def main():
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
np.random.seed(args.seed)
torch.cuda.set_device(args.gpu)
cudnn.benchmark = True
torch.manual_seed(args.seed)
cudnn.enabled = True
torch.cuda.manual_seed(args.seed)
logging.info('gpu device = %d' % args.gpu)
logging.info("args = %s", args)
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
number_of_classes = class_dict[args.dataset]
in_channels = inp_channel_dict[args.dataset]
model = Network(args.init_channels, number_of_classes, args.layers,
criterion, in_channels)
model = model.cuda()
logging.info("param size = %fMB", utils.count_parameters_in_MB(model))
optimizer = torch.optim.SGD(model.parameters(),
args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
# Get transforms to apply on data
train_transform, valid_transform = utils.get_data_transforms(args)
# Get the training queue
train_queue, valid_queue = get_training_queues(args, train_transform)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizer, float(args.epochs), eta_min=args.learning_rate_min)
architect = Architect(model, args)
for epoch in range(args.epochs):
scheduler.step()
lr = scheduler.get_lr()[0]
logging.info('epoch %d lr %e', epoch, lr)
genotype = model.genotype()
logging.info('genotype = %s', genotype)
print(F.softmax(model.alphas_normal, dim=-1))
print(F.softmax(model.alphas_reduce, dim=-1))
# training
train_acc, train_obj = train(train_queue, valid_queue, model,
architect, criterion, optimizer, lr)
logging.info('train_acc %f', train_acc)
# validation
valid_acc, valid_obj = infer(valid_queue, model, criterion)
logging.info('valid_acc %f', valid_acc)
utils.save(model, os.path.join(args.save, 'weights.pt'))
def main():
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
np.random.seed(args.seed)
torch.cuda.set_device(args.gpu)
cudnn.benchmark = True
torch.manual_seed(args.seed)
cudnn.enabled=True
torch.cuda.manual_seed(args.seed)
logging.info('gpu device = %d' % args.gpu)
logging.info("args = %s", args)
genotype = eval("genotypes.%s" % args.arch)
number_of_classes = class_dict[args.dataset]
in_channels = inp_channel_dict[args.dataset]
print(number_of_classes, in_channels)
model = Network(args.init_channels, number_of_classes, args.layers, args.auxiliary, genotype, in_channels)
model = model.cuda()
logging.info("param size = %fMB", utils.count_parameters_in_MB(model))
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
optimizer = torch.optim.SGD(
model.parameters(),
args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay
)
train_transform, valid_transform = utils.get_data_transforms(args)
train_queue, valid_queue = get_train_test_queues(args, train_transform, valid_transform)
# train_data = dset.CIFAR10(root=args.data, train=True, download=True, transform=train_transform)
# valid_data = dset.CIFAR10(root=args.data, train=False, download=True, transform=valid_transform)
# train_queue = torch.utils.data.DataLoader(
# train_data, batch_size=args.batch_size, shuffle=True, pin_memory=True, num_workers=2)
# valid_queue = torch.utils.data.DataLoader(
# valid_data, batch_size=args.batch_size, shuffle=False, pin_memory=True, num_workers=2)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, float(args.epochs))
for epoch in range(args.epochs):
scheduler.step()
logging.info('epoch %d lr %e', epoch, scheduler.get_lr()[0])
model.drop_path_prob = args.drop_path_prob * epoch / args.epochs
train_acc, train_obj = train(train_queue, model, criterion, optimizer)
logging.info('train_acc %f', train_acc)
valid_acc, valid_obj = infer(valid_queue, model, criterion)
logging.info('valid_acc %f', valid_acc)
utils.save(model, os.path.join(args.save, 'weights.pt'))
def predict_all(files, idx_to_class, model_path):
device = get_device()
model = models.resnet18()
model.fc = nn.Linear(512, num_classes)
checkpoint = torch.load(model_path, map_location=device)
model.load_state_dict(checkpoint['model_state_dict'])
model.to(device)
data_transforms = get_data_transforms()['val']
tmp = []
for f in files:
image = Image.open(f)
image = data_transforms(image)
image.unsqueeze_(dim=0)
image = image.to(device)
output = model(image)
target_idx = torch.argmax(output).item()
target_name = idx_to_class[target_idx]
head, tail = os.path.split(f)
tmp.append({'fname': tail, 'prediction': target_name})
return tmp
def main():
parser = argparse.ArgumentParser("Common Argument Parser")
parser.add_argument('--data', type=str, default='../data', help='location of the data corpus')
parser.add_argument('--dataset', type=str, default='cifar10', help='which dataset:\
cifar10, mnist, emnist, fashion, svhn, stl10, devanagari')
parser.add_argument('--batch_size', type=int, default=64, help='batch size')
parser.add_argument('--learning_rate', type=float, default=0.025, help='init learning rate')
parser.add_argument('--learning_rate_min', type=float, default=1e-8, help='min learning rate')
parser.add_argument('--lr_power_annealing_exponent_order', type=float, default=2,
help='Cosine Power Annealing Schedule Base, larger numbers make '
'the exponential more dominant, smaller make cosine more dominant, '
'1 returns to standard cosine annealing.')
