search

mlpack / models

models built with mlpack
https://models.mlpack.org/docs
BSD 3-Clause "New" or "Revised" License
33 stars 42 forks source link

Addition of MobileNet V1 #71

Closed Aakash-kaushik closed 2 years ago

Aakash-kaushik commented 3 years ago

Creating a issue to track the progress of Mobilenet and discuss things before we have a PR for this.

Created: https://github.com/mlpack/mlpack/pull/3007 to add depthwise seperable convolutions.
And we still don't have a final reference from which we are going to build mobilenet from but that's secondary for now, but i will keep looking for it.

Tasks.

  1. [ ] Get https://github.com/mlpack/mlpack/pull/3007 merged.
  2. [x] Using tensorflow reference to build MobileNet.
  3. [x] Start implemetation and create a PR for the same: #72
  4. [x] Get ReLU6 https://github.com/mlpack/mlpack/pull/3009 merged.

References:

  1. Mobilenet V1 paper

cc: @zoq, @kartikdutt18

Aakash-kaushik commented 3 years ago

Hi, So the models present on the repos we talked about yesterday doesn't have the param alpha and depthMultiplier used so we will only be able to provide weights weights for a single config with a single image size accepted, as the image size given to mobilenet matters. and the tool i talked about mmdnn that can only take static image sizes and the things are very rigid and hardcoded just because of that fact that any model weight is first converted to an IR and then to the framework. So at this point i want to move forward with mobilenet V2 if that's okay?

kartikdutt18 commented 3 years ago

I don't mind if we only provide weights only for one model config.

Aakash-kaushik commented 3 years ago

I don't mind if we only provide weights only for one model config.

Should be doable then.

Aakash-kaushik commented 3 years ago

Mobilenet has all these configs available: https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md ignore the quantized ones, but on;y if we somehow manage to import weights from TensorFlow. For now, I will start with the implementation of mobilenet and care less about how we are going to port the weights as that hinders a lot when I try to code the actual architecture.

Aakash-kaushik commented 3 years ago

I have also opened #72 to add mobilenet_v1, and also left a comment about a layer that we might need to implement. let me know what you guys think about that.

mlpack-bot[bot] commented 2 years ago

This issue has been automatically marked as stale because it has not had any recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions! :+1: