def bench(client, n):
""" Benchmark n requests """
pairs = [(x, x + 1) for x in range(n)]
started = time.time()
for pair in pairs:
res, err = client.call('add', *pair)
# assert err is None
duration = time.time() - started
print('Client stats:')
util.print_stats(n, duration)
def bench(client, n):
""" Benchmark n requests """
pairs = [(x, x + 1) for x in range(n)]
started = time.time()
for pair in pairs:
res, err = client.call('add', *pair)
# assert err is None
duration = time.time() - started
print('Client stats:')
util.print_stats(n, duration)
def bench(client, n):
""" Benchmark n requests """
items = list(range(n))
# Time client publish operations
# ------------------------------
started = time.time()
for i in items:
client.publish('test', i)
duration = time.time() - started
print('Publisher client stats:')
util.print_stats(n, duration)
def main(_):
# Configuration.
num_unrolls = FLAGS.num_steps
if FLAGS.seed:
tf.set_random_seed(FLAGS.seed)
# Problem.
problem, net_config, net_assignments = util.get_config(FLAGS.problem,
FLAGS.path, mode='test')
# Optimizer setup.
if FLAGS.optimizer == "Adam":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "L2L":
if FLAGS.path is None:
logging.warning("Evaluating untrained L2L optimizer")
optimizer = meta.MetaOptimizer(FLAGS.problem,, FLAGS.num_particle, **net_config)
meta_loss = optimizer.meta_loss(problem, 1, net_assignments=net_assignments, model_path = FLAGS.path)
loss, update, reset, cost_op, x_final, constant = meta_loss
else:
raise ValueError("{} is not a valid optimizer".format(FLAGS.optimizer))
with ms.MonitoredSession() as sess:
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
min_loss_record = []
all_time_loss_record = []
total_time = 0
total_cost = 0
x_record = [[sess.run(item) for item in x_final]]
for _ in xrange(FLAGS.num_epochs):
# Training.
time, cost, constants = util.eval_run_epoch(sess, cost_op, [update], reset,
num_unrolls, x_final, constant)
total_time += time
all_time_loss_record.append(cost)
with open('./{}/evaluate_record.pickle'.format(FLAGS.path),'wb') as l_record:
record = {'all_time_loss_record':all_time_loss_record,'loss':cost,\
'constants':[sess.run(item) for item in constants],\
}
pickle.dump(record, l_record)
# Results.
util.print_stats("Epoch {}".format(FLAGS.num_epochs), total_cost,
total_time, FLAGS.num_epochs)
def start_service(addr, n):
""" Start a service """
s = Responder(addr)
s.register('add', lambda x, y: x + y)
started = time.time()
for _ in range(n):
s.process()
duration = time.time() - started
time.sleep(0.1)
print('Service stats:')
util.print_stats(n, duration)
return
def start_service(addr, n):
""" Start a service """
s = Service(addr)
started = time.time()
for _ in range(n):
msg = s.sock.recv()
s.sock.send(msg)
s.sock.close()
duration = time.time() - started
print('Raw REP service stats:\n')
util.print_stats(n, duration)
return
def start_service(addr, n):
""" Start a service """
s = Service(addr)
s.register('add', lambda x, y: x + y)
started = time.time()
for _ in range(n):
s.process()
duration = time.time() - started
time.sleep(0.1)
print('Service stats:')
util.print_stats(n, duration)
return
def bench(client, n):
""" Benchmark n requests """
items = list(range(n))
# Time client publish operations
# ------------------------------
started = time.time()
msg = b'test line'
for i in items:
client.sock.send(msg)
duration = time.time() - started
print('Raw PUB client stats:')
util.print_stats(n, duration)
def bench(client, n):
""" Benchmark n requests """
items = list(range(n))
# Time client publish operations
# ------------------------------
started = time.time()
msg = b'test line'
for i in items:
client.socket.send(msg)
duration = time.time() - started
print('Raw PUB client stats:')
util.print_stats(n, duration)
def train_loop(agent,
model,
env,
steps_in_iteration,
n_iterations,
window_length,
history=None,
debug=False,
verbose=0):
env.swap_other_players(
[AIPlayer(agent) for i in range(len(env.other_players))])
history = PokerHistory() if history is None else history
opponent_agent = None
save_agent_weights(agent)
for i in range(n_iterations):
print('ITERATION %s' % str(i))
# Free up resources first
release_memory([opponent_agent, agent])
# Create a copy of the agent to play against us (and to free up the resources recreate our training agent as well)
agent = build_dqn_agent(
model(window_length, env.n_observation_dimensions, env.n_actions),
env.n_actions, window_length, debug)
#load_agent_weights(agent)
opponent_agent = build_dqn_agent(
model(window_length, env.n_observation_dimensions, env.n_actions),
env.n_actions, window_length, debug)
load_agent_weights(opponent_agent)
env.swap_opponent_agent(opponent_agent)
history = agent.fit(env,
nb_steps=steps_in_iteration,
visualize=debug,
log_interval=steps_in_iteration // 5,
verbose=verbose,
history=history)
print_stats(history)
save_agent_weights(agent)
release_memory([opponent_agent, agent])
# Create a copy of the agent to play against us (and to free up the resources recreate our training agent as well)
agent = build_dqn_agent(
model(window_length, env.n_observation_dimensions, env.n_actions),
env.n_actions, window_length, debug)
#load_agent_weights(agent)
opponent_agent = build_dqn_agent(
model(window_length, env.n_observation_dimensions, env.n_actions),
env.n_actions, window_length, debug)
load_agent_weights(opponent_agent)
env.swap_opponent_agent(opponent_agent)
return agent, history
def start_service(addr, n):
""" Start a service """
s = SubService(addr)
s.sock.set_string_option(nanomsg.SUB, nanomsg.SUB_SUBSCRIBE, 'test')
started = time.time()
for _ in range(n):
msg = s.sock.recv()
s.sock.close()
duration = time.time() - started
print('Raw SUB service stats:')
util.print_stats(n, duration)
return
def start_service(addr, n):
""" Start a service """
s = Subscriber(addr)
s.socket.setsockopt(zmq.SUBSCRIBE, 'test')
started = time.time()
for _ in range(n):
msg = s.socket.recv()
s.socket.close()
duration = time.time() - started
print('Raw SUB service stats:')
util.print_stats(n, duration)
return
def main(_):
# Configuration.
num_unrolls = FLAGS.num_steps
if FLAGS.seed:
tf.set_random_seed(FLAGS.seed)
# Problem.
problem, net_config, net_assignments = util.get_config(
FLAGS.problem, FLAGS.path)
# Optimizer setup.
if FLAGS.optimizer == "Adam":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.initialize_variables(problem_vars)
optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
optimizer_reset = tf.initialize_variables(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "L2L":
if FLAGS.path is None:
logging.warning("Evaluating untrained L2L optimizer")
optimizer = meta.MetaOptimizer(**net_config)
meta_loss = optimizer.meta_loss(problem,
1,
net_assignments=net_assignments)
_, update, reset, cost_op, _ = meta_loss
else:
raise ValueError("{} is not a valid optimizer".format(FLAGS.optimizer))
with ms.MonitoredSession() as sess:
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
total_time = 0
total_cost = 0
for _ in xrange(FLAGS.num_epochs):
# Training.
time, cost = util.run_epoch(sess, cost_op, [update], reset,
num_unrolls)
