![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2126876%2F4daeeaa70bf25eda3c4bb2519bcac346%2Fjk_solution_image.png?generation=1677991277683737&alt=media\")
I would like to thank the organizers for such an interesting competition!
\nWe share the Team JK's solution.
\nTeam Member: @vostankovich, @tereka, @anonamename, @yururoi, @tomo20180402
\n
I would like to thank the organizers for such an interesting competition!
\nWe share the Team JK's solution.
\nTeam Member: @vostankovich, @tereka, @anonamename, @yururoi, @tomo20180402
\n
I explain about Residual3DBlock.
\nIn this competition, 3d(t, h, w) is very important, so I decided to use any time analysis method.
\n15 frames reshape this ({-28, -24, -20}{-16, -12, -8}{-4, 0, 4}{8, 12, 16}{20, 24, 28}) and extract feature using backbone, then hidden output is applied it. here is a sample for using residual3d block(final architecture is very similar)
class Residual3DBlock(nn.Module):\n def __init__(self):\n super(Residual3DBlock, self).__init__()\n\n self.block = nn.Sequential(\n nn.Conv3d(512, 512, 3, stride=1, padding=1),\n nn.BatchNorm3d(512),\n nn.ReLU(512)\n )\n\n self.block2 = nn.Sequential(\n nn.Conv3d(512, 512, 3, stride=1, padding=1),\n nn.BatchNorm3d(512),\n )\n\n def forward(self, images):\n short_cut = images\n h = self.block(images)\n h = self.block2(h)\n\n return F.relu(h + short_cut)\n\nclass Model(nn.Module):\n def __init__(self):\n super(Model, self).__init__()\n self.backbone = timm.create_model(\"tf_efficientnet_b0_ns\", pretrained=True, num_classes=1, in_chans=3)\n self.mlp = nn.Sequential(\n nn.Linear(68, 256),\n nn.BatchNorm1d(256),\n nn.LeakyReLU(),\n nn.Dropout(0.2),\n )\n n_hidden = 1024\n\n self.conv_proj = nn.Sequential(\n nn.Conv2d(1280, 512, 1, stride=1),\n nn.BatchNorm2d(512),\n nn.ReLU(),\n )\n\n self.neck = nn.Sequential(\n nn.Linear(1024, 1024),\n nn.BatchNorm1d(1024),\n nn.LeakyReLU(),\n nn.Dropout(0.2),\n )\n\n self.triple_layer = nn.Sequential(\n Residual3DBlock(),\n )\n\n self.pool = GeM() \n\n self.fc = nn.Linear(256+1024, 1)\n\n def forward(self, images,feature, target=None, mixup_hidden = False, mixup_alpha = 0.1, layer_mix=None):\n b, t, h, w = images.shape\n images = images.view(b * t // 3, 3, h, w)\n feature_maps = self.conv_proj(self.backbone.forward_features(images))\n _, c, h, w = feature_maps.size()\n feature_maps = feature_maps.contiguous().view(b * 2,c,t // 2 // 3, h, w)\n feature_maps = self.triple_layer(feature_maps)\n middle_maps = feature_maps[:, :, 2, :, :]\n #b, h, w= middle_maps.size()\n #middle_maps = middle_maps.view(b, 1, h, w)\n nn_feature = self.neck(self.pool(middle_maps).reshape(b, -1))\n feature = self.mlp(feature)\n cat_features = torch.cat([nn_feature, feature], dim=1)\n if target is not None:\n cat_features, y_a, y_b, lam = mixup_data(cat_features, target, mixup_alpha)\n y = self.fc(cat_features)\n return y, y_a, y_b, lam\n else:\n y = self.fc(cat_features)\n return y\nI tried to run the code, but it seems to give me an error. What am I doing wrong?
\nmodel = Model()\nim = torch.randn((8, 15, 512, 512))\nfeature = torch.randn((8, 68))\nmodel(im, feature)\nOh, do you use the following because of the end and side?
\nim = torch.randn((8, 30, 512, 512))\nMay I ask a question? what is moving average? I have seen that it is mentioned several times
", "rawMarkdown": "May I ask a question? what is moving average? I have seen that it is mentioned several times", "votes": null }, { "id": "2172501", "postDate": "03/07/2023 15:04:23", "content": "yes, it's correct.
\nb, t(side + endzone), height, width.
\n", "rawMarkdown": "yes, it's correct.\nb, t(side + endzone), height, width.\n\n>im = torch.randn((8, 30, 512, 512))", "votes": null }, { "id": "2172956", "postDate": "03/08/2023 00:58:06", "content": "im = torch.randn((8, 30, 512, 512))
\n
Moving average is the average of previous and next data.
