I successfully implemented GCN, but loss was huge!!!! It is not a good idea unless you have very deep insight and know how to adjust the GCN framework for this compitition! Leave comments below if you have any idea
", "messages": [ { "id": 547938, "postDate": "2019-06-08T14:43:22.313Z", "content": "Eh, just because you had poor performance doesn't mean every one will. MPNN's are state-of-the-art when it comes to these kind of quantum property predictions. I got pretty decent performance with mine (~0.5 using the competition metric) however it still doesn't compare to the kernels currently available. Did you remember to encode which two atoms you want to target in the graph?
", "rawMarkdown": "Eh, just because you had poor performance doesn't mean every one will. MPNN's are state-of-the-art when it comes to these kind of quantum property predictions. I got pretty decent performance with mine (~0.5 using the competition metric) however it still doesn't compare to the kernels currently available. Did you remember to encode which two atoms you want to target in the graph?", "votes": 1, "replies": [ { "id": 547971, "postDate": "2019-06-08T15:19:54.050Z", "content": "I got ~0.5 too, for huge I meant that, MPNN is used for moleculer property prediction, which is graph level embedding. I am ccurious of ur implenment. If u get negative sore by GCN, please let me know.
", "rawMarkdown": "I got ~0.5 too, for huge I meant that, MPNN is used for moleculer property prediction, which is graph level embedding. I am ccurious of ur implenment. If u get negative sore by GCN, please let me know." }, { "id": 548024, "postDate": "2019-06-08T16:37:12.343Z", "content": "It was my first model, I haven't really had time to work on it any further. I basically just used the pre-built GCN from DeepChem, just added one additional feature which was the two target atoms. I imagine if you add more features like what others have shown in kernels that you can get a strong score. There seems to be a heavy bias in the atom indices, so including these alone should boost the model's performance.
", "rawMarkdown": "It was my first model, I haven't really had time to work on it any further. I basically just used the pre-built GCN from DeepChem, just added one additional feature which was the two target atoms. I imagine if you add more features like what others have shown in kernels that you can get a strong score. There seems to be a heavy bias in the atom indices, so including these alone should boost the model's performance." }, { "id": 548046, "postDate": "2019-06-08T17:05:48.903Z", "content": "Why not keep working on it? if it is that powerful as u said, I think u must give a shot and may top the leaderboard
", "rawMarkdown": "Why not keep working on it? if it is that powerful as u said, I think u must give a shot and may top the leaderboard" }, { "id": 548058, "postDate": "2019-06-08T17:24:16.377Z", "content": "Like I said, I haven't had much time, for this method or any other for that matter. And many other people are using GCNs, it's certainly not a novel approach.
", "rawMarkdown": "Like I said, I haven't had much time, for this method or any other for that matter. And many other people are using GCNs, it's certainly not a novel approach." }, { "id": 548210, "postDate": "2019-06-09T01:50:50.230Z", "content": "If u don't have much time, could u please publish ur kernel?
", "rawMarkdown": "If u don't have much time, could u please publish ur kernel?", "votes": -1 }, { "id": 548220, "postDate": "2019-06-09T02:14:15.360Z", "content": "I don't have a kernel, I ran my scripts on my computer and AWS to generate the features since it took a very long time and required more RAM than I have. I just took the Deepchem example for a GCN using the qm7 dataset here. Then made some minor modifications to data_loader.py to include the target atoms as additional features.
