def main():
# ===============
# 1. config
# ===============
print('Loading config')
config = get_config()
print(config.model_file_name)
# ===============
# 2. data
# ===============
print('Reading data from %s' % config.in_path)
data = Data(config=config)
# ===============
# 3. experiment
# ===============
exp = Experiment(data=data, config=config)
print('Start training')
exp.run(check_period=config.check_period,
early_stop=config.early_stop,
patience=config.patience)
# ===============
# 4. test
# ===============
if 'wn18' in config.in_path: # Sanity test on wn18
exp.show_link_prediction(h='06845599',
t='03754979',
r='_member_of_domain_usage',
raw=True)
if 'fb15k' in config.in_path: # Sanity test on fb15k
exp.show_link_prediction(
h='/m/08966',
t='/m/05lf_',
r=
'/travel/travel_destination/climate./travel/travel_destination_monthly_climate/month',
raw=True)
if 'KG30C' in config.in_path: # Sanity test
exp.show_link_prediction(h='7F74B998',
t='019EC1A3',
r='paper_in_domain',
raw=True)
if 'KG94C' in config.in_path: # Sanity test
exp.show_link_prediction(h='7E52972F',
t='80D75AD7',
r='author_write_paper',
raw=True)
def test_cls_init(self):
env = get_env()
exp = Experiment(
# random_seed=0,
epochs=1,
model_cls='models.transformers.JointBERT',
model_params={
'bert_model_path': env['bert_dir'] + '/bert-base-cased',
'labels_count': 3,
},
loss_func_cls='torch.nn.BCELoss', # loss,
model_output_to_loss_input=lambda ys: ys.double(),
data_helper_cls='wiki.data_helpers.JointBERTWikiDataHelper',
data_helper_params={
'wiki_relations_path': '../wiki/relations.csv',
'wiki_articles_path': '../wiki/docs.pickle',
'labels': ['employer', 'country_of_citizenship'],
# 'employer' # 'capital' # 'country_of_citizenship' #'educated_at' # 'opposite_of'
'label_col': 'relation_name',
'negative_sampling_ratio': 1.,
'train_test_split': 0.7,
'max_seq_length': 512,
'train_batch_size': 4,
'test_batch_size': 4,
'bert_model_path':
'/Volumes/data/repo/data/bert/bert-base-cased',
# 'bert_tokenizer_cls': '',
'bert_tokenizer_params': {
'do_lower_case': False,
},
'df_limit': 3,
},
tqdm_cls='tqdm.tqdm',
output_dir='../output',
)
assert isinstance(exp.model, JointBERT)
assert isinstance(exp.data_helper, JointBERTWikiDataHelper)
assert isinstance(exp.loss_func, BCELoss)
assert tqdm == exp.tqdm_cls
print(flatten(exp.to_dict()))
exp.run()
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
xorexp = Experiment(params_pso,
net_layers,
path="2in_xor.txt",
debugMode=False,
sampleMode=True)
xorexp.run()
##Single iterations of base experiments commented below
# net_single = {
# "layer1": {
# "input_count":1,
# "node_count":1,
# "activations": []
# }
# }
# cubicexp = Experiment(params_pso, net_single, path="1in_cubic.txt", debugMode=False, sampleMode=True)
# cubicexp.run()
# linearexp = Experiment(params_pso, net_single, path="1in_linear.txt", debugMode=False, sampleMode=True)
# linearexp.run()
def run_beta():
params_pso = {
"swarmsize": 40,
"alpha": 1,
"beta": 0,
"gamma": 4.1,
"delta": 0,
"jumpsize": 1,
"act_bound": 5,
"weight_bound": 10,
"bound_strat": 1,
"num_informants": 3,
"vel_range": 1,
"max_runs": 1000,
"informants_strat": 2
}
# net_layers = {
# "layer1": {
# "input_count":1,
# "node_count":1,
# "activations": []
# }
# }
# net_layers = {
# "layer1": {
# "input_count":2,
# "node_count":2,
# "activations": []
# },
# "layer2": {
# "input_count":2,
# "node_count": 1,
# "activations:":[]
# }
# }
net_layers = {
"layer1": {
"input_count":2,
"node_count":2,
"activations": []
},
"layer2": {
"input_count":2,
"node_count":2 ,
"activations:":[]
},
"layer3": {
"input_count":2,
"node_count":1 ,
"activations:":[]
}
}
best_gamma = 0
best_beta = 0
best_error = None
for j in range(0, 10):
run_beta = 0
run_gamma = 4.1
run_best = None
#first do 4.1 and 0
params_pso["beta"] = 0
params_pso["gamma"] = 4.1
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
#first do 0.5 and 3.6
params_pso["beta"] = 0.5
params_pso["gamma"] = 3.6
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
#first do 1 and 3.1
params_pso["beta"] = 1.0
params_pso["gamma"] = 3.1
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
#first do 1 and 3.1
params_pso["beta"] = 1.5
params_pso["gamma"] = 2.6
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
#first do 1 and 3.1
params_pso["beta"] = 2.05
params_pso["gamma"] = 2.05
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
#first do 1 and 3.1
params_pso["beta"] = 2.6
params_pso["gamma"] = 1.5
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
#first do 1 and 3.1
params_pso["beta"] = 3.1
params_pso["gamma"] = 1.0
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
#first do 1 and 3.1
params_pso["beta"] = 0.5
params_pso["gamma"] = 3.6
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
#first do 1 and 3.1
params_pso["beta"] = 0.0
params_pso["gamma"] = 4.1
experiment1 = Experiment(params_pso, net_layers, path="2in_complex.txt", debugMode=False, sampleMode=True)
experiment1.run()
if (run_best == None or experiment1.pso.best.fitness < run_best):
run_best = experiment1.pso.best.fitness
run_beta = params_pso["beta"]
run_gamma = params_pso["gamma"]
print("\nRun ", j, " Beta: ", run_beta, " Gamma: ", run_gamma, " Error", run_best)
print("\nOverall Beta: ", best_beta, " Gamma: ", best_gamma, " Error", best_error)
from agents import RandomSearchAgent
from experiments import Experiment
import gym
seed = 16
env = gym.make('LunarLander-v2')
env.seed(seed)
ragent = RandomSearchAgent(name='RandomSearchAgent-1',
state_dim=env.observation_space.shape[0],
action_dim=env.action_space.n,
seed=seed,
stop_search_reward=210)
exp = Experiment(env, ragent, logdir="../log", verbose=True, num_episodes=1000)
exp.run()
def run_swarmsize():
print("\Swarmsize Cubic")
print("=======================")
params_pso = {
"swarmsize": 0,
"alpha": 1,
"beta": 2.05,
"gamma": 2.05,
"delta": 0,
"jumpsize": 1,
"act_bound": 5,
"weight_bound": 10,
"bound_strat": 1,
"num_informants": 3,
"vel_range": 1,
"max_runs": 1000,
"informants_strat": 2
}
net_single = {
"layer1": {
"input_count": 1,
"node_count": 1,
"activations": []
}
}
cubic_optimal_size = 0
cubic_best = None
for j in range(0, 10):
params_pso["swarmsize"] = 0
print("\nRun ", j)
cubic_optimal_size = 0
cubic_best = None
for i in range(0, 10):
params_pso["swarmsize"] += 10
print(params_pso["swarmsize"])
experiment1 = Experiment(params_pso,
net_single,
path="1in_cubic.txt",
debugMode=False,
sampleMode=True)
experiment1.run()
if (cubic_best == None
or experiment1.pso.best.fitness < cubic_best):
cubic_best = experiment1.pso.best.fitness
cubic_optimal_size = params_pso["swarmsize"]
print("\nRun ", j, "best size", cubic_optimal_size, " produced",
cubic_best)
print("Cubic optimal size ", cubic_optimal_size, " produced", cubic_best)
print("\Swarmsize Linear")
print("=======================")
linear_optimal_size = 0
linear_best = None
for j in range(0, 1):
params_pso["swarmsize"] = 0
print("\nRun ", j)
linear_optimal_size = 0
linear_best = None
for i in range(0, 10):
params_pso["swarmsize"] += 10
experiment1 = Experiment(params_pso,
net_single,
path="1in_linear.txt",
debugMode=False,
sampleMode=True)
experiment1.run()
if (linear_best == None
or experiment1.pso.best.fitness < linear_best):
linear_best = experiment1.pso.best.fitness
linear_optimal_size = params_pso["swarmsize"]
print("\nRun ", j, "best size", linear_optimal_size, " produced",
linear_best)
print("linear optimal size ", linear_optimal_size, " produced",
linear_best)
print("\Swarmsize Sine")
print("=======================")
sine_optimal_size = 0
sine_best = None
for j in range(0, 1):
params_pso["swarmsize"] = 0
print("\nRun ", j)
sine_optimal_size = 0
sine_best = None
for i in range(0, 10):
params_pso["swarmsize"] += 10
experiment1 = Experiment(params_pso,
net_single,
path="1in_sine.txt",
debugMode=False,
sampleMode=True)
experiment1.run()
if (sine_best == None or experiment1.pso.best.fitness < sine_best):
sine_best = experiment1.pso.best.fitness
sine_optimal_size = params_pso["swarmsize"]
print("\nRun ", j, "best size", sine_optimal_size, " produced",
sine_best)
print("sine optimal size ", sine_optimal_size, " produced", sine_best)
net_layers = {
"layer1": {
"input_count": 1,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
print("\Swarmsize Tanh")
print("=======================")
tanh_optimal_size = 0
tanh_best = None
for j in range(0, 1):
params_pso["swarmsize"] = 0
print("\nRun ", j)
tanh_optimal_size = 0
tanh_best = None
for i in range(0, 10):
params_pso["swarmsize"] += 10
experiment1 = Experiment(params_pso,
net_layers,
path="1in_tanh.txt",
debugMode=False,
sampleMode=True)
experiment1.run()
if (tanh_best == None or experiment1.pso.best.fitness < tanh_best):
tanh_best = experiment1.pso.best.fitness
tanh_optimal_size = params_pso["swarmsize"]
print("\nRun ", j, "best size", tanh_optimal_size, " produced",
tanh_best)
print("tanh optimal size ", tanh_optimal_size, " produced", tanh_best)
print("\Swarmsize XOR")
print("=======================")
xor_optimal_size = 0
xor_best = None
for j in range(0, 1):
params_pso["swarmsize"] = 0
print("\nRun ", j)
xor_optimal_size = 0
xor_best = None
for i in range(0, 10):
params_pso["swarmsize"] += 10
experiment1 = Experiment(params_pso,
net_layers,
path="2in_xor.txt",
debugMode=False,
sampleMode=True)
experiment1.run()
if (xor_best == None or experiment1.pso.best.fitness < xor_best):
xor_best = experiment1.pso.best.fitness
xor_optimal_size = params_pso["swarmsize"]
print("\nRun ", j, "best size", xor_optimal_size, " produced",
xor_best)
print("xor optimal size ", xor_optimal_size, " produced", xor_best)
net_complex = {
"layer1": {
"input_count": 2,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 2,
"activations:": []
},
"layer3": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
print("\Swarmsize Complex")
print("=======================")
complex_optimal_size = 0
complex_best = None
for j in range(0, 1):
params_pso["swarmsize"] = 0
print("\nRun ", j)
complex_optimal_size = 0
complex_best = None
for i in range(0, 10):
params_pso["swarmsize"] += 10
experiment1 = Experiment(params_pso,
net_complex,
path="2in_complex.txt",
debugMode=False,
sampleMode=True)
experiment1.run()
if (complex_best == None
or experiment1.pso.best.fitness < complex_best):
complex_best = experiment1.pso.best.fitness
complex_optimal_size = params_pso["swarmsize"]
print("\nRun ", j, "best size", complex_optimal_size, " produced",
complex_best)
print("complex optimal size ", complex_optimal_size, " produced",
complex_best)
def run_final():
print("\Final Cubic")
print("=======================")
params_pso = {
"swarmsize": 40,
"alpha": 1,
"beta": 0,
"gamma": 4.1,
"delta": 0,
"jumpsize": 1,
"act_bound": 5,
"weight_bound": 10,
"bound_strat": 1,
"num_informants": 3,
"vel_range": 1,
"max_runs": 1000,
"informants_strat": 2
}
net_single = {
"layer1": {
"input_count": 1,
"node_count": 1,
"activations": []
}
}
exp1 = 0
for i in range(0, 10):
print("\nRun ", i)
experiment1 = Experiment(params_pso,
net_single,
path="1in_cubic.txt",
debugMode=False,
sampleMode=True)
experiment1.run()
exp1 += experiment1.pso.best.fitness
print("\nMse for final on cubic", exp1 / 10)
params_pso["beta"] = 0.5
params_pso["gamma"] = 3.6
print("\Final Linear")
print("=======================")
exp2 = 0
for i in range(0, 10):
print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_single,
path="1in_linear.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
exp2 += experiment2.pso.best.fitness
print("\nMse for final on linear", exp2 / 10)
params_pso["beta"] = 0
params_pso["gamma"] = 4.1
print("\Final Sine")
print("=======================")
exp3 = 0
for i in range(0, 10):
print("Run ", i, "\n")
experiment3 = Experiment(params_pso,
net_single,
path="1in_sine.txt",
debugMode=False,
sampleMode=True)
experiment3.run()
exp3 += experiment3.pso.best.fitness
print("\nMse for final on Sine", exp3 / 10)
net_layers = {
"layer1": {
"input_count": 1,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
params_pso["beta"] = 0
params_pso["gamma"] = 4.1
print("\nFinal Tanh")
print("=======================")
exp4 = 0
for i in range(0, 10):
print("Run ", i, "\n")
experiment4 = Experiment(params_pso,
net_layers,
path="1in_tanh.txt",
debugMode=False,
sampleMode=True)
experiment4.run()
exp4 += experiment4.pso.best.fitness
print("\nMse for final on Tanh", exp4 / 10)
net_layers = {
"layer1": {
"input_count": 2,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
params_pso["beta"] = 0
params_pso["gamma"] = 4.1
print("\nFinal XOR")
print("=======================")
exp6 = 0
for i in range(0, 10):
print("\nRun ", i)
experiment6 = Experiment(params_pso,
net_layers,
path="2in_xor.txt",
debugMode=False,
sampleMode=True)
experiment6.run()
exp6 += experiment6.pso.best.fitness
print("\nMse for final on XOR", exp6 / 10)
print("\nFinal Complex")
print("=======================")
net_complex = {
"layer1": {
"input_count": 2,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 2,
"activations:": []
},
"layer3": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
params_pso["beta"] = 2.05
params_pso["gamma"] = 2.05
exp5 = 0
for i in range(0, 10):
print("\nRun ", i, "\n")
experiment5 = Experiment(params_pso,
net_complex,
path="2in_complex.txt",
debugMode=False,
sampleMode=True)
experiment5.run()
exp5 += experiment5.pso.best.fitness
print("\nMse for final on Complex", exp5 / 10)
def run_informant_strat():
params_pso = {
"swarmsize": 40,
"alpha": 1,
"beta": 0.805,
"gamma": 3.295,
"delta": 0,
"jumpsize": 1,
"act_bound": 5,
"weight_bound": 10,
"bound_strat": 1,
"num_informants": 3,
"vel_range": 1,
"max_runs": 1000,
"informants_strat": 0
}
net_simple = {
"layer1": {
"input_count": 1,
"node_count": 1,
"activations": []
}
}
net_simple_2 = {
"layer1": {
"input_count": 2,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
exp2 = 0
for i in range(0, 10):
print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_simple,
path="1in_sine.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
exp2 += experiment2.pso.best.fitness
print("\nMse for strat 0 on linear", exp2 / 10)
exp2 = 0
params_pso["informants_strat"] = 1
for i in range(0, 10):
print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_simple,
path="1in_sine.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
exp2 += experiment2.pso.best.fitness
print("\nMse for strat 1 on linear", exp2 / 10)
exp2 = 0
params_pso["informants_strat"] = 2
for i in range(0, 10):
print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_simple,
path="1in_sine.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
exp2 += experiment2.pso.best.fitness
print("\nMse for strat 2 on linear", exp2 / 10)
exp2 = 0
params_pso["informants_strat"] = 3
for i in range(0, 10):
print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_simple,
path="1in_sine.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
exp2 += experiment2.pso.best.fitness
print("\nMse for strat 2 on linear", exp2 / 10)
def run_informant_count():
results = open("informantsresults.txt", "w+")
results.write("\n\nTEST")
params_pso = {
"swarmsize": 40,
"alpha": 1,
"beta": 2.05,
"gamma": 2.05,
"delta": 0,
"jumpsize": 1,
"act_bound": 5,
"weight_bound": 10,
"bound_strat": 1,
"num_informants": 0,
"vel_range": 1,
"max_runs": 1000,
"informants_strat": 2
}
net_simple = {
"layer1": {
"input_count": 1,
"node_count": 1,
"activations": []
}
}
params_pso["beta"] = 1.18
params_pso["gamma"] = 2.92
bestError = None
bestNum = 0
params_pso["num_informants"] = 0
for i in range(40):
params_pso["num_informants"] = i + 1
error = 0
for j in range(10):
