How to Have God’s Faith
“Jesus replied, “Let the faith of God be in you! Listen to the truth I speak to you: If someone says to this mountain with great faith and having no doubt, ‘Mountain, be lifted up and thrown into the midst of the sea,’ and believes that what he says will happen, it will be done.” (Mark 11:22-23 TPT)
In the passage above, Jesus said the “faith of God” and not “faith in God”. If we want to perform mighty miracles like casting a mountain into the sea, we will need to receive God’s own faith.
“For I say, through the grace that was given me, to every man who is among you, not to think of himself more highly than he ought to think; but to think reasonably, as God has apportioned to each person a measure of faith.” (Romans 12:3 WEB)
You see, there is the measure of faith that God already gave to us in the past, and also God’s own faith that He gives to believers at appointed timings.
The former is the one that we keep free from unbelief through daily renewing of the mind, while the latter is more powerful because it is pure and guaranteed to mightily prosper in its intended purpose.
“But to each one is given the manifestation of the Spirit for the profit of all. For to one is given through the Spirit the word of wisdom, and to another the word of knowledge, according to the same Spirit; to another faith, by the same Spirit; and to another gifts of healings, by the same Spirit; and to another workings of miracles; and to another prophecy; and to another discerning of spirits; to another different kinds of languages; and to another the interpretation of languages. But the one and the same Spirit produces all of these, distributing to each one separately as he desires.” (1 Corinthians 12:7-11 WEB)
Faith is one of the manifestations of the Holy Spirit. He distributes this “faith of God” at times. Every believer can receive this.
To understand when this God-faith comes, we have to know the delivery system:
“Faith, then, is birthed in a heart that responds to God’s anointed utterance of the Anointed One.” (Romans 10:17 TPT)
Many English Bible translations do a poor job of translating the verse above. They stray from the original Greek meaning. However, I like the translation I quoted above.
The words translated as “anointed utterance” in Greek is “rhematos” which means “a spoken word”. “The Anointed One” is Christou, which is Christ.
“Rhematos Christou” is not referring to God’s written Scriptures in general. That one would be the logos word, or “what God has said before”. The rhema word is talking about what Christ is currently actively saying to you through His Spirit inside you.
When Jesus speaks a rhema word to you, it carries God’s authority. When you hear and act upon that word, the “faith of God” is imparted to you for mighty power to be released. This is when miracles and breakthroughs flow.
Let us see this in action:
“Peter answered him and said, “Lord, if it is you, command me to come to you on the waters.” He said, “Come!” Peter stepped down from the boat, and walked on the waters to come to Jesus.” (Matthew 14:28-29 WEB)
Jesus said “Come,” and this rhema word carried the authority for Peter to transcend the natural laws of nature. When Peter heard the command and acted upon it, the “faith of God” was imparted to him, empowering him to walk on the waters.
You may be surprised by this, but even Jesus depended on the rhema word from His Father during His earthly ministry.
“Jesus therefore answered them, “Most certainly, I tell you, the Son can do nothing of himself, but what he sees the Father doing. For whatever things he does, these the Son also does likewise. For the Father has affection for the Son, and shows him all things that he himself does. He will show him greater works than these, that you may marvel.” (John 5:19-20 WEB)
Before doing any miracle, Jesus first received the rhema word from the Father, and He simply obeyed, acting upon what the Father showed Him.
When Jesus acted upon the rhema word, God’s power flows because the authority, faith, and will of God are working in perfect synergy.
It is therefore most crucial that we ask the Lord to give us the rhema word for our situation. To receive it, you can calm your emotions and pray about your situation, listening in silence to see what the Holy Spirit brings to your attention.
Other than that, sometimes He uses your pastor or Christian friend to speak divinely inspired words to you. The Holy Spirit is the source of the rhema word. The way He delivers it isn’t the most important thing, so be open to however He chooses to convey the rhema word to you.
When it comes, grab hold of it in your heart, and act upon it, for it is the key to unlock God’s supernatural provision in your situation!
