AttributeError: 'dict' object has no attribute 'bboxes' #622
Labels
question
Further information is requested
Comments
|
We are sorry for your inconvenience. We have improved the ChannelAnayler to adaptive yolov8. You can directly cherry-pick this pr or wait util it is merged. |
Thank you very much for your previous response, and I have successfully run the train.py. And may I ask how to export the pruned compact model after finishing the pruning training? |
|
DCFF is developing, so there is no existing export method. We suggest you use group fisher algorithm that is developed well including export (and deploy is incomming soon). |
Thanks for your reply, I'll try it out. |
When I try to use group fisher, it cames: Traceback (most recent call last):
File "tools/train.py", line 121, in <module>
main()
File "tools/train.py", line 117, in main
runner.train()
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/mmengine/runner/runner.py", line 1673, in train
self.call_hook('before_run')
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/mmengine/runner/runner.py", line 1748, in call_hook
getattr(hook, fn_name)(self, **kwargs)
File "/home/keli/Nuonepeaceyy/MMLAB/plus/mmrazor/mmrazor/implementations/pruning/group_fisher/hook.py", line 128, in before_run
original_resource = self._evaluate(model)
File "/home/keli/Nuonepeaceyy/MMLAB/plus/mmrazor/mmrazor/implementations/pruning/group_fisher/hook.py", line 170, in _evaluate
res = self.estimator.estimate(model)
File "/home/keli/Nuonepeaceyy/MMLAB/plus/mmrazor/mmrazor/models/task_modules/estimators/resource_estimator.py", line 136, in estimate
flops, params = get_model_flops_params(model, **flops_params_cfg)
File "/home/keli/Nuonepeaceyy/MMLAB/plus/mmrazor/mmrazor/models/task_modules/estimators/counters/flops_params_counter.py", line 98, in get_model_flops_params
_ = flops_params_model(batch)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1148, in _call_impl
result = forward_call(*input, **kwargs)
File "/home/keli/Nuonepeaceyy/MMLAB/plus/mmrazor/mmrazor/models/algorithms/base.py", line 144, in forward
return self._forward(inputs, data_samples)
File "/home/keli/Nuonepeaceyy/MMLAB/plus/mmrazor/mmrazor/models/algorithms/base.py", line 165, in _forward
return self.architecture(inputs, data_samples, mode='tensor')
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/mmdet/models/detectors/base.py", line 96, in forward
return self._forward(inputs, data_samples)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/mmdet/models/detectors/single_stage.py", line 132, in _forward
x = self.extract_feat(batch_inputs)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/mmdet/models/detectors/single_stage.py", line 146, in extract_feat
x = self.backbone(batch_inputs)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/home/keli/Nuonepeaceyy/MMLAB/plus/mmyolo/mmyolo/models/backbones/base_backbone.py", line 221, in forward
x = layer(x)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/torch/nn/modules/container.py", line 139, in forward
input = module(input)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/mmcv/cnn/bricks/conv_module.py", line 207, in forward
x = self.conv(x)
File "/home/keli/anaconda3/envs/mmyolo/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1151, in _call_impl
hook_result = hook(self, input, result)
File "/home/keli/Nuonepeaceyy/MMLAB/plus/mmrazor/mmrazor/models/task_modules/estimators/counters/op_counters/conv_layer_counter.py", line 82, in add_count_hook
if hasattr(mutable_channel[0], 'activated_tensor_channels'):
IndexError: list index out of rangeand here is my config file: #############################################################################
"""You have to fill these args.
_base_ (str): The path to your pretrained model checkpoint.
pretrained_path (str): The path to your pretrained model checkpoint.
interval (int): Interval between pruning two channels. You should ensure you
can reach your target pruning ratio when the training ends.
normalization_type (str): GroupFisher uses two methods to normlized the channel
importance, including ['flops','act']. The former uses flops, while the
latter uses the memory occupation of activation feature maps.
lr_ratio (float): Ratio to decrease lr rate. As pruning progress is unstable,
you need to decrease the original lr rate until the pruning training work
steadly without getting nan.
target_flop_ratio (float): The target flop ratio to prune your model.
input_shape (Tuple): input shape to measure the flops.
"""
_base_ = 'mmyolo::yolov8/yolov8_s_syncbn_fast_8xb16-500e_coco.py'
pretrained_path = '../mmyolo/checkpoints/yolov8_s_mask-refine_syncbn_fast_8xb16-500e_coco_20230216_095938-ce3c1b3f.pth'
interval = 10
normalization_type = 'act'
lr_ratio = 0.1 / 4
target_flop_ratio = 0.5
input_shape = (1, 3, 640, 640)
##############################################################################
architecture = _base_.model
if hasattr(_base_, 'data_preprocessor'):
architecture.update({'data_preprocessor': _base_.data_preprocessor})
data_preprocessor = None
architecture.init_cfg = dict(type='Pretrained', checkpoint=pretrained_path)
architecture['_scope_'] = _base_.default_scope
model = dict(
_delete_=True,
_scope_='mmrazor',
type='GroupFisherAlgorithm',
architecture=architecture,
interval=interval,
mutator=dict(
type='GroupFisherChannelMutator',
parse_cfg=dict(type='ChannelAnalyzer', tracer_type='FxTracer'),
channel_unit_cfg=dict(
type='GroupFisherChannelUnit',
default_args=dict(normalization_type=normalization_type, ),
),
),
)
model_wrapper_cfg = dict(
type='mmrazor.GroupFisherDDP',
broadcast_buffers=False,
)
optim_wrapper = dict(
optimizer=dict(lr=_base_.optim_wrapper.optimizer.lr * lr_ratio))
custom_hooks = getattr(_base_, 'custom_hooks', []) + [
dict(type='mmrazor.PruningStructureHook'),
dict(
type='mmrazor.ResourceInfoHook',
interval=interval,
demo_input=dict(
type='mmrazor.DefaultDemoInput',
input_shape=input_shape,
),
save_ckpt_thr=[target_flop_ratio],
),
] |
|
@nuonepeaceyy We have fix the bug, please pull the latest code from dev-1.x branch of mmrazor. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Prerequisite
💬 Describe the reimplementation questions
When I try to use DCFF in mmrazor to prune the yolov8 model, I first try to get the target_pruning_ratio by
python tools/pruning/get_channel_units.py configs/pruning/mmdet/new_dcff/yolov8_pruning.py, but getAttributeError: classDCFFin mmrazor/models/algorithms/pruning/dcff.py: 'dict' object has no attribute 'bboxes'; then I try to runpython tools/train.py configs/pruning/mmdet/new_dcff/yolov8_pruning.pyto train it directly, it still comes to the same error. The detail of this error as follows.and here is my prune config, and I don't change the yolov8 config file in mmyolo.
Environment
Expected results
No response
Additional information
This yolov8 model can be trained normally in mmyolo;
I print the
gt_instanceinmmyolo/mmyolo/models/dense_heads/yolov8_head.py line 423as follows:there is no attribute like
bboxesandlabelswhich they should be in there, but I have no idea about this. This may be because it is not correctly read in the label, but I can train this model in mmyolo...The text was updated successfully, but these errors were encountered: