之前在做一个手写字符的识别时遇到的问题，在此记录下详细的过程。
数据集如下：
36个文件夹，每个文件夹下3000个图片。
文件夹命名就是里面字符的实际label。
图像分辨率28*28，位深度 1 ，也就是单色图像。
如下代码是创建数据集的。
def create_dataset(path, batch_size=32):
dataset = ds.ImageFolderDataset(path,
num_parallel_workers=8,
class_indexing={"0": 0, "1": 1, "2": 2, "3": 3, "4": 4, "5": 5,
"6": 6, "7": 7, "8": 8, "9": 9, "a": 10, "b": 11,
"c": 12, "d": 13, "e": 14, "f": 15, "g": 16, "h": 17,
"i": 18, "j": 19, "k": 20, "l": 21, "m": 22, "n": 23,
"p": 24, "q": 25, "r": 26, "s": 27, "t": 28, "u": 29,
"v": 30, "w": 31, "x": 32, "y": 33, "z": 34, "space": 35})
rescale = 1.0 / 255.0
shift = 0.0
rescale_nml = 1 / 0.3081
shift_nml = -1 * 0.1307 / 0.3081
# define map operations
trans = [
transforms.Decode(),
transforms.Resize(size=32, interpolation=Inter.LINEAR),
transforms.Rescale(rescale, shift),
transforms.Rescale(rescale_nml, shift_nml),
transforms.HWC2CHW(),
]
dataset = dataset.map(operations=trans, input_columns="image", num_parallel_workers=8)
dataset = dataset.batch(batch_size, drop_remainder=True)
return dataset
数据集创建没什么问题，接下来就是放进网络里面进行训练了。
首先是创建网络，这个代码直接拿样例里面的来用就可以了。
class LeNet5(nn.Cell):
"""
LeNet-5网络结构
"""
def __init__(self, num_class=10, num_channel=1):
super(LeNet5, self).__init__()
# 卷积层，输入的通道数为num_channel，输出的通道数为6，卷积核大小为5*5
self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode='valid')
# 卷积层，输入的通道数为6，输出的通道数为16，卷积核大小为5*5
self.conv2 = nn.Conv2d(6, 16, 5, pad_mode='valid')
# 全连接层，输入个数为16*5*5，输出个数为120
self.fc1 = nn.Dense(16 * 5 * 5, 120)
# 全连接层，输入个数为120，输出个数为84
self.fc2 = nn.Dense(120, 84)
# 全连接层，输入个数为84，分类的个数为num_class
self.fc3 = nn.Dense(84, num_class)
# ReLU激活函数
self.relu = nn.ReLU()
self.relu2 = vmap(nn.ReLU(), in_axes=0, out_axes=0)
# 池化层
self.max_pool2d = nn.MaxPool2d(kernel_size=2, stride=2)
# 多维数组展平为一维数组
self.flatten = nn.Flatten()
def construct(self, x):
# 使用定义好的运算构建前向网络
x = self.conv1(x)
x = self.relu(x)
x = self.max_pool2d(x)
x = self.conv2(x)
x = self.relu(x)
x = self.max_pool2d(x)
x = self.flatten(x)
x = self.fc1(x)
x = self.relu(x)
x = self.fc2(x)
x = self.relu(x)
x = self.fc3(x)
x = self.relu2(x)
return x
network = LeNet5(num_class=36)
定义完网络后就是定义损失函数。
net_loss = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
然后定义优化器函数。
net_opt = nn.Momentum(network.trainable_params(), learning_rate=0.01, momentum=0.9)
设置模型保存参数，模型训练保存参数的step为1875。
config_ck = CheckpointConfig(save_checkpoint_steps=1875, keep_checkpoint_max=10)
应用模型保存参数。
ckpoint = ModelCheckpoint(prefix="lenet", directory="./test", config=config_ck)
然后就是初始化模型参数。
model = ms.Model(network, loss_fn=net_loss, optimizer=net_opt, metrics={'accuracy'})
最后一步就是训练网络模型，并保存为lenet-1_1875.ckpt文件
model.train(30, dataset_train, callbacks=[ckpoint, LossMonitor(0.01, 1875)])
到这里所有代码都写完了，然后就可以开始训练了。
[WARNING] DEVICE(,40c4,?):2023-6-10 19:10:0 [mindspore\ccsrc\runtime\pynative\async\async_queue.cc:66] mindspore::pynative::AsyncQueue::WorkerLoop] Run task failed, error msg:For 'Conv2D', 'C_in' of input 'x' shape divide by parameter 'group' must be equal to 'C_in' of input 'weight' shape: 1, but got 'C_in' of input 'x' shape: 3, and 'group': 1.
