def test_model(test_loader, eval_checkpoint_file, ssl_checkpoint_file, cfg,
cur_episode):
global plot_epoch_xvalues
global plot_epoch_yvalues
global plot_it_x_values
global plot_it_y_values
test_meter = TestMeter(len(test_loader))
if cfg.MODEL.TYPE == 'linear':
model = SSLEvaluator(n_input=cfg.MODEL.NUM_INPUT,
n_classes=cfg.MODEL.NUM_OUTPUT,
n_hidden=None)
else:
model = SSLEvaluator(n_input=cfg.MODEL.NUM_INPUT,
n_classes=cfg.MODEL.NUM_OUTPUT,
n_hidden=cfg.MODEL.NUM_HIDDEN)
model = cu.load_checkpoint(eval_checkpoint_file, model)
ssl_model = model_builder.build_model(cfg)
ssl_model = cu.load_checkpoint(ssl_checkpoint_file, ssl_model)
test_err = test_epoch(test_loader, model, ssl_model, test_meter,
cur_episode)
test_acc = 100. - test_err
return test_acc
def setup_model():
"""Sets up a model for training or testing and log the results."""
# Build the model
model = builders.build_model()
logger.info("Model:\n{}".format(model))
# Log model complexity
# logger.info(logging.dump_log_data(net.complexity(model), "complexity"))
if cfg.TASK == "seg" and cfg.TRAIN.DATASET == "cityscapes":
h, w = 1025, 2049
else:
h, w = cfg.TRAIN.IM_SIZE, cfg.TRAIN.IM_SIZE
if cfg.TASK == "jig":
x = torch.randn(1, cfg.JIGSAW_GRID ** 2, cfg.MODEL.INPUT_CHANNELS, h, w)
else:
x = torch.randn(1, cfg.MODEL.INPUT_CHANNELS, h, w)
macs, params = profile(model, inputs=(x, ), verbose=False)
logger.info("Params: {:,}".format(params))
logger.info("Flops: {:,}".format(macs))
# Transfer the model to the current GPU device
err_str = "Cannot use more GPU devices than available"
assert cfg.NUM_GPUS <= torch.cuda.device_count(), err_str
cur_device = torch.cuda.current_device()
model = model.cuda(device=cur_device)
# Use multi-process data parallel model in the multi-gpu setting
if cfg.NUM_GPUS > 1:
# Make model replica operate on the current device
model = torch.nn.parallel.DistributedDataParallel(
module=model, device_ids=[cur_device], output_device=cur_device
)
# Set complexity function to be module's complexity function
# model.complexity = model.module.complexity
return model
def test_model(test_loader, checkpoint_file, cfg, cur_episode):
global plot_episode_xvalues
global plot_episode_yvalues
global plot_epoch_xvalues
global plot_epoch_yvalues
global plot_it_x_values
global plot_it_y_values
test_meter = TestMeter(len(test_loader))
model = model_builder.build_model(cfg)
model = cu.load_checkpoint(checkpoint_file, model)
test_err = test_epoch(test_loader, model, test_meter, cur_episode)
test_acc = 100. - test_err
plot_episode_xvalues.append(cur_episode)
plot_episode_yvalues.append(test_acc)
plot_arrays(x_vals=plot_episode_xvalues, y_vals=plot_episode_yvalues, \
x_name="Episodes", y_name="Test Accuracy", dataset_name=cfg.DATASET.NAME, out_dir=cfg.EXP_DIR)
save_plot_values([plot_episode_xvalues, plot_episode_yvalues], \
["plot_episode_xvalues", "plot_episode_yvalues"], out_dir=cfg.EXP_DIR)
return test_acc
def ensemble_test_model(test_loader, checkpoint_file, cfg, cur_episode):
test_meter = TestMeter(len(test_loader))
model = model_builder.build_model(cfg)
model = cu.load_checkpoint(checkpoint_file, model)
test_err = test_epoch(test_loader, model, test_meter, cur_episode)
test_acc = 100. - test_err
return test_acc
def main():
config.load_cfg_fom_args("Train a classification model.")
config.assert_and_infer_cfg()
cfg.freeze()
print("building model {}".format(cfg.MODEL.TYPE))
model = build_model()
model.eval()
x = torch.randn(1, 3, 224, 224)
y = model(x)
print(y.shape)
model_complex = complexity(model)
print(model_complex)
def predict(args):
cfg.MODEL.TYPE = "regnet"
cfg.REGNET.DEPTH = 25
cfg.REGNET.SE_ON = False
cfg.REGNET.W0 = 112
cfg.REGNET.WA = 33.22
cfg.REGNET.WM = 2.27
cfg.REGNET.GROUP_W = 72
cfg.BN.NUM_GROUPS = 4
cfg.ANYNET.STEM_CHANNELS = 1
cfg.MODEL.NUM_CLASSES = 10958
net = builders.build_model()
net.load_state_dict(torch.load(args.classify_model, map_location="cpu"))
net.eval()
softmax = nn.Softmax(dim=1)
label_map = load_label_file()
# Load audio file to np.array
audio, sr = librosa.load(args.sound_file,
mono=True,
offset=1.1,
sr=CFG.sample_rate)
logmel = librosa.feature.melspectrogram(audio,
sr,
n_mels=CFG.n_mels,
fmax=CFG.fmax)
S_dB = librosa.power_to_db(logmel, ref=np.max)
aug = augment.Augment(training=False)
segs = S_dB.shape[1] // SEGMENT_SIZE
for index in range(segs):
begin = index * SEGMENT_SIZE
end = begin + SEGMENT_SIZE
if end > S_dB.shape[1]:
print(f"{end} is out of range {S_dB.shape[1]} [{args.sound_file}]")
continue
sample = S_dB[:, begin:end].copy()
sample = torch.from_numpy(sample)
sample = sample.unsqueeze(0).unsqueeze(3)
sample = aug(sample)
sample = sample.permute(0, 3, 1, 2).float()
result = net(sample)
result = softmax(result)
values, indices = torch.topk(result, 5)
print("-----------------------------------------------")
for ind, val in zip(indices[0], values[0]):
ind = ind.item()
# if ind > 0 and ind < 10950:
print(ind, label_map[ind], f"({val.item()*100:.2f}%)")
def __init__(self, num_classes=1, ckpt=None):
super(Regnet, self).__init__()
from pycls.core.config import cfg
import pycls.core.config as model_config
from pycls.core.builders import build_model
model_config.load_cfg_fom_args("Train a cls model")
cfg.freeze()
model = build_model()
if ckpt:
model.load_state_dict(torch.load(ckpt)['model_state'])
in_features = model.head.fc.in_features
fc = nn.Linear(in_features, num_classes)
self.model = model
self.model.head.fc = fc
def build_model(name, pretrained=False, cfg_list=()):
"""Constructs a predefined model (note: loads global config as well)."""
# Load the config
reset_cfg()
config_file = get_config_file(name)
cfg.merge_from_file(config_file)
cfg.merge_from_list(cfg_list)
# Construct model
model = builders.build_model()
# Load pretrained weights
if pretrained:
weights_file = get_weights_file(name)
cp.load_checkpoint(weights_file, model)
return model
def test_model(test_loader, checkpoint_file, cfg, cur_episode):
global plot_epoch_xvalues
global plot_epoch_yvalues
global plot_it_x_values
global plot_it_y_values
test_meter = TestMeter(len(test_loader))
model = model_builder.build_model(cfg)
model = cu.load_checkpoint(checkpoint_file, model)
test_err = test_epoch(test_loader, model, test_meter, cur_episode)
test_acc = 100. - test_err
return test_acc
def setup_model():
"""Sets up a model for training or testing and log the results."""
# Build the model
model = builders.build_model()
logger.info("Model:\n{}".format(model)) if cfg.VERBOSE else ()
# Log model complexity
logger.info(logging.dump_log_data(net.complexity(model), "complexity"))
# Transfer the model to the current GPU device
err_str = "Cannot use more GPU devices than available"
assert cfg.NUM_GPUS <= torch.cuda.device_count(), err_str
cur_device = torch.cuda.current_device()
model = model.cuda(device=cur_device)
# Use multi-process data parallel model in the multi-gpu setting
if cfg.NUM_GPUS > 1:
# Make model replica operate on the current device
ddp = torch.nn.parallel.DistributedDataParallel
model = ddp(module=model, device_ids=[cur_device], output_device=cur_device)
return model
def __init__(self, num_clusters, num_tiles, num_classes, ckpt):
super().__init__()
from pycls.core.config import cfg
import pycls.core.config as model_config
from pycls.core.builders import build_model
model_config.load_cfg_fom_args("Train a cls model")
cfg.freeze()
model = build_model()
if ckpt:
model.load_state_dict(torch.load(ckpt)['model_state'])
self.enc = nn.Sequential(model.stem, model.s1, model.s2, model.s3,
model.s4,
nn.AdaptiveAvgPool2d(output_size=(1, 1)),
nn.Flatten(), nn.Dropout(p=0.3))
self.nc = model.head.fc.in_features
self.netvlad = NetVLAD(cluster_size=num_clusters,
max_frames=num_tiles,
feature_size=self.nc,
truncate=False)
self.fc = nn.Linear(num_clusters * self.nc, num_classes)
def test_model():
"""Evaluates the model."""
# Setup logging
logging.setup_logging()
# Show the config
logger.info("Config:\n{}".format(cfg))
# Fix the RNG seeds (see RNG comment in core/config.py for discussion)
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Configure the CUDNN backend
torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK
# Build the model (before the loaders to speed up debugging)
model = builders.build_model()
logger.info("Model:\n{}".format(model))
logger.info(logging.dump_json_stats(net.complexity(model)))
# Compute precise time
if cfg.PREC_TIME.ENABLED:
logger.info("Computing precise time...")
loss_fun = builders.build_loss_fun()
prec_time = net.compute_precise_time(model, loss_fun)
logger.info(logging.dump_json_stats(prec_time))
net.reset_bn_stats(model)
# Load model weights
checkpoint.load_checkpoint(cfg.TEST.WEIGHTS, model)
logger.info("Loaded model weights from: {}".format(cfg.TEST.WEIGHTS))
# Create data loaders
test_loader = loader.construct_test_loader()
# Create meters
test_meter = meters.TestMeter(len(test_loader))
# Evaluate the model
test_epoch(test_loader, model, test_meter, 0)
def setup_model():
"""Sets up a model for training or testing and log the results."""
# Build the model
model = builders.build_model()
logger.info("Model:\n{}".format(model))
# Log model complexity
logger.info(logging.dump_json_stats(net.complexity(model)))
# Transfer the model to the current GPU device
err_str = "Cannot use more GPU devices than available"
assert cfg.NUM_GPUS <= torch.cuda.device_count(), err_str
cur_device = torch.cuda.current_device()
model = model.cuda(device=cur_device)
# Use multi-process data parallel model in the multi-gpu setting
if cfg.NUM_GPUS > 1:
# Make model replica operate on the current device
model = torch.nn.parallel.DistributedDataParallel(
module=model,
device_ids=[cur_device],
output_device=cur_device,
find_unused_parameters=True)
# Set complexity function to be module's complexity function
model.complexity = model.module.complexity
return model
def setup_model():
"""Sets up a model for training or testing and log the results."""
# Build the model
model = builders.build_model()
logger.info("Model:\n{}".format(model)) if cfg.VERBOSE else ()
# Log model complexity
logger.info(logging.dump_log_data(net.complexity(model), "complexity"))
# Transfer the model to the current GPU device
err_str = "Cannot use more GPU devices than available"
#assert cfg.NUM_GPUS <= torch.cuda.device_count(), err_str
assert cfg.NUM_GPUS <= torch.npu.device_count(), err_str
cur_device = torch.npu.current_device()
model = model.to(cur_device)
optimizer = optim.construct_optimizer(model)
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O2",
loss_scale=128)
if cfg.NUM_GPUS > 1:
#Make model replica operate on the current device
ddp = torch.nn.parallel.DistributedDataParallel
model = ddp(model, device_ids=[cur_device], broadcast_buffers=False)
return model, optimizer
def train_model():
"""Trains the model."""
# Setup logging
logging.setup_logging()
# Show the config
logger.info("Config:\n{}".format(cfg))
# Fix the RNG seeds (see RNG comment in core/config.py for discussion)
np.random.seed(cfg.RNG_SEED)
torch.manual_seed(cfg.RNG_SEED)
# Configure the CUDNN backend
torch.backends.cudnn.benchmark = cfg.CUDNN.BENCHMARK
# Build the model (before the loaders to speed up debugging)
model = builders.build_model()
logger.info("Model:\n{}".format(model))
logger.info(logging.dump_json_stats(net.complexity(model)))
# Define the loss function
loss_fun = builders.build_loss_fun()
# Construct the optimizer
optimizer = optim.construct_optimizer(model)
# Load checkpoint or initial weights
start_epoch = 0
if cfg.TRAIN.AUTO_RESUME and checkpoint.has_checkpoint():
last_checkpoint = checkpoint.get_last_checkpoint()
checkpoint_epoch = checkpoint.load_checkpoint(last_checkpoint, model,
optimizer)
logger.info("Loaded checkpoint from: {}".format(last_checkpoint))
start_epoch = checkpoint_epoch + 1
elif cfg.TRAIN.WEIGHTS:
checkpoint.load_checkpoint(cfg.TRAIN.WEIGHTS, model)
logger.info("Loaded initial weights from: {}".format(
cfg.TRAIN.WEIGHTS))
# Compute precise time
if start_epoch == 0 and cfg.PREC_TIME.ENABLED:
logger.info("Computing precise time...")
