DANN-Plant-Counting

Reimplementation of Unsupervised Domain Adaptation For Plant Organ Counting (UDA4POC)

Re-implementation Report

References:

• Original implementation: https://github.com/p2irc/UDA4POC

• Paper link: https://arxiv.org/abs/2009.01081

This project is a reproduction of the paper Unsupervised Domain Adaptation For Plant Organs Counting. The paper proposes a domain-adversarial learning approach for domain adaptation in object counting contexts levering the domain-adversarial neural network (DANN) framework.

Training

Datataset

• Download the CVPPP and Komatsuna Datasets: https://drive.google.com/drive/folders/1Huv7SBmLypiISoDFKWnZd_nd9w5EXYw8?usp=sharing
• Create a new /data folder
• Move the CVPPP pickle file to /data/CVPPP folder
• Move the Komatsuna pickle file to /data/KOMATSUNA folder

Configurations

• The configurations for the trainer and the model are listed in the config.json file
• Default parameters: train for 30 epochs, monitor min validation loss for early stopping, saving every 2 epochs. Below is the list of configuration parameters:

{
    "name": "Counting_Adapt",
    "n_gpu": 1,
    "UNET_ADAPT_arch": {
        "type": "CountAdapt",
        "args": {}
    },
    "data_loader_CVPPP": {
        "type": "CVPPP_DataLoader",
        "args": {
            "data_dir": "data/",
            "batch_size": 2,
            "shuffle": true,
            "validation_split": 0.1,
            "num_workers": 2
        }
    },
    "data_loader_KOMATSUNA": {
        "type": "KOMATSUNA_DataLoader",
        "args": {
            "data_dir": "data/",
            "batch_size": 2,
            "shuffle": true,
            "validation_split": 0.1,
            "num_workers": 2
        }
    },
    "optimizer_CVPPP": {
        "type": "Adam",
        "args": {
            "lr": 0.001,
            "weight_decay": 0,
            "amsgrad": true
        }
    },
    "density_loss": "mse_loss",
    "class_loss": "nll_loss",
    "domain_loss": "bce_loss",
    "metrics": [],
    "lr_scheduler": {
        "type": "StepLR",
        "args": {
            "step_size": 10,
            "gamma": 0.1
        }
    },
    "trainer": {
        "epochs": 30,
        "save_dir": "saved/",
        "save_period": 2,
        "verbosity": 2,
        "monitor": "min val_loss",
        "early_stop": 10,
        "tensorboard": true
    }
}

• Run the following command to start training

python3 train.py -c config.json

Results

These are the results for the CVPPP to KOMATSUNA Adaptation:

Adaptation Method	DiC	|DiC|	%	MSE
No Adaptation-from article	4.09	4.09	0	18.49
With Adaptation - Giuffrida, et al.	-0.78	1.04	26	1.84
With Adaptation - from article	-0.95	1.56	29.33	5.26
With Adaptation - from code	0.69	-0.07	16.67	5.13

This is a sample of training image and its dot labels:

Acknowledgements:

The code structure of this repository follow the PyTorch Template Project