mightydeveloper / deep-compression-pytorch Goto Github PK

Hello, mightydeveloper.
When I use 'weight_share.py' to compress the trained model, the error occured:
AttributeError: 'ReLU' object has no attribute 'weight' .
File "weight_share.py", line 32, in
apply_weight_sharing(model)
File "/net/quantization.py", line 16, in apply_weight_sharing
dev = module.weight.device
File "/usr/local/lib/python3.5/dist-packages/torch/nn/modules/module.py", line 539, in getattr
type(self).name, name))
AttributeError: 'ReLU' object has no attribute 'weight'
How to jump over the ReLU module?
THANKS.

Could you please mention the pytorch version this repo supports.?

i try weight sharing and huffman encode for yolov7 but it both have error:

TypeError:** expected dimension <= 2 array or matrix

Turn out coo matrix in scipy only support weight maxtrix with shape <=2. So it not flexible. Any fix for that ???.Thank you

When i run the pruning.py, the following bug was raised, what's the possible problem?
RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.

    This probably means that you are not using fork to start your
    child processes and you have forgotten to use the proper idiom
    in the main module:

        if __name__ == '__main__':
            freeze_support()
            ...

    The "freeze_support()" line can be omitted if the program
    is not going to be frozen to produce an executable.

when the num_worker of Dataloader is not zero, there would be an error. some people say it is because the multi-threads of windows is accomplished with spawn instead of fork. I am not familiar with computer science, could you please tell me what should I do to avoid this error while still getting the benifits of multi-threads?

@mightydeveloper Hi, I used the quantization script in my model .Because of the existing of convolution layer,I encountered an error named as "TypeError: expected dimension <= 2 array or matrix" in using csc_matrix . How to fix the problem? can you give me some advice ? Thank you!

Why after i use the three way, the model size is same of the three: @mightydeveloper

-rw-r--r--. 1 root root 2.1M Oct 10 02:17 initial_model.ptmodel
-rw-r--r--. 1 root root 2.1M Oct 10 02:46 model_after_retraining.ptmodel
-rw-r--r--. 1 root root 2.1M Oct 10 03:04 model_after_weight_sharing.ptmodel

When I use your`s Deep compression functin in MobileNet-V2 Model，I find some problem,

1. I need use kmeans in every weight among Layer, but the weight is different dimensions
2. When i use prune in MobileNet , I should use prune in every fc layer?
   @mightydeveloper

after prune the all parameter is 112422,when use huffman_encode,in change to 899388?? @mightydeveloper

--- After Retraining ---
fc1.weight           | nonzeros =  107208 /  235200 ( 45.58%) | total_pruned =  127992 | shape = (300, 784)
fc1.bias             | nonzeros =     300 /     300 (100.00%) | total_pruned =       0 | shape = (300,)
fc2.weight           | nonzeros =    4338 /   30000 ( 14.46%) | total_pruned =   25662 | shape = (100, 300)
fc2.bias             | nonzeros =     100 /     100 (100.00%) | total_pruned =       0 | shape = (100,)
fc3.weight           | nonzeros =     466 /    1000 ( 46.60%) | total_pruned =     534 | shape = (10, 100)
fc3.bias             | nonzeros =      10 /      10 (100.00%) | total_pruned =       0 | shape = (10,)
alive: 112422, pruned : 154188, total: 266610, Compression rate :       2.37x  ( 57.83% pruned)

Layer           |   original compressed improvement percent
----------------------------------------------------------------------
fc1.weight      |     858868     176999       4.85x  20.61%
fc1.bias        |       1200       1200       1.00x 100.00%
fc2.weight      |      35108       7696       4.56x  21.92%
fc2.bias        |        400        400       1.00x 100.00%
fc3.weight      |       3772       1102       3.42x  29.22%
fc3.bias        |         40         40       1.00x 100.00%
----------------------------------------------------------------------
total           |     899388     187437       4.80x  20.84%

I'm Trying to apply the whole compression process on LeNet5 instead of LeNet300-100
I Fixed some problems I Encountered but now in the quantization step, I can't use sparse matrices because the shape of the first layer for example is considered with the kernel size, so I'm facing the following error:
TypeError: expected dimension <= 2 array or matrix
Is there a way to fix this problem?

Please tell me what should I do if I get the error KMeans.init() got an unexpected keyword argument 'precompute_distances' when I run weight_shared.py?

Hey!
Thanks for this implementation! :)
Do you have any idea as to how we can apply Huffman encoding on darknet .weights??

@mightydeveloper hi thanks for the wonderful code base , i have following few queries
1.Can we reduce the weight size of the model with the following code base
2. can your code base be applied to custom architecture