Spaces:
Running on Zero
Running on Zero
Upload tools/tools.py with huggingface_hub
Browse files- tools/tools.py +394 -0
tools/tools.py
ADDED
|
@@ -0,0 +1,394 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os, yaml, random
|
| 2 |
+
import torch
|
| 3 |
+
import numpy as np
|
| 4 |
+
from typing import Union
|
| 5 |
+
import pickle
|
| 6 |
+
from diffusers.pipelines.flux.pipeline_flux import calculate_shift, retrieve_timesteps
|
| 7 |
+
from peft import LoraConfig
|
| 8 |
+
from peft.utils import get_peft_model_state_dict, set_peft_model_state_dict
|
| 9 |
+
|
| 10 |
+
from models.mmdit import CustomFluxTransformer2DModel
|
| 11 |
+
from models.pipeline import CustomFluxPipeline
|
| 12 |
+
from models.multiLayer_adapter import MultiLayerAdapter
|
| 13 |
+
|
| 14 |
+
def save_checkpoint(transformer, multiLayer_adater, optimizer, optimizer_adapter, scheduler, scheduler_adapter, step, save_dir):
|
| 15 |
+
import gc
|
| 16 |
+
|
| 17 |
+
trans_dir = os.path.join(save_dir, "transformer")
|
| 18 |
+
adapter_dir = os.path.join(save_dir, "adapter")
|
| 19 |
+
os.makedirs(trans_dir, exist_ok=True)
|
| 20 |
+
os.makedirs(adapter_dir, exist_ok=True)
|
| 21 |
+
|
| 22 |
+
# Get state dicts and IMMEDIATELY move to CPU to avoid GPU memory buildup
|
| 23 |
+
flux_transformer_lora_state_dict = get_peft_model_state_dict(transformer)
|
| 24 |
+
flux_transformer_lora_state_dict = {k: v.detach().cpu().to(torch.float32) for k, v in flux_transformer_lora_state_dict.items()}
|
| 25 |
+
|
| 26 |
+
flux_adapter_lora_state_dict = get_peft_model_state_dict(multiLayer_adater)
|
| 27 |
+
flux_adapter_lora_state_dict = {k: v.detach().cpu().to(torch.float32) for k, v in flux_adapter_lora_state_dict.items()}
|
| 28 |
+
|
| 29 |
+
CustomFluxPipeline.save_lora_weights(
|
| 30 |
+
os.path.join(trans_dir),
|
| 31 |
+
flux_transformer_lora_state_dict,
|
| 32 |
+
safe_serialization=True,
|
| 33 |
+
)
|
| 34 |
+
# Clear after saving
|
| 35 |
+
del flux_transformer_lora_state_dict
|
| 36 |
+
|
| 37 |
+
CustomFluxPipeline.save_lora_weights(
|
| 38 |
+
os.path.join(adapter_dir),
|
| 39 |
+
flux_adapter_lora_state_dict,
|
| 40 |
+
safe_serialization=True,
|
| 41 |
+
)
|
| 42 |
+
# Clear after saving
|
| 43 |
+
del flux_adapter_lora_state_dict
|
| 44 |
+
|
| 45 |
+
torch.save({"layer_pe": transformer.layer_pe.detach().cpu().to(torch.float32)}, os.path.join(save_dir, "layer_pe.pth"))
|
| 46 |
+
|
| 47 |
+
torch.save(optimizer.state_dict(), os.path.join(trans_dir, "optimizer.bin"))
|
| 48 |
+
torch.save(optimizer_adapter.state_dict(), os.path.join(adapter_dir, "optimizer.bin"))
|
| 49 |
+
|
| 50 |
+
torch.save(scheduler.state_dict(), os.path.join(trans_dir, "scheduler.bin"))
|
| 51 |
+
torch.save(scheduler_adapter.state_dict(), os.path.join(adapter_dir, "scheduler.bin"))
|
| 52 |
+
|
| 53 |
+
save_path = os.path.join(save_dir, f"random_states_0.pkl")
|
| 54 |
+
state = {
|
| 55 |
+
"step": step,
|
| 56 |
+
"random_state": random.getstate(),
|
| 57 |
+
