class ProLoTrain:
"""
class structure to train the BDT with a certain alpha.
This class handles training the BDT, evaluating the BDT, and saving
"""
def __init__(self,num_schedules):
self.arguments = Logger()
self.alpha = .9
self.num_schedules = num_schedules
self.home_dir = self.arguments.home_dir
self.total_loss_array = []
load_directory = '/home/ghost/PycharmProjects/bayesian_prolo/scheduling_env/datasets/' + str(
self.num_schedules) + 'dist_early_hili_pairwise.pkl'
self.data = pickle.load(open(load_directory, "rb"))
self.X, self.Y, self.schedule_array = create_new_data(self.num_schedules, self.data)
self.start_of_each_set_twenty = create_sets_of_20_from_x_for_pairwise_comparisions(self.X)
self.model = ProLoNet(input_dim=len(self.X[0]),
weights=None,
comparators=None,
leaves=16,
output_dim=1,
bayesian_embedding_dim=None,
alpha=1.5,
use_gpu=True,
vectorized=True,
is_value=True)
use_gpu = True
if use_gpu:
self.model = self.model.cuda()
print(self.model.state_dict())
self.opt = torch.optim.RMSprop(self.model.parameters())
self.num_iterations_predict_task = 0
self.total_iterations = 0
self.covergence_epsilon = .01
self.when_to_save = 1000
self.distribution_epsilon = .0001
def train(self):
"""
Trains BDT.
Randomly samples a schedule and timestep within that schedule, produces training data using x_i - x_j
and trains upon that.
:return:
"""
# loss = nn.CrossEntropyLoss()
training_done = False
criterion = torch.nn.BCELoss()
# variables to keep track of loss and number of tasks trained over
while not training_done:
# sample a timestep before the cutoff for cross_validation
set_of_twenty = np.random.choice(self.start_of_each_set_twenty)
truth = self.Y[set_of_twenty]
# find feature vector of true action taken
phi_i_num = truth + set_of_twenty
phi_i = self.X[phi_i_num]
phi_i_numpy = np.asarray(phi_i)
running_loss_predict_tasks = 0
num_iterations_predict_task = 0
# iterate over pairwise comparisons
for counter in range(set_of_twenty, set_of_twenty + 20):
if counter == phi_i_num: # if counter == phi_i_num:
label = torch.ones((1, 1))
else:
label = torch.zeros((1, 1))
phi_j = self.X[counter]
phi = np.asarray(phi_j)
feature_input = phi
if torch.cuda.is_available():
feature_input = Variable(torch.Tensor(feature_input.reshape(1, 13)).cuda())
label = Variable(torch.Tensor(label).cuda())
else:
feature_input = Variable(torch.Tensor(feature_input.reshape(1, 13)))
label = Variable(torch.Tensor(label.reshape(1, 1)))
output = self.model.forward(feature_input)
sig = torch.nn.Sigmoid()
output = sig(output)
self.opt.zero_grad()
loss = criterion(output, label)
if counter == phi_i_num:
loss *= 25
# print(self.total_iterations)
if torch.isnan(loss):
print('nan occurred at iteration ', self.total_iterations, ' at', num_iterations_predict_task)
loss.backward()
# print(self.model.state_dict())
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.5)
self.opt.step()
running_loss_predict_tasks += loss.item()
num_iterations_predict_task += 1
# add average loss to array
# print(list(self.model.parameters()))
self.total_loss_array.append(running_loss_predict_tasks / num_iterations_predict_task)
self.total_iterations += 1
if self.total_iterations % 500 == 499:
print('total loss (average for each 40, averaged) at iteration ', self.total_iterations, ' is ', np.mean(self.total_loss_array[-40:]))
if self.total_iterations > 10000 and np.mean(self.total_loss_array[-100:]) - np.mean(
self.total_loss_array[-500:]) < self.covergence_epsilon:
training_done = True
def evaluate_on_test_data(self, model, load_in_model=False):
"""
Evaluate performance of a trained network tuned upon the alpha divergence loss.
This is tested on 20% of the data and will be stored in a text file.
