def async_build_examples(
self, data_type: str,
dials: List[Tuple[str, dict]]) -> Tuple[list, list]:
"""Use multiprocessing to process raw dialogue data.
Args:
data_type: train, dev or test
dials: raw dialogues data
Returns:
new examples by all processes
"""
neg_examples = Manager().list()
pos_examples = Manager().list()
dials4single_process = (len(dials) -
1) // self.config['num_processes'] + 1
print(f'Single process have {dials4single_process} dials ...')
pool = Pool(self.config['num_processes'])
for i in range(self.config['num_processes']):
pool.apply_async(func=self.iter_dials,
args=(dials[dials4single_process *
i:dials4single_process * (i + 1)],
data_type, pos_examples, neg_examples, i))
pool.close()
pool.join()
pos_examples = list(pos_examples)
neg_examples = list(neg_examples)
return neg_examples, pos_examples
def load_async(pool: Pool,
fn: Callable,
*args,
callback: Callable = None,
**kwargs) -> Any:
"""
Load data asynchronously and serialize data via dill
Args:
pool: multiprocessing pool to use for :func:`apply_async`
fn: function to load a single sample
*args: positional arguments to dump with dill
callback: optional callback. defaults to None.
**kwargs: keyword arguments to dump with dill
Returns:
Any: reference to obtain data with :func:`get`
"""
if not DILL_AVAILABLE:
raise RuntimeError('dill is not installed. For async loading '
'please install it')
payload = dill.dumps((fn, args, kwargs))
return pool.apply_async(dill_helper, (payload, ), callback=callback)
def propagate(nnf, feat_A, feat_AP, feat_B, feat_BP, patch_size, iters=2, rand_search_radius=200):
print("\tpatch_size:{}; num_iters:{}; rand_search_radius:{}".format(patch_size, iters, rand_search_radius))
nnd = np.zeros(nnf.shape[:2])
A_size = feat_A.shape[:2]
B_size = feat_B.shape[:2]
for ay in range(A_size[0]):
for ax in range(A_size[1]):
by, bx = nnf[ay, ax]
nnd[ay, ax] = cal_dist(ay, ax, by, bx, feat_A, feat_AP, feat_B, feat_BP, A_size, B_size, patch_size)
manager = mp.Manager()
q = manager.Queue(A_size[1] * A_size[0])
cpus = min(mp.cpu_count(), A_size[0] // 20 + 1)
for i in range(iters):
p = Pool(cpus)
ay_start = 0
while ay_start < A_size[0]:
ax_start = 0
while ax_start < A_size[1]:
p.apply_async(pixelmatch, args=(q, ax_start, ay_start,
cpus,
nnf, nnd,
A_size, B_size,
feat_A, feat_AP,
feat_B, feat_BP,
patch_size,
rand_search_radius,))
ax_start += A_size[1] // cpus + 1
ay_start += A_size[0] // cpus + 1
p.close()
p.join()
while not q.empty():
ax, ay, xbest, ybest, dbest = q.get()
nnf[ay, ax] = np.array([ybest, xbest])
nnd[ay, ax] = dbest
return nnf, nnd
def match(self):
pool = Pool(self.num_processes)
results = []
total = len(self.pairs_to_match)
counter = 0
matches_pairs_fname = os.path.join(
self.input_dir, "matches_pairs" + self.suffix + ".txt")
mp_file = open(matches_pairs_fname, 'w')
for pair in tqdm(self.pairs_to_match):
orig_name1 = self.image_names[pair[0]]
orig_name2 = self.image_names[pair[1]]
name1 = os.path.join(self.input_dir, orig_name1)
name2 = os.path.join(self.input_dir, orig_name2)
image_id1 = self.image_ids[pair[0]]
image_id2 = self.image_ids[pair[1]]
if self.cuda:
# use single process
data = self.matchPairCached(orig_name1, orig_name2, image_id1,
image_id2, self.recompute,
self.cuda)
self.db.add_matches(data[0], data[1], data[2])
mp_file.write(data[3] + " " + data[4] + "\n")
counter += 1
if counter % 50000 == 0:
counter = 0
self.db.commit()
# print("Matched", counter, "out of", total)
else:
# use multiple processes
res = pool.apply_async(self.matchPairCached,
(orig_name1, orig_name2, image_id1,
image_id2, self.recompute, self.cuda))
results.append(res)
if not self.cuda:
# if more processes were used, collect the results
counter = 0
for res in tqdm(results):
