Hetero-MoE 研究档案
Collection
三个真实骨干(Qwen/BERT/ViT)跨架构对齐与融合: 维度投影空间 = 按功能槽张成的子空间之和。唯一可直接加载的模型是 HeteroMoEProjSpace-v2。 • 7 items • Updated
n_text int64 | n_image int64 | qwen_shape list | bert_shape list | vit_shape list | max_tok int64 | dtypes string | ag_news_class_counts list | food101_class_count int64 | cache_bytes int64 | cache_gb float64 | elapsed_sec float64 | device string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
10,000 | 5,000 | [
10000,
128,
896
] | [
10000,
128,
256
] | [
5000,
197,
768
] | 128 | float16 | [
2523,
2338,
2477,
2662
] | 7 | 4,462,200,640 | 4.462 | 126.5 | cuda |
v62 端到端训练与本系列多次路由实验的数据源。 存的是三个骨干在真实数据上的 最后一层隐藏状态,不是原始文本/图像。
| 文件 | 形状 | 大小 | 说明 |
|---|---|---|---|
h_qwen.npy |
(10000, 128, 896) |
2.29 GB | ag_news 文本过 Qwen2.5-0.5B |
h_bert.npy |
(10000, 128, 256) |
0.66 GB | ag_news 文本过 BERT-L4-H256 |
h_vit.npy |
(5000, 197, 768) |
1.51 GB | food101 图像过 ViT-base-16 |
labels.npy |
(10000,) |
80 KB | ag_news 4 类,分布 [2523, 2338, 2477, 2662] |
food_labels.npy |
(5000,) |
40 KB | food101 7 类 |
import numpy as np
from huggingface_hub import hf_hub_download
h_qwen = np.load(hf_hub_download('redmingwei/hetero-moe-cache', 'h_qwen.npy',
repo_type='dataset'), mmap_mode='r')
labels = np.load(hf_hub_download('redmingwei/hetero-moe-cache', 'labels.npy',
repo_type='dataset'))
v62 用的是按类别分层 80/20(train 7998 / val 2002)。 早期一次顺序切分得到 0.937,因分布偏移已作废,改用分层后的 0.9016。