prec_time = net.compute_precise_time(model, loss_fun)
logger.info(logging.dump_json_stats(prec_time))
net.reset_bn_stats(model)
# Create data loaders
train_loader = loader.construct_train_loader()
test_loader = loader.construct_test_loader()
# Create meters
train_meter = meters.TrainMeter(len(train_loader))
test_meter = meters.TestMeter(len(test_loader))
# Perform the training loop
logger.info("Start epoch: {}".format(start_epoch + 1))
for cur_epoch in range(start_epoch, cfg.OPTIM.MAX_EPOCH):
# Train for one epoch
train_epoch(train_loader, model, loss_fun, optimizer, train_meter,
cur_epoch)
# Compute precise BN stats
if cfg.BN.USE_PRECISE_STATS:
net.compute_precise_bn_stats(model, train_loader)
# Save a checkpoint
if checkpoint.is_checkpoint_epoch(cur_epoch):
checkpoint_file = checkpoint.save_checkpoint(
model, optimizer, cur_epoch)
logger.info("Wrote checkpoint to: {}".format(checkpoint_file))
# Evaluate the model
if is_eval_epoch(cur_epoch):
test_epoch(test_loader, model, test_meter, cur_epoch)
def main(cfg):
# Login to wandb
wandb.login()
# Initialize a new wandb run
wandb.init(project="rotation-pred", name=cfg.EXP_NAME)
# Setting up GPU args
use_cuda = (cfg.NUM_GPUS > 0) and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {
'num_workers': cfg.DATA_LOADER.NUM_WORKERS,
'pin_memory': cfg.DATA_LOADER.PIN_MEMORY
} if use_cuda else {}
# Using specific GPU
os.environ['NVIDIA_VISIBLE_DEVICES'] = str(cfg.GPU_ID)
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
print("Using GPU : {}.\n".format(cfg.GPU_ID))
# Getting the output directory ready (default is "/output")
cfg.OUT_DIR = os.path.join(os.path.abspath('..'), cfg.OUT_DIR)
if not os.path.exists(cfg.OUT_DIR):
os.makedirs(cfg.OUT_DIR)
# Create "DATASET" specific directory
dataset_out_dir = os.path.join(cfg.OUT_DIR, cfg.DATASET.NAME,
cfg.MODEL.TYPE)
if not os.path.exists(dataset_out_dir):
os.makedirs(dataset_out_dir)
# Creating the experiment directory inside the dataset specific directory
# all logs, labeled, unlabeled, validation sets are stroed here
# E.g., output/CIFAR10/resnet18/{timestamp or cfg.EXP_NAME based on arguments passed}
if cfg.EXP_NAME == 'auto':
now = datetime.now()
exp_dir = f'{now.year}_{now.month}_{now.day}_{now.hour}{now.minute}{now.second}'
else:
exp_dir = cfg.EXP_NAME
exp_dir = os.path.join(dataset_out_dir, exp_dir)
if not os.path.exists(exp_dir):
os.mkdir(exp_dir)
print("Experiment Directory is {}.\n".format(exp_dir))
else:
print(
"Experiment Directory Already Exists: {}. Reusing it may lead to loss of old logs in the directory.\n"
.format(exp_dir))
cfg.EXP_DIR = exp_dir
# Setup Logger
lu.setup_logging(cfg)
# Dataset preparing steps
print("\n======== PREPARING DATA AND SSL EVALUATION MODEL ========\n")
cfg.DATASET.ROOT_DIR = os.path.join(os.path.abspath('..'),
cfg.DATASET.ROOT_DIR)
data_obj = Data(cfg)
train_data, train_size = data_obj.getDataset(save_dir=cfg.DATASET.ROOT_DIR,
isTrain=True,
isDownload=True)
trainSet = [i for i in range(train_size)]
print("\n Rotation Dataset {} Loaded Sucessfully.\nTotal Train Size: {}\n".
format(cfg.DATASET.NAME, train_size))
logger.info(
"Rotation Dataset {} Loaded Sucessfully. Total Train Size: {}\n".
format(cfg.DATASET.NAME, train_size))
trainSet_path = data_obj.saveSet(setArray=trainSet,
setName='trainSet',
save_dir=cfg.EXP_DIR)
trainSet = data_obj.loadPartition(setPath=trainSet_path)
# Preparing dataloaders for initial training
trainSet_loader = data_obj.getSequentialDataLoader(
indexes=trainSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
# Initialize the evaluation model
if cfg.MODEL.TYPE == 'linear':
model = SSLEvaluator(n_input=cfg.MODEL.NUM_INPUT,
n_classes=cfg.MODEL.NUM_OUTPUT,
n_hidden=None)
else:
model = SSLEvaluator(n_input=cfg.MODEL.NUM_INPUT,
n_classes=cfg.MODEL.NUM_OUTPUT,
n_hidden=cfg.MODEL.NUM_HIDDEN)
print("Evaluation model: {}\n".format(cfg.MODEL.EVAL))
logger.info("Evalution model: {}\n".format(cfg.MODEL.EVAL))
# Initialize the SSL model
ssl_model = model_builder.build_model(cfg)
ssl_checkpoint_file = os.path.join(os.path.abspath('..'),
cfg.TEST.MODEL_PATH)
ssl_model = cu.load_checkpoint(ssl_checkpoint_file, ssl_model)
# Construct the optimizer
optimizer = optim.construct_optimizer(cfg, model)
print("optimizer: {}\n".format(optimizer))
logger.info("optimizer: {}\n".format(optimizer))
# This is to seamlessly use the code originally written for AL episodes
cfg.EPISODE_DIR = cfg.EXP_DIR
# Train model
print("======== EVALUATOR TRAINING ========")
logger.info("======== EVALUATOR TRAINING ========")
_, _, eval_checkpoint_file = train_model(trainSet_loader, None, model,
ssl_model, optimizer, cfg)
# eval_checkpoint_file = os.path.join(os.path.abspath('..'), '')
# Test best model checkpoint
print("======== EVALUATOR TESTING ========\n")
logger.info("======== EVALUATOR TESTING ========\n")
test_acc = test_model(trainSet_loader,
eval_checkpoint_file,
ssl_checkpoint_file,
cfg,
cur_episode=1)
print("Test Accuracy: {}.\n".format(round(test_acc, 4)))
logger.info("Test Accuracy {}.\n".format(test_acc))
print("================================\n\n")
logger.info("================================\n\n")
def train(args, train_loader, eval_loader):
cfg.MODEL.TYPE = "regnet"
cfg.REGNET.DEPTH = 25
cfg.REGNET.SE_ON = False
cfg.REGNET.W0 = 112
cfg.REGNET.WA = 33.22
cfg.REGNET.WM = 2.27
cfg.REGNET.GROUP_W = 72
cfg.BN.NUM_GROUPS = 4
cfg.ANYNET.STEM_CHANNELS = 1
cfg.MODEL.NUM_CLASSES = config["num_classes"]
net = builders.build_model()
net = net.cuda(device=torch.cuda.current_device())
print("net", net)
if args.resume:
print("Resuming training, loading {}...".format(args.resume))
ckpt_file = (
config["save_folder"]
+ config["ckpt_name"]
+ "_"
+ str(args.resume)
+ ".pth"
)
net.load_state_dict(torch.load(ckpt_file))
if args.finetune:
print("Finetuning......")
# Freeze all layers
for param in net.parameters():
param.requires_grad = False
# Unfreeze some layers
for layer in [net.s4.b1, net.s4.b2]:
for param in layer.parameters():
param.requies_grad = True
net.head.fc.weight.requires_grad = True
optimizer = optim.SGD(
filter(lambda param: param.requires_grad, net.parameters()),
lr=args.lr,
momentum=args.momentum,
nesterov=False,
)
else:
optimizer = optim.SGD(
net.parameters(),
lr=args.lr,
momentum=args.momentum,
nesterov=False,
)
scheduler = ReduceLROnPlateau(
optimizer,
"max",
factor=0.5,
patience=1,
verbose=True,
threshold=1e-3,
threshold_mode="abs",
)
aug = augment.Augment().cuda()
if args.fp16:
import apex.amp as amp
net, optimizer = amp.initialize(net, optimizer, opt_level="O2")
batch_iterator = iter(train_loader)
sum_accuracy = 0
step = 0
config["eval_period"] = len(train_loader.dataset) // args.batch_size
config["verbose_period"] = config["eval_period"] // 5
train_start_time = time.time()
for iteration in range(
args.resume + 1,
args.max_epoch * len(train_loader.dataset) // args.batch_size,
):
t0 = time.time()
try:
sounds, type_ids = next(batch_iterator)
except StopIteration:
batch_iterator = iter(train_loader)
sounds, type_ids = next(batch_iterator)
except Exception as ex:
print("Loading data exception:", ex)
if torch.cuda.is_available():
sounds = Variable(sounds.cuda())
type_ids = Variable(type_ids.cuda())
else:
sounds = Variable(sounds)
type_ids = Variable(type_ids)
sounds = sounds.unsqueeze(3)
sounds = sounds.permute(0, 3, 1, 2).float()
if torch.cuda.is_available():
one_hot = torch.cuda.FloatTensor(
type_ids.shape[0], config["num_classes"]
)
else:
one_hot = torch.FloatTensor(
type_ids.shape[0], config["num_classes"]
)
one_hot.fill_(0.5 / (config["num_classes"] - 1))
one_hot.scatter_(1, type_ids.unsqueeze(1), 0.5)
# augmentation
sounds = aug(sounds)
# forward
out = net(sounds)
# backprop
optimizer.zero_grad()
loss = torch.sum(-one_hot * F.log_softmax(out, -1), -1).mean()
# loss = F.cross_entropy(out, type_ids)
if args.fp16:
import apex.amp as amp
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
nn.utils.clip_grad_norm_(net.parameters(), max_norm=20, norm_type=2)
optimizer.step()
t1 = time.time()
if iteration % config["verbose_period"] == 0:
# accuracy
_, predict = torch.max(out, 1)
correct = (predict == type_ids)
accuracy = correct.sum().item() / correct.size()[0]
print(
"iter: %d loss: %.4f | acc: %.4f | time: %.4f sec."
