here is my python code implementing compute_weights function

mpl/_mpl.py

@@ -0,0 +1,44 @@
+import torch
+
+def compute_weights(losses, indices, weights, ratio, p):
+    size = losses.size(0)
+
+    # find first nonzero element
+    pos = 0
+    while losses[pos]< 1e-5:
+        pos += 1
+    n = size - pos
+    m = int(ratio * n)
+    if n <= 0 or m <= 0:
+        raise ValueError
+    q = p / (p - 1.0)
+    c = m - n + 1
+    a = [0.0 , 0.0]
+    i = pos
+    nu = 0.0
+    while i < n and nu < 1e-5:
+        loss_q = (losses[i] / losses[size - 1]) ** q
+        a[0] = a[1]
+        a[1] += loss_q
+        c += 1
+        nu = c * loss_q - a[1]
+
+    # compute alpha
+    if nu < 1e-5:
+        i += 1
+        c += 1
+        a[0] = a[1]
+    alpha = (a[0] / c) ** (1 / q) * losses[size - 1]
+
+    # compute_weights
+    tau = 1.0 / (n ** (1.0 / q)*(m **(1.0 / p)))
+    k = i
+    while k < n:
+        # maybe wrong
+        weights[indices[k]] = tau
+        k += 1
+    if alpha > -1e-5:
+        k = pos
+        while k < i:
+            weights[indices[k]] = tau * (losses[k] / alpha) ** (q - 1)
+            k += 1