def __init__(self, loader):
'''
@param loader: Loader instance, used for image retrieval
and further parameters
'''
self.loader = loader
self.log = Logger(verbose=True)
self.image_array_list = self.loader.image_array_list
self.image_list = self.loader.image_list
self.image_path_list = self.loader.image_path_list
self.sp_thread = np.zeros((len(self.image_list), 256, 256))
self.sp_thread_lock = threading.Lock()
self.max_thread = threading.Semaphore(1)
self.log_lock = threading.Lock()
self.seq_reading = self.loader.seq_reading
self.sp_array = np.array([])
self.sp_labeled = np.array([])
self.sp_labels = []
def main():
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
pin_memory = True if use_gpu else False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, 'log_train.txt'))
else:
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("========\nArgs:{}\n========".format(args))
if use_gpu:
print("Currently usimg GPU {}".format(args.gpu_devices))
os.environ['CUDA_VISIBLE_DEVICE'] = args.gpu_devices
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed) # 固定随机种子
else:
print("Currently using CPU")
dataset = data_manager.init_img_dataset(
root=args.root,
name=args.dataset,
spilt_id=args.spilt_id,
cuhk03_labeled=args.cuhk03_labeled,
cuhk03_classic_spilt=args.cuhk03_classic_spilt,
)
# dataloader & augmentation train query gallery
transform_train = T.compose([
T.Random2DTranslation(args.height, args.width),
T.RandomHorizontalFlip(),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--mode', type=str, default='train', help='train or eval')
parser.add_argument('--lr', type=float, default=0.001, help='learning rate')
parser.add_argument('--epochs', type=int, default=10, help='number of training epochs')
parser.add_argument('--batch_size', type=int, default=64, help='number of examples to process in a batch')
parser.add_argument('--num_classes', type=int, default=6, help='number of target classes')
parser.add_argument('--max_norm', type=float, default=5.0, help='max norm of gradient')
parser.add_argument('--embed_trainable', type=bool, default=True, help='finetune pre-trained embeddings')
parser.add_argument('--kernel_sizes', nargs='+', type=int, default=[2, 3, 4], help='kernel sizes for the convolution layer')
parser.add_argument('--device', type=str, default=torch.device('cuda:0' if torch.cuda.is_available() else 'cpu'))
parser.add_argument('--p', type=float, default=0.5, help='dropout rate')
parser.add_argument('--c_out', type=int, default=32, help='output channel size of the convolution layer')
args = parser.parse_args()
if args.mode == 'train':
sys.stdout = Logger(TRAIN_LOG_LOC)
print_statement('HYPERPARAMETER SETTING')
print_flags(args)
train(args, MODEL_LOC)
else:
sys.stdout = Logger(TEST_LOG_LOC)
print_statement('HYPERPARAMETER SETTING')
print_flags(args)
test(args, MODEL_LOC, LABEL_JSON_LOC)
def preprocess_pipeline(args):
# [STEP 0] load the .mat files (sample-level)
if not os.path.exists(args.path_raw):
sys.stdout = Logger(os.path.join(args.path_raw, "log_raw.txt"))
print(paint("[STEP 0] Loading the .mat files..."))
load_mat(path_data=args.path_data,
path_raw=args.path_raw,
class_map=args.class_map)
else:
print(paint("[STEP 0] Files already loaded!"))
# [STEP 1] partition the datasets (segment-level)
w, s = args.window, args.stride
if not os.path.exists(args.path_processed):
sys.stdout = Logger(
os.path.join(args.path_processed, f"log_{w}_{s}.txt"))
print(
paint(
f"[STEP 1] Partitioning the dataset (window,stride) = ({w},{s})..."
))
partition(
path_raw=args.path_raw,
path_processed=args.path_processed,
window=w,
stride=s,
class_map=args.class_map,
)
else:
print(
paint(
f"[STEP 1] Dataset already partitioned (window,stride) = ({w},{s})!"
))
def start_test(experiment_path, epoch, num_plot_samples, timesteps, plot_window_size, device):
assert(type(experiment_path) == str)
# Load the training and environment arguments, and environment
args, env_args = load_args(experiment_path)
env = env_dict[env_args['env_name']](device=device)
# Set the parameters from args
act_func_disc = act_funcs[args['act_func_disc']]
act_func_gen = act_funcs[args['act_func_gen']]
# Parameters for the test
args['device'] = device
args['do_logging'] = True
args['experiment_name'] = 'TEST_' + str(env.name)
args['show_plots'] = True
args['original_experiment_path'] = experiment_path # save the experiment path which is being tested
# Load the logger and the plotter
the_logger = Logger(args)
the_plotter = TestPlotter(args, the_logger,
num_plot_samples=num_plot_samples, linspace_size=201, timesteps=timesteps)
# Set plotter settings
if plot_window_size is 'args':
the_plotter.plot_window_size = env.plot_window_size
else:
args['plot_window'] = plot_window_size
# Get the neural network models and load them
discriminator = DiscNet(dim=env.dim, ns=args['ns'], act_func=act_func_disc, hh=args['hh'], device=device,
psi_func=env.psi_func, TT=env.TT)
generator = GenNet(dim=env.dim, ns=args['ns'], act_func=act_func_gen, hh=args['hh'], device=device, mu=None,
std=None, TT=env.TT)
disc_load_path = join(experiment_path, 'models', 'discriminator-epoch-' + str(epoch) + '.pth.tar')
gen_load_path = join(experiment_path, 'models', 'generator-epoch-' + str(epoch) + '.pth.tar')
disc_load = torch.load(disc_load_path, map_location=lambda storage, loc: storage) # Puts model on CPU
gen_load = torch.load(gen_load_path, map_location=lambda storage, loc: storage)
discriminator.load_state_dict(disc_load['model_state_dict'])
generator.load_state_dict(gen_load['model_state_dict'])
generator.mu, generator.std = gen_load['gen_mu'], gen_load['gen_std']
discriminator.eval()
generator.eval()
# The iteration to generate plots
for idx in range(0, 100):
the_plotter.make_plots(idx, generator, the_logger)
the_plotter.make_animation(generator, 60)
time.sleep(3)
the_plotter.make_new_rho0()
def main():
parser = get_parser_ens()
args = parser.parse_args()
args.method = os.path.basename(__file__).split('-')[1][:-3]
if args.aug_test:
args.method = args.method + '_augment'
torch.backends.cudnn.benchmark = True
compute = {
'CIFAR10':
['VGG16BN', 'PreResNet110', 'PreResNet164', 'WideResNet28x10'],
'CIFAR100':
['VGG16BN', 'PreResNet110', 'PreResNet164', 'WideResNet28x10'],
'ImageNet': ['ResNet50']
}
for model in compute[args.dataset]:
args.model = model
logger = Logger(base='./logs/')
print('-' * 5, 'Computing results of', model, 'on', args.dataset + '.',
'-' * 5)
loaders, num_classes = get_data(args)
targets = get_targets(loaders['test'], args)
args.num_classes = num_classes
model = get_model(args)
for run in range(1, 6):
log_probs = []
fnames = read_models(args,
base=os.path.expanduser(args.models_dir),
run=run if args.dataset != 'ImageNet' else -1)
fnames = sorted(fnames,
key=lambda a: int(a.split('-')[-1].split('.')[0]))
for ns in range(100)[:min(
len(fnames), 100 if args.dataset != 'ImageNet' else 50)]:
start = time.time()
model.load_state_dict(get_sd(fnames[ns], args))
ones_log_prob = one_sample_pred(loaders['test'], model)
log_probs.append(ones_log_prob)
logger.add_metrics_ts(ns,
log_probs,
targets,
args,
time_=start)
logger.save(args)
os.makedirs('.megacache', exist_ok=True)
logits_pth = '.megacache/logits_%s-%s-%s-%s-%s'
logits_pth = logits_pth % (args.dataset, args.model, args.method,
ns + 1, run)
log_prob = logsumexp(np.dstack(log_probs), axis=2) - np.log(ns + 1)
print('Save final logprobs to %s' % logits_pth, end='\n\n')
np.save(logits_pth, log_prob)
def main():
parser = get_parser_ens()
args = parser.parse_args()
args.method = os.path.basename(__file__).split('-')[1][:-3]
torch.backends.cudnn.benchmark = True
if args.aug_test:
args.method = args.method + '_augment'
print('Computing for all datasets!')
