Guys, I have had 2 models for this competition.
\nModel 1 has been trained for 70 epochs has a validation MSE loss of 0.33 and R2 score of 0.48.
\nModel 2 has been trained for 3 epochs, it has a validation MSE loss of 0.28 but a mere R2 score of 0.33.
\nDo you think this is due to less epochs trained? Do you think my R2 score of model 2 would increase if I trained for more epochs?
\n(PS - I did not try training model 2 for more epochs as it would take more than 10 hours of GPU)
", "messages": [ { "id": 2908945, "postDate": "2024-07-06T16:47:09.013Z", "content": "I'd review the individual R2 targets and ideally look at a plot. A lot of of us are needing to train for a long time either because the models are large or because it requires a lot of epochs to converge (or both). @jerrylin96 posted an excellent paper with strong pointers towards a decent U-net model along with training parameters. I modified a basic U-Net I already had with some of the key design decisions in the paper / NCSN origin last night and it is a lot stronger than my previous U-Net results.
", "rawMarkdown": "I'd review the individual R2 targets and ideally look at a plot. A lot of of us are needing to train for a long time either because the models are large or because it requires a lot of epochs to converge (or both). @jerrylin96 posted an excellent paper with strong pointers towards a decent U-net model along with training parameters. I modified a basic U-Net I already had with some of the key design decisions in the paper / NCSN origin last night and it is a lot stronger than my previous U-Net results. ", "replies": [ { "id": 2922382, "postDate": "2024-07-15T06:44:53.233Z", "content": "Hi @rob1080ti , which paper is this? Is this simply the ClimSim paper? or another paper? Thanks.
", "rawMarkdown": "Hi @rob1080ti , which paper is this? Is this simply the ClimSim paper? or another paper? Thanks.", "replies": [ { "id": 2922644, "postDate": "2024-07-15T11:08:32.077Z", "content": "Hi, it is posted in the above thread: https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/516605 . Note that someone posted their version of the U-Net described to the code section recently.
", "rawMarkdown": "Hi, it is posted in the above thread: https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/516605 . Note that someone posted their version of the U-Net described to the code section recently." } ] } ] }, { "id": 2908341, "postDate": "2024-07-06T08:59:34.543Z", "content": "Guys, I have had 2 models for this competition.
\nModel 1 has been trained for 70 epochs has a validation MSE loss of 0.33 and R2 score of 0.48.
\nModel 2 has been trained for 3 epochs, it has a validation MSE loss of 0.28 but a mere R2 score of 0.33.
\nDo you think this is due to less epochs trained? Do you think my R2 score of model 2 would increase if I trained for more epochs?
\n(PS - I did not try training model 2 for more epochs as it would take more than 10 hours of GPU)
", "rawMarkdown": "Guys, I have had 2 models for this competition. \n\nModel 1 has been trained for 70 epochs has a validation MSE loss of 0.33 and R2 score of 0.48. \n\nModel 2 has been trained for 3 epochs, it has a validation MSE loss of 0.28 but a mere R2 score of 0.33. \n\nDo you think this is due to less epochs trained? Do you think my R2 score of model 2 would increase if I trained for more epochs?\n\n(PS - I did not try training model 2 for more epochs as it would take more than 10 hours of GPU)" } ], "comments": [ { "id": 2908945, "author_name": "Rob Freeman", "author_url": "", "post_date": "2024-07-06T16:47:09.013000", "content": "I'd review the individual R2 targets and ideally look at a plot. A lot of of us are needing to train for a long time either because the models are large or because it requires a lot of epochs to converge (or both). @jerrylin96 posted an excellent paper with strong pointers towards a decent U-net model along with training parameters. I modified a basic U-Net I already had with some of the key design decisions in the paper / NCSN origin last night and it is a lot stronger than my previous U-Net results.
", "votes": 0, "replies": [ { "id": 2922382, "author_name": "Juan D C F", "author_url": "", "post_date": "2024-07-15T06:44:53.233000", "content": "Hi @rob1080ti , which paper is this? Is this simply the ClimSim paper? or another paper? Thanks.
", "votes": 0, "replies": [ { "id": 2922644, "author_name": "Rob Freeman", "author_url": "", "post_date": "2024-07-15T11:08:32.077000", "content": "Hi, it is posted in the above thread: https://www.kaggle.com/competitions/leap-atmospheric-physics-ai-climsim/discussion/516605 . Note that someone posted their version of the U-Net described to the code section recently.
", "votes": 0, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "2908945": "I'd review the individual R2 targets and ideally look at a plot. A lot of of us are needing to train for a long time either because the models are large or because it requires a lot of epochs to converge (or both). @jerrylin96 posted an excellent paper with strong pointers towards a decent U-net model along with training parameters. I modified a basic U-Net I already had with some of the key design decisions in the paper / NCSN origin last night and it is a lot stronger than my previous U-Net results. ", "2908341": "Guys, I have had 2 models for this competition. \n\nModel 1 has been trained for 70 epochs has a validation MSE loss of 0.33 and R2 score of 0.48. \n\nModel 2 has been trained for 3 epochs, it has a validation MSE loss of 0.28 but a mere R2 score of 0.33. \n\nDo you think this is due to less epochs trained? Do you think my R2 score of model 2 would increase if I trained for more epochs?\n\n(PS - I did not try training model 2 for more epochs as it would take more than 10 hours of GPU)" } }