\nIn short, df['pred'].rolling().mean().
\nBelow is the pseudo code.
\ntest_df['pred_ma'] = test_df.groupby(['game_play', 'nfl_player_id_1', 'nfl_player_id_2'])['pred'].rolling(3, center=True, min_periods=1).mean().to_frame('pred_ma').reset_index()['pred_ma']
I explain about Residual3DBlock.
\nIn this competition, 3d(t, h, w) is very important, so I decided to use any time analysis method.
\n15 frames reshape this ({-28, -24, -20}{-16, -12, -8}{-4, 0, 4}{8, 12, 16}{20, 24, 28}) and extract feature using backbone, then hidden output is applied it. here is a sample for using residual3d block(final architecture is very similar)
class Residual3DBlock(nn.Module):\n def __init__(self):\n super(Residual3DBlock, self).__init__()\n\n self.block = nn.Sequential(\n nn.Conv3d(512, 512, 3, stride=1, padding=1),\n nn.BatchNorm3d(512),\n nn.ReLU(512)\n )\n\n self.block2 = nn.Sequential(\n nn.Conv3d(512, 512, 3, stride=1, padding=1),\n nn.BatchNorm3d(512),\n )\n\n def forward(self, images):\n short_cut = images\n h = self.block(images)\n h = self.block2(h)\n\n return F.relu(h + short_cut)\n\nclass Model(nn.Module):\n def __init__(self):\n super(Model, self).__init__()\n self.backbone = timm.create_model(\"tf_efficientnet_b0_ns\", pretrained=True, num_classes=1, in_chans=3)\n self.mlp = nn.Sequential(\n nn.Linear(68, 256),\n nn.BatchNorm1d(256),\n nn.LeakyReLU(),\n nn.Dropout(0.2),\n )\n n_hidden = 1024\n\n self.conv_proj = nn.Sequential(\n nn.Conv2d(1280, 512, 1, stride=1),\n nn.BatchNorm2d(512),\n nn.ReLU(),\n )\n\n self.neck = nn.Sequential(\n nn.Linear(1024, 1024),\n nn.BatchNorm1d(1024),\n nn.LeakyReLU(),\n nn.Dropout(0.2),\n )\n\n self.triple_layer = nn.Sequential(\n Residual3DBlock(),\n )\n\n self.pool = GeM() \n\n self.fc = nn.Linear(256+1024, 1)\n\n def forward(self, images,feature, target=None, mixup_hidden = False, mixup_alpha = 0.1, layer_mix=None):\n b, t, h, w = images.shape\n images = images.view(b * t // 3, 3, h, w)\n feature_maps = self.conv_proj(self.backbone.forward_features(images))\n _, c, h, w = feature_maps.size()\n feature_maps = feature_maps.contiguous().view(b * 2,c,t // 2 // 3, h, w)\n feature_maps = self.triple_layer(feature_maps)\n middle_maps = feature_maps[:, :, 2, :, :]\n #b, h, w= middle_maps.size()\n #middle_maps = middle_maps.view(b, 1, h, w)\n nn_feature = self.neck(self.pool(middle_maps).reshape(b, -1))\n feature = self.mlp(feature)\n cat_features = torch.cat([nn_feature, feature], dim=1)\n if target is not None:\n cat_features, y_a, y_b, lam = mixup_data(cat_features, target, mixup_alpha)\n y = self.fc(cat_features)\n return y, y_a, y_b, lam\n else:\n y = self.fc(cat_features)\n return y\nI tried to run the code, but it seems to give me an error. What am I doing wrong?
\nmodel = Model()\nim = torch.randn((8, 15, 512, 512))\nfeature = torch.randn((8, 68))\nmodel(im, feature)\nOh, do you use the following because of the end and side?
\nim = torch.randn((8, 30, 512, 512))\nyes, it's correct.
\nb, t(side + endzone), height, width.
\n", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2172287, "author_name": "chg0901", "author_url": "", "post_date": "03/07/2023 12:29:08", "content": "im = torch.randn((8, 30, 512, 512))
\n
May I ask a question? what is moving average? I have seen that it is mentioned several times
", "votes": null, "replies": [ { "id": 2172956, "author_name": "anonamename", "author_url": "", "post_date": "03/08/2023 00:58:06", "content": "Moving average is the average of previous and next data.
\nIn short, df['pred'].rolling().mean().
\nBelow is the pseudo code.
\ntest_df['pred_ma'] = test_df.groupby(['game_play', 'nfl_player_id_1', 'nfl_player_id_2'])['pred'].rolling(3, center=True, min_periods=1).mean().to_frame('pred_ma').reset_index()['pred_ma']