compute = {
'CIFAR10': ['VGG16BN', 'WideResNet28x10do'],
'CIFAR100': ['VGG16BN', 'WideResNet28x10do']
}
for model in compute[args.dataset]:
args.model = model
logger = Logger()
print('-' * 5, 'Computing results of', model, 'on', args.dataset + '.',
'-' * 5)
loaders, num_classes = get_data(args)
targets = get_targets(loaders['test'], args)
fnames = read_models(args, base=os.path.expanduser(args.models_dir))
args.num_classes = num_classes
model = get_model(args)
for try_ in range(1, 6):
fnames = np.random.permutation(fnames)
model.load_state_dict(get_sd(fnames[0], args))
log_probs = []
for ns in range(100):
start = time.time()
ones_log_prob = one_sample_pred(loaders['test'], model)
log_probs.append(ones_log_prob)
logger.add_metrics_ts(ns,
log_probs,
targets,
args,
time_=start)
logger.save(args)
os.makedirs('./.megacache', exist_ok=True)
logits_pth = '.megacache/logits_%s-%s-%s-%s-%s'
logits_pth = logits_pth % (args.dataset, args.model, args.method,
ns + 1, try_)
log_prob = logsumexp(np.dstack(log_probs), axis=2) - np.log(ns + 1)
print('Save final logprobs to %s' % logits_pth)
np.save(logits_pth, log_prob)
print('Used weights from %s' % fnames[0], end='\n\n')
def __init__(self, access_key=None, secret_key=None):
self.logging = Logger(
self.__class__.__name__
).get_logger()
self.logging.debug(
"Initiate class for opswork environments: %s" % (self.__class__.__name__)
)
if settings.ACCESS_KEY is None or settings.SECRET_KEY is None:
self.access_key = access_key
self.secret_key = secret_key
else:
self.access_key = settings.ACCESS_KEY
self.secret_key = settings.SECRET_KEY
if self.access_key is None or self.secret_key is None:
raise ExpectedAWSKeys(
"Please, provide a secret key and acces key aws, see: http://docs.aws.amazon.com/AWSSimpleQueueService/latest/SQSGettingStartedGuide/AWSCredentials.html"
)
def main():
# get experiment arguments
args, config_dataset, config_model = get_args()
# [STEP 0 and 1] load the .mat files (sample-level) and partition the datasets (segment-level)
preprocess_pipeline(args)
if args.train_mode:
# [STEP 2] create HAR datasets
dataset = SensorDataset(**config_dataset, prefix="train")
dataset_val = SensorDataset(**config_dataset, prefix="val")
# [STEP 3] create HAR models
if torch.cuda.is_available():
model = create(args.model, config_model).cuda()
torch.backends.cudnn.benchmark = True
sys.stdout = Logger(
os.path.join(model.path_logs,
f"log_main_{args.experiment}.txt"))
# show args
print("##" * 50)
print(paint(f"Experiment: {model.experiment}", "blue"))
print(
paint(
f"[-] Using {torch.cuda.device_count()} GPU: {torch.cuda.is_available()}"
))
print(args)
get_info_params(model)
get_info_layers(model)
print("##" * 50)
# [STEP 4] train HAR models
model_train(model, dataset, dataset_val, args)
# [STEP 5] evaluate HAR models
dataset_test = SensorDataset(**config_dataset, prefix="test")
if not args.train_mode:
config_model["experiment"] = "inference"
model = create(args.model, config_model).cuda()
model_eval(model, dataset_test, args)
def main():
parser = get_parser_ens()
args = parser.parse_args()
args.method = os.path.basename(__file__).split('-')[1][:-3]
torch.backends.cudnn.benchmark = True
if args.aug_test:
args.method = args.method + '_augment'
os.makedirs('./logs', exist_ok=True)
compute = {
'CIFAR10': ['BayesVGG16BN', 'BayesPreResNet110', 'BayesPreResNet164', 'BayesWideResNet28x10'],
'CIFAR100': ['BayesVGG16BN', 'BayesPreResNet110', 'BayesPreResNet164', 'BayesWideResNet28x10'],
'ImageNet': ['BayesResNet50']
}
for model in compute[args.dataset]:
args.model = model
logger = Logger()
print('-'*5, 'Computing results of', model, 'on', args.dataset + '.', '-'*5)
loaders, num_classes = get_data(args)
targets = get_targets(loaders['test'], args)
fnames = read_models(args, base=os.path.expanduser(args.models_dir))
args.num_classes = num_classes
model = get_model(args)
for run in range(1, 6):
print('Repeat num. %s' % run)
log_probs = []
checkpoint = get_sd(fnames[0], args)
model.load_state_dict(checkpoint)
for ns in range(100 if args.dataset != 'ImageNet' else 50):
start = time.time()
ones_log_prob = one_sample_pred(loaders['test'], model)
log_probs.append(ones_log_prob)
logger.add_metrics_ts(ns, log_probs, targets, args, time_=start)
logger.save(args)
os.makedirs('.megacache', exist_ok=True)
logits_pth = '.megacache/logits_%s-%s-%s-%s-%s'
logits_pth = logits_pth % (args.dataset, args.model, args.method, ns+1, run)
log_prob = logsumexp(np.dstack(log_probs), axis=2) - np.log(ns+1)
np.save(logits_pth, log_prob)
def main():
sys.stdout = Logger(osp.join(args.save_dir, 'log_test.txt'))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU")
print("Initialize dataset {}".format(args.dataset))
dataset = h5py.File(args.dataset, 'r')
num_videos = len(dataset.keys())
test_keys = []
for key in dataset.keys():
test_keys.append(key)
print("Load model")
model = DSN(in_dim=args.input_dim,
hid_dim=args.hidden_dim,
num_layers=args.num_layers,
cell=args.rnn_cell)
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
if args.model:
print("Loading checkpoint from '{}'".format(args.model))
checkpoint = torch.load(args.model)
model.load_state_dict(checkpoint)
else:
start_epoch = 0
if use_gpu:
model = nn.DataParallel(model).cuda()
evaluate(model, dataset, test_keys, use_gpu)
print("Summary")
video2summary(os.path.join(args.save_dir, 'result.h5'), args.input,
args.save_dir) ####
target_transform = ClassLabel()
training_data = get_training_set(opt, spatial_transform, target_transform)
train_loader = torch.utils.data.DataLoader(
training_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.n_threads,
pin_memory=True)
if opt.learning_policy == '2stream':
train_logger = Logger(
os.path.join(opt.result_path, 'train.log'),
['epoch', 'loss', 'loss_cls', 'loss_box', 'OBOA', 'MAE', 'MAEP', 'MAEN', 'lr'])
train_batch_logger = Logger(
os.path.join(opt.result_path, 'train_batch.log'),
['epoch', 'batch', 'iter', 'loss', 'loss_cls', 'loss_box', 'OBOA', 'MAE', 'MAEP', 'MAEN', 'lr'])
from train_2stream import train_epoch
if opt.nesterov:
dampening = 0
else:
dampening = opt.dampening
finetune_parameters = []
d_optimizer = optim.Adam(discriminator.parameters(), lr=float(opt.lr_d), betas=(0.5, 0.999))
g_optimizer = optim.Adam(generator.parameters(), lr=float(opt.lr_g), betas=(0.5, 0.999))
loss = nn.BCELoss()
# init fixed noise
fixed_noise = torch.randn(1, nz, 1, 1, device=gpu)
# Additive noise to stabilize Training for DCGAN
initial_additive_noise_var = 0.1
add_noise_var = 0.1
# Create Logger instance
logger = Logger(model_name='LRPGAN', data_name=opt.dataset, dir_name=outf, make_fresh=True if not opt.cont else False)
print('Created Logger')
# training
for epoch in range(opt.epochs):
for n_batch, (batch_data, _) in enumerate(dataloader, 0):
batch_size = batch_data.size(0)
add_noise_var = adjust_variance(add_noise_var, initial_additive_noise_var, opt.epochs * len(dataloader) * 1 / 4)
############################
# Train Discriminator
###########################
# train with real
discriminator.zero_grad()
real_data = batch_data.to(gpu)
def main():
global args
np.random.seed(args.seed)
torch.cuda.manual_seed(args.seed)
if args.fixed_arc:
sys.stdout = Logger(filename='logs/' + args.output_filename + '_fixed.log')
else:
sys.stdout = Logger(filename='logs/' + args.output_filename + '.log')
print(args)
data_loaders = load_datasets()
controller = Controller(search_for=args.search_for,
search_whole_channels=True,
num_layers=args.child_num_layers,
num_branches=args.child_num_branches,
out_filters=args.child_out_filters,
lstm_size=args.controller_lstm_size,
lstm_num_layers=args.controller_lstm_num_layers,
tanh_constant=args.controller_tanh_constant,
temperature=None,
skip_target=args.controller_skip_target,
skip_weight=args.controller_skip_weight)
controller = controller.cuda()
shared_cnn = SharedCNN(num_layers=args.child_num_layers,
num_branches=args.child_num_branches,
out_filters=args.child_out_filters,
keep_prob=args.child_keep_prob)
shared_cnn = shared_cnn.cuda()
# https://github.com/melodyguan/enas/blob/master/src/utils.py#L218
controller_optimizer = torch.optim.Adam(params=controller.parameters(),
lr=args.controller_lr,
betas=(0.0, 0.999),
eps=1e-3)
# https://github.com/melodyguan/enas/blob/master/src/utils.py#L213
shared_cnn_optimizer = torch.optim.SGD(params=shared_cnn.parameters(),
lr=args.child_lr_max,
momentum=0.9,
nesterov=True,
weight_decay=args.child_l2_reg)
# https://github.com/melodyguan/enas/blob/master/src/utils.py#L154
shared_cnn_scheduler = CosineAnnealingLR(optimizer=shared_cnn_optimizer,
T_max=args.child_lr_T,
eta_min=args.child_lr_min)
if args.resume:
if os.path.isfile(args.resume):
print("Loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
start_epoch = checkpoint['epoch']
# args = checkpoint['args']
shared_cnn.load_state_dict(checkpoint['shared_cnn_state_dict'])
controller.load_state_dict(checkpoint['controller_state_dict'])
shared_cnn_optimizer.load_state_dict(checkpoint['shared_cnn_optimizer'])
controller_optimizer.load_state_dict(checkpoint['controller_optimizer'])
shared_cnn_scheduler.optimizer = shared_cnn_optimizer # Not sure if this actually works
print("Loaded checkpoint '{}' (epoch {})"
.format(args.resume, checkpoint['epoch']))
else:
raise ValueError("No checkpoint found at '{}'".format(args.resume))
else:
start_epoch = 0
if not args.fixed_arc:
train_enas(start_epoch,
controller,
shared_cnn,
data_loaders,
shared_cnn_optimizer,
controller_optimizer,
shared_cnn_scheduler)
else:
assert args.resume != '', 'A pretrained model should be used when training a fixed architecture.'
