我正在尝试按照教程
构建图像分类器我成功完成了教程并在 CIFAR-10 数据集上获得了良好的准确性。
我正在尝试在 imagenet 数据集上运行分类器。
我对教程中的原始代码进行了一些修改，以使模型处理 imagenet 数据集中的图像，包括将所有图像调整为形状
(128, 128, 3)下面所有的代码和回溯也都在这个笔记本
batch_size = 32 # Batch size
image_size = 128 # Image size of training data
workers = 2 # Number of parallel workers
num_classes = 1000 # Number of classes
def create_dataset(data_set, usage, resize, batch_size, workers):
trans = []
trans += [
vision.Resize((128, 128)),
vision.Rescale(1.0 / 255.0, 0.0),
vision.HWC2CHW()
]
target_trans = transforms.TypeCast(mstype.int32)
data_set = data_set.map(operations=trans,
input_columns='image',
num_parallel_workers=workers)
data_set = data_set.map(operations=target_trans,
input_columns='label',
num_parallel_workers=workers)
data_set = data_set.batch(batch_size)
return data_set
import mindspore
from mindspore.dataset import vision, transforms
import mindspore.dataset as ds
trainset = ds.ImageFolderDataset("./imagenet2012/train", decode=True)
testset = ds.ImageFolderDataset("./imagenet2012/val", decode=True)
dataset_train = create_dataset(trainset,
usage="train",
resize=image_size,
batch_size=batch_size,
workers=workers)
dataset_val = create_dataset(testset,
usage="test",
resize=image_size,
batch_size=batch_size,
workers=workers)
step_size_val = dataset_val.get_dataset_size()
step_size_train = dataset_train.get_dataset_size()
我知道我最好使用填充物或其他东西。然而，上面的代码并不完美，尽管它可以工作，我通过采样一些图像、渲染检查形状来测试数据集，一切都很好。
上面的代码成功生成了 2 个数据集，即训练集和测试集。
以下代码定义了 ResidualBlockBase 类来实现构建块结构，与教程中的相同，我认为不需要修改即可应用于 imagenet 数据集。
from typing import Type, Union, List, Optional
import mindspore.nn as nn
from mindspore.common.initializer import Normal
# Initialize the parameters of the convolutional layer and BatchNorm layer
weight_init = Normal(mean=0, sigma=0.02)
gamma_init = Normal(mean=1, sigma=0.02)
class ResidualBlockBase(nn.Cell):
expansion: int = 1 # The number of convolution kernels at the last layer is the same as that of convolution kernels at the first layer.
def __init__(self, in_channel: int, out_channel: int,
stride: int = 1, norm: Optional[nn.Cell] = None,
down_sample: Optional[nn.Cell] = None) -> None:
super(ResidualBlockBase, self).__init__()
if not norm:
self.norm = nn.BatchNorm2d(out_channel)
else:
self.norm = norm
self.conv1 = nn.Conv2d(in_channel, out_channel,
kernel_size=3, stride=stride,
weight_init=weight_init)
self.conv2 = nn.Conv2d(in_channel, out_channel,
kernel_size=3, weight_init=weight_init)
self.relu = nn.ReLU()
self.down_sample = down_sample
def construct(self, x):
"""ResidualBlockBase construct."""
identity = x # shortcut
out = self.conv1(x) # First layer of the main body: 3 x 3 convolutional layer
out = self.norm(out)
out = self.relu(out)
out = self.conv2(out) # Second layer of the main body: 3 x 3 convolutional layer
out = self.norm(out)
if self.down_sample is not None:
identity = self.down_sample(x)
out += identity # output the sum of the main body and the shortcuts
out = self.relu(out)
return out
下面的代码定义了ResidualBlock类来实现bottleneck结构。与教程中的相同，我认为它也不需要修改即可应用于 imagenet 数据集。
class ResidualBlock(nn.Cell):
expansion = 4 # The number of convolution kernels at the last layer is four times that of convolution kernels at the first layer.
