Based on this repo: https://github.com/HHTseng/video-classification\nThis model combines the Xception model in my baseline with a LSTM.\nI'm not 100% sure if this is correct but I have been able to train it normally. Any advice/fixes appreciated!\n```\n!pip install pytorchcv --quiet\nfrom pytorchcv.model_provider import get_model\nbase_model = get_model(\"xception\", pretrained=True)\nbase_model = nn.Sequential(*list(model.children())[:-1]) # Remove original output layer\nbase_model[0].final_block.pool = nn.Sequential(nn.AdaptiveAvgPool2d(1))
\n\nclass Head(torch.nn.Module):\n def init(self, in_f, out_f):\n super(Head, self).init()
\n\nself.f = nn.Flatten()\nself.l = nn.Linear(in_f, 512)\nself.d = nn.Dropout(0.75)\nself.o = nn.Linear(512, out_f)\n# self.o = nn.Linear(in_f, out_f)\nself.b1 = nn.BatchNorm1d(in_f)\nself.b2 = nn.BatchNorm1d(512)\nself.r = nn.ReLU()\ndef forward(self, x):\n x = self.f(x)\n x = self.b1(x)\n x = self.d(x)
\n\nx = self.l(x)\nx = self.r(x)\nx = self.b2(x)\nx = self.d(x)\n\nout = self.o(x)\nreturn out\nclass CNNEncoder(torch.nn.Module):\n def init(self, base, in_f, out_f):\n super(CNNEncoder, self).init()\n self.base = base\n self.h1 = Head(in_f, out_f)
\n\ndef forward(self, x_3d):\n cnn_embed_seq = []
\n\nfor i in range(x_3d.size(1)):\n x = self.base(x_3d[:,i,:,:,:])\n x = self.h1(x)\n cnn_embed_seq.append(x)\n\ncnn_embed_seq = torch.stack(cnn_embed_seq, dim=0).transpose_(0, 1)\nreturn cnn_embed_seq\nclass RNNDecoder(nn.Module):\n def init(self, in_f, out_f):\n super(RNNDecoder, self).init()\n self.LSTM = nn.LSTM(\n input_size=in_f,\n hidden_size=256,\n num_layers=3,\n batch_first=True\n )
\n\nself.f1 = nn.Linear(256, 128)\nself.f2 = nn.Linear(128, out_f)\nself.r = nn.ReLU()\nself.d = nn.Dropout(0.3)\ndef forward(self, x):\n self.LSTM.flatten_parameters()\n x, (hn,hc) = self.LSTM(x)\n x = self.d(self.r(self.f1(x[:,-1,:])))\n x = self.f2(x)\n return x
\n\nclass LRCN(nn.Module):\n def init(self, base_model, cnn_features, out_f):\n super(LRCN, self).init()\n self.cnn = CNNEncoder(base_model, 2048, cnn_features)\n self.rnn = RNNDecoder(cnn_features, out_f)
\n\ndef forward(self, x_3d):\n return self.rnn(self.cnn(x_3d))\ncnn_features = 300\nmodel = LRCN(base_model, cnn_features, 1)\n```
", "messages": [ { "id": 766156, "postDate": "2020-03-07T18:51:19.407Z", "content": "Based on this repo: https://github.com/HHTseng/video-classification\nThis model combines the Xception model in my baseline with a LSTM.\nI'm not 100% sure if this is correct but I have been able to train it normally. Any advice/fixes appreciated!\n```\n!pip install pytorchcv --quiet\nfrom pytorchcv.model_provider import get_model\nbase_model = get_model(\"xception\", pretrained=True)\nbase_model = nn.Sequential(*list(model.children())[:-1]) # Remove original output layer\nbase_model[0].final_block.pool = nn.Sequential(nn.AdaptiveAvgPool2d(1))
\n\nclass Head(torch.nn.Module):\n def init(self, in_f, out_f):\n super(Head, self).init()
