class Agent(object):
def __init__(self, param, logger):
self.logger = logger
logger.info("Start initializing Agent, mode is {}".format(param["mode"]))
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.param = param
self.session = tf.Session(config=config)
self.model = Model(self.session, self.param, self.logger) # 建立将模型graph并写入session中
self.data_manager_train = DataManager(self.param, shuffle=True) # 训练集数据生成器
self.data_manager_valid = DataManager(self.param, shuffle=False, valid=True) # 验证集数据生成器
self.tensorboard_manager = TensorboardManager(self.session, self.param, self.model, self.logger) # 使用TensorBoard进行可视化
self.trainer = Trainer(self.session, self.model, self.param, self.logger, self.tensorboard_manager) # 损失函数，优化器，以及训练策略
self.saver = Saver(self.session, self.param, self.model.checkPoint_dir, self.logger) # 用于将session保存到checkpoint或pb文件中
self.validator = Validator(self.session, self.model, self.logger) # 用于验证训练后模型的性能
logger.info("Successfully initialized")
def run(self):
if not self.param["anew"] and self.param["mode"] != "testPb":
self.saver.load_checkpoint()
if self.param["mode"] == "train_segmentation": # 训练模型分割部分
self.trainer.train_segmentation(self.data_manager_train, self.data_manager_valid, self.saver)
elif self.param["mode"] == "train_decision": # 训练模型分类部分
self.trainer.train_decision(self.data_manager_train, self.data_manager_valid, self.saver)
elif self.param["mode"] == "visualization": # 验证模型分割效果
self.validator.valid_segmentation(self.data_manager_train)
elif self.param["mode"] == "testing": # 验证模型分类效果
self.validator.valid_decision(self.data_manager_train)
self.validator.valid_decision(self.data_manager_valid)
elif self.param["mode"] == "savePb": # 保存模型到pb文件
self.saver.save_pb()
elif self.param["mode"] == "testPb": # 测试pb模型效果
self.pb_tester = PbTester(self.param, self.logger)
self.pb_tester.test_segmentation()
self.pb_tester.test_decision()
# self.pb_tester.view_timeline()
elif self.param["mode"] == "view_dataset":
# tensorboard --logdir=E:/CODES/TensorFlow_PZT/tensorboard --samples_per_plugin=images=1000
self.tensorboard_manager.vis_dataset(self.data_manager_train)
self.tensorboard_manager.vis_mask_out(self.data_manager_train)
self.session.close()
self.tensorboard_manager.close()
def main():
args = parse_args()
saver = Saver(args.model_dir)
model = SELUNet()
with warnings.catch_warnings():
warnings.simplefilter('ignore')
if args.use_cuda:
model = model.cuda()
model, _, params_dict = saver.load_checkpoint(
model, file_name=args.model_name)
model.eval()
filespan = args.filespan
idr_params, _, _ = get_optimal_params(model,
args.valspeechfolder,
args.valpeaksfolder,
args.window,
args.stride,
filespan,
numfiles=60,
use_cuda=False,
thlist=[
0.15, 0.2, 0.25, 0.3, 0.35, 0.4,
0.45, 0.5, 0.55, 0.6, 0.65, 0.7,
0.75
],
spblist=[25],
hctlist=[10, 15, 20, 25, 30])
thr = idr_params['thr']
spb = idr_params['spb']
hct = idr_params['hct']
with open('test_idr.txt', 'w') as f:
print(
'Optimal Hyperparameters\nThreshold: {} Samples Per Bin: {} Histogram Count Threshold: {}'
.format(thr, spb, hct),
file=f,
flush=True)
numfiles = len(glob(os.path.join(args.speechfolder, '*.npy')))
print('Models and Files Loaded')
metrics_list = []
for i in range(0, numfiles, filespan):
if (i + filespan) > numfiles:
break
speech_windowed_data, peak_distance, peak_indicator, indices, actual_gci_locations = create_dataset(
args.speechfolder, args.peaksfolder, args.window, args.stride,
slice(i, i + filespan))
input = to_variable(
th.from_numpy(
np.expand_dims(speech_windowed_data, 1).astype(np.float32)),
args.use_cuda, True)
with warnings.catch_warnings():
prediction = model(input)
predicted_peak_indicator = F.sigmoid(prediction[:, 1]).data.numpy()
predicted_peak_distance = (prediction[:,
0]).data.numpy().astype(np.int32)
predicted_peak_indicator_indices = predicted_peak_indicator > args.threshold
predicted_peak_indicator = predicted_peak_indicator[
predicted_peak_indicator_indices].ravel()
predicted_peak_distance = predicted_peak_distance[
predicted_peak_indicator_indices].ravel()
indices = indices[predicted_peak_indicator_indices]
assert (len(indices) == len(predicted_peak_distance))
assert (len(predicted_peak_distance) == len(predicted_peak_indicator))
positive_distance_indices = predicted_peak_distance < args.window
positive_peak_distances = predicted_peak_distance[
positive_distance_indices]
postive_predicted_peak_indicator = predicted_peak_indicator[
positive_distance_indices]
gci_locations = [
indices[i, d] for i, d in enumerate(positive_peak_distances)
]
locations_true = np.nonzero(actual_gci_locations)[0]
xaxes = np.zeros(len(actual_gci_locations))
xaxes[locations_true] = 1
ground_truth = np.row_stack(
(np.arange(len(actual_gci_locations)), xaxes))
predicted_truth = np.row_stack(
(gci_locations, postive_predicted_peak_indicator))
gx = ground_truth[0, :]
gy = ground_truth[1, :]
px = predicted_truth[0, :]
py = predicted_truth[1, :]
fs = 16000
gci = np.array(
cluster(px,
py,
threshold=thr,
samples_per_bin=spb,
histogram_count_threshold=hct))
predicted_gci_time = gci / fs
target_gci_time = np.nonzero(gy)[0] / fs
gci = np.round(gci).astype(np.int64)
gcilocs = np.zeros_like(gx)
gcilocs[gci] = 1
metric = corrected_naylor_metrics(target_gci_time, predicted_gci_time)
print(metric)
metrics_list.append(metric)
idr = np.mean([
v for m in metrics_list for k, v in m.items()
if k == 'identification_rate'
])
mr = np.mean(
[v for m in metrics_list for k, v in m.items() if k == 'miss_rate'])
far = np.mean([
v for m in metrics_list for k, v in m.items()
if k == 'false_alarm_rate'
])
se = np.mean([
v for m in metrics_list for k, v in m.items()
if k == 'identification_accuracy'
])
print('IDR: {:.5f} MR: {:.5f} FAR: {:.5f} IDA: {:.5f}'.format(
idr, mr, far, se))
with open('test_idr.txt', 'a') as f:
f.write('IDR: {:.5f} MR: {:.5f} FAR: {:.5f} IDA: {:.5f}\n'.format(
idr, mr, far, se))
def main():
args = parse_args()
speech_windowed_data, peak_distance, peak_indicator, indices, actual_gci_locations = create_dataset(
args.speechfolder, args.peaksfolder, args.window, args.stride, 10)
saver = Saver(args.model_dir)
model = SELUWeightNet
model, _, params_dict = saver.load_checkpoint(
model, file_name=args.model_name)
model.eval()
input = to_variable(
th.from_numpy(
np.expand_dims(speech_windowed_data, 1).astype(np.float32)),
args.use_cuda, True)
with warnings.catch_warnings():
if args.use_cuda:
model = model.cuda()
warnings.simplefilter('ignore')
prediction = model(input)
predicted_peak_indicator = F.sigmoid(prediction[:, 1]).data.numpy()
predicted_peak_distance = (prediction[:, 0]).data.numpy().astype(np.int32)
predicted_peak_indicator_indices = predicted_peak_indicator > args.threshold
predicted_peak_indicator = predicted_peak_indicator[
predicted_peak_indicator_indices].ravel()
predicted_peak_distance = predicted_peak_distance[
predicted_peak_indicator_indices].ravel()
indices = indices[predicted_peak_indicator_indices]
assert (len(indices) == len(predicted_peak_distance))
assert (len(predicted_peak_distance) == len(predicted_peak_indicator))
positive_distance_indices = predicted_peak_distance < args.window
positive_peak_distances = predicted_peak_distance[
positive_distance_indices]
postive_predicted_peak_indicator = predicted_peak_indicator[
positive_distance_indices]
print('Neg Peaks: {} Pos Peaks: {}'.format(
len(predicted_peak_distance) - len(positive_peak_distances),
len(positive_peak_distances)))
gci_locations = [
indices[i, d] for i, d in enumerate(positive_peak_distances)
]
locations_true = np.nonzero(actual_gci_locations)[0]
xaxes = np.zeros(len(actual_gci_locations))
xaxes[locations_true] = 1
if __debug__:
ground_truth = np.row_stack((np.arange(len(actual_gci_locations)),
xaxes))
predicted_truth = np.row_stack((gci_locations,
postive_predicted_peak_indicator))
os.makedirs(args.prediction_dir, exist_ok=True)
np.save(
os.path.join(args.prediction_dir, 'ground_truth'), ground_truth)
np.save(
os.path.join(args.prediction_dir, 'predicted'), predicted_truth)
plt.scatter(
gci_locations,
postive_predicted_peak_indicator,
color='b',
label='Predicted GCI')
plt.plot(
np.arange(len(actual_gci_locations)),
xaxes,
color='r',
label='Actual GCI')
plt.legend()
plt.show()
class Trainer:
def __init__(self):
self._setup_logger()
torch.backends.cudnn.deterministic = True
self.device = torch.device(
"cuda:%d" % (args.gpu) if torch.cuda.is_available() else "cpu")
torch.cuda.set_device(args.gpu)
self.dataset = Cifar10()
self.saver = Saver(self.save_dir)
if args.arch == 'auto':
self.Net = lambda: AutoNetwork(args.init_channels, self.dataset.
num_classes, args.layers,
nn.CrossEntropyLoss())
elif args.arch == 'VGG11':
self.Net = lambda: VGG('VGG11', self.dataset.num_classes,
nn.CrossEntropyLoss())
elif args.arch == 'mobilenet':
self.Net = lambda: MobileNetV2(self.dataset.num_classes,
nn.CrossEntropyLoss())
elif args.arch == 'simple':
self.Net = lambda: SimpleNetwork(self.dataset.num_classes,
nn.CrossEntropyLoss())
else:
raise Exception('Net not defined!')
def _setup_logger(self):
self.save_dir = os.path.join('./checkpoints', args.logdir)
if not os.path.isdir(self.save_dir):
os.makedirs(self.save_dir)
log_path = os.path.join(
self.save_dir,
datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S'))
self.logger = L.create_logger(args.logdir, log_path)
for arg in vars(args):
self.logger.info("%-25s %-20s" % (arg, getattr(args, arg)))
def test_beta(self, optimizee):
self.logger.info('Testing beta')
torch.manual_seed(2 * args.seed)
optimizee.reset_model_parameters()
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizee.optimizer,
float(args.max_epoch),
eta_min=args.learning_rate_min)
optimizee.optimizer.print_beta_greedy()
for epoch in range(0, args.max_test_epoch):
scheduler.step()
lr = scheduler.get_lr()[0]
self.train_epoch(epoch, lr, optimizee, False)
if (epoch + 1) % args.test_freq == 0:
acc = self.eval(optimizee.model)
def train(self):
optimizee = Optimizee(self.Net)
start_episode, start_epoch, best_test_acc = 0, 0, 0.
if args.checkpoint != '':
start_episode, start_epoch = self.saver.load_checkpoint(
optimizee, args.checkpoint)
# Hao: in my implementation, the optimizers for alpha and beta will preserve their states across episodes
for episode in range(start_episode, args.max_episodes):
torch.manual_seed(args.seed + episode)
optimizee.reset_model_parameters()
if args.arch == 'auto':
optimizee.reset_arch_parameters()
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
optimizee.optimizer,
float(args.max_epoch),
eta_min=args.learning_rate_min,
last_epoch=start_epoch - 1)
for epoch in range(
start_epoch,
args.max_epoch): # loop over the dataset multiple times
scheduler.step()
lr = scheduler.get_lr()[0]
training_status = self.train_epoch(epoch, lr, optimizee)
if (epoch + 1) % args.test_freq == 0:
acc = self.eval(optimizee.model)
# if accuracy not above random, discard this example and start a new one
if acc <= 0.11 or training_status == False:
self.logger.warning(
'training_status false or acc too low, break')
break
checkpoint_path = self.save_dir + '/episode_{}_epoch_{}_acc_{}'.format(
str(episode), str(epoch), str(acc))
self.saver.save_checkpoint(optimizee, epoch, episode,
checkpoint_path)
return optimizee
def train_epoch(self, epoch, lr, optimizee, train_beta=True):
if train_beta:
optimizee.optimizer.beta_temp = np.interp(
epoch + 1, [0, args.max_epoch],
[args.min_beta_temp, args.max_beta_temp])
status = self._train_epoch(epoch, lr, optimizee)
self.saver.write_beta_embedding()
optimizee.optimizer.print_beta_greedy()
else:
optimizee.optimizer.beta_temp = args.max_beta_temp
status = self._train_epoch_fix_beta(epoch, lr, optimizee)
return status
def eval(self, model):
correct, total = 0, 0
with torch.no_grad():
for images, labels in self.dataset.test_loader:
images, labels = images.to(self.device), labels.to(self.device)
outputs = model(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
self.logger.info(
'Accuracy of the network on the 10000 test images: %d %%' %
(100 * correct / total))
return float(correct) / total
def _train_epoch(self, epoch, lr, optimizee):
losses, next_step_losses, meta_update_losses = 0.0, 0.0, 0.0
model_grad_norms, beta_grad_norms = 0., 0.
