Chapter09 - Improving retention with long short-term memory networks/LSTM.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Nội dung chính\n",
    "1. Lý thuyết mạng LSTM <br/>\n",
    "    1.1 Mạng hồi quy RNN <br/>\n",
    "    1.2 Vấn đề phụ thuộc xa <br/>\n",
    "    1.3 Mạng LSTM <br/>\n",
    "2. Các biến thể LSTM\n",
    "3. Xây dựng mạng LSTM với Keras <br/>\n",
    "    3.1 LSTM cho dự đoán chuỗi thời gian đơn biến <br/>\n",
    "    3.2 LSTM cho dự đoán chuỗi thời gian sử dụng phương pháp cửa sổ (Window method) <br/>\n",
    "    3.3 Deep LSTM <br/>\n",
    "    \n",
    "Nguồn tham khảo:\n",
    "* https://dominhhai.github.io/vi/2017/10/what-is-lstm/\n",
    "* https://dominhhai.github.io/vi/2017/10/implement-gru-lstm/\n",
    "* https://medium.com/datadriveninvestor/how-do-lstm-networks-solve-the-problem-of-vanishing-gradients-a6784971a577\n",
    "* https://blog.chappiebot.com/h%C6%B0%E1%BB%9Bng-d%E1%BA%ABn-chi-ti%E1%BA%BFt-v%E1%BB%81-c%C6%A1-ch%E1%BA%BF-c%E1%BB%A7a-lstm-v%C3%A0-gru-trong-nlp-a1bd9346b209\n",
    "* https://nguyentruonglong.net/giai-thich-chi-tiet-ve-mang-long-short-term-memory-lstm.html\n",
    "* http://svail.github.io/\n",
    "* https://machinelearningmastery.com/time-series-prediction-lstm-recurrent-neural-networks-python-keras/"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 1. Lý thuyết mạng LSTM"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1.1 Mạng hồi quy RNN"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Con người không bắt đầu suy nghĩ của họ từ đầu tại tất cả các thời điểm. Cũng như bạn đang đọc bài viết này, bạn hiểu mỗi chữ ở đây dựa vào từ bạn đã hiểu các chữ trước đó chứ không phải là đọc tới đâu ném hết đi tới đó, rồi lại bắt đầu suy nghĩ lại từ đầu tới chữ bạn đang đọc. Tức là tư duy đã có một bộ nhớ để lưu lại những gì diễn ra trước đó.\n",
    "\n",
    "Tuy nhiên các mô hình mạng nơ-ron truyền thống thì không thể làm được việc đó, đó có thể coi là một khuyết điểm chính của mạng nơ-ron truyền thống. Ví dụ, bạn muốn phân loại các bối cảnh xảy ra ở tất cả các thời điểm trong một bộ phim, thì đúng là không rõ làm thế nào để có thể hiểu được một tình huống trong phim mà lại phụ thuộc vào các tình huống trước đó nếusử dụng các mạng nơ-ron truyền thống.\n",
    "\n",
    "Mạng nơ-ron hồi quy (Recurrent Neural Network) sinh ra để giải quyết vấn đề đó. Mạng này chứa các vòng lặp bên trong cho phép thông tin có thể lưu lại được.\n",
    "\n",
    "<img src=\"./images/RNN-rolled.png\" alt=\"\" width=\"30%\" height=\"30%\"/>\n",
    "\n",
    "Hình vẽ trên mô tả một đoạn của mạng nơ-ron hồi quy $A$ với đầu vào $x_t$ và đầu ra $h_t$. Một vòng lặp cho phép thông tin có thể được truyền từ bước này qua bước này qua bước khác của mạng nơ-ron.\n",
    "\n",
    "Các vòng lặp này khiến cho mạng nơ-ron hồi quy trông có vẻ khó hiểu. Tuy nhiên, nếu bạn để ý một chút thì nó không khác mấy so với các mạng nơ-ron thuần. Một mạng nơ-ron hồi quy có thể được coi là nhiều bản sao chép của cùng một mạng, trong đó mỗi đầu ra của mạng này là đầu vào của một mạng sao chép khác. Nói thì hơi khó hiểu, nhưng bạn hãy xem hình mô tả sau:\n",
    "\n",
    "![RNN unrolled](./images/RNN-unrolled.png)\n",
    "\n",
    "Chuỗi lặp lại các mạng này chính là phân giải của mạng nơ-ron hồi quy, các vòng lặp khiến chúng tạo thành một chuỗi danh sách các mạng sao chép nhau. Bạn có thấy nó khác gì một mạng nơ-ron thuần không? Không khác gì phải không? Các nút của mạng vẫn nhận đầu vào và có đầu ra hệt như mạng nơ-ron thuần.\n",
    "\n",
    "Trong vài năm gần đây, việc ứng dụng RNN đã đưa ra được nhiều kết quả không thể tin nổi trong nhiều lĩnh vực: nhận dạng giọng nói, mô hình hóa ngôn ngữ, dịch máy, mô tả ảnh,… Danh sách vẫn còn đang được mở rộng tiếp. Đằng sau sự thành công này chính là sự đóng góp của LSTM. LSTM là một dạng đặc biệt của mạng nơ-ron hồi quy, với nhiều bài toán thì nó tốt hơn mạng hồi quy thuần. Hầu hết các kết quả thú vị thu được từ mạng RNN là được sử dụng với LSTM. Trong bài viết này, ta sẽ cùng khám phá xem mạng LSTM là cái gì nhé.\n",
    "## 1.2 Vấn đề phụ thuộc xa\n",
    "Một điểm nổi bật của RNN chính là ý tưởng kết nối các thông tin phía trước để dự đoán cho hiện tại. Việc này tương tự như ta sử dụng các cảnh trước của bộ phim để hiểu được cảnh hiện thời. Nếu mà RNN có thể làm được việc đó thì chúng sẽ cực kì hữu dụng, tuy nhiên liệu chúng có thể làm được không? Câu trả lời là còn tùy.\n",
    "\n",
    "Đôi lúc ta chỉ cần xem lại thông tin vừa có thôi là đủ để biết được tình huống hiện tại. Ví dụ, ta có câu: “các đám may trên bầu trời” thì ta chỉ cần đọc tới “các đám may trên bầu” là đủ biết được chữ tiếp theo là “trời” rồi. Trong tình huống này, khoảng cách tới thông tin có được cần để dự đoán là nhỏ, nên RNN hoàn toàn có thể học được.\n",
    "\n",
    "![RNN-shorttermdepdencies](./images/RNN-shorttermdepdencies.png)\n",
    "\n",
    "Nhưng trong nhiều tình huống ta buộc phải sử dụng nhiều ngữ cảnh hơn để suy luận. Ví dụ, dự đoán chữ cuối cùng trong đoạn: “I grew up in France… I speak fluent French.”. Rõ ràng là các thông tin gần (”I speak fluent”) chỉ có phép ta biết được đằng sau nó sẽ là tên của một ngôn ngữ nào đó, còn không thể nào biết được đó là tiếng gì. Muốn biết là tiếng gì, thì ta cần phải có thêm ngữ cảnh “I grew up in France” nữa mới có thể suy luận được. Rõ ràng là khoảng cách thông tin lúc này có thể đã khá xa rồi.\n",
    "\n",
    "Thật không may là với khoảng cách càng lớn dần thì RNN bắt đầu không thể nhớ và học được nữa.\n",
    "\n",
    "Quan sát về nhược điểm của RNN, ta nhận thấy kiến trúc này không hề có cơ chế lọc những thông tin không cần thiết. Bộ nhớ của kiến trúc có hạn, nếu lưu tất cả những chi tiết không cần thiết thì sẽ dẫn đến quá tải, từ đó quên những thứ ở xa trong quá khứ. Tương tự như con người, rất dễ hiểu phải không!!!\n",
    "\n",
    "![RNN-longtermdependencies](./images/RNN-longtermdependencies.png)\n",
    "\n",
    "Về mặt lý thuyết, rõ ràng là RNN có khả năng xử lý các phụ thuộc xa (long-term dependencies). Chúng ta có thể xem xét và cài đặt các tham số sao cho khéo là có thể giải quyết được vấn đề này. Tuy nhiên, đáng tiếc trong thực tế RNN có vẻ không thể học được các tham số đó. Vấn đề này đã được khám phá khá sâu bởi <a href=\"http://people.idsia.ch/~juergen/SeppHochreiter1991ThesisAdvisorSchmidhuber.pdf\">Hochreiter (1991) [tiếng Đức]</a> và <a href=\"http://www-dsi.ing.unifi.it/~paolo/ps/tnn-94-gradient.pdf\">Bengio, et al. (1994)</a>, trong các bài báo của mình, họ đã tìm được nhưng lý do căn bản để giải thích tại sao RNN không thể học được.\n",
    "\n",
    "LSTM và GRU ra đời như bản cải tiến của RNN.\n",
    "\n",
    "![1_yBXV9o5q7L_CvY7quJt3WQ](./images/1_yBXV9o5q7L_CvY7quJt3WQ.png)\n",
    "\n",
    "Như hình trên, chúng ta có thể thấy trong mỗi module của 2 kiến trúc trên đều có trang bị các cổng (gate). Nhiệm vụ các cổng này là giúp kiến trúc đánh giá được mức độ quan trọng của thông tin, từ đó đưa ra quyết định giữ lại hay bỏ đi. Nhờ cơ chế này, các thông tin quan trọng có thể được truyền đi rất xa và vẫn có tác động đáng kể trong tương lai xa.\n",
    "\n",
    "Để hiểu rõ tại sao những kiến trúc này có thể làm được điều đó, chúng ta nên dành chút thời gian để hiểu hơn về cơ chế hoạt động của kiến trức RNN nói chung. Đầu tiên, các từ trong câu được biến đổi thành các vector. sau đó RNN sẽ xử lý các chuỗi vector này từng từ một.\n",
