train.sh

#======================================================
# TRAINING
#======================================================
# --------------------
# Load environment variables
# --------------------
export FRAMES_TRAIN_PATH=input/training/images
export MASKS_TRAIN_PATH=input/training/mask
export FRAMES_VAL_PATH=input/validation/images
export MASKS_VAL_PATH=input/validation/mask
export TRAIN_BATCH_SIZE=2	            #NIT:6 - UDD:4
export VAL_BATCH_SIZE=4               #NIT:4 - UDD:6
export TRAIN_STEPS_PER_EPOCH=900      #NIT:300 - UDD:4131    # train_len(ex. 800) = batch_size(20) * steps_per_epoch(40)
export VAL_STEPS_PER_EPOCH=286        #NIT:286 - UDD:918     # val_len(200) = batch_size(5) * steps_per_epoch(40)
export NO_EPOCHS=100
export IN_HEIGHT=320
export IN_WIDTH=320 #
export WEIGHTS_OUT_PATH=models_weights/
# If pre-trained is True, set the variable `WEIGHTS_IN_PATH` to the weights(.h5) full path
export PRE_TRAINED=False
export WEIGHTS_IN_PATH=models_weights/road_seg_nit_unet_vgg16_320_320_weights_24_02_20-00_32_AM/road_seg_nit_unet_vgg16_320_320_weights_24_02_20-00_32_AM_99.h5    # Full path of pre-trained models_weights
export PROBLEM=road_seg_udd
export LABELS_FILE=labels.txt
# If getting FLOPS is True, set the variable `FLOPS_PATH` to the CSV file fullpath
export GET_FLOPS=True
export FLOPS_PATH=notebooks/flops.csv
# MODEL:
# unet,
# fcn_8s,
# dense_unet,
# dw_unet,
# dw_dense_unet_d10, dw_dense_unet_d15, dw_dense_unet_d20
# seg_net,
# bcd_unet_d1, bcd_unet_d3,
# fcn_densenet_56, fcn_densenet_67, fcn_densenet_103,
# deeplab_v3_plus
export MODEL=deeplab_v3_plus

# BACKBONE:
# {default, vgg16, xception, mobilenetv2}
export BACKBONE=mobilenetv2

# --------------------
# Define script loader
# --------------------
python -m src.train
#======================================================