Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/Dhruv2012/Autonomous-Farm-Robot/llms.txt

Use this file to discover all available pages before exploring further.

Model training is orchestrated by main.py, which selects the model architecture and dataset, builds data generators, configures callbacks, and runs training with weighted categorical cross-entropy loss to handle class imbalance between weed, crop, and soil pixels. The script also supports reloading pre-trained weights to resume training or perform evaluation without restarting from scratch.
Update all hardcoded paths in main.py to match your local dataset locations before running training.

​
Training Configuration

The following variables at the top of main.py control the training run: 
reload_weights = True         # Set False to train from scratch

training_model = {"unet": 0, "bonnet": 1}
dataset = {"cwfid": 0, "bonirob": 1}

MODEL          = training_model["bonnet"]   # or "unet"
DATASET        = dataset["bonirob"]         # or "cwfid"
NO_OF_EPOCHS   = 1
BATCH_SIZE     = 4
results_dir    = "/path/to/results/"
save_weights_path = results_dir + "v3.h5"
When reload_weights = True, the model is initialised with architecture weights and then load_weights() is called with save_weights_path. Set it to False to start a fresh training run.

​
Dataset Paths

​
BoniRob

train_img_path  = '/path/to/BoniRob/input_img_preprocessed/train'
train_mask_path = '/path/to/BoniRob/output_img_preprocessed/train'

val_img_path    = '/path/to/BoniRob/input_img_preprocessed/val'
val_mask_path   = '/path/to/BoniRob/output_img_preprocessed/val'

test_img_path   = '/path/to/BoniRob/input_img_preprocessed/test'
test_mask_path  = '/path/to/BoniRob/output_img_preprocessed/test'
For BoniRob, main.py uses data_gen() from utils.py — a Python generator that yields batches of 10-channel input arrays and flattened one-hot masks on demand, keeping memory usage constant regardless of dataset size.

​
CWFID

path_x    = '/path/to/cwfid/WithAnnotations/images'
path_y    = '/path/to/cwfid/WithAnnotations/annotations'
path_yaml = '/path/to/cwfid/WithAnnotations/train_test_split.yaml'
For CWFID, load_cwfid_withyaml() reads the YAML split, loads all images into memory, and returns NumPy arrays directly — acceptable for this smaller dataset.

​
Loss Function

Weed pixels represent a small fraction of most farm images. Plain categorical cross-entropy would therefore be dominated by the soil class. AGRIBOT uses weighted categorical cross-entropy defined in utils.py to upweight the rare weed class: 
# Class weights for BoniRob dataset (weed is heavily upweighted)
class_weights = [0.90, 0.11, 0.1]   # [weed, crop, soil]

wcce = weighted_categorical_crossentropy(class_weights)

seg_model.compile(
    loss=wcce,
    optimizer='Adam',
    metrics=[
        'accuracy',
        iou_coef,
        keras_metrics.categorical_precision(label=0),
        keras_metrics.categorical_recall(label=0),
        keras_metrics.categorical_precision(label=1),
        keras_metrics.categorical_recall(label=1),
        keras_metrics.categorical_precision(label=2),
        keras_metrics.categorical_recall(label=2),
    ]
)
The iou_coef metric computes mean Intersection-over-Union across the batch, giving a more spatially meaningful measure of segmentation quality than pixel accuracy alone. For CWFID, use class_weights = [0.15, 0.75, 0.10] instead, reflecting the different class distribution of carrot crop imagery.

​
Callbacks

checkpoint = ModelCheckpoint(
    results_dir + 'checkpoints/bonnet_{epoch:03d}_{val_loss:.2f}.hdf5',
    monitor='val_loss',
    save_best_only=True,
    save_weights_only=True,
    mode='min',
    period=20
)

csv_logger  = CSVLogger(results_dir + 'training.csv', append=True)

tensorboard = TensorBoard(log_dir=results_dir + 'graphs')

stopping    = EarlyStopping(
    monitor='val_loss',
    min_delta=0.003,
    patience=1,
    restore_best_weights=True
)
CallbackPurpose
ModelCheckpointSaves weights every 20 epochs when validation loss improves
CSVLoggerAppends per-epoch metrics to training.csv for offline analysis
TensorBoardWrites event files to graphs/ for loss and accuracy visualisation
EarlyStoppingHalts training if val_loss does not improve by ≥ 0.003 for one epoch
Add callbacks to callbacks_list before calling fit_generator to activate them during training.

​
Running Training

cd "Autonomous-Farm-Robot/Crop_Weed_ Classification/"
python3 main.py
For BoniRob training, uncomment the fit_generator line and set the appropriate steps_per_epoch and validation_steps values: 
history = seg_model.fit_generator(
    train_gen,
    epochs=NO_OF_EPOCHS,
    steps_per_epoch=(NO_OF_TRAINING_IMAGES // BATCH_SIZE),
    validation_data=val_gen,
    validation_steps=(NO_OF_VAL_IMAGES // BATCH_SIZE),
    callbacks=callbacks_list
)

​
Evaluation Only

To evaluate an already-trained model without re-running training, keep reload_weights = True and comment out the fit_generator call. The script will run evaluate_generator on the test split and print per-class metrics: 
Loss: 0.xxxx
Mean Accuracy: 0.xxxx
Mean iou 0.xxxx
Precision (label 0): 0.xx Recall (label 0): 0.xx
Precision (label 1): 0.xx Recall (label 1): 0.xx
Precision (label 2): 0.xx Recall (label 2): 0.xx
The full numeric results for the published Bonnet model are available in Documents/readme-images/bonnet-metrics.png.

Build docs developers (and LLMs) love

Get started for freeTalk to us