"numpy_random_seed": np.random.get_state(),
|
| 58 |
+
"torch_manual_seed": torch.get_rng_state(),
|
| 59 |
+
}
|
| 60 |
+
|
| 61 |
+
if torch.cuda.is_available():
|
| 62 |
+
state["torch_cuda_manual_seed"] = torch.cuda.get_rng_state_all() # list of tensors
|
| 63 |
+
|
| 64 |
+
with open(save_path, "wb") as f:
|
| 65 |
+
pickle.dump(state, f)
|
| 66 |
+
|
| 67 |
+
# Force garbage collection and clear CUDA cache
|
| 68 |
+
gc.collect()
|
| 69 |
+
if torch.cuda.is_available():
|
| 70 |
+
torch.cuda.empty_cache()
|
| 71 |
+
|
| 72 |
+
print(f"[INFO] Saved RNG states + step {step} to {save_path}")
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
def load_checkpoint(transformer, multiLayer_adater, optimizer, optimizer_adapter, scheduler, scheduler_adapter, ckpt_dir, device="cuda"):
|
| 76 |
+
trans_dir = os.path.join(ckpt_dir, "transformer")
|
| 77 |
+
adapter_dir = os.path.join(ckpt_dir, "adapter")
|
| 78 |
+
start_step = 0
|
| 79 |
+
|
| 80 |
+
lora_path = os.path.join(trans_dir, "pytorch_lora_weights.safetensors")
|
| 81 |
+
lora_path_adapter = os.path.join(adapter_dir, "pytorch_lora_weights.safetensors")
|
| 82 |
+
if os.path.exists(lora_path):
|
| 83 |
+
lora_state_dict = CustomFluxPipeline.lora_state_dict(lora_path)
|
| 84 |
+
stripped = {k.replace("transformer.", "", 1) if k.startswith("transformer.") else k: v for k, v in lora_state_dict.items()}
|
| 85 |
+
result = set_peft_model_state_dict(transformer, stripped)
|
| 86 |
+
if result.unexpected_keys:
|
| 87 |
+
print(f"[WARN] Transformer LoRA: {len(result.unexpected_keys)} unexpected keys")
|
| 88 |
+
print(f"[INFO] Loaded Transformer LoRA weights ({len(stripped)} keys).")
|
| 89 |
+
if os.path.exists(lora_path_adapter):
|
| 90 |
+
lora_state_dict = CustomFluxPipeline.lora_state_dict(lora_path_adapter)
|
| 91 |
+
stripped = {k.replace("transformer.", "", 1) if k.startswith("transformer.") else k: v for k, v in lora_state_dict.items()}
|
| 92 |
+
result = set_peft_model_state_dict(multiLayer_adater, stripped)
|
| 93 |
+
if result.unexpected_keys:
|
| 94 |
+
print(f"[WARN] Adapter LoRA: {len(result.unexpected_keys)} unexpected keys")
|
| 95 |
+
print(f"[INFO] Loaded Adapter LoRA weights ({len(stripped)} keys).")
|
| 96 |
+
|
| 97 |
+
pe_path = os.path.join(ckpt_dir, "layer_pe.pth")
|
| 98 |
+
if os.path.exists(pe_path):
|
| 99 |
+
layer_pe = torch.load(pe_path)
|
| 100 |
+
missing_keys, unexpected_keys = transformer.load_state_dict(layer_pe, strict=False)
|
| 101 |
+
|
| 102 |
+
opt_path = os.path.join(trans_dir, "optimizer.bin")
|
| 103 |
+
opt_path_adapter = os.path.join(adapter_dir, "optimizer.bin")
|
| 104 |
+
if os.path.exists(opt_path):
|
| 105 |
+
optimizer.load_state_dict(torch.load(opt_path, map_location=device))
|
| 106 |
+
print("[INFO] Loaded optimizer state.")
|
| 107 |
+
if os.path.exists(opt_path_adapter):
|
| 108 |
+
optimizer_adapter.load_state_dict(torch.load(opt_path_adapter, map_location=device))
|
| 109 |
+
print("[INFO] Loaded optimizer state.")
|
| 110 |
+
|
| 111 |
+
sch_path = os.path.join(trans_dir, "scheduler.bin")
|
| 112 |
+
sch_path_adapter = os.path.join(adapter_dir, "scheduler.bin")
|
| 113 |
+
if os.path.exists(sch_path):
|
| 114 |
+
scheduler.load_state_dict(torch.load(sch_path, map_location=device))
|
| 115 |
+
print("[INFO] Loaded scheduler state.")
|
| 116 |
+
if os.path.exists(sch_path_adapter):
|
| 117 |
+
scheduler_adapter.load_state_dict(torch.load(sch_path_adapter, map_location=device))
|
| 118 |
+
print("[INFO] Loaded scheduler state.")
|
| 119 |
+
|
| 120 |
+
rng_file = None
|
| 121 |
+
for f in os.listdir(ckpt_dir):
|
| 122 |
+
if f.startswith("random_states_") and f.endswith(".pkl"):
|
| 123 |
+
rng_file = os.path.join(ckpt_dir, f)
|
| 124 |
+
break
|
| 125 |
+
|
| 126 |
+
if rng_file:
|
| 127 |
+
with open(rng_file, "rb") as f:
|
| 128 |
+
state = pickle.load(f)
|
| 129 |
+
start_step = state.get("step", 0)
|
| 130 |
+
|
| 131 |
+
if "random_state" in state:
|
| 132 |
+
random.setstate(state["random_state"])
|
| 133 |
+
if "numpy_random_seed" in state:
|
| 134 |
+
np.random.set_state(state["numpy_random_seed"])
|
| 135 |
+
if "torch_manual_seed" in state:
|
| 136 |
+
torch.set_rng_state(state["torch_manual_seed"])
|
| 137 |
+
if "torch_cuda_manual_seed" in state and torch.cuda.is_available():
|
| 138 |
+
torch.cuda.set_rng_state_all(state["torch_cuda_manual_seed"])
|
| 139 |
+
|
| 140 |
+
print(f"[INFO] Resumed RNG states + step {start_step}")
|
| 141 |
+
|
| 142 |
+
return start_step
|
| 143 |
+
|
| 144 |
+
|
| 145 |
+
def load_config(path):
|
| 146 |
+
with open(path, "r") as f:
|
| 147 |
+
return yaml.safe_load(f)
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
def seed_everything(seed: int):
|
| 151 |
+
random.seed(seed)
|
| 152 |
+
np.random.seed(seed)
|
| 153 |
+
torch.manual_seed(seed)
|
| 154 |
+
if torch.cuda.is_available():
|
| 155 |
+
torch.cuda.manual_seed_all(seed)
|
| 156 |
+
torch.backends.cudnn.deterministic = True
|
| 157 |
+
|
| 158 |
+
|
| 159 |
+
def get_input_box(layer_boxes, image_size=512):
|
| 160 |
+
"""
|
| 161 |
+
Quantize layer boxes to 16-pixel grid for latent space alignment.
