Training starts correctly but suddenly stops after a while

[gvalvano]

Describe the bug
The training starts correctly, but it suddenly stops with AttributeError after a while.

Reproduction

The config file:

dataset_type = 'UnconditionalImageDataset'
train_pipeline = [
    dict(type='LoadImageFromFile', key='real_img', io_backend='disk'),
    dict(
        type='Resize',
        keys=['real_img'],
        scale=(1126, 1126),
        interpolation='bicubic'),
    dict(
        type='Crop',
        keys=['real_img'],
        crop_size=(1024, 1024),
        random_crop=True),
    dict(
        type='Normalize',
        keys=['real_img'],
        mean=[127.5, 127.5, 127.5],
        std=[127.5, 127.5, 127.5],
        to_rgb=False),
    dict(type='Flip', keys=['real_img'], direction='horizontal'),
    dict(type='ImageToTensor', keys=['real_img']),
    dict(type='Collect', keys=['real_img'], meta_keys=['real_img_path'])
]
val_pipeline = [
    dict(type='LoadImageFromFile', key='real_img', io_backend='disk'),
    dict(
        type='Resize',
        keys=['real_img'],
        scale=(1024, 1024),
        interpolation='bicubic'),
    dict(
        type='Normalize',
        keys=['real_img'],
        mean=[127.5, 127.5, 127.5],
        std=[127.5, 127.5, 127.5],
        to_rgb=False),
    dict(type='ImageToTensor', keys=['real_img']),
    dict(type='Collect', keys=['real_img'], meta_keys=['real_img_path'])
]
data = dict(
    samples_per_gpu=2,
    workers_per_gpu=1,
    train=dict(
        type='RepeatDataset',
        times=100,
        dataset=dict(
            type='UnconditionalImageDataset',
            imgs_root='../data/images_collection/',
            pipeline=[
                dict(
                    type='LoadImageFromFile',
                    key='real_img',
                    io_backend='disk'),
                dict(
                    type='Resize',
                    keys=['real_img'],
                    scale=(1126, 1126),
                    interpolation='bicubic'),
                dict(
                    type='Crop',
                    keys=['real_img'],
                    crop_size=(1024, 1024),
                    random_crop=True),
                dict(
                    type='Normalize',
                    keys=['real_img'],
                    mean=[127.5, 127.5, 127.5],
                    std=[127.5, 127.5, 127.5],
                    to_rgb=False),
                dict(type='Flip', keys=['real_img'], direction='horizontal'),
                dict(type='ImageToTensor', keys=['real_img']),
                dict(
                    type='Collect',
                    keys=['real_img'],
                    meta_keys=['real_img_path'])
            ])),
    val=dict(
        type='UnconditionalImageDataset',
        imgs_root='../data/images_collection/',
        pipeline=[
            dict(type='LoadImageFromFile', key='real_img', io_backend='disk'),
            dict(
                type='Resize',
                keys=['real_img'],
                scale=(1024, 1024),
                interpolation='bicubic'),
            dict(
                type='Normalize',
                keys=['real_img'],
                mean=[127.5, 127.5, 127.5],
                std=[127.5, 127.5, 127.5],
                to_rgb=False),
            dict(type='ImageToTensor', keys=['real_img']),
            dict(
                type='Collect', keys=['real_img'], meta_keys=['real_img_path'])
        ]))
d_reg_interval = 16
g_reg_interval = 4
g_reg_ratio = 0.8
d_reg_ratio = 0.9411764705882353
model = dict(
    type='StaticUnconditionalGAN',
    generator=dict(
        type='StyleGANv2Generator', out_size=1024, style_channels=512),
