def test_default_can_play_2048(self):
n = Net()
b = Board()
data = b._tiles.flatten()
output = n.run(data)
self.assertEqual(output.shape, (n.outputs, ))
b.move(output.argmax(), suppress_invalid=True)
def __init__(
self,
net_number=None,
epoch_number=None,
experiment_name=None,
use_naive_opponent=False,
play_solo=False,
**kwargs):
self.board = BoardTTT(**kwargs)
self.opponent = None
if use_naive_opponent:
self.opponent = NaiveTTTPlayer()
elif (not play_solo
and epoch_number is not None
and experiment_name is not None
and net_number is not None):
from nets import Net
from trainer import GeneticNetTrainer
try:
net_data = GeneticNetTrainer.load_single_net(
idx=net_number,
epoch_num=epoch_number,
experiment_name=experiment_name)
# Hardcoding is hard
net_hidden_sizes = (3, 3)
# 2D Tic Tac Toe sizes
net_inputs = 9
net_outputs = 9
self.opponent = Net(
hidden_sizes=net_hidden_sizes,
weights=net_data,
inputs=net_inputs,
outputs=net_outputs)
except IOError:
pass
def __init__(self,
net_number=None,
epoch_number=None,
experiment_name=None,
use_naive_opponent=False,
play_solo=False,
**kwargs):
self.board = BoardTTT(**kwargs)
self.opponent = None
if use_naive_opponent:
self.opponent = NaiveTTTPlayer()
elif (not play_solo and epoch_number is not None
and experiment_name is not None and net_number is not None):
from nets import Net
from trainer import GeneticNetTrainer
try:
net_data = GeneticNetTrainer.load_single_net(
idx=net_number,
epoch_num=epoch_number,
experiment_name=experiment_name)
# Hardcoding is hard
net_hidden_sizes = (3, 3)
# 2D Tic Tac Toe sizes
net_inputs = 9
net_outputs = 9
self.opponent = Net(hidden_sizes=net_hidden_sizes,
weights=net_data,
inputs=net_inputs,
outputs=net_outputs)
except IOError:
pass
def get_net(name, device):
if name == 'MNIST':
return Net(MNIST_Net, params[name], device)
elif name == 'FashionMNIST':
return Net(MNIST_Net, params[name], device)
elif name == 'SVHN':
return Net(SVHN_Net, params[name], device)
elif name == 'CIFAR10':
return Net(CIFAR10_Net, params[name], device)
else:
raise NotImplementedError
def train(self):
save_path = "models/net_arcloss2.pth"
epoch = 0
train_data = torchvision.datasets.MNIST(root="./MNIST",
download=True,
train=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(
mean=[
0.5,
],
std=[
0.5,
])
]))
train_loader = data.DataLoader(train_data,
shuffle=True,
batch_size=100)
net = Net().to(self.device)
if os.path.exists(save_path):
net.load_state_dict(torch.load(save_path))
else:
print("NO Param")
optimizer = torch.optim.SGD(net.parameters(),
lr=1e-3,
momentum=0.9,
weight_decay=0.0005)
# optimizer = torch.optim.Adam(net.parameters())
scheduler = lr_scheduler.StepLR(optimizer, 20, gamma=0.8)
while True:
feat_loader = []
label_loader = []
for i, (x, y) in enumerate(train_loader):
x = x.to(self.device)
y = y.to(self.device)
feature, output = net.forward(x)
loss = self.lossfn_cls(output, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
feat_loader.append(feature)
label_loader.append(y)
if i % 600 == 0:
print("epoch:", epoch, "i:", i, "arcsoftmax_loss:",
loss.item())
feat = torch.cat(feat_loader, 0)
labels = torch.cat(label_loader, 0)
net.visualize(feat.data.cpu().numpy(),
labels.data.cpu().numpy(), epoch)
epoch += 1
torch.save(net.state_dict(), save_path)
scheduler.step()
if epoch == 150:
break
def moons_test():
print('-' * 10, 'moons test', '-' * 10)
data = datasets.make_moons(n_samples=1000)
X = data[0]
y = np.expand_dims(data[1], 1) # binary classification
print('X: {}, y: {}'.format(X.shape, y.shape))
net = Net([Linear(2, 8, act='sigmoid'),
Linear(8, 1, act='sigmoid')],
loss='mse')
print(net)
net.train(X, y, batch_size=64, epochs=100, lr=0.5)
print('-' * 35)
def test_trainer_respects_net_properties(self):
iters = 77
h_sizes = (4, )
i_size = 17
o_size = 44
t = GeneticNetTrainer(
iterations_per_model=iters,
hidden_sizes=h_sizes,
inputs=i_size,
outputs=o_size,
)
sizes = (i_size, ) + h_sizes + (o_size, )
def test_net(net, num_iterations):
self.assertEqual(num_iterations, iters)
self.assertEqual(net.sizes, sizes)
return (0, )
t.test_net = test_net
t.run_model(None)
t2 = GeneticNetTrainer(
iterations_per_model=iters,
hidden_sizes=h_sizes,
weights=Net.random_weights(sizes),
inputs=i_size,
outputs=o_size,
)
t2.test_net = test_net
t2.run_model(None)
def get_net_from_logs(logs):
"""
Best tf.keras.model
(best: best performance on validation set out of all models in the logdir)
used to choose optimum hyperparameters.
