bf6414d298d29c12913cf5c83a0277e1db63e6d6,finetune/nn/auxiliary.py,,embed_context,#Any#Any#Any#Any#,6

Before Change

    context_weighted_avg = tf.get_variable(	
        name="cwa",
        shape=[context_dim],	
        initializer=tf.random_normal_initializer(stddev=config.weight_stddev),	
    )	

    if config.train_embeddings:	
        context_embed_weights = dropout(	
            context_embed_weights, config.embed_p_drop, train	
        )	
        context_weighted_avg = dropout(context_weighted_avg, config.embed_p_drop, train)	
    else:	
        context_embed_weights = tf.stop_gradient(context_embed_weights)	

    with tf.variable_scope("context_embedding"):	
        weighted_C = tf.multiply(	
            context, context_weighted_avg	
        )  // [batch_size, seq_length, context_dim] * [context_dim] = [batch_size, seq_length, context_dim], with weighted inputs	
        c_embed = tf.tensordot(	
            weighted_C, context_embed_weights, axes=[[2], [0]]	
        )  // [batch_size, seq_length, context_dim] * [context_dim, n_embed] = [batch_size, seq_length, n_embed]	
        c_embed = norm(c_embed, tf.get_variable_scope())
        // c_embed = tf.reduce_mean(c_embed, axis=1)
    featurizer_state["context"] = c_embed
    return featurizer_state

After Change

    context_bias = tf.get_variable(
        name="ca",	
        shape=[config.n_context_embed],	
        initializer=tf.zeros_initializer(),	
    )
    with tf.variable_scope("context_embedding"):
        c_embed = tf.add(tf.multiply(context, context_weight), context_bias)

Instances