Hyperparameter Setting

[Forgetmmmm]

Hi,when I use the dataset hugadb I set bz=8 and sample_rate=4,but it does not reach the results of the paper, and it is much worse than the results of the paper.May I ask how you set these parameters?Looking forward to your reply. Thank you.

[BenjaminFiltjens]

Hi, we do not downsample hugadb (so sr=1). However, regardless of the batch size and downsampling factor (within reasonable settings ofcourse), you should get f1@50>90 for hugadb.

We use the raw acc and gyro signals so be sure to comment out get_features in batch_gen and model.predict. Hope that helps.

[Forgetmmmm]

Thank for your reply.I don't quite understand the Relationship between sr and downsampling factor.Hugadb's sampling rate in the paper is 60HZ,why set sr=1 in the code?How do I determine which datasets need to be downsampled and which datasets do not.When I use the datasets lara should I set sr=4?Looking forward to your reply. Thank you.

[BenjaminFiltjens]

Table 1 specifies the used sampling rate per dataset. I believe Hugadb IMU's were sampled at 60hz, we use 60hz. Hence, no need to downsample (sr=1). I believe motion capture in lara was sampled at 200hz , we use 50hz. Hence, downsampled by a factor of 4 (sr=4).

[Forgetmmmm]

May I ask what tp-vicon means?When I use the datasets pku-mmd and tug,there shows do not exist these layout.

[BenjaminFiltjens]

That is the 9 node graph for the FOG (proprietary) dataset. TUG and lara use the same graph (19 nodes, lara should change from 17 to 19 nodes in our pre-print). If you do not have the graph for pku-mmd then you are using an older version.

[Forgetmmmm]

Hello,I downloaded the latest codeI ,but I don't see layout= 'tug' in graph.py.When I use the dataset tug the code will report an error.

[BenjaminFiltjens]

Same graph for lara and tug

[Forgetmmmm]

I'm sorry to bother you again.When I use the dataset tug I have the following problem:
ValueError:operands could not be broadcast together with shapes(3,1830,1)(3,1830,19,1)
I used the same parameters as I used to train LARA,I'm not quite sure what the problem is.

[BenjaminFiltjens]

No problem. I will re-open the issue and have a look this weekend.

[BenjaminFiltjens]

I cannot reproduce the problem with the TUG dataset. Could you give more information?

[Forgetmmmm]

I made the following changes in the code:
1)There is no 'tug' in the code. I changed "elif layout =='lara'" to "elif layout==' lara 'or layout=='tug'"at line 56 of graph.py.
2)Set sample_rate=4 in the main.py.
3)Set the tensor dimensions in batch_gen:batch_input_tensor = torch.zeros(len(batch_input), 6, max(length_of_sequences), 19, 1, dtype=torch.float).
4)Set layout='tug' in ms_gcn.py.
Then when I run main.py,An error message is displayed:
Traceback (most recent call last):
File "/media/hz/A2/cmz/code/MS-GCN-main/main.py", line 63, in
trainer.train(model_dir, batch_gen, num_epochs=num_epochs, batch_size=bz, learning_rate=lr, device=device)
File "/media/hz/A2/cmz/code/MS-GCN-main/model.py", line 33, in train
batch_input, batch_target, mask, weight = batch_gen.next_batch(batch_size)
File "/media/hz/A2/cmz/code/MS-GCN-main/batch_gen.py", line 50, in next_batch
features = get_features(features)
File "/media/hz/A2/cmz/code/MS-GCN-main/batch_gen.py", line 10, in get_features
rel_coords = get_relative_coordinates(sample, references=(0))
File "/media/hz/A2/cmz/code/MS-GCN-main/data/signals/rel_coords.py", line 21, in get_relative_coordinates
ref_loc = sample[:, :, references, :]//sample
ValueError: operands could not be broadcast together with shapes (3,1830,1) (3,1830,19,1)

[BenjaminFiltjens]

Your settings are fine. I cannot reproduce the problem in "get_relative_coordinates".