data = res.get()
if data is None:
continue
self.db.add_matches(data[0], data[1], data[2])
name1 = data[3]
name2 = data[4]
mp_file.write(name1 + " " + name2 + "\n")
counter += 1
if counter % 50000 == 0:
counter = 0
self.db.commit()
# be sure to commit everything at the end
self.db.commit()
mp_file.close()
def save(self):
try:
mp.set_start_method('spawn')
except RuntimeError:
pass
pool = Pool(processes=4)
self.labels = []
item = 0
for label_name in label_map.keys():
images = self._listdir(os.path.join(self.path, label_name))
for i in range(len(images)):
rows, cols = [], []
files = glob.glob(
os.path.join(self.path, label_name, images[i],
str(self.magnify), '*.jpeg'))
for file in files:
filename = os.path.basename(file)
nums = filename.split('_')
row, col = int(nums[0]), int(nums[1])
rows.append(row)
cols.append(col)
num_row = max(rows) - min(rows) + 1
num_col = max(cols) - min(cols) + 1
patches = np.chararray((num_row, num_col), itemsize=1024)
for file in files:
filename = os.path.basename(file)
nums = filename.split('_')
row, col = int(nums[0]), int(nums[1])
patches[row - min(rows), col - min(cols)] = file
self.labels.append(label_map[label_name])
# Save feature vector
pool.apply_async(self.doit,
args=(item, patches, num_row, num_col),
error_callback=self.print_error)
item += 1
# Save labels
torch.save(self.labels, self._get_label_file())
pool.close()
pool.join()
print('done')
def mul_infer(args):
setuplogging()
set_start_method('spawn', force=True)
root_data_dir = os.path.join(args.root_data_dir, 'testdata')
checkpoint = torch.load(os.path.join(args.model_dir, args.load_ckpt_name),
map_location=torch.device('cpu'))
subcategory_dict = checkpoint['subcategory_dict']
category_dict = checkpoint['category_dict']
logging.info('load ckpt: {}'.format(args.load_ckpt_name))
check_preprocess_result(args,
root_data_dir,
mode='test',
category=category_dict,
subcategory=subcategory_dict)
logging.info('finish the preprocess of docfeatures')
docid_features, category_dict, subcategory_dict = read_news(
args, root_data_dir)
news_index = {}
news_feature = []
cnt = 0
for k, v in docid_features.items():
news_index[k] = cnt
news_feature.append(v)
cnt += 1
news_num = len(news_feature)
logging.info('news_num:{}'.format(news_num))
pool = Pool(processes=args.world_size)
results = []
sigle_size = news_num // args.world_size
for rank in range(args.world_size):
start = sigle_size * rank
end = sigle_size * (rank + 1)
if rank == args.world_size - 1:
end = news_num
local_features = news_feature[start:end]
result = pool.apply_async(sigle_process_infer,
args=(rank, local_features, checkpoint,
args))
results.append(result)
pool.close()
pool.join()
results = [x.get() for x in results]
news_vecs = np.concatenate(results, 0)
return news_index, news_vecs
def fit(self, train_data, train_label):
a = time()
pool = Pool(16)
results = []
for i in range(self.round):
print(i)
bag_index = np.random.choice(np.arange(train_label.shape[0]),
train_label.shape[0])
data = train_data[bag_index]
label = train_label[bag_index]
results.append(
pool.apply_async(self.parallel_fit,
args=(self.clf[i], data, label)))
pool.close()
pool.join()
for i, result in enumerate(results):
self.clf[i] = result.get()
print('Class %d cost %.1f seconds' % (i, time() - a))
def train(self,
data,
labels,
val_data=None,
val_labels=None,
warm_start=False):
"""
:param data:
:param labels:
:param val_data:
:param val_labels:
:param warm_start:
:return:
"""
# initialize variable
data = torch.from_numpy(data).float()
labels = torch.from_numpy(labels).char()
if val_data is None:
train_data, val_data = data, data
train_labels, val_labels = labels, labels
else:
train_data = data
train_labels = labels
val_data = torch.from_numpy(val_data).float()
