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()
yhat = model(x)
loss = criterion(yhat, y)
loss.backward()
optimizer.step()
running_loss += loss.item()
if i % 200 == 199: # print every 2000 mini-batches
print('[%d, %5d] loss: %.3f' %
(epoch + 1, i + 1, running_loss / 200))
running_loss = 0.0
#print(running_loss)
import matplotlib.pyplot as plt
plt.plot(loss_list)
plt.title('Model_loss vs epochs')
plt.ylabel('Loss')
plt.xlabel('epochs')
s = '../working/epochwise_loss_' + str(epochs)
plt.savefig(s)
plt.show()
plt.close()
torch.save(model.state_dict(), '../working/new_mod.pth')
PATH = '/home/prathmesh/Desktop/Mesh_Neural_Network/new_models.pth'
torch.save(model.state_dict(), PATH)
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()
optimizer.step()
# print statistics
print('[%d, %5d/%d] loss: %.3f' %
(epoch + 1, i + 1, len(train_load), loss.item()))
print('Finished Training')
torch.save(net.state_dict(), './cnn')
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
correct += (pred_t == labels_t).sum().item()
pred_temp.append(torch.cat(pt))
true_temp.append(torch.cat(tt))
test_acc = correct / total
test_acc_list.append(test_acc)
print('test Acc {}:'.format(test_acc))
best_result_index = np.argmax(np.array(test_acc_list))
loss_total.append(loss_list[best_result_index])
acc_total.append(test_acc_list[best_result_index])
pred_total.append(pred_temp[best_result_index].tolist())
true_total.append(true_temp[best_result_index].tolist())
file_name = 'res/2gram MCSC'
torch.save(net.state_dict(), file_name + '.pth')
loss_DF = pd.DataFrame(loss_total)
loss_DF.to_csv(file_name + " loss.csv")
acc_DF = pd.DataFrame(acc_total)
acc_DF.to_csv(file_name + " acc.csv")
pred_DF = pd.DataFrame(pred_total)
pred_DF.to_csv(file_name + " pred.csv")
true_DF = pd.DataFrame(true_total)
true_DF.to_csv(file_name + " true.csv")
running_loss = 0.0
st2 = time.time()
for i in range(len(DataObject_X)):
#x,y = DataObject[i]
x, y = torch.from_numpy(DataObject_X[i]), torch.tensor(
[DataObject_Y[i]])
if torch.cuda.is_available():
x = x.cuda()
y = y.cuda()
#print(y)
x = x.float()
#y = y.float()
optimizer.zero_grad()
yhat = model(x)
loss = criterion(yhat, y)
loss.backward()
optimizer.step()
running_loss += loss.item()
if i % 200 == 199: # print every 2000 mini-batches
print('[%d, %5d] loss: %.3f' %
(epoch + 1, i + 1, running_loss / 200))
running_loss = 0.0
st3 = time.time()
print(st3 - st2)
print(epochs, running_loss)
PATH = '/home/prathmesh/Desktop/Mesh_Neural_Network/new_models.pth'
torch.save(model.state_dict(), PATH)
roadImageNew1 = stacked_roadImage1.reshape(args.batch_size, 800*800)
roadImageNewf1 = roadImageNew1.type(torch.float)
sample1, road_image1 = Variable(sampleNew1).to(device), Variable(roadImageNewf1).to(device)
output1 = model(sample1)
loss_func = nn.BCELoss()
vloss = loss_func(output1, road_image1)
validation_loss += loss_func(output1, road_image1).data.item() # sum up batch loss
outputNew = output1 > 0.5
outputNew1 = outputNew.type(torch.float)
iou = compute_iou(outputNew1,road_image1)
total_iou += iou
validation_loss /= len(valLoader)
total_iou /= len(valLoader)
return validation_loss,total_iou
# In[ ]:
for epoch in range(0, args.epoch+1):
total_loss = train(epoch)
print("Epoch:" + str(epoch) +" " + "total training loss: " +str(total_loss) +" \n")
val_loss, total_iou = validation()
print("Epoch:" + str(epoch) +" " + "total validation loss: " +str(val_loss) +" total iou: " + str(total_iou) +"\n")
model_file = 'model5/resnet18_model2_' + str(epoch) + '.pth'
torch.save({'modelRoadMap_state_dict': model.state_dict()},model_file)
print('\nSaved model to ' + model_file )
inputs, labels = data[0].to(device), data[1].to(device)
outputs = model(inputs)
loss = criterion(outputs, labels)
loss_val.append(loss.item())
validation_loss_store.append([epoch, loss.item()])
val_loss += loss
val_loss = val_loss / float(i + 1)
if val_loss < min_loss:
min_loss = val_loss
no_impr_epoch = 0
#save the best model
torch.save(
model.state_dict(), 'weight/' + 'epoch_' + str(epoch + 1) +
'loss_' + str(float(val_loss.cpu().numpy())) + '.pt')
print('performance improved with validation loss ' +
str(float(val_loss.cpu().numpy())))
file.write(
'--------------------------------------------------------------------\n'
)
file.write('performance improved with validation loss = ' +
str(float(val_loss.cpu().numpy())) + '\n')
file.write('epoch = ' + str(epoch + 1) + '\t' + 'step = ' +
str(step) + '\t' + 'val_loss = ' + '\t' +
str(np.mean(loss_val)) + '\n')
file.write(
'--------------------------------------------------------------------\n\n'
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.item()
if i % 100 == 99: # print every 2000 mini-batches
print('[%d, %5d] loss: %.3f' %
(epoch + 1, i + 1, running_loss / 100))
running_loss = 0.0
i += 1
print(i)
print('Finished Training')
SAVE_PATH = './prototype_net'
torch.save(net.state_dict(), SAVE_PATH)
correct = 0
total = 0
with torch.no_grad():
for data in batch_feeder(X_val, y_val, batch_size):
images, labels = data
images = torch.from_numpy(images).float().to(device)
labels = torch.from_numpy(labels).long().to(device)
outputs = net(images)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
print('Accuracy of the network on validation images: %d %%' %
(100 * correct / total))
