本项目为辅助获得静态图和动态图权重名对应关系并基于此对应关系转换权重，流程如下：
- 静态图一层层顺序创建，故在
create_parameter时打印权重名和shape (static_print.py) - 动态图
create_parameter在__init__里，网络计算在forward里，顺序不一定一致，故在forward里打印权重名和shape (dygraph_print.py) - 从前到后依次匹配静态图和动态图的权重名，须shape能匹配上，若shape匹配不上，会自动搜索后续shape能匹配上的动态图权重名，选出候选权重名供用户手动选择 (parse.py)
- 匹配完静态图和动态图权重名对应关系存储在
weight_name_map.txt中，通过convert.py将静态图权重转换为动态图权重 (convert.py)
注意： 权重匹配过程中已自动将Conv-BN和FC的weight-bias融合匹配
可以尝试如下两条命令体验匹配过程:
YOLOv3
python parse.py yolov3_dy_print.txt yolov3_st_print.txt
MaskRCNN
python parse.py mask_fpn_dy_print.txt mask_fpn_st_print.txt
-
layers.py替换paddle包中paddle/fluid/dygraph/layers.py,layer_helper_base.py替换paddle包中paddle/fluid/layer_helper_base.py, 可以通过pip install/uninstall获取paddle包安装路径，替换过程建议用vimdiff把增量代码替换过去。 -
准备PaddleDetection静态图和动态图的代码库,
static_print.py移动至静态图库中，dynamic_print.py移动至动态图库中, 运行对应配置文件并将输出保存为文件，如：
# 静态图库中
python static_print.py -c configs/yolov3_darknet.yml 2>&1 | tee yolov3_st_print.txt
# 动态图库中
python dygraph_print.py -c configs/yolov3_darknet53_270e_coco.yml 2>&1 | tee yolov3_dy_print.txt
- 运行
parse.py如下，会解析静态图和动态图权重名对应关系，并将结果保存为weight_name_map.txt
命令行传递两个参数，分别为动态图输出文件和静态图输出文件
python parse.py yolov3_dy_print.txt yolov3_st_print.txt
若出现无法匹配的权重名(即shape匹配失败)，会有如下提示，手动选择匹配的权重名继续匹配。
bn2c_branch2c_scale matched backbone.res2.res2c.branch2c.norm.weight
bn2c_branch2c_offset matched backbone.res2.res2c.branch2c.norm.bias
bn2c_branch2c_mean matched backbone.res2.res2c.branch2c.norm._mean
bn2c_branch2c_variance matched backbone.res2.res2c.branch2c.norm._variance
res3a_branch2a_weights matched backbone.res3.res3a.branch2a.conv.weight
bn3a_branch2a_scale matched backbone.res3.res3a.branch2a.norm.weight
bn3a_branch2a_offset matched backbone.res3.res3a.branch2a.norm.bias
bn3a_branch2a_mean matched backbone.res3.res3a.branch2a.norm._mean
bn3a_branch2a_variance matched backbone.res3.res3a.branch2a.norm._variance
('*****match wrong*******', ('res3a_branch2b_weights', [128, 128, 3, 3]), ('weight', [512, 256, 1, 1], 'backbone.res3.res3a.short.conv.weight'), ', is ConvBN/FC block: ', 1)
Please select dygraph weight name for res3a_branch2b_weights:
1. backbone.res3.res3a.branch2b.conv.weight
2. backbone.res3.res3b.branch2b.conv.weight
3. backbone.res3.res3c.branch2b.conv.weight
4. backbone.res3.res3d.branch2b.conv.weight
selection:
如上输出，则为 res3a_branch2b_weights 判断后手动选择1，即backbone.res3.res3a.branch2b.conv.weight继续匹配
注意： 解析完成后最好自行check一下weight_name_map.txt中的权重对应关系，如mask_rcnn_fpn中优于fpn计算顺序动态图和静态图相反，同时shape一致，脚本无法识别出这种情况，判定匹配通过，所以会存在如下情况，可手动修改下weight_name_map.txt
fpn_inner_res2_sum_lateral_w neck.fpn_inner_res2_sum_lateral.weight
fpn_inner_res2_sum_lateral_b neck.fpn_inner_res2_sum_lateral.bias
# -------------------- 2, 3, 4, 5 顺序反了 --------------------
fpn_res5_sum_w neck.fpn_res2_sum.weight
fpn_res5_sum_b neck.fpn_res2_sum.bias
fpn_res4_sum_w neck.fpn_res3_sum.weight
fpn_res4_sum_b neck.fpn_res3_sum.bias
fpn_res3_sum_w neck.fpn_res4_sum.weight
fpn_res3_sum_b neck.fpn_res4_sum.bias
fpn_res2_sum_w neck.fpn_res5_sum.weight
fpn_res2_sum_b neck.fpn_res5_sum.bias
# ---------------------------------------------------------
conv_rpn_fpn2_w rpn_head.rpn_feat.rpn_conv.weight
conv_rpn_fpn2_b rpn_head.rpn_feat.rpn_conv.bias
rpn_cls_logits_fpn2_w rpn_head.rpn_rois_score.weight
rpn_cls_logits_fpn2_b rpn_head.rpn_rois_score.bias
若使用ResNet结构，注意确认一下shortcut分支的权重顺序，即动态图short.conv/norm应该对应静态图的branch1，若出现错位，可手动修改下weight_name_map.txt
bn2a_branch2b_mean backbone.res2.res2a.branch2b.norm._mean
bn2a_branch2b_variance backbone.res2.res2a.branch2b.norm._variance
# -------------------- short和branch2顺序反了 --------------------
res2a_branch2c_weights backbone.res2.res2a.short.conv.weight
bn2a_branch2c_scale backbone.res2.res2a.short.norm.weight
bn2a_branch2c_offset backbone.res2.res2a.short.norm.bias
bn2a_branch2c_mean backbone.res2.res2a.short.norm._mean
bn2a_branch2c_variance backbone.res2.res2a.short.norm._variance
res2a_branch1_weights backbone.res2.res2a.branch2c.conv.weight
bn2a_branch1_scale backbone.res2.res2a.branch2c.norm.weight
bn2a_branch1_offset backbone.res2.res2a.branch2c.norm.bias
bn2a_branch1_mean backbone.res2.res2a.branch2c.norm._mean
bn2a_branch1_variance backbone.res2.res2a.branch2c.norm._variance
# ----------------------------------------------------------------
res2b_branch2a_weights backbone.res2.res2b.branch2a.conv.weight
bn2b_branch2a_scale backbone.res2.res2b.branch2a.norm.weight
- 使用
convert.py和 3 中生成的weight_name_map.txt完成权重转换，命令如下：
命令行传递三个参数，为静态图权重(支持url)，3中的weight_name_map.txt和导出动态图权重文件名。
export PYTHONPATH=$PYTHONPATH:<path/to/PaddleDetection>
python convert.py https://paddlemodels.bj.bcebos.com/object_detection/yolov3_darknet.tar weight_name_map.txt new.pdparams
new.pdparams 即为转换后的动态图权重，确认下权重大小是否一致。