    discriminator=dict(type='StyleGAN2Discriminator', in_size=1024),
    gan_loss=dict(type='GANLoss', gan_type='wgan-logistic-ns'),
    disc_auxiliary_loss=dict(
        type='R1GradientPenalty',
        loss_weight=80.0,
        interval=16,
        norm_mode='HWC',
        data_info=dict(real_data='real_imgs', discriminator='disc')),
    gen_auxiliary_loss=dict(
        type='GeneratorPathRegularizer',
        loss_weight=8.0,
        pl_batch_shrink=2,
        interval=4,
        data_info=dict(generator='gen', num_batches='batch_size')))
train_cfg = dict(use_ema=True)
test_cfg = None
optimizer = dict(
    generator=dict(type='Adam', lr=0.0016, betas=(0, 0.9919919678228657)),
    discriminator=dict(
        type='Adam', lr=0.0018823529411764706, betas=(0, 0.9905854573074332)))
checkpoint_config = dict(interval=2000, by_epoch=False, max_keep_ckpts=5)
log_config = dict(
    interval=500,
    hooks=[dict(type='TextLoggerHook'),
           dict(type='TensorboardLoggerHook')])
custom_hooks = [
    dict(
        type='VisualizeUnconditionalSamples',
        output_dir='training_samples',
        interval=5000),
    dict(
        type='ExponentialMovingAverageHook',
        module_keys=('generator_ema', ),
        interval=1,
        interp_cfg=dict(momentum=0.9977843871238888),
        priority='VERY_HIGH')
]
runner = dict(
    type='DynamicIterBasedRunner',
    is_dynamic_ddp=False,
    pass_training_status=True)
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
resume_from = None
workflow = [('train', 10000)]
find_unused_parameters = True
cudnn_benchmark = True
opencv_num_threads = 0
mp_start_method = 'fork'
metrics = dict(
    fid50k=dict(
        type='FID',
        num_images=10000,
        inception_pkl=
        'results/pretrained_models/inception/ffhq-1024-50k-rgb.pkl',
        bgr2rgb=True),
    pr50k3=dict(type='PR', num_images=500, k=3),
    is50k=dict(type='IS', num_images=10000),
    ppl_zfull=dict(type='PPL', space='Z', sampling='full', num_images=10000),
    ppl_wfull=dict(type='PPL', space='W', sampling='full', num_images=10000),
    ppl_zend=dict(type='PPL', space='Z', sampling='end', num_images=10000),
    ppl_wend=dict(type='PPL', space='W', sampling='end', num_images=500),
    ms_ssim10k=dict(type='MS_SSIM', num_images=10000),
    swd16k=dict(type='SWD', num_images=10000))
run_id = 'stylegan'
data_root = '../data/images_collection/'
n_epochs = 500
n_images = 21505
total_iters = 10752500
image_resolution = 1024
inception_file_name = 'results/pretrained_models/inception/ffhq-1024-50k-rgb.pkl'
batch_size = 2
lr_config = None
ema_half_life = 10.0
_load = dict(type='LoadImageFromFile', key='real_img', io_backend='disk')
_normalize = dict(
    type='Normalize',
    keys=['real_img'],
    mean=[127.5, 127.5, 127.5],
    std=[127.5, 127.5, 127.5],
    to_rgb=False)
_resize = dict(
    type='Resize',
    keys=['real_img'],
    scale=(1126, 1126),
    interpolation='bicubic')
_resize_val = dict(
    type='Resize',
    keys=['real_img'],
    scale=(1024, 1024),
    interpolation='bicubic')
_crop = dict(
    type='Crop', keys=['real_img'], crop_size=(1024, 1024), random_crop=True)
_flip = dict(type='Flip', keys=['real_img'], direction='horizontal')
_convert = dict(type='ImageToTensor', keys=['real_img'])
_collect = dict(type='Collect', keys=['real_img'], meta_keys=['real_img_path'])
evaluation = dict(
    type='GenerativeEvalHook',
    interval=200,
    metrics=dict(
        type='FID',
        num_images=1000,
        inception_pkl=
        'results/pretrained_models/inception/ffhq-1024-50k-rgb.pkl',
        bgr2rgb=True),
    sample_kwargs=dict(sample_model='ema'))
use_ddp_wrapper = True
work_dir = 'results/logs/stylegan'
gpu_ids = range(0, 1)

