sgoel30 commited on
Commit
3340437
·
verified ·
1 Parent(s): 9b7a4b0

Upload config.py

Browse files
Files changed (1) hide show
  1. configs/config.py +63 -19
configs/config.py CHANGED
@@ -1,19 +1,63 @@
1
- ### configs/config.py
2
-
3
- ```python
4
- class Config:
5
- model_name = "facebook/esm2_t33_650M_UR50D"
6
- latent_dim = 1280 # Adjust based on ESM-2 latent dimension
7
- optim = {"lr": 1e-4}
8
- training = {
9
- "ema": 0.999,
10
- "epochs": 10,
11
- "batch_size": 32,
12
- "gpus": 8,
13
- "precision": 16, # Mixed precision training
14
- "accumulate_grad_batches": 2, # Gradient accumulation
15
- "save_dir": "./checkpoints/",
16
- }
17
- data_path = "./data/"
18
- T = 1000 # Number of diffusion steps
19
- subs_masking = False
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MODE = "train" # train, ppl_eval, sample_eval
2
+ DIFFUSION = "absorbing_state"
3
+ BACKBONE = "dit"
4
+ PARAMETERIZATION = "subs" # subs, d3pm, sedd
5
+ TIME_CONDITIONING = False
6
+ T = 1000
7
+ SUBS_MASKING = False
8
+ LATENT_DIM = 1280
9
+ MODEL_NAME = 'facebook/esm2_t33_650M_UR50D'
10
+
11
+ SEED = 1
12
+
13
+ class Loader:
14
+ BATCH_SIZE = 8
15
+ DATA_PATH = "/workspace/a03-sgoel/MDpLM/data"
16
+
17
+ class Sampling:
18
+ PREDICTOR = "ddpm_cache" # analytic, ddpm, ddpm_cache (recommended)
19
+ STEPS = 128
20
+ NOISE_REMOVAL = True
21
+ STRIDE_LENGTH = 1
22
+ NUM_STRIDES = 1
23
+ NUM_SAMPLING_BATCHES = 2
24
+
25
+ class model:
26
+ hidden_size = 1280
27
+ cond_dim = 1280
28
+ n_heads = 8
29
+ n_blocks = 2
30
+ dropout = 0.5
31
+ length = 512
32
+
33
+ class Training:
34
+ EMA = 0.9999
35
+ ANTITHETIC_SAMPLING = True
36
+ SAMPLING_EPS = 1e-3
37
+ ACCUMULATE_GRAD_BATCHES = 2
38
+ NUM_EPOCHS = 10
39
+ GRADIENT_CLIP_VAL = 1.0
40
+ PRECISION = 'bf16'
41
+ MAX_STEPS = 1,000,000
42
+ LOG_EVERY_N_STEPS: 10
43
+ GPUS = 2
44
+ SAVE_DIR = "/workspace/a03-sgoel/MDpLM/models"
45
+
46
+ class Eval:
47
+ CHECKPOINT_PATH = "/workspace/a03-sgoel/MDpLM/checkpoints"
48
+ DISABLE_EMA = False
49
+ COMPUTE_GENERATIVE_PERPLEXITY = False
50
+ PERPLEXITY_BATCH_SIZE = 8
51
+ GENERATE_SAMPLES = True
52
+
53
+ class Optim:
54
+ LR = 1e-4
55
+ BETA1 = 0.9
56
+ BETA2 = 0.999
57
+ EPS = 1e-8
58
+
59
+
60
+ # finish this when training actually works
61
+ class Wandb:
62
+ PROJECT = "MDpLM_shrey_test"
63
+ GROUP = "programmablebio"