val_labels = torch.from_numpy(val_labels).char()
orig_cols = train_data.size(1)
# counting class and get their index
self.plus_row_index = []
self.minus_row_index = []
self.orig_minus = 0
self.orig_plus = 0
for idx, value in enumerate(train_labels):
if value == 1:
self.orig_plus += 1
self.plus_row_index.append(idx)
else:
self.orig_minus += 1
self.minus_row_index.append(idx)
# balanced pick rows and cols
plus = max(2, int(self.orig_plus * self.nrows))
minus = max(2, int(self.orig_minus * self.nrows))
num_cols = max(min(5, orig_cols), int(self.nfeatures * orig_cols))
# initialize up triangle matrix and reference index
rows_sum = plus + minus
self.yp = torch.ones((self.width, rows_sum), dtype=torch.int8)
self.ref_full_index1 = torch.repeat_interleave(torch.arange(
self.updated_features * self.step.shape[0]).view((-1, 1)),
rows_sum,
dim=1)
self.ref_full_index2 = torch.repeat_interleave(torch.arange(
self.hidden_nodes * self.step.shape[0]).view((-1, 1)),
rows_sum,
dim=1)
# multi-process
c = self.round // self.num_gpus
for r in range(c + 1):
pool = Pool(self.n_jobs)
results = []
for t in range(min(self.n_jobs, self.round - r * self.num_gpus)):
if warm_start and self.w_index != []:
column_indices = self.w_index[r * self.num_gpus + t]
w1 = self.w1[:, :, r * self.num_gpus + t]
w2 = self.w2[:, r * self.num_gpus + t]
else:
column_indices = np.random.choice(np.arange(orig_cols),
num_cols,
replace=False)
# column_indices = np.arange(orig_cols)
self.w_index.append(column_indices)
results.append(
pool.apply_async(self.single_run,
args=(train_data, train_labels, plus,
minus, val_data, val_labels,
column_indices, t % self.num_gpus)))
pool.close()
pool.join()
df = pd.DataFrame(columns=[])
for i, result in enumerate(results):
temp_w1, temp_b1, temp_w2, temp_b2, temp_obj, uba, ba = result.get(
)
df['vote %d imbalanced acc' % i] = uba
df['vote %d balanced acc' % i] = ba
# temp_w1, temp_b1, temp_w2, temp_b2, temp_obj = self.single_run(train_data, train_labels, plus, minus, val_data, val_labels, w1, w2, column_indices, r % self.num_gpus)
if warm_start:
self.w1[:, :, i] = temp_w1
self.w2[:, i] = temp_w2
self.b1[:, i] = temp_b1
self.b2[i] = temp_b2
self.obj[i] = temp_obj
else:
self.w1.append(temp_w1)
self.w2.append(temp_w2)
self.b1.append(temp_b1)
self.b2.append(temp_b2)
self.obj.append(temp_obj)
del pool, results
df.to_csv('v15.csv', index=False)
if warm_start is False:
self.w1 = torch.stack(self.w1, dim=2)
self.w2 = torch.stack(self.w2, dim=1)
self.b1 = torch.stack(self.b1, dim=1)
self.b2 = torch.Tensor(self.b2)
self.obj = torch.Tensor(self.obj)
best_index = self.obj.argmax()
self.best_acc = self.obj[best_index]
self.best_w1 = self.w1[:, :, best_index]
self.best_w2 = self.w2[:, best_index]
self.best_b1 = self.b1[:, best_index]
self.best_b2 = self.b2[best_index]
self.best_w_index = self.w_index[best_index]
return
def detection_by_tracking(
frame_dir,
json_file,
tracker_model,
detection_threshold=0.9,
tracking_threshold=0.9,
save_json_file="data/demo_tracking/detection_by_tracking.tracking_json",
offset=0,
low=None,
high=None,
step=1,
parallel=False,
multithreading=False):
# Load annotations
data = json.load(open(json_file, "r"))
annotations = dict()
for annotation in data['annotations']:
if annotation['image_id'] in annotations:
annotations[annotation['image_id']] += [annotation]
else:
annotations[annotation['image_id']] = [annotation]
# Load frames
frame_files = general_utils.get_all_files(frame_dir,
keep_dir=True,
sort=True)
num_frame = len(frame_files)
tracking_data = dict()
tracking_data["images"] = data["images"]
tracking_data["categories"] = data["categories"]
tracking_data["annotations"] = list()
if low is None:
low = -int(1e9)
if high is None:
