class VisualDLWriter(Callback):
"""
Use VisualDL to log data or image
"""
def __init__(self, model):
super(VisualDLWriter, self).__init__(model)
assert six.PY3, "VisualDL requires Python >= 3.5"
try:
from visualdl import LogWriter
except Exception as e:
logger.error('visualdl not found, plaese install visualdl. '
'for example: `pip install visualdl`.')
raise e
self.vdl_writer = LogWriter(
model.cfg.get('vdl_log_dir', 'vdl_log_dir/scalar'))
self.vdl_loss_step = 0
self.vdl_mAP_step = 0
self.vdl_image_step = 0
self.vdl_image_frame = 0
def on_step_end(self, status):
mode = status['mode']
if dist.get_world_size() < 2 or dist.get_rank() == 0:
if mode == 'train':
training_staus = status['training_staus']
for loss_name, loss_value in training_staus.get().items():
self.vdl_writer.add_scalar(loss_name, loss_value,
self.vdl_loss_step)
self.vdl_loss_step += 1
elif mode == 'test':
ori_image = status['original_image']
result_image = status['result_image']
self.vdl_writer.add_image(
"original/frame_{}".format(self.vdl_image_frame),
ori_image, self.vdl_image_step)
self.vdl_writer.add_image(
"result/frame_{}".format(self.vdl_image_frame),
result_image, self.vdl_image_step)
self.vdl_image_step += 1
# each frame can display ten pictures at most.
if self.vdl_image_step % 10 == 0:
self.vdl_image_step = 0
self.vdl_image_frame += 1
def on_epoch_end(self, status):
mode = status['mode']
if dist.get_world_size() < 2 or dist.get_rank() == 0:
if mode == 'eval':
for metric in self.model._metrics:
for key, map_value in metric.get_results().items():
self.vdl_writer.add_scalar("{}-mAP".format(key),
map_value[0],
self.vdl_mAP_step)
self.vdl_mAP_step += 1
def main():
cfg = load_config(FLAGS.config)
merge_config(FLAGS.opt)
check_config(cfg)
# check if set use_gpu=True in paddlepaddle cpu version
check_gpu(cfg.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
main_arch = cfg.architecture
dataset = cfg.TestReader['dataset']
test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img)
dataset.set_images(test_images)
place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
model = create(main_arch)
startup_prog = fluid.Program()
infer_prog = fluid.Program()
with fluid.program_guard(infer_prog, startup_prog):
with fluid.unique_name.guard():
inputs_def = cfg['TestReader']['inputs_def']
inputs_def['iterable'] = True
feed_vars, loader = model.build_inputs(**inputs_def)
test_fetches = model.test(feed_vars)
infer_prog = infer_prog.clone(True)
reader = create_reader(cfg.TestReader, devices_num=1)
loader.set_sample_list_generator(reader, place)
exe.run(startup_prog)
if cfg.weights:
checkpoint.load_params(exe, infer_prog, cfg.weights)
# parse infer fetches
assert cfg.metric in ['COCO', 'VOC', 'OID', 'WIDERFACE'], \
"unknown metric type {}".format(cfg.metric)
extra_keys = []
if cfg['metric'] in ['COCO', 'OID']:
extra_keys = ['im_info', 'im_id', 'im_shape']
if cfg['metric'] == 'VOC' or cfg['metric'] == 'WIDERFACE':
extra_keys = ['im_id', 'im_shape']
keys, values, _ = parse_fetches(test_fetches, infer_prog, extra_keys)
# parse dataset category
if cfg.metric == 'COCO':
from ppdet.utils.coco_eval import bbox2out, mask2out, segm2out, get_category_info
if cfg.metric == 'OID':
from ppdet.utils.oid_eval import bbox2out, get_category_info
if cfg.metric == "VOC":
from ppdet.utils.voc_eval import bbox2out, get_category_info
if cfg.metric == "WIDERFACE":
from ppdet.utils.widerface_eval_utils import bbox2out, lmk2out, get_category_info
anno_file = dataset.get_anno()
with_background = dataset.with_background
use_default_label = dataset.use_default_label
clsid2catid, catid2name = get_category_info(anno_file, with_background,
use_default_label)
# whether output bbox is normalized in model output layer
is_bbox_normalized = False
if hasattr(model, 'is_bbox_normalized') and \
callable(model.is_bbox_normalized):
is_bbox_normalized = model.is_bbox_normalized()
# use VisualDL to log image
if FLAGS.use_vdl:
assert six.PY3, "VisualDL requires Python >= 3.5"
from visualdl import LogWriter
vdl_writer = LogWriter(FLAGS.vdl_log_dir)
vdl_image_step = 0
vdl_image_frame = 0 # each frame can display ten pictures at most.
imid2path = dataset.get_imid2path()
resultBBox = []
for iter_id, data in enumerate(loader()):
outs = exe.run(infer_prog,
feed=data,
fetch_list=values,
return_numpy=False)
res = {
k: (np.array(v), v.recursive_sequence_lengths())
for k, v in zip(keys, outs)
}
logger.info('Infer iter {}'.format(iter_id))
if 'TTFNet' in cfg.architecture:
res['bbox'][1].append([len(res['bbox'][0])])
if 'CornerNet' in cfg.architecture:
from ppdet.utils.post_process import corner_post_process
post_config = getattr(cfg, 'PostProcess', None)
corner_post_process(res, post_config, cfg.num_classes)
bbox_results = None
mask_results = None
segm_results = None
lmk_results = None
if 'bbox' in res:
bbox_results = bbox2out([res], clsid2catid, is_bbox_normalized)
if 'mask' in res:
mask_results = mask2out([res], clsid2catid,
model.mask_head.resolution)
if 'segm' in res:
segm_results = segm2out([res], clsid2catid)
if 'landmark' in res:
lmk_results = lmk2out([res], is_bbox_normalized)
# bbox 四个值：左上角坐标 + 宽度 + 高度
# {'image_id': 0, 'category_id': 0, 'bbox': [695.04443359375, 723.8153686523438, 128.288818359375, 61.5987548828125], 'score': 0.9990022778511047}
im_ids = res['im_id'][0]
image_path = imid2path[int(im_ids[0])]
prefix = image_path.split('/')[-1]
imageName = prefix.split('.')[0]
for i, result in enumerate(bbox_results):
score = result["score"]
bbox = result["bbox"]
x1 = str(int(bbox[0]))
y1 = str(int(bbox[1]))
x2 = str(int(bbox[2] + bbox[0]))
y2 = str(int(bbox[3] + bbox[1]))
if (score > 0.01):
resStr = imageName + ' ' + str(round(
score,
3)) + ' ' + x1 + ' ' + y1 + ' ' + x2 + ' ' + y2 + '\n'
resultBBox.append(resStr)
# visualize result
for im_id in im_ids:
image_path = imid2path[int(im_id)]
image = Image.open(image_path).convert('RGB')
image = ImageOps.exif_transpose(image)
# use VisualDL to log original image
if FLAGS.use_vdl:
original_image_np = np.array(image)
vdl_writer.add_image(
"original/frame_{}".format(vdl_image_frame),
original_image_np, vdl_image_step)
image = visualize_results(image, int(im_id), catid2name,
FLAGS.draw_threshold, bbox_results,
mask_results, segm_results, lmk_results)
# use VisualDL to log image with bbox
if FLAGS.use_vdl:
infer_image_np = np.array(image)
vdl_writer.add_image("bbox/frame_{}".format(vdl_image_frame),
infer_image_np, vdl_image_step)
vdl_image_step += 1
if vdl_image_step % 10 == 0:
vdl_image_step = 0
vdl_image_frame += 1
save_name = get_save_image_name(FLAGS.output_dir, image_path)
logger.info("Detection bbox results save in {}".format(save_name))
image.save(save_name, quality=95)
resulttxtPath = "/home/aistudio/work/PaddleDetection-release-2.0-beta/output/test_result.txt"
f = open(resulttxtPath, 'w+', encoding='utf-8')
for i, p in enumerate(resultBBox):
f.write(p)
f.close()
def main():
cfg = load_config(FLAGS.config)
merge_config(FLAGS.opt)
check_config(cfg)
# check if set use_gpu=True in paddlepaddle cpu version
check_gpu(cfg.use_gpu)
# check if paddlepaddle version is satisfied
check_version()
main_arch = cfg.architecture
dataset = cfg.TestReader['dataset']
test_images = get_test_images(FLAGS.infer_dir, FLAGS.infer_img)
dataset.set_images(test_images)
place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
model = create(main_arch)
startup_prog = fluid.Program()
infer_prog = fluid.Program()
with fluid.program_guard(infer_prog, startup_prog):
with fluid.unique_name.guard():
inputs_def = cfg['TestReader']['inputs_def']
inputs_def['iterable'] = True
feed_vars, loader = model.build_inputs(**inputs_def)
test_fetches = model.test(feed_vars)
infer_prog = infer_prog.clone(True)
reader = create_reader(cfg.TestReader, devices_num=1)
loader.set_sample_list_generator(reader, place)
exe.run(startup_prog)
if cfg.weights:
checkpoint.load_params(exe, infer_prog, cfg.weights)
# parse infer fetches
assert cfg.metric in ['COCO', 'VOC', 'OID', 'WIDERFACE'], \
"unknown metric type {}".format(cfg.metric)
extra_keys = []
if cfg['metric'] in ['COCO', 'OID']:
extra_keys = ['im_info', 'im_id', 'im_shape']
if cfg['metric'] == 'VOC' or cfg['metric'] == 'WIDERFACE':
extra_keys = ['im_id', 'im_shape']
keys, values, _ = parse_fetches(test_fetches, infer_prog, extra_keys)
# parse dataset category
if cfg.metric == 'COCO':
from ppdet.utils.coco_eval import bbox2out, mask2out, get_category_info
if cfg.metric == 'OID':
from ppdet.utils.oid_eval import bbox2out, get_category_info
if cfg.metric == "VOC":
from ppdet.utils.voc_eval import bbox2out, get_category_info
if cfg.metric == "WIDERFACE":
from ppdet.utils.widerface_eval_utils import bbox2out, lmk2out, get_category_info
anno_file = dataset.get_anno()
with_background = dataset.with_background
use_default_label = dataset.use_default_label
clsid2catid, catid2name = get_category_info(anno_file, with_background,
use_default_label)
# whether output bbox is normalized in model output layer
is_bbox_normalized = False
if hasattr(model, 'is_bbox_normalized') and \
callable(model.is_bbox_normalized):
is_bbox_normalized = model.is_bbox_normalized()
# use VisualDL to log image
if FLAGS.use_vdl:
assert six.PY3, "VisualDL requires Python >= 3.5"
from visualdl import LogWriter
vdl_writer = LogWriter(FLAGS.vdl_log_dir)
vdl_image_step = 0
vdl_image_frame = 0 # each frame can display ten pictures at most.
imid2path = dataset.get_imid2path()
for iter_id, data in enumerate(loader()):
outs = exe.run(infer_prog,
feed=data,
fetch_list=values,
return_numpy=False)
res = {
k: (np.array(v), v.recursive_sequence_lengths())
for k, v in zip(keys, outs)
}
logger.info('Infer iter {}'.format(iter_id))
if 'TTFNet' in cfg.architecture:
res['bbox'][1].append([len(res['bbox'][0])])
bbox_results = None
mask_results = None
lmk_results = None
if 'bbox' in res:
bbox_results = bbox2out([res], clsid2catid, is_bbox_normalized)
if 'mask' in res:
mask_results = mask2out([res], clsid2catid,
model.mask_head.resolution)
if 'landmark' in res:
lmk_results = lmk2out([res], is_bbox_normalized)
# visualize result
im_ids = res['im_id'][0]
for im_id in im_ids:
image_path = imid2path[int(im_id)]
image = Image.open(image_path).convert('RGB')
# use VisualDL to log original image
if FLAGS.use_vdl:
original_image_np = np.array(image)
vdl_writer.add_image(
"original/frame_{}".format(vdl_image_frame),
original_image_np, vdl_image_step)
image = visualize_results(image,
int(im_id), catid2name,
FLAGS.draw_threshold, bbox_results,
mask_results, lmk_results)
# use VisualDL to log image with bbox
if FLAGS.use_vdl:
infer_image_np = np.array(image)
vdl_writer.add_image("bbox/frame_{}".format(vdl_image_frame),
infer_image_np, vdl_image_step)
vdl_image_step += 1
if vdl_image_step % 10 == 0:
vdl_image_step = 0
vdl_image_frame += 1
save_name = get_save_image_name(FLAGS.output_dir, image_path)
logger.info("Detection bbox results save in {}".format(save_name))
image.save(save_name, quality=95)
def train(self):
if not self.is_parallel:
writer = LogWriter(logdir=self.result_dir + "/log/")
self.genA2B.train(), self.genB2A.train(), self.disGA.train(
), self.disGB.train(), self.disLA.train(), self.disLB.train()
start_iter = 1
if self.resume:
print(self.result_dir, self.dataset,
os.path.join(self.result_dir, self.dataset, 'model', '*.pt'))
model_list = glob(
os.path.join(self.result_dir, self.dataset, 'model', '*.pt'))
print("resuming, model_list", model_list)
if not len(model_list) == 0:
model_list.sort()
start_iter = int(model_list[-1].split('_')[-1].split('.')[0])
print("resuming, start_iter", start_iter)
self.load(os.path.join(self.result_dir, self.dataset, 'model'),
start_iter)
print(" [*] Load SUCCESS")
if self.decay_flag and start_iter > (self.iteration // 2):
self.G_optim._learning_rate -= (self.lr /
(self.iteration // 2)) * (
start_iter -
self.iteration // 2)
self.D_optim._learning_rate -= (self.lr /
(self.iteration // 2)) * (
start_iter -
self.iteration // 2)
# training loop
print('training start !')
start_time = time.time()
for step in range(start_iter, self.iteration + 1):
if self.decay_flag and step > (self.iteration // 2):
self.G_optim._learning_rate -= (self.lr /
(self.iteration // 2))
self.D_optim._learning_rate -= (self.lr /
(self.iteration // 2))
try:
real_A, _ = trainA_iter.next()
except:
trainA_iter = iter(self.trainA_loader)
real_A, _ = trainA_iter.next()
try:
real_B, _ = trainB_iter.next()
except:
trainB_iter = iter(self.trainB_loader)
real_B, _ = trainB_iter.next()
real_A = real_A[0]
real_B = real_B[0] ##some handling needed using paddle dataloader
# Update D
if hasattr(self.D_optim, "_optimizer"): # support meta optimizer
self.D_optim._optimizer.clear_gradients()
else:
self.D_optim.clear_gradients()
fake_A2B, _, _ = self.genA2B(real_A)
fake_B2A, _, _ = self.genB2A(real_B)
real_GA_logit, real_GA_cam_logit, _ = self.disGA(real_A)
real_LA_logit, real_LA_cam_logit, _ = self.disLA(real_A)
real_GB_logit, real_GB_cam_logit, _ = self.disGB(real_B)
real_LB_logit, real_LB_cam_logit, _ = self.disLB(real_B)
fake_GA_logit, fake_GA_cam_logit, _ = self.disGA(fake_B2A)
fake_LA_logit, fake_LA_cam_logit, _ = self.disLA(fake_B2A)
fake_GB_logit, fake_GB_cam_logit, _ = self.disGB(fake_A2B)
fake_LB_logit, fake_LB_cam_logit, _ = self.disLB(fake_A2B)
D_ad_loss_GA = self.MSE_loss(
real_GA_logit,
torch.ones_like(real_GA_logit).to(
self.device)) + self.MSE_loss(
fake_GA_logit,
torch.zeros_like(fake_GA_logit).to(self.device))
D_ad_cam_loss_GA = self.MSE_loss(
real_GA_cam_logit,
torch.ones_like(real_GA_cam_logit).to(
self.device)) + self.MSE_loss(
fake_GA_cam_logit,
torch.zeros_like(fake_GA_cam_logit).to(self.device))
D_ad_loss_LA = self.MSE_loss(
real_LA_logit,
torch.ones_like(real_LA_logit).to(
self.device)) + self.MSE_loss(
fake_LA_logit,
torch.zeros_like(fake_LA_logit).to(self.device))
D_ad_cam_loss_LA = self.MSE_loss(
real_LA_cam_logit,
torch.ones_like(real_LA_cam_logit).to(
self.device)) + self.MSE_loss(
fake_LA_cam_logit,
torch.zeros_like(fake_LA_cam_logit).to(self.device))
D_ad_loss_GB = self.MSE_loss(
real_GB_logit,
torch.ones_like(real_GB_logit).to(
self.device)) + self.MSE_loss(
fake_GB_logit,
torch.zeros_like(fake_GB_logit).to(self.device))
D_ad_cam_loss_GB = self.MSE_loss(
real_GB_cam_logit,
torch.ones_like(real_GB_cam_logit).to(
self.device)) + self.MSE_loss(
fake_GB_cam_logit,
torch.zeros_like(fake_GB_cam_logit).to(self.device))
D_ad_loss_LB = self.MSE_loss(
real_LB_logit,
torch.ones_like(real_LB_logit).to(
self.device)) + self.MSE_loss(
fake_LB_logit,
torch.zeros_like(fake_LB_logit).to(self.device))
D_ad_cam_loss_LB = self.MSE_loss(
real_LB_cam_logit,
torch.ones_like(real_LB_cam_logit).to(
self.device)) + self.MSE_loss(
fake_LB_cam_logit,
torch.zeros_like(fake_LB_cam_logit).to(self.device))
D_loss_A = self.adv_weight * (D_ad_loss_GA + D_ad_cam_loss_GA +
D_ad_loss_LA +
D_ad_cam_loss_LA) / self.n_gpu
D_loss_B = self.adv_weight * (D_ad_loss_GB + D_ad_cam_loss_GB +
D_ad_loss_LB +
D_ad_cam_loss_LB) / self.n_gpu
Discriminator_loss = D_loss_A + D_loss_B
Discriminator_loss.backward()
if self.is_parallel:
self.disGA.apply_collective_grads()
self.disGB.apply_collective_grads()
self.disLA.apply_collective_grads()
self.disLB.apply_collective_grads()
self.genA2B.apply_collective_grads()
self.genB2A.apply_collective_grads()
self.D_optim.minimize(Discriminator_loss)
# Update G
if hasattr(self.G_optim, "_optimizer"): # support meta optimizer
self.G_optim._optimizer.clear_gradients()
else:
self.G_optim.clear_gradients()
fake_A2B, fake_A2B_cam_logit, _ = self.genA2B(real_A)
fake_B2A, fake_B2A_cam_logit, _ = self.genB2A(real_B)
fake_A2B2A, _, _ = self.genB2A(fake_A2B)
fake_B2A2B, _, _ = self.genA2B(fake_B2A)
fake_A2A, fake_A2A_cam_logit, _ = self.genB2A(real_A)
fake_B2B, fake_B2B_cam_logit, _ = self.genA2B(real_B)
fake_GA_logit, fake_GA_cam_logit, _ = self.disGA(fake_B2A)
fake_LA_logit, fake_LA_cam_logit, _ = self.disLA(fake_B2A)
fake_GB_logit, fake_GB_cam_logit, _ = self.disGB(fake_A2B)
fake_LB_logit, fake_LB_cam_logit, _ = self.disLB(fake_A2B)
G_ad_loss_GA = self.MSE_loss(
fake_GA_logit,
torch.ones_like(fake_GA_logit).to(self.device))
G_ad_cam_loss_GA = self.MSE_loss(
fake_GA_cam_logit,
torch.ones_like(fake_GA_cam_logit).to(self.device))
G_ad_loss_LA = self.MSE_loss(
fake_LA_logit,
torch.ones_like(fake_LA_logit).to(self.device))
G_ad_cam_loss_LA = self.MSE_loss(
fake_LA_cam_logit,
torch.ones_like(fake_LA_cam_logit).to(self.device))
G_ad_loss_GB = self.MSE_loss(
fake_GB_logit,
torch.ones_like(fake_GB_logit).to(self.device))
G_ad_cam_loss_GB = self.MSE_loss(
fake_GB_cam_logit,
torch.ones_like(fake_GB_cam_logit).to(self.device))
G_ad_loss_LB = self.MSE_loss(
fake_LB_logit,
torch.ones_like(fake_LB_logit).to(self.device))
G_ad_cam_loss_LB = self.MSE_loss(
fake_LB_cam_logit,
torch.ones_like(fake_LB_cam_logit).to(self.device))
G_recon_loss_A = self.L1_loss(fake_A2B2A, real_A)
G_recon_loss_B = self.L1_loss(fake_B2A2B, real_B)
G_identity_loss_A = self.L1_loss(fake_A2A, real_A)
G_identity_loss_B = self.L1_loss(fake_B2B, real_B)