2. Did you make any modifications on the code or config? Did you understand what you have modified? --> Yes
3. What dataset did you use? --> Custom, under ../data/images_collection/

Environment

1. Please run python mmgen/utils/collect_env.py to collect necessary environment information and paste it here.

/path/to/miniconda/envs/open-mmlab/lib/python3.7/runpy.py:125: RuntimeWarning: 'mmgen.utils.collect_env' found in sys.modules after import of package 'mmgen.utils', but prior to execution of 'mmgen.utils.collect_env'; this may result in unpredictable behaviour
  warn(RuntimeWarning(msg))
fatal: not a git repository (or any parent up to mount point /)
Stopping at filesystem boundary (GIT_DISCOVERY_ACROSS_FILESYSTEM not set).
sys.platform: linux
Python: 3.7.12 | packaged by conda-forge | (default, Oct 26 2021, 06:08:21) [GCC 9.4.0]
CUDA available: True
CUDA_HOME: /usr
NVCC: Cuda compilation tools, release 9.1, V9.1.85
GPU 0: TITAN Xp
GPU 1: GeForce GTX 1080 Ti
GCC: gcc (Ubuntu 5.5.0-12ubuntu1) 5.5.0 20171010
PyTorch: 1.10.2
PyTorch compiling details: PyTorch built with:
  - GCC 7.3
  - C++ Version: 201402
  - Intel(R) oneAPI Math Kernel Library Version 2022.0-Product Build 20211112 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v2.2.3 (Git Hash 7336ca9f055cf1bfa13efb658fe15dc9b41f0740)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - LAPACK is enabled (usually provided by MKL)
  - NNPACK is enabled
  - CPU capability usage: AVX512
  - CUDA Runtime 10.2
  - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_37,code=compute_37
  - CuDNN 7.6.5
  - Magma 2.5.2
  - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=10.2, CUDNN_VERSION=7.6.5, CXX_COMPILER=/opt/rh/devtoolset-7/root/usr/bin/c++, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -DEDGE_PROFILER_USE_KINETO -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.10.2, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, 

TorchVision: 0.11.3
OpenCV: 4.5.5
MMCV: 1.4.4
MMGen: 0.5.0+
MMCV Compiler: GCC 7.3
MMCV CUDA Compiler: 10.2

2. You may add addition that may be helpful for locating the problem, such as
   • How you installed PyTorch [e.g., pip, conda, source] --> conda
   • Other environment variables that may be related (such as $PATH, $LD_LIBRARY_PATH, $PYTHONPATH, etc.) --> N.A

Error traceback
If applicable, paste the error trackback here.

