modules/maskrcnn_train.sh

#!/bin/bash

# Parameters
USE_GPU_TUNING=true   # Set this to true to enable GPU tuning
RESUME_TRAINING=false # Set this to true to resume training from last checkpoint
MAX_ITER=100000        # Default maximum iterations
CHECKPOINT_PERIOD=$(($MAX_ITER / 10))  # Set to 10% of MAX_ITER

# Default values
IMS_PER_BATCH=2
BATCH_SIZE_PER_IMAGE=512
BASE_LR=0.00025

# If USE_GPU_TUNING is true, adjust parameters based on GPU memory
if [ "$USE_GPU_TUNING" = true ]; then
    # Get the GPU memory in MB
    GPU_MEMORY=$(nvidia-smi --query-gpu=memory.total --format=csv,noheader,nounits | head -n 1)

    # Set the batch size and learning rate based on GPU memory
    if [ "$GPU_MEMORY" -ge 24576 ]; then
        IMS_PER_BATCH=8   # For 24GB+ GPUs
        BATCH_SIZE_PER_IMAGE=512
    elif [ "$GPU_MEMORY" -ge 12288 ]; then
        IMS_PER_BATCH=4   # For 12GB+ GPUs
        BATCH_SIZE_PER_IMAGE=256
    else
        IMS_PER_BATCH=2   # For smaller GPUs
        BATCH_SIZE_PER_IMAGE=128
    fi

    # Adjust learning rate based on batch size
    BASE_LR=$(echo "0.00025 * $IMS_PER_BATCH / 2" | bc -l)
fi

# Get script dir
SCRIPT_DIR=$(cd $(dirname $0); pwd)

# Set the paths for the dataset and configuration files
TRAIN_ANNOTATION=$SCRIPT_DIR/../export_coco/annotations/train.json
TRAIN_IMAGE_DIR=$SCRIPT_DIR/../export_coco/train
VAL_ANNOTATION=$SCRIPT_DIR/../export_coco/annotations/val.json
VAL_IMAGE_DIR=$SCRIPT_DIR/../export_coco/val
COCO_CONFIG_FILE=$SCRIPT_DIR/../export_coco/config/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml
BASE_CONFIG_FILE=$SCRIPT_DIR/../export_coco/config/Base-RCNN-FPN.yaml  # Base config file path
CONFIG_URL="https://raw.githubusercontent.com/facebookresearch/detectron2/main/configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"
BASE_CONFIG_URL="https://raw.githubusercontent.com/facebookresearch/detectron2/main/configs/Base-RCNN-FPN.yaml"  # Base config URL
TRAIN_NET_DIR=$SCRIPT_DIR/../export_coco/detectron2
TRAIN_NET_FILE=$TRAIN_NET_DIR/train_net.py
TRAIN_NET_URL="https://raw.githubusercontent.com/facebookresearch/detectron2/main/tools/train_net.py"
OUTPUT_DIR=$SCRIPT_DIR/../export_coco/output

# Create necessary directories
mkdir -p $(dirname $COCO_CONFIG_FILE)
mkdir -p $TRAIN_NET_DIR
mkdir -p $OUTPUT_DIR

# Download the COCO config file if it doesn't exist
if [ ! -f "$COCO_CONFIG_FILE" ]; then
    echo "Downloading Mask R-CNN configuration file..."
    wget $CONFIG_URL -O $COCO_CONFIG_FILE
fi

# Download the base config file if it doesn't exist
if [ ! -f "$BASE_CONFIG_FILE" ]; then
    echo "Downloading Base-RCNN-FPN.yaml configuration file..."
    wget $BASE_CONFIG_URL -O $BASE_CONFIG_FILE
fi

# Download train_net.py if it doesn't exist
if [ ! -f "$TRAIN_NET_FILE" ]; then
    echo "Downloading train_net.py file..."
    wget $TRAIN_NET_URL -O $TRAIN_NET_FILE
fi

# Python script to configure and run the training
python3 - <<END
import os
import json
from detectron2.data.datasets import register_coco_instances
from detectron2.data import DatasetCatalog
from detectron2.engine import DefaultTrainer
from detectron2.config import get_cfg

# Paths from Bash script
train_annotation = "$TRAIN_ANNOTATION"
train_image_dir = "$TRAIN_IMAGE_DIR"
val_annotation = "$VAL_ANNOTATION"
val_image_dir = "$VAL_IMAGE_DIR"
resume_training = $([ "$RESUME_TRAINING" = true ] && echo True || echo False)  # Convert bash boolean to Python boolean
max_iter = $MAX_ITER
checkpoint_period = $CHECKPOINT_PERIOD
output_dir = "$OUTPUT_DIR"
coco_config_file = "$COCO_CONFIG_FILE"
ims_per_batch = $IMS_PER_BATCH
batch_size_per_image = $BATCH_SIZE_PER_IMAGE
base_lr = $BASE_LR

# Load the COCO annotation file to detect number of classes
with open(train_annotation, 'r') as f:
    coco_data = json.load(f)

# Extract number of unique categories
num_classes = len(coco_data['categories'])
print(f"Detected {num_classes} classes from the dataset.")

# Register the datasets
register_coco_instances("coco_roboone_train", {}, train_annotation, train_image_dir)
register_coco_instances("coco_roboone_val", {}, val_annotation, val_image_dir)

# Confirm the datasets are registered
print("Datasets registered successfully.")
print("Available datasets:", DatasetCatalog.list())

# Set up configuration
cfg = get_cfg()
cfg.merge_from_file(coco_config_file)
cfg.DATASETS.TRAIN = ("coco_roboone_train",)
cfg.DATASETS.TEST = ("coco_roboone_val",)
cfg.MODEL.ROI_HEADS.NUM_CLASSES = num_classes
cfg.OUTPUT_DIR = output_dir

# Set solver parameters
cfg.SOLVER.MAX_ITER = max_iter
cfg.SOLVER.CHECKPOINT_PERIOD = checkpoint_period
cfg.SOLVER.IMS_PER_BATCH = ims_per_batch
cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = batch_size_per_image
cfg.SOLVER.BASE_LR = base_lr

# Train the model
trainer = DefaultTrainer(cfg)
trainer.resume_or_load(resume=resume_training)
trainer.train()

print("Mask R-CNN training completed.")
END