", "rawMarkdown": "I don't have a kernel, I ran my scripts on my computer and AWS to generate the features since it took a very long time and required more RAM than I have. I just took the Deepchem example for a GCN using the qm7 dataset [here](https://github.com/deepchem/deepchem/blob/master/examples/qm7/qm7_tensorgraph_GraphConv.py). Then made some minor modifications to [data_loader.py](https://github.com/deepchem/deepchem/blob/master/deepchem/data/data_loader.py) to include the target atoms as additional features." }, { "id": 548238, "postDate": "2019-06-09T02:44:43.043Z", "content": "Thanks a lot! I will check that out later
", "rawMarkdown": "Thanks a lot! I will check that out later" } ] }, { "id": 547666, "postDate": "2019-06-08T05:13:52.167Z", "content": "I successfully implemented GCN, but loss was huge!!!! It is not a good idea unless you have very deep insight and know how to adjust the GCN framework for this compitition! Leave comments below if you have any idea
", "rawMarkdown": "I successfully implemented GCN, but loss was huge!!!! It is not a good idea unless you have very deep insight and know how to adjust the GCN framework for this compitition! Leave comments below if you have any idea", "votes": -5 }, { "id": 548152, "postDate": "2019-06-08T21:53:56.547Z", "rawMarkdown": "", "isDeleted": true, "replies": [ { "id": 548211, "postDate": "2019-06-09T01:52:45.717Z", "content": "I use PyG to implement MPNN
", "rawMarkdown": "I use PyG to implement MPNN" } ] } ], "comments": [ { "id": 547938, "author_name": "Matt", "author_url": "", "post_date": "2019-06-08T14:43:22.313000", "content": "Eh, just because you had poor performance doesn't mean every one will. MPNN's are state-of-the-art when it comes to these kind of quantum property predictions. I got pretty decent performance with mine (~0.5 using the competition metric) however it still doesn't compare to the kernels currently available. Did you remember to encode which two atoms you want to target in the graph?
", "votes": 1, "replies": [ { "id": 547971, "author_name": "Milkshakespeare", "author_url": "", "post_date": "2019-06-08T15:19:54.050000", "content": "I got ~0.5 too, for huge I meant that, MPNN is used for moleculer property prediction, which is graph level embedding. I am ccurious of ur implenment. If u get negative sore by GCN, please let me know.
", "votes": 0, "replies": [] }, { "id": 548024, "author_name": "Matt", "author_url": "", "post_date": "2019-06-08T16:37:12.343000", "content": "It was my first model, I haven't really had time to work on it any further. I basically just used the pre-built GCN from DeepChem, just added one additional feature which was the two target atoms. I imagine if you add more features like what others have shown in kernels that you can get a strong score. There seems to be a heavy bias in the atom indices, so including these alone should boost the model's performance.
", "votes": 0, "replies": [] }, { "id": 548046, "author_name": "Milkshakespeare", "author_url": "", "post_date": "2019-06-08T17:05:48.903000", "content": "Why not keep working on it? if it is that powerful as u said, I think u must give a shot and may top the leaderboard
", "votes": 0, "replies": [] }, { "id": 548058, "author_name": "Matt", "author_url": "", "post_date": "2019-06-08T17:24:16.377000", "content": "Like I said, I haven't had much time, for this method or any other for that matter. And many other people are using GCNs, it's certainly not a novel approach.
", "votes": 0, "replies": [] }, { "id": 548210, "author_name": "Milkshakespeare", "author_url": "", "post_date": "2019-06-09T01:50:50.230000", "content": "If u don't have much time, could u please publish ur kernel?
", "votes": -1, "replies": [] }, { "id": 548220, "author_name": "Matt", "author_url": "", "post_date": "2019-06-09T02:14:15.360000", "content": "I don't have a kernel, I ran my scripts on my computer and AWS to generate the features since it took a very long time and required more RAM than I have. I just took the Deepchem example for a GCN using the qm7 dataset here. Then made some minor modifications to data_loader.py to include the target atoms as additional features.
", "votes": 0, "replies": [] }, { "id": 548238, "author_name": "Milkshakespeare", "author_url": "", "post_date": "2019-06-09T02:44:43.043000", "content": "Thanks a lot! I will check that out later
", "votes": 0, "replies": [] } ] }, { "id": 548152, "author_name": "", "author_url": "", "post_date": "2019-06-08T21:53:56.547000", "content": "", "votes": 0, "replies": [ { "id": 548211, "author_name": "Milkshakespeare", "author_url": "", "post_date": "2019-06-09T01:52:45.717000", "content": "I use PyG to implement MPNN
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "547938": "Eh, just because you had poor performance doesn't mean every one will. MPNN's are state-of-the-art when it comes to these kind of quantum property predictions. I got pretty decent performance with mine (~0.5 using the competition metric) however it still doesn't compare to the kernels currently available. Did you remember to encode which two atoms you want to target in the graph?", "547666": "I successfully implemented GCN, but loss was huge!!!! It is not a good idea unless you have very deep insight and know how to adjust the GCN framework for this compitition! Leave comments below if you have any idea", "548152": "" } }