----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore\core\ops\conv2d.cc:241 mindspore::ops::`anonymous-namespace'::Conv2dInferShape
Traceback (most recent call last):
File "D:\ai\step1\form\mylenet.py", line 102, in <module>
model.train(30, dataset_train, callbacks=[ckpoint, LossMonitor(0.01, 1875)])
File "D:\python3.9\lib\site-packages\mindspore\train\model.py", line 1056, in train
self._train(epoch,
File "D:\python3.9\lib\site-packages\mindspore\train\model.py", line 100, in wrapper
func(self, *args, **kwargs)
File "D:\python3.9\lib\site-packages\mindspore\train\model.py", line 610, in _train
self._train_process(epoch, train_dataset, list_callback, cb_params, initial_epoch, valid_infos)
File "D:\python3.9\lib\site-packages\mindspore\train\model.py", line 909, in _train_process
outputs = self._train_network(*next_element)
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 661, in __call__
raise err
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 657, in __call__
output = self._run_construct(args, kwargs)
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 445, in _run_construct
output = self.construct(*cast_inputs, **kwargs)
File "D:\python3.9\lib\site-packages\mindspore\nn\wrap\cell_wrapper.py", line 386, in construct
loss = self.network(*inputs)
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 661, in __call__
raise err
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 657, in __call__
output = self._run_construct(args, kwargs)
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 445, in _run_construct
output = self.construct(*cast_inputs, **kwargs)
File "D:\python3.9\lib\site-packages\mindspore\nn\wrap\cell_wrapper.py", line 117, in construct
out = self._backbone(data)
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 661, in __call__
raise err
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 657, in __call__
output = self._run_construct(args, kwargs)
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 445, in _run_construct
output = self.construct(*cast_inputs, **kwargs)
File "D:\ai\step1\form\mylenet.py", line 40, in construct
x = self.max_pool2d(x)
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 661, in __call__
raise err
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 657, in __call__
output = self._run_construct(args, kwargs)
File "D:\python3.9\lib\site-packages\mindspore\nn\cell.py", line 445, in _run_construct
output = self.construct(*cast_inputs, **kwargs)
File "D:\python3.9\lib\site-packages\mindspore\nn\layer\pooling.py", line 562, in construct
if x.ndim == 3:
File "D:\python3.9\lib\site-packages\mindspore\common\_stub_tensor.py", line 122, in ndim
return len(self.shape)
File "D:\python3.9\lib\site-packages\mindspore\common\_stub_tensor.py", line 90, in shape
self.stub_shape = self.stub.get_shape()
RuntimeError: For 'Conv2D', 'C_in' of input 'x' shape divide by parameter 'group' must be equal to 'C_in' of input 'weight' shape: 1, but got 'C_in' of input 'x' shape: 3, and 'group': 1.
----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore\core\ops\conv2d.cc:241 mindspore::ops::`anonymous-namespace'::Conv2dInferShape
Process finished with exit code -1073741819 (0xC0000005)
上来就是运行报错，报错的具体信息是
RuntimeError: For 'Conv2D', 'C_in' of input 'x' shape divide by parameter 'group' must be equal to 'C_in' of input 'weight' shape: 1, but got 'C_in' of input 'x' shape: 3, and 'group': 1.
然后从打印出来的堆栈里面往上翻，找到自己写的代码行里面。
File "D:\ai\step1\form\mylenet.py", line 40, in construct
x = self.max_pool2d(x)
def construct(self, x):
# 使用定义好的运算构建前向网络
x = self.conv1(x)
x = self.relu(x)
x = self.max_pool2d(x)
x = self.conv2(x)
网络在x = self.conv1(x)报错了，报错信息里面是’Conv2D’这个算子。所以问题就在这里了。
但是为什么堆栈里面出错是在x = self.max_pool2d(x)呢？这个有人能回答下吗？
先看报错信息具体什么意思先。
'C_in' of input 'x' shape divide by parameter 'group' must be equal to 'C_in' of input 'weight' shape: 1, but got 'C_in' of input 'x' shape: 3, and 'group': 1.
大概意思是x的C_in的shape 除参数group应该等于输入weight的C_in的shape 1，但是实际输入x的C_in 的shape 3，group 1，两者相除等于3，就和weight的shape 1不一致了。
所以报错了。
这个报错感觉像个绕口令，还是先看下API文档怎么写的吧.
输入Tensor的shape为 (N,Cin,Hin,Win)
self.conv1 = nn.Conv2d(num_channel, 6, 5, pad_mode=‘valid’)
在conv1前面加打印，打印输入的shape
(32, 3, 32, 32)输入的C_in是3
num_channel=1这个就是不一致的地方了。
但是图像就是1位的，怎么变成了3位的呢？
那就看下数据读取时的处理了。
transforms.Decode()
好家伙，默认rgb=True。
改成False试试。
RuntimeError: Exception thrown from dataset pipeline. Refer to 'Dataset Pipeline Error Message'.
------------------------------------------------------------------
- Dataset Pipeline Error Message:
------------------------------------------------------------------
[ERROR] map operation: [Decode] failed. The corresponding data file is: D:\ai\step1\form\classify\0\1662618914169255200_3.png. Decode: only support Decoded into RGB image, check input parameter 'rgb' first, its value should be 'True'.
------------------------------------------------------------------
- C++ Call Stack: (For framework developers)
------------------------------------------------------------------
mindspore\ccsrc\minddata\dataset\kernels\image\decode_op.cc(48).
额， 只支持RGB。
那只能修改网络的num_channel了。
def __init__(self, num_class=10, num_channel=3):
修改之后就能正常跑下去了。
本次调试到此结束。
遗留的问题就是明明是x = self.conv1(x) 报的错为什么堆栈打到了x = self.max_pool2d(x)。还有就是Conv2D的报错信息不清晰。