G_cam_loss_A = self.BCE_loss(
fake_B2A_cam_logit,
torch.ones_like(fake_B2A_cam_logit).to(
self.device)) + self.BCE_loss(
fake_A2A_cam_logit,
torch.zeros_like(fake_A2A_cam_logit).to(self.device))
G_cam_loss_B = self.BCE_loss(
fake_A2B_cam_logit,
torch.ones_like(fake_A2B_cam_logit).to(
self.device)) + self.BCE_loss(
fake_B2B_cam_logit,
torch.zeros_like(fake_B2B_cam_logit).to(self.device))
G_loss_A = (self.adv_weight *
(G_ad_loss_GA + G_ad_cam_loss_GA + G_ad_loss_LA +
G_ad_cam_loss_LA) + self.cycle_weight * G_recon_loss_A
+ self.identity_weight * G_identity_loss_A +
self.cam_weight * G_cam_loss_A) / self.n_gpu
G_loss_B = (self.adv_weight *
(G_ad_loss_GB + G_ad_cam_loss_GB + G_ad_loss_LB +
G_ad_cam_loss_LB) + self.cycle_weight * G_recon_loss_B
+ self.identity_weight * G_identity_loss_B +
self.cam_weight * G_cam_loss_B) / self.n_gpu
Generator_loss = G_loss_A + G_loss_B
Generator_loss.backward()
if self.is_parallel:
self.disGA.apply_collective_grads()
self.disGB.apply_collective_grads()
self.disLA.apply_collective_grads()
self.disLB.apply_collective_grads()
self.genA2B.apply_collective_grads()
self.genB2A.apply_collective_grads()
self.G_optim.minimize(Generator_loss)
# clip parameter of AdaILN and ILN, applied after optimizer step
self.Rho_clipper(self.genA2B)
self.Rho_clipper(self.genB2A)
if not self.is_parallel:
writer.add_scalar(tag="G/G_loss_A",
step=step,
value=G_loss_A.numpy())
writer.add_scalar(tag="G/G_loss_B",
step=step,
value=G_loss_B.numpy())
writer.add_scalar(tag="D/D_loss_A",
step=step,
value=D_loss_A.numpy())
writer.add_scalar(tag="D/D_loss_B",
step=step,
value=D_loss_B.numpy())
writer.add_scalar(tag="D/Discriminator_loss",
step=step,
value=Discriminator_loss.numpy())
writer.add_scalar(tag="D/Generator_loss",
step=step,
value=Generator_loss.numpy())
if step % 10 == 9:
writer.add_image("fake_A2B",
(porch.Tensor(fake_A2B[0] * 255)).clamp_(
0, 255).numpy().transpose(
[1, 2, 0]).astype(np.uint8), step)
writer.add_image("fake_B2A",
(porch.Tensor(fake_B2A[0] * 255)).clamp_(
0, 255).numpy().transpose(
[1, 2, 0]).astype(np.uint8), step)
writer.add_image("fake_A2B2A",
(porch.Tensor(fake_A2B[0] * 255)).clamp_(
0, 255).numpy().transpose(
[1, 2, 0]).astype(np.uint8), step)
writer.add_image("fake_B2A2B",
(porch.Tensor(fake_B2A[0] * 255)).clamp_(
0, 255).numpy().transpose(
[1, 2, 0]).astype(np.uint8), step)
print("[%5d/%5d] time: %4.4f d_loss: %.8f, g_loss: %.8f" %
(step, self.iteration, time.time() - start_time,
Discriminator_loss, Generator_loss))
if step % self.print_freq == 0:
train_sample_num = 5
test_sample_num = 5
A2B = np.zeros((self.img_size * 7, 0, 3))
B2A = np.zeros((self.img_size * 7, 0, 3))
self.genA2B.eval(), self.genB2A.eval(), self.disGA.eval(
), self.disGB.eval(), self.disLA.eval(), self.disLB.eval()
for _ in range(train_sample_num):
try:
real_A, _ = trainA_iter.next()
except:
trainA_iter = iter(self.trainA_loader)
real_A, _ = trainA_iter.next()
try:
real_B, _ = trainB_iter.next()
except:
trainB_iter = iter(self.trainB_loader)
real_B, _ = trainB_iter.next()
real_A, real_B = real_A[0], real_B[0]
fake_A2B, _, fake_A2B_heatmap = self.genA2B(real_A)
fake_B2A, _, fake_B2A_heatmap = self.genB2A(real_B)
fake_A2B2A, _, fake_A2B2A_heatmap = self.genB2A(fake_A2B)
fake_B2A2B, _, fake_B2A2B_heatmap = self.genA2B(fake_B2A)
fake_A2A, _, fake_A2A_heatmap = self.genB2A(real_A)
fake_B2B, _, fake_B2B_heatmap = self.genA2B(real_B)
A2B = np.concatenate(
(A2B,
np.concatenate(
(RGB2BGR(tensor2numpy(denorm(real_A[0]))),
cam(tensor2numpy(fake_A2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2A[0]))),
cam(tensor2numpy(fake_A2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2B[0]))),
cam(tensor2numpy(fake_A2B2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2B2A[0])))),
0)), 1)
B2A = np.concatenate(
(B2A,
np.concatenate(
(RGB2BGR(tensor2numpy(denorm(real_B[0]))),
cam(tensor2numpy(fake_B2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2B[0]))),
cam(tensor2numpy(fake_B2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2A[0]))),
cam(tensor2numpy(fake_B2A2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2A2B[0])))),
0)), 1)
for _ in range(test_sample_num):
try:
real_A, _ = testA_iter.next()
except:
testA_iter = iter(self.testA_loader)
real_A, _ = testA_iter.next()
try:
real_B, _ = testB_iter.next()
except:
testB_iter = iter(self.testB_loader)
real_B, _ = testB_iter.next()
real_A, real_B = real_A[0], real_B[0]
fake_A2B, _, fake_A2B_heatmap = self.genA2B(real_A)
fake_B2A, _, fake_B2A_heatmap = self.genB2A(real_B)
fake_A2B2A, _, fake_A2B2A_heatmap = self.genB2A(fake_A2B)
fake_B2A2B, _, fake_B2A2B_heatmap = self.genA2B(fake_B2A)
fake_A2A, _, fake_A2A_heatmap = self.genB2A(real_A)
fake_B2B, _, fake_B2B_heatmap = self.genA2B(real_B)
A2B = np.concatenate(
(A2B,
np.concatenate(
(RGB2BGR(tensor2numpy(denorm(real_A[0]))),
cam(tensor2numpy(fake_A2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2A[0]))),
cam(tensor2numpy(fake_A2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2B[0]))),