[...]
2022-02-03 10:34:04,432 - mmgen - INFO - Switch to evaluation style mode: single
2022-02-03 10:34:04,433 - mmgen - INFO - Sample 1000 fake images for evaluation
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 1000/1000, 19.6 task/s, elapsed: 51s, ETA:     0s
2022-02-03 10:34:59,248 - mmgen - INFO - Switch to train style mode: mix
2022-02-03 10:34:59,249 - mmgen - INFO - Switch to train style mode: mix
2022-02-03 10:34:59,251 - mmgen - INFO - Iter(val) [2800]       fid: 487.0529, fid_mean: 355.8218994140625, fid_cov: 131.2310
2022-02-03 10:38:12,054 - mmgen - INFO - Switch to evaluation style mode: single
2022-02-03 10:38:12,054 - mmgen - INFO - Switch to evaluation style mode: single
2022-02-03 10:38:12,055 - mmgen - INFO - Sample 1000 fake images for evaluation
[>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>] 1000/1000, 19.4 task/s, elapsed: 51s, ETA:     0s
2022-02-03 10:39:07,205 - mmgen - INFO - Switch to train style mode: mix
2022-02-03 10:39:07,206 - mmgen - INFO - Switch to train style mode: mix
2022-02-03 10:39:07,207 - mmgen - INFO - Exp name: stylegan.py
2022-02-03 10:39:07,207 - mmgen - INFO - Iter(val) [3000]       fid: 541.2982, fid_mean: 405.7856750488281, fid_cov: 135.5125
Traceback (most recent call last):
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/runpy.py", line 193, in _run_module_as_main
    "__main__", mod_spec)
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/runpy.py", line 85, in _run_code
    exec(code, run_globals)
  File "/path/to/project/MMGeneration/tools/train.py", line 204, in <module>
    main()
  File "/path/to/project/MMGeneration/tools/train.py", line 200, in main
    meta=meta)
  File "/path/to/project/MMGeneration/mmgen/apis/train.py", line 199, in train_model
    runner.run(data_loaders, cfg.workflow, cfg.total_iters)
  File "/path/to/project/MMGeneration/mmgen/core/runners/dynamic_iterbased_runner.py", line 285, in run
    iter_runner(iter_loaders[i], **kwargs)
  File "/path/to/project/MMGeneration/mmgen/core/runners/dynamic_iterbased_runner.py", line 197, in train
    data_batch = next(self.data_loader)
  File "/path/to/project/MMGeneration/mmgen/core/runners/dynamic_iterbased_runner.py", line 103, in __next__
    data = next(self.iter_loader)
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 521, in __next__
    data = self._next_data()
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 1203, in _next_data
    return self._process_data(data)
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/site-packages/torch/utils/data/dataloader.py", line 1229, in _process_data
    data.reraise()
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/site-packages/torch/_utils.py", line 434, in reraise
    raise exception
AttributeError: Caught AttributeError in DataLoader worker process 0.
Original Traceback (most recent call last):
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/site-packages/torch/utils/data/_utils/worker.py", line 287, in _worker_loop
    data = fetcher.fetch(index)
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 49, in fetch
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/path/to/miniconda/envs/open-mmlab/lib/python3.7/site-packages/torch/utils/data/_utils/fetch.py", line 49, in <listcomp>
    data = [self.dataset[idx] for idx in possibly_batched_index]
  File "/path/to/project/MMGeneration/mmgen/datasets/dataset_wrappers.py", line 31, in __getitem__
    return self.dataset[idx % self._ori_len]
  File "/path/to/project/MMGeneration/mmgen/datasets/unconditional_image_dataset.py", line 74, in __getitem__
    return self.prepare_train_data(idx)
  File "/path/to/project/MMGeneration/mmgen/datasets/unconditional_image_dataset.py", line 55, in prepare_train_data
    return self.pipeline(results)
  File "/path/to/project/MMGeneration/mmgen/datasets/pipelines/compose.py", line 57, in __call__
    data = t(data)
  File "/path/to/project/MMGeneration/mmgen/datasets/pipelines/loading.py", line 67, in __call__
    results[f'{self.key}_ori_shape'] = img.shape
AttributeError: 'NoneType' object has no attribute 'shape'

Bug fix
If you have already identified the reason, you can provide the information here. If you are willing to create a PR to fix it, please also leave a comment here and that would be much appreciated!

[LeoXing1996]

This seems like image loading failed and mmcv.imfrombytes with cv2 backend returns a NoneType.
Please make sure all images under your folder are valid.

[gvalvano]

Hello @LeoXing1996 and thank you for the support. It seems you may be right, after removing some images raising python warnings from the dataset I can train the model without obtaining the previous error anymore.

I have created a small snippet to detect problematic files, which I leave below, in case someone may find it useful :) let me know if you think it is worth including a bash-runnable version of this code in MMGeneration, I would be happy to create a pull request.

from pathlib import Path
import warnings

# use this to catch warnings
warnings.filterwarnings("error")

paths = glob.glob(str(Path(source_folder) / "*.jpg"))
paths.sort()
n_images = len(paths)
skipped = 0
for k, path in enumerate(paths):
    print(f"\nFile: {path}")
    try:
        x = PIL.Image.open(path)
        x = np.asarray(x, dtype="uint8")
        assert len(x.shape) == 3  # [w, h, c]
        assert x.shape[-1] == 3  # RGB, 3 channels
        del x
    except:
        print(f"  |--> Invalid file, moving into: {dest_folder}")
        shutil.move(path, dest_folder)
        skipped += 1

[LeoXing1996]

This is a quite useful script for user-defined datasets and you are welcome to create a PR about this.

However, I suggest this checking operation should be based on the config file. Because whether an image is valid or not is related to its loading pipeline. For example, we may allow images with single-channel (grayscale image) or 4 channels (RGBA) in some config.

One more thing, you may run the following command to install the pre-commit hooks before committing:

pre-commit install

[gvalvano]

Good idea :) I'll be happy to do so. How can I check how many channels should the input have from the config file?

[LeoXing1996]

Maybe we can directly build the loading pipeline from the config.

[zengyh1900]

@LeoXing1996 Has this issue been supported in PRs? Please consider to help @gvalvano to complete related PRs.