使用learn2learn在PyTorch中转换QuickDraw图像至84×84时遇类型错误

问题说明 #

使用learn2learn加载QuickDraw数据集时，应用调整尺寸（84×84）、随机裁剪等torchvision变换会触发TypeError: Unexpected type <class 'numpy.memmap'>错误。核心原因是QuickDraw数据集以np.memmap/.npy格式存储图像，而torchvision基于PIL实现的变换无法直接处理该类型数据。

报错堆栈 #

Traceback (most recent call last):
  File "/home/pzy2/diversity-for-predictive-success-of-meta-learning/div_src/diversity_src/dataloaders/maml_patricks_l2l.py", line 2300, in <module>
    loop_through_l2l_indexable_benchmark_with_model_test()
  File "/home/pzy2/diversity-for-predictive-success-of-meta-learning/div_src/diversity_src/dataloaders/maml_patricks_l2l.py", line 2259, in loop_through_l2l_indexable_benchmark_with_model_test
    for benchmark in [quickdraw_l2l_tasksets()]: #hdb8_l2l_tasksets(),hdb9_l2l_tasksets(), delaunay_l2l_tasksets()]:#[dtd_l2l_tasksets(), cu_birds_l2l_tasksets(), fc100_l2l_tasksets()]:
  File "/home/pzy2/diversity-for-predictive-success-of-meta-learning/div_src/diversity_src/dataloaders/maml_patricks_l2l.py", line 2216, in quickdraw_l2l_tasksets
    _transforms: tuple[TaskTransform, TaskTransform, TaskTransform] = get_task_transforms_quickdraw(_datasets,
  File "/home/pzy2/diversity-for-predictive-success-of-meta-learning/div_src/diversity_src/dataloaders/maml_patricks_l2l.py", line 2184, in get_task_transforms_quickdraw
    train_transforms: TaskTransform = DifferentTaskTransformIndexableForEachDataset(train_dataset,
  File "/home/pzy2/diversity-for-predictive-success-of-meta-learning/div_src/diversity_src/dataloaders/common.py", line 130, in __init__
    self.indexable_dataset = MetaDataset(indexable_dataset)
  File "learn2learn/data/meta_dataset.pyx", line 59, in learn2learn.data.meta_dataset.MetaDataset.__init__
  File "learn2learn/data/meta_dataset.pyx", line 96, in learn2learn.data.meta_dataset.MetaDataset.create_bookkeeping
  File "learn2learn/data/meta_dataset.pyx", line 65, in learn2learn.data.meta_dataset.MetaDataset.__getitem__
  File "/home/pzy2/miniconda3/envs/metalearning3.9/lib/python3.9/site-packages/learn2learn/vision/datasets/quickdraw.py", line 511, in __getitem__
    image = self.transform(image)
  File "/home/pzy2/miniconda3/envs/metalearning3.9/lib/python3.9/site-packages/torchvision/transforms/transforms.py", line 60, in __call__
    img = t(img)
  File "/home/pzy2/miniconda3/envs/metalearning3.9/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/pzy2/miniconda3/envs/metalearning3.9/lib/python3.9/site-packages/torchvision/transforms/transforms.py", line 900, in forward
    i, j, h, w = self.get_params(img, self.scale, self.ratio)
  File "/home/pzy2/miniconda3/envs/metalearning3.9/lib/python3.9/site-packages/torchvision/transforms/transforms.py", line 859, in get_params
    width, height = F._get_image_size(img)
  File "/home/pzy2/miniconda3/envs/metalearning3.9/lib/python3.9/site-packages/torchvision/transforms/functional.py", line 67, in _get_image_size
    return F_pil._get_image_size(img)
  File "/home/pzy2/miniconda3/envs/metalearning3.9/lib/python3.9/site-packages/torchvision/transforms/functional_pil.py", line 26, in _get_image_size
    raise TypeError("Unexpected type {}".format(type(img)))
TypeError: Unexpected type <class 'numpy.memmap'>

解决方案 #

自定义一个前置变换，先将np.memmap/numpy数组转换为PIL图像，再串联后续的尺寸调整、随机裁剪等操作。示例代码如下：

from PIL import Image
import numpy as np
from torchvision import transforms

class ToPILImageFromNumpy:
    def __call__(self, img):
        # 先将memmap转为普通numpy数组，再转成PIL灰度图（QuickDraw为单通道）
        if isinstance(img, np.memmap):
            img = np.array(img)
        return Image.fromarray(img.astype(np.uint8), mode='L')

# 构建完整的变换流水线
train_transform = transforms.Compose([
    ToPILImageFromNumpy(),
    transforms.Resize(96),  # 先放大到96，为随机裁剪留有余量
    transforms.RandomCrop(84),
    transforms.ToTensor(),
    # 可按需添加归一化等其他变换
])

# 加载QuickDraw时传入自定义变换
from learn2learn.vision.datasets import QuickDraw
train_dataset = QuickDraw(root='./data', train=True, transform=train_transform, download=True)

关键细节 #

• QuickDraw图像为单通道灰度数据，转换PIL图像时需指定mode='L'
• 先Resize到比目标尺寸大的尺寸再裁剪，避免直接缩放导致的图像拉伸变形
• 转换时先将np.memmap转为普通numpy数组，确保PIL能正确识别数据格式

内容的提问来源于stack exchange，提问作者Charlie Parker