rhodricusack/cornet_analysis

CORnet analysis

Examine usefulness of representations for classification, through development and across layers of network. Will be used to expand Anna Birbeck's project relating AoA and layer in network

• Examination of development of CORnet-S representations
• Loads a checkpoint that is the result of training
• Freezes weights in convolutional network
• Adds a logistic regression to each layer

Requirements

imagenet data, with train and val_in_folders directories GPU machine launch.json contains arguments for running of eval_linear.py with VSCODE checkpoints (e.g., from s3://neurana-imaging/rhodricusack/cornet/)

Description of procedure

(1) Ran CORnet and get models from multiple epochs of training (2) Ran eval_linear.sh (or eval_linear_2019-08-31.sh) to

• for each epoch and conv, load model and add logistic regression layer
• train up
• validate
• validate on each aoa category separately (given --aoaval) (3) Ran summarize_performance.py to summarize, make graphs and do stats

Outputs

from CORnet are in $HOME/cornet directory

from eval_linear

linearclass: was fine, but didn't record reglog layers linearclass_v3: main run. Initially, epochs 0,5,10,20,30 analysed. But will add 1,2,3,4,15,25,35 using eval_linear_2019-08-31.sh

Based on code from DeepCluster from Facebook Research https://github.com/facebookresearch/deepcluster and CORnet from Dicarlo lab at MIT https://github.com/dicarlolab/CORnet