\n\nself.f = nn.Flatten()\nself.l = nn.Linear(in_f, 512)\nself.d = nn.Dropout(0.75)\nself.o = nn.Linear(512, out_f)\n# self.o = nn.Linear(in_f, out_f)\nself.b1 = nn.BatchNorm1d(in_f)\nself.b2 = nn.BatchNorm1d(512)\nself.r = nn.ReLU()\ndef forward(self, x):\n x = self.f(x)\n x = self.b1(x)\n x = self.d(x)
\n\nx = self.l(x)\nx = self.r(x)\nx = self.b2(x)\nx = self.d(x)\n\nout = self.o(x)\nreturn out\nclass CNNEncoder(torch.nn.Module):\n def init(self, base, in_f, out_f):\n super(CNNEncoder, self).init()\n self.base = base\n self.h1 = Head(in_f, out_f)
\n\ndef forward(self, x_3d):\n cnn_embed_seq = []
\n\nfor i in range(x_3d.size(1)):\n x = self.base(x_3d[:,i,:,:,:])\n x = self.h1(x)\n cnn_embed_seq.append(x)\n\ncnn_embed_seq = torch.stack(cnn_embed_seq, dim=0).transpose_(0, 1)\nreturn cnn_embed_seq\nclass RNNDecoder(nn.Module):\n def init(self, in_f, out_f):\n super(RNNDecoder, self).init()\n self.LSTM = nn.LSTM(\n input_size=in_f,\n hidden_size=256,\n num_layers=3,\n batch_first=True\n )
\n\nself.f1 = nn.Linear(256, 128)\nself.f2 = nn.Linear(128, out_f)\nself.r = nn.ReLU()\nself.d = nn.Dropout(0.3)\ndef forward(self, x):\n self.LSTM.flatten_parameters()\n x, (hn,hc) = self.LSTM(x)\n x = self.d(self.r(self.f1(x[:,-1,:])))\n x = self.f2(x)\n return x
\n\nclass LRCN(nn.Module):\n def init(self, base_model, cnn_features, out_f):\n super(LRCN, self).init()\n self.cnn = CNNEncoder(base_model, 2048, cnn_features)\n self.rnn = RNNDecoder(cnn_features, out_f)
\n\ndef forward(self, x_3d):\n return self.rnn(self.cnn(x_3d))\ncnn_features = 300\nmodel = LRCN(base_model, cnn_features, 1)\n```
", "rawMarkdown": "Based on this repo: https://github.com/HHTseng/video-classification\nThis model combines the Xception model in my [baseline](https://www.kaggle.com/greatgamedota/xception-classifier-w-ffhq-training-lb-537) with a LSTM.\nI'm not 100% sure if this is correct but I have been able to train it normally. Any advice/fixes appreciated!\n```\n!pip install pytorchcv --quiet\nfrom pytorchcv.model_provider import get_model\nbase_model = get_model(\"xception\", pretrained=True)\nbase_model = nn.Sequential(*list(model.children())[:-1]) # Remove original output layer\nbase_model[0].final_block.pool = nn.Sequential(nn.AdaptiveAvgPool2d(1))\n\nclass Head(torch.nn.Module):\n def __init__(self, in_f, out_f):\n super(Head, self).__init__()\n \n self.f = nn.Flatten()\n self.l = nn.Linear(in_f, 512)\n self.d = nn.Dropout(0.75)\n self.o = nn.Linear(512, out_f)\n # self.o = nn.Linear(in_f, out_f)\n self.b1 = nn.BatchNorm1d(in_f)\n self.b2 = nn.BatchNorm1d(512)\n self.r = nn.ReLU()\n\n def forward(self, x):\n x = self.f(x)\n x = self.b1(x)\n x = self.d(x)\n\n x = self.l(x)\n x = self.r(x)\n x = self.b2(x)\n x = self.d(x)\n\n out = self.o(x)\n return out\n\nclass CNNEncoder(torch.nn.Module):\n def __init__(self, base, in_f, out_f):\n super(CNNEncoder, self).