|
| 162 |
+
|
| 163 |
+
Args:
|
| 164 |
+
layer_boxes: List of boxes in xyxy format [x0, y0, x1, y1]
|
| 165 |
+
image_size: Image size to clamp bounds (default 512)
|
| 166 |
+
|
| 167 |
+
Returns:
|
| 168 |
+
List of quantized boxes in xyxy format
|
| 169 |
+
"""
|
| 170 |
+
list_layer_box = []
|
| 171 |
+
for layer_box in layer_boxes:
|
| 172 |
+
min_col, min_row = layer_box[0], layer_box[1]
|
| 173 |
+
max_col, max_row = layer_box[2], layer_box[3]
|
| 174 |
+
|
| 175 |
+
# Floor for min (start of box)
|
| 176 |
+
quantized_min_row = (min_row // 16) * 16
|
| 177 |
+
quantized_min_col = (min_col // 16) * 16
|
| 178 |
+
|
| 179 |
+
# Ceiling for max (end of box) - use (val + 15) // 16 * 16 for proper ceiling
|
| 180 |
+
quantized_max_row = ((max_row + 15) // 16) * 16
|
| 181 |
+
quantized_max_col = ((max_col + 15) // 16) * 16
|
| 182 |
+
|
| 183 |
+
# Clamp to image bounds
|
| 184 |
+
quantized_min_row = max(0, quantized_min_row)
|
| 185 |
+
quantized_min_col = max(0, quantized_min_col)
|
| 186 |
+
quantized_max_row = min(image_size, quantized_max_row)
|
| 187 |
+
quantized_max_col = min(image_size, quantized_max_col)
|
| 188 |
+
|
| 189 |
+
# Ensure minimum box size of 16 pixels (1 latent token) in each dimension
|
| 190 |
+
# This prevents zero-size boxes that cause reshape errors
|
| 191 |
+
if quantized_max_col <= quantized_min_col:
|
| 192 |
+
# Expand the box, preferring to expand max if there's room
|
| 193 |
+
if quantized_min_col + 16 <= image_size:
|
| 194 |
+
quantized_max_col = quantized_min_col + 16
|
| 195 |
+
else:
|
| 196 |
+
quantized_min_col = max(0, quantized_max_col - 16)
|
| 197 |
+
quantized_max_col = quantized_min_col + 16
|
| 198 |
+
|
| 199 |
+
if quantized_max_row <= quantized_min_row:
|
| 200 |
+
# Expand the box, preferring to expand max if there's room
|
| 201 |
+
if quantized_min_row + 16 <= image_size:
|
| 202 |
+
quantized_max_row = quantized_min_row + 16
|
| 203 |
+
else:
|
| 204 |
+
quantized_min_row = max(0, quantized_max_row - 16)
|
| 205 |
+
quantized_max_row = quantized_min_row + 16
|
| 206 |
+
|
| 207 |
+
list_layer_box.append((quantized_min_col, quantized_min_row, quantized_max_col, quantized_max_row))
|
| 208 |
+
return list_layer_box
|
| 209 |
+
|
| 210 |
+
|
| 211 |
+
def set_lora_into_transformer(
|
| 212 |
+
model: Union[CustomFluxTransformer2DModel, MultiLayerAdapter],
|
| 213 |
+
lora_rank: int,
|
| 214 |
+
lora_alpha: float = 1.0,
|
| 215 |
+
lora_dropout: float = 0.1,
|
| 216 |
+
):
|
| 217 |
+
|
| 218 |
+
target_modules = [
|
| 219 |
+
"to_k", "to_q", "to_v",
|
| 220 |
+
"to_out.0",
|
| 221 |
+
"add_k_proj", "add_q_proj", "add_v_proj",
|
| 222 |
+
"to_add_out",
|
| 223 |
+
] + [f"single_transformer_blocks.{i}.proj_out" for i in range(model.config.num_single_layers)] + [f"transformer_blocks.{i}.proj_out" for i in range(model.config.num_layers)]
|
| 224 |
+
|
| 225 |
+
transformer_lora_config = LoraConfig(
|
| 226 |
+
r=lora_rank,
|
| 227 |
+
lora_alpha=lora_alpha,
|
| 228 |
+
lora_dropout=lora_dropout,
|
| 229 |
+
init_lora_weights="gaussian",
|
| 230 |
+
target_modules=target_modules,
|
| 231 |
+
)
|
| 232 |
+
|
| 233 |
