def main(BASE_DIR):
# Parse planner args
args = arguments.parse_args()
# Run PDDL translator (from TFD)
prb = args.problem
if args.domain:
domain = args.domain
task = translate.pddl.open(prb, domain)
else:
task = translate.pddl.open(prb)
domain = utils.getDomainName(prb)
# Compute initial horizon estimate querying a satisficing planner
# here: ENHSP by Enrico Scala
hplan = BASE_DIR + '/enhsp/enhsp'
val = BASE_DIR + '/bin/validate'
print('Start horizon computation...')
try:
out = subprocess.check_output([
hplan, '-o', domain, '-f', prb, '-s', 'gbfs', '-ties', 'smaller_g',
'-h', 'haddabs'
])
except subprocess.CalledProcessError as e:
sys.exit()
# Extract plan length from output of ENHSP - actions to be done
match = re.search('Plan-Length:(\d+)', out)
if match:
initial_horizon = int(match.group(1))
print('Intial horizon: {}'.format(initial_horizon))
else:
# Computing horizon with GBFS failed for some reason
print('Could not determine initial horizon with GBFS...')
# Print output of ENHSP for diagnosis and exit
print(out)
sys.exit()
# Compose encoder and search according to user flags
e = encoder.Encoder(task, modifier.LinearModifier())
s = search.LinearSearch(e, initial_horizon)
plan = s.do_search()
# Validate and print plan
try:
if plan.validate(val, domain, prb):
print('\nPlan found!')
plan.do_print()
else:
print('Plan not valid, exiting now...')
sys.exit()
except Exception as e:
print('Could not validate plan, exiting now...')
sys.exit()
def run():
args = arguments.parse_args()
cfg = K.tf.ConfigProto()
cfg.gpu_options.allow_growth = True
K.set_session(K.tf.Session(config=cfg))
if '2d' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_2d(args, args.iterations)
elif '1d' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_1d(args, args.iterations)
elif 'mlp' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_mlp(args, args.iterations)
elif '1d_anno' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_1d_anno(args, args.iterations)
elif '2d_anno' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_2d_anno(args, args.iterations)
elif 'ab_spatial' == args.mode:
ho = HyperparameterOptimizer()
pa = ho.get_baseline_2d_params()
pb = ho.get_baseline_2d_params()
pb['spatial_dropout'] = False
ho.ab_test_2d(args, pa, pb)
else:
raise ValueError('Error! Unknown hyperprameter optimizer mode:',
args.mode)
def trainable(config, name_fmt, envname, trainingconfig, evaluate_mean_n):
# Parse arguments
trial_dir = Path(tune.get_trial_dir()) if tune.get_trial_dir(
) is not None else Path.cwd()
adv_force = config["adv_force"]
name = name_fmt.format(adv_force=adv_force)
cmd_args = [
'--name', name, '--env', envname, '--log', '--trainingconfig',
str(trainingconfig), '--root',
str(trial_dir), '--monitor-dir',
str(monitor_dir_name(envname, adv_force))
]
cmd_args += ['--adv_force', str(adv_force)]
args = parse_args(cmd_args)
# Add adversarial force
logging.info(f'Running {name=} with {args=}')
def evaluate(prot, ts):
# reward = get_mean_reward_last_n_steps(evaluate_mean_n, args.monitor_dir)
# logging.info(f'{name} {reward=:.2f} {ts=}')
# tune.report(reward=reward)
robustness = eval_robustness(args, prot, envname, trainingconfig, name)
logging.info(f'{name} {robustness=:.2f} {ts=}')
tune.report(robustness=robustness)
run(args, evaluate_fn=evaluate)
def main():
logging.basicConfig(level=logging.INFO)
args = parse_args()
result = run(args)
if result is not None:
avg_reward, std_reward = result
logging.info(f'reward={avg_reward}+={std_reward}')
def run():
args = arguments.parse_args()
if '2d' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_2d(args, args.iterations)
elif '1d' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_1d(args, args.iterations)
elif 'mlp' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_mlp(args, args.iterations)
elif '1d_anno' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_1d_anno(args, args.iterations)
elif '2d_anno' == args.mode:
ho = HyperparameterOptimizer()
ho.bayesian_search_2d_anno(args, args.iterations)
elif 'ab_spatial' == args.mode:
ho = HyperparameterOptimizer()
pa = ho.get_baseline_2d_params()
pb = ho.get_baseline_2d_params()
pb['spatial_dropout'] = False
ho.ab_test_2d(args, pa, pb)
else:
raise ValueError('Error! Unknown hyperprameter optimizer mode:',
args.mode)
def saveScatter(Xs, Ys, args=None):
cnt = len(Xs)
# fig, axs = plt.subplots(1, cnt, figsize=(10, 10))
#
# if cnt == 1:
# f1_data = np.concatenate((Xs[0], Ys[0]), axis=2)
# f1_data = f1_data.reshape(f1_data.shape[0], f1_data.shape[1]*f1_data.shape[2])
# f1_data = np.transpose(f1_data, (1, 0))
# axs.scatter(f1_data[0], f1_data[1])
# else:
# for i, (x, y) in enumerate(zip(Xs, Ys)):
# f1_data = np.concatenate((x, y), axis=2)
# f1_data = f1_data.reshape(f1_data.shape[0], f1_data.shape[1]*f1_data.shape[2])
# f1_data = np.transpose(f1_data, (1, 0))
#
# axs[i].scatter(f1_data[0], f1_data[1])
colors = ['green', 'blue', 'yellow', 'red', 'black']
for i, (x, y) in enumerate(zip(Xs, Ys)):
plt.scatter(x, y, color=colors[i])
if args is None:
args = arguments.parse_args()
fig_path = fh.getStoragePath() + 'Figures/'
fh.makeDirectories(fig_path)
plt.savefig(fig_path + 'Layer_%d_Hidden_%d_Epoch_%d.png' %
(args.n_layers, args.n_hidden, args.num_epochs))
plt.clf()
def main():
args = parse_args()
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# multi-gpu training (should be after apex fp16 initialization)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
# Prepare GLUE tasks
processors = {}
output_modes = {}
label_lists = {}
num_label_list = {}
# args.task_params = open(args.task_params)
for key in args.task_params:
print(processors)
processors[key] = glue_processors[key]()
output_modes[key] = glue_output_modes[key]
label_lists[key] = processors[key].get_labels()
num_label_list[key] = len(label_lists[key])
# Configs
configs = {}
for key in args.task_params:
configs[key] = AutoConfig.from_pretrained(args.model_type,
# args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_label_list[key],
finetuning_task=key,
cache_dir=None,
)
# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(
args.model_type,
do_lower_case=args.do_lower_case,
cache_dir=None,
)
# Continual Learning
# no of task configurations specified in example_task.json
n = len(configs)
# final Accuracy of all tasks
accuracy_matrix = np.zeros((n, n))
# TODO
transfer_matrix = np.zeros((n, n))
# List of all tasks defined in example_task.json
tasks = list(args.task_params.keys())
