Spaces:
Paused
Paused
File size: 11,920 Bytes
d0d37cd |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 |
import os
import logging
from typing import Any, Optional, Tuple
import torch
import torch.nn.functional as F
from omegaconf import DictConfig, OmegaConf
from safetensors.torch import load_model, save_model
import matplotlib.pyplot as plt
import numpy as np
import seaborn as sns
from matplotlib.ticker import MaxNLocator
BOUNDING_BOX_MAX_SIZE = 1.925
def normalize_bbox(bounding_box_xyz: Tuple[float]):
#import ipdb; ipdb.set_trace()
max_l = max(bounding_box_xyz)
return [BOUNDING_BOX_MAX_SIZE * elem / max_l for elem in bounding_box_xyz]
def normalize_bboxs(bounding_box_xyz, max_xyz):
#max_l = max(bounding_box_xyz)
normalized = BOUNDING_BOX_MAX_SIZE * bounding_box_xyz / torch.tensor(max_xyz, device=bounding_box_xyz.device)
return normalized
def load_config(cfg_path: str) -> Any:
"""
Load and resolve a configuration file.
Args:
cfg_path (str): The path to the configuration file.
Returns:
Any: The loaded and resolved configuration object.
Raises:
AssertionError: If the loaded configuration is not an instance of DictConfig.
"""
cfg = OmegaConf.load(cfg_path)
OmegaConf.resolve(cfg)
assert isinstance(cfg, DictConfig)
return cfg
def parse_structured(cfg_type: Any, cfg: DictConfig) -> Any:
"""
Parses a configuration dictionary into a structured configuration object.
Args:
cfg_type (Any): The type of the structured configuration object.
cfg (DictConfig): The configuration dictionary to be parsed.
Returns:
Any: The structured configuration object created from the dictionary.
"""
scfg = OmegaConf.structured(cfg_type(**cfg))
return scfg
def load_model_weights(model: torch.nn.Module, ckpt_path: str) -> None:
"""
Load a safetensors checkpoint into a PyTorch model.
The model is updated in place.
Args:
model: PyTorch model to load weights into
ckpt_path: Path to the safetensors checkpoint file
Returns:
None
"""
assert ckpt_path.endswith(
".safetensors"
), f"Checkpoint path '{ckpt_path}' is not a safetensors file"
load_model(model, ckpt_path)
def save_model_weights(model: torch.nn.Module, save_path: str) -> None:
"""
Save a PyTorch model as safetensors format.
Args:
model: PyTorch model to save
save_path: Output path (must end with .safetensors)
"""
assert save_path.endswith(".safetensors"), "Path must end with .safetensors"
os.makedirs(os.path.dirname(save_path), exist_ok=True)
save_model(model, save_path)
assert os.path.exists(save_path), f"Failed to save to {save_path}"
def select_device() -> Any:
"""
Selects the appropriate PyTorch device for tensor allocation.
Returns:
Any: The `torch.device` object.
