m
mengsihan/deepcpgdna-smallwood2014-2i
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

multimolecule/deepcpgdna-smallwood2014-2i on Ascend NPU

模型与任务

  • 模型:multimolecule/deepcpgdna-smallwood2014-2i
  • Revision:646ae5817a52f8da18622583a02e48d047dda217
  • 适配路线:regulatory-sequence
  • 设备:npu:0

环境依赖

{
  "python": "3.11.14",
  "torch": "2.9.0+cpu",
  "torch_npu": "2.9.0.post1+gitee7ba04",
  "device_name": "Ascend910_9362",
  "device_count": 2,
  "dtype": "float32",
  "model_source": "local files of revision 646ae5817a52f8da18622583a02e48d047dda217 downloaded from the source URL",
  "revision": "646ae5817a52f8da18622583a02e48d047dda217"
}

安装:pip install -r requirements.txt

NPU 推理

默认执行:python inference.py

真实输入

{
  "source": "real 1001 bp CpG-centered human chr17 window (Ensembl REST) + official model-card 1001 bp example",
  "n_sequences": 2,
  "names": [
    "ensembl_cpg_center_chr17",
    "official_minimal_1001bp"
  ],
  "lengths_after_tokenization": [
    1001
  ],
  "primary_sequence_name": "ensembl_cpg_center_chr17",
  "primary_sequence_length_bp": 1001,
  "primary_sequence_sha256": "244fb85cbe86529b6234b0468c1b2a6fd0584519f71d4bbd640a3fcc833f93c2",
  "primary_sequence_source": "Ensembl REST API, human GRCh38, chr17:7688501-7689502 (region 17:7688401-7690401:1, window offsets 100..1101)",
  "ensembl_window_sha256": "244fb85cbe86529b6234b0468c1b2a6fd0584519f71d4bbd640a3fcc833f93c2",
  "alphabet": "ACGTN (N encoded as all-zero channel)",
  "padding": "no padding; both windows are exactly 1001 bp",
  "input_ids_shape": [
    2,
    1001
  ],
  "input_ids_dtype": "torch.int64",
  "mask": "fixed-window DNA model; no attention mask required"
}

真实推理结果

{
  "logits_shape": [
    2,
    12
  ],
  "cells": [
    "BS24_1_2I",
    "BS24_2_2I",
    "BS24_4_2I",
    "BS24_6_2I",
    "BS24_8_2I",
    "BS25_10_2I",
    "BS25_2_2I",
    "BS25_6_2I",
    "BS25_7_2I",
    "BS25_8_2I",
    "BS26_1_2I",
    "BS26_2_2I"
  ],
  "mean_logits_cpu": {
    "BS24_1_2I": 0.206,
    "BS24_2_2I": 0.2247,
    "BS24_4_2I": 0.1856,
    "BS24_6_2I": 0.2416,
    "BS24_8_2I": 0.2306,
    "BS25_10_2I": 0.7964,
    "BS25_2_2I": 0.1857,
    "BS25_6_2I": 0.317,
    "BS25_7_2I": 0.4204,
    "BS25_8_2I": 0.3498,
    "BS26_1_2I": 0.4019,
    "BS26_2_2I": 0.2711
  },
  "mean_logits_npu": {
    "BS24_1_2I": 0.2056,
    "BS24_2_2I": 0.2244,
    "BS24_4_2I": 0.1853,
    "BS24_6_2I": 0.2413,
    "BS24_8_2I": 0.2302,
    "BS25_10_2I": 0.796,
    "BS25_2_2I": 0.1853,
    "BS25_6_2I": 0.3167,
    "BS25_7_2I": 0.42,
    "BS25_8_2I": 0.3494,
    "BS26_1_2I": 0.4015,
    "BS26_2_2I": 0.2707
  },
  "primary_prediction_ensembl": {
    "name": "ensembl_cpg_center_chr17",
    "per_cell_methylation_probability": {
      "BS24_1_2I": 0.4444,
      "BS24_2_2I": 0.4578,
      "BS24_4_2I": 0.4489,
      "BS24_6_2I": 0.4724,
      "BS24_8_2I": 0.459,
      "BS25_10_2I": 0.5694,
      "BS25_2_2I": 0.5251,
      "BS25_6_2I": 0.4758,
      "BS25_7_2I": 0.4922,
      "BS25_8_2I": 0.5052,
      "BS26_1_2I": 0.506,
      "BS26_2_2I": 0.4729
    },
    "methylated_cells_at_0.5": 4
  },
  "official_minimal_prediction": {
    "name": "official_minimal_1001bp",
    "per_cell_methylation_probability": {
      "BS24_1_2I": 0.6535,
      "BS24_2_2I": 0.6498,
      "BS24_4_2I": 0.6401,
      "BS24_6_2I": 0.6441,
      "BS24_8_2I": 0.6513,
      "BS25_10_2I": 0.788,
      "BS25_2_2I": 0.5671,
      "BS25_6_2I": 0.6748,
      "BS25_7_2I": 0.705,
      "BS25_8_2I": 0.6633,
      "BS26_1_2I": 0.6855,
      "BS26_2_2I": 0.657
    },
    "methylated_cells_at_0.5": 12
  }
}

CPU-NPU 一致性

{
  "cpu_vs_npu_logits": {
    "max_abs_diff": 0.00080013,
    "mean_abs_diff": 0.00037907,
    "pearson_r": 1.0,
    "allclose_atol_rtol": "0.01/0.01",
    "allclose": true
  },
  "finite_elements": "24/24",
  "methylation_call_agreement_at_0.5": 1.0,
  "per_cell_rank_agreement": 1.0
}

性能

{
  "warmup_runs": 3,
  "stable_runs": 10,
  "mean_inference_s": 0.001092,
  "first_stable_s": 0.001155,
  "sequences_per_s": 1831.756,
  "bases_per_s": 1833587.7,
  "input_length_bp": 1001,
  "batch_size": 2,
  "dtype": "float32",
  "device": "npu:0"
}

自验证截图

Agent workflow

NPU device call

Model result

三张图片均由 xterm.js 根据本次真实日志生成。

已知限制

  • 验证覆盖固定 revision 646ae5817a52f8da18622583a02e48d047dda217、1001 bp 固定窗口、float32 与单卡 Ascend910。
  • 模型契约要求窗口必须恰为 1001 bp 且以 CpG 为中心;输入为真实人 chr17 窗口,非训练分布(Smallwood 2014 小鼠 mESC),仅作 NPU 适配一致性验证。
  • 单序列/双序列一致性为 smoke consistency,不替代完整甲基化精度评测。

标签

#NPU #Ascend #Ascend910