cam(tensor2numpy(fake_A2B2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_A2B2A[0])))),
0)), 1)
B2A = np.concatenate(
(B2A,
np.concatenate(
(RGB2BGR(tensor2numpy(denorm(real_B[0]))),
cam(tensor2numpy(fake_B2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2B[0]))),
cam(tensor2numpy(fake_B2A_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2A[0]))),
cam(tensor2numpy(fake_B2A2B_heatmap[0]),
self.img_size),
RGB2BGR(tensor2numpy(denorm(fake_B2A2B[0])))),
0)), 1)
if (not self.is_parallel
) or fluid.dygraph.parallel.Env().local_rank == 0:
cv2.imwrite(
os.path.join(self.result_dir, self.dataset, 'img',
'A2B_%07d.png' % step), A2B * 255.0)
cv2.imwrite(
os.path.join(self.result_dir, self.dataset, 'img',
'B2A_%07d.png' % step), B2A * 255.0)
self.genA2B.train(), self.genB2A.train(), self.disGA.train(
), self.disGB.train(), self.disLA.train(), self.disLB.train()
if step % self.save_freq == 0:
if (not self.is_parallel
) or fluid.dygraph.parallel.Env().local_rank == 0:
self.save(
os.path.join(self.result_dir, self.dataset, 'model'),
step)
if step % 1000 == 0:
params = {}
params['genA2B'] = self.genA2B.state_dict()
params['genB2A'] = self.genB2A.state_dict()
params['disGA'] = self.disGA.state_dict()
params['disGB'] = self.disGB.state_dict()
params['disLA'] = self.disLA.state_dict()
params['disLB'] = self.disLB.state_dict()
if (not self.is_parallel
) or fluid.dygraph.parallel.Env().local_rank == 0:
torch.save(
params,
os.path.join(self.result_dir,
self.dataset + '_params_latest.pt'))
def main():
# Step 0: preparation
writer = LogWriter(logdir="./log/scalar")
place = paddle.fluid.CUDAPlace(0)
with fluid.dygraph.guard(place):
# Step 1: Define training dataloader
image_folder = ""
image_list_file = "dummy_data/fabric_list.txt"
transform = Transform() #Normalize2() # [0,255]-->[0,1]
x_data = DataLoader(image_folder, image_list_file, transform=transform)
x_dataloader = fluid.io.DataLoader.from_generator(capacity=2,
return_list=True)
x_dataloader.set_sample_generator(x_data, args.batch_size)
total_batch = len(x_data) // args.batch_size
# Step 2: Create model
if args.net == "basic":
D = Discriminator()
G = Generator()
E = Invertor()
else:
raise NotImplementedError(
f"args.net: {args.net} is not Supported!")
# Step 3: Define criterion and optimizer
criterion = Basic_Loss
D_optim = AdamOptimizer(learning_rate=args.lr,
parameter_list=D.parameters())
G_optim = AdamOptimizer(learning_rate=args.lr,
parameter_list=G.parameters())
E_optim = AdamOptimizer(learning_rate=args.lr,
parameter_list=E.parameters())
G_loss_meter = AverageMeter()
D_loss_meter = AverageMeter()
E_loss_meter = AverageMeter()
D.train()
G.train()
E.train()
# Step 4: Slight Training
iteration = -1
is_slight_Train = True
for epoch in range(1, args.epoch_num + 1):
#optim Discriminator
for (x, x_labels) in x_dataloader():
n = x.shape[0]
if is_slight_Train:
iteration += 1
x = fluid.layers.cast(x, dtype="float32")
x = fluid.layers.transpose(x, perm=[0, 3, 1, 2])
preds_x = D(x)
preds_x_array = preds_x.numpy()
#print("D(x),1",preds_array.shape, np.mean(preds_array))
writer.add_scalar(tag="D(x)=1",
step=iteration,
value=np.mean(preds_x_array))
if np.mean(preds_x_array) >= 0.98:
is_slight_Train = False
z = np.random.rand(n, 64)
zeros = np.zeros((n, 1))
z = to_variable(z)
zeros = to_variable(zeros)
z = fluid.layers.cast(z, dtype="float32")
zeros = fluid.layers.cast(zeros, dtype="int64")
preds_fx = D(G(z))
preds_fx_array = preds_fx.numpy()
writer.add_scalar(tag="D(G(z))=0",
step=iteration,
value=np.mean(preds_fx_array))
D_loss = criterion(preds_x, x_labels) + criterion(
preds_fx, zeros)
D_loss.backward()
D_optim.minimize(D_loss)
D.clear_gradients()
D_loss_meter.update(D_loss.numpy()[0], n)
writer.add_scalar(tag="D_loss",
step=iteration,
value=D_loss_meter.avg)
print(f"EPOCH[{epoch:03d}/{args.epoch_num:03d}], " +
f"STEP{iteration}, " +
f"Average D Loss: {D_loss_meter.avg:4f}, ")
z = np.random.rand(n, 64)
ones = np.ones((n, 1))
z = to_variable(z)
ones = to_variable(ones)
z = fluid.layers.cast(z, dtype="float32")
ones = fluid.layers.cast(ones, dtype="int64")
preds = D(G(z))
preds_array = preds.numpy()
writer.add_scalar(tag="D(G(z))=1",
step=iteration,
value=np.mean(preds_array))
G_loss = criterion(preds, ones)
G_loss.backward()
G_optim.minimize(G_loss)
G.clear_gradients()
G_loss_meter.update(G_loss.numpy()[0], n)
writer.add_scalar(tag="G_loss",
step=iteration,
value=G_loss_meter.avg)
print(f"EPOCH[{epoch:03d}/{args.epoch_num:03d}], " +
f"STEP{iteration}, " +
f"Average G Loss: {G_loss_meter.avg:4f}")
if epoch % args.save_freq == 0 or epoch == args.epoch_num or not is_slight_Train:
D_model_path = os.path.join(args.checkpoint_folder,
f"D_{args.net}-Epoch-{epoch}")
G_model_path = os.path.join(args.checkpoint_folder,
f"G_{args.net}-Epoch-{epoch}")
# save model and optmizer states
model_dict = D.state_dict()
fluid.save_dygraph(model_dict, D_model_path)
optim_dict = D_optim.state_dict()
fluid.save_dygraph(optim_dict, D_model_path)
model_dict = G.state_dict()
