cheng peng

Where is ASTN?

3

It's an amazing work, I have read your paper and it impresses me a lot. But when I read your adversarial-frcnn/models/pascal_voc/VGG16/fast_rcnn_adv/train.prototxt , I can only find ASDN, and where is...

- [ ] 只用作比较, 不合并该分支

libai trainer设计文档

5

调研了一下detectron2, mmdet, ColossalAI, paddledetection ### paddledetection paddledetection则是直接定义了一个object, 比较冗长, 主要有trainer.train()函数去协调各个模块, 整体下来感官和ColoraIAI的并没有本质区别 如果要加新功能, 则在这个object里面进行改动, 改动会对旧版本造成较大的影响 [trainer.py](https://github.com/PaddlePaddle/PaddleDetection/blob/release/2.3/ppdet/engine/trainer.py) ### detectron2 && mmdet 这两者的设计思路其实差不多, 先定义一个HookBase ```python class HookBase: def before_train(self): """ Called before the first...

Thanks for sharing your code, It's an amazing work! But I'm a little confused. Suppose My testing images are already cropped , and there is only one person that occupies...

Dev export onnx

3

这个pr要做的: - [x] 支持MT5转onnx的脚本, (目前需要在分支https://github.com/Oneflow-Inc/oneflow_convert/tree/fix_t5_export_onnx_bug下进行) - [x] 支持MT5 onnx 推理的脚本. 直接运行`python libai/onnx_export/onnx_inference/t5_onnx_infer.py`即可, 但是由于onnx的输入和输出都是numpy, 所以目前generate的功能没办法从libai中迁移过来, 目前只能以`model.py`中的的model.forward()的输入转换成numpy的格式进行推理

Rwkv v4 merge main

2

rwkv 代码修正版: 跑loss对齐时的需要手动修改的代码: - 注释`libai/engine/default.py `的 729~730行: ```python # Global scheduler cfg # cfg.train.scheduler.warmup_iter = cfg.train.warmup_iter # cfg.train.scheduler.max_iter = cfg.train.train_iter ``` - 把`libai/data/build.py `下所有的 shuffle改为false, 并注释掉所有的`persistent_workers=True if num_workers > 0...

Stack in graph building

1

graph building stack here ```shell I20221201 10:32:37.597100 2644176 straighten_nodes.cpp:493] Straightening order: I20221201 10:32:37.597177 2644176 straighten_nodes.cpp:494] 3 I20221201 10:32:37.597187 2644176 straighten_nodes.cpp:494] 0 I20221201 10:32:37.816900 2644176 time_util.h:97] [count log]{"loc":"[GraphCompile]GraphBase_0 BuildTaskGraph","mem_rss":"9391.000000 MB","mem_vm":"35491.000000 MB","time_cost":"2...

flow device is different from pytorch device

3

oneflow tensor.device() in oneflow is different from torch, extra parameter : ` index=0` ```shell >>> import oneflow as flow >>> flow.zeros(2).device device(type='cpu', index=0) >>> >>> import torch >>> torch.zeros(2).device device(type='cpu')...