foldtree2/example_config.yaml at main · DessimozLab/foldtree2 · GitHub

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
# Example configuration file for learn_monodecoder.py
# This file contains all training hyperparameters and settings
# To use: python learn_monodecoder.py --config example_config.yaml
# Command-line arguments will override values in this file

# Dataset and data loading
dataset: structs_traininffttest.h5
batch_size: 20

# Model architecture
hidden_size: 256
embedding_dim: 128
num_embeddings: 40
se3_transformer: false
hetero_gae: false

# Training parameters
epochs: 100
learning_rate: 0.0001
clip_grad: false
burn_in: 0
EMA: false

# Learning rate scheduling
lr_schedule: plateau  # Options: plateau, cosine, linear, cosine_restarts, polynomial, none
lr_warmup_steps: 0    # Number of steps for linear warmup (0 = no warmup)
lr_warmup_ratio: 0.0  # Alternative: warmup as fraction of total steps (overrides lr_warmup_steps if > 0)
lr_min: 1.0e-06       # Minimum LR for cosine/linear/polynomial schedules
num_cycles: 3         # Number of cycles for cosine_restarts scheduler

# Gradient accumulation
gradient_accumulation_steps: 1  # Accumulate gradients over N steps (1 = no accumulation)
                                # Effective batch size = batch_size * gradient_accumulation_steps

# Commitment cost scheduling (for VQ-VAE encoder)
commitment_cost: 0.9  # Final commitment cost value
use_commitment_scheduling: false  # Enable commitment cost warmup from low to final value
commitment_schedule: cosine  # Options: cosine, linear, none
commitment_warmup_steps: 5000  # Number of steps to warmup commitment cost
commitment_start: 0.1  # Starting commitment cost when using scheduling

# Output options
output_fft: false
output_rt: false
output_foldx: false

# Directories and naming
output_dir: ./models/
model_name: monodecoder_model
tensorboard_dir: ./runs/
run_name: null  # Set to null for auto-generated timestamp name

# System settings
device: null  # Set to null for auto-detection, or specify like "cuda:0"
seed: 0

# Model persistence
overwrite: false  # Set to true to overwrite existing models

# Note: You can save your current command-line configuration to a file using:
# python learn_monodecoder.py [your args] --save-config my_config.yaml