% (iteration, loss.item(), accuracy, (t1 - t0)),
flush=True,
)
sum_accuracy += accuracy
step += 1
warmup_steps = config["verbose_period"]
if iteration < warmup_steps:
warmup_learning_rate(optimizer, iteration, warmup_steps)
if (
iteration % config["eval_period"] == 0
and iteration != 0
and step != 0
):
with torch.no_grad():
loss, accuracy = evaluate(net, eval_loader)
hours = int(time.time() - train_start_time) // 3600
now = datetime.datetime.now().strftime("%Y-%m-%d %H:%M")
print(
"[{}] [{}] Eval accuracy:{:6f} | Train accuracy:{:6f}".format(
now, hours, accuracy, sum_accuracy / step
),
flush=True,
)
scheduler.step(accuracy)
sum_accuracy = 0
step = 0
if iteration % config["eval_period"] == 0 and iteration != 0:
# save checkpoint
print("Saving state, iter:", iteration, flush=True)
save_ckpt(net, iteration)
# final checkpoint
save_ckpt(net, iteration)
def main(cfg):
# Setting up GPU args
use_cuda = (cfg.NUM_GPUS > 0) and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {
'num_workers': cfg.DATA_LOADER.NUM_WORKERS,
'pin_memory': cfg.DATA_LOADER.PIN_MEMORY
} if use_cuda else {}
# Using specific GPU
os.environ['CUDA_VISIBLE_DEVICES'] = str(cfg.GPU_ID)
print("Using GPU : {}.\n".format(cfg.GPU_ID))
# Getting the output directory ready (default is "/output")
cfg.OUT_DIR = os.path.join(os.path.abspath('..'), cfg.OUT_DIR)
if not os.path.exists(cfg.OUT_DIR):
os.mkdir(cfg.OUT_DIR)
# Create "DATASET" specific directory
dataset_out_dir = os.path.join(cfg.OUT_DIR, cfg.DATASET.NAME)
if not os.path.exists(dataset_out_dir):
os.mkdir(dataset_out_dir)
# Creating the experiment directory inside the dataset specific directory
# all logs, labeled, unlabeled, validation sets are stroed here
# E.g., output/CIFAR10/{timestamp or cfg.EXP_NAME based on arguments passed}
if cfg.EXP_NAME == 'auto':
now = datetime.now()
exp_dir = f'{now.year}_{now.month}_{now.day}_{now.hour}{now.minute}{now.second}'
else:
exp_dir = cfg.EXP_NAME
exp_dir = os.path.join(dataset_out_dir, exp_dir)
if not os.path.exists(exp_dir):
os.mkdir(exp_dir)
print("Experiment Directory is {}.\n".format(exp_dir))
else:
print(
"Experiment Directory Already Exists: {}. Reusing it may lead to loss of old logs in the directory.\n"
.format(exp_dir))
cfg.EXP_DIR = exp_dir
# Setup Logger
lu.setup_logging(cfg)
# Dataset preparing steps
print("\n======== PREPARING DATA AND MODEL ========\n")
cfg.DATASET.ROOT_DIR = os.path.join(os.path.abspath('..'),
cfg.DATASET.ROOT_DIR)
data_obj = Data(cfg)
train_data, train_size = data_obj.getDataset(save_dir=cfg.DATASET.ROOT_DIR,
isTrain=True,
isDownload=True)
test_data, test_size = data_obj.getDataset(save_dir=cfg.DATASET.ROOT_DIR,
isTrain=False,
isDownload=True)
print(
"\nDataset {} Loaded Sucessfully.\nTotal Train Size: {} and Total Test Size: {}\n"
.format(cfg.DATASET.NAME, train_size, test_size))
logger.info(
"Dataset {} Loaded Sucessfully. Total Train Size: {} and Total Test Size: {}\n"
.format(cfg.DATASET.NAME, train_size, test_size))
trainSet_path, valSet_path = data_obj.makeTVSets(train_split_ratio=cfg.ACTIVE_LEARNING.INIT_RATIO, \
val_split_ratio=cfg.DATASET.VAL_RATIO, data=train_data, seed_id=cfg.RNG_SEED, save_dir=cfg.EXP_DIR)
trainSet, valSet = data_obj.loadTVPartitions(trainSetPath=trainSet_path,
valSetPath=valSet_path)
print("Data Partitioning Complete. \nTrain Set: {}, Validation Set: {}\n".
format(len(trainSet), len(valSet)))
logger.info("\nTrain Set: {}, Validation Set: {}\n".format(
len(trainSet), len(valSet)))
# Preparing dataloaders for initial training
trainSet_loader = data_obj.getIndexesDataLoader(
indexes=trainSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
valSet_loader = data_obj.getIndexesDataLoader(
indexes=valSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
test_loader = data_obj.getTestLoader(data=test_data,
test_batch_size=cfg.TRAIN.BATCH_SIZE,
seed_id=cfg.RNG_SEED)
# Initialize the models
num_ensembles = cfg.ENSEMBLE.NUM_MODELS
models = []
for i in range(num_ensembles):
models.append(model_builder.build_model(cfg))
print("{} ensemble models of type: {}\n".format(cfg.ENSEMBLE.NUM_MODELS,
cfg.ENSEMBLE.MODEL_TYPE))
logger.info("{} ensemble models of type: {}\n".format(
cfg.ENSEMBLE.NUM_MODELS, cfg.ENSEMBLE.MODEL_TYPE))
# This is to seamlessly use the code originally written for AL episodes
cfg.EPISODE_DIR = cfg.EXP_DIR
# Train models
print("======== ENSEMBLE TRAINING ========")
logger.info("======== ENSEMBLE TRAINING ========")
best_model_paths = []
test_accs = []
for i in range(num_ensembles):
print("=== Training ensemble [{}/{}] ===".format(i + 1, num_ensembles))
# Construct the optimizer
optimizer = optim.construct_optimizer(cfg, models[i])
print("optimizer: {}\n".format(optimizer))
logger.info("optimizer: {}\n".format(optimizer))
# Each ensemble gets its own output directory
cfg.EPISODE_DIR = os.path.join(cfg.EPISODE_DIR,
'model_{} '.format(i + 1))
# Train the model
best_val_acc, best_val_epoch, checkpoint_file = ensemble_train_model(
trainSet_loader, valSet_loader, models[i], optimizer, cfg)
best_model_paths.append(checkpoint_file)
print("Best Validation Accuracy by Model {}: {}\nBest Epoch: {}\n".
