zhiyongc / gru-d Goto Github PK

Hello,
first of all, thanks for this repository.
in the main function, what is the reason for assigning grid input size as:

input_dim  = fea_size
hidden_dim = fea_size
output_dim = fea_size

shouldn't it be like this??

input_dim  = fea_size
hidden_dim = seq_len
output_dim = pred_len

Hi,
My understanding of Equation 11 in the GRU-D paper is that the two x values on the right side of this equation are supposed to be different values. I believe mask * x to be correct, as that is multiplying x of the current time step. However, I believe delta_x * x is incorrect, as this x should be the value of x most recently observed prior to the current time step t.

Just pointing out that hidden is appended to the left of the outputs variable, which reverses the order (on the time dimension) of the hidden states of the GRU.

Based on my understanding of the paper, the computing of time interval at time step t is based on the mask value of the previous time step t-1 [eq(2) in the paper].

It is looks like you compute time interval at time step t based on mask value at time step t

for idx in range(missing_index[0].shape[0]):
            i = missing_index[0][idx] # for 1st dim
            j = missing_index[1][idx] # for 2nd dim
            k = missing_index[2][idx] # for 3rd dim
            if j != 0:
                Delta[i,j,k] = Delta[i,j,k] + Delta[i,j-1,k] 

But, I think it suppose to be as follows:

for idx in range(missing_index[0].shape[0]):
            i = missing_index[0][idx] # for 1st dim
            j = missing_index[1][idx] # for 2nd dim
            k = missing_index[2][idx] # for 3rd dim
            if j != 0 and j!=9 :
                Delta[i,j+1,k] = Delta[i,j+1,k] + Delta[i,j,k]

I think one update function is wrong.. on line 123 in GRUD.py your update equation is:

h_tilde = F.tanh(self.hl(combined))
h = (1 - z) * h + z * h_tilde

according to the paper, they should be

new_combined = torch.cat((x, torch.mul(r, h), m), dim=2)
h_tilde = F.tanh(self.hl(new_combined))
h = (1 - z)*h + z*h_tilde

When computing h_tilde in the GRU cell, you have to first multiply the hidden state with the reset gate.

Hi, thanks for your contributions. Could you tell where the data comes from? (speed_matrix_2015 and inrix_seattle_speed_matrix_2012)

Hi,

GRUD.py initializes the gamma_h_l by setting out_features = delta_size (line 108). It's ok in this code since the hidden_size is equal to input_size. However, should it be better to set out_features = hidden_size for other cases? In other words, I think it's better to use:

self.gamma_h_l = nn.Linear(self.delta_size, self.hidden_size)

Line 281 of main.py is

loss_L1 = torch.nn.MSELoss()

but I think it should be
loss_L1 = torch.nn.L1Loss()