Support Poolformer

• [x] build poolformer model
• [x] convert pretrained weight
• [x] inference test on imagenet and update model_zoo
• [x] update docs
• [x] update changelog
• [x] pytorch speed comparison oneflow版本过慢，待解决

1. flowvision.models下面有_utils.py和utils.py。
2. IntermediateLayerGetter方法在flowvision.models._utils.py和flowvision.models.segmentation.seg_utils.py重复。

所以删除flowvision.models._utils.py，并暂时引用flowvision.models.segmentation.seg_utils.py。

when I want to use the model of vit_tiny_patch16_224 from flowvison module ,it prompt this EOFError: Ran out of input. 环境就是OneFlow实训平台的3090显卡：oneflow-0.7.0+torch-1.8.1-cu11.2-cudnn8

Support Uniformer

• [x] build uniformer model
• [x] convert pretrained weight
• [x] inference test on imagenet and update model_zoo small_plus
• [x] update docs
• [x] update changelog
• [x] pytorch speed comparison

Add LeViT

• [x] build model
• [x] update init.py in models
• [x] convert pretrained weight
• [x] inference test on imagenet and update model_zoo
• [x] update docs
• [x] update readme
• [x] update changelog
• [x] pytorch speed comparison

使用 models 中的模型时，例如 model = vgg11(pretrained=True) ，成功下载 zip 权重文件后，解压过程出错，导致解压中断、参数文件不完整。如果自行将下载的 zip 解压，就能正常使用。多个模型都有同样的问题。

Traceback (most recent call last):
  File "temp.py", line 77, in <module>
    model = vgg11(pretrained=True)
  File "/usr/local/miniconda3/lib/python3.7/site-packages/flowvision/models/vgg.py", line 182, in vgg11
    return _vgg("vgg11", "A", False, pretrained, progress, **kwargs)
  File "/usr/local/miniconda3/lib/python3.7/site-packages/flowvision/models/vgg.py", line 156, in _vgg
    state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)
  File "/usr/local/miniconda3/lib/python3.7/site-packages/flowvision/models/utils.py", line 146, in load_state_dict_from_url
    return _legacy_zip_load(cached_file, model_dir, map_location, delete_file)
  File "/usr/local/miniconda3/lib/python3.7/site-packages/flowvision/models/utils.py", line 78, in _legacy_zip_load
    f.extractall(model_dir)
  File "/usr/local/miniconda3/lib/python3.7/zipfile.py", line 1636, in extractall
    self._extract_member(zipinfo, path, pwd)
  File "/usr/local/miniconda3/lib/python3.7/zipfile.py", line 1691, in _extract_member
    shutil.copyfileobj(source, target)
  File "/usr/local/miniconda3/lib/python3.7/shutil.py", line 79, in copyfileobj
    buf = fsrc.read(length)
  File "/usr/local/miniconda3/lib/python3.7/zipfile.py", line 930, in read
    data = self._read1(n)
  File "/usr/local/miniconda3/lib/python3.7/zipfile.py", line 1006, in _read1
    data = self._decompressor.decompress(data, n)
zlib.error: Error -2 while decompressing data: inconsistent stream state

Hello, I need method for create model iresnet. I saw in documentation that flowvision has model iresnet, but when I import and use resnest50 = flowvision.models.face_recognition.iresnest50(pretrained=False, progress=True), python says module 'flowvision.models' has no attribute 'face_recognition'. What can be problem?

Add RegionViT

• [x] build model (F.unfold 算子不支持 https://github.com/Oneflow-Inc/oneflow/issues/3785)
• [x] update init.py in models
• [x] convert pretrained weight
• [x] inference test on imagenet and update model_zoo
• [x] update docs
• [x] update changelog
• [x] pytorch speed comparison

脚本运行方式：

cd ci/check
bash run_speed_test.sh

结果会输出到 当前目录下的 result 文件里

目前通过测速脚本发现的问题

import torch as flow 运行会崩

• vit
• conv_mixer
• crossformer
• cswin
• mlp_mixer
• pvt
• res_mlp
• vgg

本身运行也会报错，输入是 224x224 的时候

• efficientnet
• res2net

TODO

Model relative

• [x] freeze_bn
• [ ] unfreeze_bn
• [x] ActivationHook
• [ ] freeze_unfreeze_fn

Others

• [x] random seed

Test

• [x] test freeze_bn
• [ ] test activation_hook

oneflow/nn/init.py

from oneflow.python.ops.math_ops import fused_scale_tril from oneflow.python.ops.math_ops import fused_scale_tril_softmax_dropout from oneflow.python.ops.math_ops import relu from oneflow.python.ops.math_ops import tril

应该是 as ReLU? 难道我的oneflow版本装错了。。 flowvision-0.1.0 oneflow==0.7.0+cu102

import oneflow as flow
import torch
import numpy as np

a = np.random.randn(3,3).astype(np.float32)

b = 2

torch_a = torch.from_numpy(a)
flow_a = flow.from_numpy(a)

print(torch.div(torch_a,b,rounding_mode='floor'))
print(flow.div(flow_a,b).floor())
print(flow.div(flow_a,b,rounding_mode='floor'))

ResNet-50 训练

参照当前 vision 下的 project 复现 resnet-50 训练和精度对齐。

参考

• pytorch：resent
• timm：resnet

主要目标

• [ ] 2022.05.11 - 2022.5.12：熟悉 vision 下的分类模型训练代码，数据集配置并跑通。
• [ ] 2022.05.12 - 2022.05.20：对照 timm 和 pytorch 复现 reset-50 训练代码，对齐相关训练条件，测试并使用多卡训练。
• [ ] 2022.05.21 - 2022.05.27：对比精度差异调整并复现精度，最终将训练好的权重替换为 oneflow 版本。

项目负责人：林松 预计完成时间：2022.05.27

相关 PR

罗列对应的 PR，以为一个 issue 可能会对应多个 PR，所以这里提供的是表格

| PR | 作者 | reviewer | 日期 | | | ------------------------------------------------------------ | ---- | -------- | -------- | ---- | | 首次上传提交代码 | 林松 | zzzzzzz | 20220510 | |

目前Vision下已经有的一个可以参考的projects，迁移了Swin-T的训练代码，用于Vision下进行模型的训练，但是vision中绝大部分模型的精度复现还无法保证，所以这里开启一个完善训练的projects: 用于复现vision下实现的模型的精度，并且在后续逐渐将vision下迁移的权重替换为oneflow自身训练的权重，这里是暂时的规划，需要2-3位实习生参与完成:

可参考的projects:

• https://github.com/rwightman/pytorch-image-models
• https://github.com/microsoft/Swin-Transformer

训练的任务，以及首批需要复现精度的模型:

• 完善Vision下的这个projects: - https://github.com/Oneflow-Inc/vision/tree/main/projects/classification, 熟悉这个projects的用法（与Swin-T基本一致）
• 这里我们列举一下第一阶段在vision下需要复现精度的模型以及相关paper:

| Model | Paper | 认领人 | PR | |:----:|:----:|:----:|:----:| | ResNet50 | ResNet strikes back: An improved training procedure in timm | 林松 | | DeiT | Training data-efficient image transformers & distillation through attention | | | Swin-Transformer | Swin Transformer: Hierarchical Vision Transformer using Shifted Windows | 林德铝 | | DeiT III | DeiT III: Revenge of ViT | | |

• 需要的硬件条件：8卡V100机器，能跑得下单卡256batchsize即可