def basic_loss(self, y_true, y_pred, go_backwards=False): """y_true需要是整数形式(非one hot) """ # 导出mask并转换数据类型 if self.input_mask is None: mask = None else: mask = K.cast(self.input_mask, K.floatx()) # y_true需要重新明确一下shape和dtype y_true = K.reshape(y_true, K.shape(y_pred)[:-1]) y_true = K.cast(y_true, 'int32') # 反转相关 if self.hidden_dim is None: if go_backwards: # 是否反转序列 y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask) trans = K.transpose(self.trans) else: trans = self.trans histoty = K.gather(trans, y_true) else: if go_backwards: # 是否反转序列 y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask) r_trans, l_trans = self.l_trans, self.r_trans else: l_trans, r_trans = self.l_trans, self.r_trans histoty = K.gather(l_trans, y_true) histoty = tf.einsum('bnd,kd->bnk', histoty, r_trans) # 计算loss histoty = K.concatenate([y_pred[:, :1], histoty[:, :-1]], 1) y_pred = (y_pred + histoty) / 2 loss = K.sparse_categorical_crossentropy(y_true, y_pred, from_logits=True) if mask is None: return K.mean(loss) else: return K.sum(loss * mask) / K.sum(mask)
def compute_mask(self, inputs, mask=None): """ 保证第一个token不被mask """ mask = super(Embedding, self).compute_mask(inputs, mask) if mask is not None: mask1 = K.ones_like(mask[:, :1], dtype='bool') mask2 = mask[:, 1:] return K.concatenate([mask1, mask2], 1)
def call(self, inputs): input_shape = K.shape(inputs) batch_size, seq_len = input_shape[0], input_shape[1] pos_embeddings = self.embeddings[:seq_len] pos_embeddings = K.expand_dims(pos_embeddings, 0) if self.merge_mode == 'add': return inputs + pos_embeddings else: pos_embeddings = K.tile(pos_embeddings, [batch_size, 1, 1]) return K.concatenate([inputs, pos_embeddings])
def dense_loss(self, y_true, y_pred): """y_true需要是one hot形式 """ # 导出mask并转换数据类型 if self.input_mask is None: mask = None else: mask = K.cast(self.input_mask, K.floatx()) # 计算目标分数 target_score = self.target_score(y_true, y_pred, mask) # 递归计算log Z init_states = [y_pred[:, 0]] if mask is None: mask = K.ones_like(y_pred[:, :, :1]) else: mask = K.expand_dims(mask, 2) y_pred = K.concatenate([y_pred, mask]) log_norm, _, _ = K.rnn(self.log_norm_step, y_pred[:, 1:], init_states) # 最后一步的log Z向量 log_norm = tf.reduce_logsumexp(log_norm, 1) # logsumexp得标量 # 计算损失 -log p return log_norm - target_score
def basic_accuracy(self, y_true, y_pred, go_backwards=False): """训练过程中显示逐帧准确率的函数,排除了mask的影响 此处y_true需要是整数形式(非one hot) """ # 导出mask并转换数据类型 if self.input_mask is None: mask = None else: mask = K.cast(self.input_mask, K.floatx()) # y_true需要重新明确一下shape和dtype y_true = K.reshape(y_true, K.shape(y_pred)[:-1]) y_true = K.cast(y_true, 'int32') # 反转相关 if self.hidden_dim is None: if go_backwards: # 是否反转序列 y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask) trans = K.transpose(self.trans) else: trans = self.trans histoty = K.gather(trans, y_true) else: if go_backwards: # 是否反转序列 y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask) r_trans, l_trans = self.l_trans, self.r_trans else: l_trans, r_trans = self.l_trans, self.r_trans histoty = K.gather(l_trans, y_true) histoty = tf.einsum('bnd,kd->bnk', histoty, r_trans) # 计算逐标签accuracy histoty = K.concatenate([y_pred[:, :1], histoty[:, :-1]], 1) y_pred = (y_pred + histoty) / 2 y_pred = K.cast(K.argmax(y_pred, 2), 'int32') isequal = K.cast(K.equal(y_true, y_pred), K.floatx()) if mask is None: return K.mean(isequal) else: return K.sum(isequal * mask) / K.sum(mask)