OpenVINO support

[lutzroeder]

2018_R3_age-gender-recognition-retail-0013.zip
2018_R3_rm_lstm4f.zip
2018_R4_landmarks-regression-retail-0009-fp16.zip
2018_R4_single-image-super-resolution-0034.xml.zip
2018_R5_vehicle-license-plate-detection-barrier-0106.xml.zip
face-detection-retail-0005.zip
int4-model.zip
text-recognition-0012.zip
text-spotting-0002-recognizer-decoder.zip
ti_with_large_body.xml.zip
with_gather.xml.zip
with_nd_conv_simple.xml.zip
with_pad_simple.xml.zip
openvino_i4_matmul_model.zip

• 1. Opening rm_lstm4f.xml results in TypeError ([OpenVINO] Added pre-flight content check and fixed dot files opened #192)
• 2. dot files are not opened any more - need to fix it ([OpenVINO] Added pre-flight content check and fixed dot files opened #192)
• 3. add preflight check for invalid xml and dot content
• 6. Add test files to ./test/models.json ([OpenVINO] Added dot models to testing #195) ([OpenVINO] Adding 2018 R5 OpenVINO Model Zoo models to testing #211)
• 9. Add support for the version 3 of IR ([OpenVINO] Add v3 IR support #196)
• 10. Category color support ([OpenVINO] Support per-layer documentation  #203)
• 11. -metadata.json for coloring, documentation and attribute default filtering ([OpenVINO] Support per-layer documentation  #203).
• 5. Filter attribute defaults based on -metadata.json to show fewer attributes in the graph
• 7. Show weight tensors
• 8. Graph inputs and outputs should be exposed as Graph.inputs and Graph.outputs
• 12. Move to DOMParser
• 13. Remove dot support

[demid5111]

These are files for testing OpenVINO dot files:

i8_normalized_squeezenet_ssd.dot.zip
i8_normalized_squeezenet11.dot.zip
i8_normalized_vgg16.dot.zip

[lutzroeder]

@lutzroeder I need your advice. I am now switching to the item !7. How would you recommend to do it taking into consideration that structure of the model (only structure) is in .xml file and weights/biases (only binary data) is in .bin. I would be glad to hear your suggestions in order to not loose time with inappropriate solution from the design perspective.

Have a look at src/caffe2.js which uses a similar approach loading a binary tensor file called init_net.pb. Call context.request to load the file from the same directory as the model file. Make sure to recover if the file failed to load. For decoding, src/mxnet.js might be a good example which has an ndarray decoder for a separate parametersEntry.data byte array. For test, models.json supports comma separated lists for target and source and picks the first entry as the model file.

[lutzroeder]

@lutzroeder and also the question about coloring (!10). What can I use as a reference? Is it mandatory to define the -metadata.json file?

It is not "mandatory", it could be hardcoded in code. However, in practice you will quickly notice the benefits of having the metadata in a separate file. The format is very similar between all the frameworks and should work for yours as well. For most frameworks the color categories are manually added to -metadata.json but other information like documentation gets updated via scripts.

[lutzroeder]

@demid5111 any updates on these tasks? In particular will 12. and 13. be done by end of Jan? What's the plan for 5.?

[demid5111]

@lutzroeder can you tell more about the point 5? If it is not very complex, I will do it before points 12 and 13.
I have not yet started working on the re-writing parser. Hopefully, I will start on 29th of January, I will let you know.

[lutzroeder]

Idea for 5. is to show attributes (Attribute::visible = true) if the value is different from the default value. If the user selects Cmd+D such attributes will show when the graph is rendered. The list should only include few attributes relevant to the user. 12. and 13. should probably be done first. Currently attributes seem to not get parsed correctly (for example, 11 instead of 1,1) so not sure default filtering will work until this is fixed in the parser.

[demid5111]

@lutzroeder I need to say that your latest commit introduces degradation for the Netron support for the OpenVINO IR v4. I have checked it locally - TensorIterator layer can not be parsed correctly and the whole network can not be rendered at all. Lets revert the commit and I will build upon your version and propose as a pull request till the end of the week. How does that sound for you? Thank you for starting the work with transition to the DOM parser.

[lutzroeder]

@demid5111 Can you get a PR ready adding the TensorIterator support. It is not used by any test files so adding a test file would be great as well. Along the same lines, is there documentation for the .bin format available and are you working on item 7)?

[demid5111]

@lutzroeder no problem, I will do that. is it possible not to make the new version/release of the product without TensorIterator being fixed?

[demid5111]

@lutzroeder yes, I am working on the item 7. however, temporarily switching to the TensorIterator issue

[demid5111]

@lutzroeder the fix is on the way. I have made a half of work - it is in my fork's master to the moment. I will make a PR in a day or two with the final version. Thank you for waiting for the fix and not upgrading the release version!

[demid5111]

@lutzroeder those are files for testing the fix for the TensorIterator:

[lutzroeder]

@demid5111 any status updates on 7). See also #307.