import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from Model import Net
model = Net(10) #10 classes
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.001, momentum=0.9)
from dataset import MeshData
DataObject = MeshData(
'/home/prathmesh/Desktop/SoC-2020/ModelNet10_stl/ModelNet10')
dataLoad = torch.utils.data.DataLoader(DataObject, batch_size=1, shuffle=True)
batch = next(iter(dataLoad))
print(len(batch))
max_epochs = 30
loss_list = []
for epochs in range(max_epochs):
#print('e =',epochs)
running_loss = 0.0
for i, data in enumerate(dataLoad, 0):
x, y = data
x = x[0].float().to(device)
#y = y.float()
y = y[0].to(device)
optimizer.zero_grad()
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
train_sampler = SubsetRandomSampler(train_indices)
valid_sampler = SubsetRandomSampler(val_indices)
train_load = torch.utils.data.DataLoader(dataset,
batch_size=batch_size,
sampler=train_sampler)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
net = Net()
net.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.01, momentum=0.7)
# Training
for epoch in range(5):
for i, data in enumerate(train_load, 0):
inputs, labels, name = data
inputs = inputs.float()
inputs, labels = inputs.to(device), labels.to(device)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
class Trainer:
def __init__(self, hidden_dim, buffer_size, gamma, batch_size, device, writer):
self.env = make("connectx", debug=True)
self.device = device
self.policy = Net(self.env.configuration.columns * self.env.configuration.rows, hidden_dim,
self.env.configuration.columns).to(
device)
self.target = Net(self.env.configuration.columns * self.env.configuration.rows, hidden_dim,
self.env.configuration.columns).to(
device)
self.enemyNet = Net(self.env.configuration.columns * self.env.configuration.rows, hidden_dim,
self.env.configuration.columns).to(
device)
self.target.load_state_dict(self.policy.state_dict())
self.target.eval()
self.buffer = ExperienceReplay(buffer_size)
self.enemy = "random"
self.trainingPair = self.env.train([None, self.enemy])
self.loss_function = nn.MSELoss()
self.optimizer = optim.Adam(params=self.policy.parameters(), lr=0.001)
self.gamma = gamma
self.batch_size = batch_size
self.first = True
self.player = 1
self.writer = writer
def agent(self, observation, configuration):
with torch.no_grad():
state = torch.tensor(observation['board'], dtype=torch.float)
reshaped = self.reshape(state)
action = self.takeAction(self.enemyNet(reshaped).view(-1), reshaped, 0, False)
return action
def switch(self):
self.trainingPair = self.env.train([None, "negamax"])
self.enemy = "negamax"
def switchPosition(self):
self.env.reset()
if self.first:
self.trainingPair = self.env.train([self.enemy, None])
self.player = 2
else:
self.trainingPair = self.env.train([None, self.enemy])
self.player = 1
self.first = not self.first
def load(self, path):
self.policy.load_state_dict(torch.load(path))
def synchronize(self):
self.target.load_state_dict(self.policy.state_dict())
def save(self, name):
torch.save(self.policy.state_dict(), name)
def reset(self):
self.env.reset()
return self.trainingPair.reset()
def step(self, action):
return self.trainingPair.step(action)
def addExperience(self, experience):
self.buffer.append(experience)
def epsilon(self, maxE, minE, episode, lastEpisode):
return (maxE - minE) * max((lastEpisode - episode) / lastEpisode, 0) + minE
def change_reward(self, reward, done):
if done and reward == 1:
return 10
if done and reward == -1:
return -10
if reward is None and done:
return -20
if done:
return 1
if reward == 0:
return 1 / 42
else:
return reward
def change_reward_streak(self, reward, done, reshapedBoard, action, useStreak):
if done and reward == 1:
return 20
if done and reward == -1:
return -20
if reward is None and done:
return -40
if done:
return 1
if reward == 0 & useStreak:
return 1 / 42 + self.streakReward(self.player, reshapedBoard, action)
if reward == 0:
return 1 / 42
else:
return reward
def streakReward(self, player, reshapedBoard, action):
verticalReward = 0
horizontalReward = 0
if self.longestVerticalStreak(player, reshapedBoard, action) == 3:
verticalReward = 3
if self.longestHorizontalStreak(player, reshapedBoard, action) == 3:
horizontalReward = 3
return verticalReward + horizontalReward + self.longestDiagonalStreak(player, reshapedBoard, action)
def longestVerticalStreak(self, player, reshapedBoard, action):
count = 0
wasZero = False
for i in range(5, 0, -1):
if reshapedBoard[0][player][i][action] == 0:
wasZero = True
if reshapedBoard[0][player][i][action] == 1 & wasZero:
count = 0
wasZero = False
count += reshapedBoard[0][player][i][action]
if reshapedBoard[0][0][0][action] == 0:
return 0
return count
def longestHorizontalStreak(self, player, reshapedBoard, action):
count = 0
rowOfAction = self.rowOfAction(player, reshapedBoard, action)
wasZero = False
for i in range(7):
if reshapedBoard[0][player][rowOfAction][i] == 0:
wasZero = True
if reshapedBoard[0][player][rowOfAction][i] == 1 & wasZero:
count = 0
wasZero = False
count += reshapedBoard[0][player][rowOfAction][i]
return count
def longestDiagonalStreak(self, player, reshapedBoard, action):
rowOfAction = self.rowOfAction(player, reshapedBoard, action)
for row in range(4):
for col in range(5):
if reshapedBoard[0][player][row][col] == reshapedBoard[0][player][row + 1][col + 1] == \
reshapedBoard[0][player][row + 2][col + 2] == 1 and self.actionInDiagonal1(action, row, col,
rowOfAction):
return 3
for row in range(5, 1, -1):
for col in range(4):
if reshapedBoard[0][player][row][col] == reshapedBoard[0][player][row - 1][col + 1] == \
reshapedBoard[0][player][row - 2][col + 2] == 1 and self.actionInDiagonal2(action, row, col,
rowOfAction):
return 3
return 0
def actionInDiagonal1(self, action, row, col, rowOfAction):
return (rowOfAction == row and action == col or
rowOfAction == row + 1 and action == col + 1 or
rowOfAction == row + 2 and action == col + 2)
def actionInDiagonal2(self, action, row, col, rowOfAction):
return (rowOfAction == row and action == col or
rowOfAction == row - 1 and action == col + 1 or
rowOfAction == row - 2 and action == col + 2)
def rowOfAction(self, player, reshapedBoard, action):
rowOfAction = 10
for i in range(6):
if reshapedBoard[0][player][i][action] == 1:
rowOfAction = min(i, rowOfAction)
return rowOfAction
def policyAction(self, board, episode, lastEpisode, minEp=0.1, maxEp=0.9):
reshaped = self.reshape(torch.tensor(board))
output = self.policy(reshaped).view(-1)
return self.takeAction(output, reshaped, self.epsilon(maxEp, minEp, episode, lastEpisode))
def takeAction(self, actionList: torch.tensor, board, epsilon, train=True):
if (np.random.random() < epsilon) & train:
# invalide actions rein=geht nicht
#return torch.tensor(np.random.choice(len(actionList))).item()
return np.random.choice([i for i in range(len(actionList)) if board[0][0][0][i] == 1])
else:
for i in range(7):
if board[0][0][0][i] == 0:
actionList[i] = float('-inf')
return torch.argmax(actionList).item()
def reshape(self, board: torch.tensor, unsqz=True):
tensor = board.view(-1, 7).long()
# [0] = wo kann er reinwerfen(da wo es geht, steht eine 1), [1] = player1 (da wo es geht steht eine 0), [2] = player2 (da wo es geht steht eine 0)
a = F.one_hot(tensor, 3).permute([2, 0, 1])
b = a[:, :, :]
if unsqz:
return torch.unsqueeze(b, 0).float().to(self.device)
return b.float().to(self.device)
def preprocessState(self, state):
state = self.reshape(torch.tensor(state), True)
return state
def trainActionFromPolicy(self, state, action):
state = self.preprocessState(state)
value = self.policy(state).view(-1).to(self.device)
return value[action].to(self.device)
def trainActionFromTarget(self, next_state, reward, done):
next_state = self.preprocessState(next_state)
target = self.target(next_state)
target = torch.max(target, 1)[0].item()
target = reward + ((self.gamma * target) * (1 - done))
return torch.tensor(target).to(self.device)
def train(self):
if len(self.buffer) > self.batch_size:
self.optimizer.zero_grad()
states, actions, rewards, next_states, dones = self.buffer.sample(self.batch_size, self.device)
meanLoss = 0
for i in range(self.batch_size):
value = self.trainActionFromPolicy(states[i], actions[i])
target = self.trainActionFromTarget(next_states[i], rewards[i], dones[i])
loss = self.loss_function(value, target)
loss.backward()
meanLoss += loss
self.optimizer.step()
return meanLoss / self.batch_size
preprcess.Save()
print('wordbag: ')
print(preprcess.wordbag)
print('label_list: ')
print(preprcess.label_list)
x = np.array(preprcess.wordbag)
y = np.array(preprcess.label_list)
x = torch.from_numpy(x).type(torch.FloatTensor)
y = torch.from_numpy(y)
x, y = Variable(x), Variable(y)
net = Net()
optimizer = torch.optim.ASGD(net.parameters(), lr=0.002)
criterion = torch.nn.CrossEntropyLoss()
for t in range(10000):
out = net(x)
loss = criterion(out, y)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if t % 100 == 0:
prediction = torch.max(F.softmax(out), 1)[1]
pred_y = prediction.data.numpy().squeeze()
target_y = y.data.numpy()
accuracy = sum(pred_y == target_y) / 652
print('------------------')
acc_total = []
pred_total = []
true_total = []
for i in tqdm(range(TOTAL)):
image_shape = full_dataset.x_data.shape[1:]
device = torch.device(CUDA_N if torch.cuda.is_available() else 'cpu')
torch.manual_seed(SEED[i])
net = Net(image_shape, NUM_CLASS)
net.to(device)
print(net)
softmax = nn.Softmax()
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=LR, momentum=0.1)
loss_list = []
train_acc_list = []
test_acc_list = []
pred_temp = []
true_temp = []
for epoch in range(EPOCH):
net.train()
running_loss = 0
total = train_size
correct = 0
for step, images_labels in enumerate(train_loader):
model = nn.DataParallel(model)
if args.reloadModel:
model_fp = "model4/resnet18_model2_17.pth"
model.load_state_dict(torch.load(model_fp)['modelRoadMap_state_dict'])
print("model_loaded")
model.to(device)
# In[ ]:
# Optimizer
optimizer = optim.Adam(model.parameters(), lr=args.lr)
# In[ ]:
# loss func
def modify_output_for_loss_fn(loss_fn, output, dim):
if loss_fn == "ce":
return output
if loss_fn == "mse":
return F.softmax(output, dim=dim)
if loss_fn == "nll":
return F.log_softmax(output, dim=dim)
if loss_fn in ["bce", "wbce", "wbce1"]:
from torch.utils.data import DataLoader
from torchvision import transforms
composed = transforms.Compose([Normalize(), ToTensor()])
trainset = ImagesDataset(csv_file=root_dir + 'train.csv',
root_dir=root_dir,
transform=composed)
trainloader = DataLoader(trainset, batch_size=4, shuffle=True, num_workers=4)
"""
Train data
"""
import torch.optim as optim
criterion = nn.MSELoss()
optimizer = optim.SGD(net.parameters(),
lr=0.001,
momentum=0.9,
weight_decay=1e-6)
for epoch in range(num_epoch):
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
inputs, labels = data['image'], data['label']
inputs, labels = inputs.to(device), labels.to(device)
optimizer.zero_grad()
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
def main():
global opt, model
opt = parser.parse_args()
print(opt)
cuda = opt.cuda
if cuda and not torch.cuda.is_available():
raise Exception("No GPU found, please run without --cuda")
opt.seed = random.randint(1, 10000)
print("Random Seed: ", opt.seed)
torch.manual_seed(opt.seed)
if cuda:
torch.cuda.manual_seed(opt.seed)
cudnn.benchmark = True
print("===> Loading datasets")
train_set = DatasetFromHdf5("data/train.h5")
training_data_loader = DataLoader(dataset=train_set,
num_workers=opt.threads,
batch_size=opt.batchSize,
shuffle=True)
print("===> Building model")
model = Net()
criterion = nn.MSELoss(size_average=False)
print("===> Setting GPU")
if cuda:
model = torch.nn.DataParallel(model).cuda()
criterion = criterion.cuda()
# optionally resume from a checkpoint
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
opt.start_epoch = checkpoint["epoch"] + 1
model.load_state_dict(checkpoint["model"].state_dict())
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
# optionally copy weights from a checkpoint
if opt.pretrained:
if os.path.isfile(opt.pretrained):
print("=> loading model '{}'".format(opt.pretrained))
weights = torch.load(opt.pretrained)
model.load_state_dict(weights['model'].state_dict())
else:
print("=> no model found at '{}'".format(opt.pretrained))
print("===> Setting Optimizer")
optimizer = optim.SGD(model.parameters(),
lr=opt.lr,
momentum=opt.momentum,
weight_decay=opt.weight_decay)
print("===> Training")
for epoch in range(opt.start_epoch, opt.nEpochs + 1):
train(training_data_loader, optimizer, model, criterion, epoch)
save_checkpoint(model, epoch)
