Skip to content

Latest commit

 

History

History
50 lines (37 loc) · 1.97 KB

README_formats.md

File metadata and controls

50 lines (37 loc) · 1.97 KB
Raw

Tensor formats

In this short note, we indicate the expected format (dimensions) of tensors in our torch-based framework. Most of the code is compatible with numpy.

Dimension Abbreviations

For cleaner documentation, we denote each dimension by a specific letter. Here is the conversion table:

  • S --> Stimulus-Response pairs
  • B --> Batches
  • R --> Repeats
  • T --> Time-bins
  • F --> Frequency-bins
  • C --> Channels
  • N --> Neurons

Finally, * means any number of dimensions. You can find these abbreviations all over the code contained within this repository.

Tensors

Our framework expects a specific dimensionality for each nature of tensor. To keep the high generality, please respect these conventions when building on the project.

  • stimuli (spectrogram form): (B, C=1, F, T)
  • predicted responses: (B, R=1, T, N)
  • observed responses (groundtruth): (B, N, R, T)

Dataset classes and pytorch dataloaders are supposed to serve the data in the above formats; similarly, tensor reshapes should be made within the forward() method of model classes. This allows minimal tensor manipulation in fitting and evaluation loops, making the framework general to most datasets, models and performance metrics. It also contributes to a cleaner and much more modular code, in which you only need to change datasets and models construction.

Example

Here is an illustration of a proper usage of the conventions set by our framework.

from models import MyModelClass
model = MyModelClass(args)

B, R, T, F, C, N = 4, 10, 999, 64, 1, 1
spectrogram = torch.rand(B, C, F, T)    # a batch of 4, 1-channel spectrograms of 49 frequency bands and 999 time-bins
responses = torch.rand(B, N, R, T)      # a batch of 4 responses of 999 time-bins for 1 neuron, each with 4 repeats

prediction = model(responses)           # shape: (B, 1, T, 1)

from metrics import correlation_coefficient
cc = correlation_coefficient(prediction, responses)  # (B, ) or (0,)