models = []
# Model
# Load specific model for each task and added it to the models list
for key in args.task_params:
def main():
"""Program entrypoint"""
try:
args = parse_args()
args.command(args)
except Exception as error:
print(error)
sys.exit(1)
def run():
args = arguments.parse_args()
args.labels = defines.calling_labels
if args.mode == 'tensors':
infer_tensor(args)
elif args.mode == 'vcf':
infer_vcf(args)
else:
raise ValueError('Unknown calling mode:', args.mode)
def main():
'''Main function to be used when calling getdc.py
Keyword arguments:
args.domain -- Domain Name (required, accepted list or single value, default none)
args.nameserver -- NameServer (optional, accepted hostname/ipaddress, default none)
args.format -- Format output type, (required, accepted values [json, host, ip, hostip, zerologon)
args.verbose -- Toggle debug meesages to stdout (required, accepted values boolean)
args.exchange -- Toggle whether to look up exchange records (required, default false)
'''
# Arguments from argparse
args = arguments.parse_args()
# python logger
if args.verbose:
logging.basicConfig(level=logging.DEBUG)
else:
pass
# Hostname dictionary
host_dct = {
}
# Run func query()
for domain in args.domain:
host_dct[domain] = {} # initiate nested dict(s) from args.domain
answer_srv = query(domain, service='_kerberos.', protocol='_tcp.', recordtype='SRV', nameserver=args.nameserver) # call query(srv) to find dc hostname
# Exchange was defined
if args.exchange:
answer_srv = query(domain, service='_autodiscover.', protocol='_tcp.', recordtype='SRV', nameserver=args.nameserver) # call query(srv) to find exchange hostname
for hostname in answer_srv:
host_dct[domain][str(hostname)] = '' # populat host_dct with hostname
answer_a = query(hostname, service='', protocol='', recordtype='A', nameserver=args.nameserver) # call query(a) to find dc ipaddress
ipaddress = '\n'.join(answer_a) # convert list to string
host_dct[domain][str(hostname)] = ipaddress # populate host_dct with ipaddress
# format type output json
if args.format == 'json': # stdout json
json_data = json.dumps(host_dct, indent=4, sort_keys=True) # convert host_dct to json
print(json_data)
# format type output
for domain in args.domain:
if args.format == 'host':
for key in sorted(host_dct[domain].keys()): # stdout hostname
print(key)
if args.format == 'ip':
for value in sorted(host_dct[domain].values()): # stdout ipaddress
print(value)
if args.format == 'hostip':
for key, value in sorted(host_dct[domain].items()): # stdout hostname/ipaddress
print(key, value)
if args.format == 'zerologon':
for key, value in sorted(host_dct[domain].items()): # stdout netbios-name/ipaddress
print(key.split('.')[0], value)
def main(arglist):
params = parse_args(arglist)
print(params)
if params.epochs == 0:
assert len(params.loadmodel) == 1
modelfile = params.loadmodel[0]
else:
modelfile = run_training(params)
run_eval(params, modelfile)
def main():
'''Main function to be used when calling getdc.py
Keyword arguments:
args.domain -- Domain Name (required, accepted list or single value,default none)
args.nameserver -- NameServer (optional, accepted hostname/ipaddress, default none)
args.format -- Format output type, (required, accepted values json/host/ip/hostip'NameServer, default json)
args.verbose -- Toggle debug meesages to stdout (required, accepted values boolean)
'''
# Arguments from argparse
args = arguments.parse_args()
# python logger
if args.verbose:
logging.basicConfig(level=logging.DEBUG)
else:
pass
# DomainController dictionary
dc_dct = {}
# Run main func query()
for domain in args.domain:
dc_dct[domain] = {} # initiate nested dict(s) from args.domain
answer_srv = query(
domain, recordtype='SRV',
nameserver=args.nameserver) # call main func query(srv)
for hostname in answer_srv:
dc_dct[domain][str(hostname)] = '' # populat dc_dict with hostname
answer_a = query(
hostname, recordtype='A',
nameserver=args.nameserver) # call the func query(a)
ipaddress = '\n'.join(answer_a) # convert list to string
dc_dct[domain][str(
hostname)] = ipaddress # populate dc_dict with ipaddress
# format type output json
if args.format == 'json': # stdout json
json_data = json.dumps(dc_dct, indent=4,
sort_keys=True) # convert dc_dict to json
print(json_data)
# format type output
for domain in args.domain:
if args.format == 'host':
for key in sorted(dc_dct[domain].keys()): # stdout hostname
print(key)
if args.format == 'ip':
for value in sorted(dc_dct[domain].values()): # stdout ipaddress
print(value)
if args.format == 'hostip':
for key, value in sorted(
dc_dct[domain].items()): # stdout hostname/ipaddress
print(key, value)
def main():
args = parse_args()
set_seed(args.seed)
image_folder = args.data_dir
annotation_csv = args.annotation_dir
model_dir = args.model_dir
depth_folder = args.depth_dir
trainloader, valloader = LoadData(depth_folder, image_folder,
annotation_csv, args)
model = UNet_Encoder_Decoder(3, args)
model.to(args.device)
model = model.apply(weight_init)
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
criterion = nn.BCEWithLogitsLoss()
mse_criterion = nn.MSELoss()
num_epochs = args.num_train_epochs
if (not os.path.exists(model_dir)):
os.mkdir(model_dir)
best_eval_acc = 0.0
for epoch in tqdm(range(num_epochs)):
data_len = len(trainloader)
model.train()
running_loss, model = train_epoch(model, trainloader, args,
all_criterion)
eval_roadmap_loss, eval_depth_loss, eval_acc = evaluate(
model, valloader, args, (criterion, mse_criterion))
print(
'[%d, %5d] Loss: %.3f Eval Roadmap Loss: %.3f Eval Depth Loss: %.3f Eval ThreatScore: %.3f'
% (epoch + 1, num_epochs, running_loss /
(args.per_gpu_batch_size * data_len), eval_roadmap_loss,
eval_depth_loss, eval_acc))
torch.save(model.state_dict(),
os.path.join(model_dir, 'model_' + str(epoch) + '.pth'))
if eval_acc > best_eval_acc:
torch.save(
model.state_dict(),
os.path.join(model_dir, 'bestmodel_' + str(epoch) + '.pth'))
best_eval_acc = eval_acc
def main():
args = parse_args()
_print_args(args)
if (
os.path.isfile(downloader.MODEL_DOWNLOADER_PATH)
and os.path.isfile(downloader.MODEL_CONVERTER_PATH)
and os.path.isfile(downloader.MODEL_OPTIMIZER_PATH)
):
downloader.download_and_convert_models(
args.model_list, args.output_dir, args.force, args.dl_streamer_version
)
else:
print(
"Intel(R) distribution of OpenVINO(TM) Toolkit tools not "
"found. Please check if all dependent tools are installed and try again."