total_time += time
total_cost += cost
# Results.
util.print_stats("Epoch {}".format(FLAGS.num_epochs), total_cost,
total_time, FLAGS.num_epochs)
def bench(client, n):
""" Benchmark n requests """
items = list(range(n))
# Time client publish operations
# ------------------------------
started = time.time()
msg = b'x'
for i in items:
client.sock.send(msg)
res = client.sock.recv()
assert msg == res
duration = time.time() - started
print('Raw REQ client stats:\n')
util.print_stats(n, duration)
def start_service(addr, n):
""" Start a service """
s = SubService(addr)
def do_something(line):
pass
s.subscribe('test', do_something)
started = time.time()
for _ in range(n):
s.process()
s.sock.close()
duration = time.time() - started
print('Subscriber service stats:\n')
util.print_stats(n, duration)
return
def start_service(addr, n):
""" Start a service """
s = Subscriber(addr)
if nanomsg:
s.socket.set_string_option(nanomsg.SUB, nanomsg.SUB_SUBSCRIBE, 'test')
else:
s.socket.setsockopt(nnpy.SUB, nnpy.SUB_SUBSCRIBE, 'test')
started = time.time()
for _ in range(n):
msg = s.socket.recv()
s.socket.close()
duration = time.time() - started
print('Raw SUB service stats:')
util.print_stats(n, duration)
return
def train(device, batch_size, model, trainloader, optimizer, epoch, writer):
model.train()
criterion = nn.CrossEntropyLoss(reduction='mean')
metric_ftns = ['loss', 'correct', 'nums', 'accuracy']
train_metrics = MetricTracker(*[m for m in metric_ftns],
writer=writer,
mode='train')
train_metrics.reset()
confusion_matrix = torch.zeros(2, 2)
for batch_idx, input_tensors in enumerate(trainloader):
optimizer.zero_grad()
input_data, target = input_tensors
input_data = input_data.to(device)
target = target.to(device)
output = model(input_data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
correct, nums, acc = accuracy(output, target)
num_samples = batch_idx * batch_size + 1
_, preds = torch.max(output, 1)
for t, p in zip(target.cpu().view(-1), preds.cpu().view(-1)):
confusion_matrix[t.long(), p.long()] += 1
train_metrics.update_all_metrics(
{
'correct': correct,
'nums': nums,
'loss': loss.item(),
'accuracy': acc
},
writer_step=(epoch - 1) * len(trainloader) + batch_idx)
print_stats(epoch, batch_size, num_samples, trainloader, train_metrics)
num_samples += len(target) - 1
print_summary(epoch, num_samples, train_metrics, mode="Training")
print('Confusion Matrix\n{}\n'.format(confusion_matrix.cpu().numpy()))
return train_metrics
def mm_train(device, batch_size, a_m, i_m, trainloader, optimizer, epoch,
writer):
a_m.train()
i_m.train()
weight = torch.tensor([0.1, 0.9]).to(device)
ce_loss = nn.CrossEntropyLoss(weight=weight, reduction='mean')
alpha = 1
metric_ftns = ['loss', 'correct', 'nums', 'accuracy']
image_metrics = MetricTracker(*[m for m in metric_ftns],
writer=writer,
mode='train')
audio_metrics = MetricTracker(*[m for m in metric_ftns],
writer=writer,
mode='train')
image_metrics.reset()
audio_metrics.reset()
i_confusion_matrix = torch.zeros(2, 2)
a_confusion_matrix = torch.zeros(2, 2)
for batch_idx, (audio, a_label, img, i_label) in enumerate(trainloader):
audio, img = audio.to(device), img.to(device)
a_label, i_label = a_label.to(device), i_label.to(device)
i_output, i_feature = i_m(img)
a_output, a_feature = a_m(audio)
i_ce = ce_loss(i_output, i_label)
a_ce = ce_loss(a_output, a_label)
csa = csa_loss(a_feature, i_feature.detach(),
(a_label == i_label).float())
loss = i_ce + 0.4 * a_ce + alpha * csa
optimizer.zero_grad()
loss.backward()
optimizer.step()
num_samples = batch_idx * batch_size + 1
# image
i_correct, i_nums, i_acc = accuracy(i_output, i_label)
image_metrics.update_all_metrics(
{
'correct': i_correct,
'nums': i_nums,
'loss': loss.item(),
'accuracy': i_acc
},
writer_step=(epoch - 1) * len(trainloader) + batch_idx)
# audio
a_correct, a_nums, a_acc = accuracy(a_output, a_label)
audio_metrics.update_all_metrics(
{
'correct': a_correct,
'nums': a_nums,
'loss': loss.item(),
'accuracy': a_acc
},
writer_step=(epoch - 1) * len(trainloader) + batch_idx)
_, preds = torch.max(a_output, 1)
for t, p in zip(a_label.cpu().view(-1), preds.cpu().view(-1)):
a_confusion_matrix[t.long(), p.long()] += 1
print_stats(epoch,
batch_size,
num_samples,
trainloader,
image_metrics,
mode="Image",
acc=i_acc)
print_stats(epoch,
batch_size,
num_samples,
trainloader,
audio_metrics,
mode="Audio",
acc=a_acc)
num_samples += len(a_output) - 1
print_summary(epoch, num_samples, image_metrics, mode="Training Image")
print_summary(epoch, num_samples, audio_metrics, mode="Training Audio")
print('A_Confusion Matrix\n{}\n'.format(a_confusion_matrix.cpu().numpy()))
return audio_metrics
def main(_):
# Configuration.
num_unrolls = FLAGS.num_steps
if FLAGS.seed:
tf.set_random_seed(FLAGS.seed)
# Problem.
# problem, net_config, net_assignments = util.get_config(FLAGS.problem,
# FLAGS.path)
param_dict = {}
param_dict['bs'] = FLAGS.bs
param_dict['m'] = FLAGS.m
param_dict['n'] = FLAGS.n
print(param_dict)
problem, net_config, net_assignments = util.get_config(
FLAGS.problem,
net_name="RNNprop",
mode=FLAGS.mode, #加入mode
num_linear_heads=1,
init=FLAGS.init,
path=FLAGS.path,
param=param_dict)
# Optimizer setup.
if FLAGS.optimizer == "Adam":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "L2L":
if FLAGS.path is None:
logging.warning("Evaluating untrained L2L optimizer")
optimizer = meta.MetaOptimizer(FLAGS.num_mt, FLAGS.beta1, FLAGS.beta2,
**net_config)
# meta_loss = optimizer.meta_loss(problem, 1, net_assignments=net_assignments)
meta_loss, _, _, _, _, seq_step, \
_, _, _, _, _, _ = optimizer.meta_loss(problem, 1, net_assignments=net_assignments)
#这里原来是各种名字的变量的，但是似乎object never used就是指这些，那我就全部用下划线变量代替了
_, update, reset, cost_op, _ = meta_loss
else:
raise ValueError("{} is not a valid optimizer".format(FLAGS.optimizer))
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with ms.MonitoredSession() as sess:
# with tf.Session(config=config) as sess:
sess.run(reset)
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
total_time = 0
total_cost = 0
loss_record = []
for ep in xrange(FLAGS.num_epochs):
# Training.
time, cost = util.run_eval_epoch(sess,
cost_op, [update],
num_unrolls,
step=seq_step)
total_time += time
total_cost += sum(cost) / num_unrolls
loss_record += cost
print(ep, cost[-1])
