Thanks for hosting this fascinating competition, and congratulations to the winners! I really learned a lot through this competition.
\nTile Part:
\nMy approach to the Tile part was quite similar to the official code. I also check the public tile codes and find the validation score was already quite high , and ensembling them didn't lead to significant improvements, so I didn't invest too much time in this section.
Layout Part:
\nFor the Layout part, I primarily referred to the gst code https://github.com/kaidic/GST. However, I encountered some challenges in making it work efficiently due to GPU memory constraints. Eventually, I used a sampling method to reduce GPU usage.
Sampling Method:
\nThe raw data consumed too much memory, so I decided to sample only 500 or 1000 configurations for each sample. This significantly reduced training time and GPU memory requirements.
Training Strategy:
\nEnsembling
\nEnsembling above models proved to be effective in improving the results. Every time I trained a new model, I found that ensembling it with existing models contributed to score improvements.
Regarding the Public Score:
\nI did identify the fact mentioned in a post https://www.kaggle.com/competitions/predict-ai-model-runtime/discussion/456083 but couldn't fully understand why it worked. I exercised extreme caution in utilizing it during the private phase since I considered it very risky.
Thanks for hosting this fascinating competition, and congratulations to the winners! I really learned a lot through this competition.
\nTile Part:
\nMy approach to the Tile part was quite similar to the official code. I also check the public tile codes and find the validation score was already quite high , and ensembling them didn't lead to significant improvements, so I didn't invest too much time in this section.
Layout Part:
\nFor the Layout part, I primarily referred to the gst code https://github.com/kaidic/GST. However, I encountered some challenges in making it work efficiently due to GPU memory constraints. Eventually, I used a sampling method to reduce GPU usage.
Sampling Method:
\nThe raw data consumed too much memory, so I decided to sample only 500 or 1000 configurations for each sample. This significantly reduced training time and GPU memory requirements.
Training Strategy:
\nEnsembling
\nEnsembling above models proved to be effective in improving the results. Every time I trained a new model, I found that ensembling it with existing models contributed to score improvements.
Regarding the Public Score:
\nI did identify the fact mentioned in a post https://www.kaggle.com/competitions/predict-ai-model-runtime/discussion/456083 but couldn't fully understand why it worked. I exercised extreme caution in utilizing it during the private phase since I considered it very risky.
Congratulations and thanks for sharing your solution and method Berserker408!
", "rawMarkdown": "Congratulations and thanks for sharing your solution and method Berserker408!", "votes": null } ], "comments": [ { "id": 2533159, "author_name": "mpwolke", "author_url": "", "post_date": "11/21/2023 16:23:05", "content": "Congratulations and thanks for sharing your solution and method Berserker408!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2532640": "Thanks for hosting this fascinating competition, and congratulations to the winners! I really learned a lot through this competition.\n\n**Tile Part:**\nMy approach to the Tile part was quite similar to the official code. I also check the public tile codes and find the validation score was already quite high , and ensembling them didn't lead to significant improvements, so I didn't invest too much time in this section.\n\n**Layout Part:**\nFor the Layout part, I primarily referred to the gst code [https://github.com/kaidic/GST](url). However, I encountered some challenges in making it work efficiently due to GPU memory constraints. Eventually, I used a sampling method to reduce GPU usage. \n\n**Sampling Method:**\nThe raw data consumed too much memory, so I decided to sample only 500 or 1000 configurations for each sample. This significantly reduced training time and GPU memory requirements.\n\n**Training Strategy:**\n1. Training them all.\n2. Training seperately based on the edge & node shapes. While the test data didn't explicitly specify the model type, we could infer it based on the edge & node shapes.\n3. Using different kinds of parameters, such as graph conv type, learning rate, batchsize, layers number and hidden size. \n\n**Ensembling**\nEnsembling above models proved to be effective in improving the results. Every time I trained a new model, I found that ensembling it with existing models contributed to score improvements.\n\n**Regarding the Public Score:**\nI did identify the fact mentioned in a post [https://www.kaggle.com/competitions/predict-ai-model-runtime/discussion/456083](url) but couldn't fully understand why it worked. I exercised extreme caution in utilizing it during the private phase since I considered it very risky.", "2533159": "Congratulations and thanks for sharing your solution and method Berserker408!" }, "source": "meta" }