)
sys.exit(1)
def main():
try:
args = parse_args()
_log_options(args)
frame_source = None
frame_queue = queue.Queue()
result_queue = queue.Queue()
msp = MediaStreamProcessor(args.grpc_server_address,
args.use_shared_memory)
_, extension = os.path.splitext(args.sample_file)
if extension in ['.png', '.jpg']:
frame_source = ImageSource(args.sample_file, args.loop_count)
elif extension in ['.mp4']:
frame_source = VideoSource(args.sample_file)
else:
print("{}: unsupported file type".format(args.sample_file))
sys.exit(1)
width, height = frame_source.dimensions()
print("{} {}".format(width, height))
msp.start(width, height, frame_queue, result_queue)
with open(args.output_file, "w") as output:
image = frame_source.get_frame()
while image:
frame_queue.put(image)
while not result_queue.empty():
result = result_queue.get()
print_result(result, output)
image = frame_source.get_frame()
frame_queue.put(None)
result = result_queue.get()
while result:
print_result(result, output)
result = result_queue.get()
frame_source.close()
except Exception:
log_exception()
sys.exit(-1)
# print_rank_0(f'AutoMP: embedding_output = {embedding_output}')
def gpt2_attention_mask_func(attention_scores, ltor_mask):
attention_scores.masked_fill_(ltor_mask, -10000.0)
return attention_scores
transformer = ParallelTransformer(
attention_mask_func=gpt2_attention_mask_func,
num_layers=args.num_layers,
hidden_size=args.hidden_size,
layernorm_epsilon=args.layernorm_epsilon,
num_attention_heads=args.num_attention_heads,
attention_dropout=0.1,
hidden_dropout=0.1)
attention_mask, loss_mask, position_ids = get_ltor_masks_and_position_ids(
input_indices, vocab_size - 1)
transformer_output = transformer.forward(hidden_states=embedding_output,
attention_mask=attention_mask)
print_rank_0(f'AutoMP: transformer_output = {transformer_output}')
if __name__ == '__main__':
# Parse command line arguments
parse_args()
args = get_args()
train()
def main():
args = parse_args()
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Prepare GLUE tasks
processors = {}
output_modes = {}
label_lists = {}
num_label_list = {}
for key in args.task_params:
print(processors)
processors[key] = glue_processors[key]()
output_modes[key] = glue_output_modes[key]
label_lists[key] = processors[key].get_labels()
num_label_list[key] = len(label_lists[key])
# Configs
configs = {}
for key in args.task_params:
configs[key] = AutoConfig.from_pretrained(args.model_type,
# args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_label_list[key],
finetuning_task=key,
cache_dir=None,
)
# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(
args.model_type,
do_lower_case=args.do_lower_case,
cache_dir=None,
)
# Continual Learning
n = len(configs)
accuracy_matrix = np.zeros((n, n))
transfer_matrix = np.zeros((n, n))
tasks = list(args.task_params.keys())
models = []
# Model
for key in args.task_params:
if args.n_gpu <= 1:
# if key == 'mrpc':
# models.append((key,
# torch.load('mrpc_bert_2_5e5.ckpt')))
# # shared_model(AutoModelForSequenceClassification.from_pretrained(args.model_type))))
# accuracy_matrix = np.array([[0.86 , 0.5], [0, 0]])
# else:
models.append((key,
shared_model(AutoModelForSequenceClassification.from_pretrained(args.model_type))))
else:
models.append((key,
shared_model(torch.nn.DataParallel(
AutoModelForSequenceClassification.from_pretrained(args.model_type)))))
for i in range(n):
models[i][1].to(args.device)
save_model(args, i, models[i][1])
for i in range(len(configs)):
# if i == 0 and models[i][0] == 'mrpc':
# continue
# else:
if i > 0:
part_name = partial_name(args)
# Always load the BERT parameters of previous model
models[i][1].tmodel.load_state_dict(
torch.load(os.path.join(args.output_dir, part_name, "bert_paramters_" + str(i - 1) + ".pt")), strict=False)
models[i][1].reg_params = models[i - 1][1].reg_params
new_args = convert_dict(args.task_params[tasks[i]], args)
train_dataset = load_and_cache_examples(args, tasks[i], tokenizer, evaluate=False)
global_step, tr_loss, accuracy_matrix, new_model = train(new_args, train_dataset, tasks[i], tasks, models[i][1],
i, tokenizer, accuracy_matrix)
logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
print()
print("***** Accuracy Matrix *****")
print('regularization lambda = {}'.format(args.reg_lambda))
print()
print(accuracy_matrix)
print()
print("***** Transfer Matrix *****")
# print("Future Transfer => Upper Triangular Matrix || Backward Transfer => Lower Triangular Matrix")
print()
for i in range(n):
for j in range(n):
transfer_matrix[j][i] = accuracy_matrix[j][i] - accuracy_matrix[i][i]
print(transfer_matrix)
def main():
params = parse_args(default_files=('./gibson_submission.conf', ))
is_submission = (params.gibson_mode == 'submission')
if is_submission:
from gibson2.envs.challenge import Challenge
challenge = Challenge()
agent = MappingAgent(params)
challenge.submit(agent)
print("Done with submission.")
raise SystemExit
if params.gibson_track != '':
os.environ['SIM2REAL_TRACK'] = params.gibson_track
track = os.environ['SIM2REAL_TRACK']
# challenge.submit(agent)
models = sorted(os.listdir(gibson2.dataset_path))
skip = []
if params.gibson_split == "custom":
# # models = models[:5]
# models = models[14:15]
# # models = models[-10:]
# num_episodes_per_floor = 1
# num_repeats = 1
models = models[:-10]
num_episodes_per_floor = 5
num_repeats = 8
skip = range(708)
elif params.gibson_split == "train":
# models = models[67:-10] + models[:66]
# models = models[:66] + models[67:-10]
models = models[:-10]
num_episodes_per_floor = 5
num_repeats = 4
elif params.gibson_split == "test":
models = models[-10:]
num_episodes_per_floor = 5
num_repeats = 4
# models = models[-3:]
elif params.gibson_split == "evaltest":
models = models[-10:]
num_episodes_per_floor = 10
num_repeats = 1
# models = models[-3:]
elif params.gibson_split == "minitest":
models = models[-10:]
num_episodes_per_floor = 5
num_repeats = 1
# models = models[-3:]
else:
raise ValueError("Unknown split %s" % params.gibson_split)
models = models
if params.gibson_mode == 'gen_maps':
num_repeats = 1
print(models)
assert len(models) < 1000
print("Generating maps for %d models." % len(models))
for model_i, model_name in enumerate(models):
# p = Process(target=save_episodes_helper, args=(args, model_name, output_filename + '.{:03d}'.format(i), num_episodes_per_floor=num_episodes_per_floor))
# p.start()
# p.join()
# p.terminate()
if model_i in skip:
continue
run_in_separate_process(gen_map_helper, params, model_name)
elif params.gibson_mode in ['gen_scenarios']:
models = models * num_repeats
print(models)
assert len(models) < 1000
output_filename = './data/scenarios_{}_{}_{}.tfrecords'.format(
params.gibson_split, track,
time.strftime('%m-%d-%H-%M-%S', time.localtime()))
for model_i, model_name in enumerate(models):
if model_i in skip:
continue
run_in_separate_process(
save_episodes_helper,
model_name,
output_filename + '.{:03d}'.format(model_i),
num_episodes_per_floor=num_episodes_per_floor)
elif params.gibson_mode in ['eval']:
models = models * num_repeats
print(models)
assert len(models) < 1000
evaluate_episodes(params, models, num_episodes_per_floor)
elif params.gibson_mode in ['evalsubmission']:
models = models * num_repeats
print(models)
assert len(models) < 1000
from gibson2.envs.challenge import Challenge
for model_i, model_name in enumerate(models):
challenge = Challenge()
agent = MappingAgent(params)
challenge.submit(agent)
else:
raise ValueError("Unknown gibson_mode=%s" % params.gibson_mode)
def main():
import arguments
arguments.parse_args()
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
# parser.add_argument('n_gpu', default=1)
"""
if args.local_rank == -1 or args.no_cuda:
device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu") args.n_gpu = 0 if args.no_cuda else torch.cuda.device_count() else: # Initializes the distributed backend which will take care of sychronizing nodes/GPUs torch.cuda.set_device(args.local_rank) device = torch.device("cuda", args.local_rank) torch.distributed.init_process_group(backend="nccl") args.n_gpu = 1 args.device = device
"""
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
# try:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# except ValueError as e:
# print('Additional arguments passed not parsed by ArgumentParser: {}'.format(e))
if (
os.path.exists(training_args.output_dir)
and os.listdir(training_args.output_dir)
and training_args.do_train
and not training_args.overwrite_output_dir
):
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. Use --overwrite_output_dir to overcome."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO if training_args.local_rank in [-1, 0] else logging.WARN,
)
logger.warning(
"Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
training_args.local_rank,
training_args.device,
training_args.n_gpu,
bool(training_args.local_rank != -1),
training_args.fp16,
)
logger.info("Training/evaluation parameters %s", training_args)
args = arguments.parse_args()
# Set seed
set_seed(training_args.seed if args.seed is None else args.seed)
try:
num_labels = glue_tasks_num_labels[data_args.task_name] if random_labels == 0 else random_labels
output_mode = glue_output_modes[data_args.task_name]
except KeyError:
raise ValueError("Task not found: %s" % (data_args.task_name))
