13. Working Group Meeting about nips-2018-adversarial-vision-challenge

[Simsso]

General Items up for Discussion

• Decision on C++ implementation vs. virtual batch size method; see profiling issue #58
• Embedding vector sparsity issue and potentials solutions #51 (comment)
• Merging of working code
  • vq-layer branch release of vq_layer v1.0.0
  • vq-resnet branch alongside with submittable-resnet
• Planning / naming of requirements for the enablement tools
• Creation of an issue for structured experiments with the current VQ layer (for instance, we still don't know what an appropriate magnitude is for the learning rate, alpha, beta, and gamma)

[Simsso]

Minutes

• Pipeline works
• Logging to a file inside of the Docker container (both errors and info logs)
• avc-test-model due today
• Enablement tools are work in progress
• Replacement feature is implemented and tested
• Replacement features does not work work with virtual batches and replaces with validation samples. Needs to be improved #70.
• Submission documented and "implemented"; not tested yet (due today)
• Profiling with equations to compute memory requirements
• Logger class implemented
• Virtual batch size feature implemented
• Cosine distance is not completed, due until next time

[Simsso]

Due Today

• Logger class mini thing
• Merge accumulator #67
• PR for submittable ResNet
• Local testing avc-test-model
• Merge v0.1.0 of vq_layer #69
• Rename and open PR for resnet_base #72
• Submit to crowd.ai
• Telegram commands / enablement tools interface
• Plan structured experiments for the VQResNet
• Set up appointment for the next working group meeting

[FlorianPfisterer]

In terms of the structured VQResNet experiments: note we should add tf.flags for the parameters here as well. It's not very nice to hard-code them in the graph definition (referring to here).

[Simsso]

Hot-Fixes

• Memory requirements e7ac4ef
• Replacement feature is broken f50dd65
• Batching was broken 72f9d99