Nan
how about c++ code
how about modify ``` for (int y = 0; y < h; ++y) { for (int x = 0; x < w; ++x) { output[y * h + x] =...
> 我也发现这个问题，把1024维拉伸后，只用前两维做分类，但是我尝试做个FC效果反而不好。但是这么做很多卷积都是浪费的，不太了解原因 我的感觉是做分类的话的确浪费了 但是其实多的计算量几乎可以忽略不计； 作者保留1024的feature ， 这样懒得改代码就可以直接 用作其他用途的训练： 可以训练多分类， 可以当做提取的特征用作人脸识别等。 作者在论文中说加了FC网络容易过拟合。 可以试试加些dropout
cloned code: /data/yyrg_projects/TensorRT-LLM/examples/llama/convert_checkpoint.py from pypi: /var/lib/g02u1/.local/lib/python3.10/site-packages/tensorrt_llm/models/llama/convert.py They are from different version of tensorrt-llm ## try `cp -r /data/yyrg_projects/TensorRT-LLM/tensorrt_llm/* /var/lib/g02u1/.local/lib/python3.10/site-packages/tensorrt_llm/`
To have DLPack support for C++ API would be great, like ``` DLManagedTensor* fromNvcvTensor(const nvcv::Tensor& src) ; ... ``` An example is from pytorch: https://github.com/Tabrizian/pytorch/blob/5c5f53ef6645016794b2e0c9a267ccb095995ea5/aten/src/ATen/DLConvertor.cpp#L244
> > At this moment, CudaBuffer can be converted to torch.Tensor through `__cuda_array_interface__`. However, sometimes we have to use libtorch(at::Tensor/torch::Tensor). Is it possible to make a libtorch interface for nvcv::Tensor/ITensor...
> Can you provide more details? > > is the ask here to have a crop that takes in a list of rects, a single input image, and an output...
Hi @shiremathNV , Thanks for your reply. > If so, we have an operators for NMS, please take a look at that That would be nice though, but I cannot...
Yes, some examples need to be moved elsewhere, such as ppocr and YOLOX, as their models cannot be generated online. Meanwhile, the corresponding links in the [online documentation](https://torchpipe.github.io/docs/showcase) also need...
> Format Candidates: > [joblib](https://github.com/joblib/joblib), recommended by scikit-learn. https://scikit-learn.org/stable/model_persistence.html MyPipeline().save("myapp.someformat") That's an interesting feature. Is there any detail for how to implement this feature?