format(i + 1, round(best_val_acc, 4), best_val_epoch))
logger.info(
"Best Validation Accuracy by Model {}: {}\tBest Epoch: {}\n".
format(i + 1, round(best_val_acc, 4), best_val_epoch))
# Test the model
print("=== Testing ensemble [{}/{}] ===".format(i + 1, num_ensembles))
test_acc = ensemble_test_model(test_loader,
checkpoint_file,
cfg,
cur_episode=0)
test_accs.append(test_acc)
print("Test Accuracy by Model {}: {}.\n".format(
i + 1, round(test_acc, 4)))
logger.info("Test Accuracy by Model {}: {}.\n".format(i + 1, test_acc))
# Reset EPISODE_DIR
cfg.EPISODE_DIR = cfg.EXP_DIR
# Test each best model checkpoint and report the average
print("======== ENSEMBLE TESTING ========\n")
logger.info("======== ENSEMBLE TESTING ========\n")
mean_test_acc = np.mean(test_accs)
print("Average Ensemble Test Accuracy: {}.\n".format(
round(mean_test_acc, 4)))
logger.info("Average Ensemble Test Accuracy: {}.\n".format(mean_test_acc))
print("================================\n\n")
logger.info("================================\n\n")
def main(cfg):
# Setting up GPU args
use_cuda = (cfg.NUM_GPUS > 0) and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {
'num_workers': cfg.DATA_LOADER.NUM_WORKERS,
'pin_memory': cfg.DATA_LOADER.PIN_MEMORY
} if use_cuda else {}
# Using specific GPU
os.environ['CUDA_VISIBLE_DEVICES'] = str(cfg.GPU_ID)
print("Using GPU : {}.\n".format(cfg.GPU_ID))
# Getting the output directory ready (default is "/output")
cfg.OUT_DIR = os.path.join(os.path.abspath('..'), cfg.OUT_DIR)
if not os.path.exists(cfg.OUT_DIR):
os.mkdir(cfg.OUT_DIR)
# Create "DATASET" specific directory
dataset_out_dir = os.path.join(cfg.OUT_DIR, cfg.DATASET.NAME,
cfg.MODEL.TYPE)
if not os.path.exists(dataset_out_dir):
os.makedirs(dataset_out_dir)
# Creating the experiment directory inside the dataset specific directory
# all logs, labeled, unlabeled, validation sets are stroed here
# E.g., output/CIFAR10/resnet18/{timestamp or cfg.EXP_NAME based on arguments passed}
if cfg.EXP_NAME == 'auto':
now = datetime.now()
exp_dir = f'{now.year}_{now.month}_{now.day}_{now.hour}{now.minute}{now.second}'
else:
exp_dir = cfg.EXP_NAME
exp_dir = os.path.join(dataset_out_dir, exp_dir)
if not os.path.exists(exp_dir):
os.mkdir(exp_dir)
print("Experiment Directory is {}.\n".format(exp_dir))
else:
print(
"Experiment Directory Already Exists: {}. Reusing it may lead to loss of old logs in the directory.\n"
.format(exp_dir))
cfg.EXP_DIR = exp_dir
# Setup Logger
lu.setup_logging(cfg)
# Dataset preparing steps
print("\n======== PREPARING DATA AND MODEL ========\n")
cfg.DATASET.ROOT_DIR = os.path.join(os.path.abspath('..'),
cfg.DATASET.ROOT_DIR)
data_obj = Data(cfg)
train_data, train_size = data_obj.getDataset(save_dir=cfg.DATASET.ROOT_DIR,
isTrain=True,
isDownload=True)
test_data, test_size = data_obj.getDataset(save_dir=cfg.DATASET.ROOT_DIR,
isTrain=False,
isDownload=True)
print(
"\nDataset {} Loaded Sucessfully.\nTotal Train Size: {} and Total Test Size: {}\n"
.format(cfg.DATASET.NAME, train_size, test_size))
logger.info(
"Dataset {} Loaded Sucessfully. Total Train Size: {} and Total Test Size: {}\n"
.format(cfg.DATASET.NAME, train_size, test_size))
print("\nSampling Initial Pool using {}.".format(
str.upper(cfg.INIT_POOL.SAMPLING_FN)))
logger.info("\nSampling Initial Pool using {}.".format(
str.upper(cfg.INIT_POOL.SAMPLING_FN)))
if cfg.INIT_POOL.SAMPLING_FN == 'random':
lSet_path, uSet_path, valSet_path = data_obj.makeLUVSets(train_split_ratio=cfg.INIT_POOL.INIT_RATIO, \
val_split_ratio=cfg.DATASET.VAL_RATIO, data=train_data, seed_id=cfg.RNG_SEED, save_dir=cfg.EXP_DIR)
else:
lSet, uSet = InitialPool(cfg).sample_from_uSet(train_data)
lSet_path = f'{cfg.EXP_DIR}/lSet.npy'
np.save(lSet_path, lSet)
np.save(f'{cfg.EXP_DIR}/lSet_initial.npy', lSet)
uSet_path, valSet_path = data_obj.makeUVSets(
val_split_ratio=cfg.DATASET.VAL_RATIO,
data=uSet,
seed_id=cfg.RNG_SEED,
save_dir=cfg.EXP_DIR)
cfg.ACTIVE_LEARNING.LSET_PATH = lSet_path
cfg.ACTIVE_LEARNING.USET_PATH = uSet_path
cfg.ACTIVE_LEARNING.VALSET_PATH = valSet_path
lSet, uSet, valSet = data_obj.loadPartitions(lSetPath=cfg.ACTIVE_LEARNING.LSET_PATH, \
uSetPath=cfg.ACTIVE_LEARNING.USET_PATH, valSetPath = cfg.ACTIVE_LEARNING.VALSET_PATH)
print(
"Data Partitioning Complete. \nLabeled Set: {}, Unlabeled Set: {}, Validation Set: {}\n"
.format(len(lSet), len(uSet), len(valSet)))
logger.info(
"Labeled Set: {}, Unlabeled Set: {}, Validation Set: {}\n".format(
len(lSet), len(uSet), len(valSet)))
# Preparing dataloaders for initial training
lSet_loader = data_obj.getIndexesDataLoader(
indexes=lSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
valSet_loader = data_obj.getIndexesDataLoader(
indexes=valSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
uSet_loader = data_obj.getIndexesDataLoader(
indexes=uSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
test_loader = data_obj.getTestLoader(data=test_data,
test_batch_size=cfg.TRAIN.BATCH_SIZE,
seed_id=cfg.RNG_SEED)
# Initialize the models
num_ensembles = cfg.ENSEMBLE.NUM_MODELS
models = []
for i in range(num_ensembles):
models.append(model_builder.build_model(cfg))
print("{} ensemble models of type: {}\n".format(cfg.ENSEMBLE.NUM_MODELS,
cfg.ENSEMBLE.MODEL_TYPE))
logger.info("{} ensemble models of type: {}\n".format(
cfg.ENSEMBLE.NUM_MODELS, cfg.ENSEMBLE.MODEL_TYPE))
print("Max AL Episodes: {}\n".format(cfg.ACTIVE_LEARNING.MAX_ITER))
logger.info("Max AL Episodes: {}\n".format(cfg.ACTIVE_LEARNING.MAX_ITER))
for cur_episode in range(0, cfg.ACTIVE_LEARNING.MAX_ITER + 1):
wandb.log({"Episode": cur_episode})
print("======== EPISODE {} BEGINS ========\n".format(cur_episode))
logger.info(
"======== EPISODE {} BEGINS ========\n".format(cur_episode))
# Creating output directory for the episode
episode_dir = os.path.join(cfg.EXP_DIR, f'episode_{cur_episode}')
if not os.path.exists(episode_dir):
os.mkdir(episode_dir)
cfg.EPISODE_DIR = episode_dir
# Train models
print("======== ENSEMBLE TRAINING ========")
logger.info("======== ENSEMBLE TRAINING ========")
best_model_paths = []
test_accs = []
for i in range(num_ensembles):
print("=== Training ensemble [{}/{}] ===".format(
i + 1, num_ensembles))
# Construct the optimizer
optimizer = optim.construct_optimizer(cfg, models[i])
print("optimizer: {}\n".format(optimizer))
logger.info("optimizer: {}\n".format(optimizer))
# Each ensemble gets its own output directory
cfg.EPISODE_DIR = os.path.join(cfg.EPISODE_DIR,
'model_{}'.format(i + 1))
# Train the model
best_val_acc, best_val_epoch, checkpoint_file = ensemble_train_model(
lSet_loader, valSet_loader, models[i], optimizer, cfg)
best_model_paths.append(checkpoint_file)
print("Best Validation Accuracy by Model {}: {}\nBest Epoch: {}\n".
