开源实习 BEiT 模型微调 (#1975)

Author: 4everImmortality

llm/finetune/BEiT/BEiT finetune.md

@@ -0,0 +1,30 @@
+# Finetune microsoft beit-base-patch16-224 model
+- base model: [microsoft beit-base-patch16-224](https://huggingface.co/microsoft/beit-base-patch16-224)
+- dataset: [cifar10](https://huggingface.co/datasets/uoft-cs/cifar10)
+- pytorch version finetune [github](https://github.com/4everImmortality/microsoft-beit-cifar10-finetune)
+# requirments
+## pytorch 
+- GPU: RTX 4070ti 12G
+- cuda: 11.8
+- Python version: 3.10
+- torch version: 2.5.0
+- transformers version : 4.47.0
+## mindspore
+- Ascend: 910B
+- python: 3.9
+- mindspore: 2.3.1
+- mindnlp: 0.4.0
+# Result for finetune
+training for 3 epochs
+## torch
+| Epoch | eval_loss | eval_accuracy |
+|-------|-----------|--------------|
+| 1     | 0.193     | 94.4%        |
+| 2     | 0.157     | 95.4%        |
+| 3     | 0.117     | 96.2%        |
+## mindspore
+| Epoch | eval_loss | eval_accuracy |
+|-------|-----------|--------------|
+| 1     | 0.416     | 96.4%        |
+| 2     | 0.193     | 96.8%        |
+| 3     | 0.158     | 97.2%        |

llm/finetune/BEiT/BEiT_mind.py

@@ -0,0 +1,154 @@
+from mindspore.dataset import GeneratorDataset as ds_GeneratorDataset
+import numpy as np
+from sklearn.metrics import accuracy_score
+from mindnlp.engine import TrainingArguments, Trainer
+from mindnlp.transformers import AutoImageProcessor, BeitForImageClassification
+import mindspore
+from mindspore import Tensor
+from mindspore.dataset.transforms.py_transforms import Compose
+from mindspore.dataset.vision.py_transforms import (
+    RandomResizedCrop, RandomHorizontalFlip, Resize, CenterCrop, ToTensor, Normalize
+)
+from datasets import load_dataset
+
+# 加载数据集
+train_ds, test_ds = load_dataset(
+    'uoft-cs/cifar10', split=['train[:5000]', 'test[:2000]'])
+splits = train_ds.train_test_split(test_size=0.1)
+train_ds_hf = splits['train']
+val_ds_hf = splits['test']
+test_ds_hf = test_ds
+
+# 构造标签映射
+id2label = {id: label for id, label in enumerate(
+    train_ds_hf.features['label'].names)}
+label2id = {label: id for id, label in id2label.items()}
+
+# 初始化图像处理器
+processor = AutoImageProcessor.from_pretrained(
+    'microsoft/beit-base-patch16-224')
+image_mean, image_std = processor.image_mean, processor.image_std
+size = processor.size["height"]
+
+# 定义预处理流程
+normalize = Normalize(mean=image_mean, std=image_std)
+transform_train = Compose([
+    RandomResizedCrop(size),
+    RandomHorizontalFlip(),
+    ToTensor(),
+    normalize,
+])
+transform_val = Compose([
+    Resize(size),
+    CenterCrop(size),
+    ToTensor(),
+    normalize,
+])
+
+# 定义Hugging Face数据变换
+
+
+def train_transforms(examples):
+    examples['pixel_values'] = [transform_train(
+        image.convert("RGB")) for image in examples['img']]
+    return examples
+
+
+def val_transforms(examples):
+    examples['pixel_values'] = [transform_val(
+        image.convert("RGB")) for image in examples['img']]
+    return examples
+
+
+# 应用transform到原始数据集
+train_ds_hf.set_transform(train_transforms)
+val_ds_hf.set_transform(val_transforms)
+test_ds_hf.set_transform(val_transforms)
+
+# 创建 MindSpore Dataset
+
+
+def create_mindspore_dataset(hf_dataset):
+    def generator():
+        for example in hf_dataset:
+            # 获取图像数据
+            pixel_data = np.array(example['pixel_values'], dtype=np.float32)
+
+            # 中间打印调试
+            # print("Raw pixel_data shape:", pixel_data.shape)  #  (C, H, W)  (3, 224, 224)
+
+            # 处理图像数据维度
+            if pixel_data.ndim == 4 and pixel_data.shape[0] == 1:
+                #  (1, C, H, W)
+                pixel_data = pixel_data.squeeze(0)
+
+            yield pixel_data, np.int32(example['label'])
+
+    return ds_GeneratorDataset(
+        generator,
+        column_names=['pixel_values', 'labels'],
+        column_types=[mindspore.float32, mindspore.int32]
+    )
+
+
+#  创建数据集
+train_ds = create_mindspore_dataset(train_ds_hf).batch(
+    10, drop_remainder=True)  # 10个样本
+val_ds = create_mindspore_dataset(val_ds_hf).batch(4, drop_remainder=True)
+test_ds = create_mindspore_dataset(test_ds_hf).batch(4, drop_remainder=True)
+
+# 中间打印调试
+# for batch in train_ds.create_tuple_iterator():
+#    pixel_batch, label_batch = batch
+#    print("Batch shape:", pixel_batch.shape)  # 格式 (10, 3, 224, 224)
+#    break
+
+# 加载模型
+# 初始化训练参数
+args = TrainingArguments(
+    output_dir="checkpoints",
+    save_strategy="epoch",
+    evaluation_strategy="epoch",
+    learning_rate=2e-5,
+    per_device_train_batch_size=10,
+    per_device_eval_batch_size=4,
+    num_train_epochs=3,
+    weight_decay=0.01,
+    load_best_model_at_end=True,
+    metric_for_best_model="accuracy",
+    logging_dir='logs',
+    remove_unused_columns=False,
+    max_grad_norm=0.0,  # 禁用梯度裁剪 否则 Infer type failed.
+)
+
+# 初始化模型
+model = BeitForImageClassification.from_pretrained(
+    'microsoft/beit-base-patch16-224',
+    num_labels=10,
+    id2label=id2label,
+    label2id=label2id,
+    ignore_mismatched_sizes=True,
+
+)
+
+# 定义评估指标
+
+
+def compute_metrics(eval_pred):
+    predictions, labels = eval_pred
+    predictions = np.argmax(predictions, axis=1)
+    return {"accuracy": accuracy_score(predictions, labels)}
+
+
+# 初始化Trainer
+trainer = Trainer(
+    model=model,
+    args=args,
+    train_dataset=train_ds,
+    eval_dataset=val_ds,
+    compute_metrics=compute_metrics,
+    tokenizer=processor,
+)
+
+# 开始训练
+trainer.train()