parser.add_argument('--momentum', type=float, default=0.9, help='momentum')
parser.add_argument('--weight_decay', '--wd', dest='weight_decay', type=float, default=3e-4, help='weight decay')
parser.add_argument('--partial', default=1/8, type=float, help='partially adaptive parameter p in Padam')
parser.add_argument('--report_freq', type=float, default=50, help='report frequency')
parser.add_argument('--gpu', type=int, default=0, help='gpu device id')
parser.add_argument('--epochs', type=int, default=2000, help='num of training epochs')
parser.add_argument('--start_epoch', default=1, type=int, metavar='N',
help='manual epoch number (useful for restarts)')
parser.add_argument('--warmup_epochs', type=int, default=5, help='num of warmup training epochs')
parser.add_argument('--warm_restarts', type=int, default=20, help='warm restarts of cosine annealing')
parser.add_argument('--init_channels', type=int, default=36, help='num of init channels')
parser.add_argument('--mid_channels', type=int, default=32, help='C_mid channels in choke SharpSepConv')
parser.add_argument('--layers', type=int, default=20, help='total number of layers')
parser.add_argument('--model_path', type=str, default='saved_models', help='path to save the model')
parser.add_argument('--auxiliary', action='store_true', default=False, help='use auxiliary tower')
parser.add_argument('--mixed_auxiliary', action='store_true', default=False, help='Learn weights for auxiliary networks during training. Overrides auxiliary flag')
parser.add_argument('--auxiliary_weight', type=float, default=0.4, help='weight for auxiliary loss')
parser.add_argument('--cutout', action='store_true', default=False, help='use cutout')
parser.add_argument('--cutout_length', type=int, default=16, help='cutout length')
parser.add_argument('--autoaugment', action='store_true', default=False, help='use cifar10 autoaugment https://arxiv.org/abs/1805.09501')
parser.add_argument('--random_eraser', action='store_true', default=False, help='use random eraser')
parser.add_argument('--drop_path_prob', type=float, default=0.2, help='drop path probability')
parser.add_argument('--save', type=str, default='EXP', help='experiment name')
parser.add_argument('--seed', type=int, default=0, help='random seed')
parser.add_argument('--arch', type=str, default='DARTS', help='which architecture to use')
parser.add_argument('--ops', type=str, default='OPS', help='which operations to use, options are OPS and DARTS_OPS')
parser.add_argument('--primitives', type=str, default='PRIMITIVES',
help='which primitive layers to use inside a cell search space,'
' options are PRIMITIVES, SHARPER_PRIMITIVES, and DARTS_PRIMITIVES')
parser.add_argument('--optimizer', type=str, default='sgd', help='which optimizer to use, options are padam and sgd')
parser.add_argument('--load', type=str, default='', metavar='PATH', help='load weights at specified location')
parser.add_argument('--grad_clip', type=float, default=5, help='gradient clipping')
parser.add_argument('--flops', action='store_true', default=False, help='count flops and exit, aka floating point operations.')
parser.add_argument('-e', '--evaluate', dest='evaluate', type=str, metavar='PATH', default='',
help='evaluate model at specified path on training, test, and validation datasets')
parser.add_argument('--multi_channel', action='store_true', default=False, help='perform multi channel search, a completely separate search space')
parser.add_argument('--load_args', type=str, default='', metavar='PATH',
help='load command line args from a json file, this will override '
'all currently set args except for --evaluate, and arguments '
'that did not exist when the json file was originally saved out.')
parser.add_argument('--layers_of_cells', type=int, default=8, help='total number of cells in the whole network, default is 8 cells')
parser.add_argument('--layers_in_cells', type=int, default=4,
help='Total number of nodes in each cell, aka number of steps,'
' default is 4 nodes, which implies 8 ops')
parser.add_argument('--weighting_algorithm', type=str, default='scalar',
help='which operations to use, options are '
'"max_w" (1. - max_w + w) * op, and scalar (w * op)')
# TODO(ahundt) remove final path and switch back to genotype
parser.add_argument('--load_genotype', type=str, default=None, help='Name of genotype to be used')
parser.add_argument('--simple_path', default=True, action='store_false', help='Final model is a simple path (MultiChannelNetworkModel)')
args = parser.parse_args()
args = utils.initialize_files_and_args(args)
logger = utils.logging_setup(args.log_file_path)
if not torch.cuda.is_available():
logger.info('no gpu device available')
sys.exit(1)
np.random.seed(args.seed)
torch.cuda.set_device(args.gpu)
cudnn.benchmark = True
torch.manual_seed(args.seed)
cudnn.enabled=True
torch.cuda.manual_seed(args.seed)
logger.info('gpu device = %d' % args.gpu)
logger.info("args = %s", args)
DATASET_CLASSES = dataset.class_dict[args.dataset]
DATASET_CHANNELS = dataset.inp_channel_dict[args.dataset]
DATASET_MEAN = dataset.mean_dict[args.dataset]
DATASET_STD = dataset.std_dict[args.dataset]
logger.info('output channels: ' + str(DATASET_CLASSES))
# # load the correct ops dictionary
op_dict_to_load = "operations.%s" % args.ops
logger.info('loading op dict: ' + str(op_dict_to_load))
op_dict = eval(op_dict_to_load)
# load the correct primitives list
primitives_to_load = "genotypes.%s" % args.primitives
logger.info('loading primitives:' + primitives_to_load)
primitives = eval(primitives_to_load)
