) zPxNN`-JJw}oHQCgM4?#PyHvKyR@HjjdmqhviIdFYwljUrPtPHxyDmHK3x7E8*KV9d zVdD7Va4H>qZwShpmwxC)^L8`~y;%D8RD)HqD*f92V$nQ2Jg|?Bn#%)w@$f+d<$L#+ z%ZhdL_T5Kk{SU!7j6UQ0U;^8}w5uA|z~iOAc|Fif)m0b0X&6Ky4}hwj@=VN$!y(e* zROk!?c}yL83amqVFY6%(H0^OE3VrgKICqrksVfcFO*iLs?|rBH4i$mA6OSVFZll=<%f``8*= z8dnO@0u#y=s9c4f1|4IQ->hwBoe3Cv}H$INfS{vyd5*zbY3Z)A&eH_aIAI1Af b($Hpe0oc;GQivLu`6D1@Fqu){uPX2x5fAnS diff --git a/llm/README.md b/llm/README.md index 679a681..351b1e9 100644 --- a/llm/README.md +++ b/llm/README.md @@ -11,6 +11,7 @@ The following notebooks are actively maintained in sync with MindSpore and MindS | No. | Model | Description | | :-- | :---- | :----------------------- | | 1 | [t5](./t5/) | Includes notebooks for T5 finetuning and inference on tasks such as email summarization | +| 2 | [BERT (SWAG Multiple Choice)](./bert/finetune_bert_multiple_choice.ipynb) | Fine-tuning BERT on SWAG dataset for Multiple Choice tasks using MindSpore NLP | ### Community-Driven / Legacy Applications diff --git a/nlp/multiple_choice/finetune_bert_multiple_choice.ipynb b/llm/bert/finetune_bert_multiple_choice.ipynb similarity index 96% rename from nlp/multiple_choice/finetune_bert_multiple_choice.ipynb rename to llm/bert/finetune_bert_multiple_choice.ipynb index 1f72324..161d2b2 100644 --- a/nlp/multiple_choice/finetune_bert_multiple_choice.ipynb +++ b/llm/bert/finetune_bert_multiple_choice.ipynb @@ -11,7 +11,7 @@ "\n", "**SWAG** (Situations With Adversarial Generations) 是一个大规模的对抗性数据集，用于基于常识的自然语言推理 (NLI)。给定一个部分描述的事件作为上下文，任务是从四个选项中选择最合理的结尾。\n", "\n", - "本案例基于 **MindSpore** 框架和 **MindNLP** 套件，使用 **BERT** (Bidirectional Encoder Representations from Transformers) 预训练模型在 SWAG 数据集上进行微调 (Fine-tune)，实现多项选择任务的自动推理。\n", + "本案例基于 **MindSpore** 框架和 **MindSpore NLP** 套件，使用 **BERT** (Bidirectional Encoder Representations from Transformers) 预训练模型在 SWAG 数据集上进行微调 (Fine-tune)，实现多项选择任务的自动推理。\n", "\n", "**核心流程：**\n", "1. **环境准备**：配置 MindSpore 运行环境及 HF-Mirror 镜像加速。\n", @@ -30,7 +30,7 @@ "\n", "本案例运行环境要求如下：\n", "\n", - "| Python | MindSpore | MindNLP |\n", + "| Python | MindSpore | MindSpore NLP |\n", "| :----- | :-------- | :------ |\n", "| 3.9+ | >= 2.7.0 | >= 0.5.1 |\n", "\n", @@ -57,7 +57,7 @@ "os.environ.setdefault(\"TOKENIZERS_PARALLELISM\", \"false\")\n", "\n", "# ----------------------------\n", - "# Import mindnlp.transformers\n", + "# Import MindSpore NLP transformers\n", "# ----------------------------\n", "import mindnlp # noqa: F401\n", "\n", @@ -449,8 +449,8 @@ "为了验证模型效果，我们进行一次端到端的推理演示。\n", "\n", "**注意（关键修复）：**\n", - "在 MindNLP 环境下，为了确保推理的稳定性和跨后端兼容性，我们采取以下策略：\n", - "1. **`return_tensors=\"pt\"`**: 使用 PyTorch 兼容的 Tensor 格式（MindNLP 会自动代理到 MindTorch）。\n", + "在 MindSpore NLP 环境下，为了确保推理的稳定性和跨后端兼容性，我们采取以下策略：\n", + "1. **`return_tensors=\"pt\"`**: 使用 PyTorch 兼容的 Tensor 格式（MindSpore NLP 会自动代理到 MindTorch）。