Arguments:
dataset: instance of data.Cleaned or data.Simulated
logdir(str): path to the logging folder of the model.
The logging folder of a trained model contains:
1) folder "Training", witch contains logdirs
of individual hyperparameter searches
2) pickle "args.pickle", which contains a dict of the network args.
"""
with open(os.path.join(logdir, "args.pickle"), "rb") as f:
args = argparse.Namespace(**pickle.load(
f)) # loads dict and converts it to namespace
analysis = tune.Analysis(os.path.join(logdir, "Training"))
best_config = analysis.get_best_config(metric="valid_rmse", mode="min")
best_logdir = analysis.get_best_logdir(metric="valid_rmse", mode="min")
model = create_model(args=args, **best_config)
checkpoint_folder_name = [
s for s in os.listdir(best_logdir) if s.startswith("checkpoint")
][0]
model.load_weights(
os.path.join(best_logdir, checkpoint_folder_name, "model.h5"))
return Net(model, args)
def make_net_weights():
return Net(
hidden_sizes=self.net_hidden_sizes,
weights=None,
inputs=self.net_inputs,
outputs=self.net_outputs,
weight_spread=self.net_spread,
weight_middle=self.net_middle).weights
def test_square_can_process_input(self):
hidden_weights = (
numpy.array([[3.0, 2.0, 0.5], [3.0, 0.5, 2.0], [0.5, 3.0, 2.0]]),
numpy.array([[1.0, 1.0, 1.0], [1.0, 1.0, 1.0], [1.0, 1.0, 1.0]]),
)
n = Net(hidden_sizes=(3, ),
weights=hidden_weights,
inputs=3,
outputs=3)
# Layer one:
# 10.5 12 10.5
# Layer two:
# 33 33 33
# Softmax:
# 0.33 0.33 0.33
data = numpy.array([1, 2, 3])
output = n.run(data)
for i in output:
self.assertAlmostEqual(i, 0.333, 3)
def test_can_make_nets(self):
n = Net(hidden_sizes=(8, ), inputs=16, outputs=4)
self.assertEqual(n.sizes, (16, 8, 4))
self.assertEqual(len(n.weights), 2)
self.assertEqual(n.weights[0].shape, (16, 8))
self.assertEqual(n.weights[1].shape, (8, 4))
self.assertRaises(ValueError, Net, 3)
self.assertRaises(ValueError, Net, weights='haha')
self.assertRaises(ValueError, Net, weights=(1, ))
self.assertRaises(ValueError, Net, inputs='lol')
self.assertRaises(ValueError, Net, outputs='rofl')
def run_model(self,
weights_test,
weights_adversary,
num_iterations=None,
naive_player=None):
"""
@return: (ModelScore_test, ModelScore_adversary) where each is a list
of ModelScore tuples
Or a single list of ModelScore tuples for the non-naive player
if one is naive
"""
num_iterations = num_iterations or self.iterations_per_model
if naive_player == 1:
n1 = weights_test()
else:
n1 = Net(
hidden_sizes=self.net_hidden_sizes,
weights=weights_test,
inputs=self.net_inputs,
outputs=self.net_outputs,
weight_spread=self.net_spread,
weight_middle=self.net_middle)
if naive_player == 2:
n2 = weights_adversary()
else:
n2 = Net(
hidden_sizes=self.net_hidden_sizes,
weights=weights_adversary,
inputs=self.net_inputs,
outputs=self.net_outputs,
weight_spread=self.net_spread,
weight_middle=self.net_middle)
# We score both nets
score_params = self.test_net(n1, n2, num_iterations)
if naive_player is None:
return score_params
else:
return score_params[naive_player - 1]
def run_model(self, weights, weights_idx, num_iterations=None):
"""
@return: A ModelStats tuple
"""
num_iterations = num_iterations or self.iterations_per_model
n = Net(hidden_sizes=self.net_hidden_sizes,
weights=weights,
inputs=self.net_inputs,
outputs=self.net_outputs,
weight_spread=self.net_spread,
weight_middle=self.net_middle)
score_params = self.test_net(n, num_iterations)
return ModelStats(weights_idx=weights_idx,
score_parameters=score_params,
score=self.get_model_score(score_params))
def main(args):
'''
main function
'''
if args.logport:
args.logport = Dashboard(args.logport, 'dashboard')
EPOCH_SIZE = args.num_episode*args.num_query*args.way_train