logger.info('primitives: ' + str(primitives))
genotype = eval("genotypes.%s" % args.arch)
# create the neural network
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
if args.multi_channel:
final_path = None
if args.load_genotype is not None:
genotype = getattr(genotypes, args.load_genotype)
print(genotype)
if type(genotype[0]) is str:
logger.info('Path :%s', genotype)
# TODO(ahundt) remove final path and switch back to genotype
cnn_model = MultiChannelNetwork(
args.init_channels, DATASET_CLASSES, layers=args.layers_of_cells, criterion=criterion, steps=args.layers_in_cells,
weighting_algorithm=args.weighting_algorithm, genotype=genotype)
flops_shape = [1, 3, 32, 32]
elif args.dataset == 'imagenet':
cnn_model = NetworkImageNet(args.init_channels, DATASET_CLASSES, args.layers, args.auxiliary, genotype, op_dict=op_dict, C_mid=args.mid_channels)
flops_shape = [1, 3, 224, 224]
else:
cnn_model = NetworkCIFAR(args.init_channels, DATASET_CLASSES, args.layers, args.auxiliary, genotype, op_dict=op_dict, C_mid=args.mid_channels)
flops_shape = [1, 3, 32, 32]
cnn_model = cnn_model.cuda()
logger.info("param size = %fMB", utils.count_parameters_in_MB(cnn_model))
if args.flops:
logger.info('flops_shape = ' + str(flops_shape))
logger.info("flops = " + utils.count_model_flops(cnn_model, data_shape=flops_shape))
return
optimizer = torch.optim.SGD(
cnn_model.parameters(),
args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay
)
# Get preprocessing functions (i.e. transforms) to apply on data
train_transform, valid_transform = utils.get_data_transforms(args)
if args.evaluate:
# evaluate the train dataset without augmentation
train_transform = valid_transform
# Get the training queue, use full training and test set
train_queue, valid_queue = dataset.get_training_queues(
args.dataset, train_transform, valid_transform, args.data, args.batch_size, train_proportion=1.0, search_architecture=False)
test_queue = None
if args.dataset == 'cifar10':
# evaluate best model weights on cifar 10.1
# https://github.com/modestyachts/CIFAR-10.1
test_data = cifar10_1.CIFAR10_1(root=args.data, download=True, transform=valid_transform)
test_queue = torch.utils.data.DataLoader(
test_data, batch_size=args.batch_size, shuffle=False, pin_memory=True, num_workers=8)
if args.evaluate:
# evaluate the loaded model, print the result, and return
logger.info("Evaluating inference with weights file: " + args.load)
eval_stats = evaluate(
args, cnn_model, criterion, args.load,
train_queue=train_queue, valid_queue=valid_queue, test_queue=test_queue)
with open(args.stats_file, 'w') as f:
arg_dict = vars(args)
arg_dict.update(eval_stats)
json.dump(arg_dict, f)
logger.info("flops = " + utils.count_model_flops(cnn_model))
logger.info(utils.dict_to_log_string(eval_stats))
logger.info('\nEvaluation of Loaded Model Complete! Save dir: ' + str(args.save))
return
lr_schedule = cosine_power_annealing(
epochs=args.epochs, max_lr=args.learning_rate, min_lr=args.learning_rate_min,
warmup_epochs=args.warmup_epochs, exponent_order=args.lr_power_annealing_exponent_order)
epochs = np.arange(args.epochs) + args.start_epoch
# scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, float(args.epochs))
epoch_stats = []
stats_csv = args.epoch_stats_file
stats_csv = stats_csv.replace('.json', '.csv')
with tqdm(epochs, dynamic_ncols=True) as prog_epoch:
best_valid_acc = 0.0
best_epoch = 0
best_stats = {}
stats = {}
epoch_stats = []
weights_file = os.path.join(args.save, 'weights.pt')
for epoch, learning_rate in zip(prog_epoch, lr_schedule):
# update the drop_path_prob augmentation
cnn_model.drop_path_prob = args.drop_path_prob * epoch / args.epochs
# update the learning rate
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
# scheduler.get_lr()[0]
train_acc, train_obj = train(args, train_queue, cnn_model, criterion, optimizer)
val_stats = infer(args, valid_queue, cnn_model, criterion)
stats.update(val_stats)
stats['train_acc'] = train_acc
stats['train_loss'] = train_obj
stats['lr'] = learning_rate
stats['epoch'] = epoch
if stats['valid_acc'] > best_valid_acc:
# new best epoch, save weights
utils.save(cnn_model, weights_file)
best_epoch = epoch
best_stats.update(copy.deepcopy(stats))
best_valid_acc = stats['valid_acc']
best_train_loss = train_obj
best_train_acc = train_acc
# else:
# # not best epoch, load best weights
# utils.load(cnn_model, weights_file)
logger.info('epoch, %d, train_acc, %f, valid_acc, %f, train_loss, %f, valid_loss, %f, lr, %e, best_epoch, %d, best_valid_acc, %f, ' + utils.dict_to_log_string(stats),
epoch, train_acc, stats['valid_acc'], train_obj, stats['valid_loss'], learning_rate, best_epoch, best_valid_acc)
stats['train_acc'] = train_acc
stats['train_loss'] = train_obj
epoch_stats += [copy.deepcopy(stats)]
with open(args.epoch_stats_file, 'w') as f:
json.dump(epoch_stats, f, cls=utils.NumpyEncoder)
utils.list_of_dicts_to_csv(stats_csv, epoch_stats)
# get stats from best epoch including cifar10.1
eval_stats = evaluate(args, cnn_model, criterion, weights_file, train_queue, valid_queue, test_queue)
with open(args.stats_file, 'w') as f:
arg_dict = vars(args)
arg_dict.update(eval_stats)
json.dump(arg_dict, f, cls=utils.NumpyEncoder)
with open(args.epoch_stats_file, 'w') as f:
json.dump(epoch_stats, f, cls=utils.NumpyEncoder)
logger.info(utils.dict_to_log_string(eval_stats))
logger.info('Training of Final Model Complete! Save dir: ' + str(args.save))
def main():
global best_top1, args, logger
args.distributed = False
if 'WORLD_SIZE' in os.environ:
