bert-mini-sentence-distil-unsupervised

The model is trained by knowledge distillation between the "princeton-nlp/unsup-simcse-roberta-large" and "prajjwal1/bert-mini" on the 'ffgcc/NEWS5M'.

The model can perform inferenced by Automodel.

The model achieves 0.825 and 0.83 for pearsonr and spearmanr respectively on STS-b test dataset.

For more training detail, the training config and the pytorch forward function is as follows:

config = {
  'epoch' = 200,
  'learning_rate' = 3e-4,
  'batch_size' = 12288,
  'temperature' = 0.05
}

def forward_cos_mse_kd_unsup(self, sentences, teacher_sentence_embs):
    """forward function for the unsupervised News5M dataset"""
    _, o = self.bert(**sentences)

    # cosine similarity between the first half batch and the second half batch
    half_batch = o.size(0) // 2
    higher_half = half_batch * 2 #skip the last datapoint when the batch size number is odd
    
    cos_sim = cosine_sim(o[:half_batch], o[half_batch:higher_half])
    cos_sim_teacher = cosine_sim(teacher_sentence_embs[:half_batch], teacher_sentence_embs[half_batch:higher_half])

    # KL Divergence between student and teacher probabilities
    soft_teacher_probs = F.softmax(cos_sim_teacher / self.temperature, dim=1)
    kd_contrastive_loss = F.kl_div(F.log_softmax(cos_sim / self.temperature, dim=1),
                            soft_teacher_probs,
                            reduction='batchmean')

    # MSE loss
    kd_mse_loss = nn.MSELoss()(o, teacher_sentence_embs)/3

    # equal weight for the two losses
    total_loss = kd_contrastive_loss*0.5 + kd_mse_loss*0.5

    return total_loss, kd_contrastive_loss, kd_mse_loss