Running into an error trying to use the exported model. I have the latest version of tensorflow and online forums suggest that the .pb file needs to be updated when this error is encountered
NodeDef mentions attr 'dilations' not in Op<name=Conv2D; signature=

I have tried the Prediction API with image (Content-Type: application/octet-stream) and URL (Content-Type: application/json) and these work great! I am looking for an example where I can use a base64 encoded string representation of the image with the prediction api.

I have tried the Prediction API with image (Content-Type: application/octet-stream) and URL (Content-Type: application/json) and these work great! I am looking for an example where I can use a base64 encoded string representation of the image with the prediction api.

Hello,

Could you please provide detailed documentation regarding the integration of an exported model to CoreML. Currently, the model registers an output of a multi-dimensional array. Can the output be more defined so that it is more clear when integrated into iOS.

I see that we can change the type (negative or non-negative) using the UI interface but not with the API.

The Dockerfile, when exporting a Linux Container, has the command:

CMD python app.py

this is buffering the output of the app.py command, so you do not see the logged items as needed.

If you change this to:

CMD python -u app.py

It will not buffer the output.

Ability to give each image a name via the UI. Display that name when hovering over the image in the UI rather that displaying "open image detail". I have a lot of similar looking images and I want to name them by the date they were taken but I don't want the date to be a tag.

You also need a category of UI on your suggestions box.

Provide feature to changes subscription mapped with project.

If we moved azure custom vision resources in azure portal then it exiting project not visible in portal.

This is blocker need to fix. Project seems not working if subscription limit reached. Need to provide feature to change subscription mapping ASAP.

Thanks, Great Api. Microsoft.
Good Job.

Provide feature to download existing training images

Provide clone project feature.
Some time developer want to clone project and map with other Azure subscription.
To create multiple environment like dev, test, stage, prod

Allow a tag to have a description ie a paragraph that describes what the tag is. This is different to the name, which is the only way currently to record info about the tag. The CustomVision.ai service should allow maintenance of the description and the API should pull back the description along with the name and score when accessing the predictions API.

There should be a variable or header that can be sent with the prediction API call to specify that the image should NOT be saved past the initial classification/object detection.

My use case: Want to make a game incorporating object detection, and while I love the ability to see the predictions and modify them to improve my dataset I really don't want to see people's nude photos of them messing around or anything personally identifiable.

Aside from the result of the scan, I don't want anything to do with those images.

For certain applications (Tensorflow in Unity3D), we need to know the input_name and output_name

I found these for the models currently generated by this service to be:
input_name: Placeholder
output_name: model_outputs

I think it is useful for many devs to include this, and other information such as mean and image size when eporting to tensorflow graph.

When trying to click "export", it says I should select "compact" model, but there is no help or hint as of what that means, nor where I can do that.
Pity, because the rest of the UX is very easy.

CustomVision service has the ability to upload an image and assign a label to it (either via UI or API). It would be very convenient if there was an ability to annotate an uploaded photo and mark areas with labels so one image can have multiple labels (either via UI or API). Something like VIA or LABELIMG doing that.

And use metadata and annotations to improve predictions accuracy.

I often find that an object is detected under multiple different tags, each with a slightly different bounding box.

I'd like a feature where if there are two object detections, each with a similar bounding box location and size, only the highest confidence one is used.

This will ensure that an object will be "one or the other", rather than "both", similar to the way multiclass works in classification projects.

please I need the ability to count objects within a picture

Could not predict model has LEFT/RIGHT side, such as left arrow / right arrow.......

The 17763 release of Windows and WinML api's aren't compatible with current ONNX exports. Please update to avoid this error:

Type Error: Type (map(string,tensor(float))) of output arg (loss) of node (ZipMap) does not match expected type (seq(map(string,tensor(float))))