在最后一步
cd PaddleViT/semantic_segmentation pip3 install -r requirements.txt
报错，提示如下
(paddlevit) D:\PyWorkspace\PaddleViT\PaddleViT\semantic_segmentation>pip3 install -r requirements.txt Collecting cityscapesScripts==2.2.0 Using cached cityscapesScripts-2.2.0-py3-none-any.whl (472 kB) ERROR: Could not find a version that satisfies the requirement detail==4.0 (from versions: none) ERROR: No matching distribution found for detail==4.0
尝试直接安装detail包
(paddlevit) D:\PyWorkspace\PaddleViT\PaddleViT\semantic_segmentation>pip3 install detail==4.0 ERROR: Could not find a version that satisfies the requirement detail==4.0 (from versions: none) ERROR: No matching distribution found for detail==4.0 (paddlevit) D:\PyWorkspace\PaddleViT\PaddleViT\semantic_segmentation>pip3 install detail ERROR: Could not find a version that satisfies the requirement detail (from versions: none) ERROR: No matching distribution found for detail
更新pip
(paddlevit) D:\PyWorkspace\PaddleViT\PaddleViT\semantic_segmentation>pip install --user --upgrade pip Requirement already satisfied: pip in d:\anaconda\envs\paddlevit\lib\site-packages (21.3.1)
仍然无法安装
(paddlevit) D:\PyWorkspace\PaddleViT\PaddleViT\semantic_segmentation>pip3 install detail ERROR: Could not find a version that satisfies the requirement detail (from versions: none) ERROR: No matching distribution found for detail
百度没有找到相关包
最后安装了details-0.2.0的工具包，不知道如何使用。

I train the landcover dataset in the image segment model you provide, but the loss does not decrease.

Does this model reproduced so far ?
BEiT-L (ViT+UperNet, ImageNet-22k pretrain) mIoU-57.0%
https://paperswithcode.com/paper/beit-bert-pre-training-of-image-transformers

PaddleViT/image_classification/ViT/transformer.py Encoder 初始化创建 encoder_layer 时，深拷贝的意义是什么呢？

class Encoder(nn.Layer):

    def __init__(self,
                 embed_dim,
                 num_heads,
                 depth,
                 qkv_bias=True,
                 mlp_ratio=4.0,
                 dropout=0.,
                 attention_dropout=0.,
                 droppath=0.):
        super(Encoder, self).__init__()
        # stochatic depth decay
        depth_decay = [x.item() for x in paddle.linspace(0, droppath, depth)]
        layer_list = []
        for i in range(depth):
            encoder_layer = EncoderLayer(embed_dim,
                                         num_heads,
                                         qkv_bias=qkv_bias,
                                         mlp_ratio=mlp_ratio,
                                         dropout=dropout,
                                         attention_dropout=attention_dropout,
                                         droppath=depth_decay[i])
            layer_list.append(copy.deepcopy(encoder_layer))  # 这里对encoder_layer做深拷贝的意义是什么呢？
        self.layers = nn.LayerList(layer_list)
……

for 循环创建 encoder_layer，每一个 encoder_layer 都是不同的对象，也不存在参数共享的问题，这里对 encoder_layer 做深拷贝是基于什么样的考量呢？
我个人认为这里对 encoder_layer 的深拷贝是不必要的：

layer_list.append(encoder_layer)

可能是我想的还不够深，期待官方的答疑解惑~

Error:
Training will fail when using larger batch：
SystemError: (Fatal) Operator set_value raises an thrust::system::system_error exception. The exception content is :parallel_for failed: cudaErrorInvalidConfiguration: invalid configuration argument. (at /paddle/paddle/fluid/imperative/tracer.cc:192)

Reason:
The reason is explained by the following issues from PaddlePaddle：
PaddlePaddle/Paddle#33057 (comment)

In short, this error is raised because of cuda thrust bug, which is ignored in newer version cuda.

Solution:
install paddle dev version will fix the problem.
You will find the following instructions of how to install it:
https://www.paddlepaddle.org.cn/install/quick?docurl=/documentation/docs/zh/develop/install/pip/linux-pip.html

In detail, the problem is fixed by the following patch:
https://github.com/PaddlePaddle/Paddle/pull/33748/files/617e3eda9dfcd76cb6a7ebaa1535340f1023d3f1

Describe the bug
在使用ViT基于Imagenet2012数据集进行模型验证时，因为数据集比较大，仅下载了验证数据集。执行模型验证命令，报错，提示没有训练集数据。

经排查，
main_single_gpu.py 在进行模型验证前，需要先加载训练集和验证集。

建议单独增加一个模型验证的脚本。

Describe your feature request
Add linear_scale_lr arguments in config and control the batch_size for linear lr scale

Describe the reference code or paper
N/A

Describe the possible solution
Add argument in config.py
Add if-condition in main_single_gpu.py and main_multi_gpu.py

Additional context
N/A

Describe your feature request
Token labeling is used in VOLO model training, implement related classes and methods.

Describe the reference code or paper

1. The soft target loss and token label loss (official code: https://github.com/sail-sg/volo/blob/main/loss/cross_entropy.py)
2. Token Labeling in main (official code: https://github.com/sail-sg/volo/blob/main/main.py)
3. Token Label related classes and methods (official code: https://github.com/zihangJiang/TokenLabeling/tree/main/tlt/data)
4. The dense label map download link (9.2G) is listed in VOLO official repo: https://github.com/sail-sg/volo#4-train

Describe the possible solution
Implemented according to the official code

Additional context
N/A

I can't find package detail and yaml in PyPI.

https://github.com/BR-IDL/PaddleViT/blob/develop/semantic_segmentation/src/models/backbones/cswin_transformer.py#L537

Describe the bug
Some of the main_multi_gpu.py in image classification model miss the paddle.DataParalle(model), which may cause the multi-gpu training performs not well.

E.g., in DeiT:

Expected behavior
Add paddle.DataParallel(model)

现有的预训练参数文件都是在GitHub的项目链接里，这种方式虽然可视化效果好，但实际应用时存在诸多不便：

使用AiStudio或者本地开发时，需要将所有的权重文件提前下载（谷歌云盘或者百度网盘），之后再将该权重文件上传至指定位置，程序较繁琐；

建议

是否可以效仿PaddleSeg在保留GitHub项目链接的同时，在具体的项目配置文件中指定pretrain-model等预训练权重参数文件的下载链接，这样在模型训练时，可以依据配置文件自动下载权重，例如下图：

Describe your feature request
Check and modify the training settings for CrossViT models, add missing processing and training methods.

Describe the reference code or paper

• Paper link: https://arxiv.org/pdf/2103.14899.pdf
• Official Code: https://github.com/IBM/CrossViT

Describe the possible solution

• Paper:
  • "... based on DeiT, and apply their default hyper-parameter for training..."
  • Augmentation: rand augmentation (m9n2), mixup(0.8), cutmix(1.0), random erasing(0.25)
  • Droppath: 0.1
  • label smoothing: 0.1
  • Epochs: 300 (30 warm-ups)
  • 32 GPUs
  • Batch size: 4096
  • cosine lr decay, linear warmup
  • init lr: 0.004
  • weight decay: 0.05
  • optimizer: AdamW
  • warmup start lr: 1e-6
• Code:
  • dropout: 0.0
  • droppath: 0.1
  • clip-grad: None
  • min lr: 1e-5
• Model weights init:
  • linear layer: trunc_normal (.02)
  • layernorm layer: constant weight (1.0), bias (0.0)
  • official code here

Additional context
Add any other context or screenshots about the feature request here.

当我在 AIStudio （经典版）运行 README.md "Quick Demo for Image Classification" 中的示例：

%cd PaddleViT/image_classification/ViT/

import paddle

from config import get_config
from transformer import build_vit as build_model


# config files in ./configs/
config = get_config('./configs/vit_base_patch16_224.yaml')
# build model
model = build_model(config)
# load pretrained weights, .pdparams is NOT needed
model_state_dict = paddle.load('./vit_base_patch16_224')
model.set_dict(model_state_dict)

加载预训练权重时，出现了如下错误：

/home/aistudio/PaddleViT/image_classification/ViT
merging config from ./configs/vit_base_patch16_224.yaml
W1123 07:14:56.871081  9894 device_context.cc:447] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 10.1, Runtime API Version: 10.1
W1123 07:14:56.875849  9894 device_context.cc:465] device: 0, cuDNN Version: 7.6.
---------------------------------------------------------------------------ValueError                                Traceback (most recent call last)/tmp/ipykernel_9894/1201834454.py in <module>
     12 model = build_model(config)
     13 # load pretrained weights, .pdparams is NOT needed
---> 14 model_state_dict = paddle.load('./vit_base_patch16_224')
     15 model.set_dict(model_state_dict)
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/framework/io.py in load(path, **configs)
    983 
    984     else:
--> 985         load_result = _legacy_load(path, **configs)
    986 
    987     return load_result
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/framework/io.py in _legacy_load(path, **configs)
   1001     else:
   1002         # file prefix and directory are compatible cases
-> 1003         model_path, config = _build_load_path_and_config(path, config)
   1004         # check whether model file exists
   1005         if config.model_filename is None:
/opt/conda/envs/python35-paddle120-env/lib/python3.7/site-packages/paddle/framework/io.py in _build_load_path_and_config(path, config)
    159                 "example, it should be written as `paddle.load('model.pdparams')` instead of " \
    160                 "`paddle.load('model')`."
--> 161         raise ValueError(error_msg % path)
    162     else:
    163         if prefix_format_exist:
ValueError: The ``path`` (./vit_base_patch16_224) to load model not exists. If you want to load the results saved by `fluid.save_dygraph`, please specify the full file name, not just the file name prefix. For example, it should be written as `paddle.load('model.pdparams')` instead of `paddle.load('model')`.

错误信息显示，加载预训练权重时，需要指定预训练文件的完整文件名，即 .pdparams 是必须的。
加载预训练权重时，加上 .pdparams 后缀就没有问题了：

model_state_dict = paddle.load('./vit_base_patch16_224.pdparams')

PaddleViT/image_classification/SwinTransformer/datasets.py 下

在paddle的文档中，datasets.Cifar10 的预测mode 应该为'test'而不是val

因此我将代码改为

就可以正常跑通了

Describe your feature request
Add recompute for dygraph model training, which aims enlarge the batchsize during the training by free intermediate memories.

Describe the reference code or paper
N/A

@jarygrace Let's add this feature asap, thanks!

Describe your feature request
Now dataset.py has hard coded imagenet mean and var, which is not flexible and not easy to find for new users.
This can be set with an argument in config file, users can set the default mean, var values.

Describe your feature request
Reproduce the CvT model

Describe the reference code or paper
Paper: https://arxiv.org/pdf/2103.15808.pdf
official repo: https://github.com/VITA-Group/TransGAN

Describe the possible solution

1. implement the model (refer to official impl. here)
2. Add configs for different model (refer to official impl. here)
3. port official pretrained weights (refer to official impl. here)
4. Debug and evaluate PaddleViT impl.
5. Impl model training

Additional context
N/A

为什么我用自己的数据集按照配置单卡训练ViT，效果很差。同一份预料，我在paddle-resnet上可以达到99%精度

。

PaddleViT/image_classification/ViT/transformer.py 的 L300 创建 EncoderLayer 时，参数 qkv_bias, mlp_ratio, dropout, attention_dropout存在硬编码的问题，导致 Encoder 的 __init__ 方法传入的参数形同虚设：

class Encoder(nn.Layer):
    """Transformer encoder
    Encoder encoder contains a list of EncoderLayer, and a LayerNorm.
    Attributes:
        layers: nn.LayerList contains multiple EncoderLayers
        encoder_norm: nn.LayerNorm which is applied after last encoder layer
    """
    def __init__(self,
                 embed_dim,
                 num_heads,
                 depth,
                 qkv_bias=True,
                 mlp_ratio=4.0,
                 dropout=0.,
                 attention_dropout=0.,
                 droppath=0.):
        super(Encoder, self).__init__()
        # stochatic depth decay
        depth_decay = [x.item() for x in paddle.linspace(0, droppath, depth)]
        layer_list = []
        for i in range(depth):
            encoder_layer = EncoderLayer(embed_dim,
                                         num_heads,
                                         qkv_bias=True,
                                         mlp_ratio=4.,
                                         dropout=0.,
                                         attention_dropout=0.,
                                         droppath=depth_decay[i])
            layer_list.append(copy.deepcopy(encoder_layer))
        self.layers = nn.LayerList(layer_list)
……

应该改成：

class Encoder(nn.Layer):
    """Transformer encoder
    Encoder encoder contains a list of EncoderLayer, and a LayerNorm.
    Attributes:
        layers: nn.LayerList contains multiple EncoderLayers
        encoder_norm: nn.LayerNorm which is applied after last encoder layer
    """
    def __init__(self,
                 embed_dim,
                 num_heads,
                 depth,
                 qkv_bias=True,
                 mlp_ratio=4.0,
                 dropout=0.,
                 attention_dropout=0.,
                 droppath=0.):
        super(Encoder, self).__init__()
        # stochatic depth decay
        depth_decay = [x.item() for x in paddle.linspace(0, droppath, depth)]
        layer_list = []
        for i in range(depth):
            encoder_layer = EncoderLayer(embed_dim,
                                         num_heads,
                                         qkv_bias=qkv_bias,
                                         mlp_ratio=mlp_ratio,
                                         dropout=dropout,
                                         attention_dropout=attention_dropout,
                                         droppath=depth_decay[i])
            layer_list.append(copy.deepcopy(encoder_layer))
        self.layers = nn.LayerList(layer_list)
……

以上~

Describe your feature request
Styleformer training from scratch

Describe the reference code or paper
official repo: https://github.com/Jeeseung-Park/Styleformer

Describe the possible solution
TODO

Additional context
Add any other context or screenshots about the feature request here.

https://github.com/BR-IDL/PaddleViT/blob/develop/semantic_segmentation/configs/upernet_swin/upernet_swin_base_patch4_windown7_512x512_160k_ade20k.yaml

train with 8 cards, but the result is less than expected:

#--------------------------------------------------
2021-11-18 14:47:19 [INFO] [EVAL] #Images: 2000 mIoU: 0.0108 Acc: 0.3663 Kappa: 0.2675
2021-11-18 14:47:19 [INFO] [EVAL] Class IoU:
[2.903e-01 3.102e-01 6.999e-01 0.000e+00 3.243e-01 0.000e+00 2.000e-04
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00 0.000e+00
0.000e+00 0.000e+00 0.000e+00]
2021-11-18 14:47:19 [INFO] [EVAL] Class Acc:
[0.3054 0.3275 0.7344 0. 0.3642 0. 1. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0.
0. 0. 0. 0. 0. 0. 0. 0. 0. 0. ]
#--------------------------------------------------

Describe your feature request
Add onecycle scheduler, which is used in convmixer

Describe the reference code or paper
N/A

Describe the possible solution
refered to onecycle impl in timm

Additional context
N/A

Steps to reproduce the behavior:

1. Go to 'PaddleViT/object_detection/Swin/'
2. Run 'main_single_gpu.py'

Additional context

issue
github-PaddleViT/object_detection/Swin/中
main_single_gpu.py : line391 缺少train_loss定义
utils.py : 缺少import math

Describe your feature request
Will PaddleViT recently add Mobile-Former and release pretrained weights on ImageNet?

Describe the reference code or paper
Paper -> Mobile-Former: Bridging MobileNet and Transformer

Describe the possible solution

Additional context
Add any other context or screenshots about the feature request here.

The source code is empty. I want a reference. Please check.

PaddleViT/image_classification/ViT/transformer.py VisualTransformer 模型的输出是不是少了一个 attn:

class VisualTransformer(nn.Layer):
    ……
    def forward(self, x):
        x = self.patch_embedding(x)
        x, attn = self.encoder(x)
        logits = self.classifier(x[:, 0]) # take only cls_token as classifier
        return logits

我个人认为，模型的输出应该同时返回 attn的：

class VisualTransformer(nn.Layer):
    ……
    def forward(self, x):
        x = self.patch_embedding(x)
        x, attn = self.encoder(x)
        logits = self.classifier(x[:, 0]) # take only cls_token as classifier
        return logits, attn

理由如下：

• 其一，每层注意力注意力权重 attn，在 Attention，EncoderLayer 和 Encoder 中一直都是由返回的，如果模型输出不返回 attn，那么前面几个类的返回将会是多余的，可能毫无意义。
• 其二，每层注意力权重 attn，在后期的可视化中可能回用到。我猜前面几个类返回每层注意力权重，这样的设计可能也是基于可视化的考量的。

综上，建议模型输出同时返回每层的注意力权重 attn~

文档中多处还使用“SETR”，但是目录文件夹是小写

【原文】
CUDA_VISIBLE_DEVICES=0 python3 train.py
--config ./configs/SETR/SETR_MLA_Large_480x480_80k_pascal_context_bs_8.yaml
【应修改】
CUDA_VISIBLE_DEVICES=0 python3 train.py
--config ./configs/setr/SETR_MLA_Large_480x480_80k_pascal_context_bs_8.yaml

Describe your feature request
Add the multi-scale sampler for multi gpu batch sampling

Describe the reference code or paper
Refer to MobileViT paper: https://arxiv.org/pdf/2110.02178.pdf
Page 18.

Describe the possible solution
Refer to original paper.

Additional context
N/A

Describe your feature request
LabelSmoothingCrossEntropyLoss增加可选参数
losses.py中（以DeiT为例）LabelSmoothingCrossEntropyLoss可选参数较少

Describe the reference code or paper

Describe the possible solution
直接调用paddle.nn.functional.cross_entropy计算，可以设置更多参数。
调用前先利用paddle.nn.functional.one_hot将标签转为on-hot形式，再用 paddle.nn.functional.label_smooth将标签平滑，最后将paddle.nn.functional.cross_entropy的soft_label设为True即可实现。
代码如下：

class LabelSmoothingCrossEntropyLoss(nn.Layer):
    def __init__(self,
                 smoothing=0.1,
                 weight=None,
                 ignore_index=-100,
                 reduction='mean',
                 soft_label=True,
                 axis=-1,
                 use_softmax=True,
                 name=None):
        super(LabelSmoothingCrossEntropyLoss, self).__init__()
        assert 0 <= smoothing < 1.0
        self.smoothing = smoothing
        self.weight = weight
        self.reduction = reduction
        self.ignore_index = ignore_index
        self.soft_label = soft_label
        self.axis = axis
        self.use_softmax = use_softmax
        self.name = name

    def forward(self, input, label):
        label = paddle.nn.functional.one_hot(label, num_classes=input.shape[1])
        label = paddle.nn.functional.label_smooth(label, epsilon=self.smoothing)        
        ret = paddle.nn.functional.cross_entropy(
            input,
            label,            
            weight=self.weight,
            ignore_index=self.ignore_index,
            reduction=self.reduction,
            soft_label=self.soft_label,
            axis=self.axis,
            use_softmax=self.use_softmax,
            name=self.name)
        return ret

目前，经过简单测试结果和现有方法计算结果一致。

Additional context
Paddle ViT课程训练ResNet18作业中发现，过拟合比较严重，所以想尝试利用Label Smoothing方法缓解。但是搜索paddle api官方文档后发现没有专门的LabelSmoothingCrossEntropyLoss，利用paddle现成的one_hot和label_smooth函数实现了一下。

Add HaloNet and align ported weights performance.

Reference paper: https://arxiv.org/pdf/2103.12731.pdf

Timm impl: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/byoanet.py

如果是多卡训练，则需要初始化多卡训练环境。

if nranks > 1:
    # Initialize parallel environment if not done.
    if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized():
        logger.info("using dist training")
        # 初始化动态图模式下的并行训练环境，目前同时初始化NCCL和GLOO上下文用于通信。
        paddle.distributed.init_parallel_env()
        ddp_model = paddle.DataParallel(model)
    else:
        ddp_model = paddle.DataParallel(model)

不理解：if not paddle.distributed.parallel.parallel_helper._is_parallel_ctx_initialized(): 这个判断是什么意思，难道paddle.distributed.init_parallel_env()不应该是必须初始化的吗？

另外，还有如果要运行多机多卡需要修改什么代码吗？
谢谢

在 vit transformer 的实现中（ViT Transformer Attention），多头注意力的 attn_head_size 的计算是由传入的 embed_dim 和 num_heads 计算得到的：

self.attn_head_size = int(embed_dim / self.num_heads)

我认为这里的实现至少有两个问题：

• 其一，没有对embed_dim是否能num_heads整除做检查。当embed_dim不能被num_heads整除，或者num_heads > embed_dim时，transpose_multihead 的操作会出现异常：

    def transpose_multihead(self, x):
        new_shape = x.shape[:-1] + [self.num_heads, self.attn_head_size]
        x = x.reshape(new_shape)
        x = x.transpose([0, 2, 1, 3])
        return x

• 其二，attn_head_size 的大小受到 embed_dim 和 num_heads 的限制，当预训练模型时，不能随意设置 attn_head_size 的大小，代码不够灵活。

解决上述问题的办法，就是为 Attention 的 __init__ 方法添加一个 attn_head_size 的参数，这样即不影响现有预训练模型的加载，又可以在预训练时，灵活设置 attn_head_size 的大小。由于 attn_head_size 与输入维度 embed_dim 无关，也不需要验证 embed_dim 是否能被 num_heads 整除。
目前主流框架中，两种实现都有：
第一种，由 embed_dim 和 num_heads 参数计算 attn_head_size 的实现，包括:
PaddlePaddle: https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/nn/layer/transformer.py#L109
PyTorch: https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/transformer.py
transformers: https://github.com/huggingface/transformers/blob/master/src/transformers/models/bert/modeling_bert.py#L226
第二种，将 attn_head_size 作为参数传入的实现，包括：
TensorFlow: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/keras/layers/multi_head_attention.py#L126
TensorFlow Addons: https://github.com/tensorflow/addons/blob/master/tensorflow_addons/layers/multihead_attention.py
我个人非常推荐第二种实现方式，API 使用起来更加灵活，代码看起来也非常顺畅，更加合理。
比如，原实现中 all_head_size 的定义：

self.all_head_size = self.attn_head_size * self.num_heads

all_head_size == embed_dim，完全没有必要定义。这个变量，只在 __init__ ：

        self.qkv = nn.Linear(embed_dim,
                             self.all_head_size*3,  # weights for q, k, and v
                             weight_attr=w_attr_1,
                             bias_attr=b_attr_1 if qkv_bias else False)

和 forward ：

new_shape = z.shape[:-2] + [self.all_head_size]

中用到。__init__ 中的 qkv 映射的输出维度 self.all_head_size*3 可改为 embed_dim*3，forward中的 new_shape 用到的 self.all_head_size，可以在方法的开始，取出输入 x 的维度，修改如下:

embed_dim = x.shape[-1]
……
new_shape = z.shape[:-2] + [embed_dim]

以上是我对源码中定义 self.all_head_size 的质疑。
还有最后输出加一层 Linear Layer 的必要性：

        self.out = nn.Linear(embed_dim,
                             embed_dim,
                             weight_attr=w_attr_2,
                             bias_attr=b_attr_2)

在 forward 中，最后输出执行线性映射操作的上面由一行注释 reshape，

        z = z.reshape(new_shape)
        # reshape
        z = self.out(z)

意思应该是将维度映射回输入维度 embed_dim，方面后面的残差连接。不过既然 all_head_size == embed_dim，那何来 reshape?
所以，我认为这里对输出的线性映射是不必要的。
不过，如果我们使用第二种方式实现，将 attn_head_size 作为参数传入，不依赖 embed_size 和 num_heads 来计算，以上代码看起来就顺畅多了，合理多了。
第二种实现，将 attn_head_size 作为参数传入，只需在源代码基础上更改几行代码即可，实现如下：

from typing import Tuple, Union

import paddle
import paddle.nn as nn
from paddle import ParamAttr
from paddle import Tensor


class Attention(nn.Layer):
    """ Attention module

    Attention module for ViT, here q, k, v are assumed the same.
    The qkv mappings are stored as one single param.