EPOCH_SIZE_TEST = args.num_episode_test*args.num_query*args.way_test
'''define network'''
net = Net(args.num_in_channel, args.num_filter)
relationNet = RelationNet(args.num_filter*2, args.num_filter, 5*5*args.num_filter, args.num_fc, args.drop_prob)
if torch.cuda.is_available():
net.cuda()
relationNet.cuda()
'''
load model if needed
'''
if args.model_load_path_net!='':
net.load_state_dict(torch.load(args.model_load_path_net))
net.cuda()
relationNet.load_state_dict(torch.load(args.model_load_path_relationNet))
relationNet.cuda()
print('model loaded')
''' define loss, optimizer'''
criterion = nn.MSELoss()
params = list(net.parameters()) + list(relationNet.parameters())
optimizer = optim.Adam(params, lr=args.learning_rate)
'''get data'''
trainList = read_miniImageNet_pathonly(TESTMODE=False,
miniImageNetPath='/home/fujenchu/projects/dataset/miniImageNet_Ravi/',
imgPerCls=600)
testList = read_miniImageNet_pathonly(TESTMODE=True,
miniImageNetPath='/home/fujenchu/projects/dataset/miniImageNet_Ravi/',
imgPerCls=600)
scheduler = StepLR(optimizer, step_size=40, gamma=0.5)
''' training'''
for epoch in range(1000):
scheduler.step()
running_loss = 0.0
avg_accu_Train = 0.0
accu_Test_stats = []
net.train()
relationNet.train()
# epoch training list
trainList_combo = Producer(trainList, args.way_train, args.num_episode, "training") # combo contains [query_label, query_path ]
list_trainset = tnt.dataset.ListDataset(trainList_combo, loadImg)
trainloader = list_trainset.parallel(batch_size=args.batch_size, num_workers=args.num_workers, shuffle=True)
for i, data in enumerate(tqdm(trainloader), 0):
#for i, data in enumerate(trainloader, 0):
# get inputs
batchSize = data[0].size()[0]
labels = torch.unsqueeze(data[0], 1)
images = data[1:]
images_all = torch.cat(images).permute(0, 3, 1, 2).float()
labels_one_hot = torch.zeros(data[0].size()[0], args.way_train)
labels_one_hot.scatter_(1, labels, 1.0)
# wrap in Variable
if torch.cuda.is_available():
images_all, labels_one_hot = Variable(images_all.cuda()), Variable(labels_one_hot.cuda())
else:
images_all, labels_one_hot = Variable(images_all), Variable(labels_one_hot)
# zero gradients
optimizer.zero_grad()
# forward + backward + optimizer
feature_s_all_t0_p = net(images_all)
feature_s_all_t0_p = torch.split(feature_s_all_t0_p, batchSize, 0)
concatenatedFeat_list = [[] for _ in range(args.way_train)]
for idx in range(args.way_train):
concatenatedFeat_list[idx] = torch.cat((feature_s_all_t0_p[idx], feature_s_all_t0_p[-1]), 1)
concatenatedFeat_all = torch.cat(concatenatedFeat_list, 0)
relationScore_all = relationNet(concatenatedFeat_all)
relationScore_list = torch.split(relationScore_all, batchSize, 0)
relationScore = torch.cat(relationScore_list, 1)
#loss = criterion(relationScore, labels_one_hot)
weights = labels_one_hot.clone()
weights[labels_one_hot == 0] = 1.0/(args.way_train)
weights[labels_one_hot != 0] = (args.way_train-1.0)/(args.way_train)
loss = weighted_mse_loss(relationScore, labels_one_hot, weights)/data[0].size()[0]
loss.backward()
optimizer.step()
# summing up
running_loss += loss.data[0]
_, predicted = torch.max(relationScore.data, 1)
labels = torch.squeeze(labels, 1)
avg_accu_Train += (predicted == labels.cuda()).sum()
if i % args.log_step == args.log_step-1:
#print('[%d, %5d] train loss: %.3f train accuracy: %.3f' % (epoch + 1, i + 1, running_loss / args.log_step, avg_accu_Train/(args.log_step*batchSize)))
if args.logport:
args.logport.appendlog(running_loss / args.log_step, 'Training Loss')
args.logport.appendlog(avg_accu_Train/(args.log_step*batchSize), 'Training Accuracy')
args.logport.image((images[-1][0, :, :, :]).permute(2, 0, 1), 'query img', mode='img')