__init__()\n self.base = base\n self.h1 = Head(in_f, out_f)\n \n def forward(self, x_3d):\n cnn_embed_seq = []\n \n for i in range(x_3d.size(1)):\n x = self.base(x_3d[:,i,:,:,:])\n x = self.h1(x)\n cnn_embed_seq.append(x)\n \n cnn_embed_seq = torch.stack(cnn_embed_seq, dim=0).transpose_(0, 1)\n return cnn_embed_seq\n\n\nclass RNNDecoder(nn.Module):\n def __init__(self, in_f, out_f):\n super(RNNDecoder, self).__init__()\n self.LSTM = nn.LSTM(\n input_size=in_f,\n hidden_size=256,\n num_layers=3,\n batch_first=True\n )\n\n self.f1 = nn.Linear(256, 128)\n self.f2 = nn.Linear(128, out_f)\n self.r = nn.ReLU()\n self.d = nn.Dropout(0.3)\n \n def forward(self, x):\n self.LSTM.flatten_parameters()\n x, (hn,hc) = self.LSTM(x)\n x = self.d(self.r(self.f1(x[:,-1,:])))\n x = self.f2(x)\n return x\n\n\nclass LRCN(nn.Module):\n def __init__(self, base_model, cnn_features, out_f):\n super(LRCN, self).__init__()\n self.cnn = CNNEncoder(base_model, 2048, cnn_features)\n self.rnn = RNNDecoder(cnn_features, out_f)\n \n def forward(self, x_3d):\n return self.rnn(self.cnn(x_3d))\n\ncnn_features = 300\nmodel = LRCN(base_model, cnn_features, 1)\n```", "votes": 32 }, { "id": 766887, "postDate": "2020-03-08T22:36:03.603Z", "content": "I've been experimenting with a similar type of model. The LCRN paper mentions that they needed to use a lot of dropout (p=0.9) or the model would overfit.
\n\nThat's also been my problem: I can get the model to overfit quite easily but haven't quite figured out how to stop that and also get a good validation score.
", "rawMarkdown": "I've been experimenting with a similar type of model. The LCRN paper mentions that they needed to use a lot of dropout (p=0.9) or the model would overfit. \n\nThat's also been my problem: I can get the model to overfit quite easily but haven't quite figured out how to stop that and also get a good validation score.", "votes": 3 }, { "id": 766563, "postDate": "2020-03-08T11:34:59.077Z", "content": "I also use it, but didn't get a good score。
", "rawMarkdown": "I also use it, but didn't get a good score。 ", "votes": 2, "replies": [ { "id": 766852, "postDate": "2020-03-08T21:02:12.800Z", "content": "Same here, simple CNN performs better. Maybe CNN does better with more data and avging over a lot of frames.
", "rawMarkdown": "Same here, simple CNN performs better. Maybe CNN does better with more data and avging over a lot of frames.", "votes": 1 }, { "id": 768455, "postDate": "2020-03-10T20:24:37.227Z", "content": "I've tried something similar a few weeks ago. How many frames are you using in the sequence? And are you using consecutive frames in your sequence? I've tried with 4 frames and 10 frames (spaced at regular interval) and it did not work.
", "rawMarkdown": "I've tried something similar a few weeks ago. How many frames are you using in the sequence? And are you using consecutive frames in your sequence? I've tried with 4 frames and 10 frames (spaced at regular interval) and it did not work." }, { "id": 768461, "postDate": "2020-03-10T20:36:08.513Z", "content": "5 frames, (I used the exact model here) val loss doesn't spike randomly but it still overfit badly.
", "rawMarkdown": "5 frames, (I used the exact model here) val loss doesn't spike randomly but it still overfit badly.", "votes": 1 }, { "id": 768946, "postDate": "2020-03-11T11:30:15.370Z", "content": "I use 30 frames(skip frame) in training, and use 10 frames in kaggle, moreover, I ensamble two model results(pred = model_A * 0.5 + model_B * 0.5) ， finally I clipping the preds(preds = preds.clip(0.050.95)) to get a good score (0.36846) in LB
", "rawMarkdown": "I use 30 frames(skip frame) in training, and use 10 frames in kaggle, moreover, I ensamble two model results(pred = model_A * 0.5 + model_B * 0.5) ， finally I clipping the preds(preds = preds.clip(0.050.95)) to get a good score (0.36846) in LB", "votes": 4 }, { "id": 769073, "postDate": "2020-03-11T14:03:27.103Z", "content": "How much of a boost does clipping give you?
", "rawMarkdown": "How much of a boost does clipping give you?" }, { "id": 769090, "postDate": "2020-03-11T14:20:59.723Z", "content": "@greatgamedota I didn't test the ensamble without clipping, but single CRNN(model_B) just LB=0.43 with the same clipping and without it are 0.67932 (model_A)and 0.66504(model_B), model_A and model_B are two different imagenet CNN model. However the single CNN model's LB is 0.35
", "rawMarkdown": "@greatgamedota I didn't test the ensamble without clipping, but single CRNN(model_B) just LB=0.43 with the same clipping and without it are 0.67932 (model_A)and 0.66504(model_B), model_A and model_B are two different imagenet CNN model. However the single CNN model's LB is 0.35", "votes": 2 }, { "id": 776636, "postDate": "2020-03-17T14:23:16.797Z", "content": "@chenbaoying Have you tried this lrcn model to get the good score 0.36?
", "rawMarkdown": "@chenbaoying Have you tried this lrcn model to get the good score 0.36?" }, { "id": 776840, "postDate": "2020-03-17T17:09:21.797Z", "content": "@zhuolin yes,but I ensamble two model ,now my LCRN score is 0.35
", "rawMarkdown": "@zhuolin yes,but I ensamble two model ,now my LCRN score is 0.35" }, { "id": 776853, "postDate": "2020-03-17T17:18:06.057Z", "content": "@chenbaoying Do you mean two LRCN models? Thanks.
", "rawMarkdown": "@chenbaoying Do you mean two LRCN models? Thanks." }, { "id": 778223, "postDate": "2020-03-18T08:44:15.883Z", "content": "yes, use two LCRN to predict the videos, and the ensamble by final_pred = 0.5*model1_pred+0.5*model2*0.5
", "rawMarkdown": "yes, use two LCRN to predict the videos, and the ensamble by final_pred = 0.5*model1_pred+0.5*model2*0.5" }, { "id": 785589, "postDate": "2020-03-25T07:23:43.493Z", "content": "@chenbaoying \nHello, what is your single LRCN model's LB score? Have you tested?
", "rawMarkdown": "@chenbaoying \nHello, what is your single LRCN model's LB score? Have you tested?" } ] }, { "id": 778408, "postDate": "2020-03-18T12:16:37.300Z", "content": "I have a question @greatgamedota , since I am using your baseline Kernel . We already have the dataset created from multiple videos . How are we going to pass the information or group the sequence of 10-15 images from the videos ? Is it like process images sequentially and group it ? \nI am using the same dataset as your public kernel without ffhq . Like I have shown in my public kernel of densenet169 .
", "rawMarkdown": "I have a question @greatgamedota , since I am using your baseline Kernel . We already have the dataset created from multiple videos . How are we going to pass the information or group the sequence of 10-15 images from the videos ? Is it like process images sequentially and group it ? \nI am using the same dataset as your public kernel without ffhq . Like I have shown in my public kernel of densenet169 .", "replies": [ { "id": 778491, "postDate": "2020-03-18T13:49:09.223Z", "content": "@phoenix9032 That dataset won't work with this model since it is only 1 image per video. If you want a bigger dataset I have shared one here: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/134420\nThat dataset has ~10 images per video, some have more or less.\nTo pass multiple images into the model you just pass a (batch,frame,channel,width,height) tensor.\nI have not gotten this model to perform very well, current score is still using my public model.