train_fixed(start_epoch,
controller,
shared_cnn,
data_loaders)
def train(self,
model_dir=constant.train_config['trained_model_dir'],
model_name=constant.predict_config['best_model_name']):
iteration_step = 0
logger = Logger(self.model_name)
start_idx_epoch = 0
for epoch in range(start_idx_epoch, start_idx_epoch+self.num_epochs):
print('Executing Epoch: {}'.format(epoch))
#execute each batch
for sample in iter(self.train_batch):
#extract data and label
data = sample['feature']
label = sample['target']
#clear gradient
self.optimizer.zero_grad()
#forward propagation
batch_output = self.classifier_model.nn_model(data)
#calculate loss
loss = self.error(batch_output, label[:, 0, :])
#claculate gradient and update weight
loss.backward()
self.optimizer.step()
# Find metrics on validation dataset
iteration_step += self.batch_size
eval_metric = EvaluationMetric(self.target_num_classes)
training_loss = eval_metric.calculateLoss(self.valid_batch, self.batch_size, self.classifier_model.nn_model, self.error)
test_loss = eval_metric.calculateLoss(self.valid_batch, self.batch_size, self.classifier_model.nn_model, self.error)
precision_train, recall_train, f1_train = eval_metric.calculateEvaluationMetric(self.train_batch, self.batch_size, self.classifier_model.nn_model)
precision_valid, recall_valid, f1_valid = eval_metric.calculateEvaluationMetric(self.valid_batch, self.batch_size, self.classifier_model.nn_model)
print('Epoch: {}, F1-Score (Training Dataset): {}, F1-Score (Validation Dataset): {}, Training Loss: {}, Validation Loss: {}'
.format(epoch, f1_train, f1_valid, training_loss, test_loss))
print('Precision(Training Dataset): {}, Precision(Validation Dataset): {}, Recall(Training Dataset): {}, Recall(Validation Dataset): {}'
.format(precision_train, precision_valid, recall_train, recall_valid))
#log the metric in graph with tensorboard
logger.log(f1_train, f1_valid, training_loss, test_loss, iteration_step)
#save the model weights
model_filepath = model_dir + os.sep + 'weight_epoch-{}_loss-{}'.format(epoch, training_loss)
torch.save(self.classifier_model.nn_model.state_dict(), model_filepath)
logger.close()
for n_batch, (batch,_) in enumerate(data_loader):
print(batch.shape)
y = batch[0:BATCH_SIZE]
break
#print("ts shape ",test_samples.shape)
test_samples = y.permute(0, 1, 3, 2).numpy()
test_samples = images_to_vectors(test_samples)
test_samples = np.reshape(test_samples,(-1,28,28))
print(test_samples.shape)
fi,batch_mask = noisy_images(test_samples)
test_samples_nhwc = vectors_to_images(test_samples)
test_samples = images_to_vectors(test_samples)
logger = Logger(model_name='test', data_name='generator_output')
logger1 = Logger(model_name='test', data_name='true_image')
logger2 = Logger(model_name='test', data_name='noisy_image')
logger1.log_images(test_samples_nhwc,1, 1, 1, 1, '1',format='NHWC');
test_noise_reshape = images_to_vectors(fi)
test_noise_reshape = vectors_to_images(test_noise_reshape)
logger2.log_images(test_noise_reshape,1, 1, 1, 1,'1' ,format='NHWC');
#batch_mask = np.ones((BATCH_SIZE,28,28),np.float32)
#batch_mask[:,9:19,9:19] = 0.0
batch_mask = images_to_vectors(batch_mask)
images = tf.placeholder(tf.float32,shape = (BATCH_SIZE,28*28))
recovered_images = tf.reshape(G_sample,(BATCH_SIZE,28*28))
contextual_loss = tf.reduce_sum(tf.contrib.layers.flatten(tf.abs(tf.multiply(images,batch_mask) - tf.multiply(G_sample, batch_mask))), 1)
betas=(0.5, 0.999))
g_optimizer = optim.Adam(generator.parameters(), lr=0.0008, betas=(0.5, 0.999))
dloss = nn.CrossEntropyLoss()
gloss = nn.BCELoss()
# init fixed noise
fixed_noise = torch.randn(1, nz, 1, 1, device=gpu)
# Additive noise to stabilize Training for DCGAN
initial_additive_noise_var = 0.1
add_noise_var = 0.1
# Create Logger instance
logger = Logger(model_name='LRPGAN', data_name=opt.dataset, dir_name=outf)
print('Created Logger')
# training
for epoch in range(opt.epochs):
for n_batch, (batch_data, _) in enumerate(dataloader, 0):
batch_size = batch_data.size(0)
add_noise_var = adjust_variance(add_noise_var,
initial_additive_noise_var,
opt.epochs * len(dataloader) * 1 / 2)
############################
# Train Discriminator
###########################
# train with real
discriminator.zero_grad()
class KeysAWS(object):
_ec2_connection = None
_opsworks_conection = None
_iam_connection = None
_vpc_connection = None
def __init__(self, access_key=None, secret_key=None):
self.logging = Logger(
self.__class__.__name__
).get_logger()
self.logging.debug(
"Initiate class for opswork environments: %s" % (self.__class__.__name__)
)
if settings.ACCESS_KEY is None or settings.SECRET_KEY is None:
self.access_key = access_key
self.secret_key = secret_key
else:
self.access_key = settings.ACCESS_KEY
self.secret_key = settings.SECRET_KEY
if self.access_key is None or self.secret_key is None:
raise ExpectedAWSKeys(
"Please, provide a secret key and acces key aws, see: http://docs.aws.amazon.com/AWSSimpleQueueService/latest/SQSGettingStartedGuide/AWSCredentials.html"
)
def __name__(self):
return "{}-{}".format(self.access_key, self.secret_key)
@property
def _vpc_connection(self):
_vpc_connection = vpc.connect_to_region(
aws_access_key_id=self.access_key,
aws_secret_access_key=self.secret_key,
region_name=settings.REGION
)
self.logging.debug(
"The connection with vpc was been succesfully"
)
return _vpc_connection
@property
def _ec2_connection(self):
_ec2_connection = ec2.connect_to_region(
aws_access_key_id=self.access_key,
aws_secret_access_key=self.secret_key,
region_name=settings.REGION
)
self.logging.debug(
"The connection with ec2 was been succesfully"
)
return _ec2_connection
@property
def _iam_connection(self):
_iam_connection = iam.connection.IAMConnection(
aws_access_key_id=self.access_key,
aws_secret_access_key=self.secret_key
)
self.logging.debug(
"The connection with iam was been succesfully"
)
return _iam_connection
@property
def _opsworks_conection(self):
_opsworks_conection = opsworks.connect_to_region(
aws_access_key_id=self.access_key,
aws_secret_access_key=self.secret_key,
region_name='us-east-1'
)
self.logging.debug(
"The connection with opsworks was been succesfully"
)
return _opsworks_conection
def __del__(self):
self._ec2_connection.close()
self._opsworks_conection.close()
self._iam_connection.close()
self._vpc_connection.close()
def train():
mkdirs(config.checkpoint_path, config.best_model_path, config.logs)
# load data
src1_train_dataloader_fake, src1_train_dataloader_real, \
src2_train_dataloader_fake, src2_train_dataloader_real, \
src3_train_dataloader_fake, src3_train_dataloader_real, \
tgt_valid_dataloader = get_dataset(config.src1_data, config.src1_train_num_frames,
config.src2_data, config.src2_train_num_frames,
config.src3_data, config.src3_train_num_frames,
config.tgt_data, config.tgt_test_num_frames, config.batch_size)
best_model_ACC = 0.0
best_model_HTER = 1.0
best_model_ACER = 1.0
best_model_AUC = 0.0
# 0:loss, 1:top-1, 2:EER, 3:HTER, 4:ACER, 5:AUC, 6:threshold
valid_args = [np.inf, 0, 0, 0, 0, 0, 0, 0]
loss_classifier = AverageMeter()
classifer_top1 = AverageMeter()
net = DG_model(config.model).to(device)
ad_net_real = Discriminator().to(device)
ad_net_fake = Discriminator().to(device)
log = Logger()
log.open(config.logs + config.tgt_data + '_log_SSDG.txt', mode='a')
log.write(
"\n----------------------------------------------- [START %s] %s\n\n" %
(datetime.now().strftime('%Y-%m-%d %H:%M:%S'), '-' * 51))
print("Norm_flag: ", config.norm_flag)
log.write('** start training target model! **\n')
log.write(
'--------|------------- VALID -------------|--- classifier ---|------ Current Best ------|--------------|\n'
)
log.write(
' iter | loss top-1 HTER AUC | loss top-1 | top-1 HTER AUC | time |\n'
)
log.write(
'-------------------------------------------------------------------------------------------------------|\n'
)
start = timer()
criterion = {
'softmax': nn.CrossEntropyLoss().cuda(),
'triplet': HardTripletLoss(margin=0.1, hardest=False).cuda()
}
optimizer_dict = [
{
"params": filter(lambda p: p.requires_grad, net.parameters()),
"lr": config.init_lr
},
{
"params": filter(lambda p: p.requires_grad,
ad_net_real.parameters()),
"lr": config.init_lr
},
]
optimizer = optim.SGD(optimizer_dict,
lr=config.init_lr,
momentum=config.momentum,
weight_decay=config.weight_decay)
init_param_lr = []
for param_group in optimizer.param_groups:
init_param_lr.append(param_group["lr"])
iter_per_epoch = 10
