Runtime error: when using torch.bernoulli

1

hello, i am new to pytorch and tried to implement a simple RBM following deeplearning.net. I get a simple implementation, but with runtime error when executing at torch.bernoulli function. I have no idea how it occur, and find no answer somewhere else. can someone help me out？

my code：

import os
import torch
from torch.autograd import Variable
import torch.nn.functional as nn
import torch.optim as optim
from torchvision import datasets, transforms
import matplotlib.pyplot as plt
import matplotlib.gridspec as gridspec

# hyper_params
cd_k = 5
lr = 0.001
batch_sz = 20
test_btsz = 16

nH = 512
nV = 28 * 28

v_bias = Variable(torch.zeros(nV), requires_grad=True)
h_bias = Variable(torch.zeros(nH), requires_grad=True)

W = Variable(torch.normal(torch.zeros(nV, nH), 0.01), requires_grad=True)

params = [v_bias, h_bias, W]
solver = optim.Adam(params, lr=lr)

# data sets, Minist 28x28 train:60000 test:10000
train_loader = torch.utils.data.DataLoader(
        datasets.MNIST('Data/data', train=True, download=True,
                       transform=transforms.Compose([
                               transforms.ToTensor()
                               ])),
        batch_size=batch_sz, shuffle=True)

test_loader = torch.utils.data.DataLoader(
        dataset=datasets.MNIST('Data/data', train=True, download=True,
                               transform=transforms.Compose([
                                       transforms.ToTensor()
                                       ])),
        batch_size=test_btsz, shuffle=True)

def xw_b(x, w, b):
    return x @ W + b.repeat(x.size(0), 1)

def free_enery(v):
    xwb = xw_b(v, W, h_bias)
    vbias_term = v @ v_bias
    hidden_term = torch.sum(torch.log(1 + torch.exp(xwb)), dim=1)
    return -hidden_term - vbias_term

def sample_vfh(h):
    wx_b = h @ W.t() + v_bias.repeat(h.size(0))
    prob = nn.sigmoid(wx_b)
    return torch.bernoulli(prob)

def sample_hfv(v):
    wx_b = v @ W + h_bias.repeat(v.size(0))
    prob = nn.sigmoid(wx_b)
    return torch.bernoulli(prob)

def gibbs_chain(x, cd_k):
    v_ = x
    for _ in range(cd_k):
        h_ = sample_hfv(v_)
        v_ = sample_vfh(h_)
    return v_

def train_rbm_by_batch(x):
    chain_end = gibbs_chain(x, cd_k)
    v_ = Variable(chain_end.data, requires_grad=False)
    loss = -torch.mean(free_enery(x) - free_enery(v_))
    solver.zero_grad()
    loss.backward()
    solver.step()
    return loss

def train_rbm():
    for epoch in range(100):
        loss, batch_idx = 0, 0
        for batch_idx, data in enumerate(train_loader, 0):
            inputs, _ = data
            inputs = Variable(inputs.view(batch_sz, 28*28))
            loss = train_rbm_by_batch(inputs)
        print('Epoch-{}; loss: {} '.format(epoch, loss.data.numpy()))

if __name__ == '__main__':
    train_rbm()

error information：

File “/Users/yx/Documents/ANN/rbm_pytorch.py”, line 79, in gibbs_chain
h_ = sample_hfv(v_)

File “/Users/yx/Documents/ANN/rbm_pytorch.py”, line 73, in sample_hfv
return torch.bernoulli(prob)

File “/Users/yx/anaconda/envs/yx/lib/python3.6/site-packages/torch/autograd/variable.py”, line 705, in bernoulli
return Bernoulli()(self)

File “/Users/yx/anaconda/envs/yx/lib/python3.6/site-packages/torch/autograd/stochastic_function.py”, line 23, in _do_forward
result = super(StochasticFunction, self)._do_forward(*inputs)

File "/Users/yx/anaconda/envs/yx/lib/python3.6/site-packages/torch/autograd/functions/stochastic.py", line 41, in forward
samples = probs.new().resize_as(probs).bernoulli_(probs)

RuntimeError: must be >= 0 and <= 1 at /Users/soumith/anaconda/conda-bld/pytorch-0.1.10_1488750756897/work/torch/lib/TH/THRandom.c:270

2

RuntimeError: must be >= 0 and <= 1

As seen, the function expects input >=0 and <=1 but you have not given that.