", "rawMarkdown": "@phoenix9032 That dataset won't work with this model since it is only 1 image per video. If you want a bigger dataset I have shared one here: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/134420\nThat dataset has ~10 images per video, some have more or less.\nTo pass multiple images into the model you just pass a (batch,frame,channel,width,height) tensor.\nI have not gotten this model to perform very well, current score is still using my public model.", "votes": 1 }, { "id": 780156, "postDate": "2020-03-20T02:46:13.097Z", "content": "@greatgamedota it could work..... I'm already revealing stuff that can get you high in LB like crazy.
", "rawMarkdown": "@greatgamedota it could work..... I'm already revealing stuff that can get you high in LB like crazy." } ] }, { "id": 776600, "postDate": "2020-03-17T13:54:46.583Z", "content": "@greatgamedota Thanks for sharing. Why the num_frames is the depth parameter in the LSTM module \nself.LSTM = nn.LSTM(\n input_size=in_f,\n hidden_size=256,\n num_layers=depth,\n batch_first=True\n )?
Looked into it. I was under the impression you had to tell the LSTM how long the time distributed data was but it turns out you don't. I will rename it to just num hidden RNN layers. Thanks!
", "rawMarkdown": "Looked into it. I was under the impression you had to tell the LSTM how long the time distributed data was but it turns out you don't. I will rename it to just num hidden RNN layers. Thanks!" }, { "id": 776848, "postDate": "2020-03-17T17:15:15.023Z", "content": "Yes. When I read your code, it is strange. Maybe you will get good score after you fix this :)
", "rawMarkdown": "Yes. When I read your code, it is strange. Maybe you will get good score after you fix this :)", "votes": 1 } ] }, { "id": 766996, "postDate": "2020-03-09T02:33:40.790Z", "content": "Do we need to crop for face when using RNN model ? Or should we fixed the area where embedding is extracted ?
", "rawMarkdown": "Do we need to crop for face when using RNN model ? Or should we fixed the area where embedding is extracted ? \n\n", "replies": [ { "id": 767001, "postDate": "2020-03-09T02:42:14.930Z", "content": "You could run it with an entire frame but I'd run it with just the faces cropped.
", "rawMarkdown": "You could run it with an entire frame but I'd run it with just the faces cropped." } ] }, { "id": 769619, "postDate": "2020-03-12T05:02:40.193Z", "content": "Thanks for sharing bro
", "rawMarkdown": "Thanks for sharing bro", "votes": 1 }, { "id": 769027, "postDate": "2020-03-11T12:51:47.157Z", "content": "thanks bro
", "rawMarkdown": "thanks bro" }, { "id": 768228, "postDate": "2020-03-10T15:04:33.003Z", "content": "Thanks for sharing
", "rawMarkdown": "Thanks for sharing" }, { "id": 767243, "postDate": "2020-03-09T11:09:50.933Z", "content": "thanks
", "rawMarkdown": "thanks" }, { "id": 769655, "postDate": "2020-03-12T05:54:07.243Z", "content": "thanks a lot
", "rawMarkdown": "thanks a lot", "isDeleted": true }, { "id": 767617, "postDate": "2020-03-09T22:33:35.110Z", "content": "Thanks for sharing
", "rawMarkdown": "Thanks for sharing", "isDeleted": true } ], "comments": [ { "id": 766887, "author_name": "Human Analog", "author_url": "", "post_date": "2020-03-08T22:36:03.603000", "content": "I've been experimenting with a similar type of model. The LCRN paper mentions that they needed to use a lot of dropout (p=0.9) or the model would overfit.
\n\nThat's also been my problem: I can get the model to overfit quite easily but haven't quite figured out how to stop that and also get a good validation score.
