还是从前面的paddleocr学习笔记（二）训练 使用的文字检测训练代码学习。
python tools/train.py -c configs/det/det_mv3_db.yml -o Global.checkpoints=./output/db_mv3/latest从代码 train.py中分析。
想要了解代码是怎样运行的就要调试一下看看，在pycharm下调试设置快捷键：alt+shift+F9
入口：
调试快捷键：
Ctrl+B 跳转到函数定义
F8 单步执行
F9 执行到断点
F7 进入函数
从上面F7进入下面的部分，然后F8单步执行
如下可见配置文件读取和断点模型加载都是在program.py中完成的
配置参数说明：
是否使用GPU
训练最大轮次
每10步打印一次信息
保存模型路径
每多少迭代步评估一次
预训练模型加载路径
推理图片
保存结果路径
文字检测模型结构
损失函数
优化器
后处理
评估度量
Global:
use_gpu: true
epoch_num: 1200
log_smooth_window: 20
print_batch_step: 10
save_model_dir: ./output/db_mv3/
save_epoch_step: 1200
# evaluation is run every 2000 iterations
eval_batch_step: [0, 2000]
# if pretrained_model is saved in static mode, load_static_weights must set to True
load_static_weights: True
cal_metric_during_train: False
pretrained_model: ./pretrain_models/MobileNetV3_large_x0_5_pretrained
checkpoints:
save_inference_dir:
use_visualdl: False
infer_img: doc/imgs_en/img_10.jpg
save_res_path: ./output/det_db/predicts_db.txt
Architecture:
model_type: det
algorithm: DB
Transform:
Backbone:
name: MobileNetV3
scale: 0.5
model_name: large
Neck:
name: DBFPN
out_channels: 256
Head:
name: DBHead
k: 50
Loss:
name: DBLoss
balance_loss: true
main_loss_type: DiceLoss
alpha: 5
beta: 10
ohem_ratio: 3
Optimizer:
name: Adam
beta1: 0.9
beta2: 0.999
lr:
learning_rate: 0.001
regularizer:
name: 'L2'
factor: 0
PostProcess:
name: DBPostProcess
thresh: 0.3
box_thresh: 0.6
max_candidates: 1000
unclip_ratio: 1.5
Metric:
name: DetMetric
main_indicator: hmean训练数据的配置信息：
训练数据集与标注
加载图片增广/扩充：镜像、仿射变换、缩放、随机裁切、扩边、收缩、归一化。
图像混洗
batch大小
num_workers大小
Train:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/train_icdar2015_label.txt
ratio_list: [1.0]
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- IaaAugment:
augmenter_args:
- { 'type': Fliplr, 'args': { 'p': 0.5 } }
- { 'type': Affine, 'args': { 'rotate': [-10, 10] } }
- { 'type': Resize, 'args': { 'size': [0.5, 3] } }
- EastRandomCropData:
size: [640, 640]
max_tries: 50
keep_ratio: true
- MakeBorderMap:
shrink_ratio: 0.4
thresh_min: 0.3
thresh_max: 0.7
- MakeShrinkMap:
shrink_ratio: 0.4
min_text_size: 8
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'threshold_map', 'threshold_mask', 'shrink_map', 'shrink_mask'] # the order of the dataloader list
loader:
shuffle: True
drop_last: False
batch_size_per_card: 16
num_workers: 8
use_shared_memory: False
评估数据的配置信息：
Eval:
dataset:
name: SimpleDataSet
data_dir: ./train_data/icdar2015/text_localization/
label_file_list:
- ./train_data/icdar2015/text_localization/test_icdar2015_label.txt
transforms:
- DecodeImage: # load image
img_mode: BGR
channel_first: False
- DetLabelEncode: # Class handling label
- DetResizeForTest:
image_shape: [736, 1280]
- NormalizeImage:
scale: 1./255.
mean: [0.485, 0.456, 0.406]
std: [0.229, 0.224, 0.225]
order: 'hwc'
- ToCHWImage:
- KeepKeys:
keep_keys: ['image', 'shape', 'polys', 'ignore_tags']
loader:
shuffle: False
drop_last: False
batch_size_per_card: 1 # must be 1
num_workers: 8
use_shared_memory: False具体的训练代码也是在program.py中实现的：
def train(config,
train_dataloader,
valid_dataloader,
device,
model,
loss_class,
optimizer,
lr_scheduler,
post_process_class,
eval_class,
pre_best_model_dict,
logger,
vdl_writer=None):
cal_metric_during_train = config['Global'].get('cal_metric_during_train',
False)
log_smooth_window = config['Global']['log_smooth_window']
epoch_num = config['Global']['epoch_num']
print_batch_step = config['Global']['print_batch_step']
eval_batch_step = config['Global']['eval_batch_step']
global_step = 0
start_eval_step = 0
if type(eval_batch_step) == list and len(eval_batch_step) >= 2:
start_eval_step = eval_batch_step[0]
eval_batch_step = eval_batch_step[1]
if len(valid_dataloader) == 0:
logger.info(
'No Images in eval dataset, evaluation during training will be disabled'
)
start_eval_step = 1e111
logger.info(
"During the training process, after the {}th iteration, an evaluation is run every {} iterations".
