n-omniglot数据集的使用

[zirankexue]

作者你好，在复现过程中，我发现n-omniglot的精度下降极大(只有0.33)，对照n-omniglot和omniglot发现dvs-background和image_background对不齐，尝试修复无果，能否请教一下您使用n-omniglot和omniglot数据集的方式？

如果可以的话，能否上传份n-omniglot和omniglot数据集？

[ppx-hub]

作者你好，在复现过程中，我发现n-omniglot的精度下降极大(只有0.33)，对照n-omniglot和omniglot发现dvs-background和image_background对不齐，尝试修复无果，能否请教一下您使用n-omniglot和omniglot数据集的方式？

如果可以的话，能否上传份n-omniglot和omniglot数据集？

1. n-omniglot的精度下降大：我们提供了已经训练好的模型权重，您可以在这里找到：https://huggingface.co/xianghe/transfer_for_dvs/tree/main/SCNN-nomni-12-seed_1024-bs_16-DA_False-ls_0.0-lr_1.2-traindataratio_1.0-TET_loss_False-refined_False.
   您可以先进行推理尝试，查看是否是您那边n-omniglot数据集本身的问题。

2. 使用n-omniglot和omniglot数据集的方式：您都可以在 datasets.py 文件中找到。具体地，他们分别是get_nomni_data和get_omni_data函数。n-omniglot采用full模式，step=12

3. 对照n-omniglot和omniglot发现dvs-background和image_background对不齐：应该是您那边数据集划分的问题，我把我们使用的两个数据集进行上传，以保证复现。

4. 如果可以的话，能否上传份n-omniglot和omniglot数据集：可以，我们会提供下载链接

[ppx-hub]

作者你好，在复现过程中，我发现n-omniglot的精度下降极大(只有0.33)，对照n-omniglot和omniglot发现dvs-background和image_background对不齐，尝试修复无果，能否请教一下您使用n-omniglot和omniglot数据集的方式？

如果可以的话，能否上传份n-omniglot和omniglot数据集？

您好，您可以在以下两个链接分别找到n-omniglot和omniglot数据集：
Nomniglot.tar.gz，链接: https://pan.baidu.com/s/1-HNzhzymZ1GXFb_n9h3pKQ 提取码: 75jj
omniglot.zip，链接: https://pan.baidu.com/s/1nMFdnX0BaVBwJ8H1PpzEgA 提取码: 6nuq

[zirankexue]

作者你好，在复现过程中，我发现n-omniglot的精度下降极大(只有0.33)，对照n-omniglot和omniglot发现dvs-background和image_background对不齐，尝试修复无果，能否请教一下您使用n-omniglot和omniglot数据集的方式？
如果可以的话，能否上传份n-omniglot和omniglot数据集？

您好，您可以在以下两个链接分别找到n-omniglot和omniglot数据集： Nomniglot.tar.gz，链接: https://pan.baidu.com/s/1-HNzhzymZ1GXFb_n9h3pKQ 提取码: 75jj omniglot.zip，链接: https://pan.baidu.com/s/1nMFdnX0BaVBwJ8H1PpzEgA 提取码: 6nuq

好的，十分感谢

[zirankexue]

你好作者，我使用您给出的权重和缓存，能实现正确精度的推理，但是仍无法正确训练：

1.在使用您给出的缓存训练的时候发生了keyError，

Traceback (most recent call last):
File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 1566, in
main()
File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 884, in main
train_metrics = train_epoch(
File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 969, in train_epoch
for batch_idx, (inputs, label) in enumerate(target_loader):
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 681, in next
data = self._next_data()
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1376, in _next_data
return self._process_data(data)
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1402, in _process_data
data.reraise()
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/_utils.py", line 461, in reraise
raise exception
KeyError: Caught KeyError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop
data = fetcher.fetch(index)
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/tonic/cached_dataset.py", line 145, in getitem
data, targets = load_from_disk_cache(file_path)
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/tonic/cached_dataset.py", line 222, in load_from_disk_cache
for index in f[name].keys():
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/h5py/_hl/group.py", line 357, in getitem
oid = h5o.open(self.id, self._e(name), lapl=self._lapl)
File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper
File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper
File "h5py/h5o.pyx", line 241, in h5py.h5o.open
KeyError: 'Unable to synchronously open object (bad object header version number)'

wandb:
wandb: You can sync this run to the cloud by running:
wandb: wandb sync /home/dxm/ysh/Transfer-for-DVS-main/utils/wandb/offline-run-20250718_142906-hrxfbjrk
wandb: Find logs at: wandb/offline-run-20250718_142906-hrxfbjrk/logs