Note this function is called after training convergence
:return:
"""
# define new optimizer that only optimizes gradient
num_schedules = 100
# load in new data
load_directory = '/home/ghost/PycharmProjects/bayesian_prolo/scheduling_env/datasets/' + str(
num_schedules) + 'test_dist_early_hili_pairwise.pkl'
data = pickle.load(open(load_directory, "rb"))
X, Y, schedule_array = create_new_data(num_schedules, data)
prediction_accuracy = [0, 0]
percentage_accuracy_top1 = []
percentage_accuracy_top3 = []
if load_in_model:
model.load_state_dict(torch.load('/home/ghost/PycharmProjects/bayesian_prolo/saved_models/pairwise_saved_models/model_homog.tar')['nn_state_dict'])
for j in range(0, num_schedules):
schedule_bounds = schedule_array[j]
step = schedule_bounds[0]
while step < schedule_bounds[1]:
probability_vector = np.zeros((1, 20))
for m, counter in enumerate(range(step, step + 20)):
phi_i = X[counter]
phi_i_numpy = np.asarray(phi_i)
feature_input = phi_i_numpy
if torch.cuda.is_available():
feature_input = Variable(torch.Tensor(feature_input.reshape(1, 13)).cuda())
else:
feature_input = Variable(torch.Tensor(feature_input.reshape(1, 13)))
# push through nets
preference_prob = model.forward(feature_input)
sig = torch.nn.Sigmoid()
preference_prob = sig(preference_prob)
probability_vector[0][m] = preference_prob[0].data.detach()[
0].item()
# probability_matrix[n][m] = preference_prob[0].data.detach()[1].item()
print(probability_vector)
highest_val = max(probability_vector[0])
all_indexes_that_have_highest_val = [i for i, e in enumerate(list(probability_vector[0])) if e == highest_val]
# top 1
choice = np.random.choice(all_indexes_that_have_highest_val)
# choice = np.argmax(probability_vector)
# top 3
_, top_three = torch.topk(torch.Tensor(probability_vector), 3)
# Then do training update loop
truth = Y[step]
# index top 1
if choice == truth:
prediction_accuracy[0] += 1
# index top 3
if truth in top_three:
prediction_accuracy[1] += 1
# add average loss to array
step += 20
# schedule finished
print('Prediction Accuracy: top1: ', prediction_accuracy[0] / 20, ' top3: ', prediction_accuracy[1] / 20)
print('schedule num:', j)
percentage_accuracy_top1.append(prediction_accuracy[0] / 20)
percentage_accuracy_top3.append(prediction_accuracy[1] / 20)
prediction_accuracy = [0, 0]
self.save_performance_results(percentage_accuracy_top1, percentage_accuracy_top3, 'DDT_pointwise'+ str(self.num_schedules))
def save_trained_nets(self, name):
"""
saves the model
:return:
"""
torch.save({'nn_state_dict': self.model.state_dict(),
'parameters': self.arguments},
'/home/ghost/PycharmProjects/bayesian_prolo/saved_models/pairwise_saved_models/BNN_' + name + '.tar')
def save_performance_results(self, top1, top3, special_string):
"""
saves performance of top1 and top3
:return:
"""
print('top1_mean for ', self.alpha, ' is : ', np.mean(top1))
data = {'top1_mean': np.mean(top1),
'top3_mean': np.mean(top3),
'top1_stderr': np.std(top1) / np.sqrt(len(top1)),
'top3_stderr': np.std(top3) / np.sqrt(len(top3))}
save_pickle(file=data, file_location=self.home_dir + '/saved_models/pairwise_saved_models/', special_string=special_string)
class ProLoTrain:
"""
class structure to train the BDT with a certain alpha.
This class handles training the BDT, evaluating the BDT, and saving
"""
def __init__(self, num_schedules):
self.arguments = Logger()
self.alpha = .9
self.num_schedules = num_schedules
self.home_dir = self.arguments.home_dir
self.total_loss_array = []
load_directory = '/home/ghost/PycharmProjects/bayesian_prolo/scheduling_env/datasets/' + str(
self.num_schedules) + 'dist_early_hili_pairwise.pkl'
self.data = pickle.load(open(load_directory, "rb"))
self.X, self.Y, self.schedule_array = create_new_data(
self.num_schedules, self.data)
self.start_of_each_set_twenty = create_sets_of_20_from_x_for_pairwise_comparisions(
self.X)
self.model = ProLoNet(input_dim=len(self.X[0]),
weights=None,
comparators=None,
leaves=32,
output_dim=1,
bayesian_embedding_dim=8,
alpha=1.5,
use_gpu=True,
vectorized=True,
is_value=True)
use_gpu = True
if use_gpu:
self.model = self.model.cuda()
print(self.model.state_dict())
self.opt = torch.optim.RMSprop(
[{
'params': list(self.model.parameters())[:-1]
}, {
'params': self.model.bayesian_embedding.parameters(),
'lr': .01
}],
lr=.01)
self.num_iterations_predict_task = 0
self.total_iterations = 0
self.covergence_epsilon = .01
self.when_to_save = 1000
self.distribution_epsilon = .0001
self.embedding_list = [
torch.ones(8) * 1 / 3 for _ in range(self.num_schedules)
]
def train(self):
"""
Trains BDT.
Randomly samples a schedule and timestep within that schedule, produces training data using x_i - x_j
and trains upon that.
:return:
"""
# loss = nn.CrossEntropyLoss()
sig = torch.nn.Sigmoid()
training_done = False
criterion = torch.nn.BCELoss()
# variables to keep track of loss and number of tasks trained over
while not training_done:
# sample a timestep before the cutoff for cross_validation
set_of_twenty = np.random.choice(self.start_of_each_set_twenty)
truth = self.Y[set_of_twenty]
which_schedule = find_which_schedule_this_belongs_to(
self.schedule_array, set_of_twenty)
self.model.set_bayesian_embedding(
self.embedding_list[which_schedule])
# find feature vector of true action taken
phi_i_num = truth + set_of_twenty
phi_i = self.X[phi_i_num]
phi_i_numpy = np.asarray(phi_i)
running_loss_predict_tasks = 0
num_iterations_predict_task = 0
# iterate over pairwise comparisons
for counter in range(set_of_twenty, set_of_twenty + 20):
# positive counterfactuals
if counter == phi_i_num: # if counter == phi_i_num:
continue
else:
phi_j = self.X[counter]
phi_j_numpy = np.asarray(phi_j)
feature_input = phi_i_numpy - phi_j_numpy
if torch.cuda.is_available():
feature_input = Variable(
torch.Tensor(feature_input.reshape(1, 13)).cuda())
label = Variable(
torch.Tensor(torch.ones((1, 1))).cuda())
else:
feature_input = Variable(
torch.Tensor(feature_input.reshape(1, 13)))
label = Variable(torch.Tensor(torch.ones((1, 1))))
output = self.model.forward(feature_input)
sig = torch.nn.Sigmoid()
output = sig(output)
self.opt.zero_grad()
loss = criterion(output, label)
loss.backward()
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.5)
self.opt.step()
running_loss_predict_tasks += loss.item()
num_iterations_predict_task += 1
# Negative counterfactuals
for counter in range(set_of_twenty, set_of_twenty + 20):
if counter == phi_i_num:
continue
else:
phi_j = self.X[counter]
phi_j_numpy = np.asarray(phi_j)
feature_input = phi_j_numpy - phi_i_numpy
if torch.cuda.is_available():
feature_input = Variable(
torch.Tensor(feature_input.reshape(1, 13)).cuda())
label = Variable(
torch.Tensor(torch.zeros((1, 1))).cuda())
else:
feature_input = Variable(
torch.Tensor(feature_input.reshape(1, 13)))
label = Variable(torch.Tensor(torch.zeros((1, 1))))
output = self.model.forward(feature_input)
sig = torch.nn.Sigmoid()
output = sig(output)
loss = criterion(output, label)
loss.backward()
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.5)
self.opt.step()
running_loss_predict_tasks += loss.item()
num_iterations_predict_task += 1
# add average loss to array
# print(list(self.model.parameters()))
self.total_loss_array.append(running_loss_predict_tasks /
num_iterations_predict_task)
self.embedding_list[which_schedule] = torch.Tensor(
self.model.get_bayesian_embedding().detach().cpu().numpy()
) # very ugly
self.total_iterations += 1
if self.total_iterations % 500 == 499:
print(
'total loss (average for each 40, averaged) at iteration ',
self.total_iterations, ' is ',
np.mean(self.total_loss_array[-40:]))
if self.total_iterations > 10000 and np.mean(
self.total_loss_array[-100:]) - np.mean(
self.total_loss_array[-500:]) < self.covergence_epsilon:
training_done = True
def evaluate_on_test_data(self, model, load_in_model=False):
"""
Evaluate performance of a trained network tuned upon the alpha divergence loss.
This is tested on 20% of the data and will be stored in a text file.
Note this function is called after training convergence
:return:
"""
# define new optimizer that only optimizes gradient
num_schedules = 100
# load in new data
load_directory = '/home/ghost/PycharmProjects/bayesian_prolo/scheduling_env/datasets/' + str(
num_schedules) + 'test_dist_early_hili_pairwise.pkl'
sig = torch.nn.Sigmoid()
data = pickle.load(open(load_directory, "rb"))
X, Y, schedule_array = create_new_data(num_schedules, data)
prediction_accuracy = [0, 0]
percentage_accuracy_top1 = []
percentage_accuracy_top3 = []
embedding_optimizer = torch.optim.SGD(
[{
'params': self.model.bayesian_embedding.parameters()
}], lr=.01)
criterion = torch.nn.BCELoss()
embedding_list = [torch.ones(3) * 1 / 3 for i in range(num_schedules)]
for j in range(0, num_schedules):
schedule_bounds = schedule_array[j]
step = schedule_bounds[0]
model.set_bayesian_embedding(embedding_list[j])