src1_train_iter_real = iter(src1_train_dataloader_real)
src1_iter_per_epoch_real = len(src1_train_iter_real)
src2_train_iter_real = iter(src2_train_dataloader_real)
src2_iter_per_epoch_real = len(src2_train_iter_real)
src3_train_iter_real = iter(src3_train_dataloader_real)
src3_iter_per_epoch_real = len(src3_train_iter_real)
src1_train_iter_fake = iter(src1_train_dataloader_fake)
src1_iter_per_epoch_fake = len(src1_train_iter_fake)
src2_train_iter_fake = iter(src2_train_dataloader_fake)
src2_iter_per_epoch_fake = len(src2_train_iter_fake)
src3_train_iter_fake = iter(src3_train_dataloader_fake)
src3_iter_per_epoch_fake = len(src3_train_iter_fake)
max_iter = config.max_iter
epoch = 1
if (len(config.gpus) > 1):
net = torch.nn.DataParallel(net).cuda()
for iter_num in range(max_iter + 1):
if (iter_num % src1_iter_per_epoch_real == 0):
src1_train_iter_real = iter(src1_train_dataloader_real)
if (iter_num % src2_iter_per_epoch_real == 0):
src2_train_iter_real = iter(src2_train_dataloader_real)
if (iter_num % src3_iter_per_epoch_real == 0):
src3_train_iter_real = iter(src3_train_dataloader_real)
if (iter_num % src1_iter_per_epoch_fake == 0):
src1_train_iter_fake = iter(src1_train_dataloader_fake)
if (iter_num % src2_iter_per_epoch_fake == 0):
src2_train_iter_fake = iter(src2_train_dataloader_fake)
if (iter_num % src3_iter_per_epoch_fake == 0):
src3_train_iter_fake = iter(src3_train_dataloader_fake)
if (iter_num != 0 and iter_num % iter_per_epoch == 0):
epoch = epoch + 1
param_lr_tmp = []
for param_group in optimizer.param_groups:
param_lr_tmp.append(param_group["lr"])
net.train(True)
ad_net_real.train(True)
optimizer.zero_grad()
adjust_learning_rate(optimizer, epoch, init_param_lr,
config.lr_epoch_1, config.lr_epoch_2)
######### data prepare #########
src1_img_real, src1_label_real = src1_train_iter_real.next()
src1_img_real = src1_img_real.cuda()
src1_label_real = src1_label_real.cuda()
input1_real_shape = src1_img_real.shape[0]
src2_img_real, src2_label_real = src2_train_iter_real.next()
src2_img_real = src2_img_real.cuda()
src2_label_real = src2_label_real.cuda()
input2_real_shape = src2_img_real.shape[0]
src3_img_real, src3_label_real = src3_train_iter_real.next()
src3_img_real = src3_img_real.cuda()
src3_label_real = src3_label_real.cuda()
input3_real_shape = src3_img_real.shape[0]
src1_img_fake, src1_label_fake = src1_train_iter_fake.next()
src1_img_fake = src1_img_fake.cuda()
src1_label_fake = src1_label_fake.cuda()
input1_fake_shape = src1_img_fake.shape[0]
src2_img_fake, src2_label_fake = src2_train_iter_fake.next()
src2_img_fake = src2_img_fake.cuda()
src2_label_fake = src2_label_fake.cuda()
input2_fake_shape = src2_img_fake.shape[0]
src3_img_fake, src3_label_fake = src3_train_iter_fake.next()
src3_img_fake = src3_img_fake.cuda()
src3_label_fake = src3_label_fake.cuda()
input3_fake_shape = src3_img_fake.shape[0]
input_data = torch.cat([
src1_img_real, src1_img_fake, src2_img_real, src2_img_fake,
src3_img_real, src3_img_fake
],
dim=0)
source_label = torch.cat([
src1_label_real, src1_label_fake, src2_label_real, src2_label_fake,
src3_label_real, src3_label_fake
],
dim=0)
######### forward #########
classifier_label_out, feature = net(input_data, config.norm_flag)
######### single side adversarial learning #########
input1_shape = input1_real_shape + input1_fake_shape
input2_shape = input2_real_shape + input2_fake_shape
feature_real_1 = feature.narrow(0, 0, input1_real_shape)
feature_real_2 = feature.narrow(0, input1_shape, input2_real_shape)
feature_real_3 = feature.narrow(0, input1_shape + input2_shape,
input3_real_shape)
feature_real = torch.cat(
[feature_real_1, feature_real_2, feature_real_3], dim=0)
discriminator_out_real = ad_net_real(feature_real)
######### unbalanced triplet loss #########
real_domain_label_1 = torch.LongTensor(input1_real_shape,
1).fill_(0).cuda()
real_domain_label_2 = torch.LongTensor(input2_real_shape,
1).fill_(0).cuda()
real_domain_label_3 = torch.LongTensor(input3_real_shape,
1).fill_(0).cuda()
fake_domain_label_1 = torch.LongTensor(input1_fake_shape,
1).fill_(1).cuda()
fake_domain_label_2 = torch.LongTensor(input2_fake_shape,
1).fill_(2).cuda()
fake_domain_label_3 = torch.LongTensor(input3_fake_shape,
1).fill_(3).cuda()
source_domain_label = torch.cat([
real_domain_label_1, fake_domain_label_1, real_domain_label_2,
fake_domain_label_2, real_domain_label_3, fake_domain_label_3
],
dim=0).view(-1)
triplet = criterion["triplet"](feature, source_domain_label)
######### cross-entropy loss #########
real_shape_list = []
real_shape_list.append(input1_real_shape)
real_shape_list.append(input2_real_shape)
real_shape_list.append(input3_real_shape)
real_adloss = Real_AdLoss(discriminator_out_real, criterion["softmax"],
real_shape_list)
cls_loss = criterion["softmax"](classifier_label_out.narrow(
0, 0, input_data.size(0)), source_label)
######### backward #########
total_loss = cls_loss + config.lambda_triplet * triplet + config.lambda_adreal * real_adloss
total_loss.backward()
optimizer.step()
optimizer.zero_grad()
loss_classifier.update(cls_loss.item())
acc = accuracy(classifier_label_out.narrow(0, 0, input_data.size(0)),
source_label,
topk=(1, ))
classifer_top1.update(acc[0])
print('\r', end='', flush=True)
print(
' %4.1f | %5.3f %6.3f %6.3f %6.3f | %6.3f %6.3f | %6.3f %6.3f %6.3f | %s'
% ((iter_num + 1) / iter_per_epoch, valid_args[0], valid_args[6],
valid_args[3] * 100, valid_args[4] * 100, loss_classifier.avg,
classifer_top1.avg, float(best_model_ACC),
float(best_model_HTER * 100), float(
best_model_AUC * 100), time_to_str(timer() - start, 'min')),
end='',
flush=True)
if (iter_num != 0 and (iter_num + 1) % iter_per_epoch == 0):
# 0:loss, 1:top-1, 2:EER, 3:HTER, 4:AUC, 5:threshold, 6:ACC_threshold
valid_args = eval(tgt_valid_dataloader, net, config.norm_flag)
# judge model according to HTER
is_best = valid_args[3] <= best_model_HTER
best_model_HTER = min(valid_args[3], best_model_HTER)
threshold = valid_args[5]
if (valid_args[3] <= best_model_HTER):
best_model_ACC = valid_args[6]
best_model_AUC = valid_args[4]
save_list = [
epoch, valid_args, best_model_HTER, best_model_ACC,
best_model_ACER, threshold
]
save_checkpoint(save_list, is_best, net, config.gpus,
config.checkpoint_path, config.best_model_path)
print('\r', end='', flush=True)
log.write(
' %4.1f | %5.3f %6.3f %6.3f %6.3f | %6.3f %6.3f | %6.3f %6.3f %6.3f | %s %s'
%
((iter_num + 1) / iter_per_epoch, valid_args[0], valid_args[6],
valid_args[3] * 100, valid_args[4] * 100, loss_classifier.avg,
classifer_top1.avg, float(best_model_ACC),
float(best_model_HTER * 100), float(best_model_AUC * 100),
time_to_str(timer() - start, 'min'), param_lr_tmp[0]))
log.write('\n')
time.sleep(0.01)
Normalize([114.7748, 107.7354, 99.475], [1, 1, 1])
])
temporal_transform = TemporalCenterCrop(opt.sample_duration,
opt.downsample)
#temporal_transform = TemporalBeginCrop(opt.sample_duration)
#temporal_transform = TemporalEndCrop(opt.sample_duration)
target_transform = ClassLabel()
opt.n_val_samples = 1
validation_data = get_validation_set(opt, spatial_transform,
temporal_transform, target_transform)
data_loader = torch.utils.data.DataLoader(validation_data,
batch_size=8,
shuffle=False,
num_workers=1,
pin_memory=True)
val_logger = Logger(os.path.join(opt.result_path, 'val.log'),
['epoch', 'loss', 'acc'])
if opt.resume_path:
print('loading checkpoint {}'.format(opt.resume_path))
checkpoint = torch.load(opt.resume_path)
opt.begin_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
recorder = []
print('run')
model.eval()
batch_time = AverageMeter()
top1 = AverageMeter()
def main():
args = parse_args()
args.pretrain = False
root_path = 'exps/exp_{}'.format(args.exp)
if not os.path.exists(root_path):
os.mkdir(root_path)
os.mkdir(os.path.join(root_path, "log"))
os.mkdir(os.path.join(root_path, "model"))
base_lr = args.lr # base learning rate
train_dataset, val_dataset = build_dataset(args.dataset, args.data_root,
args.train_list)
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=1,
num_workers=args.num_workers,
pin_memory=True)
model = VNet(args.n_channels, args.n_classes).cuda()
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=0.0005)
#scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, 0.7)
model = torch.nn.DataParallel(model)
model.train()
if args.resume is None:
assert os.path.exists(args.load_path)
state_dict = model.state_dict()
print("Loading weights...")
