piEsposito
commented
2 years ago Hi. I'm gonna take a look at it and try to help you.
Em ter., 5 de out. de 2021 às 10:54, hpompom @.***> escreveu:
When I want to use blitz, I find it may not to backpropagation and tells me "forward() not find N". And I find it maybe error in "sample_elbo". I hope you can give me some help Error
TypeError Traceback (most recent call last) in ----> 1 bsde_solver.train(batch_size, N, itr, log = T) 2 #x,y = bsde_solver.model(batch_size, N)
in train(self, batch_size, N, itr, log) 94 dataloader_train = torch.utils.data.DataLoader(ds_train, batch_size=batch_size, shuffle=True) 95 for j, (datapoints, labels) in enumerate(dataloader_train): ---> 96 loss = self.model.sample_elbo(y, self.equation.g(x) ,criterion=criterion, 97 sample_nbr=10) 98
/opt/anaconda3/lib/python3.8/site-packages/blitz/utils/variational_estimator.py in sample_elbo(self, inputs, labels, criterion, sample_nbr, complexity_costweight) 63 loss = 0 64 for in range(sample_nbr): ---> 65 outputs = self(inputs) 66 loss += criterion(outputs, labels) 67 loss += self.nn_kl_divergence() * complexity_cost_weight
/opt/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, *kwargs) 1049 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks 1050 or _global_forward_hooks or _global_forward_pre_hooks): -> 1051 return forward_call(input, **kwargs) 1052 # Do not call functions when jit is used 1053 full_backward_hooks, non_full_backward_hooks = [], []
TypeError: forward() missing 1 required positional argument: 'N' code
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
class fbsde():
def __init__(self, x_0, b, sigma, f, g, T, dim_x,dim_y,dim_d): self.x_0 = x_0.to(device) self.b = b self.sigma = sigma self.f = f self.g = g self.T = T self.dim_x = dim_x self.dim_y = dim_y self.dim_d = dim_d@variational_estimator
class Model(nn.Module):
def __init__(self, equation, dim_h): super(Model, self).__init__() self.linear1 = BayesianLinear(equation.dim_x+1, dim_h)#x self.linear2 = BayesianLinear(dim_h, dim_h)#y self.linear3 = BayesianLinear(dim_h, equation.dim_y*equation.dim_d)#z self.y_0 = nn.Parameter(torch.rand(equation.dim_y, device=device))#y_0 self.equation = equation def forward(self,batch_size,N): def phi(x): x = F.relu(self.linear1(x))# x = F.relu(self.linear2(x))# return self.linear3(x).reshape(-1, self.equation.dim_y, self.equation.dim_d) delta_t = self.equation.T / N print(delta_t) W = torch.randn(batch_size, self.equation.dim_d, N, device=device) * np.sqrt(delta_t) x = self.equation.x_0+torch.zeros(W.size()[0],self.equation.dim_x,device=device) #x y = self.y_0+torch.zeros(W.size()[0],self.equation.dim_y,device=device) #y print(1) print(N) for i in range(N): u = torch.cat((x, torch.ones(x.size()[0], 1,device=device)*delta_t*i), 1) z = phi(u)# w = W[:, :, i].reshape(-1, self.equation.dim_d, 1) #pos = N[:, :, i].reshape(-1, self.equation.dim_d, 1) x = x+self.equation.b(delta_t*i, x, y)*delta_t+torch.matmul( self.equation.sigma(delta_t*i, x), w).reshape(-1, self.equation.dim_x) #x = x+self.equation.b(delta_t*i, x, y)*delta_t+torch.matmul( self.equation.sigma(delta_t*i, x), w).reshape(-1, self.equation.dim_x) + pos y = y-self.equation.f(delta_t*i, x, y, z)*delta_t + torch.matmul(z, w).reshape(-1, self.equation.dim_y) return x, y print(12)class BSDEsolver():
def __init__(self, equation, dim_h): self.model = Model(equation,dim_h).to(device) self.equation = equation def train(self, batch_size, N,itr, log): criterion = torch.nn.MSELoss().to(device)# #loss = self.model.sample_elbo(criterion=criterion, # sample_nbr=3)self.model.sample_elbo
optimizer = torch.optim.Adam(self.model.parameters())#weight_decay = 1e-2
loss_data, y0_data = [], [] print(2) for i in range(itr): print(3) x, y = self.model(batch_size,N) #loss = criterion(self.equation.g(x), y) ds_train = torch.utils.data.TensorDataset(y, self.equation.g(x)) dataloader_train = torch.utils.data.DataLoader(ds_train, batch_size=500, shuffle=True) for j, (datapoints, labels) in enumerate(dataloader_train): loss = self.model.sample_elbo(y, self.equation.g(x) ,criterion=criterion, sample_nbr=10) loss_data.append(float(loss)) y0_data.append(float(self.model.y_0)) optimizer.zero_grad() loss.backward() optimizer.step() #if(i % 100 == 0): #print("损失为{}".format(loss)) if log: np.save('loss_data', loss_data) np.save('y0_data', y0_data)import torch
import numpy as np
import os
os.environ['KMP_DUPLICATE_LIB_OK']='True'
dim_x, dim_y, dim_d, dim_h, N, itr, batch_size = 1, 1, 1, 11, 100, 3000, 1000
x_0, T = torch.ones(dim_x), 1
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def b(t, x, y):
return (1-x).reshape(batch_size, dim_x)def sigma(t, x):
return torch.sqrt(torch.abs(x)).reshape(batch_size, dim_x, dim_d)def f(t, x, y, z):
return (-y*x).reshape(batch_size, dim_y)def g(x):
return torch.ones(batch_size, dim_y,device=device)equation = fbsde(x_0, b, sigma, f, g, T,dim_x, dim_y, dim_d)
bsde_solver = BSDEsolver(equation, dim_h)
bsde_solver.train(batch_size, N, itr, log = T)
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <https://github.com/piEsposito/blitz-bayesian-deep-learning/issues/96>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ALLYRXV72HRZASJV3ULACMTUFL7QRANCNFSM5FL2UZDA> .
hpompom
When I want to use blitz, I find it may not to backpropagation and tells me "forward() not find N". And I find it maybe error in "sample_elbo". I hope you can give me some help
Error
TypeError Traceback (most recent call last)