def train(config):
train_seq = DataSequence.from_table(config.TRAIN_FP,
config.DATA_COL,
config.TARGET_COL,
config.NUM_CLASSES,
config.BATCH_SIZE,
augment=True)
val_seq = DataSequence.from_table(config.VAL_FP,
config.DATA_COL,
config.TARGET_COL,
config.NUM_CLASSES,
config.BATCH_SIZE,
augment=False)
# reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=Config.reduce_lr_factor, patience=Config.reduce_lr_patience, cooldown=0, mode='min', verbose=1, min_lr=1e-8)
early_stop = EarlyStopping(monitor='val_loss',
patience=config.EARLYSTOP_PATIENCE,
verbose=1,
mode='min')
ckpt_fp = os.path.join(config.CHECKPOINT_DIR,
"model-{epoch:02d}-{val_loss:.2f}.hdf5")
model_checkpoint = ModelCheckpoint(
filepath=ckpt_fp,
monitor='val_loss',
save_weights_only=config.SAVE_WEIGHTS_ONLY,
save_best_only=config.SAVE_BEST_ONLY,
verbose=1,
mode='min')
csv_logger = CSVLogger(config.TRAIN_LOG_FP)
callbacks = [early_stop, model_checkpoint, csv_logger]
config.MODEL.summary()
history = config.MODEL.fit_generator(generator=train_seq,
validation_data=val_seq,
steps_per_epoch=len(train_seq),
validation_steps=len(val_seq),
epochs=config.EPOCHS,
verbose=1,
callbacks=callbacks,
shuffle=True
# workers=2
)