    Attributes:
        num_heads: number of heads
        attn_head_size: feature dim of single head
        all_head_size: feature dim of all heads
        qkv: a nn.Linear for q, k, v mapping
        scales: 1 / sqrt(single_head_feature_dim)
        out: projection of multi-head attention
        attn_dropout: dropout for attention
        proj_dropout: final dropout before output
        softmax: softmax op for attention
    """
    def __init__(self,
                 embed_dim: int,
                 num_heads: int,
                 attn_head_size: int,
                 qkv_bias: Union[bool, ParamAttr],
                 dropout: float = 0.,
                 attention_dropout: float = 0.):
        super().__init__()
        """
        增加了一个attn_head_size的参数，attn_head_size和num_heads的大小不受embed_dim的限制，使API的使用更灵活。
        """
        self.num_heads = num_heads
        # self.attn_head_size = int(embed_dim / self.num_heads)
        self.attn_head_size = attn_head_size
        self.all_head_size = self.attn_head_size * self.num_heads  # Attention Layer's hidden_size

        w_attr_1, b_attr_1 = self._init_weights()
        self.qkv = nn.Linear(embed_dim,
                             self.all_head_size*3,  # weights for q, k, and v
                             weight_attr=w_attr_1,
                             bias_attr=b_attr_1 if qkv_bias else False)

        self.scales = self.attn_head_size ** -0.5

        w_attr_2, b_attr_2 = self._init_weights()
        # self.out = nn.Linear(embed_dim,
        #                      embed_dim,
        #                      weight_attr=w_attr_2,
        #                      bias_attr=b_attr_2)
        # 汇总多头注意力信息，并将维度映射回输入维度embed_dim，方便残差连接
        self.out = nn.Linear(self.all_head_size,
                             embed_dim,
                             weight_attr=w_attr_2,
                             bias_attr=b_attr_2)

        self.attn_dropout = nn.Dropout(attention_dropout)
        self.proj_dropout = nn.Dropout(dropout)
        self.softmax = nn.Softmax(axis=-1)

    def _init_weights(self) -> Tuple[ParamAttr, ParamAttr]:
        weight_attr = paddle.ParamAttr(initializer=nn.initializer.KaimingUniform())
        bias_attr = paddle.ParamAttr(initializer=nn.initializer.KaimingUniform())
        return weight_attr, bias_attr

    def transpose_multihead(self, x: Tensor) -> Tensor:
        new_shape = x.shape[:-1] + [self.num_heads, self.attn_head_size]
        x = x.reshape(new_shape)
        x = x.transpose([0, 2, 1, 3])
        return x

    def forward(self, x: Tensor) -> Tuple[Tensor, Tensor]:
        qkv = self.qkv(x).chunk(3, axis=-1)
        q, k, v = map(self.transpose_multihead, qkv)

        attn = paddle.matmul(q, k, transpose_y=True)
        attn = attn * self.scales
        attn = self.softmax(attn)
        attn_weights = attn
        attn = self.attn_dropout(attn)

        z = paddle.matmul(attn, v)
        z = z.transpose([0, 2, 1, 3])
        new_shape = z.shape[:-2] + [self.all_head_size]
        z = z.reshape(new_shape)
        # 汇总多头注意力信息，并将维度映射回输入维度embed_dim，方便残差连接
        z = self.out(z)
        z = self.proj_dropout(z)
        return z, attn_weights

测试：

def main():
    t = paddle.randn([4, 16, 96])     # [batch_size, num_patches, embed_dim]
    print('input shape = ', t.shape)

    model = Attention(embed_dim=96,
                      num_heads=8,
                      attn_head_size=128,
                      qkv_bias=False,
                      dropout=0.,
                      attention_dropout=0.)

    print(model)

    out, attn_weights = model(t)
    print(out.shape)
    print(attn_weights.shape)

    for name, param in model.named_parameters():
        print(f'param name: {name},\tparam shape: {param.shape} ')


if __name__ == "__main__":
    main()

输出：

input shape =  [4, 16, 96]
Attention(
  (qkv): Linear(in_features=96, out_features=3072, dtype=float32)
  (out): Linear(in_features=1024, out_features=96, dtype=float32)
  (attn_dropout): Dropout(p=0.0, axis=None, mode=upscale_in_train)
  (proj_dropout): Dropout(p=0.0, axis=None, mode=upscale_in_train)
  (softmax): Softmax(axis=-1)
)
[4, 16, 96]
[4, 8, 16, 16]
param name: qkv.weight,	param shape: [96, 3072] 
param name: out.weight,	param shape: [1024, 96] 
param name: out.bias,	param shape: [96] 

以上是我个人的一点儿不成熟的小建议，望官方评估采纳~

Describe the bug
There is one file named 'stat.py', which has the same name as the system file. The stat.py needs to be renamed or the VIT code would not run.