fluid.save_dygraph(model_dict, G_model_path)
optim_dict = G_optim.state_dict()
fluid.save_dygraph(optim_dict, G_model_path)
print(
f'----- Save model: {D_model_path}.pdparams, {G_model_path}.pdparams'
)
if not is_slight_Train:
break
# Step 5: full training for Generator and Discriminator
D_optim = AdamOptimizer(learning_rate=args.lr * 10,
parameter_list=D.parameters())
G_optim = AdamOptimizer(learning_rate=args.lr * 10,
parameter_list=G.parameters())
G_loss_meter = AverageMeter()
D_loss_meter = AverageMeter()
for epoch in range(1, args.epoch_num + 1):
for (x, x_labels) in x_dataloader():
n = x.shape[0]
iteration += 1
x = fluid.layers.cast(x, dtype="float32")
x = fluid.layers.transpose(x, perm=[0, 3, 1, 2])
preds1 = D(x)
preds_array = preds1.numpy()
writer.add_scalar(tag="D(x)=1",
step=iteration,
value=np.mean(preds_array))
z = np.random.rand(n, 64)
zeros = np.zeros((n, 1))
z = to_variable(z)
zeros = to_variable(zeros)
z = fluid.layers.cast(z, dtype="float32")
zeros = fluid.layers.cast(zeros, dtype="int64")
preds2 = D(G(z))
preds_array = preds2.numpy()
#print("DG(z),0:",preds_array.shape, np.mean(preds_array))
writer.add_scalar(tag="D(G(z))=0",
step=iteration,
value=np.mean(preds_array))
D_loss = criterion(preds1, x_labels) + criterion(preds2, zeros)
D_loss.backward()
D_optim.minimize(D_loss)
D.clear_gradients()
D_loss_meter.update(D_loss.numpy()[0], n)
writer.add_scalar(tag="D_loss",
step=iteration,
value=D_loss_meter.avg)
print(f"EPOCH[{epoch:03d}/{args.epoch_num:03d}], " +
f"STEP{iteration}, " +
f"Average D Loss: {D_loss_meter.avg:4f} ")
z = np.random.rand(n, 64)
ones = np.ones((n, 1))
z = to_variable(z)
ones = to_variable(ones)
z = fluid.layers.cast(z, dtype="float32")
ones = fluid.layers.cast(ones, dtype="int64")
preds = D(G(z))
preds_array = preds.numpy()
#print("DG(z),1:",preds_array.shape, np.mean(preds_array))
writer.add_scalar(tag="D(G(z))=1",
step=iteration,
value=np.mean(preds_array))
G_loss = criterion(preds, ones)
G_loss.backward()
G_optim.minimize(G_loss)
G.clear_gradients()
G_loss_meter.update(G_loss.numpy()[0], n)
writer.add_scalar(tag="G_loss",
step=iteration,
value=G_loss_meter.avg)
print(f"EPOCH[{epoch:03d}/{args.epoch_num:03d}], " +
f"STEP{iteration}, " +
f"Average G Loss: {G_loss_meter.avg:4f}")
if epoch % args.save_freq == 0 or epoch == args.epoch_num:
D_model_path = os.path.join(args.checkpoint_folder,
f"D_{args.net}-Epoch-{epoch}")
G_model_path = os.path.join(args.checkpoint_folder,
f"G_{args.net}-Epoch-{epoch}")
# save model and optmizer states
model_dict = D.state_dict()
fluid.save_dygraph(model_dict, D_model_path)
optim_dict = D_optim.state_dict()
fluid.save_dygraph(optim_dict, D_model_path)
model_dict = G.state_dict()
fluid.save_dygraph(model_dict, G_model_path)
optim_dict = G_optim.state_dict()
fluid.save_dygraph(optim_dict, G_model_path)
print(
f'----- Save model: {D_model_path}.pdparams, {G_model_path}.pdparams'
)
# Step 6: full training for Inverter
E_optim = AdamOptimizer(learning_rate=args.lr * 10,
parameter_list=E.parameters())
E_loss_meter = AverageMeter()
for epoch in range(1, args.epoch_num + 1):
for (x, x_labels) in x_dataloader():
n = x.shape[0]
iteration += 1
x = fluid.layers.cast(x, dtype="float32")
image = x.numpy()[0] * 255
writer.add_image(tag="x", step=iteration, img=image)
x = fluid.layers.transpose(x, perm=[0, 3, 1, 2])
invert_x = G(E(x))
invert_image = fluid.layers.transpose(invert_x,
perm=[0, 2, 3, 1])
invert_image = invert_image.numpy()[0] * 255
#print("D(x),1",preds_array.shape, np.mean(preds_array))
writer.add_image(tag="invert_x",
step=iteration,
img=invert_image)
print(np.max(invert_image), np.min(invert_image))
E_loss = fluid.layers.mse_loss(invert_x, x)
print("E_loss shape:", E_loss.numpy().shape)
E_loss.backward()
E_optim.minimize(E_loss)
E.clear_gradients()
E_loss_meter.update(E_loss.numpy()[0], n)
writer.add_scalar(tag="E_loss",
step=iteration,
value=E_loss_meter.avg)
print(f"EPOCH[{epoch:03d}/{args.epoch_num:03d}], " +
f"STEP{iteration}, " +
f"Average E Loss: {E_loss_meter.avg:4f}, ")
if epoch % args.save_freq == 0 or epoch == args.epoch_num:
E_model_path = os.path.join(args.checkpoint_folder,
f"E_{args.net}-Epoch-{epoch}")
# save model and optmizer states
model_dict = E.state_dict()
fluid.save_dygraph(model_dict, E_model_path)
optim_dict = E_optim.state_dict()
fluid.save_dygraph(optim_dict, E_model_path)
print(
f'----- Save model: {E_model_path}.pdparams, {E_model_path}.pdparams'
)
def train(self, loaders):
args = self.args
nets = self.nets
nets_ema = self.nets_ema
optims = self.optims
writer = LogWriter(logdir=self.args.checkpoint_dir + "/log/")
# fetch random validation images for debugging
fetcher = InputFetcher(loaders.src, loaders.ref, args.latent_dim,
'train')
fetcher_val = InputFetcher(loaders.val, None, args.latent_dim, 'val')
inputs_val = next(fetcher_val)
# resume training if necessary
if args.resume_iter > 0:
self._load_checkpoint(args.resume_iter)
# remember the initial value of ds weight
initial_lambda_ds = args.lambda_ds
print('Start training...')