def __init__(self, in_channel: int, out_channel: int,
stride: int = 1, down_sample: Optional[nn.Cell] = None) -> None:
super(ResidualBlock, self).__init__()
self.conv1 = nn.Conv2d(in_channel, out_channel,
kernel_size=1, weight_init=weight_init)
self.norm1 = nn.BatchNorm2d(out_channel)
self.conv2 = nn.Conv2d(out_channel, out_channel,
kernel_size=3, stride=stride,
weight_init=weight_init)
self.norm2 = nn.BatchNorm2d(out_channel)
self.conv3 = nn.Conv2d(out_channel, out_channel * self.expansion,
kernel_size=1, weight_init=weight_init)
self.norm3 = nn.BatchNorm2d(out_channel * self.expansion)
self.relu = nn.ReLU()
self.down_sample = down_sample
def construct(self, x):
identity = x # shortcut
out = self.conv1(x) # First layer of the main body: 1 x 1 convolutional layer
out = self.norm1(out)
out = self.relu(out)
out = self.conv2(out) # Second layer of the main body: 3 x 3 convolutional layer
out = self.norm2(out)
out = self.relu(out)
out = self.conv3(out) # Third layer of the main body: 1 x 1 convolutional layer
out = self.norm3(out)
if self.down_sample is not None:
identity = self.down_sample(x)
out += identity # The output is the sum of the main body and the shortcut.
out = self.relu(out)
return out
以下示例定义 make_layer 来构建残差块。参数如下：
last_out_channel：前一个残差网络的输出通道数
block：残差网络类型。该值可以是 ResidualBlockBase 或 ResidualBlock。
channel：残差网络的输入通道数
block_nums：堆叠残差网络块的数量
stride：卷积运动的步幅
def make_layer(last_out_channel, block: Type[Union[ResidualBlockBase, ResidualBlock]],
channel: int, block_nums: int, stride: int = 1):
down_sample = None # shortcuts
if stride != 1 or last_out_channel != channel * block.expansion:
down_sample = nn.SequentialCell([
nn.Conv2d(last_out_channel, channel * block.expansion,
kernel_size=1, stride=stride, weight_init=weight_init),
nn.BatchNorm2d(channel * block.expansion, gamma_init=gamma_init)
])
layers = []
layers.append(block(last_out_channel, channel, stride=stride, down_sample=down_sample))
in_channel = channel * block.expansion
# Stack residual networks.
for _ in range(1, block_nums):
layers.append(block(in_channel, channel))
return nn.SequentialCell(layers)
以下示例代码用于构建ResNet-50模型。您可以调用resnet50函数构建ResNet-50模型。 resnet50函数的参数如下：
num_classes：类的数量。默认值为 1000。
预训练：下载对应的训练模型，并将预训练模型中的参数加载到网络中。
from mindspore import load_checkpoint, load_param_into_net
class ResNet(nn.Cell):
def __init__(self, block: Type[Union[ResidualBlockBase, ResidualBlock]],
layer_nums: List[int], num_classes: int, input_channel: int) -> None:
super(ResNet, self).__init__()
self.relu = nn.ReLU()
# At the first convolutional layer, the number of the input channels is 3 (color image) and that of the output channels is 64.
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, weight_init=weight_init)
self.norm = nn.BatchNorm2d(64)
# Maximum pooling layer, reducing the image size
self.max_pool = nn.MaxPool2d(kernel_size=3, stride=2, pad_mode='same')
# Define each residual network structure block
self.layer1 = make_layer(64, block, 64, layer_nums[0])
self.layer2 = make_layer(64 * block.expansion, block, 128, layer_nums[1], stride=2)
self.layer3 = make_layer(128 * block.expansion, block, 256, layer_nums[2], stride=2)
self.layer4 = make_layer(256 * block.expansion, block, 512, layer_nums[3], stride=2)
# average pooling layer
self.avg_pool = nn.AvgPool2d()
# flattern layer
self.flatten = nn.Flatten()
# fully-connected layer
self.fc = nn.Dense(in_channels=input_channel, out_channels=num_classes)
def construct(self, x):
x = self.conv1(x)
x = self.norm(x)
x = self.relu(x)
x = self.max_pool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avg_pool(x)
x = self.flatten(x)
x = self.fc(x)
return x
def _resnet(model_url: str, block: 类型[Union[ResidualBlockBase, ResidualBlock]], 层：List [int]，num_classes：int，预训练：bool，pretrained_ckpt：str， 输入通道：整数）： 模型 = ResNet(块、层、num_classes、input_channel)
if pretrained:
# load pre-trained models
download(url=model_url, path=pretrained_ckpt, replace=True)
param_dict = load_checkpoint(pretrained_ckpt)
load_param_into_net(model, param_dict)
return model
def resnet50(num_classes: int = 1000, pretrained: bool = False):
"ResNet50 model"
resnet50_url = "https://obs.dualstack.cn-north-4.myhuaweicloud.com/mindspore-website/notebook/models/application/resnet50_224_new.ckpt"
resnet50_ckpt = "./LoadPretrainedModel/resnet50_224_new.ckpt"
return _resnet(resnet50_url, ResidualBlock, [3, 4, 6, 3], num_classes,
pretrained, resnet50_ckpt, 2048)
我不确定上面代码中的最后一行是否需要修改才能应用于我的新数据。因为它似乎与我遇到的错误有关。
我将
out_channels# Define the ResNet50 network.