进程已结束,退出代码1

2.“具体地，他们分别是get_nomni_data和get_omni_data函数。”我检查了代码，RGB读取使用的应该是get_transfer_omni_data，他将background和eval拼接起来，这样才和nomni的full模式类似，是否是您在后续代码做了些改动？

3.我将source_label和label打印出来，发现您给出的缓存标签之间是对齐的，我生成的则不行——根据get_transfer_omni_data和get_nomni_data，RGB是将background和eval拼接起来，并未对标签做过变动，所以source_label和label之间可能存在964的差距，您给出的缓存却没有这一问题(尽管训练开始没多久就发生（1）中的问题)

希望您能解答一下我的疑惑，非常感谢。

[ppx-hub]

你好作者，我使用您给出的权重和缓存，能实现正确精度的推理，但是仍无法正确训练：

1.在使用您给出的缓存训练的时候发生了keyError，

Traceback (most recent call last): File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 1566, in main() File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 884, in main train_metrics = train_epoch( File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 969, in train_epoch for batch_idx, (inputs, label) in enumerate(target_loader): File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 681, in next data = self._next_data() File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1376, in _next_data return self._process_data(data) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1402, in _process_data data.reraise() File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/_utils.py", line 461, in reraise raise exception KeyError: Caught KeyError in DataLoader worker process 0. Original Traceback (most recent call last): File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop data = fetcher.fetch(index) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch data = [self.dataset[idx] for idx in possibly_batched_index] File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in data = [self.dataset[idx] for idx in possibly_batched_index] File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/tonic/cached_dataset.py", line 145, in getitem data, targets = load_from_disk_cache(file_path) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/tonic/cached_dataset.py", line 222, in load_from_disk_cache for index in f[name].keys(): File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/h5py/_hl/group.py", line 357, in getitem oid = h5o.open(self.id, self._e(name), lapl=self._lapl) File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper File "h5py/h5o.pyx", line 241, in h5py.h5o.open KeyError: 'Unable to synchronously open object (bad object header version number)'

wandb: wandb: You can sync this run to the cloud by running: wandb: wandb sync /home/dxm/ysh/Transfer-for-DVS-main/utils/wandb/offline-run-20250718_142906-hrxfbjrk wandb: Find logs at: wandb/offline-run-20250718_142906-hrxfbjrk/logs

进程已结束,退出代码1

2.“具体地，他们分别是get_nomni_data和get_omni_data函数。”我检查了代码，RGB读取使用的应该是get_transfer_omni_data，他将background和eval拼接起来，这样才和nomni的full模式类似，是否是您在后续代码做了些改动？

3.我将source_label和label打印出来，发现您给出的缓存标签之间是对齐的，我生成的则不行——根据get_transfer_omni_data和get_nomni_data，RGB是将background和eval拼接起来，并未对标签做过变动，所以source_label和label之间可能存在964的差距，您给出的缓存却没有这一问题(尽管训练开始没多久就发生（1）中的问题)

希望您能解答一下我的疑惑，非常感谢。

已收到您的问题，我看一下

[ppx-hub]

你好作者，我使用您给出的权重和缓存，能实现正确精度的推理，但是仍无法正确训练：

1.在使用您给出的缓存训练的时候发生了keyError，

Traceback (most recent call last): File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 1566, in main() File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 884, in main train_metrics = train_epoch( File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 969, in train_epoch for batch_idx, (inputs, label) in enumerate(target_loader): File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 681, in next data = self._next_data() File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1376, in _next_data return self._process_data(data) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1402, in _process_data data.reraise() File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/_utils.py", line 461, in reraise raise exception KeyError: Caught KeyError in DataLoader worker process 0. Original Traceback (most recent call last): File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop data = fetcher.fetch(index) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch data = [self.dataset[idx] for idx in possibly_batched_index] File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in data = [self.dataset[idx] for idx in possibly_batched_index] File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/tonic/cached_dataset.py", line 145, in getitem data, targets = load_from_disk_cache(file_path) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/tonic/cached_dataset.py", line 222, in load_from_disk_cache for index in f[name].keys(): File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/h5py/_hl/group.py", line 357, in getitem oid = h5o.open(self.id, self._e(name), lapl=self._lapl) File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper File "h5py/h5o.pyx", line 241, in h5py.h5o.open KeyError: 'Unable to synchronously open object (bad object header version number)'

wandb: wandb: You can sync this run to the cloud by running: wandb: wandb sync /home/dxm/ysh/Transfer-for-DVS-main/utils/wandb/offline-run-20250718_142906-hrxfbjrk wandb: Find logs at: wandb/offline-run-20250718_142906-hrxfbjrk/logs