# Results.
util.print_stats("Epoch {}".format(FLAGS.num_epochs), total_cost,
total_time, FLAGS.num_epochs)
with open(
'{}/{}_eval_loss_record.pickle'.format(FLAGS.path,
FLAGS.optimizer),
'wb') as l_record:
pickle.dump(loss_record, l_record)
print("Saving evaluate loss record {}".format(FLAGS.path))
model = FrequencyModel(inverted_vocab)
model.fit(docs)
models[country] = model
eval_train = pd.read_csv("./data/DialectSim_train.csv")
eval_test = pd.read_csv("./data/DialectSim_test.csv")
train_word = eval_train["word"].tolist()
train_label = eval_train["label"].tolist()
test_word = eval_test["word"].tolist()
test_label = eval_test["label"].tolist()
pred = np.array([score(w, models) for w in train_word])
linspace = np.linspace(0, 0.5, 1000)
accs = []
for t in linspace:
accs.append(acc(train_label, pred > t))
plt.plot(linspace, accs)
plt.xlabel("Threshold")
plt.ylabel("Acc on training set")
plt.show()
t = linspace[np.argmax(accs)]
# print("Train Acc = %f" % acc(train_label, pred > t))
# print("Test Acc = %f" % acc(test_label, np.array([score(w, models) for w in test_word]) > t))
print("Training stats:")
util.print_stats(pred > t, train_label)
print("Testing stats:")
util.print_stats(
np.array([score(w, models) for w in test_word]) > t, test_label)
def process_input(in_lines, flush_stdout_):
time1 = time.time()
global flush_stdout
flush_stdout = flush_stdout_
multiline = ""
all_lines = []
double_form_cnt = 0
for line in in_lines:
if "#" in line:
line = line.split("#")[0]
if not line.strip():
continue
line = line.rstrip()
if line.endswith("\\"):
multiline += line #.replace("\\", "")
continue
else:
if multiline:
line = multiline + line
multiline = ""
if ("/v" in line and ":imperf:perf" in line) \
or ("/adjp" in line and "&adj" in line):
double_form_cnt += 1
try:
tag_lines = expand_line(line, flush_stdout)
check_lines(tag_lines)
except:
print("Exception in line: \"" + line + "\"", file=sys.stderr)
raise
if flush_stdout:
sorted_lines = util.sort_all_lines(tag_lines)
print("\n".join(sorted_lines))
sys.stdout.flush()
else:
all_lines.extend(tag_lines)
if "-time" in sys.argv:
time2 = time.time()
print("Total out_lines", len(all_lines), file=sys.stderr)
print("Processing time: {0:.3f}".format(time2-time1), file=sys.stderr)
if not flush_stdout:
# print("\n".join(line for line in all_lines if "adv:" in line), file=sys.stderr)
# print("sort 1", file=sys.stderr)
sorted_lines = util.sort_all_lines(all_lines)
# print("\n".join(line for line in sorted_lines if "adv:" in line), file=sys.stderr)
sorted_lines = post_process_sorted(sorted_lines)
if "-time" in sys.argv:
time3 = time.time()
print("Sorting time 1: {0:.3f}".format(time3-time2), file=sys.stderr)
if "-indent" in sys.argv:
# to sort newely promoted lemmas
sorted_lines = util.sort_all_lines(list(set(sorted_lines)))
if "-time" in sys.argv:
time4 = time.time()
print("Sorting time 2: {0:.3f}".format(time4-time3), file=sys.stderr)
# print ("\n-- ".join( ln for ln in sorted_lines if ln.startswith("Венед") ), file=sys.stderr)
sorted_lines = [ cleanup(line) for line in sorted_lines ]
if "-wordlist" in sys.argv:
print_word_list(sorted_lines)
if "-indent" in sys.argv:
if "-mfl" in sys.argv:
with open("dict_corp_lt.txt", "w", encoding="utf-8") as f:
f.write("\n".join(sorted_lines))
sorted_lines = util.indent_lines(sorted_lines)
if "-stats" in sys.argv:
util.print_stats(sorted_lines, double_form_cnt)
if "--log-usage" in sys.argv:
log_usage()
return sorted_lines
def main():
assert torch.cuda.device_count() == 1 # mmdet only supports single GPU testing
opts = parse_args()
assert (not opts.mask_timing) or (opts.mask_timing and not opts.no_mask)
mkdir2(opts.out_dir)
db = COCO(opts.annot_path)
class_names = [c['name'] for c in db.dataset['categories']]
n_class = len(class_names)
coco_mapping = None if opts.no_class_mapping else db.dataset.get('coco_mapping', None)
if coco_mapping is not None:
coco_mapping = np.asarray(coco_mapping)
seqs = db.dataset['sequences']
seq_dirs = db.dataset['seq_dirs']
model = init_detector(opts)
# warm up the GPU
img = db.imgs[0]
w_img, h_img = img['width'], img['height']
_ = inference_detector(model, np.zeros((h_img, w_img, 3), np.uint8))
torch.cuda.synchronize()
runtime_all = []
n_processed = 0
n_total = 0
for sid, seq in enumerate(tqdm(seqs)):
frame_list = [img for img in db.imgs.values() if img['sid'] == sid]
# load all frames in advance
frames = []
for img in frame_list:
img_path = join(opts.data_root, seq_dirs[sid], img['name'])
frames.append(imread(img_path))
n_frame = len(frames)
n_total += n_frame
timestamps = []
results_raw = []
results_parsed = []
input_fidx = []
runtime = []
last_fidx = None
if not opts.dynamic_schedule:
stride_cnt = 0
t_total = n_frame/opts.fps
t_start = perf_counter()
while 1:
t1 = perf_counter()
t_elapsed = t1 - t_start
if t_elapsed >= t_total:
break
# identify latest available frame
fidx_continous = t_elapsed*opts.fps
fidx = int(np.floor(fidx_continous))
if fidx == last_fidx:
continue
last_fidx = fidx
if opts.dynamic_schedule:
fidx_remainder = fidx_continous - fidx
if fidx_remainder > 0.5:
continue
else:
if stride_cnt % opts.det_stride == 0:
stride_cnt = 1
else:
stride_cnt += 1
continue
frame = frames[fidx]
result = inference_detector(model, frame, gpu_pre=not opts.cpu_pre, decode_mask=not opts.mask_timing)
if opts.mask_timing:
mask_encoded = result[2]
result = result[:2]
bboxes, scores, labels, _, sel = \
parse_det_result(result, coco_mapping, n_class, return_sel=True)
else:
bboxes, scores, labels, masks = \
parse_det_result(result, coco_mapping, n_class)
torch.cuda.synchronize()
t2 = perf_counter()
t_elapsed = t2 - t_start
if t_elapsed >= t_total:
break
if opts.mask_timing:
masks = decode_mask(mask_encoded, sel)
timestamps.append(t_elapsed)
results_raw.append(result)
results_parsed.append((bboxes, scores, labels, masks))
input_fidx.append(fidx)
runtime.append(t2 - t1)
out_path = join(opts.out_dir, seq + '.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump({
'results_raw': results_raw,
'results_parsed': results_parsed,
'timestamps': timestamps,
'input_fidx': input_fidx,
'runtime': runtime,
}, open(out_path, 'wb'))
runtime_all += runtime
n_processed += len(results_raw)
runtime_all_np = np.asarray(runtime_all)
n_small_runtime = (runtime_all_np < 1.0/opts.fps).sum()
out_path = join(opts.out_dir, 'time_info.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump({
'runtime_all': runtime_all,
'n_processed': n_processed,
'n_total': n_total,
'n_small_runtime': n_small_runtime,
}, open(out_path, 'wb'))
# convert to ms for display
s2ms = lambda x: 1e3*x
print(f'{n_processed}/{n_total} frames processed')
print_stats(runtime_all_np, 'Runtime (ms)', cvt=s2ms)
print(f'Runtime smaller than unit time interval: '
f'{n_small_runtime}/{n_processed} '
f'({100.0*n_small_runtime/n_processed:.4g}%)')
def main():
assert torch.cuda.device_count() == 1 # mmdet only supports single GPU testing
opts = parse_args()
mkdir2(opts.out_dir)
db = COCO(opts.annot_path)
class_names = [c['name'] for c in db.dataset['categories']]
n_class = len(class_names)
coco_mapping = db.dataset.get('coco_mapping', None)
if coco_mapping is not None:
coco_mapping = np.asarray(coco_mapping)
seqs = db.dataset['sequences']
