У меня проблема с кодом ниже. Когда я запускаю код, я получаю значение для MSE Nan. Я проверил все промежуточные функции, они прекрасно работают. Я не знаю откуда Наны входят в сеть.
def train_model(self, X, Y, lr = 0.01):
print("No of Layers: ", self.no_of_layers)
self.intermediate = []
for epoch in range(100):
overall_loss = 0
print("Epoch:",epoch+1)
for i in range(len(X)):
self.layers = []
inp = X.iloc[i,:]
output = Y.iloc[i]
## PreProcessing for one input
a = torch.tensor(inp)
# a = torch.from_numpy(trimf(a,[0,3,10]))
a = a.to(device = torch.device('cuda'))
a = a.view(-1,1)
output = torch.tensor(output, device = self.device)
# print("Input Number:",i)
## Feed Forward
for layer in range(self.no_of_layers):
if layer == 0:
inpt = a
else:
inpt = self.layers[layer-1]
## Problem is with this inp. Layer 0 tensor of 1st exaple is passed as input to all the other examples for succesive layers.
# print("CL:", inpt)
Model.compute_layer(self,self.weights[layer],inpt,layer)
# print("Overall Layers",self.layers)
pred = self.layers[-1]
pred != pred
# print(pred)
pred = torch.sigmoid(pred)
## Loss Function
loss = (pred - output)**2
overall_loss +=torch.sum(loss)
## Back Propgation
loss.backward(torch.empty(loss.size(), device = self.device), retain_graph = True)
j = 0
for w in self.weights:
# print("BU", w.data)
w.data -= (lr*((w.grad-torch.min(w.grad))/(torch.max(w.grad)-torch.min(w.grad))))
w.grad != w.grad
# w.data -= (lr*w.grad)
# w.data != w.data
# print("AU",w.data)
overall_loss = overall_loss/len(X)
print("MSE: ", overall_loss.item())
Вывод:
