haiku实现TemplatePairStack类

TemplatePairStack是实现蛋白质结构模版pair_act特征表示的类:
通过layer_stack.layer_stack(c.num_block)(block) 堆叠c.num_block(配置文件中为2)block 函数,每个block对输入pair_act 和 pair_mask执行计算流程:TriangleAttention ---> dropout ->TriangleAttention ---> dropout -> TriangleMultiplication ---> dropout -> TriangleMultiplication ---> dropout -> Transition

复制代码
import haiku as hk


class TemplatePairStack(hk.Module):
  """Pair stack for the templates.

  Jumper et al. (2021) Suppl. Alg. 16 "TemplatePairStack"
  """

  def __init__(self, config, global_config, name='template_pair_stack'):
    super().__init__(name=name)
    self.config = config
    self.global_config = global_config

  def __call__(self, pair_act, pair_mask, is_training, safe_key=None):
    """Builds TemplatePairStack module.

    Arguments:
      pair_act: Pair activations for single template, shape [N_res, N_res, c_t].
      pair_mask: Pair mask, shape [N_res, N_res].
      is_training: Whether the module is in training mode.
      safe_key: Safe key object encapsulating the random number generation key.

    Returns:
      Updated pair_act, shape [N_res, N_res, c_t].
    """

    if safe_key is None:
      safe_key = prng.SafeKey(hk.next_rng_key())

    gc = self.global_config
    c = self.config

    if not c.num_block:
      return pair_act

    def block(x):
      """One block of the template pair stack."""
      pair_act, safe_key = x

      dropout_wrapper_fn = functools.partial(
          dropout_wrapper, is_training=is_training, global_config=gc)

      safe_key, *sub_keys = safe_key.split(6)
      sub_keys = iter(sub_keys)

      pair_act = dropout_wrapper_fn(
          TriangleAttention(c.triangle_attention_starting_node, gc,
                            name='triangle_attention_starting_node'),
          pair_act,
          pair_mask,
          next(sub_keys))
      pair_act = dropout_wrapper_fn(
          TriangleAttention(c.triangle_attention_ending_node, gc,
                            name='triangle_attention_ending_node'),
          pair_act,
          pair_mask,
          next(sub_keys))
      pair_act = dropout_wrapper_fn(
          TriangleMultiplication(c.triangle_multiplication_outgoing, gc,
                                 name='triangle_multiplication_outgoing'),
          pair_act,
          pair_mask,
          next(sub_keys))
      pair_act = dropout_wrapper_fn(
          TriangleMultiplication(c.triangle_multiplication_incoming, gc,
                                 name='triangle_multiplication_incoming'),
          pair_act,
          pair_mask,
          next(sub_keys))
      pair_act = dropout_wrapper_fn(
          Transition(c.pair_transition, gc, name='pair_transition'),
          pair_act,
          pair_mask,
          next(sub_keys))

      return pair_act, safe_key

    if gc.use_remat:
      block = hk.remat(block)

    res_stack = layer_stack.layer_stack(c.num_block)(block)
    pair_act, safe_key = res_stack((pair_act, safe_key))
    return pair_act
相关推荐
ZH15455891313 小时前
Flutter for OpenHarmony Python学习助手实战:API接口开发的实现
python·学习·flutter
断眉的派大星3 小时前
均值为0,方差为1:数据的“标准校服”
人工智能·机器学习·均值算法
小宋10213 小时前
Java 项目结构 vs Python 项目结构:如何快速搭一个可跑项目
java·开发语言·python
A尘埃4 小时前
电子厂PCB板焊点缺陷检测(卷积神经网络CNN)
人工智能·神经网络·cnn
Tadas-Gao4 小时前
缸中之脑:大模型架构的智能幻象与演进困局
人工智能·深度学习·机器学习·架构·大模型·llm
中金快讯4 小时前
新视野混合净值波动有几何?贝莱德基金回撤控制策略是否命中关键?
人工智能
楚兴4 小时前
MacBook M1 安装 OpenClaw 完整指南
人工智能·后端
23遇见4 小时前
探索CANN:开源AI计算底座的关键组件与技术思想
人工智能
jl48638214 小时前
变比测试仪显示屏的“标杆“配置!如何兼顾30000小时寿命与六角矢量图精准显示?
人工智能·经验分享·嵌入式硬件·物联网·人机交互