Renamed folder

Author: kma-code

generalized_latent_equilibrium/experiments/CIFAR10/Fig6/BP/acc_lr_seeds_epochs.npy

@@ -0,0 +1,161 @@
+import argparse
+import numpy as np
+import json
+import subprocess
+from pathlib import Path
+import os
+import pickle
+import logging
+
+if __name__ == '__main__':
+
+	logging.basicConfig(format='Generating param files -- %(levelname)s: %(message)s', level=logging.INFO, force=True)
+
+	parser = argparse.ArgumentParser(description='Train a Latent Equilibrium Neural Network on CIFAR10 task.')
+	parser.add_argument('--run', action='store_true',
+													default=False,
+													help='Run parameter sweep.')
+	parser.add_argument('--gather', action='store_true',
+													default=False,
+													help='Gather results.')
+	parser.add_argument('--algorithm', required=True,
+													help='Choose algorithm: BP, FA, DFA or PAL')
+
+
+	args = parser.parse_args()
+	algo = args.algorithm
+	logging.info(f'Algorithm: {algo}')
+
+	# path to parent generalized_latent_equilbrium folder
+	PATH_parent = Path(__file__).parent.resolve().parents[2]
+	# path to folder of this file
+	PATH_runner = Path(__file__).parent.resolve() / str(algo)
+
+
+
+	OUTPUT_DIR = PATH_runner / "runs"
+
+	# create a bunch of JSON param files in subfolders (lr/seeds/)
+	seeds = [1,3,5,7,9,11,13,15,17,19]
+
+	params_arr = []
+	for lr in [None]:
+		params_per_lr = []
+		for seed in seeds:
+			if algo == 'BP':
+				params = {
+					"algorithm": algo,
+					"epochs": 50,
+					"batch_size": 128,
+					"batch_learning_multiplier": 64,
+					"lr_factors": [5.0, 0.0, 1.0, 0.0, 2.0, 0.2],
+					#"lr_factors":  [lr1, 0, lr2, 0, lr3, lr4],
+					#"lr_factors": [0.0, 0.0, 0.0, 0.0, 0.0, 0.0],
+					"wn_sigma": [0,0,0,0,0,0],
+					"n_updates": 100,
+					"target_type": "mse",
+					"seed": seed,
+					"activation": "Sigmoid",
+					"model_variant": "vanilla",
+					"with_optimizer": "store_true"
+						}
+			elif algo == 'FA':
+				params = {
+					"algorithm": algo,
+					"epochs": 50,
+					"batch_size": 128,
+					"batch_learning_multiplier": 64,
+					"lr_factors": [10.0, 0.0, 25.0, 0.0, 2.0, 0.2],
+					"wn_sigma": [0,0,0,0,0,0],
+					"n_updates": 100,
+					"target_type": "mse",
+					"seed": seed,
+					"activation": "Sigmoid",
+					"model_variant": "vanilla",
+					"with_optimizer": "store_true",
+					"rec_degs": True,	
+						}
+			elif algo == 'DFA':
+				params = {
+					"algorithm": algo,
+					"epochs": 50,
+					"batch_size": 128,
+					"batch_learning_multiplier": 64,
+					"lr_factors": [1.0, 0.0, 1.0, 0.0, 5.0, 0.2],
+					"wn_sigma": [0,0,0,0,0,0],
+					"n_updates": 100,
+					"target_type": "mse",
+					"seed": seed,
+					"activation": "Sigmoid",
+					"model_variant": "vanilla",
+					"with_optimizer": "store_true",
+					"rec_degs": True,	
+						}
+			elif algo == 'PAL':
+				params = {
+					"algorithm": algo,
+					"epochs": 50,
+					"batch_size": 128,
+					"batch_learning_multiplier": 64,
+					"lr_factors": [5.0, 0.0, 1.0, 0.0, 2.0, 0.2],
+					"wn_sigma": [0,0,0,0,0,0],
+					"n_updates": 100,
+					"target_type": "mse",
+					"seed": seed,
+					"activation": "Sigmoid",
+					"model_variant": "vanilla",
+					"with_optimizer": "store_true",
+					"rec_degs": True,				
+					"bw_lr_factors": [0,0,0,0,10,10],
+					"regularizer": [0,0,0,0,5e-5,5e-5],
+					"tau_xi": [10,10,10,10,10,10],
+					"tau_HP": [10,10,10,10,10,10],
+					"tau_LO": [0,0,0,0,0,0],
+					"sigma": [0,5e-2,0,5e-2,5e-2,5e-2]
+						}
+			params_per_lr.append(params)
+	params_arr.append(params_per_lr)
+
+
+
+	if args.run and not args.gather:
+		for i, params_per_lr in zip(range(len(params_arr)), params_arr):
+			for j, params in zip(range(len(params_per_lr)), params_per_lr):
+
+				PATH_output = OUTPUT_DIR / str('lr' + str(i) +	'/seed' + str(j))
+				params["output"] = str(PATH_output) + '/'
+
+				# create output directory if it doesn't exist
+				if not(os.path.exists(params['output'])):
+					# logging.info(f"{PATH_runner + '/' + params['output'] } doesn't exists, creating")
+					os.makedirs(params['output'] )
+
+				with open(str(params['output']) + '/params.json', 'w') as f:
+					logging.info(f"Saving to {f.name}")
+					json.dump(params, f)
+
+				sim_dir = PATH_output
+				# start runs as separate processes
+				proc_name = ['python', 'experiments/CIFAR10/le_layers_cifar10_training.py', '--params', str(sim_dir / 'params.json')]
+
+				logging.info(f"Starting run as subprocess {proc_name}.")
+				subprocess.Popen(proc_name, cwd=PATH_parent)
+
+	elif args.gather:
+		# Collect results
+		lin_acc_arr = []
+		for i, params_per_lr in zip(range(len(params_arr)), params_arr):
+				lin_acc_per_lr = []
+				for j, params in zip(range(len(params_per_lr)), params_per_lr):
+
+					sim_dir = OUTPUT_DIR / str('lr' + str(i) +		'/seed' + str(j))
+					with open(str(sim_dir) + '/val_acc.pkl', 'rb') as in_file:
+							val_acc = pickle.load(in_file)
+					lin_acc_per_lr.append(val_acc)
+				lin_acc_arr.append(lin_acc_per_lr)
+
+
+		lin_acc_output_file = PATH_runner / "acc_lr_seeds_epochs.npy"
+		np.save(lin_acc_output_file, lin_acc_arr)
+		logging.info(f"Gathered data and saved to {lin_acc_output_file}.")
+