{ "id": 398876, "title": "Leagrning GNN, how to define model using pytorch geometric?", "url": "/competitions/icecube-neutrinos-in-deep-ice/discussion/398876", "author_name": "", "post_date": "2023-04-01T07:44:46.695447Z", "votes": 9, "comment_count": 4, "views": 0, "content": "

Dear all, it is first time for me to use GNN. So to understand how it works, I am trying to create very simple own GNN using pytorch geometric with some of data from this competition.

\n

Now I prepare 10 graphs from the competition data like below.

\n

x: All sensors positions but masked if sensor does not appear in event.
\nedge_index: Use sensor ID and connecting based on time series.
\ny: simply use [azimuth, zenith]

\n

[Data(x=[5160, 3], edge_index=[2, 2967], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 36], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 58], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 998], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 32], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 55], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 62], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 86], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 66], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 36], y=[2])]

\n

Then define model like below and train it with above data. However the output of model has different shape with y 🤔

\n
\n

class GCN(torch.nn.Module):
\n def init(self):
\n super(GCN, self).init()
\n self.conv1 = GCNConv(3, 16)
\n self.conv2 = GCNConv(16, 32)
\n self.conv3 = GCNConv(32, 2)

\n
\n
def forward(self, x, edge_index):\n    x = F.relu(self.conv1(x, edge_index))\n    x = F.relu(self.conv2(x, edge_index))\n    x = self.conv3(x, edge_index)\n    return x\n
\n

Does anyone could give me an advice how should I modify it?
\nThe full code is in this notebook.
\nThank you in advance!🙏

", "messages": [ { "id": "2205045", "postDate": "04/01/2023 07:44:46", "content": "

Dear all, it is first time for me to use GNN. So to understand how it works, I am trying to create very simple own GNN using pytorch geometric with some of data from this competition.

\n

Now I prepare 10 graphs from the competition data like below.

\n

x: All sensors positions but masked if sensor does not appear in event.
\nedge_index: Use sensor ID and connecting based on time series.
\ny: simply use [azimuth, zenith]

\n

[Data(x=[5160, 3], edge_index=[2, 2967], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 36], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 58], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 998], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 32], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 55], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 62], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 86], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 66], y=[2]),
\n Data(x=[5160, 3], edge_index=[2, 36], y=[2])]

\n

Then define model like below and train it with above data. However the output of model has different shape with y 🤔

\n
\n

class GCN(torch.nn.Module):
\n def init(self):
\n super(GCN, self).init()
\n self.conv1 = GCNConv(3, 16)
\n self.conv2 = GCNConv(16, 32)
\n self.conv3 = GCNConv(32, 2)

\n
\n
def forward(self, x, edge_index):\n    x = F.relu(self.conv1(x, edge_index))\n    x = F.relu(self.conv2(x, edge_index))\n    x = self.conv3(x, edge_index)\n    return x\n
\n

Does anyone could give me an advice how should I modify it?
\nThe full code is in this notebook.
\nThank you in advance!🙏

", "rawMarkdown": "Dear all, it is first time for me to use GNN. So to understand how it works, I am trying to create very simple own GNN using pytorch geometric with some of data from this competition. \n\nNow I prepare 10 graphs from the competition data like below.\n\nx: All sensors positions but masked if sensor does not appear in event.\nedge_index: Use sensor ID and connecting based on time series.\ny: simply use [azimuth, zenith]\n\n[Data(x=[5160, 3], edge_index=[2, 2967], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 36], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 58], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 998], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 32], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 55], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 62], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 86], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 66], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 36], y=[2])]\n\nThen define model like below and train it with above data. However the output of model has different shape with y 🤔\n\n>class GCN(torch.nn.Module):\n def __init__(self):\n super(GCN, self).__init__()\n self.conv1 = GCNConv(3, 16)\n self.conv2 = GCNConv(16, 32)\n self.conv3 = GCNConv(32, 2)\n\n def forward(self, x, edge_index):\n x = F.relu(self.conv1(x, edge_index))\n x = F.relu(self.conv2(x, edge_index))\n x = self.conv3(x, edge_index)\n return x\n\nDoes anyone could give me an advice how should I modify it?\nThe full code is in [this notebook](https://www.kaggle.com/code/hechtjp/learning-gnn-with-icecube-dataset/notebook).\nThank you in advance!🙏", "votes": null }, { "id": "2205225", "postDate": "04/01/2023 10:58:34", "content": "

I am not good with coding GNN's but you can refer this playlist https://www.youtube.com/playlist?list=PLV8yxwGOxvvoNkzPfCx2i8an--Tkt7O8Z

", "rawMarkdown": "I am not good with coding GNN's but you can refer this playlist https://www.youtube.com/playlist?list=PLV8yxwGOxvvoNkzPfCx2i8an--Tkt7O8Z", "votes": null }, { "id": "2205237", "postDate": "04/01/2023 11:14:18", "content": "

Thanks @warcoder for your comment! I am checking it 👍

", "rawMarkdown": "Thanks @warcoder for your comment! I am checking it 👍", "votes": null }, { "id": "2221286", "postDate": "04/14/2023 05:46:22", "content": "

One possible solution is to add a global pooling layer after the last convolution layer, such as global_mean_pool or global_max_pool, which will aggregate the node features into a single vector. Then you can use a linear layer or a fully connected layer to map that vector to the output dimension of 2. Something like this:

\n
 (torch.nn.Module): \n (): \n    (GCN, self).init() \n    self.conv1 = GCNConv(, ) \n    self.conv2 = GCNConv(, ) \n    self.conv3 = GCNConv(, ) \n    self.pool = global_mean_pool \n    self.fc = torch.nn.Linear(, )\n\n (): \n    x = F.relu(self.conv1(x, edge_index)) \n    x = F.relu(self.conv2(x, edge_index)) \n    x = F.relu(self.conv3(x, edge_index)) \n    x = self.pool(x, edge_index) \n    x = self.fc(x) \n     x\n
", "rawMarkdown": "One possible solution is to add a global pooling layer after the last convolution layer, such as global_mean_pool or global_max_pool, which will aggregate the node features into a single vector. Then you can use a linear layer or a fully connected layer to map that vector to the output dimension of 2. Something like this:\n\n```python\nclass GCN(torch.nn.Module): \ndef init(self): \n super(GCN, self).init() \n self.conv1 = GCNConv(3, 16) \n self.conv2 = GCNConv(16, 32) \n self.conv3 = GCNConv(32, 64) \n self.pool = global_mean_pool # or global_max_pool \n self.fc = torch.nn.Linear(64, 2)\n\ndef forward(self, x, edge_index): \n x = F.relu(self.conv1(x, edge_index)) \n x = F.relu(self.conv2(x, edge_index)) \n x = F.relu(self.conv3(x, edge_index)) \n x = self.pool(x, edge_index) \n x = self.fc(x) \n return x\n```", "votes": null }, { "id": "2221746", "postDate": "04/14/2023 14:35:31", "content": "

@yus002 , thanks for your advice! I will try it out 😉

", "rawMarkdown": "yus002 , thanks for your advice! I will try it out 😉", "votes": null } ], "comments": [ { "id": 2205225, "author_name": "warcoder", "author_url": "", "post_date": "04/01/2023 10:58:34", "content": "

I am not good with coding GNN's but you can refer this playlist https://www.youtube.com/playlist?list=PLV8yxwGOxvvoNkzPfCx2i8an--Tkt7O8Z

", "votes": null, "replies": [ { "id": 2205237, "author_name": "hechtjp", "author_url": "", "post_date": "04/01/2023 11:14:18", "content": "

