When evaluating the SDUD strategy on the Beauty dataset, the only difference from the experimental setup described in the paper is that I adopted multi-GPU training with gradient accumulation due to insufficient single-GPU memory.
Surprisingly, the metrics on the test set are inferior to TIGER with frozen tokenizer.
I wonder whether there are undocumented training or testing details that I might have overlooked.
My hardware and training configuration:
GPUs: 2 × RTX 3090 (24GB)
Batch size: 64 * 4（I set micro batch size to 64 with 4 gradient accumulation steps, and I have adjusted the learning rate and corresponding hyperparameters accordingly.）
When evaluating the SDUD strategy on the Beauty dataset, the only difference from the experimental setup described in the paper is that I adopted multi-GPU training with gradient accumulation due to insufficient single-GPU memory.
Surprisingly, the metrics on the test set are inferior to TIGER with frozen tokenizer.
I wonder whether there are undocumented training or testing details that I might have overlooked.
My hardware and training configuration:
GPUs: 2 × RTX 3090 (24GB)
Batch size: 64 * 4（I set micro batch size to 64 with 4 gradient accumulation steps, and I have adjusted the learning rate and corresponding hyperparameters accordingly.）