+
model.add_adapter(transformer_lora_config)
|
| 234 |
+
return model
|
| 235 |
+
|
| 236 |
+
|
| 237 |
+
def build_layer_mask(n_layers, H_lat, W_lat, list_layer_box):
|
| 238 |
+
mask = torch.zeros((n_layers, 1, H_lat, W_lat), dtype=torch.float32)
|
| 239 |
+
for i, box in enumerate(list_layer_box):
|
| 240 |
+
if box is None:
|
| 241 |
+
continue
|
| 242 |
+
x1, y1, x2, y2 = box
|
| 243 |
+
x1_t, y1_t, x2_t, y2_t = x1 // 8, y1 // 8, x2 // 8, y2 // 8
|
| 244 |
+
x1_t, y1_t = max(0, x1_t), max(0, y1_t)
|
| 245 |
+
x2_t, y2_t = min(W_lat, x2_t), min(H_lat, y2_t)
|
| 246 |
+
if x2_t > x1_t and y2_t > y1_t:
|
| 247 |
+
mask[i, :, y1_t:y2_t, x1_t:x2_t] = 1.0
|
| 248 |
+
return mask
|
| 249 |
+
|
| 250 |
+
|
| 251 |
+
def encode_target_latents(pipeline, pixel_bchw, n_layers, list_layer_box):
|
| 252 |
+
device = pixel_bchw.device
|
| 253 |
+
dtype = pixel_bchw.dtype
|
| 254 |
+
|
| 255 |
+
vae = pipeline.vae.eval()
|
| 256 |
+
bs, n_layers_in, C, H, W = pixel_bchw.shape
|
| 257 |
+
assert n_layers_in == n_layers, f"The number of input layers {n_layers_in} does not match the specified number of layers {n_layers}"
|
| 258 |
+
|
| 259 |
+
with torch.no_grad():
|
| 260 |
+
dummy_lat = vae.encode(pixel_bchw[:,0]).latent_dist.sample()
|
| 261 |
+
_, C_lat, H_lat, W_lat = dummy_lat.shape
|
| 262 |
+
|
| 263 |
+
x0 = torch.zeros((bs, n_layers, C_lat, H_lat, W_lat), device=device, dtype=dtype)
|
| 264 |
+
|
| 265 |
+
with torch.no_grad():
|
| 266 |
+
for i in range(n_layers):
|
| 267 |
+
pixel_i = pixel_bchw[:, i]
|
| 268 |
+
lat = vae.encode(pixel_i).latent_dist.sample() # [1,C_lat,H_lat,W_lat]
|
| 269 |
+
lat = (lat - vae.config.shift_factor) * vae.config.scaling_factor
|
| 270 |
+
x0[:, i] = lat
|
| 271 |
+
|
| 272 |
+
latent_ids = pipeline._prepare_latent_image_ids(H_lat, W_lat, list_layer_box, device, dtype)
|
| 273 |
+
|
| 274 |
+
return x0, latent_ids
|
| 275 |
+
|
| 276 |
+
|
| 277 |
+
def get_timesteps(pipeline, image_seq_len, num_inference_steps, device):
|
| 278 |
+
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
|
| 279 |
+
|
| 280 |
+
mu = calculate_shift(
|
| 281 |
+
image_seq_len,
|
| 282 |
+
pipeline.scheduler.config.base_image_seq_len,
|
| 283 |
+
pipeline.scheduler.config.max_image_seq_len,
|
| 284 |
+
pipeline.scheduler.config.base_shift,
|
| 285 |
+
pipeline.scheduler.config.max_shift,
|
| 286 |
+
)
|
| 287 |
+
|
| 288 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
| 289 |
+
scheduler=pipeline.scheduler,
|
| 290 |
+
num_inference_steps=num_inference_steps,
|
| 291 |
+
device=device,
|
| 292 |
+
sigmas=sigmas,
|
| 293 |
+
mu=mu,
|
| 294 |
+
)
|
| 295 |
+
|
| 296 |
+
return timesteps
|
| 297 |
+
|
| 298 |
+
|
| 299 |
+
# ============================================================================
|
| 300 |
+
# Box utilities for Prism blended dataset
|
| 301 |
+
# ============================================================================
|
| 302 |
+
|
| 303 |
+
def scale_box_xyxy(box, source_size: int, target_size: int):
|
| 304 |
+
"""
|
| 305 |
+
Scale a box from source_size to target_size.
|
| 306 |
+
Box is already in xyxy format: [x0, y0, x1, y1].