# print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_simple,
path="1in_linear.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
error += experiment2.pso.best.fitness
avg = error / 10
results.write("\nLinear - Avg error for " + str(i) + " informants: " +
str(avg))
if (bestError == None or avg < bestError):
bestError = avg
bestNum = i + 1
results.write("\nBest Informants for linear " + str(bestNum) +
" Informants : " + str(bestError))
#Cubic
params_pso["beta"] = 1.325
params_pso["gamma"] = 2.775
bestError = None
bestNum = 0
params_pso["num_informants"] = 0
for i in range(40):
params_pso["num_informants"] = i + 1
error = 0
for j in range(10):
# print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_simple,
path="1in_cubic.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
error += experiment2.pso.best.fitness
avg = error / 10
results.write("\nCubic - Avg error for " + str(i) + " informants: " +
str(avg))
if (bestError == None or avg < bestError):
bestError = avg
bestNum = i + 1
results.write("\nBest Informants for cubic " + str(bestNum) +
" Informants : " + str(bestError))
#Sine
params_pso["beta"] = 0.91
params_pso["gamma"] = 3.19
bestError = None
bestNum = 0
params_pso["num_informants"] = 0
for i in range(40):
params_pso["num_informants"] = i + 1
error = 0
for j in range(10):
# print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_simple,
path="1in_sine.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
error += experiment2.pso.best.fitness
avg = error / 10
results.write("\nSine - Avg error for " + str(i) + " informants: " +
str(avg))
if (bestError == None or avg < bestError):
bestError = avg
bestNum = i + 1
results.write("\nBest Informants for sine " + str(bestNum) +
" Informants : " + str(bestError))
#Tanh
net_layers = {
"layer1": {
"input_count": 1,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
params_pso["beta"] = 0.805
params_pso["gamma"] = 3.295
bestError = None
bestNum = 0
params_pso["num_informants"] = 0
for i in range(40):
params_pso["num_informants"] = i + 1
error = 0
for j in range(10):
# print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_layers,
path="1in_tanh.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
error += experiment2.pso.best.fitness
avg = error / 10
results.write("\nTanh - Avg error for " + str(i) + " informants: " +
str(avg))
if (bestError == None or avg < bestError):
bestError = avg
bestNum = i + 1
results.write("\nBest Informants for tanh " + str(bestNum) +
" Informants : " + str(bestError))
#xor
net_layers = {
"layer1": {
"input_count": 2,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
params_pso["beta"] = 1.125
params_pso["gamma"] = 2.975
bestError = None
bestNum = 0
params_pso["num_informants"] = 0
for i in range(40):
params_pso["num_informants"] = i + 1
error = 0
for j in range(10):
# print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_layers,
path="2in_xor.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
error += experiment2.pso.best.fitness
avg = error / 10
results.write("\nXor - Avg error for " + str(i) + " informants: " +
str(avg))
if (bestError == None or avg < bestError):
bestError = avg
bestNum = i + 1
results.write("\nBest Informants for XOR " + str(bestNum) +
" Informants : " + str(bestError))
#Complex
net_complex = {
"layer1": {
"input_count": 2,
"node_count": 2,
"activations": []
},
"layer2": {
"input_count": 2,
"node_count": 2,
"activations:": []
},
"layer3": {
"input_count": 2,
"node_count": 1,
"activations:": []
}
}
params_pso["beta"] = 1.38
params_pso["gamma"] = 2.72
bestError = None
bestNum = 0
params_pso["num_informants"] = 0
for i in range(40):
params_pso["num_informants"] = i + 1
error = 0
for j in range(10):
# print("\nRun ", i)
experiment2 = Experiment(params_pso,
net_complex,
path="2in_complex.txt",
debugMode=False,
sampleMode=True)
experiment2.run()
error += experiment2.pso.best.fitness
avg = error / 10
