'inbatch', 'uniform', 'frequency', 'adaptive'

### 不同采样策略的应用场景 #### Inbatch 负采样 Inbatch负采样的核心理念是在每次迭代过程中，从当前批次(batch)内选取未被选作正样本的数据作为负样本[^1]。这种方法的优势在于计算效率高，因为不需要额外存储大量的负样本集合；同时能够充分利用每一批次内的数据分布特点。 ```python def get_inbatch_negatives(current_batch, positive_samples): negative_samples = [item for item in current_batch if item not in positive_samples] return negative_samples ``` #### Uniform 采样 Uniform采样意味着在整个候选集中均匀随机抽取样本，而不考虑任何先验概率或权重分配[^2]。这种方式简单直观，但在某些情况下可能导致重要信息丢失，特别是当不同类别的频率差异较大时。 ```python import random def uniform_sampling(candidates, sample_size): sampled_items = random.sample(candidates, k=sample_size) return sampled_items ``` #### Frequency-based 采样 Frequency-based采样依据各类别出现的频次来进行加权抽样操作。对于高频项给予较低的选择几率，而低频项则有更高的机会被挑选出来参与训练过程[^3]。这有助于缓解类别不平衡带来的负面影响，并使模型更加关注稀疏事件的学习效果。 ```python from collections import Counter def frequency_based_sampling(data, num_samples): freq_dist = Counter([label for _, label in data]) weights = {k: (sum(freq_dist.values()) / v)**0.75 for k, v in freq_dist.items()} total_weight = sum(weights[k] for _, k in data) probabilities = [(weights[label]/total_weight) for _, label in data] indices = range(len(data)) selected_indices = random.choices(indices, weights=probabilities, k=num_samples) result = [data[i] for i in selected_indices] return result ``` #### Adaptive 采样 Adaptive采样是一种动态调整机制，根据前一轮训练的结果反馈来决定下一次应该优先选择哪些类型的实例进行学习。这种自适应的方法可以有效提高收敛速度并优化最终性能表现。具体实现上可能涉及到在线更新参数或是引入外部奖励信号指导决策流程。 ```python class AdaptiveSampler: def __init__(self): self.weights = {} def update_weights(self, feedback): # 更新权重逻辑... pass def adaptive_sample(self, dataset, batch_size): weighted_probabilities = [] for instance in dataset: key = str(instance.features()) weight = self.weights.get(key, 1.0) weighted_probabilities.append(weight) norm_probs = [p/sum(weighted_probabilities) for p in weighted_probabilities] chosen_indices = np.random.choice(range(len(dataset)), size=batch_size, replace=False, p=norm_probs) return [dataset[idx] for idx in chosen_indices] ```

创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考

下一篇 SIFT（Scale-Invariant Feature Transform）