ImagePredictor fails if any row contains NaN as its image.

ImagePredictor currently crashes if a NaN is present in a row for the image (row contains no image)

Ideally, it should either use a dummy image or predict_proba the mean of NaN rows in train during test, or otherwise predict_proba the overall mean of the training data if no NaN rows were in train. NaN rows in train can be dropped.

               'mxnet.base.MXNetError: Traceback (most recent call last):\n'
               '  File "src/io/image_io.cc", line 223\n'
               "Imread: Check failed: file.is_open(): '' couldn't open file: "
               'No such file or directory\n',

[sxjscience]

@gradientsky Thanks for the suggestion! How do you usually determine the most frequent image? I can think of two algorithms:

• 1. Measure the checksum of all images and randomly pick one from those that have the highest counts.
• 1. Extract embeddings from these images via ResNet50 and build the cluster with K-means. After that, we can pick the image that are closest to the center of the largest cluster.

[gradientsky]

Another option for ensembles: get oof predictions from non-image models, then fill with random image from the same class.