import tqdm
start_time = time.time()
tqdm_descriptor = tqdm.trange(args.resume_iter, args.total_iters)
for i in tqdm_descriptor:
# fetch images and labels
inputs = next(fetcher)
x_real, y_org = inputs.x_src, inputs.y_src
x_ref, x_ref2, y_trg = inputs.x_ref, inputs.x_ref2, inputs.y_ref
z_trg, z_trg2 = inputs.z_trg, inputs.z_trg2
masks = nets.fan.get_heatmap(x_real) if args.w_hpf > 0 else None
# train the discriminator
d_loss, d_losses_latent = compute_d_loss(nets,
args,
x_real,
y_org,
y_trg,
z_trg=z_trg,
masks=masks)
self._reset_grad()
d_loss.backward()
optims.discriminator.minimize(d_loss)
d_loss, d_losses_ref = compute_d_loss(nets,
args,
x_real,
y_org,
y_trg,
x_ref=x_ref,
masks=masks)
self._reset_grad()
d_loss.backward()
optims.discriminator.minimize(d_loss)
# train the generator
if i - args.resume_iter > 100: ##train discriminator first
g_loss, g_losses_latent, sample_1 = compute_g_loss(
nets,
args,
x_real,
y_org,
y_trg,
z_trgs=[z_trg, z_trg2],
masks=masks)
self._reset_grad()
g_loss.backward()
optims.generator.minimize(g_loss)
optims.mapping_network.minimize(g_loss)
optims.style_encoder.minimize(g_loss)
g_loss, g_losses_ref, sample_2 = compute_g_loss(
nets,
args,
x_real,
y_org,
y_trg,
x_refs=[x_ref, x_ref2],
masks=masks)
self._reset_grad()
g_loss.backward()
optims.generator.minimize(g_loss)
# compute moving average of network parameters
moving_average(nets.generator, nets_ema.generator, beta=0.999)
moving_average(nets.mapping_network,
nets_ema.mapping_network,
beta=0.999)
moving_average(nets.style_encoder,
nets_ema.style_encoder,
beta=0.999)
# decay weight for diversity sensitive loss
if args.lambda_ds > 0:
args.lambda_ds -= (initial_lambda_ds / args.ds_iter)
# print out log info
if (i + 1) % args.print_every == 0:
elapsed = time.time() - start_time
elapsed = str(datetime.timedelta(seconds=elapsed))[:-7]
log = "Elapsed time [%s], Iteration [%i/%i], " % (
elapsed, i + 1, args.total_iters)
all_losses = dict()
for loss, prefix in zip([
d_losses_latent, d_losses_ref, g_losses_latent,
g_losses_ref
], ['D/latent_', 'D/ref_', 'G/latent_', 'G/ref_']):
for key, value in loss.items():
all_losses[prefix + key] = value
writer.add_scalar(tag=prefix + key,
step=i + 1,
value=value)
all_losses['G/lambda_ds'] = args.lambda_ds
log += ' '.join([
'%s: [%.4f]' % (key, value)
for key, value in all_losses.items()
])
tqdm_descriptor.set_description(log)
writer.add_image("x_fake",
(utils.denormalize(sample_1) *
255).numpy().transpose([1, 2, 0]).astype(
np.uint8), i + 1)
# generate images for debugging
if (i + 1) % args.sample_every == 0:
os.makedirs(args.sample_dir, exist_ok=True)
utils.debug_image(nets_ema,
args,
inputs=inputs_val,
step=i + 1)
# save model checkpoints
if (i + 1) % args.save_every == 0:
self._save_checkpoint(step=i + 1)
# compute FID and LPIPS if necessary
if (i + 1) % args.eval_every == 0:
calculate_metrics(nets_ema, args, i + 1, mode='latent')
calculate_metrics(nets_ema, args, i + 1, mode='reference')
else:
if (i + 1) % args.print_every == 0:
elapsed = time.time() - start_time
elapsed = str(datetime.timedelta(seconds=elapsed))[:-7]
log = "Elapsed time [%s], Iteration [%i/%i], " % (
elapsed, i + 1, args.total_iters)
all_losses = dict()
for loss, prefix in zip([d_losses_latent, d_losses_ref],
['D/latent_', 'D/ref_']):
for key, value in loss.items():
all_losses[prefix + key] = value
writer.add_scalar(tag=prefix + key,
step=i + 1,
value=value)
log += ' '.join([
'%s: [%.4f]' % (key, value)
for key, value in all_losses.items()
])
tqdm_descriptor.set_description(log)
writer.close()
def synthesis(text_input, args):
local_rank = dg.parallel.Env().local_rank
place = (fluid.CUDAPlace(local_rank) if args.use_gpu else fluid.CPUPlace())
with open(args.config) as f:
cfg = yaml.load(f, Loader=yaml.Loader)
# tensorboard
if not os.path.exists(args.output):
os.mkdir(args.output)
writer = LogWriter(os.path.join(args.output, 'log'))
fluid.enable_dygraph(place)
with fluid.unique_name.guard():
network_cfg = cfg['network']
model = TransformerTTS(
network_cfg['embedding_size'], network_cfg['hidden_size'],
network_cfg['encoder_num_head'], network_cfg['encoder_n_layers'],
cfg['audio']['num_mels'], network_cfg['outputs_per_step'],
network_cfg['decoder_num_head'], network_cfg['decoder_n_layers'])
# Load parameters.
global_step = io.load_parameters(
model=model, checkpoint_path=args.checkpoint_transformer)
model.eval()
# init input
text = np.asarray(text_to_sequence(text_input))
text = fluid.layers.unsqueeze(dg.to_variable(text).astype(np.int64), [0])
mel_input = dg.to_variable(np.zeros([1, 1, 80])).astype(np.float32)
pos_text = np.arange(1, text.shape[1] + 1)
pos_text = fluid.layers.unsqueeze(
dg.to_variable(pos_text).astype(np.int64), [0])
for i in range(args.max_len):
pos_mel = np.arange(1, mel_input.shape[1] + 1)
pos_mel = fluid.layers.unsqueeze(
dg.to_variable(pos_mel).astype(np.int64), [0])
mel_pred, postnet_pred, attn_probs, stop_preds, attn_enc, attn_dec = model(
text, mel_input, pos_text, pos_mel)
if stop_preds.numpy()[0, -1] > args.stop_threshold:
break
mel_input = fluid.layers.concat([mel_input, postnet_pred[:, -1:, :]],
axis=1)
global_step = 0
for i, prob in enumerate(attn_probs):
for j in range(4):
x = np.uint8(cm.viridis(prob.numpy()[j]) * 255)
writer.add_image('Attention_%d_0' % global_step, x, i * 4 + j)
if args.vocoder == 'griffin-lim':
#synthesis use griffin-lim
wav = synthesis_with_griffinlim(postnet_pred, cfg['audio'])
elif args.vocoder == 'waveflow':
# synthesis use waveflow
wav = synthesis_with_waveflow(postnet_pred, args,
args.checkpoint_vocoder, place)
else:
print(
'vocoder error, we only support griffinlim and waveflow, but recevied %s.'
% args.vocoder)
writer.add_audio(text_input + '(' + args.vocoder + ')', wav, 0,
cfg['audio']['sr'])
if not os.path.exists(os.path.join(args.output, 'samples')):
os.mkdir(os.path.join(args.output, 'samples'))
write(
os.path.join(os.path.join(args.output, 'samples'),
args.vocoder + '.wav'), cfg['audio']['sr'], wav)
print("Synthesis completed !!!")
writer.close()
def main(args):
local_rank = dg.parallel.Env().local_rank
nranks = dg.parallel.Env().nranks
parallel = nranks > 1
with open(args.config) as f:
cfg = yaml.load(f, Loader=yaml.Loader)
global_step = 0
place = fluid.CUDAPlace(local_rank) if args.use_gpu else fluid.CPUPlace()
if not os.path.exists(args.output):
os.mkdir(args.output)
writer = LogWriter(os.path.join(args.output,
'log')) if local_rank == 0 else None
fluid.enable_dygraph(place)
network_cfg = cfg['network']
model = TransformerTTS(
network_cfg['embedding_size'], network_cfg['hidden_size'],
network_cfg['encoder_num_head'], network_cfg['encoder_n_layers'],
cfg['audio']['num_mels'], network_cfg['outputs_per_step'],
network_cfg['decoder_num_head'], network_cfg['decoder_n_layers'])
model.train()
optimizer = fluid.optimizer.AdamOptimizer(
learning_rate=dg.NoamDecay(1 / (cfg['train']['warm_up_step'] *
(cfg['train']['learning_rate']**2)),
cfg['train']['warm_up_step']),
parameter_list=model.parameters(),
grad_clip=fluid.clip.GradientClipByGlobalNorm(cfg['train'][
'grad_clip_thresh']))
