model-profiler 1.1.8

Tensorflow/Keras Model Profiler: Tells you model's memory requirement, no. of parameters, flops etc.

Tensorflow/ Keras Model Profiler

Gives you some basic but important information about your tf or keras model like,

• Model Parameters
• Model memory requirement on GPU
• Memory required to store parameters model weights.
• GPU availability and GPU IDs if available

Version 1.1.8 fixes problems with custom sequential models and includes other minor improvements.

Dependencies

python >= 3.6
numpy 
tabulate
tensorflow >= 2.0.0
keras >= 2.2.4

Built and tested on tensorflow >= 2.3.1

Installation

using pip.

pip install model_profiler

Usage

Firs load any model built using keras or tensorflow. Here for simplicity we will load model from kera applications.

from tensorflow.keras.applications import EfficientNetB3

model = EfficientNetB3(include_top=True)

Now after installing model_profiler run

from model_profiler import model_profiler

Batch_size = 128
profile = model_profiler(model, Batch_size)

print(profile)

Batch_size have effect on model memory usage so GPU memory usage need batch_size, it's default value if 1.

Output

| Model Profile                    | Value         | Unit    |
|----------------------------------|---------------|---------|
| Selected GPUs                    | ['0', '1']    | GPU IDs |
| No. of FLOPs                     | 1.5651        | BFLOPs  |
| GPU Memory Requirement           | 41.7385       | GB      |
| Model Parameters                 | 12.3205       | Million |
| Memory Required by Model Weights | 46.9991       | MB      |

Using Custom Models

you can also built your own models on top of keras.application a simple example is given below. The usage is almost same the only difference is that for better calculation of FLOPs, it's better to sepcify the input_shape argument. Simple example below.

from tensorflow.keras.applications import EfficientNetB3
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Dropout
from profiler import model_profiler


custom_model = Sequential()
custom_model.add(EfficientNetB3(weights='imagenet', include_top=False,
                                input_shape=(296,256,3), pooling='avg'))

# build you custom layers on top
custom_model.add(Dropout(0.50))
custom_model.add(Dense(1000, activation='softmax'))


profile = model_profiler(custom_model, Batch_size)

Output

| Model Profile                    | Value         | Unit    |
|----------------------------------|---------------|---------|
| Selected GPUs                    | None Detected | GPU IDs |
| No. of FLOPs                     | 1.2868        | BFLOPs  |
| GPU Memory Requirement           | 34.8656       | GB      |
| Model Parameters                 | 10.7989       | Million |
| Memory Required by Model Weights | 41.1945       | MB      |

Units

Default units for the prfiler are

# in order 
use_units = ['GPU IDs', 'BFLOPs', 'GB', 'Million', 'MB']

You can change units by changing the list entry in appropriate location. For example if you want to get model FLOPs in million just change the list as follows.

# keep order 
use_units = ['GPU IDs', 'MFLOPs', 'GB', 'Million', 'MB']

Availabel units are

    'GB':memory unit gega-byte
    'MB': memory unit mega-byte
    'MFLOPs':  FLOPs unit million-flops
    'BFLOPs':  FLOPs unit billion-flops
    'Million': paprmeter count unit millions
    'Billion': paprmeter count unit billions

More Examples

For further details and more examples visit my github