CFNet

Implementation from paper "Co-Occurrent Features in Semantic Segmentation"

Paper

• Link: https://openaccess.thecvf.com/content_CVPR_2019/papers/Zhang_Co-Occurrent_Features_in_Semantic_Segmentation_CVPR_2019_paper.pdf

Main Library Features

• High Level API
• ResNet101 backbones
• Models can be used as Subclassed or Functional Model
• Implement like paper code, adding multi head attention and mixtures of softmax to CFNet with source code from TASM

Installation and Setup

Requirements Windows or Linus

1. Python 3.6 or higher
2. tensorflow >= 2.3.0 (>= 2.0.0 is sufficient if no efficientnet backbone is used)
3. numpy

Clone Repository

$ git clone https://github.com/huy105/CFNet.git

or directly install it:
Pip Install Repository

$ pip install git+https://github.com/huy105/CFNet.git

Training Pipeline

Please check that Tensorflow is installed on your computer.

To import the model just use the standard python import statement

Firstly, import function to get base model (backbone) and main head (CFNet):

from backbone import create_base_model
from main import CFNet

Then, get the backbone model, output layers, height and width:

BACKBONE_NAME = "resnet101"
WEIGHTS = "imagenet"
HEIGHT = 224
WIDTH = 224

model, layers_outputs, layer_names = create_base_model(name= BACKBONE_NAME, weights= WEIGHTS, height= HEIGHT, width= WIDTH, channels=3)

After that, create full model (base model + head) with parameter you want

CFnet_model = CFNet(n_classes = 5, base_model = model, output_layers = layers_outputs, n_heads=2, n_mix = 4,backbone_trainable = True)

If you want to use the Functional Model class define instead:

CFnet_model = CFNet(n_classes = 5, base_model = model, output_layers = layers_outputs, n_heads=2, 
n_mix = 4,backbone_trainable = True,  height= HEIGHT, width= WIDTH).model()

References

Using custom layers and main code from TensorFlow Advanced Segmentation Models (TASM).

• TensorFlow Advanced Segmentation Models, GitHub, GitHub Repository, https://github.com/JanMarcelKezmann/TensorFlow-Advanced-Segmentation-Models
• Code from author of paper, GitHub, GitHub Repository, https://github.com/zhanghang1989/PyTorch-Encoding