pretrain_state_dict = torch.load(args.load_path,
map_location="cpu")['state_dict']
for k in list(pretrain_state_dict.keys()):
if k not in state_dict:
del pretrain_state_dict[k]
model.load_state_dict(pretrain_state_dict)
print("Loaded weights")
else:
print("Resuming from {}".format(args.resume))
checkpoint = torch.load(args.resume, map_location="cpu")
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
model.load_state_dict(checkpoint['state_dict'])
logger = Logger(root_path)
saver = Saver(root_path)
for epoch in range(args.start_epoch, args.epochs):
train(model, train_loader, optimizer, logger, args, epoch)
validate(model, val_loader, optimizer, logger, saver, args, epoch)
adjust_learning_rate(args, optimizer, epoch)
# Hyper Parameters
BATCH_SIZE = 32
BATCH_SIZE_TEST = 20
LR = 0.001 # learning rate
NUM_WORKERS = 12
NUM_CLASSES = 4
LOG_FILE_NAME = './logs/log_' + time.strftime("%Y-%m-%d_%H:%M:%S", time.localtime()) + '.log'
lr_steps = [30, 60, 90, 120]
np.random.seed(42)
DATA_PATH = '/home/ubuntu5/wxp/datasets/acne4/VOCdevkit2007/VOC2007/JPEGImages_300'
log = Logger()
log.open(LOG_FILE_NAME, mode="a")
def criterion(lesions_num):
if lesions_num <= 5:
return 0
elif lesions_num <= 20:
return 1
elif lesions_num <= 50:
return 2
else:
return 3
def trainval_test(cross_val_index, sigma, lam):
class OpsWorkSetup(object):
def __init__(self, access_key=None, secret_key=None):
self.logging = Logger(
self.__class__.__name__
).get_logger()
self.logging.debug(
"Initiate class for opswork environments: %s" % (self.__class__.__name__)
)
if not settings.ACCESS_KEY or not settings.SECRET_KEY:
self.access_key = access_key
self.secret_key = secret_key
else:
self.access_key = settings.ACCESS_KEY
self.secret_key = settings.SECRET_KEY
if self.access_key is None or self.secret_key is None:
raise ExpectedAWSKeys(
"Please, provide a secret key and acces key aws, see: http://docs.aws.amazon.com/AWSSimpleQueueService/latest/SQSGettingStartedGuide/AWSCredentials.html"
)
@property
def conn(self):
_conn = opsworks.connect_to_region(
aws_access_key_id=self.access_key,
aws_secret_access_key=self.secret_key,
region_name='us-east-1'
)
self.logging.debug(
"The connection with opsworks was been succesfully"
)
return _conn
@property
def security_groups(self):
_security_groups = ec2.connect_to_region(
aws_access_key_id=self.access_key,
aws_secret_access_key=self.secret_key,
region_name=settings.REGION
)
self.logging.debug(
"The connection with ec2 was been succesfully"
)
return _security_groups
@property
def describe_subnets(self):
_describe_subnets = vpc.connect_to_region(
aws_access_key_id=self.access_key,
aws_secret_access_key=self.secret_key,
region_name=settings.REGION
)
self.logging.debug(
"The connection with vpc was been succesfully"
)
return _describe_subnets
def create_stack(self):
""" create stack for modeling environment """
stack_name = 'ElasticSearchStack-{}'.format(str(uuid.uuid4())[:8])
if (not settings.DEFAULT_INSTANCE_PROFILE_ARN or settings.DEFAULT_INSTANCE_PROFILE_ARN is None
or not settings.SERVICE_ROLE_ARN or settings.SERVICE_ROLE_ARN is None):
raise ExpectedAWSRoles("Please, provide the correct services roles, see http://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles_create.html and check README.md about the access required for this roles.")
self.stack = self.conn.create_stack(
name=stack_name,
region=settings.REGION,
default_availability_zone=settings.AVAILABLE_ZONE,
custom_json="{}".format(settings.CUSTOM_JSON_CHEF),
use_custom_cookbooks=True,
hostname_theme='Europe_Cities',
use_opsworks_security_groups=True,
custom_cookbooks_source={"Type": "git", "Url": settings.REPOSITORY_URL},
default_instance_profile_arn=settings.DEFAULT_INSTANCE_PROFILE_ARN,
service_role_arn=settings.SERVICE_ROLE_ARN,
default_ssh_key_name=settings.SSH_KEY_NAME_DEFAULT,
default_os=settings.DEFAULT_OS,
configuration_manager=settings.CONFIGURATION_MANAGER
)
self.logging.debug(
"The stack: {!r} has been created with successfull".format(
stack_name
)
)
return self.stack
def create_security_group(self, network_policies=[]):
""" get security groups and modeling template for security groups:
:network_policies (list)
e.g: [{
'protocol': 'http',
'from_port': '80',
'to_port': '80',
'cidr_ip': '172.0.0.1/16'
}, {
'protocol': 'tcp',
'from_port': '9201',
'to_port': '9201',
'cidr_ip': '172.16.0.1/16'
}]
** Observation **
Don't accepet rules with 0.0.0.0/0"""
security_groups = self.security_groups.get_all_security_groups()
for sg_attributes in security_groups:
if u"AWS-OpsWorks-Default-Server" == sg_attributes.name:
for new_rule in network_policies:
print new_rule
try:
sg_attributes.authorize(
new_rule['protocol'],
new_rule['from_port'],
new_rule['to_port'],
new_rule['cidr_ip']
)
self.logging.info(
"The new rule: {!r}, {!r}, {!r}, {!r} has been created on security group: {!r}".format(
new_rule['protocol'],
new_rule['from_port'],
new_rule['to_port'],
new_rule['cidr_ip'],
sg_attributes.name
)
)
except EC2ResponseError:
self.logging.info(
"Specified Rule already exists...skipped"
)
pass
# If put rule with "0.0.0.0/0" will be deleted.
# I decided put this code here, just to force no to have world rule for anywhere
for rule in sg_attributes.rules:
for grant in rule.grants:
if u"0.0.0.0/0" == grant.cidr_ip:
sg_attributes.revoke(
rule.ip_protocol,
rule.from_port,
rule.to_port,
grant.cidr_ip
)
self.logging.info(
"The rule: {!r}, {!r}, {!r}, {!r} has been deleted on security group: {!r}.".format(
rule.ip_protocol,
rule.from_port,
rule.to_port,
grant.cidr_ip,
sg_attributes.name
)
)
def vpc_data_network(self, protocol='tcp', cidr_ips=[]):
""" This method is just for get and management vpc informcations:
:protocol (string):
:cidr_ips (list):
e.g: [{
'protocol': 'http',
'from_port': '80',
'to_port': '80',
'cidr_ip': '172.0.0.1/16'
}, {
'protocol': 'tcp',
'from_port': '9200',
'to_port': '9200',
'cidr_ip': '172.1.0.1/16'
}]"""
network_policies = []
if not cidr_ips:
# Get default subnets on defauilt VPC (my case)
subnets = self.describe_subnets.get_all_subnets()
for subnet in subnets:
cidr_ips.append(subnet.cidr_block)
network_policies = [{
'protocol': 'tcp',
'from_port': 9300,
'to_port': 9300,
'cidr_ip': [cidr_ips[0], cidr_ips[1]]
}, {
'protocol': 'tcp',
'from_port': 9201,
'to_port': 9201,
'cidr_ip': [cidr_ips[0], cidr_ips[1]]
}, {
'protocol': 'tcp',
'from_port': 80,
'to_port': 80,
'cidr_ip': [cidr_ips[0], cidr_ips[1]]
}, {
'protocol': 'tcp',
'from_port': 443,
'to_port': 443,
'cidr_ip': [cidr_ips[0], cidr_ips[1]]
}]
else:
for cidr_ip in cidr_ips:
network_policies.append({
'protocol': cidr_ip['protocol'],
'from_port': cidr_ip['from_port'],
'to_port': cidr_ip['to_port'],
'cidr_ip': cidr_ip['cidr_ip']
})
if not network_policies:
raise ExpectedSubnetsAndVPC("Well, in this case, it's necessary to create one VPC and two subnets for this region")
self.logging.debug("will be created network policies and adjusted with parameters: {}".format(
network_policies
)
)
self.create_security_group(network_policies=network_policies)
def create_layer(self, new_stack=False, stack_id=None):
""" The method is just for create layer:
:new_stack (booblean):
:stack_id (string):"""
layer_name = 'ElasticSearchLayer-{}'.format(str(uuid.uuid4())[:8])
if new_stack and stack_id is None:
new_stack_id = self.create_stack()['StackId']
if stack_id:
new_stack_id = stack_id
self.stack['StackId'] = new_stack_id
self.layer_created = self.conn.create_layer(
stack_id=self.stack['StackId'],
type='custom',
name=layer_name,
volume_configurations=settings.EBS_VOLUM,
shortname='elasticsearchlayer',
custom_instance_profile_arn=settings.DEFAULT_INSTANCE_PROFILE_ARN,
auto_assign_elastic_ips=True,
custom_recipes=settings.RECIPES
)
self.logging.debug(
"The layer: {!r} has been created with successfull".format(
layer_name
)
)
self.vpc_data_network()
return self.layer_created
def create_instances(self, number_instances=3, subnets_list=[], new_layer=True, new_stack=True, stack_id=None, layer_id=[], cidr_ips=[], **kwargs):
"""The method is just for create instances:
:number_instances (int): Number of the instances you want create
:subnets_list (list): list with the subnets for input your instances, example:
[ 172.0.0.1/16, 172.1.0.1/16 ]
:new_layer (boolean): If you want create a new layer before or input in specific layer, expected LayerId
:new_stack (boolean): if you want create a new stack before or input in specific stack, expected StackId
:layer_id (list): If new_layer is False, i need a list with layer ids, example: [ 'foor', 'bar' ]
:cidr_ips (list): Set the ips list with arbitrary cidr_ips
:**kwargs (dict): dict with another increments for boto.opsworks method
"""
if new_layer and not layer_id:
new_layer_id = [self.create_layer(new_stack=new_stack)['LayerId']]
if layer_id:
new_layer_id = layer_id
if not new_stack:
self.stack = {
'StackId': stack_id
}
if subnets_list:
if len(subnets_list) != number_instances:
raise ParameterProblems("number instances and subnets_list needed the same lenght.")