Describe the bug
resume training error
AttributeError: 'Momentum' object has no attribute 'set_dict'

To Reproduce
Steps to reproduce the behavior:
1.Go to 'PaddleViT/object_detection/Swin/'
2.Run 'python main_single_gpu.py -resume='./output/train-20211210-09-50-43/Swin-Epoch-45'

The recovery of model can pass

Screenshots
Traceback (most recent call last): File "C:\Program Files\JetBrains\PyCharm Community Edition 2021.2.2\plugins\python-ce\helpers\pydev\pydevd.py", line 1483, in _exec pydev_imports.execfile(file, globals, locals) # execute the script File "C:\Program Files\JetBrains\PyCharm Community Edition 2021.2.2\plugins\python-ce\helpers\pydev\_pydev_imps\_pydev_execfile.py", line 18, in execfile exec(compile(contents+"\n", file, 'exec'), glob, loc) File "F:/***/pp_swin/main_single_gpu.py", line 400, in <module> main() File "F:/***/pp_swin/main_single_gpu.py", line 313, in main optimizer.set_dict(opt_state) AttributeError: 'Momentum' object has no attribute 'set_dict'

Version (please complete the following information):

• Paddle Version: [ 2.2.0]
• Python Version [3.6]
• GPU/CPU mode [ Gpu]

Describe your feature request
Current validation (config.EVAL is True) mode still create and load training dataset and dataloader, which is not flexible when users only have val set.

So main method needs support val mode without touching training set.

Describe the reference code or paper
N/A

Describe the possible solution
I have a fix in ViT model, which can be used to other classification model.
Please refer to commit 9a7c105 for details.

我发现 PaddleViT 所有模型中的 README.md 都存在两个问题（以下均以 PaddleViT/image_classification/BEiT/ BEiT 模型的 README.md 为例）：

• 其一，Usage 示例代码中，加载预训练权重时少了后缀 .pdparams，而且注释中提到 .pdparams is NOT needed 也是不对的，应该是在下面的命令行参数中 -pretrained 的值是不需要 .pdparams，二者搞混了。

from config import get_config
from beit import build_beit as build_model
# config files in ./configs/
config = get_config('./configs/beit_base_patch16_224.yaml')
# build model
model = build_model(config)
# load pretrained weights, .pdparams is NOT needed
model_state_dict = paddle.load('./beit_base_patch16_224_ft22kto1k')
model.set_dict(model_state_dict)

应该讲注释注释中的 , .pdparams is NOT needed 删去，并在模型加载时，加上后缀 .pdparams：

from config import get_config
from beit import build_beit as build_model
# config files in ./configs/
config = get_config('./configs/beit_base_patch16_224.yaml')
# build model
model = build_model(config)
# load pretrained weights
model_state_dict = paddle.load('./beit_base_patch16_224_ft22kto1k.')
model.set_dict(model_state_dict)

• 其二，在 Evaluation 和 Training 的命令行参数值多加了一个单引号，如果在终端直接执行，会出现 FileNotFoundError 错误：

FileNotFoundError: [Errno 2] No such file or directory: "'./configs/beit_base_patch16_224.yaml'"

我之前在终端预训模型训练和验证的命令时，出现过这个错误，群里也有其他同学出现了这样的问题。出现这个错误的原因是因为 argparse 在解析命令行参数时，为字符串类型的参数值自动加上了一个双引号。所以，在为命令行参数赋值时，不需要加上引号。所以，应该去掉 Evaluation 和 Training 命令行参数值中的单引号。
单 GPU 验证：

CUDA_VISIBLE_DEVICES=0 \
python main_single_gpu.py \
    -cfg='./configs/beit_base_patch16_224.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \
    -eval \
    -pretrained='./beit_base_patch16_224_ft22kto1k'

我修改为：

CUDA_VISIBLE_DEVICES=0 \
python main_single_gpu.py \
    -cfg=./configs/beit_base_patch16_224.yaml \
    -dataset=imagenet2012 \
    -batch_size=16 \
    -data_path=/path/to/dataset/imagenet/val \
    -eval \
    -pretrained=/path/to/pretrained/model/beit_base_patch16_224_ft22kto1k  # .pdparams is NOT needed

多 GPU 验证：

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python main_multi_gpu.py \
    -cfg='./configs/beit_base_patch16_224.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \
    -eval \
    -pretrained='./beit_base_patch16_224_ft22kto1k'

我修改为：

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python main_multi_gpu.py \
    -cfg=./configs/beit_base_patch16_224.yaml \
    -dataset=imagenet2012 \
    -batch_size=16 \
    -data_path=/path/to/dataset/imagenet/val \
    -eval \
    -pretrained=/path/to/pretrained/model/beit_base_patch16_224_ft22kto1k  # .pdparams is NOT needed

单 GPU 训练：

CUDA_VISIBLE_DEVICES=0 \
python main_single_gpu.py \
  -cfg='./configs/beit_base_patch16_224.yaml' \
  -dataset='imagenet2012' \
  -batch_size=32 \
  -data_path='/dataset/imagenet' \

