"0,1", data_device_id="cuda:0"
) #0, 1, 2, 3, IMPORTANT: data_device_id is set to free gpu for storing the model, e.g."cuda:1"
multi_gpu = [0, 1] #use 2 gpus
#SEED = 1234#5678#4567#3456#2345#1234
debug = False # if True, load 100 samples, False
IMG_SIZE = (256, 1600)
BATCH_SIZE = 32
NUM_WORKERS = 24
warm_start, last_checkpoint_path = False, 'checkpoint/%s_%s_v1_seed%s/best.pth.tar' % (
MODEL, IMG_SIZE, SEED)
checkpoint_path = '../checkpoint/nonzero_classifier_%s_%dx%d_v4_seed%s' % (
MODEL, IMG_SIZE[0], IMG_SIZE[1], SEED)
LOG_PATH = '../logging/nonzero_classifier_%s_%dx%d_v4_seed%s.log' % (
MODEL, IMG_SIZE[0], IMG_SIZE[1], SEED) #
#torch.cuda.manual_seed_all(SEED)
NUM_EPOCHS = 100
early_stopping_round = 10 #500#50
LearningRate = 0.02 #0.02#0.002
######### Load data #########
train_dl, val_dl = prepare_trainset(BATCH_SIZE, NUM_WORKERS, SEED, IMG_SIZE,
debug)
######### Run the training process #########
run_check_net(train_dl, val_dl, multi_gpu=multi_gpu)
print('------------------------\nComplete SEED=%d\n------------------------' %
SEED)
BATCH_SIZE = 16
NUM_WORKERS = 24
warm_start, last_checkpoint_path = False, '../checkpoint/nonzero_classifier_efficientnet-b5_512x768_v4_seed2024/best.pth.tar'
checkpoint_path = '../checkpoint/nonzero_classifier_%s_%dx%d_v4_seed%s' % (
MODEL, IMG_SIZE[0], IMG_SIZE[1], SEED)
LOG_PATH = '../logging/nonzero_classifier_%s_%dx%d_v4_seed%s.log' % (
MODEL, IMG_SIZE[0], IMG_SIZE[1], SEED) #
#torch.cuda.manual_seed_all(SEED)
NUM_EPOCHS = 30 #25
LearningRate = 0.01
MIN_LR = 0.001
FREEZE = False #True #only train final fc layer for first few epochs
early_stopping_round = 500 #500#50
######### Load data #########
train_dl, val_dl = prepare_trainset(BATCH_SIZE,
NUM_WORKERS,
SEED,
IMG_SIZE,
debug,
nonempty_only=False,
crop=False) #crop=False
######### Run the training process #########
run_check_net(train_dl, val_dl, multi_gpu=multi_gpu)
print('------------------------\nComplete SEED=%d\n------------------------' %
SEED)
BATCH_SIZE = 16 #8 #16
NUM_WORKERS = 24
warm_start, last_checkpoint_path = False, '../checkpoint/deeplabv3plus_%s_%dx%d_v2_seed%s/last.pth.tar' % (
MODEL, IMG_SIZE[0], IMG_SIZE[1], SEED)
checkpoint_path = '../checkpoint/CSAILVision_%s_%dx%d_v3_seed%s' % (
MODEL, IMG_SIZE[0], IMG_SIZE[1], SEED)
LOG_PATH = '../logging/CSAILVision_%s_%dx%d_v3_seed%s.log' % (
MODEL, IMG_SIZE[0], IMG_SIZE[1], SEED)
#torch.cuda.manual_seed_all(SEED)
NUM_EPOCHS = 50
early_stopping_round = 10 #10#500#50
LearningRate = 0.02
######### Load data #########
train_dl, val_dl = prepare_trainset(
BATCH_SIZE,
NUM_WORKERS,
SEED,
IMG_SIZE,
debug,
nonempty_only=False,
crop=False,
output_shape=None) #True: Only using nonempty-mask!
######### Run the training process #########
run_check_net(train_dl, val_dl, multi_gpu=multi_gpu, nonempty_only_loss=False)
print('------------------------\nComplete SEED=%d\n------------------------' %
SEED)