Loss functions

class compressors.distillation.losses.AttentionLoss(p: int = 2)
forward(s_hidden_states: Tuple[torch.FloatTensor], t_hidden_states: Tuple[torch.FloatTensor]) → torch.FloatTensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class compressors.distillation.losses.CRDLoss(student_dim: int, teacher_dim: int, n_data: int, feature_dim: int = 128, nce_k: int = 16384, nce_t: float = 0.07, nce_m: float = 0.5)

CRD Loss function includes two symmetric parts: (a) using teacher as anchor, choose positive and negatives over the student side (b) using student as anchor, choose positive and negatives over the teacher side

Parameters

  • student_dim – the dimension of student’s feature

  • teacher_dim – the dimension of teacher’s feature

  • feature_dim – the dimension of the projection space

  • nce_k – number of negatives paired with each positive

  • nce_t – the temperature

  • nce_m – the momentum for updating the memory buffer

  • n_data – the number of samples in the training set, therefor the memory buffer is: opt.n_data x opt.feat_dim

forward(f_s, f_t, idx, contrast_idx=None)

Forward pass.

Parameters

  • f_s – the feature of student network, size [batch_size, s_dim]

  • f_t – the feature of teacher network, size [batch_size, t_dim]

  • idx – the indices of these positive samples in the dataset, size [batch_size]

  • contrast_idx – the indices of negative samples, size [batch_size, nce_k]

Returns

The contrastive loss

class compressors.distillation.losses.KLDivLoss(temperature: float = 1.0)
forward(s_logits, t_logits)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class compressors.distillation.losses.MSEHiddenStatesLoss(normalize: bool = False, need_mapping: bool = False, teacher_hidden_state_dim: Optional[int] = None, student_hidden_state_dim: Optional[int] = None, num_layers: Optional[int] = None)
forward(s_hidden_states: Union[torch.FloatTensor, Tuple[torch.FloatTensor]], t_hidden_states: Union[torch.FloatTensor, Tuple[torch.FloatTensor]]) → torch.FloatTensor

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

compressors.distillation.losses.kl_div_loss(s_logits: torch.FloatTensor, t_logits: torch.FloatTensor, temperature: float = 1.0) → torch.FloatTensor

KL-devergence loss

Parameters

  • s_logits (FloatTensor) – output for student model.

  • t_logits (FloatTensor) – output for teacher model.

  • temperature (float, optional) – Temperature for teacher distribution. Defaults to 1.

Returns

Divergence between student and teachers distribution.

Return type

FloatTensor

compressors.distillation.losses.mse_loss(s_hidden_states: Tuple[torch.FloatTensor], t_hidden_states: Tuple[torch.FloatTensor], normalize: bool = False) → torch.FloatTensor

mse loss for hidden states

Parameters

  • s_hidden_states (Tuple[FloatTensor]) – student hiddens

  • t_hidden_states (Tuple[FloatTensor]) – teacher hiddens

  • normalize (bool, optional) – normalize embeddings. Defaults to False.

Returns

loss

Return type

FloatTensor

compressors.distillation.losses.pkt_loss(s_hidden_states: torch.FloatTensor, t_hidden_states: torch.FloatTensor, eps: float = 1e-07) → torch.FloatTensor

Loss between distributions over features similarity with cosine similarity kernel.

Parameters

  • s_hidden_states (FloatTensor) – student hidden states

  • t_hidden_states (FloatTensor) – teacher hidden states

  • eps (float, optional) – small value. Defaults to 1e-7.

Returns

loss

Return type

FloatTensor