diff --git a/notebook/research/acl2021.ipynb b/notebook/research/acl2021.ipynb index e5f53c556..18aef2b99 100644 --- a/notebook/research/acl2021.ipynb +++ b/notebook/research/acl2021.ipynb @@ -774,7 +774,7 @@ " \"learning_rate\": {\"l\": 3e-5, \"u\": 1.5e-4, \"space\": \"log\"},\n", " \"warmup_ratio\": [0.1],\n", " \"num_train_epochs\": [3],\n", - " \"per_device_train_batch_size\": [16, 32, 64],\n", + " \"per_device_train_batch_size\": [32],\n", " \"weight_decay\": [0.0],\n", " \"attention_probs_dropout_prob\": [0.1],\n", " \"hidden_dropout_prob\": [0.1],\n", @@ -786,7 +786,7 @@ "cell_type": "markdown", "metadata": {}, "source": [ - "The validation F1/accuracy we got was 93.2/90.4, where the accuracy outperforms grid search. The test F1/accuracy was 90.9/87.4. As a result, random search with 4*GST and the minimal search space `hpo_space_min` has outperformed grid search. We stop the troubleshooting process." + "The validation F1/accuracy we got was 93.5/91.1, where the accuracy outperforms grid search. The test F1/accuracy was 90.1/86.1. As a result, random search with 4*GST and the minimum space overfits. We stop the troubleshooting process because the search space cannot be further reduced." ] } ],