pytorch-op-counter-layerwise

This repository provides thoplw that is a ython module to compute MACs (multiply–accumulate operations) and number of parameters for each layer of neural network models implemented by PyTorch.

Installation

Dependencies

The thoplw module requires:

• PyTorch >= 2.0.0 (will work on the older version, but not checked)

and the sample code require:

• Torchvision >= 0.15.0

User installation

pip install thoplw

Usage

Minimal example

import torch, torchvision, thoplw

# Instanciate the target model.
model = torchvision.models.resnet18()

# Compute MACs, number of parameters, and details of each layer.
macs, params, layers_info = thoplw.profile(model, tensor=torch.randn(1, 3, 224, 224))

# Print the total MACs and number of parameters.
print("Total MACs and params:")
print("  - Macs   =", macs)
print("  - Params =", params)
print()

# Print layer details.
print(layers_info.summary())

Running the above code will give you the output below (the table has been partially omitted because it is too long to show everything in this README).

Total MACs and params:
  - Macs   = 1824010216                                                                                                                                                       
  - Params = 11699112

| Name                  | Class             | Input shape    | Output shape   | MACs       | Params   |
+-----------------------+-------------------+----------------+----------------+------------+----------+
| conv1                 | Conv2d            | 3 x 224 x 224  | 64 x 112 x 112 | 118013952  | 9408     |
| bn1                   | BatchNorm2d       | 64 x 112 x 112 | 64 x 112 x 112 | 3211264    | 256      |
| relu                  | ReLU              | 64 x 112 x 112 | 64 x 112 x 112 | 0          | 0        |
| maxpool               | MaxPool2d         | 64 x 112 x 112 | 64 x 56 x 56   | 0          | 0        |
...
| layer4.1.conv2        | Conv2d            | 512 x 7 x 7    | 512 x 7 x 7    | 115605504  | 2359296  |
| layer4.1.bn2          | BatchNorm2d       | 512 x 7 x 7    | 512 x 7 x 7    | 100352     | 2048     |
| avgpool               | AdaptiveAvgPool2d | 512 x 7 x 7    | 512 x 1 x 1    | 1024       | 0        |
| fc                    | Linear            | 512            | 1000           | 513000     | 513000   |
+-----------------------+-------------------+----------------+----------------+------------+----------+
| Total                 | ResNet            | 3 x 224 x 224  | 1000           | 1824010216 | 11699112 |

Clever formatting

The thoplw provides clever_format function that returns appropriate expressions of the given numbers likewise thop package.

macs, params = thoplw.clever_format([macs, params], "%.3f")
print("Total MACs and params:")
print("  - Macs   =", macs)
print("  - Params =", params)

The table of layers detail supports three types of formatting, raw number (default choice), clever format like the clever_format function, and ratio.

# Print the table with clever formatting.
print(layers_info.summary(kind="text", fmt="clever"))

# Print the table with ratio formatting.
print(layers_info.summary(kind="text", fmt="ratio"))

Table type

The above example prints the NN model summary as a plain text, but you can dump the summary as CSV and Markdown format too. The following is an example to save the table as CSV and Markdown format respectively.

# Save as CSV format.
with open("summary.csv") as ofp:
    ofp.write(layers_info.summary(kind="csv"))

# Save as Markdown format.
with open("summary.md") as ofp:
    ofp.write(layers_info.summary(kind="md"))

API reference

thoplw.profile

macs, params, layers_info = thoplw.profile(model: torch.nn.Module,
                                           tensor: torch.Tensor,
                                           custom_ops: dict = {},
                                           verbose: bool = True)

Computes MADDs and the number of parameters.

• Args
  • model: the target NN model.
  • tensor: input tensor for the model.
  • custom_ops: optional custom operations.
  • verbose: shows extra message on your terminal if True.
• Returns
  • macs: total MADDs of the target model and the given input tensor.
  • params: number of parameters of the target model.
  • layer_info: LayerInfo class that store the details of each layer.

thoplw.clever_format

formatted_values = thoplw.clever_format(values: int or list, fmt: str = "%6.2f")

Returns formatted string of the given integer(s).

• Args
  • values: input value, or values.
  • fmt: format specifier.
• Returns
  • formatted_values: format result.

LayerInfo class

class LayerInfo:
    ...
    def summary(self,
                kind: str = "text",
                fmt: str = "raw") -> str:
    ...

A class to store layer details. Only the summary function is opened to users.

• Args
  • kind: table type to be returned ("text" means simple table, "csv" means CSV, and "md" means Markdown).
  • fmt: output format ("raw" means raw integer, "clever" means auto formatting, and "ratio" means ratio format).
• Returns
  • formatted string.

Results of Recent Models

The following results can be obtained by running tests/test_benchmarks.py. Click the model anem to see the layer details.

Gratitude

• Developers and maintainers of pytorch-OpCounter. The author learned a lot from the repository.