The 37 recorded miracles of Jesus Christ in the four gospels show us God’s heart of love and Grace towards us. As you explore them in “Messiah’s Miracles”, faith will arise in your heart to receive miraculous breakthroughs in your own life: https://bit.ly/messiahs-miracles
attention is all you need 在 [徵文] Attention is all you need - 看板DataScience - 批踢踢實業坊 的推薦與評價
1) 論文介紹
arXiv 連結: https://arxiv.org/abs/1706.03762
這篇是 Google 發表在 NIPS2017 上的 paper，其最大的亮點是提出一個新的
encoder-decoder 架構完全依賴於 Attention 機制而完全沒用到 CNN 以及 RNN，這樣
的好處是提高了訓練時可平行化處理的部分 (RNN 依賴於序列處理，而 Attention 可以
是矩陣乘法運算) 並且讓 input, output sequences 中的文字彼此之間跨越距離的障礙
能夠找到有關聯的字詞。
在這篇論文之前，為了減少序列的計算量，被提出的方法有 ByteNet, ConvS2S 等網路架
構，都是用到了 CNN 來做 downsampling，但這樣的方法仍然會遇到距離障礙的問題: 相
距越遠的字詞會越難找到彼此的資訊，其解決方法就是使用 self-attention 取代
encode, decode 階段，讓每一個詞都去計算整個 sequence 的表示，此外其也能解決過
長的 RNN 架構可能會造成的梯度消失問題。
這篇提出的架構 Transformer 如下圖:
也是 encoder-decoder 的形式，只不過都換成了使用 attention + fully connected
layers 來實現。
對於 input, output 序列每個字詞過 Embedding 之後須加上 Positional Encoding 主
要是因為這個架構不像 RNN 是有序的，但是為了捕捉到字詞前後的關係所以須加上基於
位置的 Embedding。
Encoder:
在 Encoder 中堆疊了六個相同的 layer，每個 layer 都包含兩個 sublayer，分別是
multi-head self-attention 和 position-wise fully connected network，並且都是
以殘差連接的方式，好處是能夠加深網路，並且都過 layer normalization 加速收斂，
因此每一層 sublayer 都可以以 LayerNorm(x + SubLayer(x)) 來表示。
Decoder:
Decoder 同樣堆疊六層，但每一層包含了三個 sublayer，其同樣有目標序列的
self-attention，再加上了 decoder 向 encoder 的 attention 機制，最後同樣過全連
接層輸出，多的 sublayer 就是在負責從 encoder 藉由 attention 抓取重要資訊來作為
輸出參考，其中比較需要注意的是 decoder 的 self-attention 需要加上遮罩機制，也
就是讓位置 t 的字詞只能 attend 到自己以前的字詞，不能向後偷看。
其用到的 Attention 計算公式為 Scaled Dot-Product Attention 流程如下圖:
其中 Q 是 query (發起 attention 的 matrix)
K, V 分別是 key, value (被 attention 的 matrix)，在這篇論文中使用的 K, V 是相
同的，都是某時間點的 hidden state
寫成公式如下:
其實與 dot product attention 的計算方式幾乎一樣，使用 Q, K 進行點積得到
attention weight(知道對每個時間點的該注意的程度)，再和 value 相乘得到加權結果
。
只差在 Q, K 進行點積之後除以一個 hidden dimension (d_k) 的根號，是為了避免點積
的結果太大影響訓練穩定程度 ( 除以根號 d_k 之後可以讓方差變成 1 )
Multi-Head Attention
其使用的 attention 計算公式為上述的 scaled dot-product attention，而其使用的機
制為 multi-head attention，概念是分別將 Q, K, V 經過線性轉換(learnable的)成 h
個，再 h 個各自平行地去做 attention，最後再將 h 個結果 concatenate 在一起得到
最終的結果，通常會希望 concatenate 之後的維度與原來相同(d_k)，因此在做線性轉換
時通常會把 h 個轉換出來的結果維度為 d_k / h
寫成公式如下：
Position-wise Feed Forward Networks
其實就是兩層的 fully connected layers 搭配 Relu
總體架構來說，就是讓源句子做 self-attention 以殘差連接和 layer normalization，
接著丟去 feed forward 也以殘差連接以及 layer normalization，這個動作重複 6 層
後當作 Encoder 輸出，Decoder 階段目標句子同樣先做 self-attention (不過帶有
mask) 後，對 Encoder 的輸出做 attention 再丟進 feed forward，這三個步驟也都是
殘差連接以及 layer normalization，並也做 6 層後經過一個線性轉換以及 softmax 輸
出預測句子。
Why Self-Attention?
作者們認為主要有三大好處：
1. 降低了每一層的計算複雜度：只要 sequence length 小於 hidden dimension 就會比
RNN 複雜度低
2. 增加了可平行化處理的程度：加快訓練速度
3. 解決long-dependency的問題：字詞相距很遠難以關注到彼此的問題
Training
在訓練過程中他們也使用了 learning rate 遞減、dropout、label smoothing 等等
tips，在此不贅述。
Result
當時在機器翻譯資料集 WMT2014 取得了 state-of-the-art 的結果，值得注意的是
Transformer 的 Training Cost 是比其他模型少許多的。
這篇並沒有詳細提到 Positional Encoding 怎麼 init，因為本篇重點應該比較注重在
Attention，並且 Google 也在今年(2018)提出了一種新的將位置加入 Transformer 的
方法，詳細可以參考 https://arxiv.org/abs/1803.02155
2) 個人心得
這篇應該算是 NLP 近年來最多人關注的 paper 之一，Google 也還在針對這個架構進行
研究發展新的 paper 來改善增進 Transformer 的問題與能力。舉例來說目前
Transformer 的其中一個問題是無法像 RNN-based model 做 schedule sampling 的訓
練，讓模型在訓練階段都只能看 ground truth 而在測試階段就要看自己前一個時間點的
輸出結果。
底下附上我自己實作 Transformer 應用在 PTT Gossiping QA Dataset
(https://github.com/zake7749/Gossiping-Chinese-Corpus ) 的一些結果:
第一次寫這類型的文章，若有理解錯誤或是表達的不精確請各位大大指正~謝謝
--
※ 發信站: 批踢踢實業坊(ptt.cc), 來自: 61.218.53.138
※ 文章網址: https://www.ptt.cc/bbs/DataScience/M.1533004952.A.244.html
※ 編輯: dav1a1223 (114.136.163.75), 07/31/2018 10:43:56
... <看更多>