FAQ

What is the default image representation in nuts-ml?

The standard formats for image data in nuts-ml are Numpy arrays of shape (h,w,3) for RGB images, (h,w) for gray-scale images and (h,w,4) for RGBA image.

What image formats can nuts-ml read?

The ReadImage nut can read images in the following formats GIF, PNG, JPG, BMP, TIF, NPY, where NPY are plain Numpy arrays.

How to flatten a batch of predictions?

Assuming the output of a network prediction is a batch of labels, how can it be flattened into a flow of labels?

>>> from nutsflow import Collect, Flatten
>>> batched_labels = [(0, 1, 0), (1, 1, 0)]   # e.g. from network.predict()
>>> batched_labels >> Flatten() >> Collect()
[0, 1, 0, 1, 1, 0]

What if the batch has multiple columns, e.g. labels and probabilities

>>> from nutsflow import Collect, FlattenCol
>>> batched_preds = [((0,1,0), (0.1,0.2,0.3)), ((1,1,0), (0.4,0.5,0.6))]
>>> batched_preds >> FlattenCol((0,1)) >> Collect()
[(0, 0.1), (1, 0.2), (0, 0.3), (1, 0.4), (1, 0.5), (0, 0.6)]

Error: Only length-1 arrays can be converted to Python scalars

If you see the following error message when running network.evaluate() under Keras you need to upgrade to Keras 2.x and the latests nuts-ml version.

in compute_metric
return float(result.eval() if hasattr(result, 'eval') else result)
TypeError: only length-1 arrays can be converted to Python scalars

ImportError: No module named Tkinter

This means the computer nuts-ml is running on is not supporting the default graphical backend for matplotlib. In this case create a file ~/.config/matplotlib/matplotlibrc with the following content:

backend : Agg

Alternatively add the following lines to your code:

import matplotlib
matplotlib.use('Agg')

How to use class weights for imbalanced classes in Keras

class_weight = {0:1, 1:50}

for epoch in xrange(EPOCHS):
  t_loss = samples >> build_batch >> network.train(class_weight=class_weight) >> Mean()

If the samples are an iterable (and not an iterator that is consumed) the class weights can also be computed directly. For instance, assuming that the class labels are in the second column (index = 1) of the sample, the following code can be used

class_weight = samples >> Get(1) >> CountValues()