high = int(1e9)
start = time.time()
last_count = 0
# Set up parallel processing
if parallel:
mp.set_start_method('spawn', force=True)
mp.set_sharing_strategy('file_system')
pool = Pool()
else:
pool = None
results = [None for _ in range(num_frame)]
# Set up multithreading processing
if multithreading:
executor = ThreadPoolExecutor()
else:
executor = None
# Loop over frames
for frame_id in range(num_frame):
# Align id
frame_id += offset
num_box = len(annotations[frame_id])
# Count boxes with high confidence
count = 0
for box_id in range(num_box):
score = annotations[frame_id][box_id]["score"]
if score > detection_threshold:
count += 1
# If this frame has more boxes, track from it for certain; else check skip criteria
if count <= last_count:
last_count = count
# Skip frame
if frame_id % step != 0:
continue
else:
last_count = count
print("Process frame ", frame_id)
forward_tracker = build_tracker(tracker_model)
backward_tracker = build_tracker(tracker_model)
# Loop over detection boxes
for box_id in range(num_box):
# print("=> Process box ", box_id)
# Filter by detection score
score = annotations[frame_id][box_id]["score"]
if score < detection_threshold:
# print("==> Skip")
continue
if multithreading:
print(
f"---> Multithread tracking for box {box_id} frame {frame_id}"
)
executor.submit(single_box_in_single_frame_tracking,
(frame_files, frame_id, box_id, annotations,
tracking_threshold, forward_tracker,
backward_tracker, offset, low, high))
if parallel:
print(
f"---> Parallel tracking for box {box_id} frame {frame_id}"
)
results[frame_id - offset] = pool.apply_async(
single_box_in_single_frame_tracking, [
frame_files, frame_id, box_id, annotations,
tracking_threshold, forward_tracker, backward_tracker,
offset, low, high
])
if not multithreading and not parallel:
tracking_data[
"annotations"] += single_box_in_single_frame_tracking(
frame_files, frame_id, box_id, annotations,
tracking_threshold, forward_tracker, backward_tracker,
offset, low, high)
for result in results:
if result is not None:
tracking_data["annotations"] += result.get()
end = time.time()
print(f"Total time: {(end - start)} s")
with open(save_json_file, "w") as outfile:
json.dump(tracking_data, outfile)
def main():
opt = parse_args()
device = 'cuda' if torch.cuda.is_available() else 'cpu'
random.seed(opt.seed)
numpy.random.seed(opt.seed)
torch.manual_seed(opt.seed)
data_path = 'traffic-data/state-action-cost/data_i80_v0'
dataloader = DataLoader(None, opt, 'i80')
(
forward_model,
value_function,
policy_network_il,
policy_network_mper,
data_stats
) = load_models(opt, data_path, device)
splits = torch.load(path.join(data_path, 'splits.pth'))
if opt.u_reg > 0.0:
forward_model.train()
forward_model.opt.u_hinge = opt.u_hinge
if hasattr(forward_model, 'value_function'):
forward_model.value_function.train()
planning.estimate_uncertainty_stats(
forward_model, dataloader, n_batches=50, npred=opt.npred)
gym.envs.registration.register(
id='I-80-v1',
entry_point='map_i80_ctrl:ControlledI80',
kwargs=dict(
fps=10,
nb_states=opt.ncond,
display=False,
delta_t=0.1,
store_simulator_video=opt.save_sim_video,
show_frame_count=False,
)
)
print('Building the environment (loading data, if any)')
env_names = {
'i80': 'I-80-v1',
}
env = gym.make(env_names[opt.map])
plan_file = build_plan_file_name(opt)
print(f'[saving to {path.join(opt.save_dir, plan_file)}]')
# different performance metrics
time_travelled, distance_travelled, road_completed = [], [], []
collided, offscreen = [], []
# values saved for later inspection
action_sequences, state_sequences, cost_sequences = [], [], []
image_sequences = []
writer = utils.create_tensorboard_writer(opt)
n_test = len(splits['test_indx'])
set_start_method('spawn')