format(i + 1, round(best_val_acc, 4), best_val_epoch))
logger.info(
"EPISODE {} Best Validation Accuracy by Model {}: {}\tBest Epoch: {}\n"
.format(cur_episode, i + 1, round(best_val_acc, 4),
best_val_epoch))
# Test the model
print("=== Testing ensemble [{}/{}] ===".format(
i + 1, num_ensembles))
test_acc = ensemble_test_model(test_loader, checkpoint_file, cfg,
cur_episode)
test_accs.append(test_acc)
print("Test Accuracy by Model {}: {}.\n".format(
i + 1, round(test_acc, 4)))
logger.info("EPISODE {} Test Accuracy by Model {}: {}.\n".format(
cur_episode, i + 1, test_acc))
# Reset EPISODE_DIR
cfg.EPISODE_DIR = episode_dir
# Test each best model checkpoint and report the average
print("======== ENSEMBLE TESTING ========\n")
logger.info("======== ENSEMBLE TESTING ========\n")
mean_test_acc = np.mean(test_accs)
print("Average Ensemble Test Accuracy: {}.\n".format(
round(mean_test_acc, 4)))
logger.info("EPISODE {} Average Ensemble Test Accuracy: {}.\n".format(
cur_episode, mean_test_acc))
wandb.log({"Test Accuracy": mean_test_acc})
global plot_episode_xvalues
global plot_episode_yvalues
global plot_epoch_xvalues
global plot_epoch_yvalues
global plot_it_x_values
global plot_it_y_values
plot_episode_xvalues.append(cur_episode)
plot_episode_yvalues.append(mean_test_acc)
plot_arrays(x_vals=plot_episode_xvalues, y_vals=plot_episode_yvalues, \
x_name="Episodes", y_name="Test Accuracy", dataset_name=cfg.DATASET.NAME, out_dir=cfg.EXP_DIR)
save_plot_values([plot_episode_xvalues, plot_episode_yvalues], \
["plot_episode_xvalues", "plot_episode_yvalues"], out_dir=cfg.EXP_DIR, saveInTextFormat=True)
# No need to perform active sampling in the last episode iteration
if cur_episode == cfg.ACTIVE_LEARNING.MAX_ITER:
break
# Active Sample
print("======== ENSEMBLE ACTIVE SAMPLING ========\n")
logger.info("======== ENSEMBLE ACTIVE SAMPLING ========\n")
al_obj = ActiveLearning(data_obj, cfg)
clf_models = []
for i in range(num_ensembles):
temp = model_builder.build_model(cfg)
clf_models.append(cu.load_checkpoint(best_model_paths[i], temp))
activeSet, new_uSet = al_obj.sample_from_uSet(
None, lSet, uSet, train_data, supportingModels=clf_models)
# Save current lSet, new_uSet and activeSet in the episode directory
data_obj.saveSets(lSet, uSet, activeSet, cfg.EPISODE_DIR)
# Add activeSet to lSet, save new_uSet as uSet and update dataloader for the next episode
lSet = np.append(lSet, activeSet)
uSet = new_uSet
lSet_loader = data_obj.getIndexesDataLoader(
indexes=lSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
valSet_loader = data_obj.getIndexesDataLoader(
indexes=valSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
uSet_loader = data_obj.getSequentialDataLoader(
indexes=uSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
print(
"Ensemble Active Sampling Complete. After Episode {}:\nNew Labeled Set: {}, New Unlabeled Set: {}, Active Set: {}\n"
.format(cur_episode, len(lSet), len(uSet), len(activeSet)))
logger.info(
"Ensemble Active Sampling Complete. After Episode {}:\nNew Labeled Set: {}, New Unlabeled Set: {}, Active Set: {}\n"
.format(cur_episode, len(lSet), len(uSet), len(activeSet)))
print("================================\n\n")
logger.info("================================\n\n")
def train(args, train_loader, eval_loader):
cfg.MODEL.TYPE = "regnet"
cfg.REGNET.DEPTH = 20
cfg.REGNET.SE_ON = False
cfg.REGNET.W0 = 512
cfg.MODEL.NUM_CLASSES = config["num_classes"]
net = model_builder.build_model()
net = net.cuda(device=torch.cuda.current_device())
print("net", net)
if args.resume:
print("Resuming training, loading {}...".format(args.resume))
ckpt_file = (config["save_folder"] + config["ckpt_name"] + "_" +
str(args.resume) + ".pth")
net.load_state_dict(torch.load(ckpt_file))
if args.finetune:
print("Finetuning......")