results.write("\nComplex - Avg error for " + str(i) + " informants: " +
str(avg))
if (bestError == None or avg < bestError):
bestError = avg
bestNum = i + 1
results.write("\nBest Informants for Complex " + str(bestNum) +
" Informants : " + str(bestError))
results.close()
name='FullDQNAgent-1',
state_dim=env.observation_space.shape[0],
action_dim=env.action_space.n,
epsdecay=0.975,
buffersize=500000,
samplesize=32,
minsamples=1000,
gamma=0.99,
update_target_freq=600,
nnparams={ # Improved DQN setting
'hidden_layers': [(50, 'relu'), (40, 'relu')],
'loss': hubert_loss,
'optimizer': Adam(lr=0.0005),
'target_network': True
})
# Create an experiment with the LunarLander env and improved DQN agent for 500 train/test episodes
exp = Experiment(env, ragent, logdir="../log", verbose=True, num_episodes=500)
# Training trials
exp.run(testmode=False)
# Test trials
exp.run(testmode=True)
# Evaluate the agent in test trials
evaluate('../log/LunarLander_FullDQNAgent-1_test_data.csv',
'../log/LunarLander_FullDQNAgent-1_test_evaluation.png', 'Experiment')
#plt.show()
if __name__ == '__main__':
# TODO Use 'SimpleBaselineBinaryNetTanh' if you want labels [-1,1]
baseline = 'SimpleBaselineNet'
dataset = 'CIFAR_10'
distortion_type = 'AE_shift'
property = 'entropy'
property_perc = [0.25, 0.5]
classes = range(0, 10)
distortion_mean = [2, 3]
for p in property_perc:
experiment = Experiment(baseline=baseline, dataset=dataset, distortion_type=distortion_type,
mean=0, std=0, random=False, boosting_perc=1.0,
property='cross_entropy', property_perc=p, most=True, alternate=False,
classes=classes, classes_to_distort=None)
experiment.run()
experiment = Experiment(baseline=baseline, dataset=dataset, distortion_type=distortion_type,
mean=0, std=0, random=False, boosting_perc=1.0,
property='cross_entropy', property_perc=p, most=False, alternate=False,
classes=classes, classes_to_distort=None)
experiment.run()
experiment = Experiment(baseline=baseline, dataset=dataset, distortion_type=distortion_type,
mean=0, std=0, random=True, boosting_perc=1.0,
property='cross_entropy', property_perc=p, most=True, alternate=False,
classes=classes, classes_to_distort=None)
experiment.run()
exit()
#
# for p in property_perc:
def run(output_dir, gpu_id: int, config_name, **override_config):
"""
:param output_dir:
:param gpu_id: GPU (-1 == CPU)
:param config_name: Predefined experiment config
:param override_config: Use kwargs to override config variables, e.g., --foo__bar=1 (nested dict with __)
:return:
"""
output_dir = os.path.join(output_dir, config_name)
logger.info(f'Starting... {config_name}')
if os.path.exists(output_dir):
logger.error(f'Output dir exist already: {output_dir}')
sys.exit(1)
# GPU
if gpu_id < 0:
logger.info('GPU is disabled')
else:
os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id) # check with nvidia-smi
import torch
if not torch.cuda.is_available():
logger.error('CUDA is not available!')
sys.exit(1)
# Predefined configs
config_name = 'experiments.predefined.' + config_name
try:
package, module_name = config_name.rsplit('.', 1)
module = import_module(package)
config = getattr(module, module_name)
assert isinstance(config, dict) == True
except ValueError:
logger.error(f'Cannot load experiment config from: {config_name}')
sys.exit(1)
# Override config
from experiments.predefined import update
from experiments.utils import unflatten
if override_config:
override_config = unflatten(override_config)
logger.info(f'Override config with: {override_config}')
config = update(config, override_config)
from experiments import Experiment
exp = Experiment(**config)
exp.run(mode=2)
# save
os.makedirs(output_dir)
exp.output_dir = output_dir
with open(os.path.join(output_dir, 'experiment.pickle'), 'wb') as f:
# json.dump(exp.to_dict(), f)
pickle.dump(exp.to_dict(), f)
with open(os.path.join(output_dir, 'reports.json'), 'w') as f:
json.dump(exp.reports, f)
exp.save()
logger.info('Done')