# Load pretrained model and tokenizer
#
# Distributed training:
# The .from_pretrained methods guarantee that only one local process can concurrently
# download model & vocab.
config = SANConfig(
model_args.config_name if model_args.config_name else model_args.model_name_or_path,
num_labels=num_labels,
hidden_act="gelu",
intermediate_size=args.hidden_dim,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
)
config.max_position_embeddings = -1
config.train_loss_l = []
# Can enable test_width to observe effects of widths.
test_width = False
if test_width:
depth = True
depth = False
if depth:
print("DEPTH")
config.num_hidden_layers = 3 # 8 #8
config.num_attention_heads = 3 # 8
dim_per_head = 51
else:
print("BREADTH")
config.num_hidden_layers = 6 # 5
config.num_attention_heads = 1 # 1 #10 6
dim_per_head = 64
print(config.num_hidden_layers, config.num_attention_heads)
config.hidden_size = config.num_attention_heads * dim_per_head
config.intermediate_size = 2 * config.hidden_size
if do_paths:
config.num_hidden_layers = args.depth # 1
config.num_attention_heads = args.width # 1
config.hidden_dropout_prob = 0.1
config.hidden_size = args.hidden_dim
config.intermediate_size = config.hidden_size # *2
# hidden_size is total hidden dimension, to be divided across heads, num_heads is number of heads
tokenizer = AutoTokenizer.from_pretrained(
model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path,
cache_dir=model_args.cache_dir,
)
if do_paths or test_width:
# model = AutoModelForSequenceClassification.from_config(config=config)
model = SANForSequenceClassification(config=config)
else:
model = AutoModelForSequenceClassification.from_pretrained(
model_args.model_name_or_path,
from_tf=bool(".ckpt" in model_args.model_name_or_path),
config=config,
cache_dir=model_args.cache_dir,
)
if do_paths or test_width:
# count params
total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
for name, p in model.named_parameters():
print("p {} size {}".format(name, p.numel()))
print("+++TOTAL model params {}+++".format(total_params))
model_path = "modelNLP{}d{}_{}_{}.pt".format(
config.num_attention_heads, config.num_hidden_layers, config.hidden_size, int(training_args.num_train_epochs)
) # modelNLP4d6_256_10.pt
if os.path.exists(model_path):
state_dict = torch.load(model_path, map_location="cpu")
model.load_state_dict(state_dict)
model = model.to(training_args.device)
training_args.num_train_epochs = 0
# Get datasets
train_dataset = (
GlueDataset(data_args, tokenizer=tokenizer, cache_dir=model_args.cache_dir) if training_args.do_train else None
)
eval_dataset = (
GlueDataset(data_args, tokenizer=tokenizer, mode="dev", cache_dir=model_args.cache_dir)
if training_args.do_eval
else None
)
test_dataset = (
GlueDataset(data_args, tokenizer=tokenizer, mode="test", cache_dir=model_args.cache_dir)
if training_args.do_predict
else None
)
def build_compute_metrics_fn(task_name: str) -> Callable[[EvalPrediction], Dict]:
def compute_metrics_fn(p: EvalPrediction):
if output_mode == "classification":
preds = np.argmax(p.predictions, axis=1)
elif output_mode == "regression":
preds = np.squeeze(p.predictions)
return glue_compute_metrics(task_name, preds, p.label_ids) if random_labels == 0 else 0
return compute_metrics_fn
if random_labels:
print("Note: random labels assigned to data!")
target_feat_len = min(len(train_dataset.features), args.n_train_data)
train_dataset.features = train_dataset.features[:target_feat_len]
# traindataset consists of InputFeatures objects, each having label, attention mask, input_ids, etc.
# torch.manual_seed(44)
rand = torch.randint(random_labels, (len(train_dataset),))
n_tok = len(train_dataset[0].input_ids)
# Learn and evaluate at token-level.
test_token = True
for i, feat in enumerate(train_dataset):
if test_token:
rand_labels = torch.randint(random_labels, (n_tok,)) # len(feat.input_ids)
rand_labels[torch.Tensor(feat.attention_mask) == 0] = -1
feat.label = rand_labels
else:
feat.label = rand[i]
training_args.logging_steps = 100
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
compute_metrics=build_compute_metrics_fn(data_args.task_name),
args_sort=args,
)
# Training
if training_args.do_train:
trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
trainer.save_model()
# For convenience, we also re-save the tokenizer to the same directory,
# so that you can share your model easily on huggingface.co/models =)
if trainer.is_world_master():
tokenizer.save_pretrained(training_args.output_dir)
# Evaluation
eval_results = {}
if training_args.do_eval:
logger.info("*** Evaluate ***")
# Loop to handle MNLI double evaluation (matched, mis-matched)
##eval_datasets = [eval_dataset]
eval_datasets = [train_dataset]
if data_args.task_name == "mnli":
mnli_mm_data_args = dataclasses.replace(data_args, task_name="mnli-mm")
eval_datasets.append(
GlueDataset(mnli_mm_data_args, tokenizer=tokenizer, mode="dev", cache_dir=model_args.cache_dir)
)
# config.no_sub_path = args.no_sub_path
# generate paths
# convex_hull.create_path(path_len, args, all_heads=args.all_heads )
for eval_dataset in eval_datasets:
trainer.compute_metrics = build_compute_metrics_fn(eval_dataset.args.task_name)
trainer.train(
model_path=model_args.model_name_or_path if os.path.isdir(model_args.model_name_or_path) else None
)
# eval_result = trainer.evaluate(eval_dataset=eval_dataset)
output_eval_file = os.path.join(
training_args.output_dir, f"eval_results_{eval_dataset.args.task_name}.txt"
)
if trainer.is_world_master():
with open(output_eval_file, "w") as writer:
logger.info("***** Eval results {} *****".format(eval_dataset.args.task_name))
for key, value in eval_result.items():
logger.info(" %s = %s", key, value)
writer.write("%s = %s\n" % (key, value))
eval_results.update(eval_result)
if training_args.do_predict:
logging.info("*** Test ***")
test_datasets = [test_dataset]
if data_args.task_name == "mnli":
mnli_mm_data_args = dataclasses.replace(data_args, task_name="mnli-mm")
test_datasets.append(
GlueDataset(mnli_mm_data_args, tokenizer=tokenizer, mode="test", cache_dir=model_args.cache_dir)
)
for test_dataset in test_datasets:
predictions = trainer.predict(test_dataset=test_dataset).predictions
if output_mode == "classification":
predictions = np.argmax(predictions, axis=1)
output_test_file = os.path.join(
training_args.output_dir, f"test_results_{test_dataset.args.task_name}.txt"
)
if trainer.is_world_master():
with open(output_test_file, "w") as writer:
logger.info("***** Test results {} *****".format(test_dataset.args.task_name))
writer.write("index\tprediction\n")
for index, item in enumerate(predictions):
if output_mode == "regression":
writer.write("%d\t%3.3f\n" % (index, item))
else:
item = test_dataset.get_labels()[item]
writer.write("%d\t%s\n" % (index, item))
return eval_results
for late in lates:
print(late.value)
return []
def dist(start, end):
(ys, xs), (ye, xe) = start, end
return abs(ys - ye) + abs(xs - xe)