for idx in range(args.way_train):
args.logport.image((images[idx][0, :, :, :]).permute(2, 0, 1), 'support img', mode='img')
running_loss = 0.0
avg_accu_Train = 0.0
if (i+1) % args.save_step == 0:
torch.save(net.state_dict(),
os.path.join(args.model_path,
'net-model-%d-%d.pkl' %(epoch+1, i+1)))
torch.save(relationNet.state_dict(),
os.path.join(args.model_path,
'relationNet-model-%d-%d.pkl' %(epoch+1, i+1)))
net.eval()
relationNet.eval()
# epoch training list
testList_combo = Producer(testList, args.way_test, args.num_episode_test, "testing") # combo contains [query_label, query_path ]
list_testset = tnt.dataset.ListDataset(testList_combo, loadImg_testing)
testloader = list_testset.parallel(batch_size=args.batch_size_test, num_workers=args.num_workers, shuffle=False)
#for i, data in enumerate(tqdm(testloader), 0):
for i, data in enumerate(testloader, 0):
# get inputs
batchSize = data[0].size()[0]
labels = torch.unsqueeze(data[0], 1)
images = data[1:]
images_all = torch.cat(images).permute(0, 3, 1, 2).float()
labels_one_hot = torch.zeros(batchSize, args.way_test)
labels_one_hot.scatter_(1, labels, 1.0)
# wrap in Variable
if torch.cuda.is_available():
images_all, labels_one_hot = Variable(images_all.cuda(), volatile=True), Variable(labels_one_hot.cuda(), volatile=True)
else:
images_all, labels_one_hot = Variable(images_all, volatile=True), Variable(labels_one_hot, volatile=True)
# forward
feature_s_all_t0_p = net(images_all)
feature_s_all_t0_p = torch.split(feature_s_all_t0_p, batchSize, 0)
concatenatedFeat_list = [[] for _ in range(args.way_test)]
for idx in range(args.way_test):
concatenatedFeat_list[idx] = torch.cat((feature_s_all_t0_p[idx], feature_s_all_t0_p[-1]), 1)
concatenatedFeat_all = torch.cat(concatenatedFeat_list, 0)
relationScore_all = relationNet(concatenatedFeat_all)
relationScore_list = torch.split(relationScore_all, batchSize, 0)
relationScore = torch.cat(relationScore_list, 1)
_, predicted = torch.max(relationScore.data, 1)
#avg_accu_Test += (predicted == torch.squeeze(labels, 1).cuda()).sum()
accu_Test_stats.append((predicted == torch.squeeze(labels, 1).cuda()).sum()/float(batchSize))
m, h = mean_confidence_interval(np.asarray(accu_Test_stats), confidence=0.95)
print('[epoch %3d] test accuracy with 0.95 confidence: %.4f, +-: %.4f' % (epoch + 1, m, h))
#avg_accu_Test = 0.0
accu_Test_stats = []
result = sess.run(accuracy, feed_dict={img_input: x_data})
return result
mnist = input_data.read_data_sets("MNIST_data", one_hot="true")
slim = tf.contrib.slim
# img_input的placeholder
img_input = tf.placeholder(tf.float32, shape=(None, 784))
img_reshape = tf.reshape(img_input, shape=(-1, 28, 28, 1))
# 载入模型
sess = tf.Session()
Conv_Net = Net.Conv_Net()
prediction = Conv_Net.net(img_reshape)
# 载入模型
ckpt_filename = './logs/model.ckpt-20000'
# 初始化所有变量
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
# 恢复
saver.restore(sess, ckpt_filename)
print(compute_accuracy(mnist.test.images, mnist.test.labels))
def __init__(self,
# Genetic parameters
breeding_algorithm=DEFAULT_BREEDING_ALGORITHM,
generations=DEFAULT_GENERATIONS,
generation_cutoff=DEFAULT_GENERATION_CUTOFF,
generation_size=DEFAULT_GENERATION_SIZE,
mutation_chance=DEFAULT_MUTATION_CHANCE,
mutation_difference=DEFAULT_MUTATION_DIFFERENCE,
# Model parameters
iterations_per_model=DEFAULT_MODEL_ITERATIONS,
# Epoch parameters
epoch_num=DEFAULT_EPOCH_NUM,
epoch_size=DEFAULT_EPOCH_SIZE,
experiment_name=DEFAULT_EXPERIMENT_NAME,
**kwargs):
if (not isinstance(generations, int) or
not isinstance(generation_size, int) or
generations < 1 or generation_size < 1):
raise ValueError('Generations and generation size must be positive'