args.distributed = int(os.environ['WORLD_SIZE']) > 1
# commented because it is now set as an argparse param.
# args.gpu = 0
args.world_size = 1
if args.distributed:
args.gpu = args.local_rank % torch.cuda.device_count()
torch.cuda.set_device(args.gpu)
torch.distributed.init_process_group(backend='nccl',
init_method='env://')
args.world_size = torch.distributed.get_world_size()
# note the gpu is used for directory creation and log files
# which is needed when run as multiple processes
args = utils.initialize_files_and_args(args)
logger = utils.logging_setup(args.log_file_path)
if args.fp16:
assert torch.backends.cudnn.enabled, "fp16 mode requires cudnn backend to be enabled."
if args.static_loss_scale != 1.0:
if not args.fp16:
logger.info(
"Warning: if --fp16 is not used, static_loss_scale will be ignored."
)
# # load the correct ops dictionary
op_dict_to_load = "operations.%s" % args.ops
logger.info('loading op dict: ' + str(op_dict_to_load))
op_dict = eval(op_dict_to_load)
# load the correct primitives list
primitives_to_load = "genotypes.%s" % args.primitives
logger.info('loading primitives:' + primitives_to_load)
primitives = eval(primitives_to_load)
logger.info('primitives: ' + str(primitives))
# create model
genotype = eval("genotypes.%s" % args.arch)
# get the number of output channels
classes = dataset.class_dict[args.dataset]
# create the neural network
if args.dataset == 'imagenet':
model = NetworkImageNet(args.init_channels,
classes,
args.layers,
args.auxiliary,
genotype,
op_dict=op_dict,
C_mid=args.mid_channels)
flops_shape = [1, 3, 224, 224]
else:
model = NetworkCIFAR(args.init_channels,
classes,
args.layers,
args.auxiliary,
genotype,
op_dict=op_dict,
C_mid=args.mid_channels)
flops_shape = [1, 3, 32, 32]
model.drop_path_prob = 0.0
# if args.pretrained:
# logger.info("=> using pre-trained model '{}'".format(args.arch))
# model = models.__dict__[args.arch](pretrained=True)
# else:
# logger.info("=> creating model '{}'".format(args.arch))
# model = models.__dict__[args.arch]()
if args.flops:
model = model.cuda()
logger.info("param size = %fMB", utils.count_parameters_in_MB(model))
logger.info("flops_shape = " + str(flops_shape))
logger.info("flops = " +
utils.count_model_flops(model, data_shape=flops_shape))
return
if args.sync_bn:
import apex
logger.info("using apex synced BN")
model = apex.parallel.convert_syncbn_model(model)
model = model.cuda()
if args.fp16:
model = network_to_half(model)
if args.distributed:
# By default, apex.parallel.DistributedDataParallel overlaps communication with
# computation in the backward pass.
# model = DDP(model)
# delay_allreduce delays all communication to the end of the backward pass.
model = DDP(model, delay_allreduce=True)
# define loss function (criterion) and optimizer
criterion = nn.CrossEntropyLoss().cuda()
# Scale learning rate based on global batch size
args.learning_rate = args.learning_rate * float(
args.batch_size * args.world_size) / 256.
init_lr = args.learning_rate / args.warmup_lr_divisor
optimizer = torch.optim.SGD(model.parameters(),
init_lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
# epoch_count = args.epochs - args.start_epoch
# scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, float(epoch_count))
# scheduler = warmup_scheduler.GradualWarmupScheduler(
# optimizer, args.warmup_lr_divisor, args.warmup_epochs, scheduler)
if args.fp16:
optimizer = FP16_Optimizer(optimizer,
static_loss_scale=args.static_loss_scale,
dynamic_loss_scale=args.dynamic_loss_scale)
# Optionally resume from a checkpoint
if args.resume or args.evaluate:
if args.evaluate:
args.resume = args.evaluate
# Use a local scope to avoid dangling references
def resume():
if os.path.isfile(args.resume):
logger.info("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(
args.resume,
map_location=lambda storage, loc: storage.cuda(args.gpu))
args.start_epoch = checkpoint['epoch']
if 'best_top1' in checkpoint:
best_top1 = checkpoint['best_top1']
model.load_state_dict(checkpoint['state_dict'])
# An FP16_Optimizer instance's state dict internally stashes the master params.
optimizer.load_state_dict(checkpoint['optimizer'])
# TODO(ahundt) make sure scheduler loading isn't broken
if 'lr_scheduler' in checkpoint:
scheduler.load_state_dict(checkpoint['lr_scheduler'])
elif 'lr_schedule' in checkpoint:
lr_schedule = checkpoint['lr_schedule']
logger.info("=> loaded checkpoint '{}' (epoch {})".format(
args.resume, checkpoint['epoch']))
else:
logger.info("=> no checkpoint found at '{}'".format(
args.resume))
resume()
