msr action

[weiyutao886]

Hello, I used your code to experiment on MSR action dataset, and the accuracy did not reach the level in your paper. Why? My parameters are as follows: num_epoch: 300 work_dir: ./work_dir/baseline/ batch_size: 8 test_batch_size: 8 num_worker: 10

empty for cpu

device: 1 log_interval: 50 eval_interval: 5 save_interval: 5

weights: ./work_dir/pointlstm/epoch200_model.pt

framesize: &framesize 32 pts_size: &frame_pts_size 128

optimizer_args: optimizer: Adam base_lr: 0.0001 step: [ 100, 160, 180] weight_decay: 0.005 start_epoch: 0 nesterov: False

[weiyutao886]

and i follow this： Yes, this is correct, the corresponding subject splits are [ ([1, 2, 3, 4, 5], [6, 7, 8, 9, 10]), ([1, 3, 5, 7, 9], [2, 4, 6, 8, 10]), ([1, 4, 7, 10, 3], [2, 5, 6, 8, 9]), ([1, 5, 9, 3, 7], [2, 4, 6, 8, 10]), ([1, 6, 2, 7, 3], [4, 5, 8, 9, 10]) ]. We just find a bug that the second group is the same as the fourth group (because we adopt (subject_idx+offset)%10 to generate the splits), and the corresponding results of PointLSTM-late are [ 93.13, 87.59, 97.07, 89.08, 92.78, 94.36, 97.07, 89.08, 91.11, 91.64], and the corrected accuracy is 92.10 \pm 2.78.

I use the same strategy as you, but I can't reach each accuracy you said

[ycmin95]

Can you provide more details about the experiments (e.g., the experimental log)? What kind of performance can you achieve?

[weiyutao886]

Thank you for your reply. I currently use subjet Sprint ([1, 2, 3, 4, 5], [6, 7, 8, 9, 10]), and the corresponding accuracy should be 93.13, 87.59 you mentioned. I achieved the current effect by adjusting the parameters bitchsize to 8 and framesize to 24, as follows

{'work_dir': 'result17', 'config': 'pointlstm.yaml', 'device': '0', 'phase': 'train', 'random_fix': True, 'random_seed': 0, 'save_interval': 5, 'eval_inte train subjet[1, 2, 3, 4, 5] test [6, 7, 8, 9, 10]) result： [ Fri Apr 1 23:36:10 2022 ] Epoch 135, Test, Evaluation: prec1 90.5724, prec5 96.6330 [ Fri Apr 1 23:38:48 2022 ] Epoch 140, Test, Evaluation: prec1 89.5623, prec5 96.6330 [ Fri Apr 1 23:41:25 2022 ] Epoch 145, Test, Evaluation: prec1 90.2357, prec5 96.6330 [ Fri Apr 1 23:44:03 2022 ] Epoch 150, Test, Evaluation: prec1 88.2155, prec5 96.6330 [ Fri Apr 1 23:46:39 2022 ] Epoch 155, Test, Evaluation: prec1 88.8889, prec5 96.6330 [ Fri Apr 1 23:49:15 2022 ] Epoch 160, Test, Evaluation: prec1 90.9091, prec5 96.6330 [ Fri Apr 1 23:51:51 2022 ] Epoch 165, Test, Evaluation: prec1 89.8990, prec5 96.6330 [ Fri Apr 1 23:54:29 2022 ] Epoch 170, Test, Evaluation: prec1 89.5623, prec5 96.6330

best result [ Fri Apr 1 23:49:15 2022 ] Epoch 160, Test, Evaluation: prec1 90.9091, prec5 96.6330 Your result is 93.13,

train subjet[6, 7, 8, 9, 10]) test[1, 2, 3, 4, 5] result： [ Fri Apr 1 10:25:15 2022 ] Epoch 145, Test, Evaluation: prec1 85.1852, prec5 98.5185 [ Fri Apr 1 10:29:01 2022 ] Epoch 150, Test, Evaluation: prec1 87.0370, prec5 97.7778 [ Fri Apr 1 10:32:47 2022 ] Epoch 155, Test, Evaluation: prec1 84.4444, prec5 98.8889 [ Fri Apr 1 10:36:33 2022 ] Epoch 160, Test, Evaluation: prec1 87.0370, prec5 97.7778 [ Fri Apr 1 10:40:18 2022 ] Epoch 165, Test, Evaluation: prec1 87.7778, prec5 97.4074 [ Fri Apr 1 10:44:04 2022 ] Epoch 170, Test, Evaluation: prec1 86.6667, prec5 98.1481 [ Fri Apr 1 10:47:50 2022 ] Epoch 175, Test, Evaluation: prec1 87.0370, prec5 98.1481

