def __init__(self, parent=None):#параметры по умолчанию класса MainWindow
QtGui.QMainWindow.__init__(self, parent)
#self — описание окна, определение переменных виджетов
self.setGeometry(300, 300, 1500, 800)
mainWidget = QtGui.QWidget()
self.setCentralWidget(mainWidget)
#Настройка меню
menu = self.menuBar()
list1 = menu.addMenu('File')
list1.addAction('exit')
list1.addAction('save')
list1.addAction('save as')
#Настройка панелей инструментов
toolbar = self.addToolBar('Exit')
toolbar.addAction('exit')
toolbar.addAction('save')
toolbar.addAction('save as')
#определение переменных кнопок, редакторов, сеток.
self.rButton = QtGui.QPushButton("Right")
self.lButton = QtGui.QPushButton("Left")
self.dButton = QtGui.QPushButton("Download")
self.label1 = QtGui.QLabel()
self.label2 = QtGui.QLabel()
self.textEdit = QtGui.QTextBrowser(readOnly=1)
self.textEdit.setOpenExternalLinks(True)
grid = QtGui.QGridLayout(mainWidget)
#положение на сетке
#grid.setSpacing(100)
grid.setColumnMinimumWidth(0, 400)
grid.setColumnMinimumWidth(1, 400)
grid.setColumnMinimumWidth(2, 200)
grid.addWidget(self.label1, 0, 0)
grid.addWidget(self.label2, 0, 1)
grid.addWidget(self.lButton, 1, 0)
grid.addWidget(self.rButton, 1, 1)
grid.addWidget(self.dButton, 1, 2)
grid.addWidget(self.textEdit, 0, 2)
self.person=functions.unpickle()
#self.person=self.person[:10]
self.rButton.clicked.connect(self.rbutton)
self.lButton.clicked.connect(self.lbutton)
self.dButton.clicked.connect(self.dbutton)
#self.person[13].elo=200
self.update(2)
self.setGeometry(300, 300, 250, 150)
def dbutton(self):
functions.download()
self.person=functions.unpickle()
self.update(2)
def train_fc(learning_rate=1e-4,
num_epochs=50,
decay_every=50,
gamma=0.1,
batch_size=256,
num_batches=100,
display_info=True,
hidden_width=1536):
TRAIN_SET_SIZE = batch_size * num_batches
file_name = "data/cifar-10-batches-py/data_batch_1"
data = f.unpickle(file_name)
ims = f.open_data_ims(50000)
x, y = f.seperate_xy_ims(ims[0:TRAIN_SET_SIZE])
# figure our device here
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
if torch.cuda.is_available():
torch.set_default_tensor_type('torch.cuda.FloatTensor')
print("Using device:{}, cuda:{}, pytorch:{}".format(
device, torch.version.cuda, torch.__version__))
m = fc.FC_1(1536, hidden_width, 1536)
if torch.cuda.is_available():
m = m.cuda() # transfer model to cuda as needed
lossFunction = nn.L1Loss()
optimizer = torch.optim.Adam(m.parameters(), lr=learning_rate)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=decay_every,
gamma=gamma)
x = torch.tensor(x, dtype=torch.float)
y = torch.tensor(y, dtype=torch.float)
x, y = x / 256.0, y / 256.0 # regularize
# introduce a dataset class
train_dataset = datasets.VectorizedDataset(x, y)
train_loader = datasets.DataLoader(dataset=train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=0)
# remember loss graph
losses = np.zeros(num_epochs, dtype=np.float)
print("Started training.")
train_start_time = time.time()
for epoch in range(num_epochs):
if epoch != 0 and epoch % int(num_epochs / 10) == 0:
print("{}% done! loss:{}".format(epoch / num_epochs * 100.0,
losses[epoch - 1]))
epoch_losses = 0.0
# perform MINI-BATCH grad_descent here
for i, data in enumerate(train_loader, 0):
x_batch, y_batch = data
y_pred = m.forward(x_batch)
loss = lossFunction(y_pred, y_batch)
epoch_losses += loss
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses[epoch] = epoch_losses / num_batches
print("Training took {} seconds!".format(time.time() - train_start_time))
if display_info:
# try to write this first layer to a video
v = m.linear1.weight[:, 0:512].detach().cpu().numpy()
num_frames = v.shape[0]
v = v.reshape(num_frames, 16, 32)
v = np.absolute(v)
v = v * (1.0 / v.max()) * 255.0 # normailize 0->255
f.save_video(v)
# print loss graph
print(losses)
plt.plot(losses)
plt.show()
# print results
y_pred = m.forward(x)
x, y_pred = x.cpu().detach().numpy(), y_pred.detach().cpu().numpy()
# x, y_pred = np.asarray(x.data), np.asarray(y_pred.data)
x, y_pred = x * 256.0, y_pred * 256.0
x, y_pred = x[0:50], y_pred[0:50]
fuse_xy_pred = f.combine_xy_ims(x, y_pred)
big_im = f.comb_ims(fuse_xy_pred, 32, 32)
f.show_im_std(big_im)
# print image of the last FIRST layer weights
s = m.linear1.weight[0, 0:512].detach().cpu().numpy()
s = s.reshape(16, 32)
s = np.absolute(s)
s *= (1.0 / s.max()) # normalize 0->1
plt.imshow(s, interpolation='nearest')
plt.show()
return losses[num_epochs - 1]
