[Unreproducible Universal DA] Mismatch between paper and implementation

[NaJeehye]

https://github.com/mims-harvard/Raincoat/tree/main/trainers/trainer_uni.py

Paper Algorithm 1. Overview of RAINCOAT

[Stage 3 Inference]
c_list = self.learn_t(dis2proto_a, dis2proto_c) print(c_list) self.trg_true_labels = tar_uni_label_test acc, f1, H = self.detect_private(dis2proto_a_test, dis2proto_c_test, tar_uni_label_test, c_list)

(Mismatch 1)

Num 24
Paper: for c in C^{s} do.
Github: for c in { all possible labels at WISDM} do
def detect_private(self, d1, d2, tar_uni_label, c_list): diff = np.abs(d2-d1) for i in range(6): # <= not C^{s}

(Mismatch 2)

Num 27: CLUSTER(d^{ac}|y^{hat} = c) => CLUSTER is based on the predicted value.
In Paper, However, in github code, the ground truth value is used. (i.e. the function, learn_t, shouldn't be available at inference)

def learn_t(self,d1,d2): diff = np.abs(d2-d1) c_list= [] for i in range(6): cat = np.where(self.trg_train_dl.dataset.y_data==i) cc = diff[cat]

Question

While looking at the code, I found that the algorithm and the GitHub code in the paper do not match.
Therefore, we modified the mismatch based on the corresponding paper. However, we could not get the same H-Scoure in Paper Table 1.

How do I get the same experimental results of RAINCOAT in Universal DA?
I'd like to ask if there's any updated code.

[ayhamo]

Hello, any update on this?