best result [ Fri Apr 1 10:40:18 2022 ] Epoch 165, Test, Evaluation: prec1 87.7778, prec5 97.4074 your result is 87.59

my model：pointlstm-late

stage 3: inter-frame, middle, applying lstm in this stage

    in_dims = fea2.shape[1] * 2 - 4
    pts_num //= self.downsample[1]
    # output = self.lstm(fea2.permute(0, 2, 1, 3))

    # fea3 = output[0][0].squeeze(-1).permute(0, 2, 1, 3)

    ret_group_array3 = self.group.st_group_points(fea2, 3, [0, 1, 2], self.knn[2], 3)
    ret_array3, inputs, ind = self.select_ind(ret_group_array3, inputs,
                                              batchsize, in_dims, timestep, pts_num)

    fea3 = self.pool3(self.stage3(ret_array3)).view(batchsize, -1, timestep, pts_num)

    # fea3 = fea3.gather(-1, ind.unsqueeze(1).expand(-1, fea3.shape[1], -1, -1))
    fea3 = torch.cat((inputs, fea3), dim=1)
    print('fea3===', fea3.shape)
    # stage 4: inter-frame, late
    in_dims = fea3.shape[1] * 2 - 4
    pts_num //= self.downsample[2]
    output = self.lstm(fea3.permute(0, 2, 1, 3))

    fea4 = output[0][0].squeeze(-1).permute(0, 2, 1, 3)
    print('lstm333=', fea4.shape)
    ret_group_array4 = self.group.st_group_points(fea3, 3, [0, 1, 2], self.knn[3], 3)
    ret_array4, inputs, _ = self.select_ind(ret_group_array4, inputs,
                                            batchsize, in_dims, timestep, pts_num)

    # fea4 = self.pool4(self.stage4(ret_array4)).view(batchsize, -1, timestep, pts_num)
    fea4 = fea4.gather(-1, _.unsqueeze(1).expand(-1, fea4.shape[1], -1, -1))
    print('outfea4=', fea4.shape)

[ycmin95]

Hi, it has been a long time from this experiment, and I remember that we select this dataset to show the generalization and do not perform ablation study on this dataset. I do not know how the changes of frames affect the performance on this dataset. I find the relevant logs, and you can find your information. Another difference is that we set offset=True for MSR Action 3D experiments, the motion information can help recognition.

For train subjet[1, 2, 3, 4, 5] test [6, 7, 8, 9, 10]): train.txt log.txt

For train subjet[6, 7, 8, 9, 10]) test[1, 2, 3, 4, 5]: train.txt log.txt

[weiyutao886]

thank you, i will try it again.

[weiyutao886]

Hello, I conducted the experiment again according to your parameters, but the effect is not ideal. As follows: I only conducted the experiment in train subjet [1, 2, 3, 4, 5] test [6, 7, 8, 9, 10]) [ Wed Apr 6 21:59:49 2022 ] Epoch 170, Test, Evaluation: prec1 87.9725, prec5 98.2818 best result is 87.9 your best result is 93 My parameters are as follows [ Thu Apr 7 22:00:00 2022 ] Parameters: {'work_dir': 'result19', 'config': 'pointlstm.yaml', 'device': '0', 'phase': 'train', 'random_fix': True, 'random_seed': 0, 'save_interval': 5, 'eval_interval': 5, 'print_log': True, 'log_interval': 50, 'dataloader': 'data load1.SHRECLoader', 'num_worker': 0, 'framesize': 32, 'pts_size': 128, 'train_loader_args': {'phase': 'train', 'framerate': 32}, 'test_loader_args': {'phase': 'test', 'framerate': 32}, 'valid_loader_args': {}, 'model': 'models.motion10.Motion', 'model_args': {'pts_size': 128, 'num_classes': 20, 'knn': [16, 24, 48, 12], 'offsets': True, 'topk': 16}, 'weights': None, 'ignore_weights': [], 'batch_size': 8, 'test_batch_size': 8, 'optimizer_args': {'optimizer': 'Adam', 'base_lr': 0.0001, 'step': [100, 160, 180], 'weight_decay': 0.005, 'start_epoch': 0, 'nesterov': False}, 'num_epoch': 300}