Thanks @warcoder for your comment! I am checking it 👍

", "votes": null, "replies": [] } ] }, { "id": 2221286, "author_name": "yus002", "author_url": "", "post_date": "04/14/2023 05:46:22", "content": "

One possible solution is to add a global pooling layer after the last convolution layer, such as global_mean_pool or global_max_pool, which will aggregate the node features into a single vector. Then you can use a linear layer or a fully connected layer to map that vector to the output dimension of 2. Something like this:

\n
 (torch.nn.Module): \n (): \n    (GCN, self).init() \n    self.conv1 = GCNConv(, ) \n    self.conv2 = GCNConv(, ) \n    self.conv3 = GCNConv(, ) \n    self.pool = global_mean_pool \n    self.fc = torch.nn.Linear(, )\n\n (): \n    x = F.relu(self.conv1(x, edge_index)) \n    x = F.relu(self.conv2(x, edge_index)) \n    x = F.relu(self.conv3(x, edge_index)) \n    x = self.pool(x, edge_index) \n    x = self.fc(x) \n     x\n
", "votes": null, "replies": [ { "id": 2221746, "author_name": "hechtjp", "author_url": "", "post_date": "04/14/2023 14:35:31", "content": "

@yus002 , thanks for your advice! I will try it out 😉

", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2205045": "Dear all, it is first time for me to use GNN. So to understand how it works, I am trying to create very simple own GNN using pytorch geometric with some of data from this competition. \n\nNow I prepare 10 graphs from the competition data like below.\n\nx: All sensors positions but masked if sensor does not appear in event.\nedge_index: Use sensor ID and connecting based on time series.\ny: simply use [azimuth, zenith]\n\n[Data(x=[5160, 3], edge_index=[2, 2967], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 36], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 58], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 998], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 32], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 55], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 62], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 86], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 66], y=[2]),\n Data(x=[5160, 3], edge_index=[2, 36], y=[2])]\n\nThen define model like below and train it with above data. However the output of model has different shape with y 🤔\n\n>class GCN(torch.nn.Module):\n def __init__(self):\n super(GCN, self).__init__()\n self.conv1 = GCNConv(3, 16)\n self.conv2 = GCNConv(16, 32)\n self.conv3 = GCNConv(32, 2)\n\n def forward(self, x, edge_index):\n x = F.relu(self.conv1(x, edge_index))\n x = F.relu(self.conv2(x, edge_index))\n x = self.conv3(x, edge_index)\n return x\n\nDoes anyone could give me an advice how should I modify it?\nThe full code is in [this notebook](https://www.kaggle.com/code/hechtjp/learning-gnn-with-icecube-dataset/notebook).\nThank you in advance!🙏", "2205225": "I am not good with coding GNN's but you can refer this playlist https://www.youtube.com/playlist?list=PLV8yxwGOxvvoNkzPfCx2i8an--Tkt7O8Z", "2205237": "Thanks @warcoder for your comment! I am checking it 👍", "2221286": "One possible solution is to add a global pooling layer after the last convolution layer, such as global_mean_pool or global_max_pool, which will aggregate the node features into a single vector. Then you can use a linear layer or a fully connected layer to map that vector to the output dimension of 2. Something like this:\n\n```python\nclass GCN(torch.nn.Module): \ndef init(self): \n super(GCN, self).init() \n self.conv1 = GCNConv(3, 16) \n self.conv2 = GCNConv(16, 32) \n self.conv3 = GCNConv(32, 64) \n self.pool = global_mean_pool # or global_max_pool \n self.fc = torch.nn.Linear(64, 2)\n\ndef forward(self, x, edge_index): \n x = F.relu(self.conv1(x, edge_index)) \n x = F.relu(self.conv2(x, edge_index)) \n x = F.relu(self.conv3(x, edge_index)) \n x = self.pool(x, edge_index) \n x = self.fc(x) \n return x\n```", "2221746": "yus002 , thanks for your advice! I will try it out 😉" }, "source": "meta" }