' integers greater than zero')
if (not isinstance(generation_cutoff, float) or
generation_cutoff <= 0.0 or generation_cutoff >= 1.0):
raise ValueError('Generation cutoff must be a positive float '
' between 0.0 and 1.0 (exclusive)')
if (not isinstance(breeding_algorithm, basestring) or
breeding_algorithm not in self.BREEDING_ALGORITHM_CHOICES):
raise ValueError('Unknown net breeding algorithm: {}'.format(
breeding_algorithm))
if (not isinstance(mutation_chance, float) or
not isinstance(mutation_difference, float) or
mutation_chance < 0.0 or mutation_chance > 1.0 or
mutation_difference < 0.0 or mutation_difference > 1.0):
raise ValueError('Mutation chance, and mutation difference must be'
' positive floats between 0.0 and 1.0'
' (inclusive)')
if (not isinstance(iterations_per_model, int) or
iterations_per_model <= 0):
raise ValueError('Number of iterations per model must be a positive'
' integer')
if (not isinstance(epoch_size, int) or not isinstance(epoch_num, int)
or epoch_size <= 0 or epoch_num < 0):
raise ValueError('Epoch size must be an integer greater than zero'
' and epoch num must be a non-negative integer')
if not experiment_name or not isinstance(experiment_name, basestring):
raise ValueError('Epoch name must be a non-empty string')
# Nets
net_hidden_sizes = kwargs.pop('hidden_sizes', Net.DEFAULT_HIDDEN_SIZES)
net_weights = kwargs.pop('weights', Net.DEFAULT_WEIGHTS)
net_inputs = kwargs.pop('inputs', Net.DEFAULT_INPUTS)
net_outputs = kwargs.pop('outputs', Net.DEFAULT_OUTPUTS)
net_spread = kwargs.pop('weight_spread', Net.DEFAULT_WEIGHT_SPREAD)
net_middle = kwargs.pop('weight_middle', Net.DEFAULT_WEIGHT_MIDDLE)
# Test to ensure nothing breaks
Net(hidden_sizes=net_hidden_sizes, weights=net_weights,
inputs=net_inputs, outputs=net_outputs,
weight_spread=net_spread, weight_middle=net_middle)
# Debug variables
self.debug = kwargs.pop('debug', None)
self.get_model_score = kwargs.pop('score_algorithm',
self.get_model_score)
# If kwargs not empty, extra key words were passed
if len(kwargs.keys()) > 0:
raise ValueError('Unnecessary kwargs passed to GeneticNetTrainer:'
' {}'.format(','.join(kwargs.keys())))
# Genetic variables
self.breeding_algorithm = breeding_algorithm
self.generations = generations
self.generation_size = generation_size
self.generation_cutoff = generation_cutoff
self.mutation_chance = mutation_chance
self.mutation_difference = mutation_difference
# Model variables
self.iterations_per_model = iterations_per_model
# Net variables
self.net_hidden_sizes = net_hidden_sizes
self.net_weights = net_weights
self.net_inputs = net_inputs
self.net_outputs = net_outputs
self.net_spread = net_spread
self.net_middle = net_middle
# Epoch variables
self.epoch_size = epoch_size
self.epoch_num = epoch_num
self.experiment_name = experiment_name
def test_nets_can_use_callable_weights(self):
n = Net(weights=Net.random_weights)
data = numpy.random.random((n.inputs, ))
output = n.run(data)
self.assertEqual(output.shape, (n.outputs, ))
def test_default_can_process_input(self):
n = Net()
data = numpy.random.random((n.inputs, ))
output = n.run(data)
self.assertEqual(output.shape, (n.outputs, ))
def empty_network():
return Net(D=1, H1=10, H2=20, class_count=10).cuda()
class Game(object):
def __init__(self,
net_number=None,
epoch_number=None,
experiment_name=None,
use_naive_opponent=False,
play_solo=False,
**kwargs):
self.board = BoardTTT(**kwargs)
self.opponent = None
if use_naive_opponent:
self.opponent = NaiveTTTPlayer()
elif (not play_solo and epoch_number is not None
and experiment_name is not None and net_number is not None):
from nets import Net
from trainer import GeneticNetTrainer
try:
net_data = GeneticNetTrainer.load_single_net(
