Open weirman opened 9 months ago
通过resize_short字段控制的哈,建议参考一下源码https://github.com/PaddlePaddle/PaddleClas/blob/develop/ppcls/data/preprocess/ops/operators.py#L224,并且应该在yml文件中去掉crop相关部分,即可达到你想要的效果哈
我在训练过程中使用的yml文件信息如下:
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: /
cls_label_path: /mnt/cls_train/train_new.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- ResizeImage:
resize_short: 48
- RandFlipImage:
flip_code: 1
- TimmAutoAugment:
prob: 1.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: [320, 48]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 1.0
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 512
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
但是有错误提示
Traceback (most recent call last):
File "/root/miniconda3/envs/myconda/lib/python3.8/threading.py", line 932, in _bootstrap_inner
self.run()
File "/root/miniconda3/envs/myconda/lib/python3.8/threading.py", line 870, in run
self._target(*self._args, **self._kwargs)
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/dataloader_iter.py", line 604, in _thread_loop
batch = self._get_data()
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/dataloader_iter.py", line 752, in _get_data
batch.reraise()
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/worker.py", line 178, in reraise
raise self.exc_type(msg)
ValueError: DataLoader worker(6) caught ValueError with message:
Traceback (most recent call last):
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/worker.py", line 363, in _worker_loop
batch = fetcher.fetch(indices)
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/fetcher.py", line 86, in fetch
data = self.collate_fn(data)
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/collate.py", line 75, in default_collate_fn
return [default_collate_fn(fields) for fields in zip(*batch)]
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/collate.py", line 75, in <listcomp>
return [default_collate_fn(fields) for fields in zip(*batch)]
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/collate.py", line 56, in default_collate_fn
batch = np.stack(batch, axis=0)
File "<__array_function__ internals>", line 180, in stack
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/numpy/core/shape_base.py", line 426, in stack
raise ValueError('all input arrays must have the same shape')
ValueError: all input arrays must have the same shape
Traceback (most recent call last):
File "tools/train.py", line 32, in <module>
engine.train()
File "/mnt/PaddleClas-release-2.5_240124/PaddleClas-release-2.5/ppcls/engine/engine.py", line 356, in train
self.train_epoch_func(self, epoch_id, print_batch_step)
File "/mnt/PaddleClas-release-2.5_240124/PaddleClas-release-2.5/ppcls/engine/train/train.py", line 24, in train_epoch
for iter_id, batch in enumerate(engine.train_dataloader):
File "/root/miniconda3/envs/myconda/lib/python3.8/site-packages/paddle/io/dataloader/dataloader_iter.py", line 825, in __next__
self._reader.read_next_list()[0]
SystemError: (Fatal) Blocking queue is killed because the data reader raises an exception.
[Hint: Expected killed_ != true, but received killed_:1 == true:1.] (at ../paddle/fluid/operators/reader/blocking_queue.h:175)
应该是图片虽然被按照 resize_short 进行了缩放,但是没有填充,导致输入图像的宽度不在相同尺寸导致的吧?
那crop加上在看下呢,能提供一下训练的哪个模型吗
在训练文种分类模型,crop是指的添加CropImage中的size么。 看代码仅在RandomResizedCrop仅仅在里面有CropWithPadding
建议参考这个配置文件修改一下配置试试哈https://github.com/PaddlePaddle/PaddleClas/blob/4092cabf77fcbb066823560fb117ed8bca60c924/ppcls/configs/metric_learning/adaface_ir18.yaml#L61
好的,我测试一下,感觉这个模型很难训练,尤其是长宽比出现了很大的变化。
建议参考这个配置文件修改一下配置试试哈
这个文件是否已经被弃用了呢?
我的配置文件如下:
DataLoader:
Train:
dataset:
name: ImageNetDataset
image_root: /
cls_label_path: /mnt/cls_train/train.txt
transform_ops:
- DecodeImage:
to_rgb: True
channel_first: False
- CropWithPadding:
prob: 0.2
padding_num: 0
size: [112, 112]
scale: [0.2, 1.0]
ratio: [0.75, 1.3333333333333333]
- RandFlipImage:
flip_code: 1
- TimmAutoAugment:
prob: 1.0
config_str: rand-m9-mstd0.5-inc1
interpolation: bicubic
img_size: [320, 48]
- NormalizeImage:
scale: 1.0/255.0
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: ''
- RandomErasing:
EPSILON: 1.0
sl: 0.02
sh: 1.0/3.0
r1: 0.3
attempt: 10
use_log_aspect: True
mode: pixel
sampler:
name: DistributedBatchSampler
batch_size: 512
drop_last: False
shuffle: True
loader:
num_workers: 8
use_shared_memory: True
错误信息:with msg: 'CropWithPadding' object has no attribute '_get_param',另外我发现这个文件中使用的是transform
,其他文件都是使用的transform_ops
。
错误原因:_get_param 被引用了,但是没有声明,这里应该是一个bug。
把完整的配置提供一下吧,我们帮你复现下问题
我使用的是PULC language的默认参数,仅仅修改了
DataLoader:
- CropWithPadding:
prob: 0.2
padding_num: 0
size: [112, 112]
scale: [0.2, 1.0]
ratio: [0.75, 1.3333333333333333]
其他都没有改动。问题出现在_get_param(),可以发现在CropWithPadding
类中调用了self._get_param
,但是其实没有对_get_param
进行定义。
好的,我们已经记录,后续会进行测试
请问下,分类数据集图像尺寸差异过大,应该如何设置 ResizeImage 的相关配置。 我希望将图像的短边resize到统一尺寸,长边按照短边进行缩放。 目前看到这个回答。 当我设置
提示