else:
subnets_list=None
for loop in range(0, number_instances):
if subnets_list:
new_subnets_list = subnets_list[loop]
else:
new_subnets_list = None
instance_created = self.conn.create_instance(
stack_id=self.stack['StackId'],
layer_ids=new_layer_id,
root_device_type='ebs',
instance_type=settings.INSTANCE_TYPE,
subnet_id=new_subnets_list,
**kwargs
)
self.logging.debug(
"The {!r} instance(s) has been created with successfull: stack_id: {!r}, layer_id: {!r}, instance_type: {!r}, subnets: {!r}".format(
number_instances,
self.stack['StackId'],
new_layer_id,
settings.INSTANCE_TYPE,
new_subnets_list
)
)
self.conn.start_instance(instance_created['InstanceId'])
if cidr_ips:
rules=[]
for cidr_ip in cidr_ips:
rules.append({
'protocol': 'tcp',
'from_port': 80,
'to_port': 80,
'cidr_ip': cidr_ip
})
rules.append({
'protocol': 'tcp',
'from_port': 9201,
'to_port': 9201,
'cidr_ip': cidr_ip
})
rules.append({
'protocol': 'tcp',
'from_port': 443,
'to_port': 443,
'cidr_ip': cidr_ip
})
rules.append({
'protocol': 'tcp',
'from_port': 9300,
'to_port': 9300,
'cidr_ip': cidr_ip
})
self.vpc_data_network(cidr_ips=rules)
def managament_instance(self, instance_id, action='stop'):
""" This class, is just for management instances (stop, start etc...)
:instance_id (string):
:action (string) - specific strings expected, are options:
'stop', 'start', 'delete'
"""
status = None
if action == 'stop':
status = self.conn.stop_instance(instance_id)
if action == 'start':
status = self.conn.start_instance(instance_id)
if action == 'delete':
status = self.conn.delete_instance(instance_id=instance_id, delete_elastic_ip=True, delete_volumns=True)
if not status:
raise UnrecognizedComand("Plase, try again with: 'stop', 'start', or 'delete' command.")
return status
learning_rate=0.001,
batch_size=args.batch_size_vae,
epochs=args.epochs_vae)
config_temporal = dict(model_name='LSTM',
dynamic_latent_dim=args.dynamic_latent_dim,
learning_rate=0.001,
batch_size=args.batch_size_lstm,
epochs=args.epochs_lstm)
experiment_dir.mkdir(
parents=True,
exist_ok=True) # Create experiment dir if it doesn't exist yet
experiment_dir.joinpath('logs').mkdir(
parents=True, exist_ok=True) # Create logs dir if it doesn't exist yet
runId = datetime.datetime.now().isoformat()
sys.stdout = Logger(f'{experiment_dir}/logs/run_{runId}.txt'
) # Redirect print to log file as well as the terminal
# __ __ ______
# \ \ / /\ | ____|
# \ \ / / \ | |__
# \ \/ / /\ \ | __|
# \ / ____ \| |____
# \/_/ \_\______|
# SETUP VAE
spatial_model = models.Image32x32_VAE(latent_dim=config_spatial['latent_dim'],
learn_prior=False)
kvae_dataset = ImageSequencesDataset(NPZ_ImageSequencesLoader(
spatial_data_filepath.absolute()).get_as_tensor(),
return_sequences=False)
def main():
import warnings
warnings.filterwarnings("ignore")
args = parse_args()
os.environ['CUDA_VISIBLE_DEVICES']= f'{args.gpu}'
utils.prepare_train_directories(args)
log = Logger()
log.open(args.log_dir + '/' + args.model_name + f'/fold_{args.fold}' + '/findlr_log.txt', mode='a')
log.write('*'*30)
log.write('\n')
log.write('Logging arguments!!\n')
log.write('*'*30)
log.write('\n')
for arg, value in sorted(vars(args).items()):
log.write(f'{arg}: {value}\n')
log.write('*'*30)
log.write('\n')
run(args, log)
print('success!')
import re
from etc import settings
from utils.utils import Logger, APP_ROOT
from lib.exceptions import UnknowCIDRRange
from lib.opswork_setup import OpsWorkSetup
from lib.iam import AWSPolicies
__author__ = "Rondineli G. de Araujo"
__copyright__ = "Copyright (C) 2015 Rondineli G. Araujo"
__version__ = "0.0.1"
logging = Logger("OpsWorks Setup").get_logger()
logging.debug("Lunch opsworks setup with elasticSearch Cluster")
def call(args, parse):
if args.access_key:
settings.ACCESS_KEY = args.access_key
if args.secret_key:
settings.SECRET_KEY = args.secret_key
if args.service_role_arn:
settings.SERVICE_ROLE_ARN = args.service_role_arn
if args.instance_arn_role:
settings.DEFAULT_INSTANCE_PROFILE_ARN = args.instance_arn_role
def main():
args = parse_args()
if args.turnon < 0:
args.pretrain = True
else:
args.pretrain = False
print("Using GPU: {}".format(args.local_rank))
root_path = 'exps/exp_{}'.format(args.exp)
if args.local_rank == 0 and not os.path.exists(root_path):
os.mkdir(root_path)
os.mkdir(os.path.join(root_path, "log"))
os.mkdir(os.path.join(root_path, "model"))
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
train_dataset, val_dataset = build_dataset(args.dataset,
args.data_root,
args.train_list,
sampling=args.sampling)
args.world_size = len(args.gpu.split(","))
if args.world_size > 1:
os.environ['MASTER_PORT'] = args.port
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group('nccl')
device = torch.device('cuda:{}'.format(args.local_rank))
train_sampler = torch.utils.data.distributed.DistributedSampler(
train_dataset,
num_replicas=len(args.gpu.split(",")),
rank=args.local_rank)
else:
train_sampler = None
train_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=(train_sampler is None),
sampler=train_sampler,
num_workers=args.num_workers,
pin_memory=True)
val_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=1,
num_workers=args.num_workers,
pin_memory=True)
model = VNet(args.n_channels, args.n_classes, input_size=64,
pretrain=True).cuda(args.local_rank)
model_ema = VNet(args.n_channels,
args.n_classes,
input_size=64,
pretrain=True).cuda(args.local_rank)
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=0.9,
weight_decay=0.0005)
if args.world_size > 1:
model = DDP(model,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
model_ema = DDP(model_ema,
device_ids=[args.local_rank],
output_device=args.local_rank,
find_unused_parameters=True)
model.train()
model_ema.load_state_dict(model.state_dict())
print("Loaded weights")
logger = Logger(root_path)
saver = Saver(root_path, save_freq=args.save_freq)
if args.sampling == 'default':
contrast = RGBMoCo(128, K=4096,
T=args.temperature).cuda(args.local_rank)
elif args.sampling == 'layerwise':
contrast = RGBMoCoNew(128, K=4096,
T=args.temperature).cuda(args.local_rank)
else:
raise ValueError("unsupported sampling method")
criterion = torch.nn.CrossEntropyLoss()
flag = False
for epoch in range(args.start_epoch, args.epochs):
train_sampler.set_epoch(epoch)
train(model, model_ema, train_loader, optimizer, logger, saver, args,
epoch, contrast, criterion)
validate(model_ema, val_loader, optimizer, logger, saver, args, epoch)
adjust_learning_rate(args, optimizer, epoch)
model_name = 'efficientnet-b1'
writeFile = '/data1/cby/temp/output_my_aug/logs/' + model_name
store_name = '/data1/cby/temp/output_my_aug/weights/' + model_name
if not os.path.isdir(store_name):
os.makedirs(store_name)
model_path = None
# model_path = '/data1/cby/temp/output_my_aug/weights/efficientnet-b1/efn-b1_LS_9_loss_0.1610.pth'
model = get_efficientnet(model_name=model_name)
if model_path is not None:
# model = torch.load(model_path)
model.load_state_dict(torch.load(model_path, map_location='cpu'))
print('Model found in {}'.format(model_path))
else:
print('No model found, initializing random model.')
