File size: 7,065 Bytes
f44b9a6 bbc3095 f44b9a6 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 | # =============================================================================
# Cell 1: Install dependencies
# =============================================================================
"""
try:
!pip uninstall -qy sd15-flow-trainer[dev]
except:
pass
!pip install "sd15-flow-trainer[dev] @ git+https://github.com/AbstractEyes/sd15-flow-trainer.git" -q
"""
# =============================================================================
# Cell 2: Pre-encode 50k object-relations dataset
# =============================================================================
import torch
import gc, os
from sd15_trainer_geo.pipeline import load_pipeline
pipe = load_pipeline(device="cuda", dtype=torch.float16)
from sd15_trainer_geo.trainer import pre_encode_hf_dataset
CACHE_PATH = "/content/latent_cache/object_relations_schnell_512_2.pt"
pre_encode_hf_dataset(
pipe,
dataset_name="AbstractPhil/synthetic-object-relations",
subset="schnell_512_2",
split="train",
image_column="image",
prompt_column="prompt",
output_path=CACHE_PATH,
image_size=512,
batch_size=16,
max_samples=50_000,
)
del pipe.vae, pipe.clip
gc.collect()
torch.cuda.empty_cache()
print(f"VRAM after encoding cleanup: {torch.cuda.memory_allocated()/1e9:.1f} GB")
# =============================================================================
# Cell 3: Load pipeline + Lune UNet, baseline samples
# =============================================================================
from sd15_trainer_geo.pipeline import load_pipeline
from sd15_trainer_geo.generate import generate, save_images, show_images
pipe = load_pipeline(device="cuda", dtype=torch.float16)
pipe.unet.load_pretrained(
"AbstractPhil/tinyflux-experts",
subfolder="",
filename="sd15-flow-lune-unet.safetensors",
)
spatial_prompts = [
"a red cup on top of a blue book",
"a cat sitting beside a vase of flowers",
"a small ball inside a glass bowl on a table",
"a pair of shoes next to an umbrella by the door",
]
novel_prompts = [
"a guitar leaning against a piano in a dim room",
"three candles arranged in a triangle on a wooden tray",
"a telescope pointed at the moon through an open window",
"a child's drawing pinned to a refrigerator with magnets",
]
print("=" * 60)
print("BASELINE (before geo_prior training)")
print("=" * 60)
baseline_spatial = generate(pipe, spatial_prompts, shift=2.5, seed=42, num_steps=30)
save_images(baseline_spatial, "/content/samples_baseline_spatial")
baseline_novel = generate(pipe, novel_prompts, shift=2.5, seed=42, num_steps=30)
save_images(baseline_novel, "/content/samples_baseline_novel")
show_images(baseline_spatial)
show_images(baseline_novel)
# =============================================================================
# Cell 4: Train geo_prior on 50k object-relations
# =============================================================================
from sd15_trainer_geo.trainer import Trainer, TrainConfig, LatentDataset
config = TrainConfig(
num_steps=8333,
batch_size=6,
base_lr=5e-5,
min_lr=1e-6,
lr_scheduler="cosine",
warmup_steps=200,
# Flow matching
shift=2.5,
cfg_dropout=0.1,
min_snr_gamma=5.0,
# Geometric regularization
geo_loss_weight=0.01,
geo_loss_warmup=400,
# Logging
log_every=100,
sample_every=2000,
save_every=2000,
sample_prompts=spatial_prompts[:2] + novel_prompts[:2],
seed=42,
output_dir="/content/geo_prior_object_relations",
)
dataset = LatentDataset(CACHE_PATH)
trainer = Trainer(pipe, config)
trainer.fit(dataset)
# =============================================================================
# Cell 5: Push trained weights to hub
# =============================================================================
from sd15_trainer_geo.pipeline import push_geo_to_hub
push_geo_to_hub(
pipe,
repo_id="AbstractPhil/sd15-geoflow-object-association",
base_repo="sd-legacy/stable-diffusion-v1-5",
commit_message="geo_prior v1: 1 epoch 50k object-relations schnell_512_2",
extra={
"dataset": "AbstractPhil/synthetic-object-relations (schnell_512_2)",
"samples": 50000,
"epochs": 1,
"steps": 8333,
"shift": 2.5,
"base_lr": 5e-5,
"min_snr_gamma": 5.0,
"cfg_dropout": 0.1,
"batch_size": 6,
"geo_loss_weight": 0.01,
"loss_final": trainer.log_history[-1]["loss"] if trainer.log_history else "n/a",
},
)
# =============================================================================
# Cell 6: Compare before/after
# =============================================================================
print("=" * 60)
print("AFTER TRAINING — Spatial Prompts (in-distribution)")
print("=" * 60)
trained_spatial = generate(pipe, spatial_prompts, shift=2.5, seed=42, num_steps=30)
save_images(trained_spatial, "/content/samples_trained_spatial")
show_images(trained_spatial)
print("=" * 60)
print("AFTER TRAINING — Novel Prompts (out-of-distribution)")
print("=" * 60)
trained_novel = generate(pipe, novel_prompts, shift=2.5, seed=42, num_steps=30)
save_images(trained_novel, "/content/samples_trained_novel")
show_images(trained_novel)
hard_spatial = [
"a book on top of a cup",
"a lamp beneath a table",
"a knife to the left of a fork on a plate",
"a hat resting on a basketball",
"a key inside a shoe next to the door",
"a red apple behind a green bottle",
]
print("=" * 60)
print("HARD SPATIAL (never seen, complex relations)")
print("=" * 60)
hard_out = generate(pipe, hard_spatial, shift=2.5, seed=42, num_steps=30)
save_images(hard_out, "/content/samples_hard_spatial")
show_images(hard_out)
# =============================================================================
# Cell 7: Training summary
# =============================================================================
print("\n" + "=" * 60)
print("TRAINING SUMMARY")
print("=" * 60)
if trainer.log_history:
first = trainer.log_history[0]
last = trainer.log_history[-1]
mid = trainer.log_history[len(trainer.log_history) // 2]
print(f"Steps: {last.get('step', config.num_steps)}")
print(f"Loss (start): {first['loss']:.4f}")
print(f"Loss (mid): {mid['loss']:.4f}")
print(f"Loss (final): {last['loss']:.4f}")
print(f"Task (final): {last.get('task_loss', 'n/a')}")
print(f"Geo (final): {last.get('geo_loss', 'n/a')}")
stats = pipe.unet.get_geometry_stats()
if stats:
print(f"\nGeometry:")
print(f" Blend: {stats.get('blend', 'n/a')}")
for i in range(4):
vol = stats.get(f'layer_{i}/vol_sq', 'n/a')
ent = stats.get(f'layer_{i}/entropy', 'n/a')
ds = stats.get(f'layer_{i}/deform_scale', 'n/a')
if isinstance(vol, float):
print(f" Layer {i}: vol²={vol:.4e}, entropy={ent:.2f}, δ={ds:.4f}")
print(f"\nCheckpoints: /content/geo_prior_object_relations/")
print(f"Hub: https://huggingface.co/AbstractPhil/sd15-geoflow-object-association") |