seq_dirs = db.dataset['seq_dirs']
img = db.imgs[0]
w_img, h_img = img['width'], img['height']
mp.set_start_method('spawn')
frame_recv, frame_send = mp.Pipe(False)
det_res_recv, det_res_send = mp.Pipe(False)
det_proc = mp.Process(target=det_process, args=(opts, frame_recv, det_res_send, w_img, h_img))
det_proc.start()
if opts.dynamic_schedule:
runtime = pickle.load(open(opts.runtime, 'rb'))
runtime_dist = dist_from_dict(runtime, opts.perf_factor)
mean_rtf = runtime_dist.mean()*opts.fps
n_total = 0
t_det_all = []
t_send_frame_all = []
t_recv_res_all = []
t_assoc_all = []
t_forecast_all = []
with torch.no_grad():
kf_F = torch.eye(8)
kf_Q = torch.eye(8)
kf_R = 10*torch.eye(4)
kf_P_init = 100*torch.eye(8).unsqueeze(0)
for sid, seq in enumerate(tqdm(seqs)):
frame_list = [img for img in db.imgs.values() if img['sid'] == sid]
frame_list = [join(opts.data_root, seq_dirs[sid], img['name']) for img in frame_list]
n_frame = len(frame_list)
n_total += n_frame
timestamps = []
results_parsed = []
input_fidx = []
processing = False
fidx_t2 = None # detection input index at t2
fidx_latest = None
tkidx = 0 # track starting index
kf_x = torch.empty((0, 8, 1))
kf_P = torch.empty((0, 8, 8))
n_matched12 = 0
# let detector process to read all the frames
frame_send.send(frame_list)
# it is possible that unfetched results remain in the pipe
while 1:
msg = det_res_recv.recv() # wait till the detector is ready
if msg == 'ready':
break
elif isinstance(msg, Exception):
raise msg
t_total = n_frame/opts.fps
t_unit = 1/opts.fps
t_start = perf_counter()
while 1:
t1 = perf_counter()
t_elapsed = t1 - t_start
if t_elapsed >= t_total:
break
# identify latest available frame
fidx_continous = t_elapsed*opts.fps
fidx = int(np.floor(fidx_continous))
if fidx == fidx_latest:
# algorithm is fast and has some idle time
wait_for_next = True
else:
wait_for_next = False
if opts.dynamic_schedule:
if mean_rtf >= 1:
# when runtime < 1, it should always process every frame
fidx_remainder = fidx_continous - fidx
if mean_rtf < np.floor(fidx_remainder + mean_rtf):
# wait till next frame
wait_for_next = True
if wait_for_next:
# sleep
continue
if not processing:
t_start_frame = perf_counter()
frame_send.send((fidx, t_start_frame))
fidx_latest = fidx
processing = True
# wait till query - forecast-rt-ub
wait_time = t_unit - opts.forecast_rt_ub
if det_res_recv.poll(wait_time): # wait
# new result
result = det_res_recv.recv()
if isinstance(result, Exception):
raise result
result, t_send_frame, t_start_res = result
bboxes_t2, scores_t2, labels_t2, _ = \
parse_det_result(result, coco_mapping, n_class)
processing = False
t_det_end = perf_counter()
t_det_all.append(t_det_end - t_start_frame)
t_send_frame_all.append(t_send_frame)
t_recv_res_all.append(t_det_end - t_start_res)
# associate across frames
t_assoc_start = perf_counter()
if len(kf_x):
dt = fidx_latest - fidx_t2
kf_F = make_F(kf_F, dt)
kf_Q = make_Q(kf_Q, dt)
kf_x, kf_P = batch_kf_predict(kf_F, kf_x, kf_P, kf_Q)
bboxes_f = x2bbox(kf_x)
fidx_t2 = fidx_latest
n = len(bboxes_t2)
if n:
# put high scores det first for better iou matching
score_argsort = np.argsort(scores_t2)[::-1]
bboxes_t2 = bboxes_t2[score_argsort]
scores_t2 = scores_t2[score_argsort]
labels_t2 = labels_t2[score_argsort]
ltrb2ltwh_(bboxes_t2)
updated = False
if len(kf_x):
order1, order2, n_matched12, tracks, tkidx = iou_assoc(
bboxes_f, labels, tracks, tkidx,
bboxes_t2, labels_t2, opts.match_iou_th,
no_unmatched1=True,
)
if n_matched12:
kf_x = kf_x[order1]
kf_P = kf_P[order1]
kf_x, kf_P = batch_kf_update(
bbox2z(bboxes_t2[order2[:n_matched12]]),
kf_x,
kf_P,
kf_R,
)
kf_x_new = bbox2x(bboxes_t2[order2[n_matched12:]])
n_unmatched2 = len(bboxes_t2) - n_matched12
kf_P_new = kf_P_init.expand(n_unmatched2, -1, -1)
kf_x = torch.cat((kf_x, kf_x_new))
kf_P = torch.cat((kf_P, kf_P_new))
labels = labels_t2[order2]
scores = scores_t2[order2]
updated = True
if not updated:
# start from scratch
kf_x = bbox2x(bboxes_t2)
kf_P = kf_P_init.expand(len(bboxes_t2), -1, -1)
labels = labels_t2
scores = scores_t2
tracks = np.arange(tkidx, tkidx + n, dtype=np.uint32)
tkidx += n
t_assoc_end = perf_counter()
t_assoc_all.append(t_assoc_end - t_assoc_start)
# apply forecasting for the current query
t_forecast_start = perf_counter()
query_pointer = fidx + opts.eta + 1
if len(kf_x):
dt = (query_pointer - fidx_t2)
kf_x_np = kf_x[:, :, 0].numpy()
bboxes_t3 = kf_x_np[:n_matched12, :4] + dt*kf_x_np[:n_matched12, 4:]
if n_matched12 < len(kf_x):
bboxes_t3 = np.concatenate((bboxes_t3, kf_x_np[n_matched12:, :4]))
bboxes_t3, keep = extrap_clean_up(bboxes_t3, w_img, h_img, lt=True)
labels_t3 = labels[keep]
scores_t3 = scores[keep]
tracks_t3 = tracks[keep]
else:
bboxes_t3 = np.empty((0, 4), dtype=np.float32)
scores_t3 = np.empty((0,), dtype=np.float32)
labels_t3 = np.empty((0,), dtype=np.int32)
tracks_t3 = np.empty((0,), dtype=np.int32)
t_forecast_end = perf_counter()
t_forecast_all.append(t_forecast_end - t_forecast_start)
t3 = perf_counter()
t_elapsed = t3 - t_start
if t_elapsed >= t_total:
break
if len(bboxes_t3):
ltwh2ltrb_(bboxes_t3)
if fidx_t2 is not None:
timestamps.append(t_elapsed)
results_parsed.append((bboxes_t3, scores_t3, labels_t3, None, tracks_t3))
input_fidx.append(fidx_t2)
out_path = join(opts.out_dir, seq + '.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump({
'results_parsed': results_parsed,
'timestamps': timestamps,
'input_fidx': input_fidx,
}, open(out_path, 'wb'))
# terminates the child process
frame_send.send(None)
out_path = join(opts.out_dir, 'time_info.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump({
'n_total': n_total,
't_det': t_det_all,
't_send_frame': t_send_frame_all,
't_recv_res': t_recv_res_all,
't_assoc': t_assoc_all,
't_forecast': t_forecast_all,
}, open(out_path, 'wb'))
# convert to ms for display
s2ms = lambda x: 1e3*x
print_stats(t_det_all, 'Runtime detection (ms)', cvt=s2ms)
print_stats(t_send_frame_all, 'Runtime sending the frame (ms)', cvt=s2ms)
print_stats(t_recv_res_all, 'Runtime receiving the result (ms)', cvt=s2ms)
print_stats(t_assoc_all, "Runtime association (ms)", cvt=s2ms)
print_stats(t_forecast_all, "Runtime forecasting (ms)", cvt=s2ms)
with open(uk_dict_file, 'wb') as handle:
pickle.dump(uk_dict, handle, protocol=pickle.HIGHEST_PROTOCOL)
# print(us_dict)
# input()
# print(uk_dict)
# input()
pred = np.array([
manhattan_distance(w, us_dict, uk_dict)
for w in tqdm(train_words, desc="AdaGram")
])
linspace = np.linspace(0, 2.0, 1000)
metrics_scores = []
for t in linspace:
metrics_scores.append(acc(train_labels, pred > t))
plt.plot(linspace, metrics_scores)
plt.xlabel("Threshold")
plt.ylabel("Acc on training set")
plt.show()
t = linspace[np.argmax(metrics_scores)]
print("Training stats:")
util.print_stats(pred > t, train_labels)
print("Testing stats:")
util.print_stats(
np.array([manhattan_distance(w, us_dict, uk_dict)
for w in test_words]) > t, test_labels)
print(COUNT)
def main():
assert torch.cuda.device_count(
) == 1 # mmdet only supports single GPU testing
opts = parse_args()
mp.set_start_method('spawn')
frame_recv, frame_send = mp.Pipe(False)
det_res_recv, det_res_send = mp.Pipe(False)
det_proc = mp.Process(target=det_process,
