This is a public repository of pretrained ELL models.
The ELL models repository contains a number of pretrained models suited for different device footprints. These models were generated by parameterizing a few neural network models that are known to work well, and generating a variety of related models based on these starting points. The model files themselves observe the following naming convention:
The filename will start with a short prefix identifying the general model architecture used to train the network. The
prefixes are v
, d
, and a
.
Following the prefix will the an underscore (_
), followed by an encoding of the input dimension,
followed by encodings of the layers present in the model. This encoding uses a single upper-case
letter to represent a layerThe layer types and their encodings are:
Layer | Abbrev |
---|---|
Input | I |
Convolution | C |
Binary convolution | B |
Variable-bit conv | V |
Max pooling | M |
Average pooling | A |
Fully-connected | F |
Softmax | S |
The parameters of the various layers will generally be determined by the architecture, but where necessary, parameters may be added immediately following the layer abbreviation. For instance, a convolutional layer in the 'd' model type uses a 3x3 window for its convolutional layers, except for the last, which uses a 1x1 spatial window. The identifier for a full 'd' model would look like:
d_I224x224x3CMCMCMCMCMCMC1AS
Since many of the models we use alter the structure of the base model they're inspired by, we need to encode the parameters that vary as well. An example of a 'v' model with a single fully-connected layer, and of size 3072, would be:
v_I112x112x3BBMCCMCCCMCCCMCBBMF3072
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.