idx=net_number,
epoch_num=epoch_number,
experiment_name=experiment_name)
# Hardcoding is hard
net_hidden_sizes = (3, 3)
# 2D Tic Tac Toe sizes
net_inputs = 9
net_outputs = 9
self.opponent = Net(hidden_sizes=net_hidden_sizes,
weights=net_data,
inputs=net_inputs,
outputs=net_outputs)
except IOError:
pass
def run(self):
last_move = None
last_last_move = None
player_to_move = 1
try:
while True:
self.board.display()
print("{} to move".format(
self.board.player_to_tile[player_to_move]))
if player_to_move == 2 and self.opponent:
# AI always see themselves as player one
data = self.board.normalize_board(player_to_move)
output = self.opponent.run(data)
print("{} to take {}".format(
self.board.player_to_tile[player_to_move],
output.argmax()))
if self.board.loop(output.argmax(),
player_to_move,
suppress_invalid=True):
# 2 player game
player_to_move = (player_to_move % 2) + 1
else:
print(
"The AI made an invalid move and probably can't"
" find a valid move. Sorry. :(")
self.board.game_over(invalid_move=True)
continue
s = raw_input().strip().lower()
is_number = False
temp_move = -1
try:
temp_move = int(s)
is_number = True
except ValueError:
pass
print("")
if s.startswith('m'):
print("Moves: {}".format(self.board.valid_moves))
continue
elif s.startswith('u'):
if self.board.undo():
last_move = last_last_move
continue
elif s.startswith('q'):
self.board.game_over()
elif (s.startswith('h') or not is_number):
print("Commands:")
print("- [H]elp (this menu)")
print("- [M]oves (# of moves)")
print("- [U]ndo (revert last move)")
print("- [Q]uit (leave game)")
print("- Type a number to place your symbol there")
continue
if self.board.loop(temp_move, player_to_move):
last_last_move = last_move
last_move = temp_move
player_to_move = (player_to_move % 2) + 1
except GameOver as go:
self.board.display()
print("Game over! {} wins".format(
self.board.player_to_tile.get(int(str(go)), "Nobody")))
def main(args):
'''
main function
'''
if args.logport:
args.logport = Dashboard(args.logport, 'dashboard')
EPOCH_SIZE = args.num_episode * args.num_query * args.way_train
EPOCH_SIZE_TEST = args.num_episode_test * args.num_query * args.way_test
SM_CONSTANT = 50
'''define network'''
net = Net(args.num_in_channel, args.num_filter)
if torch.cuda.is_available():
net.cuda()
'''
load model if needed
'''
if args.model_load_path_net != '':
net.load_state_dict(torch.load(args.model_load_path_net))
net.cuda()
print('model loaded')
''' define loss, optimizer'''
criterion = nn.CrossEntropyLoss()
params = list(net.parameters())
optimizer = optim.Adam(params, lr=args.learning_rate)
'''get data'''
trainList = read_miniImageNet_pathonly(
TESTMODE=False,
miniImageNetPath='/media/fujenchu/data/miniImageNet_Ravi/',
imgPerCls=600)
testList = read_miniImageNet_pathonly(
TESTMODE=True,
miniImageNetPath='/media/fujenchu/data/miniImageNet_Ravi/',
imgPerCls=600)
scheduler = StepLR(optimizer, step_size=40, gamma=0.5)
''' training'''
for epoch in range(1000):
scheduler.step()
running_loss = 0.0
avg_accu_Train = 0.0
accu_Test_stats = []
net.train()
# epoch training list
trainList_combo = Producer(
trainList, args.way_train, args.num_episode,
"training") # combo contains [query_label, query_path ]
list_trainset = tnt.dataset.ListDataset(trainList_combo, loadImg)
trainloader = list_trainset.parallel(batch_size=args.batch_size,
num_workers=args.num_workers,
shuffle=True)
for i, data in enumerate(tqdm(trainloader), 0):
#for i, data in enumerate(trainloader, 0):
# get inputs
batchSize = data[0].size()[0]
labels = torch.unsqueeze(data[0], 1)
images = data[1:]
images_all = torch.cat(images).permute(0, 3, 1, 2).float()
labels_one_hot = torch.zeros(data[0].size()[0], args.way_train)
labels_one_hot.scatter_(1, labels, 1.0)
# wrap in Variable
if torch.cuda.is_available():
images_all, labels = Variable(images_all.cuda()), Variable(
labels.cuda())
else:
images_all, labels = Variable(images_all), Variable(labels)
# zero gradients
optimizer.zero_grad()
# forward + backward + optimizer
feature_s_all_t0_p = net(images_all)
feature_s_all_t0_p = torch.split(feature_s_all_t0_p, batchSize, 0)
cosineDist_list = [[] for _ in range(args.way_train)]
for idx in range(args.way_train):
cosineDist_list[idx] = SM_CONSTANT * torch.sum(torch.mul(
feature_s_all_t0_p[-1].div(
torch.norm(
feature_s_all_t0_p[-1], p=2, dim=1,
keepdim=True).expand_as(feature_s_all_t0_p[-1])),
feature_s_all_t0_p[idx].div(
torch.norm(
feature_s_all_t0_p[idx], p=2, dim=1,
keepdim=True).expand_as(feature_s_all_t0_p[idx]))),
dim=1,
keepdim=True)
cosineDist_all = torch.cat(cosineDist_list, 1)
labels = labels.squeeze(1)
loss = criterion(cosineDist_all, labels)
loss.backward()
optimizer.step()
# summing up
running_loss += loss.data[0]
_, predicted = torch.max(cosineDist_all.data, 1)
avg_accu_Train += (predicted == labels.data).sum()
if i % args.log_step == args.log_step - 1:
#print('[%d, %5d] train loss: %.3f train accuracy: %.3f' % (epoch + 1, i + 1, running_loss / args.log_step, avg_accu_Train/(args.log_step*batchSize)))
if args.logport:
args.logport.appendlog(running_loss / args.log_step,
'Training Loss')
args.logport.appendlog(
avg_accu_Train / (args.log_step * batchSize),
'Training Accuracy')
args.logport.image(
(images[-1][0, :, :, :]).permute(2, 0, 1),
'query img',
mode='img')
for idx in range(args.way_train):
args.logport.image(
(images[idx][0, :, :, :]).permute(2, 0, 1),
'support img',
mode='img')
running_loss = 0.0
avg_accu_Train = 0.0
if (i + 1) % args.save_step == 0:
torch.save(
net.state_dict(),
os.path.join(args.model_path,
'net-model-%d-%d.pkl' % (epoch + 1, i + 1)))
net.eval()
# epoch training list
testList_combo = Producer(
testList, args.way_test, args.num_episode_test,
"testing") # combo contains [query_label, query_path ]
list_testset = tnt.dataset.ListDataset(testList_combo, loadImg_testing)
testloader = list_testset.parallel(batch_size=args.batch_size_test,
num_workers=args.num_workers,
shuffle=False)
#for i, data in enumerate(tqdm(testloader), 0):
for i, data in enumerate(testloader, 0):
# get inputs
batchSize = data[0].size()[0]
labels = torch.unsqueeze(data[0], 1)
images = data[1:]
images_all = torch.cat(images).permute(0, 3, 1, 2).float()
labels_one_hot = torch.zeros(batchSize, args.way_test)
labels_one_hot.scatter_(1, labels, 1.0)
# wrap in Variable
if torch.cuda.is_available():
images_all, labels = Variable(
images_all.cuda(), volatile=True), Variable(labels.cuda(),
volatile=True)
else:
images_all, labels = Variable(
images_all, volatile=True), Variable(labels, volatile=True)
# forward
feature_s_all_t0_p = net(images_all)
feature_s_all_t0_p = torch.split(feature_s_all_t0_p, batchSize, 0)
cosineDist_list = [[] for _ in range(args.way_train)]
for idx in range(args.way_train):
cosineDist_list[idx] = SM_CONSTANT * torch.sum(torch.mul(
feature_s_all_t0_p[-1].div(
torch.norm(
feature_s_all_t0_p[-1], p=2, dim=1,
keepdim=True).expand_as(feature_s_all_t0_p[-1])),
feature_s_all_t0_p[idx].div(
torch.norm(
feature_s_all_t0_p[idx], p=2, dim=1,
keepdim=True).expand_as(feature_s_all_t0_p[idx]))),
dim=1,
keepdim=True)
cosineDist_all = torch.cat(cosineDist_list, 1)
_, predicted = torch.max(cosineDist_all.data, 1)
accu_Test_stats.append(
(predicted == torch.squeeze(labels, 1).data.cuda()).sum() /
float(batchSize))
equality = (predicted != torch.squeeze(labels, 1).data.cuda())
equality_s = (predicted == torch.squeeze(labels, 1).data.cuda())
