# DNN feature maps info ## Data directory structure AlexNet, ResNet-50, CORnet-S and MoCo PCA downsampled feature maps. Each `.zip` file is organized in the following directory structure: /alexnet │ └───pretrained-False │ │ │ └───layers-all │ │ └───pca_feature_maps_ilsvrc2012_test.npy │ │ └───pca_feature_maps_ilsvrc2012_val.npy │ │ └───pca_feature_maps_test.npy │ │ └───pca_feature_maps_training.npy │ │ │ └───layers-single │ └───pca_feature_maps_ilsvrc2012_test.npy │ └───pca_feature_maps_ilsvrc2012_test.npy │ └───pca_feature_maps_test.npy │ └───pca_feature_maps_training.npy │ └───pretrained-True │ └───layers-all │ └───pca_feature_maps_ilsvrc2012_test.npy │ └───pca_feature_maps_ilsvrc2012_val.npy │ └───pca_feature_maps_test.npy │ └───pca_feature_maps_training.npy │ └───layers-single └───pca_feature_maps_ilsvrc2012_test.npy └───pca_feature_maps_ilsvrc2012_val.npy └───pca_feature_maps_test.npy └───pca_feature_maps_training.npy * The `pretrained-False` folders contain feature maps of untrained (randomly initialized) DNNs. * The `pretrained-True` folders contain feature maps of trained DNNs. * The `layers-all` folders contain feature maps of combined DNN layers. * The `layers-single` folders contain feature maps of individual DNN layers. ## Data files All data files are Python dictionaries with DNN layer names as keys and the corresponding feature maps activations as values. Following is a description of each file's array dimensions: * `pca_feature_maps_ilsvrc2012_test.npy`: [ILSVRC-2012 test images][ilsvrc2012] feature maps in a 2-dimensional array of shape [100,000 ILSVRC-2012 test image conditions × 1000 components]. * `pca_feature_maps_ilsvrc2012_val.npy`: [ILSVRC-2012 validation images][ilsvrc2012] feature maps in a 2-dimensional array of shape [50,000 ILSVRC-2012 validation image conditions × 1000 components]. * `pca_feature_maps_test.npy`: [THINGS test images][things_part] feature maps in a 2-dimensional array of shape [200 test image conditions × 3000 components]. * `pca_feature_maps_training.npy`: [THINGS training images][things_part] feature maps in a 2-dimensional array of shape [16,540 training image conditions × 3000 components]. ## Additional info You can donwnload the feature maps from [Figshare+][figshare]. For a detailed description of the feature maps extraction and PCA downsampling please refer to our [paper][paper_link] and [code][code_link]. [figshare]: https://plus.figshare.com/articles/dataset/Deep_neural_network_feature_maps/21514590 [paper_link]: https://doi.org/10.1016/j.neuroimage.2022.119754 [code_link]: https://github.com/gifale95/eeg_encoding/tree/main/02_dnn_feature_maps_extraction [things_part]: https://osf.io/y63gw/ [ilsvrc2012]: https://www.image-net.org/download.php