network = resnet50(pretrained=True)
# Size of the input layer of the fully-connected layer
in_channel = network.fc.in_channels
fc = nn.Dense(in_channels=in_channel, out_channels=1000)
# Reset the fully-connected layer.
network.fc = fc
我认为以下超参数不需要修改，所以我保持不变。
# Set the learning rate
num_epochs = 5
lr = nn.cosine_decay_lr(min_lr=0.00001, max_lr=0.001, total_step=step_size_train * num_epochs,
step_per_epoch=step_size_train, decay_epoch=num_epochs)
# Define optimizer and loss function
opt = nn.Momentum(params=network.trainable_params(), learning_rate=lr, momentum=0.9)
loss_fn = nn.SoftmaxCrossEntropyWithLogits(sparse=True, reduction='mean')
def forward_fn(inputs, targets):
logits = network(inputs)
loss = loss_fn(logits, targets)
return loss
grad_fn = ms.value_and_grad(forward_fn, None, opt.parameters)
def train_step(inputs, targets):
loss, grads = grad_fn(inputs, targets)
opt(grads)
return loss
我认为以下代码不需要修改，所以我保持不变。
import os
# Creating Iterators
data_loader_train = dataset_train.create_tuple_iterator(num_epochs=num_epochs)
data_loader_val = dataset_val.create_tuple_iterator(num_epochs=num_epochs)
# Optimal model storage path
best_acc = 0
best_ckpt_dir = "./BestCheckpoint"
best_ckpt_path = "./BestCheckpoint/resnet50-best.ckpt"
if not os.path.exists(best_ckpt_dir):
os.mkdir(best_ckpt_dir)
import mindspore.ops as ops
def train(data_loader, epoch):
"""Model taining"""
losses = []
network.set_train(True)
for i, (images, labels) in enumerate(data_loader):
loss = train_step(images, labels)
if i % 100 == 0 or i == step_size_train - 1:
print('Epoch: [%3d/%3d], Steps: [%3d/%3d], Train Loss: [%5.3f]' %
(epoch + 1, num_epochs, i + 1, step_size_train, loss))
losses.append(loss)
return sum(losses) / len(losses)
def evaluate(data_loader):
"""Model Evaluation"""
network.set_train(False)
correct_num = 0.0 # Number of correct predictions
total_num = 0.0 # Total number of predictions
for images, labels in data_loader:
logits = network(images)
pred = logits.argmax(axis=1) # Prediction results
correct = ops.equal(pred, labels).reshape((-1, ))
correct_num += correct.sum().asnumpy()
total_num += correct.shape[0]
acc = correct_num / total_num # Accuracy
return acc
以下代码抛出 ValueError
对于“MatMul”，输入维度必须相等，但得到“x1_col”：32768 和“x2_row”：2048。
print("Start Training Loop ...")
for epoch in range(num_epochs):
curr_loss = train(data_loader_train, epoch)
curr_acc = evaluate(data_loader_val)
print("-" * 50)
print("Epoch: [%3d/%3d], Average Train Loss: [%5.3f], Accuracy: [%5.3f]" % (
epoch+1, num_epochs, curr_loss, curr_acc
))
print("-" * 50)
# Save the model that has achieved the highest prediction accuracy
if curr_acc > best_acc:
best_acc = curr_acc
ms.save_checkpoint(network, best_ckpt_path)
print("=" * 80)
print(f"End of validation the best Accuracy is: {best_acc: 5.3f}, "
f"save the best ckpt file in {best_ckpt_path}", flush=True)
我什至不知道从哪里开始检查我的代码，因为我没有对教程中的原始代码进行太多更改
任何有关提示或解决方案的建议将不胜感激。
这是回溯
ValueError Traceback (most recent call last)
Cell In[18], line 5
2 print("Start Training Loop ...")
