-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtest.yaml
107 lines (84 loc) · 4.46 KB
/
test.yaml
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
### Experiment configuration
## General settings
split: 'test' # data split; choices ['train', 'test']
debug: false # if true, load only a few data samples
seed: 999
## Computational resources
accelerator: 'gpu'
devices: [0,1] # indices of cuda devices to use
## Paths
dir_name: 'train_experiment' # name of the directory of the current experiment
data_dir: './sktDD' # path to the data
exp_dir: './checkpoints' # path to the directory that will contain the current experiment directory
test_path: './sktDD/testing/test_frame_mask' # path to the test data
load_ckpt: 'epoch=80-step=34587.ckpt' # name of the checkpoint to load at inference time
create_experiment_dir: false # if true, create a new directory for the current experiment
## WANDB configuration
use_wandb: false
project_name: "project_name"
wandb_entity: "entity_name"
group_name: "group_name"
use_ema: false
##############################
### Model's configuration
## U-Net's configuration
embedding_dim: 16 # dimension of the embedding of the UNet
dropout: 0. # probability of dropout
conditioning_strategy: 'inject' # choices ['inject' (add2layers), 'concat' (cat), 'inbetween_imp' (interleave), 'random_imp' (random_indices), 'no_condition' (none)]
## Conditioning network's configuration
conditioning_architecture: 'AE' # choices ['AE', 'E', 'E_unet']
conditioning_indices: [0,1,2] # If conditioning_strategy=random_imp, it must be int and it is used as the number of random indices that will be selected
# If an int is given and conditioning_strategy=[inject|concat], n_frames//conditioning_indices will be used as the number of conditioning indices
# If a list is given, it will be used as the conditioning indices
h_dim: 32 # dimension of the bottleneck at the end of the encoder of the conditioning network
latent_dim: 16 # dimension of the latent space of the conditioning encoder
channels: [32,16,32] # channels for the encoder
##############################
### Training's configuration
## Diffusion's configuration
noise_steps: 10 # how many diffusion steps to perform
### Optimizer and scheduler's configuration
n_epochs: 100
opt_lr: 0.001
## Losses' configuration
loss_fn: 'smooth_l1' # loss function; choices ['mse', 'l1', 'smooth_l1']
rec_weight: 0.01 # weight of the reconstruction loss
##############################
### Inference's configuration
n_generated_samples: 50 # number of samples to generate
model_return_value: 'loss' # choices ['loss', 'poses', 'all']; if 'loss', the model will return the loss;
# if 'poses', the model will return the generated poses;
# if 'all', the model will return both the loss and the generated poses
aggregation_strategy: 'best' # choices ['best', 'mean', 'median', 'random']; if 'best', the best sample will be selected;
# if 'mean', the mean of loss of the samples will be selected;
# if 'median', the median of the loss of the samples will be selected;
# if 'random', a random sample will be selected;
# if 'mean_pose', the mean of the generated poses will be selected;
# if 'median_pose', the median of the generated poses will be selected;
# if 'all', all the generated poses will be selected
filter_kernel_size: 30 # size of the kernel to use for smoothing the anomaly score of each clip
frames_shift: 6 # it compensates the shift of the anomaly score due to the sliding window;
# in conjuction with pad_size and filter_kernel_size, it strongly depends on the dataset
save_tensors: false # if true, save the generated tensors for faster inference
load_tensors: false # if true, load the generated tensors for faster inference
##############################
### Dataset's configuration
## Important parameters
dataset_choice: 'sktDD'
seg_len: 4 #length of the window (cond+noised)
vid_res: [1920,1080]
batch_size: 1024
pad_size: 12 # size of the padding
## Other parameters
headless: false # remove the keypoints of the head
hip_center: false # center the keypoints on the hip
kp18_format: false # use the 18 keypoints format
normalization_strategy: 'robust' # use 'none' to avoid normalization, 'robust' otherwise
num_coords: 2 # number of coordinates to use
num_transform: 5 # number of transformations to apply
num_workers: 8
seg_stride: 1
seg_th: 0
start_offset: 0
symm_range: true
use_fitted_scaler: false