# # Data loading code
# traindir = os.path.join(args.data, 'train')
# valdir = os.path.join(args.data, 'val')
# if(args.arch == "inception_v3"):
# crop_size = 299
# val_size = 320 # I chose this value arbitrarily, we can adjust.
# else:
# crop_size = 224
# val_size = 256
# train_dataset = datasets.ImageFolder(
# traindir,
# transforms.Compose([
# transforms.RandomResizedCrop(crop_size),
# transforms.RandomHorizontalFlip(),
# autoaugment.ImageNetPolicy(),
# # transforms.ToTensor(), # Too slow, moved to data_prefetcher()
# # normalize,
# ]))
# val_dataset = datasets.ImageFolder(valdir, transforms.Compose([
# transforms.Resize(val_size),
# transforms.CenterCrop(crop_size)
# ]))
# train_sampler = None
# val_sampler = None
# if args.distributed:
# train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
# val_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset)
# train_loader = torch.utils.data.DataLoader(
# train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None),
# num_workers=args.workers, pin_memory=True, sampler=train_sampler, collate_fn=fast_collate)
# val_loader = torch.utils.data.DataLoader(
# val_dataset,
# batch_size=args.batch_size, shuffle=False,
# num_workers=args.workers, pin_memory=True,
# sampler=val_sampler,
# collate_fn=fast_collate)
# Get preprocessing functions (i.e. transforms) to apply on data
# normalize_as_tensor = False because we normalize and convert to a
# tensor in our custom prefetching function, rather than as part of
# the transform preprocessing list.
train_transform, valid_transform = utils.get_data_transforms(
args, normalize_as_tensor=False)
# Get the training queue, select training and validation from training set
train_loader, val_loader = dataset.get_training_queues(
args.dataset,
train_transform,
valid_transform,
args.data,
args.batch_size,
train_proportion=1.0,
collate_fn=fast_collate,
distributed=args.distributed,
num_workers=args.workers)
if args.evaluate:
if args.dataset == 'cifar10':
# evaluate best model weights on cifar 10.1
# https://github.com/modestyachts/CIFAR-10.1
train_transform, valid_transform = utils.get_data_transforms(args)
# Get the training queue, select training and validation from training set
# Get the training queue, use full training and test set
train_queue, valid_queue = dataset.get_training_queues(
args.dataset,
train_transform,
valid_transform,
args.data,
args.batch_size,
train_proportion=1.0,
search_architecture=False)
test_data = cifar10_1.CIFAR10_1(root=args.data,
download=True,
transform=valid_transform)
test_queue = torch.utils.data.DataLoader(
test_data,
batch_size=args.batch_size,
shuffle=False,
pin_memory=True,
num_workers=args.workers)
eval_stats = evaluate(args,
model,
criterion,
train_queue=train_queue,
valid_queue=valid_queue,
test_queue=test_queue)
with open(args.stats_file, 'w') as f:
# TODO(ahundt) fix "TypeError: 1869 is not JSON serializable" to include arg info, see train.py
# arg_dict = vars(args)
# arg_dict.update(eval_stats)
# json.dump(arg_dict, f)
json.dump(eval_stats, f)
logger.info("flops = " + utils.count_model_flops(model))
logger.info(utils.dict_to_log_string(eval_stats))
logger.info('\nEvaluation of Loaded Model Complete! Save dir: ' +
str(args.save))
else:
validate(val_loader, model, criterion, args)
return
lr_schedule = cosine_power_annealing(
epochs=args.epochs,
max_lr=args.learning_rate,
min_lr=args.learning_rate_min,
warmup_epochs=args.warmup_epochs,
exponent_order=args.lr_power_annealing_exponent_order,
restart_lr=args.restart_lr)
epochs = np.arange(args.epochs) + args.start_epoch
stats_csv = args.epoch_stats_file
stats_csv = stats_csv.replace('.json', '.csv')
with tqdm(epochs,
dynamic_ncols=True,
disable=args.local_rank != 0,
leave=False) as prog_epoch:
best_stats = {}
stats = {}
epoch_stats = []
best_epoch = 0
for epoch, learning_rate in zip(prog_epoch, lr_schedule):
if args.distributed and train_loader.sampler is not None:
train_loader.sampler.set_epoch(int(epoch))
# if args.distributed:
# train_sampler.set_epoch(epoch)
# update the learning rate
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
# scheduler.step()
model.drop_path_prob = args.drop_path_prob * float(epoch) / float(
args.epochs)
# train for one epoch
train_stats = train(train_loader, model, criterion, optimizer,
int(epoch), args)
if args.prof:
break
# evaluate on validation set
top1, val_stats = validate(val_loader, model, criterion, args)
stats.update(train_stats)
stats.update(val_stats)
# stats['lr'] = '{0:.5f}'.format(scheduler.get_lr()[0])
stats['lr'] = '{0:.5f}'.format(learning_rate)
stats['epoch'] = epoch
# remember best top1 and save checkpoint
if args.local_rank == 0:
is_best = top1 > best_top1
best_top1 = max(top1, best_top1)
stats['best_top1'] = '{0:.3f}'.format(best_top1)
if is_best:
best_epoch = epoch
best_stats = copy.deepcopy(stats)
stats['best_epoch'] = best_epoch
stats_str = utils.dict_to_log_string(stats)
