From 55d80c317edd0fb5847e599a1802d394f34a3141 Mon Sep 17 00:00:00 2001
From: Chilicyy <lichuyi@meituan.com>
Date: Thu, 24 Aug 2023 20:36:25 +0800
Subject: [PATCH] upload convert model tutorial

---
 docs/Convert_model.md | 74 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 74 insertions(+)
 create mode 100644 docs/Convert_model.md

diff --git a/docs/Convert_model.md b/docs/Convert_model.md
new file mode 100644
index 00000000..efdd2051
--- /dev/null
+++ b/docs/Convert_model.md
@@ -0,0 +1,74 @@
+# 模型转换
+### 此文档将展示如何把 在3.0代码库下训练出来的模型 转换为 可在4.0代码库训练测试的模型。
+### 具体步骤如下：
+
+### 0. 分别拉取3.0代码库和4.0代码库
+
+### 1. 基于4.0代码库下训练相同配置的模型，当运行一个epoch后会在保存目录下保存模型文件，即可停止训练
+通过 `torch.load`加载模型文件，并将模型里所有权重的键值 `model.keys()` 保存下来.
+
+### 2. 在3.0代码库下进行模型权重键值替换
+其中newnames是步骤1保存的 `model.keys()`，需要提前得到并替换进来，运行得到转换keys后的model. 这里以 YOLOv6m 转换为例：
+
+```python
+import json
+import pickle
+import numpy
+import torch
+import sys
+import os
+ROOT = os.getcwd()
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))
+# Run in original code
+param_state_dict = torch.load('../weights/yolov6_v3/p5/yolov6m.pt')
+
+old_state_dict = {}
+param_state_dict = param_state_dict['model']
+conv_bias = None
+for k in param_state_dict.float().state_dict():
+    if 'conv.bias' in k:
+        conv_bias = param_state_dict.float().state_dict()[k].numpy()
+    if 'conv.bias' not in k:
+        if conv_bias is not None and 'bn.running_mean' in k:
+            old_state_dict[k] = param_state_dict.float().state_dict()[k].numpy() - conv_bias
+            conv_bias = None
+        else:
+            old_state_dict[k] = param_state_dict.float().state_dict()[k].numpy()
+
+# load new model with the latest code and paste new model.state_dict().keys() 
+newnames = ['backbone.stem.rbr_dense.conv.weight', 'backbone.stem.rbr_dense.bn.weight', ...]
+print(f'num of newnames:{len(newnames)}')
+print(f'num of old weights: {len(old_state_dict.keys())}')
+for i in range(len(newnames)):
+    print(newnames[i] + "   ******     " + list(old_state_dict.keys())[i])
+
+
+tweights = {}
+oldnames = list(old_state_dict.keys())
+for i in range(len(newnames)):
+    tweights[newnames[i]] = torch.tensor(old_state_dict[oldnames[i]])
+ckpt = {'model': tweights}
+torch.save(ckpt, 'weights/YOLOv6m_newtensor.pt')
+```
+
+
+### 3. 在4.0代码库下进行模型权重替换
+
+```python
+import torch
+import sys
+import os
+ROOT = os.getcwd()
+if str(ROOT) not in sys.path:
+    sys.path.append(str(ROOT))
+# Run in latest code
+m = torch.load('runs/train/coco_yolov6m_distill_retry/weights/last_ckpt.pt') #步骤1保存的模型路径
+state_dict = torch.load('../yolov6_3.0/weights/YOLOv6m_newtensor.pt') #步骤2转换权重键值后保存的模型路径
+m['model'].load_state_dict(state_dict['model'])
+m['ema'] = None
+m['updates'] = None
+m['optimizer'] = None
+m['epoch'] = 299
+torch.save(m, './weights/yolov6m.pt') # 保存最终转换后的模型文件
+```