def build_times(demand):
def check(first, second):
ride = dist(first[0], first[1]) + dist(first[1], second[0])
return first[2] + ride - second[2]
times = [[check(first, second) for second in demand] for first in demand]
return times
# --------------------------- Argument parsing ---------------------------------
if __name__ == '__main__':
parsed_args = parse_args(True)
file_name = parsed_args.pop('file_name')
if file_name == 'all':
for file_name in file_names:
main(**parsed_args)
else:
main(**parsed_args)
def main():
args = parse_args()
# Set the GPU to use
torch.cuda.set_device(args.gpu)
transform = transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
vqa_loader = dataset.get_train_dataloader(osp.expanduser(args.annotations),
osp.expanduser(args.questions),
args.images,
args,
raw_images=args.raw_images,
transforms=transform)
# We always use the vocab from the training set
vocab = vqa_loader.dataset.vocab
maps = {
"vocab": vocab,
"word_to_wid": vqa_loader.dataset.word_to_wid,
"wid_to_word": vqa_loader.dataset.wid_to_word,
"ans_to_aid": vqa_loader.dataset.ans_to_aid,
"aid_to_ans": vqa_loader.dataset.aid_to_ans,
}
val_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
])
val_loader = dataset.get_val_dataloader(osp.expanduser(
args.val_annotations),
osp.expanduser(args.val_questions),
args.val_images,
args,
raw_images=args.raw_images,
maps=maps,
vocab=vocab,
shuffle=False,
transforms=val_transform)
arch = Models[args.arch].value
model = arch(len(vocab),
output_dim=args.top_answer_limit,
raw_images=args.raw_images)
if args.resume:
state = torch.load(args.resume)
model.load_state_dict(state["model"])
model.cuda()
criterion = nn.CrossEntropyLoss().cuda()
optimizer = optim.Adam(model.parameters(),
lr=args.lr,
betas=tuple(args.betas),
weight_decay=args.weight_decay)
scheduler = lr_scheduler.StepLR(optimizer,
step_size=args.decay_interval,
gamma=args.lr_decay)
if args.visualize:
vis = visualize.Visualizer(args.port)
else:
vis = None
print("Beginning training")
print("#" * 80)
for epoch in range(args.start_epoch, args.epochs):
scheduler.step()
trainer.train(model,
vqa_loader,
criterion,
optimizer,
epoch,
args,
vis=vis)
trainer.evaluate(model, val_loader, criterion, epoch, args, vis=vis)
print("Training complete!")
def compile_results(settings_list):
batch_exp_name = settings_list[0]["experiment_id"]
# if batch_exp_name == "2":
RL_DIR = "rl_runs"
# elif batch_exp_name == "1":
# EVO_DIR = "evo_runs_06-13"
# RL_DIR = "evo_runs_06-14"
# ignored_keys = set(
# (
# "exp_name",
# "evaluate",
# "show_vis",
# "visualize",
# "render_levels",
# "multi_thread",
# "play_level",
# "evaluate",
# "save_levels",
# "cascade_reward",
# "model",
# "n_generations",
# "render",
# "infer",
# )
# )
# keys = []
# for k in settings_list[0].keys():
# if k not in ignored_keys:
# keys.append(k)
keys = [
"problem",
"representation",
"conditionals",
"alp_gmm",
"change_percentage"
]
columns = None
data = []
vals = []
for i, settings in enumerate(settings_list):
val_lst = []
controllable = False
for k in keys:
v = settings[k]
if k == 'conditionals':
if k != ['NONE']:
controllable = True
if isinstance(settings[k], list):
if len(settings[k]) < 2:
val_lst.append("-".join(settings[k]))
else:
val_lst.append(newline(settings[k][0]+'-', v[1]))
elif k == 'alp_gmm':
if not controllable:
v = ''
elif v:
v = 'learning'
else:
v = 'random'
val_lst.append(v)
else:
val_lst.append(v)
args = parse_args(load_args=settings)
arg_dict = vars(args)
# FIXME: well this is stupid
arg_dict["cond_metrics"] = arg_dict.pop("conditionals")
exp_name = get_exp_name(
arg_dict.pop("problem"), arg_dict.pop("representation"), **arg_dict
) + "_{}_log".format(batch_exp_name)
# NOTE: For now, we run this locally in a special directory, to which we have copied the results of eval on
# relevant experiments.
exp_name = os.path.join(RL_DIR, exp_name)
stats_f = os.path.join(exp_name, "eval", "scores_ctrlTrgs.json")
fixTrgs_stats_f = os.path.join(exp_name, "eval", "scores_fixTrgs.json")
if not (os.path.isfile(stats_f) and os.path.isfile(fixTrgs_stats_f)):
print(stats_f)
print(
"skipping evaluation of experiment due to missing stats file(s): {}".format(
exp_name
)
)
continue
vals.append(tuple(val_lst))
data.append([])
stats = json.load(open(stats_f, "r"))
fixLvl_stats = json.load(open(fixTrgs_stats_f, "r"))
flat_stats = flatten_stats(fixLvl_stats)
flat_stats.update(flatten_stats(stats, controllable=True))
if columns is None:
columns = list(flat_stats.keys())
for j, c in enumerate(columns):
if c not in flat_stats:
data[-1].append("N/A")
else:
data[-1].append(flat_stats[c])
tuples = vals
# Rename headers
new_keys = []
for k in keys:
if k in header_text:
new_keys.append(header_text[k])
else:
new_keys.append(k)
for (i, lst) in enumerate(tuples):
new_lst = []
for v in lst:
if v in header_text:
new_lst.append(header_text[v])
else:
new_lst.append(v)
tuples[i] = new_lst
index = pd.MultiIndex.from_tuples(tuples, names=new_keys)
# df = index.sort_values().to_frame(index=True)
df = pd.DataFrame(data=data, index=index, columns=columns).sort_values(by=new_keys)
# print(index)
csv_name = r"{}/cross_eval_{}.csv".format(RL_DIR, batch_exp_name)
html_name = r"{}/cross_eval_{}.html".format(RL_DIR, batch_exp_name)
df.to_csv(csv_name)
df.to_html(html_name)
print(df)
# tex_name = r"{}/zelda_empty-path_cell_{}.tex".format(OVERLEAF_DIR, batch_exp_name)
# FIXME: F*****G ROUND YOURSELF DUMB FRIEND
# df = df.round(2)
for p in ["binary", "zelda", "sokoban"]:
tex_name = "{}/{}_{}.tex".format(RL_DIR, p, batch_exp_name)
df_tex = df.loc[p, "narrow"]
p_name = p + '_ctrl'
lcl_conds = ['None'] + ['-'.join(pi) if len(pi) < 2 else newline(pi[0]+'-',pi[1]) for pi in local_controls[p_name]]
print(lcl_conds)
df_tex = df_tex.loc[lcl_conds]
# df_tex = df_tex.sort_values(by=['ALP GMM'])
z_cols = [
header_text["net_score (mean)"],
header_text["diversity_score (mean)"],
header_text["(controls) net_score (mean)"],
# header_text["(controls) ctrl_score (mean)"],
# header_text["(controls) fixed_score (mean)"],
header_text["(controls) diversity_score (mean)"],
]
# df_tex = df.drop(columns=z_cols)
df_tex = df_tex.loc[:, z_cols]
df_tex = df_tex * 100
df_tex = df_tex.round(0)
dual_conds = ['None', lcl_conds[1]]
for k in z_cols:
if k in df_tex:
# df_tex.loc[dual_conds][k] = df_tex.loc[dual_conds][k].apply(
# lambda data: bold_extreme_values(data, data_max=df_tex.loc[dual_conds][k].max())
# )
df_tex[k] = df_tex[k].apply(
lambda data: bold_extreme_values(data, data_max=df_tex[k].max())
)
# df_tex = df_tex.round(2)
# df_tex.reset_index(level=0, inplace=True)
print(df_tex)
with open(tex_name, "w") as tex_f:
col_widths = "p{0.5cm}p{0.5cm}p{0.5cm}p{0.5cm}p{0.5cm}p{0.5cm}p{0.8cm}p{0.8cm}p{0.8cm}"
df_tex.to_latex(
tex_f,
index=True,
columns=z_cols,
multirow=True,
# column_format=col_widths,
escape=False,