    "\n",
    "![1_AQ52bwW55GsJt6HTxPDuMA](./images/1_AQ52bwW55GsJt6HTxPDuMA.gif)\n",
    "\n",
    "Trong lúc xử lý, RNN đưa thông tin về trạng thái ẩn (hidden state) được tính toán khi tính toán trong quá khứ ($w_{t-2}, w_{t-1},$ vv) thành đầu vào cho quá trình xử lý ở hiện tại ($w_{t}$). Trạng thái ẩn này mang thông tin của từ hiện tại và các từ trước đó để truyền lại cho các từ tiếp theo. Đây chính là cơ chế giúp RNN xử lý dữ liệu tuần tự: các từ phía trước sẽ ảnh hưởng đầu ra của từ phía sau.\n",
    "\n",
    "![1_o-Cq5U8-tfa1_ve2Pf3nfg](./images/1_o-Cq5U8-tfa1_ve2Pf3nfg.gif)\n",
    "\n",
    "Bây giờ hãy nhìn mộ cách chi tiết hơn làm sao RNN tính toán được trạng thái ẩn. Trước hết, vector của từ hiện tại và trạng thái ẩn của từ phía trước được kết hợp lại, có thể bằng phép concatenation. Vector tổng này sau đó sẽ đi qua hàm tanh. Có thể thấy là tất cả thông tin của các từ phía trước được dồn vào trạng thái ẩn, đây rõ ràng là một nút thắt.\n",
    "\n",
    "![1_WMnFSJHzOloFlJHU6fVN-g](./images/1_WMnFSJHzOloFlJHU6fVN-g.gif)\n",
    "\n",
    "Hàm kích hoạt tanh được dùng để điều chỉnh dòng thông tin đi qua hệ thống. Mọi giá trị sẽ được chiếu về khoảng (-1, 1). Khi một vector đi qua mạng neuron, chúng trải qua rất nhiều phép tính và trong quá trình đó sẽ có thành phần nào đó trở nên quá to và khiến các thành phần khác trở nên không đáng kể. Hàm tanh lúc này sẽ giúp tinh chỉnh sự chênh lệch này để giúp \"các thành phần yếu thế vẫn có tiếng nói trong xã hội\"\n",
    "\n",
    "![1_gFC2bTg3uihp1klknWU0qg](./images/1_gFC2bTg3uihp1klknWU0qg.gif)\n",
    "\n",
    "Đó là những thứ căn bản về RNN. Nó hoạt động tương đối hiệu quả với các chuỗi ngắn và tiết kiệm tài nguyên, nhất là khi so sánh với các mô hình trong kiến trúc CNN."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 1.3 Mạng LSTM"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Mọi mạng hồi quy đều có dạng là một chuỗi các mô-đun lặp đi lặp lại của mạng nơ-ron. Với mạng RNN chuẩn, các mô-dun này có cấu trúc rất đơn giản, thường là một tầng tanh.\n",
    "\n",
    "![LSTM3-SimpleRNN](./images/LSTM3-SimpleRNN.png)\n",
    "\n",
    "LSTM cũng có kiến trúc dạng chuỗi như vậy, nhưng các mô-đun trong nó có cấu trúc khác với mạng RNN chuẩn. Thay vì chỉ có một tầng mạng nơ-ron, chúng có tới 4 tầng tương tác với nhau một cách rất đặc biệt.\n",
    "\n",
    "![LSTM3-chain](./images/LSTM3-chain.png)\n",
    "\n",
    "![LSTM2-notation](./images/LSTM2-notation.png)\n",
    "\n",
    "LTSM có các cổng giúp lọc thông tin. Trạng thái của cell có thể coi là trí nhớ của toàn hệ thống. Muốn được cập nhất vào hệ thống này, các thành phần thông tin phải đi qua các cổng. Vậy có thể thấy cơ chế này linh hoạt hơn khi mà các thành phần thông tin được chọn lọc mới có thể được đưa vào bộ nhớ, nhờ đó các thành phần có ích từ thời điểm rất lâu trong quá khứ vẫn có thể có tác dụng điều chỉnh.\n",
    "\n",
    "![1_0f8r3Vd-i4ueYND1CUrhMA](./images/1_0f8r3Vd-i4ueYND1CUrhMA.png)\n",
    "\n",
    "Trong LSTM sử dùng hàm sigmoid, hàm này có tác dụng như hàm tanh, nhưng vùng chiếu của nó chỉ là $(0,1)$ thay vì $(-1,1)$. Nếu phần thông tin nào không quan trọng, ta chiếu nó về 0, nó coi như bị \"lãng quên\".\n",
    "\n",
    "![1_rOFozAke2DX5BmsX2ubovw](./images/1_rOFozAke2DX5BmsX2ubovw.gif)\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.3.1 Cổng quên"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Cổng này quyết định xem thông tin nào trong bộ nhớ hiện tại được giữ và thông tin nào. Đầu vào là $h_{t-1}$ và $x_t$ rồi đưa qua hàm sigmoid sẽ được giá trị $f_t$ trong khoảng [0, 1]. Đầu ra 1 thể hiện việc giữ toàn bộ thông tin, đầu ra 0 thể hiện toàn bộ thông tin sẽ bị bỏ đi.\n",
    "\n",
    "![1_GjehOa513_BgpDDP6Vkw2Q](./images/1_GjehOa513_BgpDDP6Vkw2Q.gif)\n",
    "\n",
    "Công thức cho tầng cổng quên:\n",
    "$$f_t = \\sigma \\left( W_f.[h_{t-1}, x_t] + b_f \\right)$$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.3.2 Cổng vào"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Cổng này dùng để cập nhật thông tin mới. Ở đây có xuất hiện 2 hàm sigmoid và hàm tanh. Tác dụng của chúng cũng như trên. Output từ hàm sigmoid sẽ có tác dụng lọc thông tin đã qua xử lý từ output hàm tanh.\n",
    "\n",
    "![1_TTmYy7Sy8uUXxUXfzmoKbA](./images/1_TTmYy7Sy8uUXxUXfzmoKbA.gif)\n",
    "\n",
    "Công thức:\n",
    "$$i_t = \\sigma \\left( W_i . [h_{t-1}, x_t] + b_i \\right)$$\n",
    "$$\\tilde{C}_t = \\text{tanh} \\left( W_c . [h_{t-1}, x_t] + b_C \\right)$$ \n",
    "\n",
    "Giờ là lúc cập nhật trạng thái tế bào cũ $C_{t-1}$ thành trạng thái mới $C_t$. Ta sẽ nhân trạng thái cũ với $f_t$ để bỏ đi những thông tin ta quyết định quên lúc trước. Sau đó cộng thêm $i_t * \\tilde{C}_t$. Trạng thái mới thu được này phụ thuộc vào việc ta quyết định cập nhập mỗi giá trị trạng thái ra sao.\n",
    "\n",
    "![1_S0rXIeO_VoUVOyrYHckUWg](./images/1_S0rXIeO_VoUVOyrYHckUWg.gif)\n",
    "\n",
    "Công thức:\n",
    "$$C_t = f_t * C_{t-1} + i_t * \\tilde{C}_t $$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 1.3.3 Cổng ra"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Giá trị đầu ra sẽ dựa vào trạng thái tế bào, nhưng sẽ được tiếp tục sàng lọc. Đầu tiên, ta chạy một tầng sigmoid để quyết định phần nào của trạng thái tế bào ta muốn xuất ra. Sau đó, ta đưa nó trạng thái tế bảo qua một hàm tanh để co giá trị nó về khoảng $[-1, 1]$, và nhân nó với đầu ra của cổng sigmoid để được giá trị đầu ra ta mong muốn. Chú ý rằng cả kết quả đầu ra và cả trạng thái cell đều được đưa vào bước tiếp theo.\n",
    "\n",
    "![1_VOXRGhOShoWWks6ouoDN3Q](./images/1_VOXRGhOShoWWks6ouoDN3Q.gif)\n",
    "\n",
    "Công thức:\n",
    "$$o_t = \\sigma \\left( W_o . [h_{t-1}, x_t] + b_o \\right)$$\n",
    "$$h_t = o_t * \\text{tanh}(C_t)$$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 2. Các biến thể LSTM"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Những thứ ta vừa mô tả ở trên là một LSTM khá bình thường. Nhưng không phải tất cả các LTSM đều giống như vậy. Thực tế, các bài báo về LTSM đều sử dụng một phiên bản hơi khác so với mô hình LTSM chuẩn. Sự khác nhau không lớn, nhưng chúng giúp giải quyết phần nào đó trong cấu trúc của LTSM.\n",
    "\n",
    "Một dạng LTSM phổ biến được giới thiệu bởi <a href=\"ftp://ftp.idsia.ch/pub/juergen/TimeCount-IJCNN2000.pdf\">Gers & Schmidhuber (2000)</a> được thêm các đường kết nối “peephole connections”, làm cho các tầng cổng nhận được giá trị đầu vào là trạng thái tế bào.\n",
    "\n",
    "![LSTM3-var-peepholes](./images/LSTM3-var-peepholes.png)\n",
    "\n",
    "Hình trên mô tả các đường được thêm vào mọi cổng, nhưng cũng có những bài báo chỉ thêm cho một vài cổng mà thôi.\n",
    "\n",
    "Một biến thể khác là nối 2 cổng loại trừ và đầu vào với nhau. Thay vì phân tách các quyết định thông tin loại trừ và thông tin mới thêm vào, ta sẽ quyết định chúng cùng với nhau luôn. Ta chỉ bỏ đi thông tin khi mà ta thay thế nó bằng thông tin mới đưa vào. Ta chỉ đưa thông tin mới vào khi ta bỏ thông tin cũ nào đó đi.\n",
    "\n",
    "![LSTM3-var-tied](./images/LSTM3-var-tied.png)\n",
    "\n",
    "Một biến thể khá thú vị khác của LSTM là Gated Recurrent Unit, hay GRU được giới thiệu bởi <a href=\"http://arxiv.org/pdf/1406.1078v3.pdf\">Cho, et al. (2014)</a>. Nó kết hợp các cổng loại trừ và đầu vào thành một cổng “cổng cập nhập” (update gate). Nó cũng hợp trạng thái tế bào và trạng thái ẩn với nhau tạo ra một thay đổi khác. Kết quả là mô hình của ta sẽ đơn giản hơn mô hình LSTM chuẩn và ngày càng trở nên phổ biến.\n",