# Freeze all layers
for param in net.parameters():
param.requires_grad = False
# Unfreeze some layers
for layer in [net.s1.b18, net.s1.b19, net.s1.b20]:
for param in layer.parameters():
param.requies_grad = True
net.head.fc.weight.requires_grad = True
optimizer = optim.SGD(
filter(lambda param: param.requires_grad, net.parameters()),
lr=args.lr,
momentum=args.momentum,
nesterov=False,
)
else:
optimizer = optim.SGD(
net.parameters(),
lr=args.lr,
momentum=args.momentum,
nesterov=False,
)
scheduler = ReduceLROnPlateau(
optimizer,
"max",
factor=0.5,
patience=2,
verbose=True,
threshold=1e-3,
threshold_mode="abs",
)
if args.fp16:
net, optimizer = amp.initialize(net, optimizer, opt_level="O2")
aug = augmentations.Augmentations().cuda()
batch_iterator = iter(train_loader)
sum_accuracy = 0
step = 0
for iteration in range(
args.resume + 1,
args.max_epoch * len(train_loader.dataset) // args.batch_size,
):
t0 = time.time()
try:
images, type_ids = next(batch_iterator)
except StopIteration:
batch_iterator = iter(train_loader)
images, type_ids = next(batch_iterator)
except Exception as e:
print("Loading data exception:", e)
images = Variable(images.cuda()).permute(0, 3, 1, 2).float()
type_ids = Variable(type_ids.cuda())
one_hot = torch.cuda.FloatTensor(type_ids.shape[0],
config["num_classes"])
one_hot.fill_((1 - 0.5) / config["num_classes"])
one_hot.scatter_(1, type_ids.unsqueeze(1), 0.5)
# augmentation
if not args.finetune:
images = aug(images)
# forward
out = net(images)
loss = (torch.sum(-one_hot * F.log_softmax(out, -1), -1).mean() /
args.iter_size)
if args.fp16:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
nn.utils.clip_grad_norm_(net.parameters(), max_norm=20, norm_type=2)
if iteration != 0 and iteration % args.iter_size == 0:
# backprop
optimizer.step()
optimizer.zero_grad()
t1 = time.time()
if iteration % config["verbose_period"] == 0:
# accuracy
_, predict = torch.max(out, 1)
correct = predict == type_ids
accuracy = correct.sum().item() / correct.size()[0]
print(
"iter: %d loss: %.4f | acc: %.4f | time: %.4f sec." %
(iteration, loss.item(), accuracy, (t1 - t0)),
flush=True,
)
sum_accuracy += accuracy
step += 1
warmup_steps = config["verbose_period"] * 8 * args.iter_size
if iteration < warmup_steps:
warmup_learning_rate(optimizer, iteration, warmup_steps)
if (iteration % config["eval_period"] == 0 and iteration != 0
and step != 0):
loss, accuracy = evaluate(net, eval_loader)
now = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
print(
f"[{now}] Eval accuracy: {accuracy:.4f} | Train accuracy: {sum_accuracy/step:.4f}",
flush=True,
)
scheduler.step(accuracy)
sum_accuracy = 0
step = 0
if iteration % config["save_period"] == 0 and iteration != 0:
# save checkpoint
print("Saving state, iter:", iteration, flush=True)
save_ckpt(net, iteration)
# final checkpoint
save_ckpt(net, iteration)
def main(cfg):
# Setting up GPU args
use_cuda = (cfg.NUM_GPUS > 0) and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {
'num_workers': cfg.DATA_LOADER.NUM_WORKERS,
'pin_memory': cfg.DATA_LOADER.PIN_MEMORY
} if use_cuda else {}
# Auto assign a RNG_SEED when not supplied a value
if cfg.RNG_SEED is None:
cfg.RNG_SEED = np.random.randint(100)
# Using specific GPU
# os.environ['NVIDIA_VISIBLE_DEVICES'] = str(cfg.GPU_ID)
# os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# print("Using GPU : {}.\n".format(cfg.GPU_ID))
# Getting the output directory ready (default is "/output")
cfg.OUT_DIR = os.path.join(os.path.abspath('..'), cfg.OUT_DIR)
if not os.path.exists(cfg.OUT_DIR):
os.mkdir(cfg.OUT_DIR)
# Create "DATASET/MODEL TYPE" specific directory
dataset_out_dir = os.path.join(cfg.OUT_DIR, cfg.DATASET.NAME,
cfg.MODEL.TYPE)
if not os.path.exists(dataset_out_dir):
os.makedirs(dataset_out_dir)
# Creating the experiment directory inside the dataset specific directory
# all logs, labeled, unlabeled, validation sets are stroed here
# E.g., output/CIFAR10/resnet18/{timestamp or cfg.EXP_NAME based on arguments passed}
if cfg.EXP_NAME == 'auto':
now = datetime.now()
exp_dir = f'{now.year}_{now.month}_{now.day}_{now.hour}{now.minute}{now.second}'
else:
exp_dir = cfg.EXP_NAME
exp_dir = os.path.join(dataset_out_dir, exp_dir)
if not os.path.exists(exp_dir):
os.mkdir(exp_dir)
print("Experiment Directory is {}.\n".format(exp_dir))
else:
print(
"Experiment Directory Already Exists: {}. Reusing it may lead to loss of old logs in the directory.\n"
.format(exp_dir))
cfg.EXP_DIR = exp_dir
# Save the config file in EXP_DIR
dump_cfg(cfg)
# Setup Logger
lu.setup_logging(cfg)
# Dataset preparing steps
print("\n======== PREPARING DATA AND MODEL ========\n")
cfg.DATASET.ROOT_DIR = os.path.join(os.path.abspath('..'),
cfg.DATASET.ROOT_DIR)
data_obj = Data(cfg)
train_data, train_size = data_obj.getDataset(save_dir=cfg.DATASET.ROOT_DIR,
isTrain=True,
isDownload=True)
test_data, test_size = data_obj.getDataset(save_dir=cfg.DATASET.ROOT_DIR,
isTrain=False,
isDownload=True)
print(
"\nDataset {} Loaded Sucessfully.\nTotal Train Size: {} and Total Test Size: {}\n"
.format(cfg.DATASET.NAME, train_size, test_size))
logger.info(
"Dataset {} Loaded Sucessfully. Total Train Size: {} and Total Test Size: {}\n"
.format(cfg.DATASET.NAME, train_size, test_size))
lSet_path, uSet_path, valSet_path = data_obj.makeLUVSets(train_split_ratio=cfg.ACTIVE_LEARNING.INIT_L_RATIO, \
val_split_ratio=cfg.DATASET.VAL_RATIO, data=train_data, seed_id=cfg.RNG_SEED, save_dir=cfg.EXP_DIR)
cfg.ACTIVE_LEARNING.LSET_PATH = lSet_path
cfg.ACTIVE_LEARNING.USET_PATH = uSet_path
cfg.ACTIVE_LEARNING.VALSET_PATH = valSet_path
lSet, uSet, valSet = data_obj.loadPartitions(lSetPath=cfg.ACTIVE_LEARNING.LSET_PATH, \
uSetPath=cfg.ACTIVE_LEARNING.USET_PATH, valSetPath = cfg.ACTIVE_LEARNING.VALSET_PATH)
print(
"Data Partitioning Complete. \nLabeled Set: {}, Unlabeled Set: {}, Validation Set: {}\n"
.format(len(lSet), len(uSet), len(valSet)))
logger.info(
"Labeled Set: {}, Unlabeled Set: {}, Validation Set: {}\n".format(
len(lSet), len(uSet), len(valSet)))
# Preparing dataloaders for initial training
lSet_loader = data_obj.getIndexesDataLoader(
indexes=lSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
valSet_loader = data_obj.getIndexesDataLoader(
indexes=valSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
test_loader = data_obj.getTestLoader(data=test_data,
test_batch_size=cfg.TRAIN.BATCH_SIZE,
seed_id=cfg.RNG_SEED)