caption=("Performance of controllable {}-generating agents with learning-progress-informed and uniform-random control regimes and baseline (single-objective) agents with various change percentage allowances.".format(p)),
label={"tbl:{}".format(p)},
)
def main():
args = parse_args()
print('Called with args:')
print(args)
cfg = set_configs(args)
timers = defaultdict(Timer)
### --------------------------------------------------------------------------------
### Dataset Training ###
### --------------------------------------------------------------------------------
timers['roidb_training'].tic()
roidb_training, ratio_list_training, ratio_index_training, category_to_id_map, prd_category_to_id_map = combined_roidb_for_training(
cfg.TRAIN.DATASETS)
timers['roidb_training'].toc()
roidb_size_training = len(roidb_training)
logger.info('{:d} training roidb entries'.format(roidb_size_training))
logger.info('Takes %.2f sec(s) to construct training roidb',
timers['roidb_training'].average_time)
batch_size = cfg.NUM_GPUS * cfg.TRAIN.IMS_PER_BATCH
dataset_training = RoiDataLoader(roidb_training,
cfg.MODEL.NUM_CLASSES,
training=True,
dataset=cfg.TRAIN.DATASETS)
dataloader_training = torch.utils.data.DataLoader(
dataset_training,
batch_size=batch_size,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch,
shuffle=True,
drop_last=True)
dataiterator_training = iter(dataloader_training)
### --------------------------------------------------------------------------------
### Dataset Validation ###
### --------------------------------------------------------------------------------
timers['roidb_val'].tic()
roidb_val, ratio_list_val, ratio_index_val, _, _ = combined_roidb_for_training(
cfg.VAL.DATASETS)
timers['roidb_val'].toc()
roidb_size_val = len(roidb_val)
logger.info('{:d} val roidb entries'.format(roidb_size_val))
logger.info('Takes %.2f sec(s) to construct val roidb',
timers['roidb_val'].average_time)
dataset_val = RoiDataLoader(roidb_val,
cfg.MODEL.NUM_CLASSES,
training=False,
dataset=cfg.VAL.DATASETS)
dataloader_val = torch.utils.data.DataLoader(
dataset_val,
batch_size=batch_size,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch,
drop_last=True)
### --------------------------------------------------------------------------------
### Dataset Test ###
### --------------------------------------------------------------------------------
timers['roidb_test'].tic()
roidb_test, ratio_list_test, ratio_index_test, _, _ = combined_roidb_for_training(
cfg.TEST.DATASETS)
timers['roidb_test'].toc()
roidb_size_test = len(roidb_test)
logger.info('{:d} test roidb entries'.format(roidb_size_test))
logger.info('Takes %.2f sec(s) to construct test roidb',
timers['roidb_test'].average_time)
dataset_test = RoiDataLoader(roidb_test,
cfg.MODEL.NUM_CLASSES,
training=False,
dataset=cfg.TEST.DATASETS)
dataloader_test = torch.utils.data.DataLoader(
dataset_test,
batch_size=batch_size,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch,
drop_last=True)
### --------------------------------------------------------------------------------
### Dataset Unseen ###
### --------------------------------------------------------------------------------
if args.dataset == 'vhico':
timers['roidb_unseen'].tic()
roidb_unseen, ratio_list_unseen, ratio_index_unseen, _, _ = combined_roidb_for_training(
cfg.UNSEEN.DATASETS)
timers['roidb_unseen'].toc()
roidb_size_unseen = len(roidb_unseen)
logger.info('{:d} test unseen roidb entries'.format(roidb_size_unseen))
logger.info('Takes %.2f sec(s) to construct test roidb',
timers['roidb_unseen'].average_time)
dataset_unseen = RoiDataLoader(roidb_unseen,
cfg.MODEL.NUM_CLASSES,
training=False,
dataset=cfg.UNSEEN.DATASETS)
dataloader_unseen = torch.utils.data.DataLoader(
dataset_unseen,
batch_size=batch_size,
num_workers=cfg.DATA_LOADER.NUM_THREADS,
collate_fn=collate_minibatch,
drop_last=True)
### --------------------------------------------------------------------------------
### Model ###
### --------------------------------------------------------------------------------
maskRCNN = Generalized_RCNN(category_to_id_map=category_to_id_map,
prd_category_to_id_map=prd_category_to_id_map,
args=args)
if cfg.CUDA:
maskRCNN.cuda()
### --------------------------------------------------------------------------------
### Optimizer ###
# record backbone params, i.e., conv_body and box_head params
### --------------------------------------------------------------------------------
gn_params = []
backbone_bias_params = []
backbone_bias_param_names = []
prd_branch_bias_params = []
prd_branch_bias_param_names = []
backbone_nonbias_params = []
backbone_nonbias_param_names = []
prd_branch_nonbias_params = []
prd_branch_nonbias_param_names = []
for key, value in dict(maskRCNN.named_parameters()).items():
if value.requires_grad:
if 'gn' in key:
gn_params.append(value)
elif 'Conv_Body' in key or 'Box_Head' in key or 'Box_Outs' in key or 'RPN' in key:
if 'bias' in key:
backbone_bias_params.append(value)
backbone_bias_param_names.append(key)
else:
backbone_nonbias_params.append(value)
backbone_nonbias_param_names.append(key)
else:
if 'bias' in key:
prd_branch_bias_params.append(value)
prd_branch_bias_param_names.append(key)
else:
prd_branch_nonbias_params.append(value)
prd_branch_nonbias_param_names.append(key)
# Learning rate of 0 is a dummy value to be set properly at the start of training
params = [{
'params': backbone_nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
}, {
'params':
backbone_bias_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
}, {
'params': prd_branch_nonbias_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY
}, {
'params':
prd_branch_bias_params,
'lr':
0 * (cfg.SOLVER.BIAS_DOUBLE_LR + 1),
'weight_decay':
cfg.SOLVER.WEIGHT_DECAY if cfg.SOLVER.BIAS_WEIGHT_DECAY else 0
}, {
'params': gn_params,
'lr': 0,
'weight_decay': cfg.SOLVER.WEIGHT_DECAY_GN
}]
if cfg.SOLVER.TYPE == "SGD":
optimizer = torch.optim.SGD(params, momentum=cfg.SOLVER.MOMENTUM)
elif cfg.SOLVER.TYPE == "Adam":
optimizer = torch.optim.Adam(params)
### --------------------------------------------------------------------------------
### Load checkpoint
### --------------------------------------------------------------------------------
if args.load_ckpt:
load_name = args.load_ckpt
logging.info("loading checkpoint %s", load_name)
checkpoint = torch.load(load_name,
map_location=lambda storage, loc: storage)
net_utils.load_ckpt(maskRCNN, checkpoint['model'])
print(
'--------------------------------------------------------------------------------'
)
print('loading checkpoint %s' % load_name)
print(
'--------------------------------------------------------------------------------'
)
if args.resume:
print('resume')
args.start_step = checkpoint['step'] + 1
misc_utils.load_optimizer_state_dict(optimizer,
checkpoint['optimizer'])
del checkpoint
torch.cuda.empty_cache()
else:
print('args.load_ckpt', args.load_ckpt)
lr = optimizer.param_groups[2][
'lr'] # lr of non-backbone parameters, for commmand line outputs.