pool = Pool(opt.num_processes)
async_results = []
time_started = time.time()
total_images = 0
for j in range(n_test):
# print(type(splits), len(splits['test_indx']), splits['test_indx'].shape, list(dataloader.car_sizes.keys())[0:5], list(dataloader.car_sizes[list(dataloader.car_sizes.keys())[0]].keys())[0:5],dataloader.car_sizes[list(dataloader.car_sizes.keys())[0]][list(dataloader.car_sizes[list(dataloader.car_sizes.keys())[0]].keys())[0]])
car_path = dataloader.ids[splits['test_indx'][j]]
timeslot, car_id = utils.parse_car_path(car_path)
car_sizes = torch.tensor(
dataloader.car_sizes[sorted(list(dataloader.car_sizes.keys()))[
timeslot]][car_id]
)[None, :]
async_results.append(
pool.apply_async(
process_one_episode, (
opt,
env,
car_path,
forward_model,
policy_network_il,
data_stats,
plan_file,
j,
car_sizes
)
)
)
for j in range(n_test):
simulation_result = async_results[j].get()
time_travelled.append(simulation_result.time_travelled)
distance_travelled.append(simulation_result.distance_travelled)
road_completed.append(simulation_result.road_completed)
action_sequences.append(torch.from_numpy(
simulation_result.action_sequence))
state_sequences.append(torch.from_numpy(
simulation_result.state_sequence))
# image_sequences.append(torch.from_numpy(
# simulation_result.image_sequence))
cost_sequences.append(simulation_result.cost_sequence)
total_images += time_travelled[-1]
collided.append(simulation_result.has_collided)
offscreen.append(simulation_result.off_screen)
log_string = ' | '.join((
f'ep: {j + 1:3d}/{n_test}',
f'time: {time_travelled[-1]}',
f'distance: {distance_travelled[-1]:.0f}',
f'success: {road_completed[-1]:d}',
f'mean time: {torch.Tensor(time_travelled).mean():.0f}',
f'mean distance: {torch.Tensor(distance_travelled).mean():.0f}',
f'mean success: {torch.Tensor(road_completed).mean():.3f}',
))
print(log_string)
utils.log(path.join(opt.save_dir, f'{plan_file}.log'), log_string)
if writer is not None:
# writer.add_video(
# f'Video/success={simulation_result.road_completed:d}_{j}',
# simulation_result.images.unsqueeze(0),
# j
# )
writer.add_scalar('ByEpisode/Success',
simulation_result.road_completed, j)
writer.add_scalar('ByEpisode/Collision',
simulation_result.has_collided, j)
writer.add_scalar('ByEpisode/OffScreen',
simulation_result.off_screen, j)
writer.add_scalar('ByEpisode/Distance',
simulation_result.distance_travelled, j)
pool.close()
pool.join()
diff_time = time.time() - time_started
print('avg time travelled per second is', total_images / diff_time)
torch.save({"road_completed" : road_completed, "collided": collided, "offscreen": offscreen}, path.join(opt.save_dir, f'{plan_file}.others'))
torch.save(action_sequences, path.join(
opt.save_dir, f'{plan_file}.actions'))
torch.save(state_sequences, path.join(opt.save_dir, f'{plan_file}.states'))
# torch.save(image_sequences, path.join(opt.save_dir, f'{plan_file}.images'))
torch.save(cost_sequences, path.join(opt.save_dir, f'{plan_file}.costs'))
if writer is not None:
writer.close()
with open(os.path.join(
output_root, "%s_list.Score" % tgtname), "a") as F:
F.write(score_output)
pbar3.update()
return
for i in range(datasize):
tplname = template_name_list[i]
observations = torch.load(
os.path.join(obs_path, '%s-%s.pth' % (tplname, tgt['name'])),
pickle_module=pickle)
observations = observations.float()
observations = torch.mul(observations, Node_Weight)
pool.apply_async(compute_alignment,
args=(tplname + tpl_type, observations,
transitions, pair_distance,
disc_method.tolist(),
ADMM_ITERATION, edge_type, Node_Weight),
callback=getoutput)
pool.close()
pool.join()
pbar3.close()
print('time: %.3f s' % (time.time()-start))
print('finish admm algorithm')
print('start generate output..')