进程已结束,退出代码1

2.“具体地，他们分别是get_nomni_data和get_omni_data函数。”我检查了代码，RGB读取使用的应该是get_transfer_omni_data，他将background和eval拼接起来，这样才和nomni的full模式类似，是否是您在后续代码做了些改动？

3.我将source_label和label打印出来，发现您给出的缓存标签之间是对齐的，我生成的则不行——根据get_transfer_omni_data和get_nomni_data，RGB是将background和eval拼接起来，并未对标签做过变动，所以source_label和label之间可能存在964的差距，您给出的缓存却没有这一问题(尽管训练开始没多久就发生（1）中的问题)

希望您能解答一下我的疑惑，非常感谢。

您好，我刚刚调试了程序。get_transfer_omni_data是用于"dataset = torch.utils.data.ConcatDataset([train_dataset, test_dataset])"，train_dataset和test_dataset的样本数目分别为19280，13180，因此dataset 共有32460个样本。get_nomni_data用于读取nomni数据，训练样本和测试样本数目分别为24345和8115，因此共有32460个样本。请您先检查这两个数据集的总样本数是否相同。

[ppx-hub]

另外，您指的“所以source_label和label之间可能存在964的差距”是什么意思？为什么会有964

[zirankexue]

另外，您指的“所以source_label和label之间可能存在964的差距”是什么意思？为什么会有964
在NOmniglotfull类中：
testSet = NOmniglot(root=root, train=False, frames_num=frames_num, data_type=data_type,
transform=transform, target_transform=lambda x: x + 964,
use_npz=use_npz, crop=crop, create=create)
手动讲test数据集的标签向后移动了964，RGB未作类似处理
RGB直接讲train和test拼接，故标签会存在重叠
train_dataset = datasets.Omniglot(
root="/home/hexiang/data/datasets/", background=True, download=True, transform=transform
)
test_dataset = datasets.Omniglot(
root="/home/hexiang/data/datasets/", background=False, download=True, transform=transform
)
dataset = torch.utils.data.ConcatDataset([train_dataset, test_dataset])

[zirankexue]

