Mind

[joshgev]

These modifications add a supplier and associated sources to kur in order to support the Stanford mind reading data set.

[ajsyp]

Looks awesome! Can you add documentation for this before we merge?

[noajshu]

We can add documentation in another pull request. Here is an example Kurfile:

---

settings:

  # Backend to use
  backend:
    name: keras
    backend: tensorflow
    parallel: 4

  # Data provider settings
  provider: &provider
    batch_size: 128

  # Hyperparameters
  cnn:
    kernels: [32, 64, 128]
    size: [[5, 5], [2, 2], [2, 2]]
    strides: [[1, 1], [2, 2], [2, 2]]

# The model itself.
# This is parsed immediately after the "parameters" block.
model:
  - input: brain_probe_image
  - expand: -1
  - for:
      range: "{{ cnn.kernels|length }}"
      iterate:
        - convolution:
            kernels: "{{ cnn.kernels[index] }}"
            size: "{{ cnn.size[index] }}"
            strides: "{{ cnn.strides[index] }}"
        - activation: relu
  - flatten:
  - dense: 6
  - activation: softmax
    name: category_label

# The training section
train:
  data:
    - mind:
        path: train

  provider:
    <<: *provider
  log: mindreading_category.log
  epochs: 50
  weights:
    initial: mindreading_category.best.valid.w
    best: mindreading_category.best.train.w
    last: mindreading_category.last.w

  optimizer:
    name: adam
    learning_rate: 0.0001

The validation section
validate:
  data:
    - mind:
        path: validate

  provider:
    <<: *provider
  weights: mindreading_category.best.valid.w

# The loss function selection
loss:
  # The target is the label you want your model to guess and get correct
  - target: category_label
    name: categorical_crossentropy

...