# Load parameters.
global_step = io.load_parameters(
model=model,
optimizer=optimizer,
checkpoint_dir=os.path.join(args.output, 'checkpoints'),
iteration=args.iteration,
checkpoint_path=args.checkpoint)
print("Rank {}: checkpoint loaded.".format(local_rank))
if parallel:
strategy = dg.parallel.prepare_context()
model = fluid.dygraph.parallel.DataParallel(model, strategy)
reader = LJSpeechLoader(
cfg['audio'],
place,
args.data,
cfg['train']['batch_size'],
nranks,
local_rank,
shuffle=True).reader
iterator = iter(tqdm(reader))
global_step += 1
while global_step <= cfg['train']['max_iteration']:
try:
batch = next(iterator)
except StopIteration as e:
iterator = iter(tqdm(reader))
batch = next(iterator)
character, mel, mel_input, pos_text, pos_mel, stop_tokens = batch
mel_pred, postnet_pred, attn_probs, stop_preds, attn_enc, attn_dec = model(
character, mel_input, pos_text, pos_mel)
mel_loss = layers.mean(
layers.abs(layers.elementwise_sub(mel_pred, mel)))
post_mel_loss = layers.mean(
layers.abs(layers.elementwise_sub(postnet_pred, mel)))
loss = mel_loss + post_mel_loss
stop_loss = cross_entropy(
stop_preds, stop_tokens, weight=cfg['network']['stop_loss_weight'])
loss = loss + stop_loss
if local_rank == 0:
writer.add_scalar('training_loss/mel_loss',
mel_loss.numpy(),
global_step)
writer.add_scalar('training_loss/post_mel_loss',
post_mel_loss.numpy(),
global_step)
writer.add_scalar('stop_loss', stop_loss.numpy(), global_step)
if parallel:
writer.add_scalar('alphas/encoder_alpha',
model._layers.encoder.alpha.numpy(),
global_step)
writer.add_scalar('alphas/decoder_alpha',
model._layers.decoder.alpha.numpy(),
global_step)
else:
writer.add_scalar('alphas/encoder_alpha',
model.encoder.alpha.numpy(),
global_step)
writer.add_scalar('alphas/decoder_alpha',
model.decoder.alpha.numpy(),
global_step)
writer.add_scalar('learning_rate',
optimizer._learning_rate.step().numpy(),
global_step)
if global_step % cfg['train']['image_interval'] == 1:
for i, prob in enumerate(attn_probs):
for j in range(cfg['network']['decoder_num_head']):
x = np.uint8(
cm.viridis(prob.numpy()[j * cfg['train'][
'batch_size'] // nranks]) * 255)
writer.add_image(
'Attention_%d_0' % global_step,
x,
i * 4 + j)
for i, prob in enumerate(attn_enc):
for j in range(cfg['network']['encoder_num_head']):
x = np.uint8(
cm.viridis(prob.numpy()[j * cfg['train'][
'batch_size'] // nranks]) * 255)
writer.add_image(
'Attention_enc_%d_0' % global_step,
x,
i * 4 + j)
for i, prob in enumerate(attn_dec):
for j in range(cfg['network']['decoder_num_head']):
x = np.uint8(
cm.viridis(prob.numpy()[j * cfg['train'][
'batch_size'] // nranks]) * 255)
writer.add_image(
'Attention_dec_%d_0' % global_step,
x,
i * 4 + j)
if parallel:
loss = model.scale_loss(loss)
loss.backward()
model.apply_collective_grads()
else:
loss.backward()
optimizer.minimize(loss)
model.clear_gradients()
# save checkpoint
if local_rank == 0 and global_step % cfg['train'][
'checkpoint_interval'] == 0:
io.save_parameters(
os.path.join(args.output, 'checkpoints'), global_step, model,
optimizer)
global_step += 1
if local_rank == 0:
writer.close()
def main(args):
config = Config(args.config)
cfg = config(vars(args), mode=['infer', 'init'])
scale = cfg['infer']['scale']
mdname = cfg['infer']['model']
imgname = ''.join(mdname) # + '/' + str(scale)
#dirname = ''.join(mdname) + '_' + str(scale)
sz = cfg['infer']['sz']
infer_size = cfg['infer']['infer_size']
#save_path = os.path.join(args.save_dir, cfg['init']['result'])
list = cfg['infer']['infer_size']
save_path = create_path(args.save_dir, cfg['init']['result'])
save_path = create_path(save_path, imgname)
save_path = create_path(save_path, str(scale))
tif_path = create_path(save_path, cfg['infer']['lab'])
color_path = create_path(save_path, cfg['infer']['color'])
gray_path = create_path(save_path, cfg['infer']['gray'])
vdl_dir = os.path.join(args.save_dir, cfg['init']['vdl_dir'])
palette = cfg['infer']['palette']
palette = np.array(palette, dtype=np.uint8)
num_class = cfg['init']['num_classes']
batchsz = cfg['infer']['batchsz']
infer_path = os.path.join(cfg['infer']['root_path'], cfg['infer']['path'])
tagname = imgname + '/' + str(scale)
vdl_dir = os.path.join(vdl_dir, 'infer')
writer = LogWriter(logdir=vdl_dir)
infer_ds = TeDataset(path=cfg['infer']['root_path'],
fl=cfg['infer']['path'],
sz=sz)
total = len(infer_ds)
# select model
#addresult = np.zeros((total//batchsz,batchsz,num_class,sz,sz))
addresult = np.zeros((total, num_class, sz, sz))
for mnet in mdname:
result_list = []
net = modelset(mode=mnet, num_classes=cfg['init']['num_classes'])
# load moel
input = InputSpec([None, 3, 64, 64], 'float32', 'x')
label = InputSpec([None, 1, 64, 64], 'int64', 'label')
model = paddle.Model(net, input, label)
model.load(path=os.path.join(args.save_dir, mnet) + '/' + mnet)
model.prepare()
result = model.predict(
infer_ds,
batch_size=batchsz,
num_workers=cfg['infer']['num_workers'],
stack_outputs=True # [160,2,64,64]
)
addresult = result[0] + scale * addresult
pred = construct(addresult, infer_size, sz=sz)
# pred = construct(addresult,infer_size,sz = sz)
# # 腐蚀膨胀
# read vdl
file_list = os.listdir(infer_path)
file_list.sort(key=lambda x: int(x[-5:-4]))
step = 0
for i, fl in enumerate(file_list):
name, _ = fl.split(".")
# save pred
lab_img = Image.fromarray(pred[i].astype(np.uint8)).convert("L")
saveimg(lab_img, tif_path, name=name, type='.tif')
# gray_label
label = colorize(pred[i], palette)
writer.add_image(tag=tagname,
img=saveimg(label,
gray_path,
name=name,
type='.png',
re_out=True),
step=step,
dataformats='HW')
step += 1
# color_label
file = os.path.join(infer_path, fl)
out = blend_image(file, label, alpha=0.25)
writer.add_image(tag=tagname,
img=saveimg(out,
color_path,
name=name,
type='.png',
re_out=True),
step=step,
dataformats='HWC')
step += 1
writer.close()
