Any extra parameters or losses are used during 3-s pretraining and downstream?

The cluster diagram in your readme is very attractive. Is there a corresponding part of the code that can draw this class diagram

I am very interested in your work and thanks for your sharing. I'd like to know how to finetune and semi-supervised evaluate AimCLR? Have the two evaluation been included in your code? If not , could you please share them for me? My email is :[email protected]

when i run:
$ python main.py linear_evaluation --config config/ntu60/linear_eval/linear_eval_aimclr_xview_joint.yaml
out:
Traceback (most recent call last):
File "main.py", line 50, in
p = Processor(sys.argv[2:])
File "/home/xxx/project1/AimCLR/processor/processor.py", line 43, in init
self.init_environment()
File "/home/xxx/project1/AimCLR/processor/processor.py", line 76, in init_environment
super().init_environment()
File "/home/xxx/project1/AimCLR/processor/io.py", line 55, in init_environment
self.io.save_arg(self.arg)
File "/home/xxx/anaconda3/envs/TEST2/lib/python3.7/site-packages/torchlight-1.0-py3.7.egg/torchlight/io.py", line 116, in save_arg
File "/home/xxx/anaconda3/envs/TEST2/lib/python3.7/os.py", line 211, in makedirs
makedirs(head, exist_ok=exist_ok)
File "/home/xxx/anaconda3/envs/TEST2/lib/python3.7/os.py", line 211, in makedirs
makedirs(head, exist_ok=exist_ok)
File "/home/xxx/anaconda3/envs/TEST2/lib/python3.7/os.py", line 211, in makedirs
makedirs(head, exist_ok=exist_ok)
[Previous line repeated 1 more time]
File "/home/xxx/anaconda3/envs/TEST2/lib/python3.7/os.py", line 221, in makedirs
mkdir(name, mode)
PermissionError: [Errno 13] Permission denied: '/data/gty'

my .yaml:
work_dir: /data/gty/AAAI_github/ntu60_cv/aimclr_joint/linear_eval

weights: /data/gty/released_model/ntu60_xview_joint.pt

weights: /data/gty/AAAI_github/ntu60_cv/aimclr_joint/pretext/epoch300_model.pt

ignore_weights: [encoder_q.fc, encoder_k, queue]

feeder

train_feeder: feeder.ntu_feeder.Feeder_single
train_feeder_args:
data_path: /home/wuyushan/project1/data/NTU60_frame50/xview/train_position.npy
label_path: /home/wuyushan/project1/data/NTU60_frame50/xview/train_label.pkl
shear_amplitude: -1
temperal_padding_ratio: -1
mmap: True
test_feeder: feeder.ntu_feeder.Feeder_single
test_feeder_args:
data_path: /home/xxx/project1/data/NTU60_frame50/xview/val_position.npy
label_path: /home/xxx/project1/data/NTU60_frame50/xview/val_label.pkl
shear_amplitude: -1
temperal_padding_ratio: -1
mmap: True

model

model: net.aimclr.AimCLR
model_args:
base_encoder: net.st_gcn.Model
pretrain: False

feature_dim: 128

queue_size: 32768

momentum: 0.999

Temperature: 0.07

mlp: True

in_channels: 3
hidden_channels: 16
hidden_dim: 256
num_class: 60
dropout: 0.5
graph_args:
layout: 'ntu-rgb+d'
strategy: 'spatial'
edge_importance_weighting: True

optim

nesterov: False
weight_decay: 0.0
base_lr: 3.
optimizer: SGD
step: [80]

training

device: [1] # 3
batch_size: 64 # 128
test_batch_size: 64 # 128
num_epoch: 100
stream: 'joint'

log

save_interval: -1
eval_interval: 5

i don't know why,can you help me,please!

Your paper has classified as "Notably, the KNN classifier does not require learning extra weights compared with the
linear classifier". What is the "extra weights"？ Can you explain it clearly?

[04.26.22|13:33:35] Can not find weights [encoder_q.fc.bias].
[04.26.22|13:33:35] Can not find weights [encoder_q.fc.weight].
Traceback (most recent call last):
File "main.py", line 50, in
p = Processor(sys.argv[2:])
File "/home/lailai/lailai_file/action/AimCLR/processor/processor.py", line 60, in init
self.load_data()
File "/home/lailai/lailai_file/action/AimCLR/processor/processor.py", line 92, in load_data
dataset=train_feeder(**self.arg.train_feeder_args),
File "/home/lailai/lailai_file/action/AimCLR/feeder/ntu_feeder.py", line 17, in init
self.load_data(mmap)
File "/home/lailai/lailai_file/action/AimCLR/feeder/ntu_feeder.py", line 22, in load_data
self.sample_name, self.label = pickle.load(f)
_pickle.UnpicklingError: STACK_GLOBAL requires str

Hello, for the papers,“For all the reported results of three streams, we use the weights of [0.6, 0.6, 0.4] for weighted fusion like other multi-stream GCN methods.”
But I still don’t understand, can you teach me how to calculate the comprehensive results of the three streams?
For example, how did the 83.8 in the last line of the fourth column in Table II calculated?

Hi, Thank you for sharing your work! You used KNN evaluation method in your paper, but I did not find the related implementation details in the article, nor did I find the corresponding code. Would please provide some more details of the KNN evaluation? And as to the finetune setting, did you adjust the weight decay when doing finetune? Looking forward to your reply

Thanks for your time, I encounter some problems using Google Drive to download PKUMMD dataset. It can help a lot if both parts of the datasets can be shared on BaiduYun if it is convenient for you, I appreciate it a lot for your kindness.

By the way, is the preprocessing and experiment settings same for part1 and part2? Looking forward to your reply.

How to fine-tune the model after the linear evaluation is completed? I don't see the corresponding yaml file in the project

Hello, thank you for presenting your excellent work!

While pretraining aimclr, how do you measure the progress with respect to the loss function train_mean_loss, what is a good value demonstrating that we can stop the aimclr pretraining to yield a good encoder?

Thanks

Hi, thank you for sharing your work with the community!
I was intersted in training a model using 150 frames long sequences, but after preprocessing the NTU dataset I ended up with just .npy files, while .pkl are needed for the label. Could you please share the code to convert the .npy files into .pkl?
Thank you,
Paolo

Hello, what is the initial value of the learning rate set for fine tuning?