我的get_nomni_data发生了很大偏差，分别只有10935和3645，不清楚是哪一步处理的问题 孜然萪学 ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "Brain-Cog-Lab/Transfer-for-DVS" ***@***.***>; 发送时间: 2025年7月18日(星期五) 下午4:01 ***@***.***>; ***@***.******@***.***>; 主题: Re: [Brain-Cog-Lab/Transfer-for-DVS] n-omniglot数据集的使用 (Issue #3) ppx-hub left a comment (Brain-Cog-Lab/Transfer-for-DVS#3) 你好作者，我使用您给出的权重和缓存，能实现正确精度的推理，但是仍无法正确训练： 1.在使用您给出的缓存训练的时候发生了keyError， Traceback (most recent call last): File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 1566, in main() File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 884, in main train_metrics = train_epoch( File "/home/dxm/ysh/Transfer-for-DVS-main/utils/main_transfer.py", line 969, in train_epoch for batch_idx, (inputs, label) in enumerate(target_loader): File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 681, in next data = self._next_data() File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1376, in _next_data return self._process_data(data) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/dataloader.py", line 1402, in _process_data data.reraise() File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/_utils.py", line 461, in reraise raise exception KeyError: Caught KeyError in DataLoader worker process 0. Original Traceback (most recent call last): File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/worker.py", line 302, in _worker_loop data = fetcher.fetch(index) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch data = [self.dataset[idx] for idx in possibly_batched_index] File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/torch/utils/data/_utils/fetch.py", line 49, in data = [self.dataset[idx] for idx in possibly_batched_index] File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/tonic/cached_dataset.py", line 145, in getitem data, targets = load_from_disk_cache(file_path) File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/tonic/cached_dataset.py", line 222, in load_from_disk_cache for index in f[name].keys(): File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper File "/home/dxm/anaconda3/envs/MLA/lib/python3.8/site-packages/h5py/_hl/group.py", line 357, in getitem oid = h5o.open(self.id, self._e(name), lapl=self._lapl) File "h5py/_objects.pyx", line 54, in h5py._objects.with_phil.wrapper File "h5py/_objects.pyx", line 55, in h5py._objects.with_phil.wrapper File "h5py/h5o.pyx", line 241, in h5py.h5o.open KeyError: 'Unable to synchronously open object (bad object header version number)' wandb: wandb: You can sync this run to the cloud by running: wandb: wandb sync /home/dxm/ysh/Transfer-for-DVS-main/utils/wandb/offline-run-20250718_142906-hrxfbjrk wandb: Find logs at: wandb/offline-run-20250718_142906-hrxfbjrk/logs 进程已结束,退出代码1 2.“具体地，他们分别是get_nomni_data和get_omni_data函数。”我检查了代码，RGB读取使用的应该是get_transfer_omni_data，他将background和eval拼接起来，这样才和nomni的full模式类似，是否是您在后续代码做了些改动？ 3.我将source_label和label打印出来，发现您给出的缓存标签之间是对齐的，我生成的则不行——根据get_transfer_omni_data和get_nomni_data，RGB是将background和eval拼接起来，并未对标签做过变动，所以source_label和label之间可能存在964的差距，您给出的缓存却没有这一问题(尽管训练开始没多久就发生（1）中的问题) 希望您能解答一下我的疑惑，非常感谢。 您好，我刚刚调试了程序。get_transfer_omni_data是用于"dataset = torch.utils.data.ConcatDataset([train_dataset, test_dataset])"，train_dataset和test_dataset的样本数目分别为19280，13180，因此dataset 共有32460个样本。get_nomni_data用于读取nomni数据，训练样本和测试样本数目分别为24345和8115，因此共有32460个样本。请您先检查这两个数据集的总样本数是否相同。 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[ppx-hub]

964是对的，我刚刚调试了torchvision中train_dataset中的类别数，即源码中的

        self._characters: List[str] = sum(
            ([join(a, c) for c in sorted(list_dir(join(self.target_folder, a)))] for a in self._alphabets), []
        )

所以加964是为了标签不重叠（因为nomni的测试集label一开始是从零开始的）

[ppx-hub]

我的get_nomni_data发生了很大偏差，分别只有10935和3645，不清楚是哪一步处理的问题

孜然萪学
@.***

 
…

这个数是不对的，The Omniglot dataset (Lake, Salakhutdinov, and Tenenbaum 2015) consists of 1,623 handwritten char�acters from 50 different languages, each with 20 different handwritings, N-Omniglot采用了75%作为训练集，所以应该为：1623 * 20 * 0.75=24345和1623 * 20 * 0.25=8115.

[ppx-hub]

您可以再检查一下omni数据集的数目吗

我的get_nomni_data发生了很大偏差，分别只有10935和3645，不清楚是哪一步处理的问题

孜然萪学
@.***

 
…

[zirankexue]

964是对的，我刚刚调试了torchvision中train_dataset中的类别数，即源码中的

        self._characters: List[str] = sum(
            ([join(a, c) for c in sorted(list_dir(join(self.target_folder, a)))] for a in self._alphabets), []
        )

所以加964是为了标签不重叠（因为nomni的测试集label一开始是从零开始的）

但是rgb数据并未做此处理，其eval中的标签也是从0开始，所以和dvs的标签会存在偏差，使用您的缓存时标签却是对的上的

[zirankexue]

您可以再检查一下omni数据集的数目吗

我的get_nomni_data发生了很大偏差，分别只有10935和3645，不清楚是哪一步处理的问题
孜然萪学
@.***
 
…

omni数据集的数目是19280，13180，完全正确

[ppx-hub]

964是对的，我刚刚调试了torchvision中train_dataset中的类别数，即源码中的

        self._characters: List[str] = sum(
            ([join(a, c) for c in sorted(list_dir(join(self.target_folder, a)))] for a in self._alphabets), []
        )

所以加964是为了标签不重叠（因为nomni的测试集label一开始是从零开始的）

但是rgb数据并未做此处理，其eval中的标签也是从0开始，所以和dvs的标签会存在偏差，使用您的缓存时标签却是对的上的

“但是rgb数据并未做此处理，其eval中的标签也是从0开始”，没错。标签对的上是什么意思呢？source_label_list?和谁对的上呢？