    "\n",
    "![LSTM3-var-GRU](./images/LSTM3-var-GRU.png)\n",
    "\n",
    "GRU chỉ có 2 cổng để chắt lọc thông tin, cổng reset và cổng update. Nó không có trạng thái cell mà chỉ có đầu ra vừa dùng để đưa ra quyết định vừa dùng để thông tin cho các bước tiếp theo.\n",
    "\n",
    "![1_jhi5uOm9PvZfmxvfaCektw](./images/1_jhi5uOm9PvZfmxvfaCektw.png)\n",
    "\n",
    "Do mất đi trạng thái cell, nên tác dụng của 2 cổng trên khá khó để phân biệt rõ ràng, có thể nói là cả 2 cổng đều có tác dụng chắt lọc thông tin từ các đầu vào của cell để đưa ra một output thỏa mãn cả 2 tiêu chí là lưu giữ thông tin quá khứ và có khả năng đưa ra quyết định hiện tại một cách chính xác nhất. Do giảm bớt tính phức tạp nên GRU hoạt động nhanh hơn đôi chút so với LSTM. Về tính hiệu quả thì rất khó để đưa ra kết luận chính xác.\n",
    "\n",
    "Trên đây chỉ là một vài biến thế được chú ý nhiều nhất thôi, thực tế có rất nhiều các biến thể khác nhau của LSTM như Depth Gated RNNs của <a href=\"http://arxiv.org/pdf/1508.03790v2.pdf\">Yao, et al. (2015)</a>. Cũng có những biến thể mà chiến lực xử lý phụ thuộc xa hoàn toàn khác như Clockwork RNNs của <a href=\"http://arxiv.org/pdf/1402.3511v1.pdf\">Koutnik, et al. (2014)</a>.\n",
    "\n",
    "Nếu bạn muốn tìm hiểu xem biến thể nào là tốt nhất và chúng khác nhau thế nào, thì có thể đọc bài so sánh khá hay này của <a href=\"http://arxiv.org/pdf/1503.04069.pdf\">Greff, et al. (2015)</a>. Ngoài ra thì <a href=\"http://jmlr.org/proceedings/papers/v37/jozefowicz15.pdf\">Jozefowicz, et al. (2015)</a> thậm chí còn thử hàng chục nghìn kiến trúc RNN khác nhau và tìm ra một vài mô hình hoạt động tốt hơn cả LSTM ở một số bài toán."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 3. Xây dựng mạng LSTM với Keras\n",
    "Các vấn đề dự đoán chuỗi thời gian (time series) là một vấn đề khó khăn. Trong phần này, chúng ta sẽ khám phá cách triển khai mạng LSTM bằng thư viện deep learning Keras để giải quyết vấn đề dự đoán chuỗi thời gian.\n",
    "\n",
    "Sau khi hoàn thành hướng dẫn này, bạn sẽ biết cách triển khai và phát triển mạng LSTM cho các vấn đề dự đoán chuỗi thời gian của riêng bạn và các vấn đề tương tự khác.\n",
    "\n",
    "Chúng ta sẽ tìm hiểu về bài toán **dự đoán chuỗi thời gian của hành khách hàng không quốc tê**. Trong hướng dẫn này, tôi sẽ phát triển một số dạng LSTM cho một vấn đề dự đoán chuỗi thời gian khác nhau. Vấn đề và cấu hình được chọn cho các mạng LSTM chỉ nhằm mục đích trình diễn mà chúng không được tối ưu hóa.\n",
    "\n",
    "Những ví dụ này sẽ cho bạn thấy chính xác làm thế nào bạn có thể phát triển các mạng LSTM có cấu trúc khác nhau của riêng bạn cho các vấn đề mô hình dự báo chuỗi thời gian.\n",
    "\n",
    "## 3.1 LSTM cho dự đoán chuỗi thời gian đơn biến\n",
    "Chúng ta xem xét data về ngày tháng và lượng khách hàng của một hãng hàng không quốc tế (data `airline-passengers.csv` trong thư mục data cùng jupyter này). Mỗi quan sát được coi là cách nhau 1 tháng nên chúng ta có thể loại trừ cột đầu tiên."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Passengers</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>112</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>118</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>132</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>129</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>121</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>...</th>\n",
       "      <td>...</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>139</th>\n",
       "      <td>606</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>140</th>\n",
       "      <td>508</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>141</th>\n",
       "      <td>461</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>142</th>\n",
       "      <td>390</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>143</th>\n",
       "      <td>432</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>144 rows × 1 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "     Passengers\n",
       "0           112\n",
       "1           118\n",
       "2           132\n",
       "3           129\n",
       "4           121\n",
       "..          ...\n",
       "139         606\n",
       "140         508\n",
       "141         461\n",
       "142         390\n",
       "143         432\n",
       "\n",
       "[144 rows x 1 columns]"
      ]
     },
     "execution_count": 1,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import pandas as pd\n",
    "import matplotlib.pyplot as plt\n",
    "dataset = pd.read_csv('./data/airline-passengers.csv', usecols=[1])\n",
    "dataset"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(dataset)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Để áp dụng LSTM cho bài toán này, chúng ta có thể diễn đạt vấn đề như sau. Đó là, với số lượng khách hàng (tính theo đơn vị hàng ngàn) trong tháng này, dự đoán số lượng khách hàng trong tháng tới là bao nhiêu?\n",
    "\n",
    "Chúng ta có thể viết 1 hàm đơn giản để chuyển đổi cột dữ liệu duy nhất thành tập dữ liệu gồm 2 cột: cột đầu tiên chứa số lượng khách hàng của tháng (t) và cột thứ 2 chứa dữ liệu dự đoán (tháng t+1)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset = dataset.values\n",
    "dataset = dataset.astype('float32')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "LSTM rất nhạy cảm với độ lớn của dữ liệu đầu vào, đặc biệt khi các hàm sigmoid hay tanh được sử dụng. Nên ta có thể scale dữ liệu của ta lại thành khoảng 0 đến 1, hay còn gọi là chuẩn hoá. Chúng ta dễ dàng chuẩn hoá tập dữ liệu bằng cách sử dụng lớp **MinMaxScaler** từ thư viện scikit-learn."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.preprocessing import MinMaxScaler\n",
    "scaler = MinMaxScaler(feature_range=(0, 1))\n",
    "dataset = scaler.fit_transform(dataset)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Với dữ liệu thời gian, chuỗi giá trị rất quan trọng. Chúng ta sẽ chia dữ liệu 67% dùng để train và 33% dùng để test mô hình"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "96 48\n"
     ]
    }
   ],
   "source": [
    "train_size = int(len(dataset)*0.67)\n",
    "test_size = len(dataset) - train_size\n",
    "train, test = dataset[0: train_size, :], dataset[train_size:len(dataset), :]\n",
    "print(len(train), len(test))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Bây giờ chúng ta có thể định nghĩa một hàm để tạo một tập dữ liệu mới, như được mô tả ở trên. Hàm lấy hai tham số là **dataset** là mảng numpy mà chsng ta muốn đổi thành tập dữ liệu và **look_back** là số bước thời gian trước sử dụng làm biến đầu vào để dự đoán khoảng thời gian tiếp theo - mặc định là 1.\n",
    "\n",
    "Mặc định này sẽ tạo ra một tập dữ liệu trong đó X là số lượng hành khách tại thời điểm t và Y là số lượng hành khách tại thời điểm t+1."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "def create_dataset(dataset, look_back=1):\n",
    "    dataX, dataY = [], []\n",
    "    for i in range(len(dataset)-look_back-1):\n",
    "        a = dataset[i:(i+look_back), 0]\n",
    "        dataX.append(a)\n",