while step < schedule_bounds[1]:
probability_matrix = np.zeros((20, 20))
for m, counter in enumerate(range(step, step + 20)):
phi_i = X[counter]
phi_i_numpy = np.asarray(phi_i)
# for each set of twenty
for n, second_counter in enumerate(range(step, step + 20)):
# fill entire array with diagnols set to zero
if second_counter == counter: # same as m = n
continue
phi_j = X[second_counter]
phi_j_numpy = np.asarray(phi_j)
feature_input = phi_i_numpy - phi_j_numpy
if torch.cuda.is_available():
feature_input = Variable(
torch.Tensor(feature_input.reshape(1,
13)).cuda())
else:
feature_input = Variable(
torch.Tensor(feature_input.reshape(1, 13)))
# push through nets
preference_prob = model.forward(feature_input)
sig = torch.nn.Sigmoid()
preference_prob = sig(preference_prob)
probability_matrix[m][n] = preference_prob[
0].data.detach()[0].item(
) # TODO: you can do a check if only this line leads to the same thing as the line below
# probability_matrix[n][m] = preference_prob[0].data.detach()[1].item()
# Set of twenty is completed
column_vec = np.sum(probability_matrix, axis=1)
embedding_list[j] = torch.Tensor(
self.model.get_bayesian_embedding().detach().cpu().numpy(
)) # very ugly
# top 1
# given all inputs, and their liklihood of being scheduled, predict the output
highest_val = max(column_vec)
all_indexes_that_have_highest_val = [
i for i, e in enumerate(list(column_vec))
if e == highest_val
]
if len(all_indexes_that_have_highest_val) > 1:
print('length of indexes greater than 1: ',
all_indexes_that_have_highest_val)
# top 1
choice = np.random.choice(all_indexes_that_have_highest_val)
# choice = np.argmax(probability_vector)
# top 3
_, top_three = torch.topk(torch.Tensor(column_vec), 3)
# Then do training update loop
truth = Y[step]
# index top 1
if choice == truth:
prediction_accuracy[0] += 1
# index top 3
if truth in top_three:
prediction_accuracy[1] += 1
# Then do training update loop
phi_i_num = truth + step
phi_i = X[phi_i_num]
phi_i_numpy = np.asarray(phi_i)
# iterate over pairwise comparisons
for counter in range(step, step + 20):
if counter == phi_i_num:
continue
else:
phi_j = X[counter]
phi_j_numpy = np.asarray(phi_j)
feature_input = phi_i_numpy - phi_j_numpy
if torch.cuda.is_available():
feature_input = Variable(
torch.Tensor(feature_input.reshape(1,
13)).cuda())
label = Variable(
torch.Tensor(torch.ones((1, 1))).cuda())
else:
feature_input = Variable(
torch.Tensor(feature_input.reshape(1, 13)))
label = Variable(torch.Tensor(torch.ones((1, 1))))
output = model(feature_input)
output = sig(output)
loss = criterion(output, label)
# prepare optimizer, compute gradient, update params
embedding_optimizer.zero_grad()
loss.backward()
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.5)
embedding_optimizer.step()
# print(model.EmbeddingList.state_dict())
for counter in range(step, step + 20):
if counter == phi_i_num:
continue
else:
phi_j = X[counter]
phi_j_numpy = np.asarray(phi_j)
feature_input = phi_j_numpy - phi_i_numpy
if torch.cuda.is_available():
feature_input = Variable(
torch.Tensor(feature_input.reshape(1,
13)).cuda())
label = Variable(
torch.Tensor(torch.zeros((1, 1))).cuda())
else:
feature_input = Variable(
torch.Tensor(feature_input.reshape(1, 13)))
label = Variable(torch.Tensor(torch.zeros((1, 1))))
output = model.forward(feature_input)
output = sig(output)
embedding_optimizer.zero_grad()
loss = criterion(output, label)
loss.backward()
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.5)
embedding_optimizer.step()
# print(model.EmbeddingList.state_dict())
# add average loss to array
step += 20
# schedule finished
print('Prediction Accuracy: top1: ', prediction_accuracy[0] / 20,
' top3: ', prediction_accuracy[1] / 20)
print('schedule num:', j)
percentage_accuracy_top1.append(prediction_accuracy[0] / 20)
percentage_accuracy_top3.append(prediction_accuracy[1] / 20)
prediction_accuracy = [0, 0]
self.save_performance_results(
percentage_accuracy_top1, percentage_accuracy_top3,
'PDDT_pairwise' + str(self.num_schedules))
def save_trained_nets(self, name):
"""
saves the model
:return:
"""
torch.save(
{
'nn_state_dict': self.model.state_dict(),
'parameters': self.arguments
},