logger.info(stats_str)
save_checkpoint(
{
'epoch': epoch,
'arch': args.arch,
'state_dict': model.state_dict(),
'best_top1': best_top1,
'optimizer': optimizer.state_dict(),
# 'lr_scheduler': scheduler.state_dict()
'lr_schedule': lr_schedule,
'stats': best_stats
},
is_best,
path=args.save)
prog_epoch.set_description(
'Overview ***** best_epoch: {0} best_valid_top1: {1:.2f} ***** Progress'
.format(best_epoch, best_top1))
epoch_stats += [copy.deepcopy(stats)]
with open(args.epoch_stats_file, 'w') as f:
json.dump(epoch_stats, f, cls=utils.NumpyEncoder)
utils.list_of_dicts_to_csv(stats_csv, epoch_stats)
stats_str = utils.dict_to_log_string(best_stats, key_prepend='best_')
logger.info(stats_str)
with open(args.stats_file, 'w') as f:
arg_dict = vars(args)
arg_dict.update(best_stats)
json.dump(arg_dict, f, cls=utils.NumpyEncoder)
with open(args.epoch_stats_file, 'w') as f:
json.dump(epoch_stats, f, cls=utils.NumpyEncoder)
utils.list_of_dicts_to_csv(stats_csv, epoch_stats)
logger.info('Training of Final Model Complete! Save dir: ' +
str(args.save))
def main():
if not torch.cuda.is_available():
logger.info('no gpu device available')
sys.exit(1)
np.random.seed(args.seed)
torch.cuda.set_device(args.gpu)
cudnn.benchmark = True
torch.manual_seed(args.seed)
cudnn.enabled = True
torch.cuda.manual_seed(args.seed)
logger.info('gpu device = %d' % args.gpu)
logger.info("args = %s", args)
# # load the correct ops dictionary
op_dict_to_load = "operations.%s" % args.ops
logger.info('loading op dict: ' + str(op_dict_to_load))
op_dict = eval(op_dict_to_load)
# load the correct primitives list
primitives_to_load = "genotypes.%s" % args.primitives
logger.info('loading primitives:' + primitives_to_load)
primitives = eval(primitives_to_load)
logger.info('primitives: ' + str(primitives))
criterion = nn.CrossEntropyLoss()
criterion = criterion.cuda()
if args.multi_channel:
final_path = None
if args.final_path is not None:
final_path = np.load(args.final_path)
genotype = None
if args.load_genotype is not None:
genotype = getattr(genotypes, args.load_genotype)
cnn_model = model_search.MultiChannelNetwork(
args.init_channels,
CIFAR_CLASSES,
layers=args.layers_of_cells,
criterion=criterion,
steps=args.layers_in_cells,
primitives=primitives,
op_dict=op_dict,
weighting_algorithm=args.weighting_algorithm,
genotype=genotype)
#save_graph(cnn_model.G, os.path.join(args.save, 'network_graph.pdf'))
if args.load_genotype is not None:
# TODO(ahundt) support other batch shapes
data_shape = [1, 3, 32, 32]
batch = torch.zeros(data_shape)
cnn_model(batch)
logger.info("loaded genotype_raw_weights = " +
str(cnn_model.genotype('raw_weights')))
logger.info("loaded genotype_longest_path = " +
str(cnn_model.genotype('longest_path')))
logger.info("loaded genotype greedy_path = " +
str(gen_greedy_path(cnn_model.G, strategy="top_down")))
logger.info(
"loaded genotype greedy_path_bottom_up = " +
str(gen_greedy_path(cnn_model.G, strategy="bottom_up")))
# TODO(ahundt) support other layouts
else:
cnn_model = model_search.Network(
args.init_channels,
CIFAR_CLASSES,
layers=args.layers_of_cells,
criterion=criterion,
steps=args.layers_in_cells,
primitives=primitives,
op_dict=op_dict,
weights_are_parameters=args.no_architect,
C_mid=args.mid_channels,
weighting_algorithm=args.weighting_algorithm)
cnn_model = cnn_model.cuda()
logger.info("param size = %fMB", utils.count_parameters_in_MB(cnn_model))
if args.load:
logger.info('loading weights from: ' + args.load)
utils.load(cnn_model, args.load)
optimizer = torch.optim.SGD(cnn_model.parameters(),
args.learning_rate,
momentum=args.momentum,
weight_decay=args.weight_decay)
# Get preprocessing functions (i.e. transforms) to apply on data
train_transform, valid_transform = utils.get_data_transforms(args)
# Get the training queue, select training and validation from training set
train_queue, valid_queue = dataset.get_training_queues(
args.dataset,
train_transform,
valid_transform,
args.data,
args.batch_size,
args.train_portion,
search_architecture=True)
lr_schedule = cosine_power_annealing(
epochs=args.epochs,
max_lr=args.learning_rate,
min_lr=args.learning_rate_min,
warmup_epochs=args.warmup_epochs,
exponent_order=args.lr_power_annealing_exponent_order)
epochs = np.arange(args.epochs) + args.start_epoch
# scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
# optimizer, float(args.epochs), eta_min=args.learning_rate_min)
if args.no_architect:
architect = None
else:
architect = Architect(cnn_model, args)
epoch_stats = []
stats_csv = args.epoch_stats_file
stats_csv = stats_csv.replace('.json', '.csv')
with tqdm(epochs, dynamic_ncols=True) as prog_epoch:
best_valid_acc = 0.0
best_epoch = 0
# state_dict = {}
# og_state_keys = set()
# updated_state_keys = set()
#saving state_dict for debugging weights by comparison