    "        dataY.append(dataset[i+look_back, 0])\n",
    "    return np.array(dataX), np.array(dataY)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "look_back = 1\n",
    "trainX, trainY = create_dataset(train, look_back)\n",
    "testX, testY = create_dataset(test, look_back)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[0.01544401],\n",
       "       [0.02702703],\n",
       "       [0.05405405],\n",
       "       [0.04826255],\n",
       "       [0.03281853],\n",
       "       [0.05984557],\n",
       "       [0.08494207],\n",
       "       [0.08494207],\n",
       "       [0.06177607],\n",
       "       [0.02895753],\n",
       "       [0.        ],\n",
       "       [0.02702703],\n",
       "       [0.02123553],\n",
       "       [0.04247104],\n",
       "       [0.07142857],\n",
       "       [0.05984557],\n",
       "       [0.04054055],\n",
       "       [0.08687258],\n",
       "       [0.12741312],\n",
       "       [0.12741312],\n",
       "       [0.10424709],\n",
       "       [0.05598456],\n",
       "       [0.01930502],\n",
       "       [0.06949806],\n",
       "       [0.07915059],\n",
       "       [0.08880308],\n",
       "       [0.14285713],\n",
       "       [0.11389962],\n",
       "       [0.13127413],\n",
       "       [0.14285713],\n",
       "       [0.18339768],\n",
       "       [0.18339768],\n",
       "       [0.15444016],\n",
       "       [0.11196911],\n",
       "       [0.08108109],\n",
       "       [0.1196911 ],\n",
       "       [0.12934363],\n",
       "       [0.14671814],\n",
       "       [0.17181468],\n",
       "       [0.14864865],\n",
       "       [0.15250966],\n",
       "       [0.22007722],\n",
       "       [0.24324325],\n",
       "       [0.26640925],\n",
       "       [0.2027027 ],\n",
       "       [0.16795367],\n",
       "       [0.13127413],\n",
       "       [0.17374519],\n",
       "       [0.17760617],\n",
       "       [0.17760617],\n",
       "       [0.25482625],\n",
       "       [0.25289574],\n",
       "       [0.24131274],\n",
       "       [0.26833975],\n",
       "       [0.3088803 ],\n",
       "       [0.32432434],\n",
       "       [0.25675675],\n",
       "       [0.20656371],\n",
       "       [0.14671814],\n",
       "       [0.18725869],\n",
       "       [0.19305018],\n",
       "       [0.16216215],\n",
       "       [0.25289574],\n",
       "       [0.23745173],\n",
       "       [0.25096524],\n",
       "       [0.3088803 ],\n",
       "       [0.38223937],\n",
       "       [0.36486486],\n",
       "       [0.2992278 ],\n",
       "       [0.24131274],\n",
       "       [0.1911197 ],\n",
       "       [0.24131274],\n",
       "       [0.26640925],\n",
       "       [0.24903473],\n",
       "       [0.31467178],\n",
       "       [0.3185328 ],\n",
       "       [0.32046333],\n",
       "       [0.4073359 ],\n",
       "       [0.5019305 ],\n",
       "       [0.46911195],\n",
       "       [0.40154442],\n",
       "       [0.32818535],\n",
       "       [0.25675675],\n",
       "       [0.3359073 ],\n",
       "       [0.34749034],\n",
       "       [0.33397684],\n",
       "       [0.41119692],\n",
       "       [0.4034749 ],\n",
       "       [0.4131274 ],\n",
       "       [0.52123547],\n",
       "       [0.5965251 ],\n",
       "       [0.58108103],\n",
       "       [0.484556  ],\n",
       "       [0.3899614 ]], dtype=float32)"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "trainX"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(94, 1, 1)"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "trainX.shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Mạng LSTM hy vọng dữ liệu đầu vào (X) sẽ được cung cấp với cấu trúc mảng cụ thể dưới dạng: *[sample, timestep, feature ]*\n",
    "\n",
    "Hiện tại, dữ liệu của tôi ở dạng: *[ sample, featuree ]* và tôi đang đóng khung vấn đề dưới dạng một bước cho mỗi mẫu. Chúng ta có thể chuyển đổi dữ liệu đầu vào và kiểm tra thành cấu trúc dự kiến bằng cách sử dụng **numpy.reshape()** như sau:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "# reshape input to be [samples, time steps, features]\n",
    "trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))\n",
    "testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[[0.01544401]],\n",
       "\n",
       "       [[0.02702703]],\n",
       "\n",
       "       [[0.05405405]],\n",
       "\n",
       "       [[0.04826255]],\n",
       "\n",
       "       [[0.03281853]],\n",
       "\n",
       "       [[0.05984557]],\n",
       "\n",
       "       [[0.08494207]],\n",
       "\n",
       "       [[0.08494207]],\n",
       "\n",
       "       [[0.06177607]],\n",
       "\n",
       "       [[0.02895753]],\n",
       "\n",
       "       [[0.        ]],\n",
       "\n",
       "       [[0.02702703]],\n",
       "\n",
       "       [[0.02123553]],\n",
       "\n",
       "       [[0.04247104]],\n",
       "\n",
       "       [[0.07142857]],\n",
       "\n",
       "       [[0.05984557]],\n",
       "\n",
       "       [[0.04054055]],\n",
       "\n",
       "       [[0.08687258]],\n",
       "\n",
       "       [[0.12741312]],\n",
       "\n",
       "       [[0.12741312]],\n",
       "\n",
       "       [[0.10424709]],\n",
       "\n",
       "       [[0.05598456]],\n",
       "\n",
       "       [[0.01930502]],\n",
       "\n",
       "       [[0.06949806]],\n",
       "\n",
       "       [[0.07915059]],\n",
       "\n",
       "       [[0.08880308]],\n",
       "\n",
       "       [[0.14285713]],\n",
       "\n",
       "       [[0.11389962]],\n",
       "\n",
       "       [[0.13127413]],\n",
       "\n",
       "       [[0.14285713]],\n",
       "\n",
       "       [[0.18339768]],\n",
       "\n",
       "       [[0.18339768]],\n",
       "\n",
       "       [[0.15444016]],\n",
       "\n",
       "       [[0.11196911]],\n",
       "\n",
       "       [[0.08108109]],\n",
       "\n",
       "       [[0.1196911 ]],\n",
       "\n",
       "       [[0.12934363]],\n",
       "\n",
       "       [[0.14671814]],\n",
       "\n",
       "       [[0.17181468]],\n",
       "\n",
       "       [[0.14864865]],\n",
       "\n",
       "       [[0.15250966]],\n",
       "\n",
       "       [[0.22007722]],\n",
       "\n",
       "       [[0.24324325]],\n",
       "\n",
       "       [[0.26640925]],\n",
       "\n",
       "       [[0.2027027 ]],\n",
       "\n",
       "       [[0.16795367]],\n",
       "\n",
       "       [[0.13127413]],\n",
       "\n",
       "       [[0.17374519]],\n",
       "\n",
       "       [[0.17760617]],\n",
       "\n",
       "       [[0.17760617]],\n",
       "\n",
       "       [[0.25482625]],\n",
       "\n",
       "       [[0.25289574]],\n",
       "\n",
       "       [[0.24131274]],\n",
       "\n",
       "       [[0.26833975]],\n",
       "\n",
       "       [[0.3088803 ]],\n",
       "\n",
       "       [[0.32432434]],\n",
       "\n",
       "       [[0.25675675]],\n",
       "\n",
       "       [[0.20656371]],\n",
       "\n",
       "       [[0.14671814]],\n",
       "\n",
       "       [[0.18725869]],\n",
       "\n",
       "       [[0.19305018]],\n",
       "\n",
       "       [[0.16216215]],\n",
       "\n",
       "       [[0.25289574]],\n",
       "\n",
       "       [[0.23745173]],\n",
       "\n",
       "       [[0.25096524]],\n",
       "\n",
       "       [[0.3088803 ]],\n",
       "\n",
       "       [[0.38223937]],\n",
       "\n",
       "       [[0.36486486]],\n",
       "\n",
       "       [[0.2992278 ]],\n",
       "\n",
       "       [[0.24131274]],\n",
       "\n",
       "       [[0.1911197 ]],\n",
       "\n",
       "       [[0.24131274]],\n",
       "\n",
       "       [[0.26640925]],\n",
       "\n",
       "       [[0.24903473]],\n",
       "\n",
       "       [[0.31467178]],\n",
       "\n",
       "       [[0.3185328 ]],\n",
       "\n",
       "       [[0.32046333]],\n",
       "\n",
       "       [[0.4073359 ]],\n",
       "\n",
       "       [[0.5019305 ]],\n",
       "\n",
       "       [[0.46911195]],\n",
       "\n",
       "       [[0.40154442]],\n",
       "\n",
       "       [[0.32818535]],\n",
       "\n",
       "       [[0.25675675]],\n",
       "\n",
       "       [[0.3359073 ]],\n",
       "\n",
       "       [[0.34749034]],\n",
       "\n",
       "       [[0.33397684]],\n",
       "\n",
       "       [[0.41119692]],\n",
       "\n",
       "       [[0.4034749 ]],\n",
       "\n",
       "       [[0.4131274 ]],\n",
       "\n",
       "       [[0.52123547]],\n",
       "\n",
       "       [[0.5965251 ]],\n",
       "\n",
       "       [[0.58108103]],\n",
       "\n",
       "       [[0.484556  ]],\n",
       "\n",
       "       [[0.3899614 ]]], dtype=float32)"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "trainX"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Bây giờ chúng ta đã sẵn sàng để thiết kế và phù hợp với mạng LSTM của chúng ta cho vấn đề này.\n",
    "\n",
    "Mạng có một lớp có thể nhìn thấy với 1 đầu vào, một lớp ẩn với 4 khối LSTM hoặc nơ-ron và một lớp đầu ra tạo ra một dự đoán giá trị duy nhất. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Để đảm bảo kết quả có thể lặp lại (ngẫu nhiên nhưng lần chạy sau vẫn giữ như vậy), chúng ta dùng lệnh sau:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "# fix random seed for reproducibility\n",
    "np.random.seed(7)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WARNING:tensorflow:From c:\\program files\\python37\\lib\\site-packages\\tensorflow_core\\python\\ops\\resource_variable_ops.py:1630: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version.\n",
      "Instructions for updating:\n",
      "If using Keras pass *_constraint arguments to layers.\n",
      "Model: \"sequential_1\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "lstm_1 (LSTM)                (None, 4)                 96        \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 1)                 5         \n",
      "=================================================================\n",
      "Total params: 101\n",
      "Trainable params: 101\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "WARNING:tensorflow:From c:\\program files\\python37\\lib\\site-packages\\keras\\backend\\tensorflow_backend.py:422: The name tf.global_variables is deprecated. Please use tf.compat.v1.global_variables instead.\n",
      "\n",
      "Epoch 1/50\n",
      "94/94 [==============================] - 1s 12ms/step - loss: 0.0409\n",
      "Epoch 2/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0197\n",
      "Epoch 3/50\n",
      "94/94 [==============================] - 1s 5ms/step - loss: 0.0142\n",
      "Epoch 4/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0127\n",
      "Epoch 5/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0117\n",
      "Epoch 6/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0106\n",
      "Epoch 7/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0097\n",
      "Epoch 8/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0087\n",
      "Epoch 9/50\n",
      "94/94 [==============================] - 0s 5ms/step - loss: 0.0076\n",
      "Epoch 10/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0065\n",
      "Epoch 11/50\n",
      "94/94 [==============================] - 1s 5ms/step - loss: 0.0057 ETA: 0s - loss:\n",
      "Epoch 12/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0048\n",
      "Epoch 13/50\n",
      "94/94 [==============================] - 0s 5ms/step - loss: 0.0041\n",
      "Epoch 14/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0035\n",
      "Epoch 15/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0030\n",
      "Epoch 16/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0027\n",
      "Epoch 17/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0025\n",
      "Epoch 18/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0023\n",
      "Epoch 19/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0022\n",
      "Epoch 20/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 21/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 22/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 23/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 24/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 25/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 26/50\n",
      "94/94 [==============================] - ETA: 0s - loss: 0.002 - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 27/50\n",
      "94/94 [==============================] - 0s 5ms/step - loss: 0.0020\n",
      "Epoch 28/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020A: 0s - los\n",
      "Epoch 29/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 30/50\n",
      "94/94 [==============================] - 1s 7ms/step - loss: 0.0021\n",
      "Epoch 31/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 32/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 33/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 34/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 35/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 36/50\n",
      "94/94 [==============================] - 1s 5ms/step - loss: 0.0020\n",
      "Epoch 37/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 38/50\n",
      "94/94 [==============================] - 1s 8ms/step - loss: 0.0020\n",
      "Epoch 39/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 40/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 41/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 42/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 43/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 44/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 45/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0021\n",
      "Epoch 46/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 47/50\n",
      "94/94 [==============================] - 1s 7ms/step - loss: 0.0020\n",
      "Epoch 48/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 49/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n",
      "Epoch 50/50\n",
      "94/94 [==============================] - 1s 6ms/step - loss: 0.0020\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.callbacks.History at 0x17ec7c85c88>"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from keras.models import Sequential\n",
    "from keras.layers import Dense\n",
    "from keras.layers import LSTM\n",
    "# create and fit the LSTM network\n",
    "model = Sequential()\n",
    "model.add(LSTM(4, input_shape=(1, look_back)))\n",
    "model.add(Dense(1))\n",
    "model.compile(loss='mean_squared_error', optimizer='adam')\n",