I wonder if there is a problem with data processing or my model. What do you think

[ycmin95]

Sorry for late reply, you can visualize the point cloud sequence at different stages, which should be likely to Figure 4. If you will, I can send the source data processing code to you via email, and you can create a PR after you reimplement the results.

[weiyutao886]

This is my dataloader and data processing code. You are welcome to correct my mistakes. I would be grateful if I could use your code for reference. And this is my email [935628178@qq.com](mailto:935628178@qq.com msraction_process.zip data load1.zip )

[weiyutao886]

我通过您发的trian的日志文件，发现您报告的准确率不是里面最高的，请问您每个subject取准确率的标准是什么

[ Thu Mar 5 23:16:02 2020 ] Epoch 110, Test Evaluation: prec1 93.4708, prec5 98.6254 [ Thu Mar 5 23:18:14 2020 ] Epoch 115, Test Evaluation: prec1 92.7835, prec5 98.6254 [ Thu Mar 5 23:20:27 2020 ] Epoch 120, Test Evaluation: prec1 92.0962, prec5 98.6254 [ Thu Mar 5 23:22:39 2020 ] Epoch 125, Test Evaluation: prec1 93.1272, prec5 98.6254 [ Thu Mar 5 23:24:52 2020 ] Epoch 130, Test Evaluation: prec1 93.1272, prec5 98.6254 [ Thu Mar 5 23:27:04 2020 ] Epoch 135, Test Evaluation: prec1 93.4708, prec5 98.6254 这里最高的准确率是93.4，但是您报告的准确率是93.1

[ Thu Mar 5 23:29:19 2020 ] Epoch 130, Test Evaluation: prec1 87.5940, prec5 99.6241 [ Thu Mar 5 23:31:41 2020 ] Epoch 135, Test Evaluation: prec1 87.5940, prec5 99.6241 [ Thu Mar 5 23:34:04 2020 ] Epoch 140, Test Evaluation: prec1 88.3459, prec5 99.6241 [ Thu Mar 5 23:36:27 2020 ] Epoch 145, Test Evaluation: prec1 88.3459, prec5 99.2481 [ Thu Mar 5 23:38:50 2020 ] Epoch 150, Test Evaluation: prec1 87.9699, prec5 99.6241 [ Thu Mar 5 23:41:13 2020 ] Epoch 155, Test Evaluation: prec1 89.0977, prec5 99.6241 [ Thu Mar 5 23:43:35 2020 ] Epoch 160, Test Evaluation: prec1 87.9699, prec5 98.8722 [ Thu Mar 5 23:45:58 2020 ] Epoch 165, Test Evaluation: prec1 87.2180, prec5 98.8722 [ Thu Mar 5 23:48:21 2020 ] Epoch 170, Test Evaluation: prec1 87.2180, prec5 99.2481 这里最高的准确率是89.1，而您报告的是87.59，再次麻烦您非常抱歉。

[ycmin95]

The accuracy of the last epoch.

[weiyutao886]

This is my dataloader and data processing code. You are welcome to correct my mistakes. I would be grateful if I could use your code for reference. And this is my email  @.***

------------------ 原始邮件 ------------------ 发件人: "ycmin95/pointlstm-gesture-recognition-pytorch" @.>; 发送时间: 2022年4月11日(星期一) 上午10:12 @.>; @.>;"State @.>; 主题: Re: [ycmin95/pointlstm-gesture-recognition-pytorch] msr action (Issue #23)

Sorry for late reply, you can visualize the point cloud sequence at different stages, which should be likely to Figure 4. If you will, I can send the source data processing code to you via email, and you can create a PR after you reimplement the results.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you modified the open/close state.Message ID: @.***>