【深度学习】图像多标签分类任务，百度PaddleClas_百度单图多标签标注

百度PaddleClas

百度PaddleClas GitHub链接：https://github.com/PaddlePaddle/PaddleClas。
在项目中PaddleClas/docs/en/advanced_tutorials/multilabel/multilabel_en.md目录下面是对多标签分类任务的指导，版本迭代找不到的话就搜索multilabel应该就能找到。本文跟着这个指导实战一遍。
我下载的分支是release/2.4，文章使用的代码和数据我都传百度云了一份：https://pan.baidu.com/s/19d7dSK075Vs_KzwmhwxGzA?pwd=e22x 。

Docker环境

之前pip安装paddle环境被坑惨了，这次直接上Docker得了，显卡V100找CUDA11版本，去PaddleClas路径执行：

sudo docker run --gpus all -v $PWD:/paddle -v /ssd/xiedong/datasets:/ssd/xiedong/datasets --shm-size=64G --network=host -it paddlepaddle/paddle:2.1.0-gpu-cuda11.2-cudnn8 /bin/bash
1

创建docker的时候默认shm大小为64M，所以要给shm-size=8G 。共享内存的一个介绍。

dockerHub：https://hub.docker.com/r/paddlepaddle/paddle/tags?page=1&name=gpu

If you want to use PaddlePaddle on GPU, you can use the following command to install PaddlePaddle.

pip install paddlepaddle-gpu --upgrade -i https://mirror.baidu.com/pypi/simple
1

If you want to use PaddlePaddle on CPU, you can use the following command to install PaddlePaddle.

pip install paddlepaddle --upgrade -i https://mirror.baidu.com/pypi/simple
1

数据准备

原始数据：https://lms.comp.nus.edu.sg/wp-content/uploads/2019/research/nuswide/NUS-WIDE.html

我使用百度处理好的NUS-WIDE-SCENE数据就舒服多了，在docker容器中执行：

cd /paddle/dataset

mkdir NUS-WIDE-SCENE
cd NUS-WIDE-SCENE
wget https://paddle-imagenet-models-name.bj.bcebos.com/data/NUS-SCENE-dataset.tar
tar -xf NUS-SCENE-dataset.tar
1
2
3
4
5
6

最终路径：

训练Train

首先修个BUG才行：
https://github.com/PaddlePaddle/PaddleClas/issues/2136

在docker容器中执行：

unset GREP_OPTIONS

cd /paddle && python -m pip install -i https://pypi.tuna.tsinghua.edu.cn/simple --upgrade pip && pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple && pip install -r requirements.txt
1
2
3

单gpu训练：

export CUDA_VISIBLE_DEVICES=0
python3 tools/train.py -c ./ppcls/configs/quick_start/professional/MobileNetV1_multilabel.yaml
1
2

多gpu训练：

export CUDA_VISIBLE_DEVICES=0,1,2
python3 -m paddle.distributed.launch --gpus="0,1,2" tools/train.py -c ./ppcls/configs/quick_start/professional/MobileNetV1_multilabel.yaml
1
2