", "votes": 3, "replies": [] }, { "id": 766563, "author_name": "BokingChen", "author_url": "", "post_date": "2020-03-08T11:34:59.077000", "content": "I also use it, but didn't get a good score。
", "votes": 2, "replies": [ { "id": 766852, "author_name": "GreatGameDota", "author_url": "", "post_date": "2020-03-08T21:02:12.800000", "content": "Same here, simple CNN performs better. Maybe CNN does better with more data and avging over a lot of frames.
", "votes": 1, "replies": [] }, { "id": 768455, "author_name": "MPWARE", "author_url": "", "post_date": "2020-03-10T20:24:37.227000", "content": "I've tried something similar a few weeks ago. How many frames are you using in the sequence? And are you using consecutive frames in your sequence? I've tried with 4 frames and 10 frames (spaced at regular interval) and it did not work.
", "votes": 0, "replies": [] }, { "id": 768461, "author_name": "GreatGameDota", "author_url": "", "post_date": "2020-03-10T20:36:08.513000", "content": "5 frames, (I used the exact model here) val loss doesn't spike randomly but it still overfit badly.
", "votes": 1, "replies": [] }, { "id": 768946, "author_name": "BokingChen", "author_url": "", "post_date": "2020-03-11T11:30:15.370000", "content": "I use 30 frames(skip frame) in training, and use 10 frames in kaggle, moreover, I ensamble two model results(pred = model_A * 0.5 + model_B * 0.5) ， finally I clipping the preds(preds = preds.clip(0.050.95)) to get a good score (0.36846) in LB
", "votes": 4, "replies": [] }, { "id": 769073, "author_name": "GreatGameDota", "author_url": "", "post_date": "2020-03-11T14:03:27.103000", "content": "How much of a boost does clipping give you?
", "votes": 0, "replies": [] }, { "id": 769090, "author_name": "BokingChen", "author_url": "", "post_date": "2020-03-11T14:20:59.723000", "content": "@greatgamedota I didn't test the ensamble without clipping, but single CRNN(model_B) just LB=0.43 with the same clipping and without it are 0.67932 (model_A)and 0.66504(model_B), model_A and model_B are two different imagenet CNN model. However the single CNN model's LB is 0.35
", "votes": 2, "replies": [] }, { "id": 776636, "author_name": "zjiang", "author_url": "", "post_date": "2020-03-17T14:23:16.797000", "content": "@chenbaoying Have you tried this lrcn model to get the good score 0.36?
", "votes": 0, "replies": [] }, { "id": 776840, "author_name": "BokingChen", "author_url": "", "post_date": "2020-03-17T17:09:21.797000", "content": "@zhuolin yes,but I ensamble two model ,now my LCRN score is 0.35
", "votes": 0, "replies": [] }, { "id": 776853, "author_name": "zjiang", "author_url": "", "post_date": "2020-03-17T17:18:06.057000", "content": "@chenbaoying Do you mean two LRCN models? Thanks.
", "votes": 0, "replies": [] }, { "id": 778223, "author_name": "BokingChen", "author_url": "", "post_date": "2020-03-18T08:44:15.883000", "content": "yes, use two LCRN to predict the videos, and the ensamble by final_pred = 0.5*model1_pred+0.5*model2*0.5
", "votes": 0, "replies": [] }, { "id": 785589, "author_name": "xmaocai", "author_url": "", "post_date": "2020-03-25T07:23:43.493000", "content": "@chenbaoying \nHello, what is your single LRCN model's LB score? Have you tested?
", "votes": 0, "replies": [] } ] }, { "id": 778408, "author_name": "Nirjhar Roy", "author_url": "", "post_date": "2020-03-18T12:16:37.300000", "content": "I have a question @greatgamedota , since I am using your baseline Kernel . We already have the dataset created from multiple videos . How are we going to pass the information or group the sequence of 10-15 images from the videos ? Is it like process images sequentially and group it ? \nI am using the same dataset as your public kernel without ffhq . Like I have shown in my public kernel of densenet169 .