'/home/ghost/PycharmProjects/bayesian_prolo/saved_models/pairwise_saved_models/BNN_'
+ name + '.tar')
def save_performance_results(self, top1, top3, special_string):
"""
saves performance of top1 and top3
:return:
"""
print('top1_mean for ', self.alpha, ' is : ', np.mean(top1))
data = {
'top1_mean': np.mean(top1),
'top3_mean': np.mean(top3),
'top1_stderr': np.std(top1) / np.sqrt(len(top1)),
'top3_stderr': np.std(top3) / np.sqrt(len(top3))
}
save_pickle(file=data,
file_location=self.home_dir +
'/saved_models/pairwise_saved_models/',
special_string=special_string)
class Trainer:
def __init__(self):
self.loss_array = []
checkpoint = torch.load(
'/home/ghost/PycharmProjects/scheduling_environment/learn_action_embeddings.pkl'
)
self.action_embedding_list = checkpoint['embedding_list']
self.all_data_train_dir = '/home/ghost/PycharmProjects/scheduling_environment' + '/training_encoding_states_starcraft'
self.mmr = '/home/ghost/PycharmProjects/scheduling_environment' + '/games_that_have_an_win_loss.pkl'
self.list_of_games_mmr_train = pickle.load(open(self.mmr, "rb"))
self.size_of_training_set = len(self.list_of_games_mmr_train)
self.all_data_test_dir = '/home/ghost/PycharmProjects/scheduling_environment/testing_encoding_states_starcraft'
self.gamma = .9
self.criterion = torch.nn.BCELoss()
self.not_converged = True
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self.embedding_list_Bayesian_neur_net = [
torch.ones(12) * 1 / 2 for i in range(4000)
]
self.model = ProLoNet(input_dim=256 + 36,
weights=None,
comparators=None,
leaves=64,
output_dim=1,
bayesian_embedding_dim=12,
alpha=1.5,
use_gpu=True,
vectorized=True,
is_value=True)
self.optimizer_main_net = torch.optim.Adam(
[{
'params': list(self.model.parameters())[:-1]
}, {
'params': self.model.bayesian_embedding.parameters(),
'lr': .01
}],
lr=.001)
self.embedding_optimizer_bnn = torch.optim.SGD(
self.model.bayesian_embedding.parameters(), lr=.1)
# noinspection PyArgumentList
def train(self, num_steps_for_batch):
sig = torch.nn.Sigmoid()
iteration = 0
while self.not_converged:
# files in training directory
files_in_training_directory = os.listdir(self.all_data_train_dir)
# choose a batch of data
batch = np.random.choice(files_in_training_directory)
# open batch
set_of_games = pickle.load(
open(os.path.join(self.all_data_train_dir, batch), 'rb'))
games = set_of_games['state_embeddings'] # state embedding dict
actions = set_of_games['actions_at_each_frame']
players = set_of_games['player_per_game']
big_loss = set_of_games['big_loss'] # maybe games to skip
for l in range(num_steps_for_batch):
file_with_low_loss_found = False
filename = None
# loop to find a filename with low loss (fairly inefficient)
while not file_with_low_loss_found:
filename = np.random.choice(list(games.keys()))
if len(big_loss[filename]) < 50:
file_with_low_loss_found = True
states_filename = games[filename]
actions_filename = actions[filename]
player_filename = players[filename]
length_of_game = len(states_filename)
# print("iteration ", iteration)
# choose a random frame
frame = np.random.randint(1, length_of_game)
player_id = int(re.findall('\d+', batch)[0]) + list(
games.keys()).index(filename)
self.model.set_bayesian_embedding(
self.embedding_list_Bayesian_neur_net[player_id])
X = states_filename[frame] # input - a
actions_taken_at_frame = actions_filename[
frame] # set of actions taken
actions_taken_list = [
i for i, e in enumerate(actions_taken_at_frame) if e != 0
]
set_of_non_actions = self.compute_set_of_non_actions(
actions_taken_list)
# True
Y = torch.ones((1, 1)) # label
if torch.cuda.is_available():
X = Variable(torch.Tensor(X).cuda())
Y = Variable(torch.Tensor(Y).cuda())
else:
X = Variable(torch.Tensor(X))
Y = Variable(torch.Tensor(Y))
running_loss = []
for a in actions_taken_list:
action_embedding_a = self.action_embedding_list[int(a)]
for i, non_action in enumerate(set_of_non_actions):
action_embeddimg_a_prime = self.action_embedding_list[
non_action]
subtracted_input = action_embedding_a.cuda(
) - action_embeddimg_a_prime.cuda()
prediction = self.model(
torch.cat([X, subtracted_input.reshape(36)],
dim=0).reshape(1, 292))
prediction = sig(prediction)
prediction_loss = self.criterion(prediction, Y)
self.optimizer_main_net.zero_grad()
# if int(a) != 0:
# prediction_loss *= 5
prediction_loss.backward()
self.optimizer_main_net.step()
running_loss.append(prediction_loss.item())
# False
Y = torch.zeros((1, 1)) # label
if torch.cuda.is_available():
Y = Variable(torch.Tensor(Y).cuda())
else:
Y = Variable(torch.Tensor(Y))
for a in actions_taken_list:
action_embedding_a = self.action_embedding_list[int(a)]
for i, non_action in enumerate(set_of_non_actions):
action_embeddimg_a_prime = self.action_embedding_list[
non_action]
subtracted_input = action_embeddimg_a_prime.cuda(
) - action_embedding_a.cuda()
prediction = self.model(
torch.cat([X, subtracted_input.reshape(36)],
dim=0).reshape(1, 292))
prediction = sig(prediction)
# print(prediction)
prediction_loss = self.criterion(prediction, Y)
self.optimizer_main_net.zero_grad()
prediction_loss.backward()
self.optimizer_main_net.step()
running_loss.append(prediction_loss.item())
if iteration > 55 and iteration % 50 == 49:
print('Mean loss for iteration: ', iteration, ' is ',
np.mean(running_loss[-50:]))
self.loss_array.append(np.mean(running_loss))
self.embedding_list_Bayesian_neur_net[
player_id] = torch.Tensor(
self.model.get_bayesian_embedding().detach().cpu(
).numpy()) # very ugly
iteration += 1
if iteration % 1000 == 999:
# self.plot_networks()
self.save()
if iteration > 6000 and np.mean(
self.loss_array[-500:]) - np.mean(
self.loss_array[-1000:]) < .001:
self.not_converged = False
# noinspection PyArgumentList
def test(self, load_in_model=False):
iteration = 0
tot_test_games = 0
max_frames = 0
loss_lists_per_timestep_nn = []
sig = torch.nn.Sigmoid()
game_accuracies = []
embedding_list = [torch.ones(12) * 1 / 2 for i in range(4000)]
if load_in_model == True:
checkpoint = torch.load(
'/home/ghost/PycharmProjects/bayesian_prolo/starcraft/pairwise_sc2_PDDT.pkl'
)
self.model.load_state_dict(checkpoint['state_dict'])
files_in_testing_directory = os.listdir(self.all_data_test_dir)
for each_batch in files_in_testing_directory:
print(each_batch)
set_of_games = pickle.load(
open(os.path.join(self.all_data_test_dir, each_batch), 'rb'))
games = set_of_games['state_embeddings'] # state embedding dict
actions = set_of_games['actions_at_each_frame']
players = set_of_games['player_per_game']
big_loss = set_of_games['big_loss'] # maybe games to skip
tot_test_games += len(games.keys())
for game_num, filename in enumerate(games.keys()):
if len(big_loss[filename]) > 50:
continue
states_filename = games[filename]
actions_filename = actions[filename]
player_filename = players[filename]
player_id = int(re.findall('\d+', each_batch)[0]) + list(
games.keys()).index(filename)
self.model.set_bayesian_embedding(embedding_list[player_id])
print("iteration ", iteration)
frame = 0
iteration += 1
length_of_game = len(states_filename)
# choose a random frame
while frame < length_of_game - 2:
self.probability_matrix_nn = np.zeros((40, 40))
print('reached frame ', frame)
# choose a random frame
X = states_filename[frame] # input - a
actions_taken_at_frame = actions_filename[
frame] # set of actions taken
actions_taken_list = [
i for i, e in enumerate(actions_taken_at_frame)
if e != 0
]
set_of_non_actions = self.compute_set_of_non_actions(
actions_taken_list)
if torch.cuda.is_available():
X = Variable(torch.Tensor(X).cuda())
else:
X = Variable(torch.Tensor(X))
for i in range(40):
action_embedding_a = self.action_embedding_list[int(i)]
for j in range(40):
if i == j:
continue
else:
action_embedding_a_prime = self.action_embedding_list[
int(j)]
subtracted_input = action_embedding_a.cuda(
) - action_embedding_a_prime.cuda()
prediction = self.model.forward(
torch.cat(
[X, subtracted_input.reshape(36)],
dim=0).reshape(1, 292))
prediction = sig(prediction)
# add all these to matrixes
self.probability_matrix_nn[i][
j] = prediction.item()
column_vec_nn = np.sum(self.probability_matrix_nn, axis=1)
loss = torch.nn.BCELoss()
soft = nn.Softmax(dim=0)
column_vec_nn = soft(torch.Tensor(column_vec_nn))
loss_nn = loss(
column_vec_nn,
Variable(torch.Tensor(actions_taken_at_frame)))
