"Failed building wheel for onnx" Error while install_prerequisites.bat about open_model_zoo

[snosov1]

The root cause seems to be the missing Protobuf (see the line Protobuf compiler not found). Not sure where it should be fixed.

@vmatveic Slava, do you think it's an issue of the installer or the environment?

[Bahramudin]

@snosov1 Thanks for the reply! But when I want to install the Protobuf, gives me this:

pip3 install Protobuf
Requirement already satisfied: Protobuf in c:\users\bahra\appdata\roaming\python\python37\site-packages (3.6.1)
Requirement already satisfied: six>=1.9 in c:\users\bahra\appdata\roaming\python\python37\site-packages (from Protobuf) (1.12.0)
Requirement already satisfied: setuptools in c:\users\bahra\appdata\local\programs\python\python37\lib\site-packages (from Protobuf) (40.8.0)

[alalek]

pip3 install Protobuf

pip3 install onnx is your target command.

protobuf python is not the same as protobuf C++ which is in mentioned error log.

Building wheels

Building of Python packages from sources usually doesn't work well on Windows platform.

Try to downgrade Python or downgrade onnx version - to ensure that it is loaded by pip via wheels (prebuilt wheels may not be available for "latest" configurations).

---

This issue can't be fixed by patching code in this repository.

[Bahramudin]

@alalek Yes, the problem was with Python version. I have one another question, but I don't know to ask it here or open a new issue, OK I will prefer to ask it here.

That is: In the latest version the pre-trained models (intel_models) are not bundled into the download package, so now the intel_models there is only the description of the model.

So I found a script which is used to download the models.
But here I found some problem which some of the models cannot be downloaded, the connection timed out or the connection cannot be established, the host not responding and so on problems, so it makes hard to download the models, the problem may be due to some restriction to some addresses it not possible to have access to some of the links.

It is the result of my download, after many tries

###############|| Post processing ||###############

========= Changing input dimensions in vgg19.prototxt =========
========= Changing input dimensions in vgg16.prototxt =========
========= Changing input dimensions in googlenet-v1.prototxt =========
========= Changing input dimensions in googlenet-v2.prototxt =========
========= Moving to new Caffe layer presentation googlenet-v2.prototxt =========
========= Changing input dimensions in alexnet.prototxt =========
FAILED:
Sphereface
googlenet-v4
inception-resnet-v2
mobilenet-ssd
mobilenet-v2-1.4-224
mtcnn-o
mtcnn-p
mtcnn-r
resnet-101
resnet-152
resnet-50
se-inception
se-resnet-101
se-resnet-152
se-resnet-50
se-resnext-101
se-resnext-50
squeezenet1.0
squeezenet1.1
ssd300
ssd512

One error example:

Error Connecting: HTTPSConnectionPool(host='docs.google.com', port=443): Max retries exceeded with url: /uc?export=download&id=0B_geeR2lTMegb2F6dmlmOXhWaVk (Caused by NewConnectionError('<urllib3.connection.VerifiedHTTPSConnection object at 0x00000202B81BA160>: Failed to establish a new connection:[WinError 10060]

Most occurred error hosts are:

HTTPSConnectionPool(host='docs.google.com', port=443)
HTTPSConnectionPool(host='onedrive.live.com', port=443)

So is it not possible to bundle all the models into a single package and download at once?
I think the better way is to provide three download ways:

• Bundle all them into a single package

• Bundle each model in a separate package and let the user download it manually

• And download by the script

So now everyone will be able to download whatever models they want.

---

One small question:
What does the value exactly mean in after of the each model's names, e.g:
instance-segmentation-security-0033
instance-segmentation-security-0049
AND
face-detection-adas-0001
face-detection-retail-0004

Thanks in advance!!

Update:
One another problem with download script is: when running it for the second time, then it will download the already downloaded models again, so it will be good to check for the model existing first, if already exist, skip and go for the next one. So if some models fail to download and when want to run the second time you cannot run it again, it is a big waste of time and of bandwidth for both sides Server and the Client.

[alalek]

numbers

It is like a version (of used model's scheme, training processes, including used datasets).
May be faster and/or more accurate.
There is no "gold" solution, so models are always evolving. Some interesting variants goes public.

face-detection-adas-0001
face-detection-retail-0004

Perhaps they are trained on different datasets. And/or based on different scheme of models. Or have different purposes. Refer to attached description/readme files of these models.

Evaluate all applicable variants on your data and pick the best based on these metrics:

• accuracy
• performance

[snosov1]

On the value interpretation - @alalek is correct. It's simply an ID, so we can distinguish between different models.

On the download issue - @atsyplen can you comment? In the meantime, @Bahramudin can you try model_downloader from this repo?

[atsyplen]

@Bahramudin regarding HTTPS errors, first of all please check that the correct date and time settings are set. Secondly make sure that you use a correct https proxy settings, if you need it for your network.
Regarding requested features - you can find them in the latest version of model downloader from this repo.

[Bahramudin]

@snosov1 I am just using the one you mentioned to download the models. I think it is the reason which some of the links is not accessible may be restricted, anyway, I have tried many ways such as using VPN, but also was not working well, but VPN is not very stable after a while it is interrupted.

And also this script has a problem which when run it for the second time, then it will start downloading the already downloaded models, it will be good to check if the file already exists.

So I think it will be good to provide three ways of download which I mentioned in the above comment.

About the version, I checked the lower the heavier bin file, so the heavier may be more accurate and also works slow.

[snosov1]

@Bahramudin please, reopen if it's still relevant