Compute loss for model. If both `labels` and `mask` are None, it degenerates to SimCLR unsupervised loss: https://arxiv.org/pdf/2002.0

Compute loss for model. If both `labels` and `mask` are None, it degenerates to SimCLR unsupervised loss: https://arxiv.org/pdf/2002.0

Args: audio_x: tensor of shape (batch_size, sequence_len, audio_in) video_x: tensor of shape (batch_size, sequence_le

Args: audio: tensor of shape (batch_size, seqlen, audio_in) video: tensor of shape (batch_size, seqlen, video_in)

audio_feat: tensor of shape (batch, seqlen1, audio_in) video_feat: tensor of shape (batch, seqlen2, video_in) tex

audio_feat: tensor of shape (batch, seqlen1, audio_in) video_feat: tensor of shape (batch, seqlen2, video_in) tex

audio_feat: tensor of shape (batch, seqlen1, audio_in) video_feat: tensor of shape (batch, seqlen2, video_in) tex

audio_feat: tensor of shape (batch, seqlen1, audio_in) video_feat: tensor of shape (batch, seqlen2, video_in) tex

Args: audio_x: tensor of shape (batch_size, audio_dim) video_x: tensor of shape (batch_size, video_dim) t

Args: audio_x: tensor of shape (batch_size, audio_dim) video_x: tensor of shape (batch_size, video_dim) t

Args: audio_x: tensor of shape (batch_size, audio_in) video_x: tensor of shape (batch_size, video_in) tex

Forward lld (gaussian prior) and entropy estimation, partially refers the implementation of https://github.com/Linear95/CLUB/blob/master/MI_D

Calulate the score

Args: x: tensor of shape (batch_size, in_size)

support feat_type: utt | frm-align | frm-unalign

audio_feat: tensor of shape (batch, seqlen1, audio_in) video_feat: tensor of shape (batch, seqlen2, video_in) tex

simulating word-align network (for seq_len_T == seq_len_A == seq_len_V) audio_x: tensor of shape (batch, seqlen, audio_in) vi

audio_feat: tensor of shape (batch, seqlen1, audio_in) text_feat: tensor of shape (batch, seqlen2, text_in) vide

x : (batch_size, sequence_len, in_size)

# 感觉用aligned feature能够方便计算重建loss吧 text_feat: [batch, seqlen, feat1] audio_feat: [batch, seqlen, feat2] video_feat: [

Args: x: tensor of shape (batch_size, in_size)

Args: x: tensor of shape (batch_size, sequence_len, in_size) 因为用的是 final_states ，所以特征的 padding 是放在前面的

Input is expected to be of size [bsz x seqlen].

Args: x_in (FloatTensor): embedded input of shape `(src_len, batch, embed_dim)` x_in_k (FloatTensor): embedded input

Args: x (Tensor): input to the layer of shape `(seq_len, batch, embed_dim)` encoder_padding_mask (ByteTensor): binary

Input shape: Time x Batch x Channel Self-attention can be implemented by passing in the same arguments for query, key and value. Times