File size: 6,589 Bytes
8245f38 | 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 | import argparse
import json
import os
import random
from collections import OrderedDict
from typing import List, Sequence, Tuple
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
try:
from transformers import AutoTokenizer, AutoConfig, XLMRobertaModel
except ImportError:
AutoTokenizer = None
HFBertModel = None
class XLMRobertaLanguageBackbone(nn.Module):
def __init__(
self,
ckpt_path,
frozen_modules: Sequence[str] = (),
dropout: float = 0.0,
init_cfg= None,
) -> None:
super().__init__()
if 'base' in ckpt_path:
self.head = nn.Linear(768, 768, bias=True) # XLarge
model_name = "./xlm-roberta-base/"
elif 'large' in ckpt_path:
self.head = nn.Linear(1024, 768, bias=True) # XLarge
model_name = "./xlm-roberta-large/"
self.frozen_modules = frozen_modules
cfg = AutoConfig.from_pretrained(model_name)
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
self.model = XLMRobertaModel(cfg)
self.language_dim = cfg.hidden_size
# 加载 model 权重
new_state_dict = OrderedDict()
state_dict = torch.load(
ckpt_path,
map_location="cpu",
weights_only=False,
)['state_dict']
for k, v in state_dict.items():
if k.startswith('backbone.text_model.'):
name = k.split("backbone.text_model.")[-1]
new_state_dict[name] = v
msg = self.load_state_dict(new_state_dict, strict=True)
print(msg)
print("TEXT-ENCODER xlm-roberta-base LOADING WEIGHTS !!!!")
def forward(self, text: List[str], max_seq_len: int = 32):
text = self.tokenizer(text=text, return_tensors="pt",
padding="max_length", max_length=max_seq_len)
text = text.to(device=self.model.device)
txt_feats = self.model(**text)["last_hidden_state"][:, 0]
txt_feats = self.head(txt_feats)
return txt_feats
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--wedetect_checkpoint', type=str, default='checkpoints/wedetect_base.pth')
parser.add_argument('--classname_file', type=str, default='data/texts/coco_zh_class_texts.json')
parser.add_argument('--max-seq-len', type=int, default=32,
help='Fixed token length (must match ONNX export).')
parser.add_argument('--num-classes-per-group', type=int, default=4,
help='Number of classes per group npy.')
parser.add_argument('--num-groups', type=int, default=64,
help='Number of random groups to generate.')
parser.add_argument('--calib-dir', type=str, default='calib_data',
help='Directory for text-encoder quantisation calibration data.')
args = parser.parse_args()
with open(args.classname_file) as f:
name_chinese = json.load(f)
name_chinese = [name[0] for name in name_chinese]
language_encoder = XLMRobertaLanguageBackbone(args.wedetect_checkpoint).cuda()
# Generate random groups: each group picks 4 random classes → shape (1, 4, 768)
total_classes = len(name_chinese)
print(f"Total classes: {total_classes} → Generating {args.num_groups} random groups")
# Generate calibration data for text-encoder quantisation
# Directories: calib_dir/input_ids/ calib_dir/attention_mask/
calib_input_ids = os.path.join(args.calib_dir, "input_ids")
calib_attn_mask = os.path.join(args.calib_dir, "attention_mask")
for d in (calib_input_ids, calib_attn_mask):
os.makedirs(d, exist_ok=True)
tokenizer = language_encoder.tokenizer
for g in range(args.num_groups):
idx = random.sample(range(total_classes), args.num_classes_per_group)
group_texts = [name_chinese[i] for i in idx]
tokens = tokenizer(group_texts, padding="max_length",
max_length=args.max_seq_len, return_tensors="np")
np.save(os.path.join(calib_input_ids, f"{g:03d}.npy"),
tokens["input_ids"].astype(np.int64))
np.save(os.path.join(calib_attn_mask, f"{g:03d}.npy"),
tokens["attention_mask"].astype(np.int64))
print(f"calib [{g:03d}] input_ids: {tokens['input_ids'].shape} "
f"classes: {group_texts}")
# Compress each subdirectory to .tar.gz
import tarfile
for sub_name in ("input_ids", "attention_mask"):
sub_dir = os.path.join(args.calib_dir, sub_name)
tar_path = os.path.join(args.calib_dir, f"{sub_name}.tar.gz")
with tarfile.open(tar_path, "w:gz") as tar:
for fname in sorted(os.listdir(sub_dir)):
tar.add(os.path.join(sub_dir, fname), arcname=fname)
print(f"Compressed: {tar_path}")
print(f"Saved calibration data to {args.calib_dir}/")
# -------------------------------------------------------------------
# Generate 64 random 4-class text embeddings → class_embedding_4cls/
# Each file: (1, 4, 768) float32, L2-normalised (same as the ONNX
# image encoder expects via text_features input).
# -------------------------------------------------------------------
embed_dir = os.path.join(args.calib_dir, "class_embedding_4cls")
os.makedirs(embed_dir, exist_ok=True)
print(f"\nGenerating {args.num_groups} random {args.num_classes_per_group}-class "
f"text embeddings → {embed_dir}/")
for g in range(args.num_groups):
idx = random.sample(range(total_classes), args.num_classes_per_group)
group_texts = [name_chinese[i] for i in idx]
with torch.no_grad():
feats = language_encoder(group_texts, max_seq_len=args.max_seq_len)
feats = F.normalize(feats, dim=-1).unsqueeze(0) # (1, 4, 768)
fpath = os.path.join(embed_dir, f"{g:03d}.npy")
np.save(fpath, feats.cpu().numpy().astype(np.float32))
if (g + 1) % 16 == 0 or g == args.num_groups - 1:
print(f" [{g + 1:3d}/{args.num_groups}] shape={feats.shape} "
f"classes: {group_texts}")
# Compress
tar_path = os.path.join(args.calib_dir, "class_embedding_4cls.tar.gz")
with tarfile.open(tar_path, "w:gz") as tar:
for fname in sorted(os.listdir(embed_dir)):
tar.add(os.path.join(embed_dir, fname), arcname=fname)
print(f"Compressed: {tar_path}")
print(f"Saved calibration data to {args.calib_dir}/")
|