我修改为：

CUDA_VISIBLE_DEVICES=0 \
python main_single_gpu.py \
  -cfg=./configs/beit_base_patch16_224.yaml \
  -dataset=imagenet2012 \
  -batch_size=32 \
  -data_path=/path/to/dataset/imagenet/train \

多 GPU 训练：

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python main_multi_gpu.py \
    -cfg='./configs/beit_base_patch16_224.yaml' \
    -dataset='imagenet2012' \
    -batch_size=16 \
    -data_path='/dataset/imagenet' \ 

我修改为：

CUDA_VISIBLE_DEVICES=0,1,2,3 \
python main_multi_gpu.py \
    -cfg=./configs/beit_base_patch16_224.yaml \
    -dataset=imagenet2012 \
    -batch_size=16 \
    -data_path=/path/to/dataset/imagenet/train \ 

一会儿，我再提交个 PR，请官方审查~

Describe your feature request
Add and debug the multi scale sampler

Describe the reference code or paper
Refer to the source code listed in the original paper, from here

Describe the possible solution
Now has an initial version, needs debug and test

Additional context
N/A

Describe your feature request
Cascade Mask R-CNN

Describe the reference code or paper
Swin detection official code (from mmdet) here

Describe the possible solution
Mask R-CNN is already implemented in PaddleViT here

Additional context
Add any other context or screenshots about the feature request here.

https://github.com/BR-IDL/PaddleViT/blob/develop/image_classification/SwinTransformer/swin_transformer.py line224
The size of the tensor should be '2, window_h * window_w, window_h * window_w',not '2, window_h * window_w, window_h * window_h'

Describe the question
In the classification directory, when MIXUP prerequisites are performed in main_single_gpu / main_multi_gpu.py, the NUM_CLASSES parameter is not passed, resulting in a non-IMAGENET1K data set to load, the existing main_single_gpu / main_multi_gpu.py will result in an error——since the number of categories of the current model is not equal to 1000, the Mixup default classification number is still 1000, which will cause loss calculations that cannot be performed at this time.

Expected behavior
Introducing the class number parameters, so that the training code is easier to use.

Screenshots

Describe the bug
PaddleViT-Seg的requirements.txt文档中需要安装的依赖库有如下：

• cityscapesScripts==2.2.0
• detail==4.0
• numpy==1.20.3
• opencv-python==4.5.2.52
• scipy==1.6.3
• yacs==0.1.8

问题一：
没有detail 这个库

自己解决方式：删除detail

问题二：
opencv-python库没有4.5.2.52这个版本

自己解决方式：更换opencv-python版本

Change ViT training script according to new version:
https://github.com/BR-IDL/PaddleViT/blob/develop/image_classification/SwinTransformer/main_multi_gpu.py

代码中说Dataset related classes and methods for ViT training and validation
Cifar10, Cifar100 and ImageNet2012 are supported。
我想问下，我想训练自己的数据集需要自己重新写datase类吗？

Describe the bug
I found that the code was copied from the source author's code and some pytorch's code was transformed to paddle error.

To Reproduce
1:
VIT_custom.py line59

return input * paddle.rsqrt(paddle.mean(input ** 2, dim=2, keepdim=True) + 1e-8)

should be

return input * paddle.rsqrt(paddle.mean(input ** 2, axis=2, keepdim=True) + 1e-8)

2:
VIT_custom.py line88

class CustomAct(nn.Layer):
    """ CustomAct layer
    Custom act method set, defalut "gelu"
    """
    def __init__(self, act_layer):
        super().__init__()
        if act_layer == "gelu":
            self.act_layer = gelu
        elif act_layer == "leakyrelu":
            self.act_layer = leakyrelu
        else:
            self.act_layer = gelu

which

leakyrelu has not been defined

Describe your feature request
TransGAN: training from scratch

Describe the reference code or paper
official repo: https://github.com/VITA-Group/TransGAN

Describe the possible solution
TODO

Additional context
N/A

Describe the bug
A train function has error in main_single_gpu.py

Expected behavior
delete the excess function definition.

Screenshots

Describe the bug
For the classification model which has EMA training,
error occurs when start resuming training.

The EMA model name and loading is not correct.

To Reproduce
Steps to reproduce the behavior:

1. Start model training with -resume enabled and EMA = True
2. In the main_single_gpu.py/main_multi_gpu.py, model_ema name and loading raises error

Expected behavior
Model EMA should be loaded without error

Additional context

1. Model name can be solved by using default names for EMA models, and add exist check
2. Model loading can be solved by used model_ema.modual.set_state_dict.

do you have any model for this paper "line Segment detection using transformer without edges"

交叉熵loss应该计算预测分布概率与真实分布之间的距离，按道理是应该在softmax计算之后再进行Loss计算的，但是训练代码中都是直接计算loss loss=criterion(output, label)，后面才再计算softmax。请问不应该是先计算softmax再计算Loss，还是其实这两者的顺序对训练没有区别