", "votes": 0, "replies": [ { "id": 778491, "author_name": "GreatGameDota", "author_url": "", "post_date": "2020-03-18T13:49:09.223000", "content": "@phoenix9032 That dataset won't work with this model since it is only 1 image per video. If you want a bigger dataset I have shared one here: https://www.kaggle.com/c/deepfake-detection-challenge/discussion/134420\nThat dataset has ~10 images per video, some have more or less.\nTo pass multiple images into the model you just pass a (batch,frame,channel,width,height) tensor.\nI have not gotten this model to perform very well, current score is still using my public model.
", "votes": 1, "replies": [] }, { "id": 780156, "author_name": "Shangqiu Li", "author_url": "", "post_date": "2020-03-20T02:46:13.097000", "content": "@greatgamedota it could work..... I'm already revealing stuff that can get you high in LB like crazy.
", "votes": 0, "replies": [] } ] }, { "id": 776600, "author_name": "zjiang", "author_url": "", "post_date": "2020-03-17T13:54:46.583000", "content": "@greatgamedota Thanks for sharing. Why the num_frames is the depth parameter in the LSTM module \nself.LSTM = nn.LSTM(\n input_size=in_f,\n hidden_size=256,\n num_layers=depth,\n batch_first=True\n )?
Looked into it. I was under the impression you had to tell the LSTM how long the time distributed data was but it turns out you don't. I will rename it to just num hidden RNN layers. Thanks!
", "votes": 0, "replies": [] }, { "id": 776848, "author_name": "zjiang", "author_url": "", "post_date": "2020-03-17T17:15:15.023000", "content": "Yes. When I read your code, it is strange. Maybe you will get good score after you fix this :)
", "votes": 1, "replies": [] } ] }, { "id": 766996, "author_name": "yukiya", "author_url": "", "post_date": "2020-03-09T02:33:40.790000", "content": "Do we need to crop for face when using RNN model ? Or should we fixed the area where embedding is extracted ?
", "votes": 0, "replies": [ { "id": 767001, "author_name": "GreatGameDota", "author_url": "", "post_date": "2020-03-09T02:42:14.930000", "content": "You could run it with an entire frame but I'd run it with just the faces cropped.
", "votes": 0, "replies": [] } ] }, { "id": 769619, "author_name": "Abhishek Kumar Pandit", "author_url": "", "post_date": "2020-03-12T05:02:40.193000", "content": "Thanks for sharing bro
", "votes": 1, "replies": [] }, { "id": 769027, "author_name": "Asd_123", "author_url": "", "post_date": "2020-03-11T12:51:47.157000", "content": "thanks bro
", "votes": 0, "replies": [] }, { "id": 768228, "author_name": "Kranti Kumar", "author_url": "", "post_date": "2020-03-10T15:04:33.003000", "content": "Thanks for sharing
", "votes": 0, "replies": [] }, { "id": 767243, "author_name": "mustafa mahmutoglu", "author_url": "", "post_date": "2020-03-09T11:09:50.933000", "content": "thanks
", "votes": 0, "replies": [] }, { "id": 769655, "author_name": "", "author_url": "", "post_date": "2020-03-12T05:54:07.243000", "content": "thanks a lot
", "votes": 0, "replies": [] }, { "id": 767617, "author_name": "", "author_url": "", "post_date": "2020-03-09T22:33:35.110000", "content": "Thanks for sharing
