# Error about hooks in implementing customized modules

Hi, I just implemented an customized modules as follows:

class Loss(torch.nn.Module):
’’‘
Implement the loss function from output from RNN.
Ref paper: https://arxiv.org/abs/1308.0850
’’‘
def init(self):
’’‘
x is sequence of coordinates with dim (batch, seq_length, 3).
Parameters are sequence of output from rnn with dim (batch, seq_length, 128).
’’'
self.e = [] # predicted end of stroke probability scalar
self.m1 = [] # vector of means for x1 with len 20
self.m2 = [] # vector of means for x2 with len 20
self.pi = [] # vector of mixture density network coefficients with len 20
self.rho = [] # vector of correlation with len 20
self.s1 = [] # vector of standard deviation for x1 with len 20
self.s2 = [] # vector of standard deviation for x2 with len 20
self.x1 = [] # x1 coordinate at t+1
self.x2 = [] # x2 coordinates at t + 1
self.et = [] # end of probability indicator from ground truth
self.batch = 0 # batch size
self.seq_length = 0 # reduce by 1 because loss is caculated at t+1 timestamp
self.parameters = []

``````def forward(self, x, para):
'''
Implement eq 26 of ref paper for each batch.
Input:
para: dim(seq_len, batch, 121)
x:    dim(seq_len, batch, 3)
'''
if x.size()[0] == para.size()[0]:
self.seq_length = x.size()[0] - 1
total_loss = 0
for i in range(self.seq_length):
# prepare parameters
self.__get_para(i, x, para)
normalpdf = self.__para2normal(self.x1, self.x2, self.m1, self.m2, self.s1, self.s2, self.rho) #dim (n_batch, 20)
single_loss = self.__singleLoss(normalpdf)
total_loss += single_loss
else:
raise Exception("x and para don't match")

def __get_para(self, i, x, para):
'''
Slice and process parameters to the right form.
Implementing eq 18-23 of ref paper.
'''
self.batch = x.size()[1]
self.e = torch.sigmoid(-para[i,:,0])  # eq 18
self.parameters = para

# slice remaining parameters and training inputs
self.pi, self.m1, self.m2, self.s1, self.s2, self.rho = torch.split(self.parameters[i,:,1:], 20, dim = 1) # dim(batch, 20)
self.x1 = x[i+1,:,0].resize(self.batch, 1)  # dim(batch, 1)
self.x2 = x[i+1,:,1].resize(self.batch, 1)
self.et = x[i+1,:,2].resize(self.batch, 1)

## process parameters
# pi
max_pi = torch.max(self.pi, dim = 1)[0]
max_pi = max_pi.expand_as(self.pi)
diff = self.pi - max_pi
red_sum = torch.sum(diff, dim = 1).expand_as(self.pi)
self.pi = diff.div(red_sum)

# sd
self.s1 = self.s1.exp()
self.s2 = self.s2.exp()

# rho
self.rho = self.rho.tanh()

# reshape ground truth x1, x2 to match m1, m2 because broadcasting is currently not supported by pytorch
self.x1 = self.x1.expand_as(self.m1)
self.x2 = self.x2.expand_as(self.m2)

def __para2normal(self, x1, x2, m1, m2, s1, s2, rho):
'''
Implement eq 24, 25 of ref paper.
All input with dim(1, batch, 20)
'''
norm1 = x1.sub(m1)
norm2 = x2.sub(m2)
s1s2 = torch.mul(s1, s2)
z = torch.pow(torch.div(norm1, s1), 2) + torch.pow(torch.div(norm2, s2), 2) - \
2*torch.div(torch.mul(rho, torch.mul(norm1, norm2)), s1s2)
negRho = 1 - torch.pow(rho, 2)
expPart = torch.exp(torch.div(-z, torch.mul(negRho, 2)))
coef = 2*np.pi*torch.mul(s1s2, torch.sqrt(negRho))
result = torch.div(expPart, coef)
return result

def __singleLoss(self, normalpdf):
'''
Calculate loss for single time stamp. eq 26
Input: normalpdf (1,n_batch, 20).
'''
epsilon = 1e-20  # floor of loss from mixture density component since initial loss could be zero
mix_den_loss = torch.mul(self.pi, normalpdf)
red_sum_loss = torch.sum(torch.log(mix_den_loss))  # sum for all batch
end_loss = torch.sum(torch.log(torch.mul(self.e, self.et) + torch.mul(1-self.e, 1 - self.et)))
total_loss = -red_sum_loss - end_loss

``````

when I call loss(x, para), the following error comes up:
‘Loss’ object has no attribute ‘_forward_hooks’

What does the error message imply ?
Specifically, what is hook used for ?
Did I get the error because I have no parameters in this module ?

Hi,

Firstly let me say it would be nice if You could format all of your code, including `init` method.
Also, it would be helpful if You could provide a working example of how to replicate the error (the code doesn’t run)

Nevertheless, your error is caused by the fact that you don’t call `Module` class `__init__`. Your code should be:

``````class Loss(torch.nn.Module):
def __init__(self):
super(torch.nn.Module, self).__init__()

... # rest of the __init__ here
``````
1 Like