    "model.summary()\n",
    "model.fit(trainX, trainY, epochs=50, batch_size=1, verbose=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Cuối cùng, chúng ta có thể tạo dự đoán bằng cách sử dụng mô hình cho cả tập dữ liệu thử nghiệm và train để có được một dấu hiệu trực quan về kỹ năng của mô hình.  hiển thị tập dữ liệu gốc màu xanh lam, các dự đoán cho tập dữ liệu huấn luyện có màu xanh lục và các dự đoán về tập dữ liệu kiểm tra không nhìn thấy có màu đỏ."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "trainPredict = model.predict(trainX)\n",
    "testPredict = model.predict(testX)\n",
    "# invert predictions\n",
    "trainPredict = scaler.inverse_transform(trainPredict)\n",
    "testPredict = scaler.inverse_transform(testPredict)\n",
    "# shift train predictions for plotting\n",
    "trainPredictPlot = np.empty_like(dataset)\n",
    "trainPredictPlot[:, :] = np.nan\n",
    "trainPredictPlot[look_back:len(trainPredict)+look_back, :] = trainPredict\n",
    "# shift test predictions for plotting\n",
    "testPredictPlot = np.empty_like(dataset)\n",
    "testPredictPlot[:, :] = np.nan\n",
    "testPredictPlot[len(trainPredict)+(look_back*2)+1:len(dataset)-1, :] = testPredict\n",
    "# plot baseline and predictions\n",
    "plt.plot(scaler.inverse_transform(dataset))\n",
    "plt.plot(trainPredictPlot)\n",
    "plt.plot(testPredictPlot)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Chúng ta có thể thấy rằng mô hình đã làm tuyệt vời để dự đoán phù hợp với cả bộ dữ liệu đào tạo và thử nghiệm."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 3.2 LSTM cho dự đoán chuỗi thời gian sử dụng phương pháp cửa sổ (Window method)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Chúng ta có thể diễn đạt vấn đề bằng cách sử dụng nhiều bước thời gian gần đây để đưa ra dự đoán cho bước tiếp theo. Đây được gọi là cửa sổ và kích thước cửa sổ là một tham số có thể điều chỉnh. Ví dụ với thời gian hiện tại t, chúng ta muốn dự đoán giá trị ở lần tiếp theo t+1, chúng ta có thể sử dụng thời gian hiện tại t, và có thể sử dụng cả 2 lần bước trước là t-1 và t-2 là các biến đầu vào.\n",
    "\n",
    "Hàm **created_dataset()** mà chúng ta đã tạo trong phần trước cho phép chúng ta tạo công thức này cho bài toán chuỗi thời gian bằng cách tăng đối số **look_back** từ 1 lên 3.\n",
    "\n",
    "Chúng ta có thể chạy lại ví dụ trong phần trước với kích thước cửa sổ lớn hơn."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential_3\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "lstm_3 (LSTM)                (None, 4)                 128       \n",
      "_________________________________________________________________\n",
      "dense_3 (Dense)              (None, 1)                 5         \n",
      "=================================================================\n",
      "Total params: 133\n",
      "Trainable params: 133\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Epoch 1/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0540\n",
      "Epoch 2/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0256\n",
      "Epoch 3/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0164\n",
      "Epoch 4/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0142\n",
      "Epoch 5/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0126\n",
      "Epoch 6/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0113\n",
      "Epoch 7/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0101\n",
      "Epoch 8/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0091\n",
      "Epoch 9/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0078\n",
      "Epoch 10/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0071\n",
      "Epoch 11/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0064\n",
      "Epoch 12/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0059\n",
      "Epoch 13/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0055\n",
      "Epoch 14/50\n",
      "92/92 [==============================] - 1s 5ms/step - loss: 0.0051\n",
      "Epoch 15/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0049\n",
      "Epoch 16/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0048\n",
      "Epoch 17/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0045\n",
      "Epoch 18/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0045\n",
      "Epoch 19/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0044\n",
      "Epoch 20/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0043\n",
      "Epoch 21/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0043\n",
      "Epoch 22/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0042\n",
      "Epoch 23/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0041\n",
      "Epoch 24/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0041\n",
      "Epoch 25/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0040\n",
      "Epoch 26/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0041\n",
      "Epoch 27/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0040\n",
      "Epoch 28/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0040\n",
      "Epoch 29/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0039\n",
      "Epoch 30/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0038\n",
      "Epoch 31/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0038\n",
      "Epoch 32/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0037\n",
      "Epoch 33/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0036\n",
      "Epoch 34/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0036\n",
      "Epoch 35/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0036\n",
      "Epoch 36/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0034\n",
      "Epoch 37/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0035\n",
      "Epoch 38/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0034\n",
      "Epoch 39/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0034\n",
      "Epoch 40/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0033\n",
      "Epoch 41/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0034\n",
      "Epoch 42/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0033\n",
      "Epoch 43/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0032\n",
      "Epoch 44/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0031\n",
      "Epoch 45/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0030\n",
      "Epoch 46/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0032\n",
      "Epoch 47/50\n",
      "92/92 [==============================] - 0s 5ms/step - loss: 0.0029\n",
      "Epoch 48/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0029\n",
      "Epoch 49/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0029\n",
      "Epoch 50/50\n",
      "92/92 [==============================] - 1s 6ms/step - loss: 0.0028\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# convert an array of values into a dataset matrix\n",
    "def create_dataset(dataset, look_back=1):\n",
    "\tdataX, dataY = [], []\n",
    "\tfor i in range(len(dataset)-look_back-1):\n",
    "\t\ta = dataset[i:(i+look_back), 0]\n",
    "\t\tdataX.append(a)\n",
    "\t\tdataY.append(dataset[i + look_back, 0])\n",
    "\treturn np.array(dataX), np.array(dataY)\n",
    "# fix random seed for reproducibility\n",