4 for epoch in range(num_epochs):
----> 5 curr_loss = train(data_loader_train, epoch)
6 curr_acc = evaluate(data_loader_val)
8 print("-" * 50)
Cell In[17], line 10, in train(data_loader, epoch)
7 network.set_train(True)
9 for i, (images, labels) in enumerate(data_loader):
---> 10 loss = train_step(images, labels)
11 if i % 100 == 0 or i == step_size_train - 1:
12 print('Epoch: [%3d/%3d], Steps: [%3d/%3d], Train Loss: [%5.3f]' %
13 (epoch + 1, num_epochs, i + 1, step_size_train, loss))
Cell In[15], line 20, in train_step(inputs, targets)
19 def train_step(inputs, targets):
---> 20 loss, grads = grad_fn(inputs, targets)
21 opt(grads)
22 return loss
File ~/miniconda3/lib/python3.9/site-packages/mindspore/ops/composite/base.py:620, in _Grad.__call__.<locals>.after_grad(*args, **kwargs)
619 def after_grad(*args, **kwargs):
--> 620 return grad_(fn_, weights)(*args, **kwargs)
File ~/miniconda3/lib/python3.9/site-packages/mindspore/common/api.py:106, in _wrap_func.<locals>.wrapper(*arg, **kwargs)
104 @wraps(fn)
105 def wrapper(*arg, **kwargs):
--> 106 results = fn(*arg, **kwargs)
107 return _convert_python_data(results)
File ~/miniconda3/lib/python3.9/site-packages/mindspore/ops/composite/base.py:595, in _Grad.__call__.<locals>.after_grad(*args, **kwargs)
593 @_wrap_func
594 def after_grad(*args, **kwargs):
--> 595 res = self._pynative_forward_run(fn, grad_, weights, args, kwargs)
596 _pynative_executor.grad(fn, grad_, weights, grad_position, *args, **kwargs)
597 out = _pynative_executor()
File ~/miniconda3/lib/python3.9/site-packages/mindspore/ops/composite/base.py:645, in _Grad._pynative_forward_run(self, fn, grad, weights, args, kwargs)
643 _pynative_executor.set_grad_flag(True)
644 _pynative_executor.new_graph(fn, *args, **new_kwargs)
--> 645 outputs = fn(*args, **new_kwargs)
646 _pynative_executor.end_graph(fn, outputs, *args, **new_kwargs)
647 return outputs
Cell In[15], line 11, in forward_fn(inputs, targets)
10 def forward_fn(inputs, targets):
---> 11 logits = network(inputs)
12 loss = loss_fn(logits, targets)
13 return loss
File ~/miniconda3/lib/python3.9/site-packages/mindspore/nn/cell.py:662, in Cell.__call__(self, *args, **kwargs)
660 except Exception as err:
661 _pynative_executor.clear_res()
--> 662 raise err
664 if isinstance(output, Parameter):
665 output = output.data
File ~/miniconda3/lib/python3.9/site-packages/mindspore/nn/cell.py:659, in Cell.__call__(self, *args, **kwargs)
657 _pynative_executor.new_graph(self, *args, **kwargs)
658 output = self._run_construct(args, kwargs)
--> 659 _pynative_executor.end_graph(self, output, *args, **kwargs)
660 except Exception as err:
661 _pynative_executor.clear_res()
File ~/miniconda3/lib/python3.9/site-packages/mindspore/common/api.py:1304, in _PyNativeExecutor.end_graph(self, obj, output, *args, **kwargs)
1291 def end_graph(self, obj, output, *args, **kwargs):
1292 """
1293 Clean resources after building forward and backward graph.
1294
(...)
1302 None.
1303 """
-> 1304 self._executor.end_graph(obj, output, *args, *(kwargs.values()))
ValueError: For 'MatMul' the input dimensions must be equal, but got 'x1_col': 32768 and 'x2_row': 2048.
----------------------------------------------------
- C++ Call Stack: (For framework developers)
----------------------------------------------------
mindspore/core/ops/mat_mul.cc:101 InferShape
For 'MatMul' the input dimensions must be equal, but got 'x1_col': 32768 and 'x2_row': 2048.如果您查看错误消息中的值，您可能会发现
32768 = 128x128x22048 = 32x32x2def resnet50(num_classes: int = 1000, pretrained: bool = False):
...
return _resnet(resnet50_url, ResidualBlock, [3, 4, 6, 3], num_classes, pretrained, resnet50_ckpt, 2048)