\n", "2. **`move_inputs_to_device`**: 显式将输入数据移动到模型参数所在的设备，避免 \"All tensor arguments must be on the same device\" 错误。\n" ] }, @@ -464,7 +464,7 @@ "# 10. Inference Demo (FIXED)\n", "print(\"\\n>>> Running Inference Demo...\")\n", "\n", - "# 推理阶段需要 torch/no_grad；在 mindnlp 环境下 torch 会被代理到 mindtorch\n", + "# 推理阶段需要 torch/no_grad；在 MindSpore NLP 环境下 torch 会被代理到 mindtorch\n", "import torch\n", "\n", "model.eval()\n", diff --git a/nlp/.DS_Store b/nlp/.DS_Store deleted file mode 100644 index fa701c2d7ccacf251c5f37c3558b0bc6fdb8145f..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 6148 zcmeHKK~LK-6n-uR5+Sr6n8Y#Cu86e^G$bxrR}MSuLeO?#Q)!xPL<@1%r0b!ol=Jpu zcHJ-8-`T!r8w7My4@hW2?@7WR?*G%yC9AFkkU7~9>%$jk;$LQEBNkjV~Ccr z<7qi^?4R){&&sOZ{-qi#&DFJaZ{6GQK1LUM8dZa8Qg#R9Gp?O#oyS-4Abyvn(|&8~ zwJxebR-|J?$dVK(AKqt0qNiOwDU#A~BirHmp5Je6&t{#yy`5lxzcb$nW`{4^h+iDc z=f3y!+4G~5{+Il+(BI5FNZ}hcw(0N+z9U$o(Wj&w#&~NUVoB=PXk6sF_>4@>_UVL< zA!?7_0(Y?#$sNIeAnT^!sEUaKqJSuHKLy-==WX0?7-5S7qQE0mfcJ+0&KNqZEt;(Z zg}DL%n`kzMn12$P6F3YV))vtNQ_i_ib#f8YOalcXjJhyssF0oCk9y)LHY&eny=@mU+d&*5yG nS6jTMATUQUa``CUh8shBzyUCHSX)F3OnwB63{r>!530Z)BL#RA diff --git a/nlp/README.md b/nlp/README.md index 6c5b393..203c638 100644 --- a/nlp/README.md +++ b/nlp/README.md @@ -4,9 +4,7 @@ This directory contains ready-to-use Natural Language Processing application not ## Application List -| No. | Model | Description | -| :-- | :---- | :---------- | -| 1 | [BERT (SWAG Multiple Choice)](./multiple_choice/finetune_bert_multiple_choice.ipynb) | Fine-tuning BERT on SWAG dataset for Multiple Choice tasks using Mind NLP. | +Currently no notebooks are listed in this directory. ## Contributing New NLP Applications diff --git a/nlp/multiple_choice/.DS_Store b/nlp/multiple_choice/.DS_Store deleted file mode 100644 index 2a48537ade332bec09c484a35dcc1fa53beca53b..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 6148 zcmeH~Jx;?w5QX1Dibxcsq}&rk;s#5U6x195Fd&2!kqC$oojZ;VZ+<{nhJuCyy^&_$ z?s&#)e}%^jpe%OR$G{B0lx~Vw8^&hW%}2JEkwT^O9_Q$Ag9kieKFa<+V6}&M#4|2& zJE*_I7B9F*ce>tOb?Z&XOnb}|`uxp}Gnoj8fCz|y2#CP%2w3l>O%Fmfihu}+z$XFw zK2*ACOYKAb)4{=y0MrS?Zk+pBf?6#=ZK-`IBeW8w)F>@^#3)hDc+I?)+J{Ctito(( z Date: Wed, 31 Dec 2025 13:02:32 +0800 Subject: [PATCH 6/6] Restore nlp README --- nlp/README.md | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/nlp/README.md b/nlp/README.md index 203c638..a2c9dae 100644 --- a/nlp/README.md +++ b/nlp/README.md @@ -4,7 +4,9 @@ This directory contains ready-to-use Natural Language Processing application not ## Application List -Currently no notebooks are listed in this directory. +| No. | Model | Description | +| :-- | :---- | :------------------------------ | +| 1 | / | This section is empty for now — feel free to contribute your first application! | ## Contributing New NLP Applications