Simple multimodal compact bilinear pooling

gistfile1.txt

import torch
import numpy as np
seed = 2
torch.manual_seed(seed)
np.random.seed(seed)

x_img = torch.tensor([8,1,3,0,2])
x_txt = torch.tensor([ 0,  7, 10,  1,  1])
out_dim = 3

def psi(x, out_dim):
    in_dim = x.size()
    y = torch.zeros(out_dim)
    #initialize h and s to be input size. h has indicies corresponding to output dim
    #e.g out_dim = 3 so h can look like: [1, 1, 0, 2, 1]
    h = torch.LongTensor(in_dim).random_(0, out_dim)
    #s is same size as input and we fill it with -1 and 1s
    s = 2*torch.LongTensor(in_dim).random_(0, 2)-1
    #Ψ function in paper
    for i in range(len(x)):
        j = h[i]
        y[j] += s[i]*x[i]
    return y

y_img = psi(x_img, out_dim)#v1'
y_txt = psi(x_txt, out_dim)#v2'
print(y_img)
print(y_txt)

y_img_prime = np.fft.fft(y_img)#FFT(v1')
y_txt_prime = np.fft.fft(y_txt)#FFT(v2')
phi = np.fft.ifft(np.multiply(y_img_prime,y_txt_prime))
print(phi)