[2020-05-04T03:39:44.167Z] -- Prepare external packages for TVM...

[2020-05-04T03:39:44.167Z]   % Total    % Received % Xferd  Average Speed   Time    Time     Time  Current

[2020-05-04T03:39:44.167Z]                                  Dload  Upload   Total   Spent    Left  Speed

[2020-05-04T03:39:44.167Z] 
  0     0    0     0    0     0      0      0 --:--:-- --:--:-- --:--:--     0

Example scripts contains some dependencies not listed for Horovod, and in some cases require datasets without explaining how to obtain them. We should provide a README file along with a set of packages (requirements.txt) for successfully running the examples.

我发现examples/retinaface.cpp中，如果开启OMP加速的话似乎在检测到人脸时会发生内存泄漏，但我定位不了这个问题的具体原因。

值得注意的时，如果将qsort_descent_inplace函数中的OMP指令注释掉这个问题就会消失掉。

static void qsort_descent_inplace(std::vector<FaceObject>& faceobjects, int left, int right)
{
    int i = left;
    int j = right;
    float p = faceobjects[(left + right) / 2].prob;
    ...
    // #pragma omp parallel sections
    {
        // #pragma

Several parts of the op sec like the main op description, attributes, input and output descriptions become part of the binary that consumes ONNX e.g. onnxruntime causing an increase in its size due to strings that take no part in the execution of the model or its verification.

Setting __ONNX_NO_DOC_STRINGS doesn't really help here since (1) it's not used in the SetDoc(string) overload (s

i had following the step to do face detection using retina face and got rectangle bounding boxes output as like on example, could someone please guide me to do realtime recognition like shown in video? i'm currently pretty new at face recognition task, thank you.

Platform (like ubuntu 16.04/win10): Windows 10
Python version: 3.7.4, mmdnn==0.2.5

Running scripts: mmconvert -f caffe -df keras -om test

I know that this command is not supposed to run without passing an input file, but the error message is incorrect and should be improved:

mmconvert: error: argument --srcFramework/-f: invalid choice: 'None' (choose from 'caffe', 'caffe2', 'cn

Hi, thanks for the great code!

I wonder do you have plans to support resuming from checkpoints for classification? As we all know, in terms of training ImageNet, the training process is really long and it can be interrupted somehow, but I haven't notice any code related to "resume" in scripts/classification/train_imagenet.py.

Maybe @hetong007 ? Thanks in advance.

does simpledet support for negative examples without any foreground instances? and there is no
relative description in the dataset.md.

Feature motivation

Azure storage client is backward incompatible, there's an issue #757 for upgrading the azure connection to the latest client. The new version has an async interface that can be leveraged for the streams module, at least for downloads.

N.B. This is likely a 3~4h work and not a pressing need, so just putting this in the backlog.

Feature implementation

There's a

I tried building the docs, but was met with a graphviz error. Typically this means I can spend a few hours pecking away at the dependencies until I get stable build... or someone that has it working can export their environment, and publish an environment.yml that we can use with the build instructions.
I was going off of the d2l book since that's a dep here, but their [environment.yml](https://g

I happened to read an article about openmp yesterday. https://zhuanlan.zhihu.com/p/118604153

I followed the suggestions in the article and set OMP_WAIT_POLICY=PASSIVE for both workers and servers. and the throughput surged around 25%. the profile shows this is because a huge reduction in time taken by PULL operation. so i suppose the flag is beneficial to servers. i guess it eases the content

I think we can remove the check of mutable_args in the get_model API for BERT, GPT, Language model and transformer. We can introduce a separate flag "allow_override" if users want to override any configuration. Otherwise overriding any configuration is forbidden.

I've been using rllib for a while and quite like the Trainable and Tune (hyperaram search) abstractions.
Does coach contain something similar?

I really like coach's documentation. Can someone provide a comparison between the two and when it is applicable to use each?
Thanks.

In the xlearning job web UI, the metrics graph can not be seen for these is no js for it.

This was reported to me by @max-andr. Most of the differences are actually explicitly mentioned in comments in our implementation, but we should check again if we can match the reference implementation more closely and possible mention deviations in the docs, not just in comments.

@max-andr might create a PR to fix this

Changes made in #1874 to test/models/mxnet/1rnn_layer_3lstm.json reveal an error in the fusion of a sequence of LSTM cells into a single layer RNN for gpu.

The fused result should look like:

Instead, we currently see:
![rnn_fusion_issue](https://user-images.githubusercontent.com/48

1. 31页，第一行
   
   应改成
   

2. 35 页， 正定矩阵的定义 （公式 2-15）
   标准正定矩阵的定义，应该没有 “对称矩阵” 这一条件吧？ 只是您在本书大多讨论的是“对称”正定矩阵，所以加上了 “对称” 这一条件？

3. 38页， 奇异值分解定义（公式 2-16）
   ![default](https://

What did you find confusing? Please describe.
I was looking for a way to get information about the inference endpoint from the inference endpoint code at runtime (eg. within a Flask request handler). Similar to how training and processing have access to JSON config in a file under the /opt/ml/ structure or an env var.

Describe how documentation can be improved
Need docs for the righ

The documentation in DJL was originally written with the expectation that users are reasonably familiar with deep learning. So, it does not go out of the way to define and explain some of the key concepts. To help users who are newer to deep learning, we created a [documentation convention](https://github.com/awslabs/djl/blob/master/docs/development/development_guideline.md#documentation-conventio