    "np.random.seed(7)\n",
    "# load the dataset\n",
    "dataframe = pd.read_csv('./data/airline-passengers.csv', usecols=[1])\n",
    "dataset = dataframe.values\n",
    "dataset = dataset.astype('float32')\n",
    "# normalize the dataset\n",
    "scaler = MinMaxScaler(feature_range=(0, 1))\n",
    "dataset = scaler.fit_transform(dataset)\n",
    "# split into train and test sets\n",
    "train_size = int(len(dataset) * 0.67)\n",
    "test_size = len(dataset) - train_size\n",
    "train, test = dataset[0:train_size,:], dataset[train_size:len(dataset),:]\n",
    "# reshape into X=t and Y=t+1\n",
    "look_back = 3\n",
    "trainX, trainY = create_dataset(train, look_back)\n",
    "testX, testY = create_dataset(test, look_back)\n",
    "# reshape input to be [samples, time steps, features]\n",
    "trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))\n",
    "testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))\n",
    "# create and fit the LSTM network\n",
    "model = Sequential()\n",
    "model.add(LSTM(4, input_shape=(1, look_back)))\n",
    "model.add(Dense(1))\n",
    "model.compile(loss='mean_squared_error', optimizer='adam')\n",
    "model.summary()\n",
    "model.fit(trainX, trainY, epochs=50, batch_size=1, verbose=1)\n",
    "# make predictions\n",
    "trainPredict = model.predict(trainX)\n",
    "testPredict = model.predict(testX)\n",
    "# invert predictions\n",
    "trainPredict = scaler.inverse_transform(trainPredict)\n",
    "testPredict = scaler.inverse_transform(testPredict)\n",
    "# shift train predictions for plotting\n",
    "trainPredictPlot = np.empty_like(dataset)\n",
    "trainPredictPlot[:, :] = np.nan\n",
    "trainPredictPlot[look_back:len(trainPredict)+look_back, :] = trainPredict\n",
    "# shift test predictions for plotting\n",
    "testPredictPlot = np.empty_like(dataset)\n",
    "testPredictPlot[:, :] = np.nan\n",
    "testPredictPlot[len(trainPredict)+(look_back*2)+1:len(dataset)-1, :] = testPredict\n",
    "# plot baseline and predictions\n",
    "plt.plot(scaler.inverse_transform(dataset))\n",
    "plt.plot(trainPredictPlot)\n",
    "plt.plot(testPredictPlot)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([[[0.01544401, 0.02702703, 0.05405405]],\n",
       "\n",
       "       [[0.02702703, 0.05405405, 0.04826255]],\n",
       "\n",
       "       [[0.05405405, 0.04826255, 0.03281853]],\n",
       "\n",
       "       [[0.04826255, 0.03281853, 0.05984557]],\n",
       "\n",
       "       [[0.03281853, 0.05984557, 0.08494207]],\n",
       "\n",
       "       [[0.05984557, 0.08494207, 0.08494207]],\n",
       "\n",
       "       [[0.08494207, 0.08494207, 0.06177607]],\n",
       "\n",
       "       [[0.08494207, 0.06177607, 0.02895753]],\n",
       "\n",
       "       [[0.06177607, 0.02895753, 0.        ]],\n",
       "\n",
       "       [[0.02895753, 0.        , 0.02702703]],\n",
       "\n",
       "       [[0.        , 0.02702703, 0.02123553]],\n",
       "\n",
       "       [[0.02702703, 0.02123553, 0.04247104]],\n",
       "\n",
       "       [[0.02123553, 0.04247104, 0.07142857]],\n",
       "\n",
       "       [[0.04247104, 0.07142857, 0.05984557]],\n",
       "\n",
       "       [[0.07142857, 0.05984557, 0.04054055]],\n",
       "\n",
       "       [[0.05984557, 0.04054055, 0.08687258]],\n",
       "\n",
       "       [[0.04054055, 0.08687258, 0.12741312]],\n",
       "\n",
       "       [[0.08687258, 0.12741312, 0.12741312]],\n",
       "\n",
       "       [[0.12741312, 0.12741312, 0.10424709]],\n",
       "\n",
       "       [[0.12741312, 0.10424709, 0.05598456]],\n",
       "\n",
       "       [[0.10424709, 0.05598456, 0.01930502]],\n",
       "\n",
       "       [[0.05598456, 0.01930502, 0.06949806]],\n",
       "\n",
       "       [[0.01930502, 0.06949806, 0.07915059]],\n",
       "\n",
       "       [[0.06949806, 0.07915059, 0.08880308]],\n",
       "\n",
       "       [[0.07915059, 0.08880308, 0.14285713]],\n",
       "\n",
       "       [[0.08880308, 0.14285713, 0.11389962]],\n",
       "\n",
       "       [[0.14285713, 0.11389962, 0.13127413]],\n",
       "\n",
       "       [[0.11389962, 0.13127413, 0.14285713]],\n",
       "\n",
       "       [[0.13127413, 0.14285713, 0.18339768]],\n",
       "\n",
       "       [[0.14285713, 0.18339768, 0.18339768]],\n",
       "\n",
       "       [[0.18339768, 0.18339768, 0.15444016]],\n",
       "\n",
       "       [[0.18339768, 0.15444016, 0.11196911]],\n",
       "\n",
       "       [[0.15444016, 0.11196911, 0.08108109]],\n",
       "\n",
       "       [[0.11196911, 0.08108109, 0.1196911 ]],\n",
       "\n",
       "       [[0.08108109, 0.1196911 , 0.12934363]],\n",
       "\n",
       "       [[0.1196911 , 0.12934363, 0.14671814]],\n",
       "\n",
       "       [[0.12934363, 0.14671814, 0.17181468]],\n",
       "\n",
       "       [[0.14671814, 0.17181468, 0.14864865]],\n",
       "\n",
       "       [[0.17181468, 0.14864865, 0.15250966]],\n",
       "\n",
       "       [[0.14864865, 0.15250966, 0.22007722]],\n",
       "\n",
       "       [[0.15250966, 0.22007722, 0.24324325]],\n",
       "\n",
       "       [[0.22007722, 0.24324325, 0.26640925]],\n",
       "\n",
       "       [[0.24324325, 0.26640925, 0.2027027 ]],\n",
       "\n",
       "       [[0.26640925, 0.2027027 , 0.16795367]],\n",
       "\n",
       "       [[0.2027027 , 0.16795367, 0.13127413]],\n",
       "\n",
       "       [[0.16795367, 0.13127413, 0.17374519]],\n",
       "\n",
       "       [[0.13127413, 0.17374519, 0.17760617]],\n",
       "\n",
       "       [[0.17374519, 0.17760617, 0.17760617]],\n",
       "\n",
       "       [[0.17760617, 0.17760617, 0.25482625]],\n",
       "\n",
       "       [[0.17760617, 0.25482625, 0.25289574]],\n",
       "\n",
       "       [[0.25482625, 0.25289574, 0.24131274]],\n",
       "\n",
       "       [[0.25289574, 0.24131274, 0.26833975]],\n",
       "\n",
       "       [[0.24131274, 0.26833975, 0.3088803 ]],\n",
       "\n",
       "       [[0.26833975, 0.3088803 , 0.32432434]],\n",
       "\n",
       "       [[0.3088803 , 0.32432434, 0.25675675]],\n",
       "\n",
       "       [[0.32432434, 0.25675675, 0.20656371]],\n",
       "\n",
       "       [[0.25675675, 0.20656371, 0.14671814]],\n",
       "\n",
       "       [[0.20656371, 0.14671814, 0.18725869]],\n",
       "\n",
       "       [[0.14671814, 0.18725869, 0.19305018]],\n",
       "\n",
       "       [[0.18725869, 0.19305018, 0.16216215]],\n",
       "\n",
       "       [[0.19305018, 0.16216215, 0.25289574]],\n",
       "\n",
       "       [[0.16216215, 0.25289574, 0.23745173]],\n",
       "\n",
       "       [[0.25289574, 0.23745173, 0.25096524]],\n",
       "\n",
       "       [[0.23745173, 0.25096524, 0.3088803 ]],\n",
       "\n",
       "       [[0.25096524, 0.3088803 , 0.38223937]],\n",
       "\n",
       "       [[0.3088803 , 0.38223937, 0.36486486]],\n",
       "\n",
       "       [[0.38223937, 0.36486486, 0.2992278 ]],\n",
       "\n",
       "       [[0.36486486, 0.2992278 , 0.24131274]],\n",
       "\n",
       "       [[0.2992278 , 0.24131274, 0.1911197 ]],\n",
       "\n",
       "       [[0.24131274, 0.1911197 , 0.24131274]],\n",
       "\n",
       "       [[0.1911197 , 0.24131274, 0.26640925]],\n",
       "\n",
       "       [[0.24131274, 0.26640925, 0.24903473]],\n",
