The L2 EyeRobot1 group is sharing preprocessed data [DOWNLOAD HERE]These features are learned and extracted using our deep Convolutional neural network that we described in our earlier reports. They represent the activations of the last fully connect layer of our network for each image of the three datasets.
Please note that this layer has a dropout subroutine with a probability of 0.3, which means that 30% of those features are zeros, which is ok. The dataset is organized as follow: Training data : newdata.csv - 400 features and the real class of the image separated by commas. - each feature is a floating number founded to %0.4f. Training data : newval.csv - 400 features and a dummy class separated by commas. - each feature is a floating number founded to %0.4f. - the dummy class is set to zero. Training data : newtest.csv - 400 features and a dummy class separated by commas. - each feature is a floating number founded to %0.4f. - the dummy class is set to zero. |
Details
AuthorIsabelle Guyon. Chaired professor of "big data", Paris-Saclay University. President of ChaLearn.org. Archives
April 2018
Categories |