backbone_lr = optimizer.param_groups[0][
'lr'] # lr of backbone parameters, for commmand line outputs.
maskRCNN = mynn.DataParallel(maskRCNN,
cpu_keywords=['im_info', 'roidb'],
minibatch=True)
### --------------------------------------------------------------------------------
### Training Setups ###
### --------------------------------------------------------------------------------
args.run_name = args.out_dir
output_dir = misc_utils.get_output_dir(args, args.out_dir)
args.cfg_filename = os.path.basename(args.cfg_file)
if not args.no_save:
if not os.path.exists(output_dir):
os.makedirs(output_dir)
blob = {'cfg': yaml.dump(cfg), 'args': args}
with open(os.path.join(output_dir, 'config_and_args.pkl'), 'wb') as f:
pickle.dump(blob, f, pickle.HIGHEST_PROTOCOL)
if args.use_tfboard:
from tensorboardX import SummaryWriter
tblogger = SummaryWriter(output_dir)
### --------------------------------------------------------------------------------
### Training Loop ###
### --------------------------------------------------------------------------------
maskRCNN.train()
# Set index for decay steps
decay_steps_ind = None
for i in range(1, len(cfg.SOLVER.STEPS)):
if cfg.SOLVER.STEPS[i] >= args.start_step:
decay_steps_ind = i
break
if decay_steps_ind is None:
decay_steps_ind = len(cfg.SOLVER.STEPS)
training_stats = TrainingStats(
args, args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None, True)
val_stats = ValStats(
args, args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None, False)
test_stats = TestStats(
args, args.disp_interval,
tblogger if args.use_tfboard and not args.no_save else None, False)
best_total_loss = np.inf
best_eval_result = 0
### --------------------------------------------------------------------------------
### EVAL ###
### --------------------------------------------------------------------------------
if cfg.EVAL_SUBSET == 'unseen':
print('testing unseen ...')
is_best, best_eval_result = run_eval(args,
cfg,
maskRCNN,
dataloader_unseen,
step=0,
output_dir=output_dir,
test_stats=test_stats,
best_eval_result=best_eval_result,
eval_subset=cfg.EVAL_SUBSET)
return
elif cfg.EVAL_SUBSET == 'test':
print('testing ...')
is_best, best_eval_result = run_eval(args,
cfg,
maskRCNN,
dataloader_test,
step=0,
output_dir=output_dir,
test_stats=test_stats,
best_eval_result=best_eval_result,
eval_subset=cfg.EVAL_SUBSET)
return
### --------------------------------------------------------------------------------
### TRAIN ###
### --------------------------------------------------------------------------------
try:
logger.info('Training starts !')
step = args.start_step
for step in range(args.start_step, cfg.SOLVER.MAX_ITER):
# Warm up
if step < cfg.SOLVER.WARM_UP_ITERS:
method = cfg.SOLVER.WARM_UP_METHOD
if method == 'constant':
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR
elif method == 'linear':
alpha = step / cfg.SOLVER.WARM_UP_ITERS
warmup_factor = cfg.SOLVER.WARM_UP_FACTOR * (1 -
alpha) + alpha
else:
raise KeyError(
'Unknown SOLVER.WARM_UP_METHOD: {}'.format(method))
lr_new = cfg.SOLVER.BASE_LR * warmup_factor
net_utils.update_learning_rate_rel(optimizer, lr, lr_new)
lr = optimizer.param_groups[2]['lr']
backbone_lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
elif step == cfg.SOLVER.WARM_UP_ITERS:
net_utils.update_learning_rate_rel(optimizer, lr,
cfg.SOLVER.BASE_LR)
lr = optimizer.param_groups[2]['lr']
backbone_lr = optimizer.param_groups[0]['lr']
assert lr == cfg.SOLVER.BASE_LR
# Learning rate decay
if decay_steps_ind < len(cfg.SOLVER.STEPS) and \
step == cfg.SOLVER.STEPS[decay_steps_ind]:
logger.info('Decay the learning on step %d', step)
lr_new = lr * cfg.SOLVER.GAMMA
net_utils.update_learning_rate_rel(optimizer, lr, lr_new)
lr = optimizer.param_groups[2]['lr']
backbone_lr = optimizer.param_groups[0]['lr']
assert lr == lr_new
decay_steps_ind += 1
#########################################################################################################################
## train
#########################################################################################################################
training_stats.IterTic()
optimizer.zero_grad()
for inner_iter in range(args.iter_size):
try:
input_data = next(dataiterator_training)
except StopIteration:
print('recurrence data loader')
dataiterator_training = iter(dataloader_training)
input_data = next(dataiterator_training)
for key in input_data:
if key != 'roidb': # roidb is a list of ndarrays with inconsistent length
input_data[key] = list(map(Variable, input_data[key]))
net_outputs = maskRCNN(**input_data)
training_stats.UpdateIterStats(net_outputs['gt_label'],
inner_iter)
loss = net_outputs['gt_label']['total_loss']
loss.backward()
optimizer.step()
training_stats.IterToc()
training_stats.LogIterStats(step, lr, backbone_lr)
if (step + 1) % cfg.SAVE_MODEL_ITER == 0:
save_ckpt(output_dir, args, step, batch_size, maskRCNN,
optimizer, False, best_total_loss)
# ---- Training ends ----
save_ckpt(output_dir, args, step, batch_size, maskRCNN, optimizer,
False, best_total_loss)
except (RuntimeError, KeyboardInterrupt):
del dataiterator_training
logger.info('Save ckpt on exception ...')
save_ckpt(output_dir, args, step, batch_size, maskRCNN, optimizer,
False, best_total_loss)
logger.info('Save ckpt done.')
stack_trace = traceback.format_exc()
print(stack_trace)
finally:
if args.use_tfboard and not args.no_save:
tblogger.close()
self.image_channels = 3
self.image_size = 84
super(fc100, self).__init__(options)
class miniImagenet(FewShotDataset):
def __init__(self, options):
self.dataset_root = abspath(
getattr_or_default(options, 'dataset_root',
'/home/aditya.bharti/mini-imagenet'))
self.mean = (0.4802, 0.4481, 0.3975)
self.std = (0.2302, 0.2265, 0.2262)
self.image_channels = 3
self.image_size = 84
super(miniImagenet, self).__init__(options)
if __name__ == "__main__":
from arguments import parse_args
opts = parse_args()
for dataset in [fc100, miniImagenet, cifar100fs, Imagenet1k]:
d = dataset(opts)
print(str(d), "loaded")
if (hasattr(d, "trainval_set")):
print(str(d), "has trainval")
else:
print(str(d), "couldn't find trainval")
tb_writer = SummaryWriter('./log/tb_{0}'.format(args.log_name))
# Set seed
set_seed(args.seed, cudnn=args.make_deterministic)
# Set sampler
sampler = BatchSampler(args, log)
# Set policy
policy = CaviaMLPPolicy(
input_size=int(np.prod(sampler.observation_space.shape)),
output_size=int(np.prod(sampler.action_space.shape)),
hidden_sizes=(args.hidden_size, ) * args.num_layers,
num_context_params=args.num_context_params,
device=args.device)
# Initialise baseline
baseline = LinearFeatureBaseline(
int(np.prod(sampler.observation_space.shape)))
# Initialise meta-learner
metalearner = MetaLearner(sampler, policy, baseline, args, tb_writer)
# Begin train
train(sampler, metalearner, args, log, tb_writer)
if __name__ == '__main__':
args = parse_args()
main(args)
#!/usr/bin/env python3
import seacow
import arguments
arguments.parser.add_argument('--host', '-x', default='kxgames.net')
arguments.parser.add_argument('--port', '-p', default=53351, type=int)
arguments.parse_args()
game = seacow.ClientLoop('Client', arguments.host, arguments.port)
game.play()
eval_correct_characters += correct
bar.update(batch_id + 1,
exact=[("eval score", eval_score / eval_total_instances)
])
logger.info('{} dataset accuracy: {}'.format(
dataset_name, eval_correct_characters / eval_total_characters))
def save_model(model):
filename = os.path.join(args.output_dir, 'model.pt')
utils.ensure_folder(filename)
torch.save(model, filename)
args = arguments.parse_args()
logger = init_logger()
logger.info(args)
def main():
training_data, dev_data, test_data, char_to_ix, tag_to_ix = load_datasets()
START_TAG = "<START>"
STOP_TAG = "<STOP>"
char_to_ix, [START_CHAR_ID,
STOP_CHAR_ID] = complete_collection(char_to_ix,
[START_TAG, STOP_TAG])
tag_to_ix, [START_TAG_ID,
STOP_TAG_ID] = complete_collection(tag_to_ix,
[START_TAG, STOP_TAG])
def main():
args = parse_args()
print(args)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
set_seed(args.seed, args.n_gpu)
# Prepare GLUE tasks
processors = {}
output_modes = {}
label_lists = {}
num_label_list = {}
for key in args.task_params:
print(processors)
processors[key] = glue_processors[key]()
output_modes[key] = glue_output_modes[key]
label_lists[key] = processors[key].get_labels()
num_label_list[key] = len(label_lists[key])
# Configs
configs = {}
for key in args.task_params:
configs[key] = AutoConfig.from_pretrained(
args.model_type,
# args.config_name if args.config_name else args.model_name_or_path,
num_labels=num_label_list[key],
finetuning_task=key,
cache_dir=None,
)
# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(
args.model_type,
do_lower_case=args.do_lower_case,
cache_dir=None,
)
# Continual Learning
n = len(configs)
accuracy_matrix = np.zeros((n, n))
transfer_matrix = np.zeros((n, n))
tasks = list(args.task_params.keys())
models = []
# Model
for key in args.task_params:
if args.n_gpu <= 1:
if args.mas:
models.append(
(key,
shared_model(
AutoModelForSequenceClassification.from_pretrained(
args.model_type, config=configs[key]))))
else:
models.append(
(key,
AutoModelForSequenceClassification.from_pretrained(
args.model_type, config=configs[key])))
else:
if args.mas:
models.append((
key,
shared_model(
torch.nn.DataParallel(
AutoModelForSequenceClassification.from_pretrained(
args.model_type, config=configs[key])))))
else:
models.append(
(key,
torch.nn.DataParallel(
AutoModelForSequenceClassification.from_pretrained(
args.model_type, config=configs[key]))))
for i in range(n):
models[i][1].to(args.device)
save_model(args, i, models[i][1])
if args.cuda == torch.device('cuda') and torch.cuda.device_count() > 1:
for key in models:
models[key] = torch.nn.DataParallel(models[key])
consolidate_fisher = {}
consolidate_mean = {}
for i in range(len(configs)):
if (i > 0):
part_name = partial_name(args)
# Always load the BERT parameters of previous model
if args.mas:
models[i][1].tmodel.load_state_dict(torch.load(
os.path.join(args.output_dir, part_name,
"bert_paramters_" + str(i - 1) + ".pt")),
strict=False)
models[i][1].reg_params = models[i - 1][1].reg_params
else:
models[i][1].load_state_dict(torch.load(
os.path.join(args.output_dir, part_name,
"bert_paramters_" + str(i - 1) + ".pt")),
strict=False)
new_args = convert_dict(args.task_params[tasks[i]], args)
train_dataset = load_and_cache_examples(args,
tasks[i],
tokenizer,
evaluate=False)
# --- No randomness till here ---
global_step, tr_loss, accuracy_matrix, model = train(
new_args, train_dataset, tasks[i], tasks, models[i][1], i,
tokenizer, accuracy_matrix, consolidate_fisher, consolidate_mean)
logger.info(" global_step = %s, average loss = %s", global_step,
tr_loss)
if args.ewc:
if (i == (len(configs) - 1)):
print(
"\n ***** Last Task : No need to estimate fisher matrix ***** \n"
)
break
else:
print()
print(
"***** Estimating Diagonals of Fisher Information Matrix *****"
)
print()
consolidate_fisher, consolidate_mean = estimate_fisher_mean(
new_args, train_dataset, model, tasks[i],
consolidate_fisher, consolidate_mean,
args.per_gpu_batch_size)
print("\n ***** Accuracy Matrix *****\n")
print(accuracy_matrix)
print("\n***** Transfer Matrix *****")
print(
"Future Transfer => Upper Triangular Matrix || Backward Transfer => Lower Triangular Matrix\n"
)
for i in range(n):
for j in range(n):
transfer_matrix[j][
i] = accuracy_matrix[j][i] - accuracy_matrix[i][i]
print(transfer_matrix)
import file_handler as fh
import evaluation_handler as eh
import learn_handler as lh
from vectorizer import Vectorizer
if __name__ == '__main__':
# Load Data
# senti_dir = './KOSAC_sample/'
# # {'sentences':sentences, 'tokens':tokens, 'polarity':polarity}
# corpus = fh.getSentiCorpus(senti_dir)
# contents = corpus['sentences']
comments = fh.getComments()
contents = comments['sentences']
# Vectorizing
# vec = Vectorizer()
# tokenized_contents = vec.tokenizing(contents)
# # sg=1(skip-gram), 0(CBOW)
# model_path = 'models/word2vec_ko.model'
# vectorized_contents = vec.vectorize(model_path, contents, dims=100)
#
# fh.saveVectorizedContents(vectorized_contents)
vectorized_contents = fh.loadVectorizedContents()
# Train ML and Evaluation
args = arguments.parse_args('Bidirectional_LSTM')
eh.evaluations(vectorized_contents, comments['scores'],
lh.Bidirectional_LSTM(args))
from plot import plot_acc, plot_gan_losses, plot_confusion_matrix
from arguments import parse_args
import random
import torch
import torch.backends.cudnn as cudnn
import os
import numpy as np
from dataset import FeaturesCls, FeaturesGAN
from train_cls import TrainCls
from train_gan import TrainGAN
from generate import load_unseen_att, load_all_att
from mmdetection.splits import get_unseen_class_labels
opt = parse_args()
try:
os.makedirs(opt.outname)
except OSError:
pass
if opt.manualSeed is None:
opt.manualSeed = random.randint(1, 10000)
for arg in vars(opt): print(f"###################### {arg}: {getattr(opt, arg)}")
print("Random Seed: ", opt.manualSeed)
random.seed(opt.manualSeed)