ResNet50 - Computing Sparsity

grid_world · May 11, 2021, 5:53pm

I have trained a ResNet-50 model on CIFAR-10 using transfer learning with some modifications. After which, I am using torch.nn.utils.prune to prune weights using global magnitude pruning in an iterative manner. The model architecture is as follows:

print(fine_tuned_model)
ResNet(
  (conv1): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
  (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (relu): ReLU(inplace=True)
  (maxpool): MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
  (layer1): Sequential(
    (0): Bottleneck(
      (conv1): Conv2d(64, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (downsample): Sequential(
        (0): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): Bottleneck(
      (conv1): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
    (2): Bottleneck(
      (conv1): Conv2d(256, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(64, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
  )
  (layer2): Sequential(
    (0): Bottleneck(
      (conv1): Conv2d(256, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (downsample): Sequential(
        (0): Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): Bottleneck(
      (conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
    (2): Bottleneck(
      (conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
    (3): Bottleneck(
      (conv1): Conv2d(512, 128, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(128, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
  )
  (layer3): Sequential(
    (0): Bottleneck(
      (conv1): Conv2d(512, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (downsample): Sequential(
        (0): Conv2d(512, 1024, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): Bottleneck(
      (conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
    (2): Bottleneck(
      (conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
    (3): Bottleneck(
      (conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
    (4): Bottleneck(
      (conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
    (5): Bottleneck(
      (conv1): Conv2d(1024, 256, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(256, 1024, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(1024, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
  )
  (layer4): Sequential(
    (0): Bottleneck(
      (conv1): Conv2d(1024, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
      (downsample): Sequential(
        (0): Conv2d(1024, 2048, kernel_size=(1, 1), stride=(2, 2), bias=False)
        (1): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
    )
    (1): Bottleneck(
      (conv1): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
    (2): Bottleneck(
      (conv1): Conv2d(2048, 512, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn1): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv2): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
      (bn2): BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (conv3): Conv2d(512, 2048, kernel_size=(1, 1), stride=(1, 1), bias=False)
      (bn3): BatchNorm2d(2048, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      (relu): ReLU(inplace=True)
    )
  )
  (avgpool): AdaptiveAvgPool2d(output_size=(1, 1))
  (fc): Linear(in_features=2048, out_features=10, bias=True)
)

To count the number of pruned connections (or, number of zeroes), number of total connections and the resulting sparsity, the following two functions are defined:

def measure_module_sparsity(module, weight = True, bias = False, use_mask = False):

    num_zeros = 0
    num_elements = 0

    if use_mask == True:
        for buffer_name, buffer in module.named_buffers():
            if "weight_mask" in buffer_name and weight == True:
                num_zeros += torch.sum(buffer == 0).item()
                num_elements += buffer.nelement()
            if "bias_mask" in buffer_name and bias == True:
                num_zeros += torch.sum(buffer == 0).item()
                num_elements += buffer.nelement()
    else:
        for param_name, param in module.named_parameters():
            if "weight" in param_name and weight == True:
                num_zeros += torch.sum(param == 0).item()
                num_elements += param.nelement()
            if "bias" in param_name and bias == True:
                num_zeros += torch.sum(param == 0).item()
                num_elements += param.nelement()

    sparsity = num_zeros / num_elements

    return num_zeros, num_elements, sparsity



def measure_global_sparsity(
    model, weight = True,
    bias = False, conv2d_use_mask = False,
    linear_use_mask = False):

    num_zeros = 0
    num_elements = 0

    for module_name, module in model.named_modules():

        if isinstance(module, torch.nn.Conv2d):
            module_num_zeros, module_num_elements, _ = measure_module_sparsity(
                module, weight=weight, bias=bias, use_mask=conv2d_use_mask)
            num_zeros += module_num_zeros
            num_elements += module_num_elements

        elif isinstance(module, torch.nn.Linear):
            module_num_zeros, module_num_elements, _ = measure_module_sparsity(
                module, weight=weight, bias=bias, use_mask=linear_use_mask)
            num_zeros += module_num_zeros
            num_elements += module_num_elements

    sparsity = num_zeros / num_elements

    return num_zeros, num_elements, sparsity

To call the function-

num_zeros, num_elements, sparsity = measure_global_sparsity(
    fine_tuned_model, weight = True,
    bias = False, conv2d_use_mask = True,
    linear_use_mask = False)

this gives the error:

ZeroDivisionError Traceback (most recent call last)
in
----> 1 num_zeros, num_elements, sparsity = measure_global_sparsity(
2 fine_tuned_model, weight = True,
3 bias = False, conv2d_use_mask = True,
4 linear_use_mask = False)
5

in measure_global_sparsity(model, weight, bias, conv2d_use_mask, linear_use_mask)
10
11 if isinstance(module, torch.nn.Conv2d):
—> 12 module_num_zeros, module_num_elements, _ = measure_module_sparsity(
13 module, weight=weight, bias=bias, use_mask=conv2d_use_mask)
14 num_zeros += module_num_zeros

in measure_module_sparsity(module, weight, bias, use_mask)
21 num_elements += param.nelement()
22
—> 23 sparsity = num_zeros / num_elements
24
25 return num_zeros, num_elements, sparsity

ZeroDivisionError: division by zero

What’s the problem OR is there another way to count the: number of zeroes and the total number of parameters within this ResNet-50 model?

randinoo · May 11, 2021, 10:15pm

why you put 64 of conv1??instead of 10 or 20 etc

KFrank · May 12, 2021, 12:22am

Hi Arjun!

grid_world:

def measure_module_sparsity(module, weight = True, bias = False, use_mask = False):

    num_zeros = 0
    num_elements = 0

    if use_mask == True:
        for buffer_name, buffer in module.named_buffers():
            ...
    else:
        for param_name, param in module.named_parameters():
            ...

    sparsity = num_zeros / num_elements

    return num_zeros, num_elements, sparsity

 22
 —> 23 sparsity = num_zeros / num_elements
 24
 25     return num_zeros, num_elements, sparsity

ZeroDivisionError: division by zero

What’s the problem … ?

