DCN_AlexNet - a deep convolutional network for image classification

Dean Pospisil and Wyeth Bair


This Topic Page describes how to run the DCN_AlexNet model from a linux (or Mac OS X) command line using the WM model command. You will first need to download and compile the WM software (follow the iModel Download link).

This implementation of AlexNet provides access to a pre-trained model that can be tested with visual stimuli. Responses from all of the internal units can be studied. Training has not been implemented here, but is available from other sources, for example the Caffe deep learning framework.

Measuring responses to a test image. The following command will run the DCN_AlexNet model on a test image of a kitten:

wm mod DCN_AlexNet.moo image.stm all_center.rsp
where the following parameter and data files are required: The first three files listed above are the main model (.moo), stimulus (.stm) and response (.rsp) files. The last two files are auxiliary data files. The file 'anw.txt' is a large file containing all the weights and biases for the model, and it is referenced from the model file (DCN_AlexNet.moo). The file 'image_kitten_fst.txt' contains the RGB values for an image of a kitten, and it is refrenced from 'image.stm'. These two auxiliary data files are referenced with a path of "./data/" (relative to the working directory from which the wm command is to be executed).

The output is written to a file named "zz.nd," which contains response values for each of the units at the spatially centered position in the network. The values can be extracted to a text file with the following command:

  nda resp.nda zz.nd out.txt
where 'resp.nda' is a text script that instructs the 'nda' command and references the list of unit names given in 'chanlist_center_all.txt'. This command will write the output text file 'out.txt', which will contain all of the unit names followed by the response value associated with that unit.