args=(opts, frame_recv, det_res_send))
det_proc.start()
mkdir2(opts.out_dir)
db = COCO(opts.annot_path)
class_names = [c['name'] for c in db.dataset['categories']]
n_class = len(class_names)
coco_mapping = db.dataset.get('coco_mapping', None)
if coco_mapping is not None:
coco_mapping = np.asarray(coco_mapping)
seqs = db.dataset['sequences']
seq_dirs = db.dataset['seq_dirs']
runtime_all = []
n_processed = 0
n_total = 0
t_send_frame_all = []
t_recv_res_all = []
for sid, seq in enumerate(tqdm(seqs)):
frame_list = [img for img in db.imgs.values() if img['sid'] == sid]
frame_list = [
join(opts.data_root, seq_dirs[sid], img['name'])
for img in frame_list
]
n_frame = len(frame_list)
n_total += n_frame
timestamps = []
results_raw = []
results_parsed = []
input_fidx = []
runtime = []
last_fidx = None
stride_cnt = 0
# let detector process to read all the frames
frame_send.send(frame_list)
init_error = det_res_recv.recv() # wait till the detector is ready
if init_error is not None:
raise init_error
t_total = n_frame / opts.fps
t_start = perf_counter()
while 1:
t1 = perf_counter()
t_elapsed = t1 - t_start
if t_elapsed >= t_total:
break
# identify latest available frame
fidx = int(np.floor(t_elapsed * opts.fps))
# t_elapsed/t_total *n_frame
# = t_elapsed*opts.fps
if fidx == last_fidx:
continue
last_fidx = fidx
if stride_cnt % opts.det_stride == 0:
stride_cnt = 1
else:
stride_cnt += 1
continue
t_start_frame = perf_counter()
frame_send.send((fidx, t_start_frame))
result = det_res_recv.recv() # wait
if isinstance(result, Exception):
raise result
result, t_send_frame, t_start_res = result
bboxes, scores, labels, masks = \
parse_det_result(result, coco_mapping, n_class)
t2 = perf_counter()
t_send_frame_all.append(t_send_frame)
t_recv_res_all.append(t2 - t_start_res)
t_elapsed = t2 - t_start
if t_elapsed >= t_total:
break
timestamps.append(t_elapsed)
results_raw.append(result)
results_parsed.append((bboxes, scores, labels, masks))
input_fidx.append(fidx)
runtime.append(t2 - t1)
out_path = join(opts.out_dir, seq + '.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump(
{
'results_raw': results_raw,
'results_parsed': results_parsed,
'timestamps': timestamps,
'input_fidx': input_fidx,
'runtime': runtime,
}, open(out_path, 'wb'))
runtime_all += runtime
n_processed += len(results_raw)
# terminates the child process
frame_send.send(None)
runtime_all_np = np.asarray(runtime_all)
n_small_runtime = (runtime_all_np < 1.0 / opts.fps).sum()
out_path = join(opts.out_dir, 'time_info.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump(
{
'runtime_all': runtime_all,
'n_processed': n_processed,
'n_total': n_total,
'n_small_runtime': n_small_runtime,
}, open(out_path, 'wb'))
# convert to ms for display
s2ms = lambda x: 1e3 * x
print(f'{n_processed}/{n_total} frames processed')
print_stats(runtime_all_np, 'Runtime (ms)', cvt=s2ms)
print(f'Runtime smaller than unit time interval: '
f'{n_small_runtime}/{n_processed} '
f'({100.0*n_small_runtime/n_processed:.4g}%)')
print_stats(t_send_frame_all,
'Time spent on sending the frame (ms)',
cvt=s2ms)
print_stats(t_recv_res_all,
'Time spent on receiving the result (ms)',
cvt=s2ms)
def main():
assert torch.cuda.device_count(
) == 1 # mmdet only supports single GPU testing
opts = parse_args()
mkdir2(opts.out_dir)
vis_out = bool(opts.vis_dir)
if vis_out:
mkdir2(opts.vis_dir)
db = COCO(opts.annot_path)
class_names = [c['name'] for c in db.dataset['categories']]
n_class = len(class_names)
coco_mapping = None if opts.no_class_mapping else db.dataset.get(
'coco_mapping', None)
if coco_mapping is not None:
coco_mapping = np.asarray(coco_mapping)
seqs = db.dataset['sequences']
seq_dirs = db.dataset['seq_dirs']
model = init_detector(opts)
if opts.weights_base is not None:
# for distillation purpose
load_checkpoint(model, opts.weights_base)
if opts.cpu_pre:
img_transform = ImageTransform(
size_divisor=model.cfg.data.test.size_divisor,
**model.cfg.img_norm_cfg)
else:
img_transform = ImageTransformGPU(
size_divisor=model.cfg.data.test.size_divisor,
**model.cfg.img_norm_cfg)
device = next(model.parameters()).device # model device
n_history = model.cfg.data.train.n_history if opts.n_history is None else opts.n_history
n_future = model.cfg.data.train.n_future if opts.n_future is None else opts.n_future
results_ccf = [] # instance based
runtime_all = []
for sid, seq in enumerate(tqdm(seqs)):
# print(seq)
frame_list = [img for img in db.imgs.values() if img['sid'] == sid]
n_frame = len(frame_list)
# load all frames in advance
frames = []
for img in frame_list:
img_path = join(opts.data_root, seq_dirs[sid], img['name'])
frames.append(imread(img_path))
with torch.no_grad():
preprocessed = []
for i in range(n_history):
data = _prepare_data(frames[i], img_transform, model.cfg,
device)
preprocessed.append(data)
for ii in range(n_history, n_frame - n_future):
# target frame
iid = frame_list[ii + n_future]['id']
img_name = frame_list[ii + n_future]['name']
I = frames[ii + n_future]
t_start = perf_counter()
# input frame
data = _prepare_data(frames[ii], img_transform, model.cfg,
device)
# if n_history == 0:
# data_merge = data
# # print(data['img'])
# # print(data['img'][0].shape)
# # print(data['img'][0][0][0][300][300:305])
# # import sys
# # sys.exit()
# else:
preprocessed.append(data)
# print(preprocessed[0]['img'][0].data_ptr())
# print(preprocessed[2]['img'][0].data_ptr())
# print(torch.all(preprocessed[0]['img'][0] == preprocessed[2]['img'][0]))
imgs = [d['img'][0] for d in preprocessed]
imgs = torch.cat(imgs, 0)
imgs = imgs.unsqueeze(0)
data_merge = {
'img': [imgs],
'img_meta': data['img_meta'],
}
# print(data_merge['img'][0][0][2][0][300][300:305])
# import sys
# sys.exit()
result = model(return_loss=False,
rescale=True,
numpy_res=True,
**data_merge)
bboxes, scores, labels, masks = \
parse_det_result(result, coco_mapping, n_class)
# if ii == 2:
# print(ii, scores)
# import sys
# sys.exit()
# if n_history != 0:
del preprocessed[0]
t_end = perf_counter()
runtime_all.append(t_end - t_start)
if vis_out:
vis_path = join(opts.vis_dir, seq, img_name[:-3] + 'jpg')
if opts.overwrite or not isfile(vis_path):
vis_det(I,
bboxes,
labels,
class_names,
masks,
scores,
out_scale=opts.vis_scale,
out_file=vis_path)
# convert to coco fmt
n = len(bboxes)
if n:
bboxes[:, 2:] -= bboxes[:, :2]
for i in range(n):
result_dict = {
'image_id': iid,
'bbox': bboxes[i],
'score': scores[i],
'category_id': labels[i],
}
if masks is not None:
result_dict['segmentation'] = masks[i]
results_ccf.append(result_dict)
out_path = join(opts.out_dir, 'time_info.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump({
'runtime_all': runtime_all,
'n_total': len(runtime_all),
}, open(out_path, 'wb'))
# convert to ms for display
s2ms = lambda x: 1e3 * x
print_stats(runtime_all, 'Runtime (ms)', cvt=s2ms)
out_path = join(opts.out_dir, 'results_ccf.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump(results_ccf, open(out_path, 'wb'))
if not opts.no_eval:
eval_summary = eval_ccf(db, results_ccf)
out_path = join(opts.out_dir, 'eval_summary.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump(eval_summary, open(out_path, 'wb'))
if vis_out:
print(f'python vis/make_videos.py "{opts.vis_dir}"')
def main(_):
# Configuration.
num_unrolls = FLAGS.num_steps // FLAGS.unroll_length
# Problem.
problem, net_config, net_assignments = util.get_config(FLAGS.problem)
# Optimizer setup.
optimizer = meta.MetaOptimizer(**net_config)
minimize = optimizer.meta_minimize(
problem, FLAGS.unroll_length,
learning_rate=FLAGS.learning_rate,
net_assignments=net_assignments,
second_derivatives=FLAGS.second_derivatives)
step, loss, update, reset, cost_op, farray, lropt, _ = minimize
with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:
# Prevent accidental changes to the graph.
graph_writer = tf.summary.FileWriter(logs_path, sess.graph)
sess.run(tf.global_variables_initializer())
best_evaluation = float("inf")
start = timer()
losstrain = []
lrtrain = []
losseval = []
plotlosstrain = []
plotlrtrain = []
plotlosseval = []
for e in range(FLAGS.num_epochs):
cost, trainloss, lropttrain = util.run_epoch(sess, cost_op, farray, lropt, [step, update], reset, num_unrolls)
print(cost)
losstrain.append(cost)
lrtrain.append(lropttrain)
util.print_stats("Training Epoch {}".format(e), trainloss, timer() - start)
saver = tf.train.Saver()
if (e + 1) % FLAGS.logging_period == 0:
plotlosstrain.append(cost)
plotlrtrain.append(lropttrain)
if (e + 1) % FLAGS.evaluation_period == 0:
for _ in range(FLAGS.evaluation_epochs):
evalcost, evaloss, _ = util.run_epoch(sess, cost_op, farray, lropt, [update], reset, num_unrolls)
losseval.append(evalcost)
if save_path is not None and evaloss < best_evaluation:
print("Saving meta-optimizer to {}".format(save_path))
saver.save(sess, save_path + '/model.ckpt', global_step=e + 1)
best_evaluation = evaloss
plotlosseval.append(evalcost)
slengths = np.arange(FLAGS.num_steps)
slengthlr = np.arange(FLAGS.num_steps - num_unrolls)
np.savetxt(save_path + '/plotlosstrain.out', plotlosstrain, delimiter=',')
np.savetxt(save_path + '/plotlrtrain.out', plotlrtrain, delimiter=',')
np.savetxt(save_path + '/plotlosseval.out', plotlosseval, delimiter=',')
np.savetxt(save_path + '/losstrain.out', losstrain, delimiter=',')
np.savetxt(save_path + '/lrtrain.out', plotlosstrain, delimiter=',')
np.savetxt(save_path + '/losseval.out', losseval, delimiter=',')
plt.figure(figsize=(8, 5))
plt.plot(slengths, np.mean(plotlosstrain, 0), 'r-', label='Training Loss')
plt.xlabel('Epoch')
plt.ylabel('Training Loss')
plt.legend()
savefig(save_path + '/Training.png')
plt.close()
plt.figure(figsize=(8, 5))
plt.plot(slengths, np.mean(plotlosseval, 0), 'b-', label='Validation Loss')
plt.xlabel('Epoch')
plt.ylabel('Validation Loss')
plt.legend()
savefig(save_path + '/Validation.png')
plt.close()
plt.figure(figsize=(8, 5))
plt.plot(slengthlr, np.mean(plotlrtrain, 0), 'r-', label='Learning Rate')
plt.xlabel('Epoch')
plt.ylabel('Average Learning Rate')
plt.legend()
savefig(save_path + '/LearningRate.png')
plt.close()
graph_writer.close()
meta_loss = optimizer.meta_loss(problem, 1, net_assignments=net_assignments)
_, update, reset, cost_op, problem_vars = meta_loss
else:
raise ValueError("{} is not a valid optimizer".format(optimizer))
sess = ms.MonitoredSession()# as sess:
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
total_time = 0
total_cost = 0
for i in xrange(num_epochs):
# Training.
time, cost, problem_res = util.run_epoch_test(sess, cost_op, problem_vars, [update], reset, num_unrolls)
total_time += time
total_cost += cost
print('Cost at Iteration: '+str(i)+' is: '+str(total_cost))
# Results.
util.print_stats("Epoch {}".format(num_epochs), total_cost,
total_time, num_epochs)
#%% print the results
# not quiet clear how the shaping is done..
# e.g. problem_res.shape = (20, 2, 100, 4, 100) # (nsteps, (real/imag), nx, nz, ny) - > REAL/IMAG are two seperate variables which we try to optimize here
for i in range(problem_res.shape[0]):
mycurrent_result = np.abs(problem_res[i,0,:,1,:]+1j*problem_res[i,1,:,1,:])
plt.imshow(np.abs(np.squeeze(mycurrent_result))), plt.show()
def main(_):
# Configuration.
num_unrolls = FLAGS.num_steps
if FLAGS.seed:
tf.set_random_seed(FLAGS.seed)
# Problem.
problem, net_config, net_assignments = util.get_config(FLAGS.problem,
FLAGS.path,
net_name="RNNprop")
# Optimizer setup.
if FLAGS.optimizer == "Adam":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "L2L":
if FLAGS.path is None:
logging.warning("Evaluating untrained L2L optimizer")
optimizer = meta.MetaOptimizer(FLAGS.beta1, FLAGS.beta2, **net_config)
meta_loss, _, _, step = optimizer.meta_loss(
problem, 1, net_assignments=net_assignments)
_, update, reset, cost_op, _ = meta_loss
else:
raise ValueError("{} is not a valid optimizer".format(FLAGS.optimizer))
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with ms.MonitoredSession() as sess:
sess.run(reset)
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
total_time = 0
total_cost = 0
loss_record = []
for e in xrange(FLAGS.num_epochs):
# Training.
time, cost = util.run_eval_epoch(sess,
cost_op, [update],
num_unrolls,
step=step,
unroll_len=1)
total_time += time
total_cost += sum(cost) / num_unrolls
loss_record += cost
# Results.
util.print_stats("Epoch {}".format(FLAGS.num_epochs), total_cost,
total_time, FLAGS.num_epochs)
if FLAGS.output_path is not None:
if not os.path.exists(FLAGS.output_path):
os.mkdir(FLAGS.output_path)
output_file = '{}/{}_eval_loss_record.pickle-{}'.format(
FLAGS.output_path, FLAGS.optimizer, FLAGS.problem)
with open(output_file, 'wb') as l_record:
pickle.dump(loss_record, l_record)
print("Saving evaluate loss record {}".format(output_file))
def main(_):
# Configuration.
num_unrolls = FLAGS.num_steps // FLAGS.unroll_length
problem, net_config, net_assignments = util.get_config(FLAGS.problem)
optimizer = meta.MetaOptimizer(**net_config)
if FLAGS.save_path is not None:
if not os.path.exists(FLAGS.save_path):
os.mkdir(FLAGS.save_path)
path = None
# raise ValueError("Folder {} already exists".format(FLAGS.save_path))
else:
if os.path.exists('{}/loss-record.pickle'.format(FLAGS.save_path)):
path = FLAGS.save_path
else:
path = None
# Problem.
# Optimizer setup.
minimize = optimizer.meta_minimize(
problem,
FLAGS.unroll_length,
learning_rate=FLAGS.learning_rate,
net_assignments=net_assignments,
model_path=path,
second_derivatives=FLAGS.second_derivatives)
step, update, reset, cost_op, x_final, test, fc_weights, fc_bias, fc_va = minimize
# saver=tf.train.Saver()
with ms.MonitoredSession() as sess:
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
# Step=[step for i in range(len(cost_op))]
best_evaluation = float("inf")
total_time = 0
total_cost = 0
loss_record = []
constants = []
for e in xrange(FLAGS.num_epochs):
# Training.
time, cost, constant, Weights = util.run_epoch(
sess, cost_op, [update, step], reset, num_unrolls, test,
[fc_weights, fc_bias, fc_va])
cost = sum(cost) / len(cost_op)
total_time += time
total_cost += cost
loss_record.append(cost)
constants.append(constant)
# Logging.
if (e + 1) % FLAGS.log_period == 0:
util.print_stats("Epoch {}".format(e + 1), total_cost,
total_time, FLAGS.log_period)
total_time = 0
total_cost = 0
# Evaluation.
if (e + 1) % FLAGS.evaluation_period == 0:
eval_cost = 0
eval_time = 0
for _ in xrange(FLAGS.evaluation_epochs):
time, cost, constant, weights = util.run_epoch(
sess, cost_op, [update, step], reset, num_unrolls,
test, [fc_weights, fc_bias, fc_va])
# cost/=len(cost_op)
eval_time += time
eval_cost += sum(cost) / len(cost_op)
util.print_stats("EVALUATION", eval_cost, eval_time,
FLAGS.evaluation_epochs)
if FLAGS.save_path is not None and eval_cost < best_evaluation:
print("Removing previously saved meta-optimizer")
for f in os.listdir(FLAGS.save_path):
os.remove(os.path.join(FLAGS.save_path, f))
print("Saving meta-optimizer to {}".format(
FLAGS.save_path))
# saver.save(sess,'./quadratic/quadratic.ckpt',global_step = e)
optimizer.save(sess, FLAGS.save_path)
with open(FLAGS.save_path + '/loss_record.pickle',
'wb') as l_record:
record = {'loss_record':loss_record, 'fc_weights':sess.run(weights[0]), \
'fc_bias':sess.run(weights[1]), 'fc_va':sess.run(weights[2]), 'constant':sess.run(constant)}
pickle.dump(record, l_record)
best_evaluation = eval_cost
def main(_):
# Configuration.
num_unrolls = FLAGS.num_steps // FLAGS.unroll_length
# if FLAGS.save_path is not None:
# if os.path.exists(FLAGS.save_path):
# raise ValueError("Folder {} already exists".format(FLAGS.save_path))
# else:
# os.mkdir(FLAGS.save_path)
# Problem.
problem, net_config, net_assignments = util.get_config(
FLAGS.problem, main_parade_path, first_batch_parade_path)
# Optimizer setup.
optimizer = meta.MetaOptimizer(**net_config)
minimize = optimizer.meta_minimize(
problem, FLAGS.unroll_length,
learning_rate=FLAGS.learning_rate,
net_assignments=net_assignments,
second_derivatives=FLAGS.second_derivatives)
step, update, reset, cost_op, _ = minimize
with ms.MonitoredSession() as sess:
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
writer = tf.summary.FileWriter('summary')
writer.add_graph(tf.get_default_graph())
best_evaluation = float("inf")
total_time = 0
total_cost = 0
for e in xrange(FLAGS.num_epochs):
# Training.
time, cost = util.run_epoch(sess, cost_op, [update, step], reset,
num_unrolls)
total_time += time
total_cost += cost
# Logging.
if (e + 1) % FLAGS.log_period == 0:
util.print_stats("Epoch {}".format(e + 1), total_cost, total_time,
FLAGS.log_period)
total_time = 0
total_cost = 0
# Evaluation.
if (e + 1) % FLAGS.evaluation_period == 0:
eval_cost = 0
eval_time = 0
for _ in xrange(FLAGS.evaluation_epochs):
time, cost = util.run_epoch(sess, cost_op, [update], reset,
num_unrolls)
eval_time += time
eval_cost += cost
util.print_stats("EVALUATION", eval_cost, eval_time,
FLAGS.evaluation_epochs)
if FLAGS.save_path is not None and eval_cost < best_evaluation:
print("Removing previously saved meta-optimizer")
for f in os.listdir(FLAGS.save_path):
os.remove(os.path.join(FLAGS.save_path, f))
print("Saving meta-optimizer to {}".format(FLAGS.save_path))
optimizer.save(sess, FLAGS.save_path)
best_evaluation = eval_cost
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
best_evaluation = float("inf")
total_time = 0
total_cost = 0
for e in xrange(num_epochs):
# Training.
time, cost = util.run_epoch(sess, cost_op, [update, step], reset,
num_unrolls)
total_time += time
total_cost += cost
# Logging.
if (e + 1) % log_period == 0:
util.print_stats("Epoch {}".format(e + 1), total_cost, total_time,
log_period)
total_time = 0
total_cost = 0
# Evaluation.
if (e + 1) % evaluation_period == 0:
eval_cost = 0
eval_time = 0
for _ in xrange(evaluation_epochs):
time, cost = util.run_epoch(sess, cost_op, [update], reset,
num_unrolls)
eval_time += time
eval_cost += cost
util.print_stats("EVALUATION", eval_cost, eval_time,
evaluation_epochs)
def main():
opts = parse_args()
mkdir2(opts.out_dir)
db = COCO(opts.annot_path)
class_names = [c['name'] for c in db.dataset['categories']]
n_class = len(class_names)
coco_mapping = None if opts.no_class_mapping else db.dataset.get('coco_mapping', None)
if coco_mapping is not None:
coco_mapping = np.asarray(coco_mapping)
seqs = db.dataset['sequences']
seq_dirs = db.dataset['seq_dirs']
if opts.cached_res:
cache_in_ccf = '_ccf' in basename(opts.cached_res)
if cache_in_ccf:
# speed up based on the assumption of sequential storage
cache_end_idx = 0
cached_res = pickle.load(open(opts.cached_res, 'rb'))
else:
assert torch.cuda.device_count() == 1 # mmdet only supports single GPU testing
model = init_detector(opts)
np.random.seed(opts.seed)
runtime = pickle.load(open(opts.runtime, 'rb'))
runtime_dist = dist_from_dict(runtime, opts.perf_factor)
runtime_all = []
n_processed = 0
n_total = 0
for sid, seq in enumerate(tqdm(seqs)):
frame_list = [img for img in db.imgs.values() if img['sid'] == sid]
n_frame = len(frame_list)
n_total += n_frame
if not opts.cached_res:
# load all frames in advance
frames = []
for img in frame_list:
img_path = join(opts.data_root, seq_dirs[sid], img['name'])
frames.append(imread(img_path))
timestamps = []
results_parsed = []
input_fidx = []
runtime = []
last_fidx = None
if opts.cached_res and cache_in_ccf:
results_raw = None
else:
results_raw = []
t_total = n_frame/opts.fps
t_elapsed = 0
if opts.dynamic_schedule:
mean_rtf = runtime_dist.mean()*opts.fps
else:
stride_cnt = 0
while 1:
if t_elapsed >= t_total:
break
# identify latest available frame
fidx_continous = t_elapsed*opts.fps
fidx = int(np.floor(fidx_continous))
if fidx == last_fidx:
# algorithm is fast and has some idle time
fidx += 1
if fidx == n_frame:
break
t_elapsed = fidx/opts.fps
last_fidx = fidx
if opts.dynamic_schedule:
if mean_rtf > 1:
# when runtime <= 1, it should always process every frame
fidx_remainder = fidx_continous - fidx
if mean_rtf < np.floor(fidx_remainder + mean_rtf):
# wait till next frame
continue
else:
if stride_cnt % opts.det_stride == 0:
stride_cnt = 1
else:
stride_cnt += 1
continue
if opts.cached_res:
img = frame_list[fidx]
if cache_in_ccf:
cache_end_idx, bboxes, scores, labels, masks = \
result_from_ccf(cached_res, img['id'], cache_end_idx)
ltwh2ltrb_(bboxes)
else:
result = cached_res[img['id']]
bboxes, scores, labels, masks = \
parse_det_result(result, coco_mapping, n_class)
else:
frame = frames[fidx]
result = inference_detector(model, frame)
bboxes, scores, labels, masks = \
parse_det_result(result, coco_mapping, n_class)
rt_this = runtime_dist.draw()
t_elapsed += rt_this
if t_elapsed >= t_total:
break
timestamps.append(t_elapsed)
if results_raw is not None:
results_raw.append(result)
results_parsed.append((bboxes, scores, labels, masks))
input_fidx.append(fidx)
runtime.append(rt_this)
out_path = join(opts.out_dir, seq + '.pkl')
if opts.overwrite or not isfile(out_path):
out_dict = {
'results_parsed': results_parsed,
'timestamps': timestamps,
'input_fidx': input_fidx,
'runtime': runtime,
}
if results_raw is not None:
out_dict['results_raw'] = results_raw
pickle.dump(out_dict, open(out_path, 'wb'))
runtime_all += runtime
n_processed += len(results_parsed)
runtime_all_np = np.array(runtime_all)
n_small_runtime = (runtime_all_np < 1.0/opts.fps).sum()
out_path = join(opts.out_dir, 'time_info.pkl')
if opts.overwrite or not isfile(out_path):
pickle.dump({
'runtime_all': runtime_all,
'n_processed': n_processed,
'n_total': n_total,
'n_small_runtime': n_small_runtime,
}, open(out_path, 'wb'))
# convert to ms for display
s2ms = lambda x: 1e3*x
print(f'{n_processed}/{n_total} frames processed')
print_stats(runtime_all_np, 'Runtime (ms)', cvt=s2ms)
print(f'Runtime smaller than unit time interval: '
f'{n_small_runtime}/{n_processed} '
f'({100.0*n_small_runtime/n_processed:.4g}%)')
def main(_):