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "766156": "Based on this repo: https://github.com/HHTseng/video-classification\nThis model combines the Xception model in my [baseline](https://www.kaggle.com/greatgamedota/xception-classifier-w-ffhq-training-lb-537) with a LSTM.\nI'm not 100% sure if this is correct but I have been able to train it normally. Any advice/fixes appreciated!\n```\n!pip install pytorchcv --quiet\nfrom pytorchcv.model_provider import get_model\nbase_model = get_model(\"xception\", pretrained=True)\nbase_model = nn.Sequential(*list(model.children())[:-1]) # Remove original output layer\nbase_model[0].final_block.pool = nn.Sequential(nn.AdaptiveAvgPool2d(1))\n\nclass Head(torch.nn.Module):\n def __init__(self, in_f, out_f):\n super(Head, self).__init__()\n \n self.f = nn.Flatten()\n self.l = nn.Linear(in_f, 512)\n self.d = nn.Dropout(0.75)\n self.o = nn.Linear(512, out_f)\n # self.o = nn.Linear(in_f, out_f)\n self.b1 = nn.BatchNorm1d(in_f)\n self.b2 = nn.BatchNorm1d(512)\n self.r = nn.ReLU()\n\n def forward(self, x):\n x = self.f(x)\n x = self.b1(x)\n x = self.d(x)\n\n x = self.l(x)\n x = self.r(x)\n x = self.b2(x)\n x = self.d(x)\n\n out = self.o(x)\n return out\n\nclass CNNEncoder(torch.nn.Module):\n def __init__(self, base, in_f, out_f):\n super(CNNEncoder, self).__init__()\n self.base = base\n self.h1 = Head(in_f, out_f)\n \n def forward(self, x_3d):\n cnn_embed_seq = []\n \n for i in range(x_3d.size(1)):\n x = self.base(x_3d[:,i,:,:,:])\n x = self.h1(x)\n cnn_embed_seq.append(x)\n \n cnn_embed_seq = torch.stack(cnn_embed_seq, dim=0).transpose_(0, 1)\n return cnn_embed_seq\n\n\nclass RNNDecoder(nn.Module):\n def __init__(self, in_f, out_f):\n super(RNNDecoder, self).__init__()\n self.LSTM = nn.LSTM(\n input_size=in_f,\n hidden_size=256,\n num_layers=3,\n batch_first=True\n )\n\n self.f1 = nn.Linear(256, 128)\n self.f2 = nn.Linear(128, out_f)\n self.r = nn.ReLU()\n self.d = nn.Dropout(0.3)\n \n def forward(self, x):\n self.LSTM.flatten_parameters()\n x, (hn,hc) = self.LSTM(x)\n x = self.d(self.r(self.f1(x[:,-1,:])))\n x = self.f2(x)\n return x\n\n\nclass LRCN(nn.Module):\n def __init__(self, base_model, cnn_features, out_f):\n super(LRCN, self).__init__()\n self.cnn = CNNEncoder(base_model, 2048, cnn_features)\n self.rnn = RNNDecoder(cnn_features, out_f)\n \n def forward(self, x_3d):\n return self.rnn(self.cnn(x_3d))\n\ncnn_features = 300\nmodel = LRCN(base_model, cnn_features, 1)\n```", "766887": "I've been experimenting with a similar type of model. The LCRN paper mentions that they needed to use a lot of dropout (p=0.9) or the model would overfit. \n\nThat's also been my problem: I can get the model to overfit quite easily but haven't quite figured out how to stop that and also get a good validation score.", "766563": "I also use it, but didn't get a good score。 ", "778408": "I have a question @greatgamedota , since I am using your baseline Kernel . We already have the dataset created from multiple videos . How are we going to pass the information or group the sequence of 10-15 images from the videos ? Is it like process images sequentially and group it ? \nI am using the same dataset as your public kernel without ffhq . Like I have shown in my public kernel of densenet169 .", "776600": "@greatgamedota Thanks for sharing. Why the `num_frames` is the depth parameter in the LSTM module \n`self.LSTM = nn.LSTM(\n input_size=in_f,\n hidden_size=256,\n num_layers=depth,\n batch_first=True\n )`?", "766996": "Do we need to crop for face when using RNN model ? Or should we fixed the area where embedding is extracted ? \n\n", "769619": "Thanks for sharing bro", "769027": "thanks bro", "768228": "Thanks for sharing", "767243": "thanks", "769655": "thanks a lot", "767617": "Thanks for sharing" } }