|
| 307 |
+
|
| 308 |
+
Args:
|
| 309 |
+
box: [x0, y0, x1, y1] in source_size coordinates
|
| 310 |
+
source_size: Original data size (e.g., 512)
|
| 311 |
+
target_size: Target inference size (e.g., 512)
|
| 312 |
+
|
| 313 |
+
Returns:
|
| 314 |
+
(x0, y0, x1, y1) in target_size coordinates
|
| 315 |
+
"""
|
| 316 |
+
scale = target_size / source_size
|
| 317 |
+
x0, y0, x1, y1 = box
|
| 318 |
+
|
| 319 |
+
x0_s = int(x0 * scale)
|
| 320 |
+
y0_s = int(y0 * scale)
|
| 321 |
+
x1_s = int(x1 * scale)
|
| 322 |
+
y1_s = int(y1 * scale)
|
| 323 |
+
|
| 324 |
+
# Clamp to valid range
|
| 325 |
+
x0_s = max(0, x0_s)
|
| 326 |
+
y0_s = max(0, y0_s)
|
| 327 |
+
x1_s = min(target_size, x1_s)
|
| 328 |
+
y1_s = min(target_size, y1_s)
|
| 329 |
+
|
| 330 |
+
return (x0_s, y0_s, x1_s, y1_s)
|
| 331 |
+
|
| 332 |
+
|
| 333 |
+
def quantize_box_16(box, target_size: int):
|
| 334 |
+
"""
|
| 335 |
+
Quantize box to 16-pixel grid for latent space alignment.
|
| 336 |
+
Box is in xyxy format.
|
| 337 |
+
"""
|
| 338 |
+
x0, y0, x1, y1 = box
|
| 339 |
+
|
| 340 |
+
# Quantize to 16-pixel grid
|
| 341 |
+
x0_q = (x0 // 16) * 16
|
| 342 |
+
y0_q = (y0 // 16) * 16
|
| 343 |
+
x1_q = ((x1 + 15) // 16) * 16
|
| 344 |
+
y1_q = ((y1 + 15) // 16) * 16
|
| 345 |
+
|
| 346 |
+
# Clamp to image bounds
|
| 347 |
+
x0_q = max(0, x0_q)
|
| 348 |
+
y0_q = max(0, y0_q)
|
| 349 |
+
x1_q = min(target_size, x1_q)
|
| 350 |
+
y1_q = min(target_size, y1_q)
|
| 351 |
+
|
| 352 |
+
return (x0_q, y0_q, x1_q, y1_q)
|
| 353 |
+
|
| 354 |
+
|
| 355 |
+
def get_prism_layer_boxes_xyxy(layers, source_size: int, target_size: int):
|
| 356 |
+
"""
|
| 357 |
+
Extract and scale layer boxes from prism blended metadata.
|
| 358 |
+
|
| 359 |
+
Note: Our blended dataset uses xyxy format [x0, y0, x1, y1].
|
| 360 |
+
|
| 361 |
+
Args:
|
| 362 |
+
layers: List of layer metadata dicts with 'box' field (xyxy format)
|
| 363 |
+
source_size: Size the data was generated at (e.g., 512)
|
| 364 |
+
target_size: Size to run inference at (e.g., 512)
|
| 365 |
+
|
| 366 |
+
Returns:
|
| 367 |
+
List of quantized boxes in xyxy format
|
| 368 |
+
"""
|
| 369 |
+
boxes = []
|
| 370 |
+
|
| 371 |
+
for layer in layers:
|
| 372 |
+
box = layer.get('box', [0, 0, source_size, source_size])
|
| 373 |
+
|
| 374 |
+
# Scale from source to target size (box is already xyxy)
|
| 375 |
+
scaled_box = scale_box_xyxy(box, source_size, target_size)
|
| 376 |
+
|
| 377 |
+
# Quantize to 16-pixel grid
|
| 378 |
+
quantized_box = quantize_box_16(scaled_box, target_size)
|
| 379 |
+
|
| 380 |
+
boxes.append(quantized_box)
|
| 381 |
+
|
| 382 |
+
return boxes
|
| 383 |
+
|
| 384 |
+
|
| 385 |
+
def xywh_to_xyxy(box):
|
| 386 |
+
"""Convert (x, y, w, h) to (x0, y0, x1, y1)."""
|
| 387 |
+
x, y, w, h = box
|
| 388 |
+
return (x, y, x + w, y + h)
|
| 389 |
+
|
| 390 |
+
|
| 391 |
+
def xyxy_to_xywh(box):
|
| 392 |
+
"""Convert (x0, y0, x1, y1) to (x, y, w, h)."""
|
| 393 |
+
x0, y0, x1, y1 = box
|
| 394 |
+
return (x0, y0, x1 - x0, y1 - y0)
|