print('sort by index: %d' % sort_col)
sortoutput(tgt['name'],
os.path.join(output_root, tgt['name'] + '_list.Score'),
args.k, sort_col)
print("finish %d alignment generation and save them in %s" %
(datasize, output_root))
def train(self,
data,
labels,
val_data=None,
val_labels=None,
warm_start=False):
"""
:param data:
:param labels:
:param val_data:
:param val_labels:
:param warm_start:
:return:
"""
# print('start train')
# initialize variable
data = torch.from_numpy(data).float()
labels = torch.from_numpy(labels).char()
if val_data is None:
train_data, val_data = data, data
train_labels, val_labels = labels, labels
else:
train_data = data
train_labels = labels
val_data = torch.from_numpy(val_data).float()
val_labels = torch.from_numpy(val_labels).char()
orig_cols = train_data.size(1)
# counting class and get their index
self.plus_row_index = []
self.minus_row_index = []
self.orig_minus = 0
self.orig_plus = 0
for idx, value in enumerate(train_labels):
if value == 1:
self.orig_plus += 1
self.plus_row_index.append(idx)
else:
self.orig_minus += 1
self.minus_row_index.append(idx)
# balanced pick rows and cols
plus = max(2, int(self.orig_plus * self.nrows))
minus = max(2, int(self.orig_minus * self.nrows))
num_cols = max(min(5, orig_cols), int(self.nfeatures * orig_cols))
# initialize up triangle matrix and reference index
rows_sum = plus + minus
if self.adv_train:
rows_sum = rows_sum * 2
# plus = plus * 2
# minus = minus * 2
self.yp = torch.ones((self.width, rows_sum), dtype=torch.int8)
self.ref_full_index = torch.repeat_interleave(torch.arange(
self.updated_features * self.step.shape[0]).view((-1, 1)),
rows_sum,
dim=1)
# multi-process
c = self.round // self.num_gpus
results = []
logs = {}
# print('enter pool')
for r in range(c + 1):
pool = Pool(self.n_jobs)
results = []
for t in range(min(self.n_jobs, self.round - r * self.num_gpus)):
if warm_start and self.w_index != []:
column_indices = self.w_index[r * self.num_gpus + t]
w1 = self.w1[:, :, r * self.num_gpus + t]
w2 = self.w2[:, r * self.num_gpus + t]
else:
column_indices = np.random.choice(np.arange(orig_cols),
num_cols,
replace=False)
# column_indices = np.arange(orig_cols)
self.w_index.append(column_indices)
results.append(
pool.apply_async(self.single_run,
args=(train_data, train_labels, plus,
minus, val_data, val_labels,
column_indices, t % self.num_gpus)))
pool.close()
pool.join()
for i, result in enumerate(results):
temp_w, temp_b, temp_obj = result.get()
# logs['vote%d_train' % i] = train_log
# logs['vote%d_test' % i] = test_log
if warm_start:
self.w[:, i] = temp_w
self.b[:, i] = temp_b
self.obj[i] = temp_obj
else:
self.w.append(temp_w.view((-1, 1)))
self.b.append(temp_b.view((1, 1)))
self.obj.append(temp_obj)
del pool, results
if warm_start is False:
self.w = torch.cat(self.w, dim=1)
self.b = torch.cat(self.b, dim=1)
self.obj = torch.Tensor(self.obj)
best_index = self.obj.argmax()
self.best_acc = self.obj[best_index]
self.best_w = self.w[:, best_index]
self.best_b = self.b[:, best_index]
self.best_w_index = self.w_index[best_index]
del self.yp, self.ref_full_index
return
def learn_selfplay_client(self):
"""
Process that continuously generates self-play data
"""
manager = mp.Manager()
sharedQ = manager.Queue()
statedict_name = "Default"
# Create num_selfplay_procs queues for sending nn eval results to selfplay procs.
queues = []
for j in range(self.args.num_selfplay_procs):
queues.append(manager.Queue())
