Code Structure

Square_Pooling (relating to Section 5)
	o	Main.py  For training Baseline and Foveated models by changing the flag “flag_fullres” in patchconvnet_models.py
	o	Analysis  To extract all the results from the checkpoint file
			Step_1_throughput.py  To get image throughput
			Step_2_TrainStats_main  Collecting training statistics for Dynamic-Stop
			Step_3_DS_main  Running the Dynamic-Stop using the training statistics
			Step_5_RG_main  To find a moderately difficult data subset
			Step_7_AdvRobustness_main  Computing adversarial robustness
			Comparing fixation guidance mechanisms
			•	Step_9_RG_main  Guided by self-attention or Random
			•	Step_9a_RG_main  Guided by Itti-Koch or GBVS
			•	Step_9b_RG_main  Guided by Deep Gaze II

Radial_Polar_Pooling (relating to Section 6)
	o	Scene_Classification  For calibrating the pooling regions
			FullRes 
			•	Step_4_main  Finetune the Baseline model on this dataset
			•	Step_4_main_Evaluate  Run the model on images seen by humans
			Foveated_0.84
			•	Implement_log_polar_PR_a  Convert pooling region’s mat files into numpy files
			•	Step_3a_main  Finetune the Foveated model on this dataset
			•	Step_3a_main_Evaluate  Run the model on images seen by humans
			Step_6b_SC_V2_Single_Fixation_analysis  Compute the agreement between model decisions and human decisions.
	o	ImageNet  Evaluating the selected configuration on ImageNet
			Step_3a_main  Finetune the Foveated model on ImageNet (using radial-polar)
			Step_7_AdvRobustness_main  Compute adversarial robustness
			Step_8_TrainStats_main  Compute training statistics
			Step_8_DS_main  Running Dynamic-Stop using the training statistics