# Initialize the model.
model = model_builder.build_model(cfg)
print("model: {}\n".format(cfg.MODEL.TYPE))
logger.info("model: {}\n".format(cfg.MODEL.TYPE))
# Construct the optimizer
optimizer = optim.construct_optimizer(cfg, model)
print("optimizer: {}\n".format(optimizer))
logger.info("optimizer: {}\n".format(optimizer))
print("AL Query Method: {}\nMax AL Episodes: {}\n".format(
cfg.ACTIVE_LEARNING.SAMPLING_FN, cfg.ACTIVE_LEARNING.MAX_ITER))
logger.info("AL Query Method: {}\nMax AL Episodes: {}\n".format(
cfg.ACTIVE_LEARNING.SAMPLING_FN, cfg.ACTIVE_LEARNING.MAX_ITER))
for cur_episode in range(0, cfg.ACTIVE_LEARNING.MAX_ITER + 1):
print("======== EPISODE {} BEGINS ========\n".format(cur_episode))
logger.info(
"======== EPISODE {} BEGINS ========\n".format(cur_episode))
# Creating output directory for the episode
episode_dir = os.path.join(cfg.EXP_DIR, f'episode_{cur_episode}')
if not os.path.exists(episode_dir):
os.mkdir(episode_dir)
cfg.EPISODE_DIR = episode_dir
# Train model
print("======== TRAINING ========")
logger.info("======== TRAINING ========")
best_val_acc, best_val_epoch, checkpoint_file = train_model(
lSet_loader, valSet_loader, model, optimizer, cfg)
print("Best Validation Accuracy: {}\nBest Epoch: {}\n".format(
round(best_val_acc, 4), best_val_epoch))
logger.info(
"EPISODE {} Best Validation Accuracy: {}\tBest Epoch: {}\n".format(
cur_episode, round(best_val_acc, 4), best_val_epoch))
# Test best model checkpoint
print("======== TESTING ========\n")
logger.info("======== TESTING ========\n")
test_acc = test_model(test_loader, checkpoint_file, cfg, cur_episode)
print("Test Accuracy: {}.\n".format(round(test_acc, 4)))
logger.info("EPISODE {} Test Accuracy {}.\n".format(
cur_episode, test_acc))
# No need to perform active sampling in the last episode iteration
if cur_episode == cfg.ACTIVE_LEARNING.MAX_ITER:
# Save current lSet, uSet in the final episode directory
data_obj.saveSet(lSet, 'lSet', cfg.EPISODE_DIR)
data_obj.saveSet(uSet, 'uSet', cfg.EPISODE_DIR)
break
# Active Sample
print("======== ACTIVE SAMPLING ========\n")
logger.info("======== ACTIVE SAMPLING ========\n")
al_obj = ActiveLearning(data_obj, cfg)
clf_model = model_builder.build_model(cfg)
clf_model = cu.load_checkpoint(checkpoint_file, clf_model)
activeSet, new_uSet = al_obj.sample_from_uSet(clf_model, lSet, uSet,
train_data)
# Save current lSet, new_uSet and activeSet in the episode directory
data_obj.saveSets(lSet, uSet, activeSet, cfg.EPISODE_DIR)
# Add activeSet to lSet, save new_uSet as uSet and update dataloader for the next episode
lSet = np.append(lSet, activeSet)
uSet = new_uSet
lSet_loader = data_obj.getIndexesDataLoader(
indexes=lSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
valSet_loader = data_obj.getIndexesDataLoader(
indexes=valSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
uSet_loader = data_obj.getSequentialDataLoader(
indexes=uSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
print(
"Active Sampling Complete. After Episode {}:\nNew Labeled Set: {}, New Unlabeled Set: {}, Active Set: {}\n"
.format(cur_episode, len(lSet), len(uSet), len(activeSet)))
logger.info(
"Active Sampling Complete. After Episode {}:\nNew Labeled Set: {}, New Unlabeled Set: {}, Active Set: {}\n"
.format(cur_episode, len(lSet), len(uSet), len(activeSet)))
print("================================\n\n")
logger.info("================================\n\n")
def main(cfg):
# Setting up GPU args
use_cuda = (cfg.NUM_GPUS > 0) and torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {
'num_workers': cfg.DATA_LOADER.NUM_WORKERS,
'pin_memory': cfg.DATA_LOADER.PIN_MEMORY
} if use_cuda else {}
# Using specific GPU
os.environ['NVIDIA_VISIBLE_DEVICES'] = str(cfg.GPU_ID)
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
print("Using GPU : {}.\n".format(cfg.GPU_ID))
# Getting the output directory ready (default is "/output")
cfg.OUT_DIR = os.path.join(os.path.abspath('..'), cfg.OUT_DIR)
if not os.path.exists(cfg.OUT_DIR):
os.makedirs(cfg.OUT_DIR)
# Create "DATASET" specific directory
dataset_out_dir = os.path.join(cfg.OUT_DIR, cfg.DATASET.NAME,
cfg.MODEL.TYPE)
if not os.path.exists(dataset_out_dir):
os.makedirs(dataset_out_dir)
# Creating the experiment directory inside the dataset specific directory
# all logs, labeled, unlabeled, validation sets are stroed here
# E.g., output/CIFAR10/resnet18/{timestamp or cfg.EXP_NAME based on arguments passed}
if cfg.EXP_NAME == 'auto':
now = datetime.now()
exp_dir = f'{now.year}_{now.month}_{now.day}_{now.hour}{now.minute}{now.second}'
else:
exp_dir = cfg.EXP_NAME
exp_dir = os.path.join(dataset_out_dir, exp_dir)
if not os.path.exists(exp_dir):
os.mkdir(exp_dir)
print("Experiment Directory is {}.\n".format(exp_dir))
else:
print(
"Experiment Directory Already Exists: {}. Reusing it may lead to loss of old logs in the directory.\n"
.format(exp_dir))
cfg.EXP_DIR = exp_dir
# Setup Logger
lu.setup_logging(cfg)
# Dataset preparing steps
print("\n======== PREPARING DATA AND MODEL ========\n")
cfg.DATASET.ROOT_DIR = os.path.join(os.path.abspath('..'),
cfg.DATASET.ROOT_DIR)
data_obj = Data(cfg)
train_data, train_size = data_obj.getDataset(save_dir=cfg.DATASET.ROOT_DIR,
isTrain=True,
isDownload=True)
train_data = RotNetDataset(cfg.DATASET.NAME, train_data)
train_size = len(train_data)
print("\n Rotation Dataset {} Loaded Sucessfully.\nTotal Train Size: {}\n".
format(cfg.DATASET.NAME, train_size))
logger.info(
"Rotation Dataset {} Loaded Sucessfully. Total Train Size: {}\n".
format(cfg.DATASET.NAME, train_size))
trainSet_path, valSet_path = data_obj.makeTVSets(val_split_ratio=cfg.DATASET.VAL_RATIO, data=train_data,\
seed_id=cfg.RNG_SEED, save_dir=cfg.EXP_DIR)
cfg.INIT_POOL.TRAINSET_PATH = trainSet_path
cfg.INIT_POOL.VALSET_PATH = valSet_path
trainSet, valSet = data_obj.loadTVPartitions(
trainSetPath=cfg.INIT_POOL.TRAINSET_PATH,
valSetPath=cfg.INIT_POOL.VALSET_PATH)
print("Data Partitioning Complete. \nTrain Set: {}, Validation Set: {}\n".
format(len(trainSet), len(valSet)))
logger.info("Train Set: {}, Validation Set: {}\n".format(
len(trainSet), len(valSet)))
# Preparing dataloaders for initial training
trainSet_loader = data_obj.getSequentialDataLoader(
indexes=trainSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
valSet_loader = data_obj.getSequentialDataLoader(
indexes=valSet, batch_size=cfg.TRAIN.BATCH_SIZE, data=train_data)
# Initialize the model.
model = model_builder.build_model(cfg)
print("model: {}\n".format(cfg.MODEL.TYPE))
logger.info("model: {}\n".format(cfg.MODEL.TYPE))
# Construct the optimizer
optimizer = optim.construct_optimizer(cfg, model)
print("optimizer: {}\n".format(optimizer))
logger.info("optimizer: {}\n".format(optimizer))
# This is to seamlessly use the code originally written for AL episodes
cfg.EPISODE_DIR = cfg.EXP_DIR
# Train model
print("======== ROTATION TRAINING ========")
logger.info("======== ROTATION TRAINING ========")
best_val_acc, best_val_epoch, checkpoint_file = train_model(
trainSet_loader, valSet_loader, model, optimizer, cfg)
print("Best Validation Accuracy: {}\nBest Epoch: {}\n".format(
round(best_val_acc, 4), best_val_epoch))
logger.info("Best Validation Accuracy: {}\tBest Epoch: {}\n".format(
round(best_val_acc, 4), best_val_epoch))
# Test best model checkpoint
print("======== ROTATION TESTING ========\n")
logger.info("======== ROTATION TESTING ========\n")
test_acc = test_model(trainSet_loader, checkpoint_file, cfg, cur_episode=1)
print("Test Accuracy: {}.\n".format(round(test_acc, 4)))
logger.info("Test Accuracy {}.\n".format(test_acc))
print("================================\n\n")
logger.info("================================\n\n")