format(start_eval_step, eval_batch_step))
save_epoch_step = config['Global']['save_epoch_step']
save_model_dir = config['Global']['save_model_dir']
if not os.path.exists(save_model_dir):
os.makedirs(save_model_dir)
main_indicator = eval_class.main_indicator
best_model_dict = {main_indicator: 0}
best_model_dict.update(pre_best_model_dict)
train_stats = TrainingStats(log_smooth_window, ['lr'])
model_average = False
model.train()
use_srn = config['Architecture']['algorithm'] == "SRN"
if 'start_epoch' in best_model_dict:
start_epoch = best_model_dict['start_epoch']
else:
start_epoch = 1
for epoch in range(start_epoch, epoch_num + 1):
train_dataloader = build_dataloader(
config, 'Train', device, logger, seed=epoch)
train_batch_cost = 0.0
train_reader_cost = 0.0
batch_sum = 0
batch_start = time.time()
for idx, batch in enumerate(train_dataloader):
train_reader_cost += time.time() - batch_start
if idx >= len(train_dataloader):
break
lr = optimizer.get_lr()
images = batch[0]
if use_srn:
others = batch[-4:]
preds = model(images, others)
model_average = True
else:
preds = model(images)
loss = loss_class(preds, batch)
avg_loss = loss['loss']
avg_loss.backward()
optimizer.step()
optimizer.clear_grad()
train_batch_cost += time.time() - batch_start
batch_sum += len(images)
if not isinstance(lr_scheduler, float):
lr_scheduler.step()
# logger and visualdl
stats = {k: v.numpy().mean() for k, v in loss.items()}
stats['lr'] = lr
train_stats.update(stats)
if cal_metric_during_train: # only rec and cls need
batch = [item.numpy() for item in batch]
post_result = post_process_class(preds, batch[1])
eval_class(post_result, batch)
metric = eval_class.get_metric()
train_stats.update(metric)
if vdl_writer is not None and dist.get_rank() == 0:
for k, v in train_stats.get().items():
vdl_writer.add_scalar('TRAIN/{}'.format(k), v, global_step)
vdl_writer.add_scalar('TRAIN/lr', lr, global_step)
if dist.get_rank(
) == 0 and global_step > 0 and global_step % print_batch_step == 0:
logs = train_stats.log()
strs = 'epoch: [{}/{}], iter: {}, {}, reader_cost: {:.5f} s, batch_cost: {:.5f} s, samples: {}, ips: {:.5f}'.format(
epoch, epoch_num, global_step, logs, train_reader_cost /
print_batch_step, train_batch_cost / print_batch_step,
batch_sum, batch_sum / train_batch_cost)
logger.info(strs)
train_batch_cost = 0.0
train_reader_cost = 0.0
batch_sum = 0
# eval
if global_step > start_eval_step and \
(global_step - start_eval_step) % eval_batch_step == 0 and dist.get_rank() == 0:
if model_average:
Model_Average = paddle.incubate.optimizer.ModelAverage(
0.15,
parameters=model.parameters(),
min_average_window=10000,
max_average_window=15625)
Model_Average.apply()
cur_metric = eval(
model,
valid_dataloader,
post_process_class,
eval_class,
use_srn=use_srn)
cur_metric_str = 'cur metric, {}'.format(', '.join(
['{}: {}'.format(k, v) for k, v in cur_metric.items()]))
logger.info(cur_metric_str)
# logger metric
if vdl_writer is not None:
for k, v in cur_metric.items():
if isinstance(v, (float, int)):
vdl_writer.add_scalar('EVAL/{}'.format(k),
cur_metric[k], global_step)
if cur_metric[main_indicator] >= best_model_dict[
main_indicator]:
best_model_dict.update(cur_metric)
best_model_dict['best_epoch'] = epoch
save_model(
model,
optimizer,
save_model_dir,
logger,
is_best=True,
prefix='best_accuracy',
best_model_dict=best_model_dict,
epoch=epoch)
best_str = 'best metric, {}'.format(', '.join([
'{}: {}'.format(k, v) for k, v in best_model_dict.items()
]))
logger.info(best_str)
# logger best metric
if vdl_writer is not None:
vdl_writer.add_scalar('EVAL/best_{}'.format(main_indicator),
best_model_dict[main_indicator],
global_step)
global_step += 1
optimizer.clear_grad()
batch_start = time.time()
if dist.get_rank() == 0:
save_model(
model,
optimizer,
save_model_dir,
logger,
is_best=False,
prefix='latest',
best_model_dict=best_model_dict,
epoch=epoch)
if dist.get_rank() == 0 and epoch > 0 and epoch % save_epoch_step == 0:
save_model(
model,
optimizer,
save_model_dir,
logger,
is_best=False,
prefix='iter_epoch_{}'.format(epoch),
best_model_dict=best_model_dict,
epoch=epoch)
best_str = 'best metric, {}'.format(', '.join(
['{}: {}'.format(k, v) for k, v in best_model_dict.items()]))
logger.info(best_str)
if dist.get_rank() == 0 and vdl_writer is not None:
vdl_writer.close()
return