If you call measure_module_sparsity() on a Module such as ReLU
that contains neither buffers nor parameters, then num_elements (and
num_zeros) will remain zero (because they never get incremented)
so you get the ZeroDivisionError.

Best.

K. Frank

grid_world · May 12, 2021, 2:42am

I am referring to ResNet paper by Kaiming He et al. where in Table 1, the first conv layer has 64 filters in it. I just changed filter size from 7 to 3 and stride from 2 to 1.

Probably, removing a max pool layer might also make sense since CIFAR-10 dataset has small images to begin with which don’t need spatial reduction right after the first conv layer.

grid_world · May 12, 2021, 2:48am

I don’t think that “measure_module_sparsity()” function will get called on a module such as “ReLU” because within “measure_global_sparsity()” function, the for loop:

for module_name, module in model.named_modules():
    if isinstance(module, torch.nn.Conv2d):
        module_num_zeros, module_num_elements, _ = measure_module_sparsity(
                module, weight=weight, bias=bias, use_mask=conv2d_use_mask)
        num_zeros += module_num_zeros
        num_elements += module_num_elements

will make sure that “measure_module_sparsity()” is only called for Conv2d and/or Linear layers.

At least, that’s what my understanding of the code is.

KFrank · May 12, 2021, 3:40pm

Hi Arjun!

Yes, you are correct. My mistake.

All I can say is that your code works for me (on a stock torchvision
resnet50).

It is possible to instantiate a degenerate Linear (or Conv2d) that has
no elements in its parameters, but that would be weird, and doesn’t
happen in torchvision’s resnet50 and does not appear to be happening
in your model.

All I can suggest is that you check (and tell us) what version of pytorch
you are using, and instrument your code so that it prints out the name
(and details) of any module for which num_elements is zero before
performing the division.

Here is a script that runs your function on a stock resnet50 and also
illustrates a degenerate Linear:

import torch
print (torch.__version__)

import torchvision
print (torchvision.__version__)

_ = torch.manual_seed (2021)

def measure_module_sparsity(module, weight = True, bias = False, use_mask = False):
    
    num_zeros = 0
    num_elements = 0
    
    if use_mask == True:
        for buffer_name, buffer in module.named_buffers():
            if "weight_mask" in buffer_name and weight == True:
                num_zeros += torch.sum(buffer == 0).item()
                num_elements += buffer.nelement()
            if "bias_mask" in buffer_name and bias == True:
                num_zeros += torch.sum(buffer == 0).item()
                num_elements += buffer.nelement()
    else:
        for param_name, param in module.named_parameters():
            if "weight" in param_name and weight == True:
                num_zeros += torch.sum(param == 0).item()
                num_elements += param.nelement()
            if "bias" in param_name and bias == True:
                num_zeros += torch.sum(param == 0).item()
                num_elements += param.nelement()
    
    sparsity = num_zeros / num_elements
    
    return num_zeros, num_elements, sparsity


def measure_global_sparsity(
    model, weight = True,
    bias = False, conv2d_use_mask = False,
    linear_use_mask = False):
    
    num_zeros = 0
    num_elements = 0
    
    for module_name, module in model.named_modules():
        
        if isinstance(module, torch.nn.Conv2d):
            module_num_zeros, module_num_elements, _ = measure_module_sparsity(
                module, weight=weight, bias=bias, use_mask=conv2d_use_mask)
            num_zeros += module_num_zeros
            num_elements += module_num_elements
        
        elif isinstance(module, torch.nn.Linear):
            module_num_zeros, module_num_elements, _ = measure_module_sparsity(
                module, weight=weight, bias=bias, use_mask=linear_use_mask)
            num_zeros += module_num_zeros
            num_elements += module_num_elements
    
    sparsity = num_zeros / num_elements
    
    return num_zeros, num_elements, sparsity

mod = torchvision.models.resnet50()

print ('type (mod) =', type (mod))
print ('measure_global_sparsity (mod) =', measure_global_sparsity (mod))

lin = torch.nn.Linear (1, 0)
print ('lin =', lin)
print ('calling measure_global_sparsity (lin) ...')
measure_global_sparsity (lin)

And here is its output:

1.9.0.dev20210504
0.10.0.dev20210504
type (mod) = <class 'torchvision.models.resnet.ResNet'>
measure_global_sparsity (mod) = (4, 25502912, 1.5684483403307042e-07)
lin = Linear(in_features=1, out_features=0, bias=True)
calling measure_global_sparsity (lin) ...
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "<string>", line 70, in <module>
  File "<string>", line 53, in measure_global_sparsity
  File "<string>", line 31, in measure_module_sparsity
ZeroDivisionError: division by zero

Best.

K. Frank

grid_world · May 12, 2021, 3:59pm

PyTorch Version: 1.8.1
Torchvision Version: 0.9.1

Good idea. Let me try your suggestion and get back.

I am using Global, unstructured, absolute magnitude & iterative pruning by using “torch.nn.utils.prune” for which computing global sparsity is important. The algo used involves using a trained model as follows:

prune p% of smallest magnitude weights globally (p = 20%)
Fine-tune for n epochs to recover from the damage caused due to pruning