MMOE - Replicate the Original paper Chapter 3.2 (Synthetic Data)

[Steven-1124]

Hi there!

I try hard to replicate the original paper for MMOE Arch while have no idea to get the package run.

Same as the original paper, my data is Synthetic Data like x is a random vector, y1, and y2 are two task vectors generated based on x. I want to reshow Figure 4 in the authors's paper, thus having to calculate loss as well.

Below is how I generate the data:

`
def generate_orthogonal_vectors(d):
u1 = np.random.rand(d)
u1 = u1 / np.linalg.norm(u1)
u2 = np.random.rand(d)
u2 = u2 - np.dot(u2, u1) * u1
u2 = u2 / np.linalg.norm(u2)
return u1, u2

def generate_weight_vectors(u1, u2, c, p):
w1 = c * u1
w2 = c * p * u1 + np.sqrt(1 - p**2) * u2
return w1, w2

def generate_labels(data, w, alpha, beta, m, epsilon):
y = np.dot(data, w)
for i in range(m):
y += np.sin(alpha[i] * np.dot(data, w) + beta[i])
y += epsilon * np.random.randn(len(data))
return y
`

The example you posted is more complex data compared to mine and I do not have GPU device. Also, I did not find your main.py

Could you show me how to run the package in Jupiter?

Thanks so much!

[Baijiong-Lin]

We will add it in the future.