训练完成的日志：

[2022/07/06 11:15:51] ppcls INFO: [Train][Epoch 10/10][Iter: 220/273]lr(CosineAnnealingDecay): 0.00008615, HammingDistance: 0.04837, AccuracyScore: 0.95163, MultiLabelLoss: 0.12250, loss: 0.12250, batch_cost: 0.11690s, reader_cost: 0.00017, ips: 547.48235 samples/s, eta: 0:00:06
[2022/07/06 11:15:53] ppcls INFO: [Train][Epoch 10/10][Iter: 230/273]lr(CosineAnnealingDecay): 0.00005571, HammingDistance: 0.04836, AccuracyScore: 0.95164, MultiLabelLoss: 0.12258, loss: 0.12258, batch_cost: 0.11685s, reader_cost: 0.00017, ips: 547.68878 samples/s, eta: 0:00:05
[2022/07/06 11:15:54] ppcls INFO: [Train][Epoch 10/10][Iter: 240/273]lr(CosineAnnealingDecay): 0.00003187, HammingDistance: 0.04824, AccuracyScore: 0.95176, MultiLabelLoss: 0.12239, loss: 0.12239, batch_cost: 0.11688s, reader_cost: 0.00017, ips: 547.55123 samples/s, eta: 0:00:03
[2022/07/06 11:15:55] ppcls INFO: [Train][Epoch 10/10][Iter: 250/273]lr(CosineAnnealingDecay): 0.00001466, HammingDistance: 0.04826, AccuracyScore: 0.95174, MultiLabelLoss: 0.12251, loss: 0.12251, batch_cost: 0.11687s, reader_cost: 0.00017, ips: 547.61131 samples/s, eta: 0:00:02
[2022/07/06 11:15:56] ppcls INFO: [Train][Epoch 10/10][Iter: 260/273]lr(CosineAnnealingDecay): 0.00000406, HammingDistance: 0.04826, AccuracyScore: 0.95174, MultiLabelLoss: 0.12253, loss: 0.12253, batch_cost: 0.11674s, reader_cost: 0.00017, ips: 548.24794 samples/s, eta: 0:00:01
[2022/07/06 11:15:57] ppcls INFO: [Train][Epoch 10/10][Iter: 270/273]lr(CosineAnnealingDecay): 0.00000006, HammingDistance: 0.04833, AccuracyScore: 0.95167, MultiLabelLoss: 0.12271, loss: 0.12271, batch_cost: 0.11677s, reader_cost: 0.00017, ips: 548.08781 samples/s, eta: 0:00:00
[2022/07/06 11:15:58] ppcls INFO: [Train][Epoch 10/10][Avg]HammingDistance: 0.04835, AccuracyScore: 0.95165, MultiLabelLoss: 0.12271, loss: 0.12271
[2022/07/06 11:16:00] ppcls INFO: [Eval][Epoch 10][Iter: 0/69]MultiLabelLoss: 0.09744, loss: 0.09744, HammingDistance: 0.03527, AccuracyScore: 0.96473, batch_cost: 2.53691s, reader_cost: 2.42421, ips: 100.91014 images/sec
[2022/07/06 11:16:05] ppcls INFO: [Eval][Epoch 10][Iter: 10/69]MultiLabelLoss: 0.12671, loss: 0.12671, HammingDistance: 0.05005, AccuracyScore: 0.94995, batch_cost: 0.38076s, reader_cost: 0.24564, ips: 672.34270 images/sec
[2022/07/06 11:16:10] ppcls INFO: [Eval][Epoch 10][Iter: 20/69]MultiLabelLoss: 0.11945, loss: 0.11945, HammingDistance: 0.04848, AccuracyScore: 0.95152, batch_cost: 0.49853s, reader_cost: 0.36578, ips: 513.50869 images/sec
[2022/07/06 11:16:15] ppcls INFO: [Eval][Epoch 10][Iter: 30/69]MultiLabelLoss: 0.12125, loss: 0.12125, HammingDistance: 0.04789, AccuracyScore: 0.95211, batch_cost: 0.47429s, reader_cost: 0.34168, ips: 539.75512 images/sec
[2022/07/06 11:16:20] ppcls INFO: [Eval][Epoch 10][Iter: 40/69]MultiLabelLoss: 0.11817, loss: 0.11817, HammingDistance: 0.04819, AccuracyScore: 0.95181, batch_cost: 0.49539s, reader_cost: 0.36272, ips: 516.76400 images/sec
[2022/07/06 11:16:24] ppcls INFO: [Eval][Epoch 10][Iter: 50/69]MultiLabelLoss: 0.10450, loss: 0.10450, HammingDistance: 0.04773, AccuracyScore: 0.95227, batch_cost: 0.47807s, reader_cost: 0.34755, ips: 535.48813 images/sec
[2022/07/06 11:16:30] ppcls INFO: [Eval][Epoch 10][Iter: 60/69]MultiLabelLoss: 0.11237, loss: 0.11237, HammingDistance: 0.04781, AccuracyScore: 0.95219, batch_cost: 0.49771s, reader_cost: 0.36855, ips: 514.35504 images/sec
[2022/07/06 11:16:32] ppcls INFO: [Eval][Epoch 10][Avg]MultiLabelLoss: 0.12195, loss: 0.12195, HammingDistance: 0.04765, AccuracyScore: 0.95235
[2022/07/06 11:16:32] ppcls INFO: [Eval][Epoch 10][best metric: 0.05279040170066246]
[2022/07/06 11:16:32] ppcls INFO: Already save model in ./output/MobileNetV1/epoch_10
[2022/07/06 11:16:33] ppcls INFO: Already save model in ./output/MobileNetV1/latest

