Code related to our paper:

Jinsi* O, Henderson* MM, Tarr MJ (2023) Early experience with low-pass filtered images facilitates visual category learning in a neural network model. PLoS ONE 18(1): e0280145. https://doi.org/10.1371/journal.pone.0280145

To download our trained models, see: https://kilthub.cmu.edu/articles/dataset/Early_experience_with_low-pass_filtered_images_facilitates_visual_category_learning_in_a_neural_network_model/23972115

Contents of this repository

Data folders (all will be empty initially):

• csvFiles: Contains important data to run models. Should be empty initially, need to run createDictEcoset.py and createDictImgNet.py to fill folder with necessary files.
• trials: Contains training logs and best models for each trial.
• labelAccuracy: Contains data on accuracy between basic- and subordinate-level categories. To get data, you'll need to run getLabelAccuracyBW.py and getLabelAccuracyColored.py AFTER finishing all trials. If you decided to run more/less than 10 trials, you'll need to alter the script to accomodate for such changes.

Code files:

General

• bOrS.csv: ImageNet categories labeled (basic or subordinate)
• createDictEcoset.py: important data preprocessing for Ecoset data
• createDictImgNet.py: important data preprocessing for Imagenet data
• getLabelAccuracyBW.py: get accuracy of basic vs subordinate categories using grayscale models from experiment 2
• getLabelAccuracyColored.py: get accuracy of basic vs subordinate categories using colored models from experiment 2

Training for Experiment 1

• BWLinearBlurModel.py: blurs linearly over 50 epochs, trains on grayscale Ecoset images
• BWNoBlurModel.py: control model (no blur applied), trains on grayscale Ecoset images
• BWNonLinearBlurModel2.py: blurs logarithmically (with a base of 2) over 50 epochs, trains on grayscale Ecoset data
• ColoredLinearBlurModel.py: blurs linearly over 50 epochs, trains on colored Ecoset images
• ColoredNoBlurModel.py: control model (no blur applied), trains on colored Ecoset images
• ColoredNonLinearBlurModel2.py: blurs logarithmically (with a base of 2) over 50 epochs, trains on colored Ecoset data

Training for Experiment 2

• BWImgNetScratch.py: control model (no pretrained model from experiment 1), trains on grayscale ImageNet images
• BWLinearBlurModelImgNet.py: uses pretrained models from BWLinearBlurModel.py, trains on grayscale ImageNet images
• BWNoBlurModelImgNet.py: uses pretrained models from BWNoBlurModel, trains on grayscale ImageNet images
• BWNonLinearBlurModel2ImgNet.py: uses pretrained models from BWNonLinearBlurModel2.py, trains on grayscale ImageNet images
• ColoredImgNetScratch.py: control model (no pretrained model from experiment 1), trains on colored ImageNet images
• ColoredLinearBlurModelImgNet.py: uses pretrained models from ColoredLinearBlurModel.py, trains on colored ImageNet images
• ColoredNoBlurModelImgNet.py: uses pretrained models from ColoredNoBlurModel, trains on colored ImageNet images
• ColoredNonLinearBlurModel2ImgNet.py: uses pretrained models from ColoredNonLinearBlurModel2.py, trains on colored ImageNet images

Statistics and plotting code (in subfolder "analysis")

• graphAccuracyEcoset.ipynb: run stats and make plots for models from Experiment 1 (makes our Figure 2)
• graphAccuracyImgNet.ipynb: run stats and make plots for models from Experiment 2 (makes our Figure 3)
• graphAccuracyImgNet_BasicSub.ipynb: plot basic vs. subordinate accuracy (makes our Figure 4)

Steps to run the code

1. Clone this repository
   • git clone https://github/com/ojinsi/startingblurry
2. Download the Ecoset and ImageNet datasets
   • Ecoset: https://codeocean.com/capsule/9570390/tree/v1
   • ImageNet: https://www.image-net.org/download.php
3. Prepare the ImageNet database
   • On initial download, the validation images will all be in a single folder rather than separated into sub-folders by category. We'll need them in subfolders to continue. To do this, use the script "valprep.sh" provided in this repo (this should be run from the "val" folder in your ImageNet directory).
4. Preprocess the data
   • Change the path names at the top of createDictEcoset.py and createDictImgNet.py to the folders where you've downloaded Ecoset and ImageNet, respectively.
   • Run createDictEcoset.py and createDictImgNet.py. These scripts will do important data preprocessing, such as assigning ecoset/imgnet labels to int labels and choosing a random subset of images for each epoch.
5. Start running your models!
   • Use the code provided to run models of each type (details of each code file described above)
   • Note that all model training will require PyTorch and access to at least one GPU (4 GPUs recommended).
   • By default, each script has "trialNumber" set to 1, which will run the first trial only. Make sure to change the trialNumber variable in your scripts whenever you run a new model, else you'll rewrite your previous trial run!
   • Remember, you can only start running your Experiment 2 models after you finish running your Experiment 1 models.
   • The results for Experiment 1 should be saved under BW/Colored and the results for Experiment 2 should be saved under BWImgNet/ColoredImgNet. In experiment 1, there should be a total of 60 models (2 color conditions * 3 blur conditions * 10 trials) while in experiment 2, there should be a total of 80 models (2 color conditions * 4 pretrained model conditions * 10 trials).
6. Alternatively, download our pre-trained model files:
   • https://kilthub.cmu.edu/articles/dataset/Early_experience_with_low-pass_filtered_images_facilitates_visual_category_learning_in_a_neural_network_model/23972115
7. Analyze results
   • To examine the difference between accuracy between basic- and subordinate-level categories in the models trained in Experiment 2, you can run getLabelAccuracyBW.py and getLabelAccuracyColored.py. Results can be found under the folder "labelAccuracy".
   • Use the provided jupyter notebooks (in folder "analysis") to perform statistics and visualize results.