|
| 1 | +import random |
| 2 | +import mindspore as ms |
| 3 | +from mindspore import nn, ops, Tensor |
| 4 | +from mindspore.dataset import GeneratorDataset |
| 5 | +from mindnlp.transformers import AlbertTokenizer, AlbertForSequenceClassification |
| 6 | +from mindnlp.engine import Trainer, TrainingArguments |
| 7 | +from datasets import load_dataset |
| 8 | +import numpy as np |
| 9 | +import os |
| 10 | +import evaluate |
| 11 | + |
| 12 | +# 1. 加载预训练模型和分词器 |
| 13 | +model_name = "albert-base-v1" |
| 14 | +tokenizer = AlbertTokenizer.from_pretrained(model_name) |
| 15 | +model = AlbertForSequenceClassification.from_pretrained( |
| 16 | + model_name, num_labels=2) |
| 17 | + |
| 18 | +# 2. 加载IMDb数据集 |
| 19 | +dataset = load_dataset("stanfordnlp/imdb", trust_remote_code=True) |
| 20 | +print("dataset:", dataset) |
| 21 | +# 3. 数据预处理函数 |
| 22 | + |
| 23 | + |
| 24 | +def tokenize_function(examples): |
| 25 | + tokenized = tokenizer( |
| 26 | + examples["text"], |
| 27 | + padding="max_length", |
| 28 | + truncation=True, |
| 29 | + max_length=512 |
| 30 | + ) |
| 31 | + # 添加标签到返回字典 |
| 32 | + tokenized["labels"] = examples["label"] |
| 33 | + return tokenized |
| 34 | + |
| 35 | + |
| 36 | +# 应用预处理 |
| 37 | +tokenized_datasets = dataset.map(tokenize_function, batched=True) |
| 38 | + |
| 39 | +# 检查标签分布（修正后的代码） |
| 40 | +print("\n==== 数据分布验证 ====") |
| 41 | + |
| 42 | +# 检查训练集 |
| 43 | +train_labels = np.array(tokenized_datasets["train"]["labels"]) |
| 44 | +print("训练集标签统计:") |
| 45 | +print("- 唯一值:", np.unique(train_labels)) |
| 46 | +print("- 分布:", np.bincount(train_labels)) |
| 47 | + |
| 48 | +# 检查测试集 |
| 49 | +test_labels = np.array(tokenized_datasets["test"]["labels"]) |
| 50 | +print("\n测试集标签统计:") |
| 51 | +print("- 唯一值:", np.unique(test_labels)) |
| 52 | +print("- 分布:", np.bincount(test_labels)) |
| 53 | +# 4. 转换数据集格式 |
| 54 | + |
| 55 | +def create_dataset(data, batch_size=8): |
| 56 | + # 将数据转换为列表以便打乱 |
| 57 | + data_list = list(data) |
| 58 | + random.shuffle(data_list) # 打乱数据顺序 |
| 59 | + |
| 60 | + def generator(): |
| 61 | + for item in data_list: # 遍历打乱后的数据 |
| 62 | + yield item["input_ids"], item["attention_mask"], Tensor(item["labels"], dtype=ms.int32) |
| 63 | + |
| 64 | + return GeneratorDataset(generator(), ["input_ids", "attention_mask", "labels"]).batch(batch_size) |
| 65 | + |
| 66 | + |
| 67 | +train_dataset = create_dataset(tokenized_datasets["train"]) |
| 68 | +eval_dataset = create_dataset(tokenized_datasets["test"]) |
| 69 | + |
| 70 | +# 5. 加载评估指标 |
| 71 | +accuracy = evaluate.load("accuracy") |
| 72 | +f1 = evaluate.load("f1") |
| 73 | +precision = evaluate.load("precision") |
| 74 | +recall = evaluate.load("recall") |
| 75 | + |
| 76 | +sample = next(iter(train_dataset)) |
| 77 | +print("Input IDs:", sample[0]) |
| 78 | +print("Attention Mask:", sample[1]) |
| 79 | +print("Labels:", sample[2]) |
| 80 | + |
| 81 | +# 自定义指标计算函数 |
| 82 | +def compute_metrics(eval_pred): |
| 83 | + logits, labels = eval_pred # 直接解包为logits和labels |
| 84 | + predictions = np.argmax(logits, axis=-1) |
| 85 | + |
| 86 | + return { |
| 87 | + "accuracy": accuracy.compute(predictions=predictions, references=labels)["accuracy"], |
| 88 | + "f1": f1.compute(predictions=predictions, references=labels, average="binary")["f1"], |
| 89 | + "precision": precision.compute(predictions=predictions, references=labels, average="binary")["precision"], |
| 90 | + "recall": recall.compute(predictions=predictions, references=labels, average="binary")["recall"] |
| 91 | + } |
| 92 | + |
| 93 | + |
| 94 | +# 6. 配置训练参数 |
| 95 | +training_args = TrainingArguments( |
| 96 | + num_train_epochs=3, |
| 97 | + per_device_train_batch_size=8, |
| 98 | + per_device_eval_batch_size=8, |
| 99 | + learning_rate=1e-5, |
| 100 | + weight_decay=0.01, |
| 101 | + output_dir="./results", |
| 102 | + logging_dir="./logs", |
| 103 | + logging_steps=10, |
| 104 | + evaluation_strategy="epoch", |
| 105 | + save_strategy="epoch", |
| 106 | + load_best_model_at_end=True, |
| 107 | + metric_for_best_model="accuracy", # 根据准确率选择最佳模型 |
| 108 | + greater_is_better=True, # 准确率越高越好 |
| 109 | +) |
| 110 | + |
| 111 | +# 7. 初始化并运行训练 |
| 112 | +trainer = Trainer( |
| 113 | + model=model, |
| 114 | + args=training_args, |
| 115 | + train_dataset=train_dataset, |
| 116 | + eval_dataset=eval_dataset, |
| 117 | + compute_metrics=compute_metrics, # 添加指标计算函数 |
| 118 | +) |
| 119 | + |
| 120 | +trainer.train() |
| 121 | + |
| 122 | +# 8. 评估模型 |
| 123 | +eval_results = trainer.evaluate(eval_dataset) |
| 124 | +print(f"Evaluation results: {eval_results}") |
| 125 | +print("\nFinal evaluation results:") |
| 126 | +print(f"Accuracy: {eval_results['eval_accuracy']:.4f}") |
| 127 | +print(f"F1 Score: {eval_results['eval_f1']:.4f}") |
| 128 | +print(f"Precision: {eval_results['eval_precision']:.4f}") |
| 129 | +print(f"Recall: {eval_results['eval_recall']:.4f}") |
| 130 | + |
0 commit comments