# loss_nn = np.linalg.norm(column_vec_nn - actions_taken_at_frame)
loss_lists_per_timestep_nn.append(loss_nn.item())
# True
Y = torch.ones((1, 1)) # label
if torch.cuda.is_available():
Y = Variable(torch.Tensor(Y).cuda())
else:
Y = Variable(torch.Tensor(Y))
for a in actions_taken_list:
action_embedding_a = self.action_embedding_list[int(a)]
for i, non_action in enumerate(set_of_non_actions):
action_embeddimg_a_prime = self.action_embedding_list[
non_action]
subtracted_input = action_embedding_a.cuda(
) - action_embeddimg_a_prime.cuda()
prediction = self.model(
torch.cat([X, subtracted_input.reshape(36)],
dim=0).reshape(1, 292))
prediction = sig(prediction)
prediction_loss = self.criterion(prediction, Y)
self.embedding_optimizer_bnn.zero_grad()
prediction_loss.backward()
self.embedding_optimizer_bnn.step()
# False
Y = torch.zeros((1, 1)) # label
if torch.cuda.is_available():
Y = Variable(torch.Tensor(Y).cuda())
else:
Y = Variable(torch.Tensor(Y))
for a in actions_taken_list:
action_embedding_a = self.action_embedding_list[int(a)]
for i, non_action in enumerate(set_of_non_actions):
action_embeddimg_a_prime = self.action_embedding_list[
non_action]
subtracted_input = action_embeddimg_a_prime.cuda(
) - action_embedding_a.cuda()
prediction = self.model(
torch.cat([X, subtracted_input.reshape(36)],
dim=0).reshape(1, 292))
prediction = sig(prediction)
prediction_loss = self.criterion(prediction, Y)
self.embedding_optimizer_bnn.zero_grad()
prediction_loss.backward()
self.embedding_optimizer_bnn.step()
frame += 1
game_accuracies.append(np.mean(loss_lists_per_timestep_nn))
loss_lists_per_timestep_nn = []
self.print_and_store_accs(game_accuracies)
if len(game_accuracies) == 5:
self.save_embeddings(embedding_list)
if len(game_accuracies) == 15:
self.print_and_store_accs(game_accuracies)
self.save_embeddings()
exit()
# game has ended
# do end of schedule tings
# a batch has ended
# finished errthang
print(tot_test_games)
print(max_frames)
# self.plot_with_errorbars()
def print_and_store_accs(self, game_accs):
print('Loss: {}'.format(np.mean(game_accs)))
file = open('starcraft_learning_results.txt', 'a')
file.write('PDDT_pairwise: mean: ' + str(np.mean(game_accs)) +
', std: ' + str(np.std(game_accs)) + '\n')
file.close()
def save(self):
torch.save({
'state_dict': self.model.state_dict()
}, '/home/ghost/PycharmProjects/bayesian_prolo/starcraft/pairwise_sc2_PDDT.pkl'
)
def save_embeddings(self, embeddings):
torch.save(
{
'state_dict': self.model.state_dict(),
'embeddings': embeddings
},
'/home/ghost/PycharmProjects/bayesian_prolo/starcraft/pairwise_sc2_PDDT_embeddings.pkl'
)
def compute_set_of_non_actions(self, actions):
set_of_non_actions = list(range(40))
for each_action in actions:
set_of_non_actions.remove(each_action)
return set_of_non_actions
class ProLoTrain:
"""
class structure to train the BDT with a certain alpha.
This class handles training the BDT, evaluating the BDT, and saving
"""
def __init__(self, num_schedules):
self.arguments = Logger()
self.alpha = .9
self.num_schedules = num_schedules
self.home_dir = self.arguments.home_dir
self.total_loss_array = []
load_directory = '/home/ghost/PycharmProjects/bayesian_prolo/scheduling_env/datasets/' + str(
self.num_schedules) + 'dist_early_hili_naive.pkl'
self.data = pickle.load(open(load_directory, "rb"))
self.X, self.Y, self.schedule_array = create_new_dataset(num_schedules=self.num_schedules, data=self.data)
for i, each_element in enumerate(self.X):
self.X[i] = each_element + list(range(20))
self.model = ProLoNet(input_dim=len(self.X[0]),
weights=None,
comparators=None,
leaves=64,
output_dim=20,
bayesian_embedding_dim=8,
alpha=1.5,
use_gpu=True,
vectorized=True,
is_value=False)
use_gpu = True
if use_gpu:
self.model = self.model.cuda()
print(self.model.state_dict())
self.opt = torch.optim.RMSprop([{'params': list(self.model.parameters())[:-1]}, {'params': self.model.bayesian_embedding.parameters(), 'lr': .01}])
self.num_iterations_predict_task = 0
self.total_iterations = 0
self.covergence_epsilon = .01
self.when_to_save = 1000
self.distribution_epsilon = .0001
self.max_depth = 10
self.embedding_list = [torch.ones(8) * 1 / 3 for _ in range(self.num_schedules)]
def train(self):
"""
Trains PDDT.
