musco-ai / musco-pytorch Goto Github PK

Relying on copy.deepcopy to copy the model in:

• compression_conf_fc.ipynb
• compress.py

Does not work for Faster R-CNN models which use mixed serialization & non-script modules.

This results in the error: _pickle.PickleError: ScriptModules cannot be saved using torch.save. Mixed serialization of script and non-script modules is not supported. For purely script modules use my_script_module.save(<filename>) instead.

Refer to PyTorch issue 18106.

What Faster R-CNN Resnet50 model were you using for this to work? I have been trying with the torchvision pretrained model with no success: torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)

Recently, Pip changed its internal implementation.
So in the last pip versions, I met error with setup.py

it can be debugged by replacing

return [str(item.req) for item in requirements]

by

try:
    return [str(item.req) for item in requirements]
except:
    return [str(item.requirement) for item in requirements]

for compatible with both the versions.

source:
https://stackoverflow.com/questions/62114945/attributeerror-parsedrequirement-object-has-no-attribute-req

Hi,
could ResNet-18 on ILSVRC-12 validation dataset in 6.2 of your paper be released, which experience only 0.47 percent top-1 accuracy drop while achieving 2.42 flops reduction.

when I run the compressed VGG16 model, the runtime is even longer than VGG16 before compressed. how to run the compressed models getting the same results (Table 5) in the paper https://arxiv.org/abs/1903.09973?

Hi, users are unable to run musco-pytorch due to dependency conflict with numpy package. As shown in the following full dependency graph of musco-pytorch, musco-pytorch requires numpy *，while scikit-tensor-py3 requires numpy ==1.16.*.

According to pip’s “first found wins” installation strategy, numpy 1.17.4 is the actually installed version. However, numpy 1.17.4 does not satisfy ==1.16.*.

Dependency tree-----------

musco-pytorch - 1.0.3
| +- absl-py(install version:0.8.1 version range:*)
| +- flopco-pytorch(install version:0.1.3 version range:*)
| +- numpy(install version:1.17.4 version range:*)
| +- scikit-tensor-py3(install version:0.4.1 version range:*)
| | +- numpy(install version:1.16.5 version range:==1.16.*)
| | +- scipy(install version:1.3.3 version range:==1.3.)
| +- scipy(install version:1.3.3 version range:*)
| +- tensorly-musco(install version:0.4.5 version range:*)

Thanks for your help.
Best,
Neolith

Hello , have you ever tried to compress not only backbone but also FPN ?

Hi author! Thx for ur sharing!

I was just trying your iterative compression algorithms using vbmf for compressing the faster rcnn model (exactly the same code mentioned in your paper), but i found it great difficulty doing the fine-tuning work. The more layers I compressed, the less mAP it achieved. Finally, it is approximately 8~10 points lost, which is far below your performance.

Can u tell me how I should do the fine-tuning part better? (like dataset, lr, epoch, etc.) Or can u tell me some of your opinions in terms of it? Thank u!

Like I already mentioned in Issue 13, the demo code seems to crash with an error.

from torchvision.models import resnet50
from flopco import FlopCo
from musco.pytorch import CompressorVBMF, CompressorPR, CompressorManual

model = resnet50(pretrained = True)
model.cuda()
model_stats = FlopCo(model, device = 'cuda')

compressor = CompressorVBMF(model,
                            model_stats,
                            ft_every=5, 
                            nglobal_compress_iters=2)
while not compressor.done:
    compressor.compression_step()
compressed_model = compressor.compressed_model

~/anaconda3/lib/python3.8/site-packages/numpy/linalg/linalg.py in _raise_linalgerror_svd_nonconvergence(err, flag)
    104 
    105 def _raise_linalgerror_svd_nonconvergence(err, flag):
--> 106     raise LinAlgError("SVD did not converge")
    107 
    108 def _raise_linalgerror_lstsq(err, flag):

LinAlgError: SVD did not converge

or

~/anaconda3/lib/python3.8/site-packages/numpy/lib/function_base.py in asarray_chkfinite(a, dtype, order)
    495     a = asarray(a, dtype=dtype, order=order)
    496     if a.dtype.char in typecodes['AllFloat'] and not np.isfinite(a).all():
--> 497         raise ValueError(
    498             "array must not contain infs or NaNs")
    499     return a

ValueError: array must not contain infs or NaNs

The output seems to be random and one of both, if code gets run multiple times.

I've tried to compress my resnet18 and failed. SVDDecomposedConvLayer has no min_rank field. I'm getting exception.

Anaconda3\envs\musco\lib\site-packages\musco\pytorch\compressor\decompositions\svd_layer.py in __init__(self, layer, layer_name, rank_selection, rank, pretrained, vbmf_weaken_factor, param_reduction_rate)
    157 
    158         if rank_selection == 'vbmf':
--> 159             self.rank = estimate_vbmf_ranks(self.weight, vbmf_weaken_factor, min_rank = self.min_rank)
    160         elif rank_selection == 'manual':
    161             self.rank = rank

AttributeError: 'SVDDecomposedConvLayer' object has no attribute 'min_rank'

because of 6e4c947