model = model.cuda(device_id)
train_logger = Logger(model_name=writeFile,
header=['epoch', 'loss', 'acc', 'lr'])
# criterion = nn.CrossEntropyLoss()
criterion = LabelSmoothing(smoothing=0.05).cuda(device_id)
# optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9)
# optimizer = optim.Adam(model.parameters(), lr=lr)
optimizer = optim.AdamW(model.parameters(), lr=lr)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=2, gamma=0.9)
is_train = True
if is_train:
xdl = DeeperForensicsDatasetNew(real_npys=train_real_paths_npy,
fake_npys=train_fake_paths_npy,
is_one_hot=True,
transforms=None)
train_loader = DataLoader(xdl,
return pow((1 - ((epoch) / args['n_epochs'])), 0.9)
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer,
lr_lambda=lambda_,
)
# clustering
cluster = Cluster()
# Visualizer
visualizer = Visualizer(('image', 'pred', 'sigma', 'seed'))
# Logger
logger = Logger(('train', 'val', 'iou'), 'loss')
# resume
start_epoch = 0
best_iou = 0
if args['resume_path'] is not None and os.path.exists(args['resume_path']):
print('Resuming model from {}'.format(args['resume_path']))
state = torch.load(args['resume_path'])
start_epoch = state['epoch'] + 1
best_iou = state['best_iou']
model.load_state_dict(state['model_state_dict'], strict=True)
optimizer.load_state_dict(state['optim_state_dict'])
logger.data = state['logger_data']
def train(epoch):
def main(args=None):
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
# torch.cuda.set_device(0)
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, args.log_train))
else:
sys.stdout = Logger(osp.join(args.save_dir, args.log_test))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset)
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = False
queryloader = DataLoader(
VideoDataset(dataset.query,
data_name=args.dataset,
seq_len=args.seq_len,
sample='dense',
transform=transform_test,
type="test"),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
data_name=args.dataset,
seq_len=args.seq_len,
sample='dense',
transform=transform_test,
type="test"),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing models: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
final_dim=args.feat_dim)
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
checkpoint = torch.load(args.model_path)
model.load_state_dict(checkpoint['state_dict'])
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
atest(model, queryloader, galleryloader, use_gpu)
return
def super_parsing(argv):
'''
@param method: SuperPixel Method to test
Possible Methods are:
"fh", "qs", "slic"
@param params: Params Array for given SuperPixels method
implements the superParsing Pipeline
'''
#handling Arguments
try:
opts, args = getopt.getopt(argv, "h", ["help"])
except:
usage()
sys.exit(2)
for opt, arg in opts:
if opt in ("-h", "--help"):
usage()
sys.exit(1)
if len(args)>0:
if args[0]=='fh': #Graph Based
method = 'fh'
elif args[0]=='qs': #Quick Shift
if len(args)==4:
method = args[0]
else:
usage()
sys.exit(2)
elif args[0]=='slic': #SLIC
if len(args)==3:
method = args[0]
else:
usage()
sys.exit(2)
elif args[0]=='gt': #Ground Truth
if len(args)==1:
method = args[0]
else:
usage()
sys.exit(2)
elif args[0]=='sal': #Saliency
if len(args)==2:
method = args[0]
else:
usage()
sys.exit(2)
elif args[0]=='grid': #GRID
if len(args)==2:
method = args[0]
else:
usage()
sys.exit(2)
else:
method = 'fh'
#Setting up Environment
log = Logger()
sfe = segmentFeatureExtractor()
using_ubuntu = False
#setting paths
#Ubuntu Paths
if using_ubuntu:
home_base = base_ubuntu
home_folder = home_base + sub_experiments
#Debian Paths
else:
home_base = base_gnome
home_folder = home_base + sub_experiments
seg_method = {}
############## Segmentation Methods ################
if method == 'qs':
#################### Quick Shift ###################
seg_method['method'] = "Quick_Shift"
seg_method['ratio'] = float(args[1])#0.05
seg_method['kernelsize'] = int(args[2])#2
seg_method['maxdist'] = int(args[3])#48
if method == 'slic':
#################### SLIC ########################
seg_method['method'] = "SLIC"
seg_method['region_size'] = int(args[1])#60
seg_method['regularizer'] = int(args[2])#100
if method == 'gt':
#################### Ground Truth ##################
seg_method['method'] = "Ground_Truth"
if method == 'sal':
#################### Saliency #######################
seg_method['method'] = "Saliency"
seg_method['k'] = int(args[1])
if method == 'grid':
#################### GRID ###########################
seg_method['method'] = 'GRID'
seg_method['k'] = int(args[1])
elif method == 'fh':
####################### FH ##########################
seg_method['method'] = "SP"
seg_method['desc'] = "Desc"
seg_method['k'] = "k200"
#"SP_Desc_k200"
############################ PATH SETTINGS #################################
seg_folder = seg_method['method']
seg_params = [seg_method[key]
for key in np.sort(np.array(seg_method.keys()))
if key!='method']
for i in range(len(seg_params)):
seg_folder = seg_folder + "_"+str(seg_params[i])
if not os.path.exists(home_folder + seg_folder):
os.makedirs(home_folder + seg_folder)
home_images = home_base + "Images"
home_label_sets = (home_folder + seg_folder+"/GeoLabels",
home_folder + seg_folder+"/SemanticLabels")
label_sets = ("GeoLabels", "SemanticLabels")
home_sp_label_folder = (home_label_sets[0] + "/SP_Desc_k200",
home_label_sets[1] + "/SP_Desc_k200")
# home_sp_label_folder = (home_folder + "SPGeoLabels",
# home_folder + "SPSemanticLabels")
home_data = home_folder + seg_folder + "/Data"
home_descriptors = home_data + "/Descriptors"
home_descriptors_global = home_descriptors + "/Global"
descriptors_global_folder = np.array(["coHist", "colorGist",
"SpatialPyrDenseScaled",
"Textons/mr_filter",
"Textons/sift_textons"])
global_descriptors = np.array(['coHist', 'colorGist',
'SpatialPyrDenseScaled', 'Textons'])
home_descriptors_segment = home_descriptors + "/SP_Desc_k200"# + "/" + seg_folder
if not os.path.exists(home_descriptors_segment):
os.makedirs(home_descriptors_segment)
home_super_pixels = home_descriptors_segment + "/super_pixels"
if not os.path.exists(home_super_pixels):
os.makedirs(home_super_pixels)
save_path = home_folder + seg_folder + "/Python/"
if not os.path.exists(save_path):
os.makedirs(save_path)
test_set_file_path = home_folder + "TestSet1.txt"
if method == 'gt':
seg_method['label_folder'] = home_label_sets[1]
if method == 'sal':
seg_method['saliency_folder'] = home_folder + seg_folder + '/saliency'
############################################################################
################# Sequential Reading #######################################
#variable to setup a sequential reading of mat files (saving ram power)
seq_reading = True
#loading Images
loader = Loader.Loader.DataLoader(save_path,save_mode = False, seq_reading=seq_reading)
loader.load_images(home_images)
#loading/Calculate SuperPixels from/as mat files
loader.load_super_pixel(home_super_pixels, seg_method)
def main(args=None):
torch.manual_seed(args.seed)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu_devices
# torch.cuda.set_device(0)
use_gpu = torch.cuda.is_available()
if args.use_cpu: use_gpu = False
if not args.evaluate:
sys.stdout = Logger(osp.join(args.save_dir, args.log_train))
else:
sys.stdout = Logger(osp.join(args.save_dir, args.log_test))
print("==========\nArgs:{}\n==========".format(args))
if use_gpu:
print("Currently using GPU {}".format(args.gpu_devices))
cudnn.benchmark = True
torch.cuda.manual_seed_all(args.seed)
else:
print("Currently using CPU (GPU is highly recommended)")
print("Initializing dataset {}".format(args.dataset))
dataset = data_manager.init_dataset(name=args.dataset)
transform_train = T.Compose([
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
T.RandomErasing(),
])
transform_train2 = T.Compose([
T.Resize((args.height, args.width)),
T.Random2DTranslation(args.height, args.width),
])
transform_test = T.Compose([
T.Resize((args.height, args.width)),
T.ToTensor(),
T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
pin_memory = False
trainloader = DataLoader(
VideoDataset(dataset.train,
data_name=args.dataset,
seq_len=args.seq_len,
sample='random',
transform=transform_train,
transform2=transform_train2,
type="train"),
sampler=RandomIdentitySampler(dataset.train,
num_instances=args.num_instances),
batch_size=args.train_batch,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=True,
)
queryloader = DataLoader(
VideoDataset(dataset.query,
data_name=args.dataset,
seq_len=args.seq_len,
sample='dense',
transform=transform_test,
type="test"),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
galleryloader = DataLoader(
VideoDataset(dataset.gallery,
data_name=args.dataset,
seq_len=args.seq_len,
sample='dense',
transform=transform_test,
type="test"),
batch_size=args.test_batch,
shuffle=False,
num_workers=args.workers,
pin_memory=pin_memory,
drop_last=False,
)
print("Initializing models: {}".format(args.arch))
model = models.init_model(name=args.arch,
num_classes=dataset.num_train_pids,
final_dim=args.feat_dim)
print("Model size: {:.5f}M".format(
sum(p.numel() for p in model.parameters()) / 1000000.0))
crossEntropyLoss = CrossEntropyLabelSmooth(
num_classes=dataset.num_train_pids, use_gpu=use_gpu)
tripletLoss = TripletLoss(margin=args.margin)
regularLoss = RegularLoss(use_gpu=use_gpu)
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
scheduler = WarmupMultiStepLR(optimizer, args.stepsize, args.gamma,
args.warmup_factor, args.warmup_items,
args.warmup_method)
start_epoch = args.start_epoch
if use_gpu:
model = nn.DataParallel(model).cuda()
if args.evaluate:
print("Evaluate only")
atest(model, queryloader, galleryloader, use_gpu)
return
start_time = time.time()
best_rank1 = -np.inf
for epoch in range(start_epoch, args.max_epoch):
print("==> Epoch {}/{}".format(epoch + 1, args.max_epoch))
train(model, crossEntropyLoss, tripletLoss, regularLoss, optimizer,
trainloader, use_gpu)
# if args.stepsize > 0:
scheduler.step()
if (epoch + 1) >= 200 and (epoch + 1) % args.eval_step == 0:
print("==> Test")
rank1 = atest(model, queryloader, galleryloader, use_gpu)
is_best = rank1 > best_rank1
if is_best: best_rank1 = rank1
if use_gpu:
state_dict = model.module.state_dict()
else:
state_dict = model.state_dict()
save_checkpoint({
'state_dict': state_dict,
}, is_best,
osp.join(
args.save_dir,
args.model_name + str(epoch + 1) + '.pth.tar'))
elapsed = round(time.time() - start_time)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}".format(elapsed))
def main():
# Load parameters.
parser = argparse.ArgumentParser()
parser.add_argument('--classifier',
type=str,
default='TextCNN',
help='classifier to use "LSTM/TextCNN"')
parser.add_argument('--pretrained',
type=bool,
default=False,
help='finetune pre-trained classifier')
parser.add_argument('--mode',
type=str,
default='train',
help='train or eval')
parser.add_argument('--epochs',
type=int,
default=50,
help='number of training epochs')
parser.add_argument('--batch_size',
type=int,
default=64,
help='number of examples to process in a batch')
parser.add_argument('--max_norm',
type=float,
default=5.0,
help='max norm of gradient')
parser.add_argument('--embed_trainable',
type=bool,
default=True,
help='finetune pre-trained embeddings')
parser.add_argument(
'--device',
type=str,
default=torch.device('cuda:0' if torch.cuda.is_available() else 'cpu'))