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

评估Evaluation

评估Evaluation:

python tools/eval.py     -c ./ppcls/configs/quick_start/professional/MobileNetV1_multilabel.yaml     -o Arch.pretrained="./output/MobileNetV1/best_model"
1

日志：

[2022/07/06 11:37:38] ppcls INFO: train with paddle 2.1.0 and device CUDAPlace(0)
W0706 11:37:38.849289  1058 device_context.cc:404] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 11.2
W0706 11:37:38.854460  1058 device_context.cc:422] device: 0, cuDNN Version: 8.1.
[2022/07/06 11:37:45] ppcls INFO: [Eval][Epoch 0][Iter: 0/69]MultiLabelLoss: 0.11629, loss: 0.11629, HammingDistance: 0.04096, AccuracyScore: 0.95904, batch_cost: 3.46316s, reader_cost: 2.33148, ips: 73.92095 images/sec
[2022/07/06 11:37:48] ppcls INFO: [Eval][Epoch 0][Iter: 10/69]MultiLabelLoss: 0.14044, loss: 0.14044, HammingDistance: 0.05525, AccuracyScore: 0.94475, batch_cost: 0.45712s, reader_cost: 0.32888, ips: 560.03236 images/sec
[2022/07/06 11:37:54] ppcls INFO: [Eval][Epoch 0][Iter: 20/69]MultiLabelLoss: 0.13331, loss: 0.13331, HammingDistance: 0.05347, AccuracyScore: 0.94653, batch_cost: 0.51752s, reader_cost: 0.38610, ips: 494.66824 images/sec
[2022/07/06 11:37:58] ppcls INFO: [Eval][Epoch 0][Iter: 30/69]MultiLabelLoss: 0.14082, loss: 0.14082, HammingDistance: 0.05283, AccuracyScore: 0.94717, batch_cost: 0.48707s, reader_cost: 0.35611, ips: 525.59345 images/sec
[2022/07/06 11:38:04] ppcls INFO: [Eval][Epoch 0][Iter: 40/69]MultiLabelLoss: 0.13737, loss: 0.13737, HammingDistance: 0.05317, AccuracyScore: 0.94683, batch_cost: 0.50345s, reader_cost: 0.37316, ips: 508.48973 images/sec
[2022/07/06 11:38:08] ppcls INFO: [Eval][Epoch 0][Iter: 50/69]MultiLabelLoss: 0.12236, loss: 0.12236, HammingDistance: 0.05288, AccuracyScore: 0.94712, batch_cost: 0.48819s, reader_cost: 0.35907, ips: 524.38639 images/sec
[2022/07/06 11:38:14] ppcls INFO: [Eval][Epoch 0][Iter: 60/69]MultiLabelLoss: 0.13099, loss: 0.13099, HammingDistance: 0.05294, AccuracyScore: 0.94706, batch_cost: 0.50024s, reader_cost: 0.37077, ips: 511.75699 images/sec
[2022/07/06 11:38:16] ppcls INFO: [Eval][Epoch 0][Avg]MultiLabelLoss: 0.13865, loss: 0.13865, HammingDistance: 0.05279, AccuracyScore: 0.94721
1
2
3
4
5
6
7
8
9
10
11

预测Prediction 推理Infer

预测Prediction 推理Infer：

python3 tools/infer.py -c ./ppcls/configs/quick_start/professional/MobileNetV1_multilabel.yaml  -o Arch.pretrained="./output/MobileNetV1/best_model"
1

日志：

W0706 11:42:24.642689  1302 device_context.cc:404] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 11.2, Runtime API Version: 11.2
W0706 11:42:24.648531  1302 device_context.cc:422] device: 0, cuDNN Version: 8.1.
[{'class_ids': [6, 13, 17, 23, 30], 'scores': [0.99138, 0.83019, 0.5909, 0.99387, 0.91533], 'file_name': './deploy/images/0517_2715693311.jpg', 'label_names': []}]