:return:
"""
threshold = .05
training_done = False
while not training_done:
# sample a timestep before the cutoff for cross_validation
rand_timestep_within_sched = np.random.randint(len(self.X))
input_nn = self.X[rand_timestep_within_sched]
truth_nn = self.Y[rand_timestep_within_sched]
which_schedule = find_which_schedule_this_belongs_to(self.schedule_array, rand_timestep_within_sched)
self.model.set_bayesian_embedding(self.embedding_list[which_schedule])
if torch.cuda.is_available():
input_nn = Variable(torch.Tensor(np.asarray(input_nn).reshape(1, 242)).cuda()) # change to 5 to increase batch size
truth = Variable(torch.Tensor(np.asarray(truth_nn).reshape(1)).cuda().long())
else:
input_nn = Variable(torch.Tensor(np.asarray(input_nn).reshape(1, 242)))
truth = Variable(torch.Tensor(np.asarray(truth_nn).reshape(1)).long())
self.opt.zero_grad()
output = self.model.forward(input_nn)
loss = F.cross_entropy(output, truth)
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.5)
self.opt.step()
self.embedding_list[which_schedule] = torch.Tensor(self.model.get_bayesian_embedding().detach().cpu().numpy()) # very ugly
# add average loss to array
# print(list(self.model.parameters()))
self.total_loss_array.append(loss.item())
self.total_iterations += 1
if self.total_iterations > 25 and self.total_iterations % 50 == 1:
print('total iterations is', self.total_iterations)
print('total loss (average for each 40, averaged)', np.mean(self.total_loss_array[-40:]))
# print(self.model.state_dict())
if self.total_iterations > 0 and self.total_iterations % self.when_to_save == self.when_to_save - 1:
self.save_trained_nets('PDDT' + str(self.num_schedules))
threshold -= .1
if self.total_iterations > 100000 and np.mean(self.total_loss_array[-100:]) - np.mean(
self.total_loss_array[-500:]) < self.covergence_epsilon:
training_done = True
def evaluate_on_test_data(self, load_in_model=False):
"""
Evaluate performance of a trained network tuned upon the alpha divergence loss.
Note this function is called after training convergence
:return:
"""
# define new optimizer that only optimizes gradient
num_schedules = 100
load_directory = '/home/ghost/PycharmProjects/bayesian_prolo/scheduling_env/datasets/' + str(
num_schedules) + 'test_dist_early_hili_naive.pkl'
data = pickle.load(open(load_directory, "rb"))
X, Y, schedule_array = create_new_dataset(num_schedules=num_schedules, data=data)
for i, each_element in enumerate(X):
X[i] = each_element + list(range(20))
embedding_optimizer = torch.optim.SGD([{'params': self.model.bayesian_embedding.parameters()}], lr=.1)
embedding_list = [torch.ones(8) * 1 / 3 for _ in range(num_schedules)]
prediction_accuracy = [0, 0]
percentage_accuracy_top1 = []
percentage_accuracy_top3 = []
if load_in_model:
self.model.load_state_dict(torch.load('/home/ghost/PycharmProjects/bayesian_prolo/saved_models/pairwise_saved_models/model_homog.tar')['nn_state_dict'])
for i, schedule in enumerate(schedule_array):
self.model.set_bayesian_embedding(self.embedding_list[i])
for count in range(schedule[0], schedule[1] + 1):
net_input = X[count]
truth = Y[count]
if torch.cuda.is_available():
input_nn = Variable(torch.Tensor(np.asarray(net_input).reshape(1, 242)).cuda()) # change to 5 to increase batch size
truth = Variable(torch.Tensor(np.asarray(truth).reshape(1)).cuda().long())
else:
input_nn = Variable(torch.Tensor(np.asarray(net_input).reshape(1, 242)))
truth = Variable(torch.Tensor(np.asarray(truth).reshape(1)).long())
#####forward#####
output = self.model.forward(input_nn)
embedding_optimizer.zero_grad()
loss = F.cross_entropy(output, truth)
loss.backward()
# torch.nn.utils.clip_grad_norm_(self.model.parameters(), 0.5)
embedding_optimizer.step()
index = torch.argmax(output).item()
# top 3
_, top_three = torch.topk(output, 3)
if index == truth.item():
prediction_accuracy[0] += 1
if truth.item() in top_three.detach().cpu().tolist()[0]:
prediction_accuracy[1] += 1
# add average loss to array
embedding_list[i] = torch.Tensor(self.model.get_bayesian_embedding().detach().cpu().numpy()) # very ugly
# schedule finished
print('Prediction Accuracy: top1: ', prediction_accuracy[0] / 20, ' top3: ', prediction_accuracy[1] / 20)
print('schedule num:', i)
percentage_accuracy_top1.append(prediction_accuracy[0] / 20)
percentage_accuracy_top3.append(prediction_accuracy[1] / 20)
prediction_accuracy = [0, 0]
self.save_performance_results(percentage_accuracy_top1, percentage_accuracy_top3, 'DDT_w_embedding')
def save_trained_nets(self, name):
"""
saves the model
:return:
"""
torch.save({'nn_state_dict': self.model.state_dict(),
'parameters': self.arguments},
'/home/ghost/PycharmProjects/bayesian_prolo/saved_models/naive_saved_models/BNN_' + name + '.tar')
def save_performance_results(self, top1, top3, special_string):
"""
saves performance of top1 and top3
:return:
"""
print('top1_mean for ', self.alpha, ' is : ', np.mean(top1))
data = {'top1_mean': np.mean(top1),
'top3_mean': np.mean(top3),
'top1_stderr': np.std(top1) / np.sqrt(len(top1)),
'top3_stderr': np.std(top3) / np.sqrt(len(top3))}
save_pickle(file=data, file_location=self.home_dir + '/saved_models/naive_saved_models/', special_string=special_string)