当前位置：首页 > article >正文

pytorch训练权重转化为tensorflow模型的教训

article 2025/3/16 5:07:46

模型构建时候有时候在工程量比较大的时候，不可避免使用迭代算法，迭代算法本身会让错误的追踪更加困难，因此掌握基本的框架之间的差异非常重要。以下均是在模型转换过程中出现的错误。

shuffle operation(shuffle 操作)

这个操作原本是用来将各个通道之间的信息进行打乱后，此时面临重要的问题就是，如果将通道打乱，在pytorch里面与tensorflow中间，两种通道排序是不一样的，是采用不同的通道数据排列进行的。

import tensorflow as tf

def channel_shuffle(x, groups):
    _, h, w, c = x.shape
    # c 通道进行划分
    x = tf.reshape(x, [-1, h, w, groups, c // groups])
    # 通道为基本单位的情况下，多group均采样重组
    x = tf.transpose(x, [0, 1, 2, 4, 3])  # 调整通道维度顺序

    # 混洗采样重组后再reshape变成之前的通道
    x = tf.reshape(x, [-1, h, w, c])
    return x

# 示例张量
x = tf.random.normal((2, 3, 3, 8))
print("Original tensor:\n", x.numpy())

# 进行通道混洗
shuffled_x = channel_shuffle(x, groups=2)
print("Shuffled tensor:\n", shuffled_x.numpy())

Pytorch下的GSBottleneck采用的Sequential具有差异

class GSBottleneck(nn.Module):
    # GS Bottleneck https://github.com/AlanLi1997/slim-neck-by-gsconv
    def __init__(self, c1, c2, k=3, s=1):
        super().__init__()
        c_ = c2 // 2
        # for lighting
        self.conv_lighting = nn.Sequential(
            GSConv(c1, c_, 1, 1),
            GSConv(c_, c2, 1, 1, act=False))
        # for receptive field
        self.conv = nn.Sequential(
            GSConv(c1, c_, 3, 1),
            GSConv(c_, c2, 3, 1, act=False))
        self.shortcut = nn.Identity()

    def forward(self, x):
        return self.conv_lighting(x)

我遇到的坑为：

class TFGSBottleneck(tf.keras.layers.Layer):
    # GS Bottleneck https://github.com/AlanLi1997/slim-neck-by-gsconv
    def __init__(self, c1, c2, k=3, s=1,w=None):
        super().__init__()
        c_ = c2 // 2
        # example
        # self.cv1 = TFConv(c1, c_, 1, 1, w=w.cv1)
        # self.cv2 = TFConv(c_, c2, 3, 1, g=g, w=w.cv2)

        self.conv_lighting = tf.keras.Sequential(
            TFGSConv(c1, c_, 1, 1,w=w.conv_lighting[0]),
            TFGSConv(c_, c2, 1, 1, act=False, w=w.conv_lighting[1])
        )

        # for receptive field
        self.conv = tf.keras.Sequential(
            TFGSConv(c1, c_, 3, 1,w=w.conv[0]),
            TFGSConv(c_, c2, 3, 1, act=False,w=w.conv[1])
        )

        self.shortcut = tf.keras.layers.Lambda(lambda x: x)

    def call(self, x):
        print("TFGSBottleneck input: ",x.shape)
        print("TFGSBottleneck output: ", self.conv_lighting(x).shape)
        return self.conv_lighting(x)

有以下错误

Traceback (most recent call last):
  File "D:\TEST\yolov5\models\tf.py", line 1078, in <module>
    main(opt)
  File "D:\TEST\yolov5\models\tf.py", line 1073, in main
    run(**vars(opt))
  File "D:\TEST\yolov5\models\tf.py", line 1044, in run
    _ = tf_model.predict(im)  # inference
        ^^^^^^^^^^^^^^^^^^^^
  File "D:\TEST\yolov5\models\tf.py", line 922, in predict
    x = m(x)  # run
        ^^^^
  File "C:\Users\Zhuliang\.conda\envs\exportyolo2\Lib\site-packages\keras\utils\traceback_utils.py", line 70, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "D:\TEST\yolov5\models\tf.py", line 383, in call
    m_x1 = self.m(x1)
           ^^^^^^^^^^
  File "D:\TEST\yolov5\models\tf.py", line 361, in call
    print("TFGSBottleneck output: ", self.conv_lighting(x).shape)
                                     ^^^^^^^^^^^^^^^^^^^^^
ValueError: Exception encountered when calling layer 'tfgs_bottleneck' (type TFGSBottleneck).

name for name_scope must be a string.

Call arguments received by layer 'tfgs_bottleneck' (type TFGSBottleneck):
  • x=tf.Tensor(shape=(1, 136, 136, 8), dtype=float32)

解决方案为

错误地将两个层作为位置参数传递给tf.keras.Sequential构造函数，导致第二个参数被误解为name参数，而name必须是一个字符串，但用户传递了一个层实例。这导致在调用层时，TensorFlow无法正确创建name_scope，从而引发错误。解决方法是把这两个层放在一个列表中，作为Sequential构造函数的第一个参数。

要解决此错误，需要将传递给tf.keras.Sequential的层放在列表中，确保它们被正确解析为层序列而不是其他参数。以下是修改后的代码：

tf.keras.Sequential与tensorflow中的pytorch

class TFGSBottleneck(tf.keras.layers.Layer):
    # GS Bottleneck https://github.com/AlanLi1997/slim-neck-by-gsconv
    def __init__(self, c1, c2, k=3, s=1, w=None):
        super().__init__()
        c_ = c2 // 2

        # 使用列表包裹层以正确传递
        self.conv_lighting = tf.keras.Sequential([
            TFGSConv(c1, c_, 1, 1, w=w.conv_lighting[0]),
            TFGSConv(c_, c2, 1, 1, act=False, w=w.conv_lighting[1])
        ])

        self.conv = tf.keras.Sequential([
            TFGSConv(c1, c_, 3, 1, w=w.conv[0]),
            TFGSConv(c_, c2, 3, 1, act=False, w=w.conv[1])
        ])

        self.shortcut = tf.keras.layers.Lambda(lambda x: x)

    def call(self, x):
        return self.conv_lighting(x)

查看全文

http://www.kler.cn/a/586503.html