# rationale specific parameters.
parser.add_argument('--lr_enc',
type=float,
default=1e-3,
help='learning rate for the encoder')
parser.add_argument('--lr_gen',
type=float,
default=1e-3,
help='learning rate for the generator')
parser.add_argument(
'--num_hidden_rationale',
type=int,
default=64,
help='number of hidden units for the PreGenerator LSTM for rationale')
parser.add_argument(
'--lstm_layer_rationale',
type=int,
default=2,
help='number of layers for the PreGenerator LSTM for rationale')
parser.add_argument(
'--lstm_bidirectional_rationale',
type=bool,
default=True,
help='bi-direction for the PreGenerator LSTM for rationale')
parser.add_argument('--lambda_1',
type=float,
default=1e-2,
help='regularizer of the length of selected words')
parser.add_argument('--lambda_2',
type=float,
default=1e-3,
help='regularizer of the local coherency of words')
parser.add_argument(
'--agg_mode',
type=str,
default='fc',
help='aggregation mode chosen after the pregenerator LSTM layer')
# LSTM specific parameters.
parser.add_argument('--num_hidden',
type=int,
default=256,
help='number of hidden units in the LSTM classifier')
parser.add_argument('--lstm_layer',
type=int,
default=2,
help='number of layers of lstm')
parser.add_argument('--lstm_bidirectional',
type=bool,
default=True,
help='bi-direction of lstm')
# TextCNN specific parameters.
parser.add_argument('--num_classes',
type=int,
default=6,
help='number of target classes')
parser.add_argument('--kernel_sizes',
nargs='+',
type=int,
default=[2, 3, 4],
help='kernel sizes for the convolution layer')
parser.add_argument('--p', type=float, default=0.5, help='dropout rate')
parser.add_argument('--c_out',
type=int,
default=32,
help='output channel size of the convolution layer')
args = parser.parse_args()
# Create log object.
if args.mode == 'train':
sys.stdout = Logger(TRAIN_LOG_LOC)
print_statement('HYPERPARAMETER SETTING')
print_flags(args)
train(args, GEN_MODEL_LOC, LSTM_MODEL_LOC, TCN_MODEL_LOC,
LABEL_JSON_LOC)
else:
sys.stdout = Logger(TEST_LOG_LOC)
print_statement('HYPERPARAMETER SETTING')
print_flags(args)
test(args, GEN_MODEL_LOC, LSTM_MODEL_LOC, TCN_MODEL_LOC,
LABEL_JSON_LOC)
def main(env, args):
# Initiate args useful for training
start_episode = 0
args.current_frame = 0
args.eval_start = 0
args.test_num = 0
args.test_time = False
args.best_avg_return = -1
# Make checkpoint path if there is none
if not os.path.isdir(args.checkpoint):
mkdir_p(args.checkpoint)
# Instantiate metric tracking for progress bar
# TODO: Add any other metrics we need to track
# TODO: Still need to keep track of time for progress bar
args.rewards = AverageMeter()
args.returns = AverageMeter()
args.episode_lengths = AverageMeter()
args.losses = AverageMeter()
# Model & experiences
print("==> creating model '{}' with '{}' noise".format(
args.alg, args.noise))
if args.alg == 'dqn':
args.epsilon_greed_init = args.epsilon_greed
args.initial_threshold = -math.log(
1 - args.epsilon_greed + args.epsilon_greed / args.action_dim)
model = DQN(action_space=env.action_space,
noise=args.noise,
initial_threshold=args.initial_threshold)
target_model = DQN(action_space=env.action_space,
noise=args.noise,
initial_threshold=args.initial_threshold)
target_model.load_state_dict(model.state_dict())
args.memory = ReplayBuffer(args.replay_memory, args.use_cuda)
else:
model = PPO(action_space=env.action_space,
noise=args.noise,
clip_epsilon=args.clip_epsilon)
# TODO: Instantiate RolloutStorage
# rollouts = RolloutStorage(args.horizon, arg.processes?,...)
# House models on GPU if needed
if args.use_cuda:
model.cuda()
if args.alg == 'dqn':
target_model.cuda()
# Criterions and optimizers
value_criterion = nn.functional.mse_loss
if args.alg == 'dqn':
if args.noise == 'adaptive':
optimizer = optim.Adam(model.parameters(), lr=1e-4)
else:
optimizer = optim.RMSprop(model.parameters(),
lr=2.5e-4,
momentum=0.95,
alpha=0.95,
eps=1e-2)
else:
policy_criterion = model.surrogate_loss
# TODO: revisit the choices here. Might be best to just go with defaults from PPO paper
if args.noise == 'learned':
optimizer = optim.RMSprop(model.parameters(),
lr=2.5e-4,
momentum=0.95,
alpha=0.95,
eps=1e-2)
else:
optimizer = optim.Adam(model.parameters(), lr=3e-4)
# Resume
# Unload status, meters, and previous state_dicts from checkpoint
print("==> resuming from '{}' at frame {}".
format(args.resume, args.start_frame) if args.resume else
"==> starting from scratch at frame {}".format(args.start_frame))
title = '{}-{}'.format(args.noise, args.env_id)
if args.resume:
# Load checkpoint.
assert os.path.isfile(
args.resume), 'Error: no checkpoint directory found!'
args.checkpoint = os.path.dirname(args.resume)
checkpoint = torch.load(args.resume)
start_episode = checkpoint['episode'] + 1
args.current_frame = checkpoint['frame'] + 1
model.load_state_dict(checkpoint['state_dict'])
if args.alg == 'dqn':
target_model.load_state_dict(checkpoint['target_state_dict'])
args.returns = checkpoint['returns']
args.best_avg_return = checkpoint['best_avg_return']
args.episode_lengths = checkpoint['episode_lengths']
optimizer.load_state_dict(checkpoint['optimizer'])
args.logger = Logger(os.path.join(args.checkpoint,
'{}-log.txt'.format(title)),
title=title,
resume=True)
args.test_logger = Logger(os.path.join(
args.checkpoint, 'eval-{}-log.txt'.format(title)),
title=title,
resume=True)
else:
args.logger = Logger(os.path.join(args.checkpoint,
'{}-log.txt'.format(title)),
title=title)
args.logger.set_names(
['Episode', 'Frame', 'EpLen', 'AvgLoss', 'Return'])
args.test_logger = Logger(os.path.join(
args.checkpoint, 'eval-{}-log.txt'.format(title)),
title=title)
args.test_logger.set_names(['Frame', 'EpLen', 'Return'])
# We need at least one experience in the replay buffer for DQN
if args.alg == 'dqn':
true_warmup = min(args.memory_warmup, args.replay_memory)
print("==> filling replay buffer with {} transition(s)".format(
true_warmup))
state = env.reset()
for i in range(true_warmup):
action = random.randrange(args.action_dim)
successor, reward, done, _ = env.step(action)
args.memory.add(state, action, reward, successor, done)
state = successor if not done else env.reset()
# Need next reset to be a true reset (due to EpisodicLifeEnv)
env.was_real_done = True
# Initialize bars
args.bar = Bar('Training', max=args.n_frames)
print("==> beginning training for {} frames".format(args.n_frames))
for episode in itertools.count(start_episode):
# Train model
if args.alg == 'dqn':
env, model, target_model, optimizer, args = trainDQN(
env, model, target_model, optimizer, value_criterion, args)
else:
env, model, optimizer, args = trainPPO(env, model, optimizer,
value_criterion,
policy_criterion, args)
# Checkpoint model to disk
is_best = args.returns.avg > args.best_avg_return
if is_best:
args.best_avg_return = args.returns.avg
save_checkpoint(
{
'episode':
episode,
'frame':
args.current_frame,
'state_dict':
model.state_dict(),
'target_state_dict':
target_model.state_dict() if args.alg == 'dqn' else None,
'rewards':
args.rewards,
'returns':
args.returns,
'best_avg_return':
args.best_avg_return,
'episode_lengths':
args.episode_lengths,
'losses':
args.losses,
'optimizer':
optimizer.state_dict()
}, is_best, title)
# Log metrics (episode, frame, episode length, average loss, return)
args.logger.append([
episode, args.current_frame, args.episode_lengths.val,
args.losses.avg, args.returns.val
])
# Reset frame-level meters
args.losses.reset()
args.rewards.reset()
# Handle testing
if args.test_time:
# For testing only
print("==> evaluating agent for {} frames at frame {}".format(
args.eval_period, args.current_frame))
args.eval_start = args.current_frame
args.testing_frame = args.current_frame
args.test_bar = Bar('Testing', max=args.eval_period)
args.test_rewards = AverageMeter()
args.test_returns = AverageMeter()
args.test_episode_lengths = AverageMeter()
# Main testing loop
while args.testing_frame - args.eval_start < args.eval_period:
if args.alg == 'dqn':
env, args = testDQN(env, model, args)
else:
env, args = testPPO(env, model, args)
args.test_logger.append([
args.testing_frame - args.eval_start,
args.test_episode_lengths.val, args.test_returns.val
])
args.test_episode_lengths.reset()
args.test_rewards.reset()
# For testing only:
#break
# Need next reset to be a true reset
env.was_real_done = True
# Need to turn off testing for next episode
args.test_time = False
args.test_num += 1
args.test_bar.finish()
if args.current_frame > args.n_frames:
break
# For testing only:
# if episode >= 100:
# break
#print('episode: {}'.format(episode))
# TODO: Handle cleanup
args.bar.finish()
args.logger.close()
args.test_logger.close()
args.logger.plot()
env.close()