# Configuration.
num_unrolls = FLAGS.num_steps
if FLAGS.seed:
tf.set_random_seed(FLAGS.seed)
# Problem.
problem, net_config, net_assignments = util.get_config(
FLAGS.problem, FLAGS.path, FLAGS.problem_path)
# Optimizer setup.
if FLAGS.optimizer == "Adam":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
x_op = problem_vars
optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "SGD":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
x_op = problem_vars
optimizer = tf.train.GradientDescentOptimizer(FLAGS.learning_rate)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "RMSProp":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
x_op = problem_vars
optimizer = tf.train.RMSPropOptimizer(FLAGS.learning_rate)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "Momentum":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
x_op = problem_vars
optimizer = tf.train.MomentumOptimizer(FLAGS.learning_rate,
momentum=0.9)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "NAG":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
x_op = problem_vars
optimizer = tf.train.MomentumOptimizer(FLAGS.learning_rate,
momentum=0.1,
use_nesterov=True)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "L2L":
if FLAGS.path is None:
logging.warning("Evaluating untrained L2L optimizer")
optimizer = meta.MetaOptimizer(**net_config)
meta_loss = optimizer.meta_loss(problem,
1,
net_assignments=net_assignments)
_, update, reset, cost_op, x_op = meta_loss
else:
raise ValueError("{} is not a valid optimizer".format(FLAGS.optimizer))
with ms.MonitoredSession() as sess:
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
total_time = 0
total_cost = 0
for _ in xrange(FLAGS.num_epochs):
# Training.
time, cost, x_values = util.run_epoch(sess, cost_op, x_op,
[update], reset, num_unrolls)
total_time += time
total_cost += cost
x_values = np.swapaxes(np.squeeze(x_values), 0, 1)
if FLAGS.problem.find('wav') != -1:
np.save(os.path.join('results', '{}_wav'.format(FLAGS.optimizer)),
x_values)
else:
np.save(os.path.join('results', '{}'.format(FLAGS.optimizer)),
x_values)
# print("x_values shape: {}".format(x_values.shape))
# print("x_values: {}".format(x_values))
# np.savetxt(os.path.join('results', '{}.txt'.format(FLAGS.optimizer)), x_values, fmt='%f')
# Results.
util.print_stats(
"Epoch {}, Optimizer {}".format(FLAGS.num_epochs, FLAGS.optimizer),
total_cost, total_time, FLAGS.num_epochs)
def main(_):
# Configuration.
if FLAGS.seed:
tf.set_random_seed(FLAGS.seed)
# Problem.
problem, net_config, net_assignments = util.get_config(
FLAGS.problem, FLAGS.path)
state = None
# Optimizer setup.
if FLAGS.optimizer == "Adam":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "SGD_MOM":
cost_op = problem()
problem_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
problem_reset = tf.variables_initializer(problem_vars)
optimizer = tf.train.MomentumOptimizer(FLAGS.learning_rate, 0.9)
optimizer_reset = tf.variables_initializer(optimizer.get_slot_names())
update = optimizer.minimize(cost_op)
reset = [problem_reset, optimizer_reset]
elif FLAGS.optimizer == "L2L":
if FLAGS.path is None:
logging.warning("Evaluating untrained L2L optimizer")
cost_op = problem()
optimizer = meta.MetaOptimizer(**net_config)
if FLAGS.load_trained_model:
# optimizer.load_states([pickle.load(open(os.path.join(FLAGS.model_path, "states.p"), "rb"))])
# optimizer.load_states([pickle.load(open("./init_state.p", "rb"))])
meta_loss = optimizer.meta_loss(problem, 1, net_assignments=net_assignments, load_states=False)
# _, update, reset, cost_op, _ = meta_loss
_, update, reset, _, _, state = meta_loss
else:
meta_loss = optimizer.meta_loss(problem, 1, net_assignments=net_assignments, load_states=False)
# _, update, reset, cost_op, _ = meta_loss
_, update, reset, _, _, state = meta_loss
else:
raise ValueError("{} is not a valid optimizer".format(FLAGS.optimizer))
process_id = os.getpid()
exp_folder = os.path.join("exp", str(process_id))
if not os.path.isdir(exp_folder):
os.mkdir(exp_folder)
writer = tf.summary.FileWriter(exp_folder)
summaries = tf.summary.merge_all()
if FLAGS.problem == "mnist":
var_name_mlp = [
"mlp/linear_0/w:0", "mlp/linear_0/b:0", "mlp/linear_1/w:0",
"mlp/linear_1/b:0"
]
else:
var_name_mlp = []
problem_vars = tf.get_collection(tf.GraphKeys.VARIABLES)
if var_name_mlp:
saver_vars = [vv for vv in problem_vars if vv.name in var_name_mlp]
else:
saver_vars = problem_vars
saver = tf.train.Saver(saver_vars)
with ms.MonitoredSession() as sess:
# a quick hack!
regular_sess = sess._sess._sess._sess._sess
# Prevent accidental changes to the graph.
tf.get_default_graph().finalize()
# print("Initial loss = {}".format(sess.run(cost_op)))
# raw_input("wait")
if FLAGS.load_trained_model == True:
print("We are loading trained model here!")
saver.restore(regular_sess, os.path.join(FLAGS.model_path, "model"))
# init_state = regular_sess.run(optimizer.init_state)
# cost_val = regular_sess.run(cost_op)
# import pdb; pdb.set_trace()
total_time = 0
total_cost = 0
for step in xrange(FLAGS.num_epochs):
time, cost = util.run_epoch_eval(
sess,
cost_op,
[update],
reset,
FLAGS.num_steps,
summary_op=summaries,
summary_writer=writer,
run_reset=False)
writer.flush()
total_time += time
total_cost += cost
saver.save(regular_sess, os.path.join(exp_folder, "model"))
pickle.dump(final_states, open(os.path.join(exp_folder, "states.p"), "wb"))
# Results.
util.print_stats("Epoch {}".format(FLAGS.num_epochs), total_cost,
total_time, FLAGS.num_epochs)
# we have to run in the end to skip the error
regular_sess.run(reset)