for i, data in enumerate(self.dataset.train_loader, 0):
x_train, y_train = data[0].to(self.device), data[1].to(self.device)
val_data = next(iter(self.dataset.val_loader))
x_val, y_val = val_data[0].to(self.device), val_data[1].to(
self.device)
# Derive \frac{\partial L}{\partial \theta}
optimizee.model.zero_grad()
loss = optimizee.model.forward_pass(x_train, y_train)
loss.backward()
model_grad_norms += nn.utils.clip_grad_norm_(
optimizee.model.parameters(), args.model_grad_norm)
param_updates = optimizee.optimizer.step()
# do a differentiable step of update over optimizee.symbolic_model
optimizee.differentiable_update(param_updates)
# check the next-step loss after update using the symbolic model
# so that it is on the computational graph
# TODO use validation data? not sure
next_step_loss = optimizee.symbolic_model.forward_pass(
x_train, y_train)
if (math.isnan(next_step_loss.item())):
self.logger.error('next step loss becomes NaN, break')
raise Exception
next_step_losses += next_step_loss.item()
losses += loss.item()
meta_update_losses += next_step_loss
# do a non-differentiable update over optimizee.model if next_step_loss is smaller
if (next_step_loss.item() <
loss.item()): # This is still important for performance
optimizee.update(param_updates)
# forsee bptt_steps then update beta
if (i + 1) % args.bptt_step == 0:
beta_grad_norms += optimizee.beta_step(
meta_update_losses).item()
# let saver save the grads for beta
with torch.no_grad():
self.saver.add_beta_grads([
b.grad.abs().mean() for b in optimizee.optimizer.beta
])
meta_update_losses = 0
optimizee.sync_symbolic_model()
if args.arch_training and (i + 1) % args.update_alpha_step == 0:
optimizee.alpha_step(
x_train, y_train, x_val, y_val,
lr) # TODO: make sure alpha step won't change weights
optimizee.sync_symbolic_model(skip_weights=True)
if (i + 1) % args.log_freq == 0: # print every 2000 mini-batches
beta_out = optimizee.optimizer.beta[-1].data.cpu().numpy()
alpha_out = optimizee.model.alphas_normal[-1].data.cpu().numpy(
) if args.arch_training else ''
self.logger.info(
'[%d, %5d] loss: %.3f/%.3f, next step loss: %.3f, beta[-1]/L2(g): %s/%s, alpha[-1]: %s'
% (epoch + 1, i + 1, losses / args.log_freq,
model_grad_norms / args.log_freq,
next_step_losses / args.log_freq, beta_out,
beta_grad_norms / args.log_freq, alpha_out))
losses, next_step_losses, model_grad_norms, beta_grad_norms = 0., 0., 0., 0.
self.saver.write_beta(optimizee.optimizer.beta)
return True
def _train_epoch_fix_beta(self, epoch, lr, optimizee):
losses, model_grad_norms = 0.0, 0.0
for i, data in enumerate(self.dataset.train_loader, 0):
x_train, y_train = data[0].to(self.device), data[1].to(self.device)
val_data = next(iter(self.dataset.val_loader))
x_val, y_val = val_data[0].to(self.device), val_data[1].to(
self.device)
optimizee.model.zero_grad()
loss = optimizee.model.forward_pass(x_train, y_train)
loss.backward()
model_grad_norms += nn.utils.clip_grad_norm_(
optimizee.model.parameters(), args.model_grad_norm)
param_updates = optimizee.optimizer.step(do_update=True)
losses += loss.item()
if args.use_darts_arch and args.train_alpha:
optimizee.alpha_step(x_train, y_train, x_val, y_val, lr)
if (i + 1) % args.log_freq == 0: # print every 2000 mini-batches
beta_out = optimizee.optimizer.beta[-1].data.cpu().numpy()
alpha_out = optimizee.model.alphas_normal[-1].data.cpu().numpy(
) if args.arch_training else 'None'
self.logger.info(
'[%d, %5d] loss: %.3f/%.3f, beta[-1]: %s, alpha[-1]: %s' %
(epoch + 1, i + 1, losses / args.log_freq,
model_grad_norms / args.log_freq, beta_out, alpha_out))
losses, model_grad_norms = 0., 0.
return True