"""
return torch.device(
"cuda"
if torch.cuda.is_available()
else "mps"
if torch.backends.mps.is_available()
else "cpu"
)
def mask_cross_entropy(p_st, p_ed, p_max, logits, target, shift):
p_range = torch.arange(p_st, p_ed, device=logits.device)
p_range_expanded = p_range.unsqueeze(0).repeat(p_max.shape[0], 1)
valid_p_mask = p_range_expanded <= p_max.unsqueeze(1)+p_st
valid_p_mask = valid_p_mask.unsqueeze(1).expand(-1, logits.shape[1], -1)
logits_masked = logits.clone()
logits_masked[:,:,p_st:p_ed][~valid_p_mask] = float('-inf')
p_loss = F.cross_entropy(
logits_masked[:, :-1, p_st:p_ed].permute(0, 2, 1),
target[:, shift:, p_st:p_ed].argmax(-1),
)
return p_loss
def positional_encoding(x, num_freqs):
freqs = 2.0 ** torch.arange(num_freqs, device=x.device) # [num_freqs]
angles = x.unsqueeze(-1) * freqs # [..., num_freqs]
sin_cos = torch.cat([angles.sin(), angles.cos()], dim=-1) # [..., 2*num_freqs]
return sin_cos.flatten(-2)
def visualize_token_probabilities(
probs,
cut_idx,
sample_idx=0,
tokens_per_page=10, # 每页显示的token数量
figsize=(12, 20), # 单页图表大小
save_dir=None # 保存图片的目录(None则直接显示)
):
"""
分页展示所有有效token的概率分布(每页10个,一行一个token)
参数:
- probs: 概率张量,形状为 (batch_size, seq_len, num_classes)
- cut_idx: 有效区域的截止索引
- sample_idx: 要可视化的batch样本索引
- tokens_per_page: 每页显示的token数量
- figsize: 单页图表大小
- save_dir: 保存图片的目录(若为None则直接显示)
"""
# 转换为numpy数组
if isinstance(probs, torch.Tensor):
probs = probs.cpu().detach().numpy()
# 获取单个样本的概率分布
sample_probs = probs[sample_idx] # (seq_len, num_classes)
seq_len, num_classes = sample_probs.shape
# 处理cut_idx,确定有效区域并提取有效token
if isinstance(cut_idx, torch.Tensor):
cut_idx = cut_idx.cpu().detach().numpy()
valid_length = min(int(cut_idx[sample_idx] if not np.isscalar(cut_idx) else cut_idx), seq_len)
valid_probs = sample_probs[:valid_length, :] # 只取有效区域内的token
num_valid_tokens = valid_probs.shape[0]
if num_valid_tokens == 0:
print(f"警告:没有有效token可显示(有效区域长度:{valid_length})")
return None
# 创建保存目录(如果需要)
if save_dir is not None and not os.path.exists(save_dir):
os.makedirs(save_dir)
# 计算总页数
total_pages = (num_valid_tokens + tokens_per_page - 1) // tokens_per_page
print(f"共{num_valid_tokens}个有效token,分为{total_pages}页展示")
# 分页生成图表
figures = []
for page in range(total_pages):
# 计算当前页的token范围
start = page * tokens_per_page
end = min(start + tokens_per_page, num_valid_tokens)
page_tokens = end - start
# 创建当前页的画布
fig, axes = plt.subplots(page_tokens, 1, figsize=(figsize[0], 2*page_tokens))
fig.suptitle(
f'Token Probability Distributions (Sample {sample_idx}) - Page {page+1}/{total_pages}',
fontsize=16,
y=1.02
)
# 为当前页的每个token绘制分布
for i in range(page_tokens):
token_idx = start + i
token_probs = valid_probs[i] # 当前页内的相对索引
ax = axes[i] if page_tokens > 1 else axes # 处理单token情况
# 绘制条形图
class_indices = np.arange(num_classes)
bars = ax.bar(class_indices, token_probs, width=0.8, color='skyblue', edgecolor='black')
# 突出显示最高概率的类别
max_prob_idx = np.argmax(token_probs)
max_prob_value = token_probs[max_prob_idx]
bars[max_prob_idx].set_color('orange')
# 标注概率>5%的类别
for j, (bar, prob) in enumerate(zip(bars, token_probs)):
height = bar.get_height()
if prob > 0.05:
ax.text(bar.get_x() + bar.get_width()/2., height + 0.01,
f'{prob:.2f}', ha='center', va='bottom', fontsize=9)
# 设置子图标题和坐标轴
ax.set_title(
f'Token {token_idx} (Max: Class {max_prob_idx} = {max_prob_value:.2f})',