equality_idx = equality.nonzero()
equality_idx_s = equality_s.nonzero()
if i % args.log_step == args.log_step - 1:
if args.logport:
pred_np = predicted.cpu().numpy()
labels_np = labels.cpu().data.numpy()
batch_idx = equality_idx[0].cpu().numpy().astype(int)
bb = batch_idx[0]
args.logport.image(
(images[-1][bb, :, :, :]).permute(2, 0, 1),
np.array_str(labels_np[bb]) +
np.array_str(pred_np[bb]) + ' query img',
mode='img-test')
support_image = []
for idx in range(args.way_train):
support_image.append(images[idx][bb, :, :, :].permute(
2, 0, 1))
args.logport.image(torch.cat(support_image, 2),
'support img',
mode='img-test')
batch_idx = equality_idx_s[0].cpu().numpy().astype(int)
bb = batch_idx[0]
args.logport.image(
(images[-1][bb, :, :, :]).permute(2, 0, 1),
np.array_str(labels_np[bb]) +
np.array_str(pred_np[bb]) + ' query img',
mode='img-test')
support_image = []
for idx in range(args.way_train):
support_image.append(images[idx][bb, :, :, :].permute(
2, 0, 1))
args.logport.image(torch.cat(support_image, 2),
'support img',
mode='img-test')
m, h = mean_confidence_interval(np.asarray(accu_Test_stats),
confidence=0.95)
print(
'[epoch %3d] test accuracy with 0.95 confidence: %.4f, +-: %.4f' %
(epoch + 1, m, h))
#avg_accu_Test = 0.0
accu_Test_stats = []
class Game(object):
def __init__(
self,
net_number=None,
epoch_number=None,
experiment_name=None,
use_naive_opponent=False,
play_solo=False,
**kwargs):
self.board = BoardTTT(**kwargs)
self.opponent = None
if use_naive_opponent:
self.opponent = NaiveTTTPlayer()
elif (not play_solo
and epoch_number is not None
and experiment_name is not None
and net_number is not None):
from nets import Net
from trainer import GeneticNetTrainer
try:
net_data = GeneticNetTrainer.load_single_net(
idx=net_number,
epoch_num=epoch_number,
experiment_name=experiment_name)
# Hardcoding is hard
net_hidden_sizes = (3, 3)
# 2D Tic Tac Toe sizes
net_inputs = 9
net_outputs = 9
self.opponent = Net(
hidden_sizes=net_hidden_sizes,
weights=net_data,
inputs=net_inputs,
outputs=net_outputs)
except IOError:
pass
def run(self):
last_move = None
last_last_move = None
player_to_move = 1
try:
while True:
self.board.display()
print("{} to move".format(
self.board.player_to_tile[player_to_move]))
if player_to_move == 2 and self.opponent:
# AI always see themselves as player one
data = self.board.normalize_board(player_to_move)
output = self.opponent.run(data)
print("{} to take {}".format(
self.board.player_to_tile[player_to_move],
output.argmax()))
if self.board.loop(
output.argmax(),
player_to_move,
suppress_invalid=True):
# 2 player game
player_to_move = (player_to_move % 2) + 1
else:
print("The AI made an invalid move and probably can't"
" find a valid move. Sorry. :(")
self.board.game_over(invalid_move=True)
continue
s = raw_input().strip().lower()
is_number = False
temp_move = -1
try:
temp_move = int(s)
is_number = True
except ValueError:
pass
print("")
if s.startswith('m'):
print("Moves: {}".format(self.board.valid_moves))
continue
elif s.startswith('u'):
if self.board.undo():
last_move = last_last_move
continue
elif s.startswith('q'):
self.board.game_over()
elif (s.startswith('h')
or not is_number):
print("Commands:")
print("- [H]elp (this menu)")
print("- [M]oves (# of moves)")
print("- [U]ndo (revert last move)")
print("- [Q]uit (leave game)")
print("- Type a number to place your symbol there")
continue
if self.board.loop(temp_move, player_to_move):
last_last_move = last_move
last_move = temp_move
player_to_move = (player_to_move % 2) + 1
except GameOver as go:
self.board.display()
print("Game over! {} wins".format(
self.board.player_to_tile.get(int(str(go)),
"Nobody")))