/DSPS24

Primary LanguageJupyter Notebook

DSPS24 Student Competition

TEAM JAPA

DSPS 24

"""This is the third student competition on the application of AI for pavement condition monitoring.
Participants will use novel machine learning algorithms to predict the pavement condition index (PCI)
for road sections based on images captured from infrastructure - mounted sensors. 
Training datasets provided consists of Top-down views of pavement image data and corresponding pavement condition indices. 
Participants are free to annotate training datasets and use any model architecture to predict the PCI
of the road section."""

In this competion, we ensemble 3 deep learning models using percentile-based aggregation method.

The Three models ensembled:
ResNet50
ResNet101
YOLOv8l-cls

To train the ResNet models use this notebook
Train the YOLOv8l-cls model using this notebook

Model Checkpoints :

Download the model checkpoints from the following links:

Model Checkpoint
ResNet50 model
ResNet101 model checkpoint
YOLOv8l-cls checkpoint

You can downlaod the model checkpoins using gdown

!gdown --id '1Jk10bgNx9w4FoJJDi-F2nS6kUhRR_Iv3' #ResNet50
!gdown --id '1m-DWqJTdERL_G9M1nRbbVxav8cokTb2a' # ResNet101
!gdown --id '1q9hR1XHXMjwb68VOOM83ZZBYNnOj2MvR' #yolov8-cls

To load any if the pretrained ResNet models:
E.g: To load ResNet50 : 

import torch
import torchvision.models as models
# Load the checkpoint
PATH_TO_MODEL_CHECKPOINT = '<path_to_downloaded_model_checkpoint>'
model = models.resnet50(weights=True)
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, 1)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
checkpoint = torch.load(PATH_TO_MODEL_CHECKPOINT)

model.load_state_dict(checkpoint['model_state_dict'])
model = model.to(device)

To load a YOLO model

  • Install ultralytics using pip e.g. pip install ultralytics
from ultralytics import YOLO
PATH_TO_MODEL_CHECKPOINT = '<path_to_downloaded_model_checkpoint>'
model = YOLO(PATH_TO_MODEL_CHECKPOINT)

To generate our final submission.json file for the competition :
Run the cells in the Inference Notebook