1
2
3
4

images/0517_2715693311.jpg 这张图：

打印出数据里的NUS_labels.txt中的第6, 13, 17, 23, 30行（从0开始计数）： sed -n '7p;14p;18p;24p;31p' NUS_labels.txt （从1开始计数），即是模型得到的这五个类别：

clouds
lake
ocean
sky
water
1
2
3
4
5

而这张图的实际标签是什么呢，使用cat multilabel_train_list.txt |grep 0517_2715693311得到：

0517_2715693311.jpg     0,0,0,0,0,0,1,0,0,0,0,0,0,1,0,1,0,0,0,0,0,0,0,1,0,0,1,0,0,0,1,0,0
1

也就是7 14 16 24 27 31 （从1开始计数），sed -n '7p;14p;16p;24p;27p;31p' NUS_labels.txt ：

clouds
lake
mountain
sky
sunset
water
1
2
3
4
5
6

在这张图上效果一般。

导出模型Export model到onnx再到mnn

到onnx

官网介绍了一些转成paddle产品的：

python3 tools/export_model.py \
    -c ./ppcls/configs/quick_start/professional/MobileNetV1_multilabel.yaml \
    -o Arch.pretrained="./output/MobileNetV1/best_model"

cd ./deploy

python3 python/predict_cls.py \
     -c configs/inference_multilabel_cls.yaml

1
2
3
4
5
6
7
8
9

而我想转成onnx，看了Paddle2ONNX文档才发现，上个步骤也是需要的，需要模型结构文件：inference.pdmodel，模型参数文件inference.pdiparams。

安装：pip install paddle2onnx onnx onnx-simplifier onnxruntime-gpu

导出模型：paddle2onnx --model_dir inference/ --model_filename inference.pdmodel --params_filename inference.pdiparams --save_file model.onnx --opset_version 10 --enable_dev_version True --enable_onnx_checker True

参数选项

• 使用onnxruntime验证转换模型, 请注意安装最新版本（最低要求1.10.0）：

如你有ONNX模型优化的需求，推荐使用onnx-simplifier，也可使用如下命令对模型进行优化：

python -m paddle2onnx.optimize --input_model model.onnx --output_model new_model.onnx
1

如需要修改导出的模型输入形状，如改为静态shape:

python -m paddle2onnx.optimize --input_model model.onnx \
                               --output_model new_model.onnx \
                               --input_shape_dict "{'x':[1,3,224,224]}"
1
2
3

再到mnn

编译安装mnn：https://blog.csdn.net/x1131230123/article/details/125536750
转化onnx到mnn：~/MNN/build/MNNConvert -f ONNX --modelFile model.onnx --MNNModel model.mnn --bizCode MNN，看到成功提示：

inputTensors : [ x, ]
outputTensors: [ save_infer_model/scale_0.tmp_1, ]
Converted Success!
1
2
3

python 3.7 环境 pip install mnn opencv-python numpy 之后执行测试代码：

从头训练模型

生成别的label文件走一遍之前的训练和val流程，效果差不多：

sed -n '1,15000p' multilabel_train_list.txt > train.txt
sed -n '15001,17462p' multilabel_train_list.txt > test.txt
1
2

损失函数

import paddle
import paddle.nn as nn
import paddle.nn.functional as F


def ratio2weight(targets, ratio):
    pos_weights = targets * (1. - ratio)
    neg_weights = (1. - targets) * ratio
    weights = paddle.exp(neg_weights + pos_weights)

    # for RAP dataloader, targets element may be 2, with or without smooth, some element must great than 1
    weights = weights - weights * (targets > 1)

    return weights


class MultiLabelLoss(nn.Layer):
    """
    Multi-label loss
    """

    def __init__(self, epsilon=None, size_sum=False, weight_ratio=False):
        super().__init__()
        if epsilon is not None and (epsilon <= 0 or epsilon >= 1):
            epsilon = None
        self.epsilon = epsilon
        self.weight_ratio = weight_ratio
        self.size_sum = size_sum

    def _labelsmoothing(self, target, class_num):
        if target.ndim == 1 or target.shape[-1] != class_num:
            one_hot_target = F.one_hot(target, class_num)
        else:
            one_hot_target = target
        soft_target = F.label_smooth(one_hot_target, epsilon=self.epsilon)
        soft_target = paddle.reshape(soft_target, shape=[-1, class_num])
        return soft_target