# Create num_gpu_procs queues for sending state_dict update info to nn procs.
nn_update_pipes1 = []
nn_update_pipes2 = []
for j in range(self.args.num_gpu_procs):
c1, c2 = mp.Pipe()
nn_update_pipes1.append(c1)
nn_update_pipes2.append(c2)
# Create num_gpu_procs nnProcess
nnProcs = []
for j in range(self.args.num_gpu_procs):
# Run nnProc
nnProc = mp.Process(
target=JanggiCoach.nnProcess,
args=[(self.game, nn_update_pipes1[j], sharedQ,
self.args.gpus_to_use[j % len(self.args.gpus_to_use)],
queues, self.args.checkpoint_folder)])
nnProc.daemon = True
nnProc.start()
nnProcs.append(nnProc)
# Create a queue for receiving info of finished jobs
nextSelfplayQ = manager.Queue()
# Create self-play process pool
selfplayPool = Pool(self.args.num_selfplay_procs)
# Run the first num_selfplay_procs process
ibs = pickle.loads(
requests.get(url=self.args.request_base_url + "/getIBS").content)
for j in range(self.args.num_selfplay_procs):
selfplayPool.apply_async(
JanggiCoach.executeEpisode,
[(Game(self.game.c1, self.game.c2, mode=ibs), self.args,
sharedQ, queues[j], j, nextSelfplayQ, None)])
# Continuously generate self-plays
while True:
# Check for any network updates
new_sd = pickle.loads(
requests.get(url=self.args.request_base_url +
"/getSD").content)
if statedict_name != new_sd:
statedict_name = new_sd
sharedStateDictFile = JanggiCoach.getSharedStateDictFile(
self.args.remote_checkpoint_folder)
if (self.args.scp_base_url != None):
JanggiCoach.checkpointSCP(
self.args.scp_base_url + ":" + sharedStateDictFile,
sharedStateDictFile)
for q in nn_update_pipes2:
q.send(statedict_name)
q.recv()
log.info('Alerted the nn procs to update the network')
# Wait for a selfplay result
is_selfplay, q_data = nextSelfplayQ.get()
if is_selfplay:
data, finished_id = q_data
self.selfPlaysPlayed += 1
log.info(
str(self.selfPlaysPlayed) +
' selfplay games played. Data length = ' + str(len(data)))
requests.post(url=self.args.request_base_url + "/postData",
data=pickle.dumps(data))
else:
checkpoint, is_rp, did_win = q_data
log.info("Evaluated (" + str(checkpoint) + ", " + str(is_rp) +
", " + str(did_win) + ")")
requests.post(url=self.args.request_base_url +
"/uploadEvalRes",
data=pickle.dumps((checkpoint, is_rp, did_win)))
# Run new selfplay
ibs = pickle.loads(
requests.get(url=self.args.request_base_url +
"/getIBS").content)
next_game = pickle.loads(
requests.get(url=self.args.request_base_url +
"/getNextGame").content)
if next_game == None:
selfplayPool.apply_async(
JanggiCoach.executeEpisode,
[(Game(self.game.c1, self.game.c2,
mode=ibs), self.args, sharedQ, queues[finished_id],
finished_id, nextSelfplayQ, None)])
else:
checkpoint, is_rp, is_p1 = next_game
assert False
def train(self,
data,
labels,
val_data=None,
val_labels=None,
warm_start=False):
"""
:param data:
:param labels:
:param val_data:
:param val_labels:
:param warm_start:
:return:
"""
# initialize variable
data = torch.from_numpy(data).float()
labels = torch.from_numpy(labels).char()
if val_data is None:
# rows = labels.shape[0]
# index = np.random.permutation(rows)
# val_size = rows // 5
# val_data = data[index[:val_size]]
# val_labels = labels[index[:val_size]]
# train_data = data[index[val_size:]]
# train_labels = labels[index[val_size:]]
train_data, val_data = data, data
train_labels, val_labels = labels, labels
else:
train_data = data
train_labels = labels
val_data = torch.from_numpy(val_data).float()
val_labels = torch.from_numpy(val_labels).char()
orig_cols = train_data.size(1)
# counting class and get their index
self.plus_row_index = []
self.minus_row_index = []
self.orig_minus = 0
self.orig_plus = 0
for idx, value in enumerate(train_labels):
if value == 1:
self.orig_plus += 1
self.plus_row_index.append(idx)
else:
self.orig_minus += 1
self.minus_row_index.append(idx)
# balanced pick rows and cols
plus = max(2, int(self.orig_plus * self.nrows))
minus = max(2, int(self.orig_minus * self.nrows))
num_cols = max(min(5, orig_cols), int(self.nfeatures * orig_cols))
# initialize up triangle matrix and reference index
rows_sum = plus + minus
if self.adv_train:
rows_sum = rows_sum * 2
# plus = plus * 2
# minus = minus * 2
self.yp = torch.ones((rows_sum, rows_sum), dtype=torch.int8).triu_(0)
self.ref_full_index = torch.repeat_interleave(torch.arange(
self.updated_features * self.step.shape[0]).view((-1, 1)),
rows_sum,
dim=1)
# multi-process
pool = Pool(self.n_jobs)
results = []
for r in range(self.round):
if warm_start and self.w_index != []:
column_indices = self.w_index[r]
w = self.w[:, r]
else:
column_indices = np.random.choice(np.arange(orig_cols),
num_cols,
replace=False)
self.w_index.append(column_indices)
w = np.random.uniform(-1, 1,
size=(num_cols, )).astype(np.float32)
results.append(
pool.apply_async(
self.single_run_adv if self.adv_train else self.single_run,
args=(train_data, train_labels, plus, minus, val_data,
val_labels, w, column_indices, r % self.num_gpus)))
pool.close()
pool.join()
for i, result in enumerate(results):
temp_w, temp_b, temp_obj = result.get()
if warm_start:
self.w[:, i] = temp_w
self.b[:, i] = temp_b
self.obj[i] = temp_obj
else:
self.w.append(temp_w.view((-1, 1)))
self.b.append(temp_b.view((1, 1)))
self.obj.append(temp_obj)
if warm_start is False:
self.w = torch.cat(self.w, dim=1)
self.b = torch.cat(self.b, dim=1)
self.obj = torch.Tensor(self.obj)
best_index = self.obj.argmax()
self.best_acc = self.obj[best_index]
self.best_w = self.w[:, best_index]
self.best_b = self.b[:, best_index]
self.best_w_index = self.w_index[best_index]
tplname + '-' + tgtname + '.fasta'), 'w') as f:
f.write(alignment_output)
with open(os.path.join(outputname, "%s_list.Score" % tgt['name']),
"a") as F:
F.write(score_output)
pbar.update()
return
for i in range(datasize):
tplname = template_name_list[i]
observations = torch.load(os.path.join(
obs_path, "%s-%s.pth" % (tplname, tgt['name'])),
pickle_module=pickle)
observations = observations.float()
pool.apply_async(generateAlign,
args=(tplname + tpl_type, tgt['name'] + tgt_type,
observations, transitions),
callback=getoutput)
pool.close()
pool.join()
pbar.close()
print("finish calculating alignment in %.2fs" % (time.time() - start))
sortoutput(tgt['name'],
os.path.join(outputname, tgt['name'] + '_list.Score'), args.k)
print("finish %d alignment generation and save them in %s" %
(datasize, outputname))
print("Date: %s" %
(datetime.datetime.now().strftime("%Y-%m-%d %H:%M")))
print("Command: %s" % " ".join(sys.argv))
print("Query name: %s" % tgt['name'])
print("Template_list: %s" % os.path.basename(args.l))
print("Output path: %s" % outputname)
pbar.update()
return
for i in range(datasize):
if len(pair_name_list[i].split('-')) == 3:
tgtname, domainID, tplname = pair_name_list[i].split('-')
tgtname = "%s-%s" % (tgtname, domainID)
else:
tgtname, tplname = pair_name_list[i].split('-')
observations = torch.load(os.path.join(
s1_path, "%s-%s.DRNF.Score.pkl" % (tplname, tgtname)),
pickle_module=pickle)
observations = observations.float()
pool0.apply_async(generateAlign,
args=(tplname + tpl_type, tgtname + tgt_type,
observations, transitions),
callback=getoutput_init)
pool0.close()
pool0.join()
pbar.close()
print("finish initial alignment in %.2fs" % (time.time() - start))
# empty the cache
if args.s1 == "" and args.s2 == '':
del observation, featdata, seqX, seqY, maskX, maskY
del obsmodel, model1
del data_generator, AlignmentSet
with torch.cuda.device(GPU):
torch.cuda.empty_cache()
del obs_group, crf_group, crfmodel