# for key in cnn_model.state_dict():
# state_dict[key] = cnn_model.state_dict()[key].clone()
# # logger.info('layer = {}'.format(key))
# logger.info('Total keys in state_dict = {}'.format(len(cnn_model.state_dict().keys())))
# og_state_keys.update(cnn_model.state_dict().keys())
best_stats = {}
weights_file = os.path.join(args.save, 'weights.pt')
for epoch, learning_rate in zip(prog_epoch, lr_schedule):
# scheduler.step()
# lr = scheduler.get_lr()[0]
for param_group in optimizer.param_groups:
param_group['lr'] = learning_rate
genotype = None
if args.final_path is None:
genotype = cnn_model.genotype()
logger.info('genotype = %s', genotype)
if not args.multi_channel:
# the genotype is the alphas in the multi-channel case
# print the alphas in other cases
logger.info('alphas_normal = %s', cnn_model.arch_weights(0))
logger.info('alphas_reduce = %s', cnn_model.arch_weights(1))
# training
train_acc, train_obj = train(train_queue, valid_queue, cnn_model,
architect, criterion, optimizer,
learning_rate)
if args.multi_channel and args.final_path is None:
# TODO(ahundt) remove final path and switch back to genotype, and save out raw weights plus optimal path
optimal_path = nx.algorithms.dag.dag_longest_path(cnn_model.G)
optimal_path_filename = os.path.join(
args.save, 'longest_path_layer_sequence.npy')
logger.info('Saving model layer sequence object: ' +
str(optimal_path_filename))
np.save(optimal_path_filename, optimal_path)
graph_filename = os.path.join(
args.save, 'network_graph_' + str(epoch) + '.graph')
logger.info('Saving updated weight graph: ' +
str(graph_filename))
nx.write_gpickle(cnn_model.G, graph_filename)
logger.info('optimal_path : %s', optimal_path)
# validation
valid_acc, valid_obj = infer(valid_queue, cnn_model, criterion)
if valid_acc > best_valid_acc:
# new best epoch, save weights
utils.save(cnn_model, weights_file)
if args.multi_channel:
graph_filename = os.path.join(
args.save,
'network_graph_best_valid' + str(epoch) + '.graph')
logger.info('Saving updated weight graph: ' +
str(graph_filename))
best_epoch = epoch
best_valid_acc = valid_acc
prog_epoch.set_description(
'Overview ***** best_epoch: {0} best_valid_acc: {1:.2f} ***** Progress'
.format(best_epoch, best_valid_acc))
logger.info(
'epoch, %d, train_acc, %f, valid_acc, %f, train_loss, %f, valid_loss, %f, lr, %e, best_epoch, %d, best_valid_acc, %f',
epoch, train_acc, valid_acc, train_obj, valid_obj,
learning_rate, best_epoch, best_valid_acc)
stats = {
'epoch': epoch,
'train_acc': train_acc,
'valid_acc': valid_acc,
'train_loss': train_obj,
'valid_loss': valid_obj,
'lr': learning_rate,
'best_epoch': best_epoch,
'best_valid_acc': best_valid_acc,
'genotype': str(genotype),
'arch_weights': str(cnn_model.arch_weights)
}
epoch_stats += [copy.deepcopy(stats)]
with open(args.epoch_stats_file, 'w') as f:
json.dump(epoch_stats, f, cls=utils.NumpyEncoder)
utils.list_of_dicts_to_csv(stats_csv, epoch_stats)
# print the final model
if args.final_path is None:
genotype = cnn_model.genotype()
logger.info('genotype = %s', genotype)
logger.info('Search for Model Complete! Save dir: ' + str(args.save))
def fetch_entire_dataloader(
source,
data_dir,
val_scale,
seed,
batch_size,
num_workers,
gaze_task=None, #either none, data augment, cam reg, cam reg convex
ood_set=None,
ood_shift=None,
subclass=False,
gan=True,
label_class=None):
transforms = get_data_transforms("cxr",
normalization_type="train_images",
gan=gan)
datasets = []
for split in ["train", "val", "test"]:
if ood_set is not None:
if split == "test":
source = f"{ood_set}/{source}/{ood_shift}"
dataset = RoboGazeDataset(source=source,
data_dir=data_dir,
split_type=split,
gaze_task=gaze_task,
transform=transforms[split],
val_scale=val_scale,
seed=seed,
subclass=subclass,
gan=gan)
datasets.append(dataset)
concat_datasets = torch.utils.data.ConcatDataset(datasets)
if label_class is not None:
full_dataloader = DataLoader(
dataset=concat_datasets,
shuffle=False,
batch_size=1,
num_workers=num_workers,
)
class_indices = []
for idx, (img, label) in enumerate(full_dataloader):
if label.item() == label_class:
class_indices.append(idx)
class_dataset = torch.utils.data.Subset(concat_datasets, class_indices)
return DataLoader(
dataset=class_dataset,
shuffle=True,
batch_size=batch_size,
num_workers=num_workers,
)
else:
return DataLoader(
dataset=concat_datasets,
shuffle=True,
batch_size=batch_size,
num_workers=num_workers,
)
def fetch_dataloaders(
source,
data_dir,
val_scale,
seed,
batch_size,
num_workers,
gaze_task=None, #either none, data augment, cam reg, cam reg convex actdiff
ood_set=None,
ood_shift=None,
subclass=False,
gan_positive=None,
gan_negative=None,
gan_type=None,
args=None):
transforms = get_data_transforms("cxr", normalization_type="train_images")
dataloaders = {}
for split in ["train", "val", "test"]:
if ood_set is not None:
if split == "test":
source = f"{args.machine}/{ood_set}/{source}/{ood_shift}"
#source = f"{ood_set}/{source}/{ood_shift}"
if split == 'train':
if gan_positive is not None:
original_dataset = RoboGazeDataset(source=source,
data_dir=data_dir,
split_type=split,
gaze_task=gaze_task,
transform=transforms[split],
val_scale=val_scale,
seed=seed,
subclass=subclass,
args=args)
## get positive and negative class amounts
original_dataloader = DataLoader(
dataset=original_dataset,
shuffle=False,
batch_size=1,
num_workers=num_workers,
)
class_amounts = [0, 0]
for img, label, _ in original_dataloader:
class_amounts[label.item()] += 1
if gan_type == "gan":
pos_generator = gan_generator.Generator_Advanced_224(
).cuda()
neg_generator = gan_generator.Generator_Advanced_224(
).cuda()
noise_size = 100
pos_generator.load_state_dict(
torch.load(gan_positive + '/generator_best_ckpt.pt'))
neg_generator.load_state_dict(
torch.load(gan_negative + '/generator_best_ckpt.pt'))
neg_noise = torch.randn(class_amounts[0], noise_size, 1,
1).cuda()
pos_noise = torch.randn(class_amounts[1], noise_size, 1,
1).cuda()
# Feed noise into the generator to create new images
neg_images = []
for i in range(class_amounts[0]):
neg_images.append(
neg_generator(
neg_noise[i].unsqueeze(dim=0)).detach().cpu())
neg_images = torch.cat(neg_images)
pos_images = []
for i in range(class_amounts[1]):
pos_images.append(
pos_generator(
pos_noise[i].unsqueeze(dim=0)).detach().cpu())
pos_images = torch.cat(pos_images)
#neg_images = neg_generator(neg_noise).detach()
#pos_images = pos_generator(pos_noise).detach()
neg_labels = torch.zeros(neg_images.shape[0]).cpu().numpy()
pos_labels = torch.ones(pos_images.shape[0]).cpu().numpy()
neg_gaze_attr = torch.zeros(
neg_images.shape[0]).cpu().numpy()
pos_gaze_attr = torch.zeros(
neg_images.shape[0]).cpu().numpy()
positive_fake_data = GanDataset(images=pos_images,
labels=pos_labels,
gaze_attr=pos_gaze_attr)
negative_fake_data = GanDataset(images=neg_images,
labels=neg_labels,
gaze_attr=neg_gaze_attr)
elif gan_type == "acgan":
noise_size = 110
pos_generator = acgan_generator.Generator_Advanced_224(
1, noise_size).cuda()
neg_generator = acgan_generator.Generator_Advanced_224(
1, noise_size).cuda()
pos_generator.load_state_dict(
torch.load(gan_positive + '/generator_best_ckpt.pt'))
neg_generator.load_state_dict(
torch.load(gan_negative + '/generator_best_ckpt.pt'))
neg_noise = torch.randn(class_amounts[0], noise_size, 1,
1).cuda()
pos_noise = torch.randn(class_amounts[1], noise_size, 1,
1).cuda()
# Feed noise into the generator to create new images
neg_images = []
for i in range(class_amounts[0]):
neg_images.append(
neg_generator(
neg_noise[i].unsqueeze(dim=0)).detach().cpu())
neg_images = torch.cat(neg_images)
pos_images = []
for i in range(class_amounts[1]):
pos_images.append(
pos_generator(
pos_noise[i].unsqueeze(dim=0)).detach().cpu())
pos_images = torch.cat(pos_images)
#neg_images = neg_generator(neg_noise).detach()
#pos_images = pos_generator(pos_noise).detach()
neg_labels = torch.zeros(neg_images.shape[0]).cpu().numpy()
pos_labels = torch.ones(pos_images.shape[0]).cpu().numpy()
neg_gaze_attr = torch.zeros(
neg_images.shape[0]).cpu().numpy()
pos_gaze_attr = torch.zeros(
neg_images.shape[0]).cpu().numpy()
positive_fake_data = GanDataset(images=pos_images,
labels=pos_labels,
gaze_attr=pos_gaze_attr)
negative_fake_data = GanDataset(images=neg_images,
labels=neg_labels,
gaze_attr=neg_gaze_attr)
dataset = torch.utils.data.ConcatDataset(
[original_dataset, positive_fake_data, negative_fake_data])
else:
if gaze_task == "actdiff" and source == "synth":
dataset = SyntheticDataset(split=split, blur=0.817838)
else:
dataset = RoboGazeDataset(source=source,
data_dir=data_dir,
split_type=split,
gaze_task=gaze_task,
transform=transforms[split],
val_scale=val_scale,
seed=seed,
subclass=subclass,
args=args)
else:
if gaze_task == "actdiff" and source == "synth":
dataset = SyntheticDataset(split=split, blur=0.817838)
else:
dataset = RoboGazeDataset(source=source,
data_dir=data_dir,
split_type=split,
gaze_task=gaze_task,
transform=transforms[split],
val_scale=val_scale,
seed=seed,
subclass=subclass,
args=args)
dataloaders[split] = (DataLoader(
dataset=dataset,
shuffle=split == "train",
batch_size=batch_size,
num_workers=num_workers,
))
return dataloaders
type=int,
help='Number of images to use for evaluation')
parser.add_argument('--using-pretrained',
default=False,
action='store_true',
help='Train from scratch or just the final layer?')
args = parser.parse_args()
# CPU or GPU device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("Using device {}".format(device))
# Args for loading data
data_dir = args.data_dir
batch_size = args.batch_size
data_transforms = get_data_transforms()
# Load in and transform data
image_datasets = {
x: datasets.ImageFolder(os.path.join(data_dir, x), data_transforms[x])
for x in ['train', 'val', 'test']
}
# Maps class_idx to its string label
classes_to_idx = {
x: {
class_idx: class_name
for class_name, class_idx in
image_datasets[x].class_to_idx.items()
}
for x in ['train', 'val', 'test']