Upload configs/vae_pretrain.yaml
Browse files- configs/vae_pretrain.yaml +58 -0
configs/vae_pretrain.yaml
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Stage 1: SLAT-Interior VAE Pre-training
|
| 2 |
+
|
| 3 |
+
model:
|
| 4 |
+
name: "SLAT-Interior-VAE"
|
| 5 |
+
latent_dim: 64
|
| 6 |
+
base_resolution: 256
|
| 7 |
+
target_resolution: 1024
|
| 8 |
+
channels: 64
|
| 9 |
+
num_layers: 8
|
| 10 |
+
|
| 11 |
+
optimizer:
|
| 12 |
+
type: AdamW
|
| 13 |
+
lr: 1.0e-4
|
| 14 |
+
weight_decay: 0.01
|
| 15 |
+
betas: [0.9, 0.999]
|
| 16 |
+
|
| 17 |
+
scheduler:
|
| 18 |
+
type: cosine_with_restarts
|
| 19 |
+
warmup_steps: 10000
|
| 20 |
+
|
| 21 |
+
training:
|
| 22 |
+
batch_size: 8 # per GPU
|
| 23 |
+
num_gpus: 8
|
| 24 |
+
effective_batch_size: 64
|
| 25 |
+
gradient_accumulation: 1
|
| 26 |
+
max_steps: 200000
|
| 27 |
+
mixed_precision: bf16
|
| 28 |
+
save_every: 5000
|
| 29 |
+
log_every: 100
|
| 30 |
+
|
| 31 |
+
curriculum:
|
| 32 |
+
- resolution: 256
|
| 33 |
+
steps: 50000
|
| 34 |
+
lr: 1.0e-4
|
| 35 |
+
- resolution: 512
|
| 36 |
+
steps: 100000
|
| 37 |
+
lr: 1.0e-4
|
| 38 |
+
- resolution: 1024
|
| 39 |
+
steps: 50000
|
| 40 |
+
lr: 5.0e-5
|
| 41 |
+
|
| 42 |
+
data:
|
| 43 |
+
dataset: "InteriorFusion-Train"
|
| 44 |
+
num_workers: 8
|
| 45 |
+
pin_memory: true
|
| 46 |
+
|
| 47 |
+
loss:
|
| 48 |
+
reconstruction:
|
| 49 |
+
weight: 1.0
|
| 50 |
+
type: l1
|
| 51 |
+
kl_divergence:
|
| 52 |
+
weight: 1.0e-3
|
| 53 |
+
depth_consistency:
|
| 54 |
+
weight: 0.5
|
| 55 |
+
type: l1
|
| 56 |
+
normal_consistency:
|
| 57 |
+
weight: 0.3
|
| 58 |
+
type: cosine
|