Boost Model Performance with Only 5 Lines of Pseudo‑Label Code

Many algorithm engineers claim that with a few well‑labeled samples they can quickly solve a problem, but labeling data is costly and time‑consuming. In real scenarios, large amounts of unlabeled data are easy to obtain while labeled data are scarce, which motivates the use of semi‑supervised learning.

Semi‑supervised learning tackles situations where a small labeled set coexists with a large unlabeled set. A strong baseline within this field is pseudo‑label learning, which generates approximate (soft) labels for the unlabeled data and then incorporates them back into training.

Step‑by‑Step Pseudo‑Label Procedure

The process can be implemented with just five lines of code:

model1.fit(train_set, label, val=validation_set)  # step1
pseudo_label = model1.predict(test_set)        # step2
new_label = concat(pseudo_label, label)        # step3
new_train_set = concat(test_set, train_set)    # step4
model2.fit(new_train_set, new_label, val=validation_set)  # step5
final_predict = model2.predict(test_set)

The accompanying diagram (shown below) visualizes the data flow.

Detailed Steps

Step 1 & 2: Split the labeled data into train_set and validation_set, then train model1 on train_set.

Step 3: Use model1 to predict the unlabeled test_set, producing pseudo‑labels.

Step 4: Concatenate the original train_set with the pseudo‑labeled test_set to form new_train_set, and train a second model model2 on this combined set.

Step 5: Predict the final results on test_set with model2, while still evaluating performance on the untouched validation_set to avoid data leakage.

Important note: The validation_set must never be used during training; it serves solely for unbiased evaluation, preventing label leakage.