fontsize=11
)
ax.set_xlabel('Class Index')
ax.set_ylabel('Probability')
ax.set_ylim(0, 1.1)
ax.xaxis.set_major_locator(MaxNLocator(integer=True))
ax.grid(True, alpha=0.3, axis='y')
# 除最后一个子图外隐藏x轴标签
if i != page_tokens - 1:
ax.set_xlabel('')
plt.tight_layout()
figures.append(fig)
# 保存或显示图表
if save_dir is not None:
save_path = os.path.join(save_dir, f'token_probs_page_{page+1}.png')
fig.savefig(save_path, dpi=300, bbox_inches='tight')
print(f"已保存第{page+1}页至: {save_path}")
else:
plt.show()
plt.close(fig) # 关闭当前页图表,释放内存
return figures
def visualize_max_prob_distribution(
probs,
cut_idx=None, # 不再需要,因为已提前过滤
sample_idx=0,
bins=20,
figsize=(12, 6)
):
# 转换为numpy数组
if isinstance(probs, torch.Tensor):
probs = probs.cpu().detach().numpy()
# 获取单个样本的概率分布并计算最大概率
sample_probs = probs[sample_idx]
max_probs_per_token = np.max(sample_probs, axis=1) # 所有token都是已过滤的有效token
# 创建画布
fig, ax = plt.subplots(figsize=figsize)
# 绘制直方图
n, bins, patches = ax.hist(
max_probs_per_token,
bins=bins,
range=(0, 1),
edgecolor='black',
alpha=0.7,
color='skyblue'
)
# 标注数量
for count, patch in zip(n, patches):
height = patch.get_height()
if height > 0:
ax.text(
patch.get_x() + patch.get_width()/2.,
height + 0.5,
f'{int(count)}',
ha='center',
va='bottom',
fontsize=9
)
# 统计指标
mean_prob = np.mean(max_probs_per_token)
median_prob = np.median(max_probs_per_token)
max_count = int(np.max(n)) if len(n) > 0 else 0
# 设置标题和坐标轴
ax.set_title(
f'Distribution of Maximum Probabilities (All Valid Tokens from 5 Iterations)\n'
f'Total tokens: {len(max_probs_per_token)} | Mean: {mean_prob:.2f} | Median: {median_prob:.2f}',
fontsize=14
)
ax.set_xlabel('Maximum Probability Value (0-1)')
ax.set_ylabel('Number of Tokens (Frequency)')
ax.set_xlim(0, 1)
ax.set_ylim(0, max_count + 2)
ax.xaxis.set_major_locator(MaxNLocator(nbins=11))
ax.yaxis.set_major_locator(MaxNLocator(integer=True))
ax.grid(True, alpha=0.3, axis='y')
plt.tight_layout()
return fig
def top_k_prob_mask(probs, cut_idx, top_percent=0.15, visualize=False):
max_probs = probs.permute(0, 2, 1).max(dim=1).values # (batch_size, seq_len)
batch_size, seq_len = max_probs.shape
# 1. 生成基础mask:cut_idx前面为True,后面为False
if isinstance(cut_idx, (int, float)):
cut_idx = torch.tensor([cut_idx] * batch_size, device=max_probs.device)
base_mask = (torch.arange(seq_len, device=max_probs.device)[None, :] < cut_idx[:, None])
valid_count = base_mask.sum().item()
# 处理无有效位置的情况
if valid_count == 0:
empty_mask = torch.zeros_like(max_probs, dtype=torch.bool)
return empty_mask, empty_mask
# 2. 计算原始目标mask(cut内前N%高概率True)
valid_probs = max_probs[base_mask]
total_valid = valid_probs.numel()
k = max(min(int(total_valid * top_percent), total_valid), 1)
_, top_valid_indices = torch.topk(valid_probs, k)
# 原始mask:cut内top k为True,其余全False
valid_area_original = torch.zeros(total_valid, dtype=torch.bool, device=max_probs.device)
valid_area_original[top_valid_indices] = True
original_mask = torch.zeros_like(max_probs, dtype=torch.bool)
original_mask[base_mask] = valid_area_original
# 3. 计算反向mask(cut内非top k为True,cut外全False)
valid_area_reverse = ~valid_area_original # 与原始有效区域完全相反
reverse_mask = torch.zeros_like(max_probs, dtype=torch.bool)
reverse_mask[base_mask] = valid_area_reverse # cut外保持False
return original_mask, reverse_mask # 返回两个mask |