    def _binary_crossentropy(self, input, target, class_num):
        if self.weight_ratio:
            target, label_ratio = target[:, 0, :], target[:, 1, :]
        if self.epsilon is not None:
            target = self._labelsmoothing(target, class_num)
        cost = F.binary_cross_entropy_with_logits(
            logit=input, label=target, reduction='none')

        if self.weight_ratio:
            targets_mask = paddle.cast(target > 0.5, 'float32')
            weight = ratio2weight(targets_mask, paddle.to_tensor(label_ratio))
            weight = weight * (target > -1)
            cost = cost * weight

        if self.size_sum:
            cost = cost.sum(1).mean() if self.size_sum else cost.mean()

        return cost

    def forward(self, x, target):
        if isinstance(x, dict):
            x = x["logits"]
        class_num = x.shape[-1]
        loss = self._binary_crossentropy(x, target, class_num)
        loss = loss.mean()
        return {"MultiLabelLoss": loss}

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65

试图看一些源码

engine

engine会启动模型，config即是对./ppcls/configs/quick_start/professional/MobileNetV1_multilabel.yaml的json：

model

得到loss、metric后，由这句得到模型：

>>>self.model
MobileNet(
  (conv): ConvBNLayer(
    (conv): Conv2D(3, 32, kernel_size=[3, 3], stride=[2, 2], padding=1, data_format=NCHW)
    (bn): BatchNorm()
    (relu): ReLU()
  )
  (blocks): Sequential(
    (0): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(32, 32, kernel_size=[3, 3], padding=1, groups=32, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(32, 64, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (1): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(64, 64, kernel_size=[3, 3], stride=[2, 2], padding=1, groups=64, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(64, 128, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (2): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(128, 128, kernel_size=[3, 3], padding=1, groups=128, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(128, 128, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (3): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(128, 128, kernel_size=[3, 3], stride=[2, 2], padding=1, groups=128, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(128, 256, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (4): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(256, 256, kernel_size=[3, 3], padding=1, groups=256, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(256, 256, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (5): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(256, 256, kernel_size=[3, 3], stride=[2, 2], padding=1, groups=256, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(256, 512, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (6): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[3, 3], padding=1, groups=512, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (7): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[3, 3], padding=1, groups=512, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (8): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[3, 3], padding=1, groups=512, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (9): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[3, 3], padding=1, groups=512, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (10): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[3, 3], padding=1, groups=512, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (11): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 512, kernel_size=[3, 3], stride=[2, 2], padding=1, groups=512, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(512, 1024, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
    (12): DepthwiseSeparable(
      (depthwise_conv): ConvBNLayer(
        (conv): Conv2D(1024, 1024, kernel_size=[3, 3], padding=1, groups=1024, data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
      (pointwise_conv): ConvBNLayer(
        (conv): Conv2D(1024, 1024, kernel_size=[1, 1], data_format=NCHW)
        (bn): BatchNorm()
        (relu): ReLU()
      )
    )
  )
  (avg_pool): AdaptiveAvgPool2D(output_size=1)
  (flatten): Flatten()
  (fc): Linear(in_features=1024, out_features=33, dtype=float32)
)

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170

train

万事俱备后，调用train：

替换backbone

ppcls.arch.backbone中含有很有backbone，而在arch的build_model中依赖getattr获取backbone，即下图：

这么多backbone，也可以大概看看：

paddle.summary(mod.MobileNetV2(),(1,3,224,224))
1

想来以后也得看，直接给个帖子汇总一下：https://blog.csdn.net/x1131230123/article/details/125661156。

修改backbone只需要修改这里的Arch里的东西即可：

onnx输入输出对比

略

推理的图片预处理

图片在输入模型之前做了什么处理？

所以把这里的方法对照给到函数里就行

参考链接

https://stats.stackexchange.com/questions/207794/what-loss-function-for-multi-class-multi-label-classification-tasks-in-neural-n

pytorch损失函数binary_cross_entropy和binary_cross_entropy_with_logits的区别

torch.nn.BCELoss用法

Depthwise卷积与Pointwise卷积