       "\n",
       "       [[0.26640925, 0.24903473, 0.31467178]],\n",
       "\n",
       "       [[0.24903473, 0.31467178, 0.3185328 ]],\n",
       "\n",
       "       [[0.31467178, 0.3185328 , 0.32046333]],\n",
       "\n",
       "       [[0.3185328 , 0.32046333, 0.4073359 ]],\n",
       "\n",
       "       [[0.32046333, 0.4073359 , 0.5019305 ]],\n",
       "\n",
       "       [[0.4073359 , 0.5019305 , 0.46911195]],\n",
       "\n",
       "       [[0.5019305 , 0.46911195, 0.40154442]],\n",
       "\n",
       "       [[0.46911195, 0.40154442, 0.32818535]],\n",
       "\n",
       "       [[0.40154442, 0.32818535, 0.25675675]],\n",
       "\n",
       "       [[0.32818535, 0.25675675, 0.3359073 ]],\n",
       "\n",
       "       [[0.25675675, 0.3359073 , 0.34749034]],\n",
       "\n",
       "       [[0.3359073 , 0.34749034, 0.33397684]],\n",
       "\n",
       "       [[0.34749034, 0.33397684, 0.41119692]],\n",
       "\n",
       "       [[0.33397684, 0.41119692, 0.4034749 ]],\n",
       "\n",
       "       [[0.41119692, 0.4034749 , 0.4131274 ]],\n",
       "\n",
       "       [[0.4034749 , 0.4131274 , 0.52123547]],\n",
       "\n",
       "       [[0.4131274 , 0.52123547, 0.5965251 ]],\n",
       "\n",
       "       [[0.52123547, 0.5965251 , 0.58108103]],\n",
       "\n",
       "       [[0.5965251 , 0.58108103, 0.484556  ]],\n",
       "\n",
       "       [[0.58108103, 0.484556  , 0.3899614 ]]], dtype=float32)"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "trainX"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 3.3 Deep LSTM\n",
    "Các mạng LSTM có thể được xếp chồng lên nhau trong Keras giống như cách các loại lớp khác có thể được xếp chồng lên nhau. Một bổ sung cho cấu hình được yêu cầu là một lớp LSTM trước mỗi lớp LSTM tiếp theo phải trả về chuỗi. Điều này có thể được thực hiện bằng cách đặt tham số **return_sequences** trên lớp thành **True**."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"sequential_4\"\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "lstm_4 (LSTM)                (None, 1, 4)              128       \n",
      "_________________________________________________________________\n",
      "lstm_5 (LSTM)                (None, 4)                 144       \n",
      "_________________________________________________________________\n",
      "dense_4 (Dense)              (None, 1)                 5         \n",
      "=================================================================\n",
      "Total params: 277\n",
      "Trainable params: 277\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n",
      "Epoch 1/50\n",
      "92/92 [==============================] - 2s 18ms/step - loss: 0.0423\n",
      "Epoch 2/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0192\n",
      "Epoch 3/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0151\n",
      "Epoch 4/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0134\n",
      "Epoch 5/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0120\n",
      "Epoch 6/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0104\n",
      "Epoch 7/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0090\n",
      "Epoch 8/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0073\n",
      "Epoch 9/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0061\n",
      "Epoch 10/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0053\n",
      "Epoch 11/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0047\n",
      "Epoch 12/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0043\n",
      "Epoch 13/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0042\n",
      "Epoch 14/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0042\n",
      "Epoch 15/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0041\n",
      "Epoch 16/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0039\n",
      "Epoch 17/50\n",
      "92/92 [==============================] - 1s 11ms/step - loss: 0.0039\n",
      "Epoch 18/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0039\n",
      "Epoch 19/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0037\n",
      "Epoch 20/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0038\n",
      "Epoch 21/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0037\n",
      "Epoch 22/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0036\n",
      "Epoch 23/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0037\n",
      "Epoch 24/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0036\n",
      "Epoch 25/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0036\n",
      "Epoch 26/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0034\n",
      "Epoch 27/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0035\n",
      "Epoch 28/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0035\n",
      "Epoch 29/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0033\n",
      "Epoch 30/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0034\n",
      "Epoch 31/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0035\n",
      "Epoch 32/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0033\n",
      "Epoch 33/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0032\n",
      "Epoch 34/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0031\n",
      "Epoch 35/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0032\n",
      "Epoch 36/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0030\n",
      "Epoch 37/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0032\n",
      "Epoch 38/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0031\n",
      "Epoch 39/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0030\n",
      "Epoch 40/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0030\n",
      "Epoch 41/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0030\n",
      "Epoch 42/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0030\n",
      "Epoch 43/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0029\n",
      "Epoch 44/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0029\n",
      "Epoch 45/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0027\n",
      "Epoch 46/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0027\n",
      "Epoch 47/50\n",
      "92/92 [==============================] - 1s 10ms/step - loss: 0.0027\n",
      "Epoch 48/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0027\n",
      "Epoch 49/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0025\n",
      "Epoch 50/50\n",
      "92/92 [==============================] - 1s 9ms/step - loss: 0.0026\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.callbacks.History at 0x17f98dd4c88>"
      ]
     },
     "execution_count": 29,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model = Sequential()\n",
    "model.add(LSTM(4, input_shape=(1, look_back), return_sequences=True))\n",
    "model.add(LSTM(4))\n",
    "model.add(Dense(1))\n",
    "model.compile(loss='mean_squared_error', optimizer='adam')\n",
    "model.summary()\n",
    "model.fit(trainX, trainY, epochs=50, batch_size=1, verbose=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "# make predictions\n",
    "trainPredict = model.predict(trainX)\n",
    "testPredict = model.predict(testX)\n",
    "# invert predictions\n",
    "trainPredict = scaler.inverse_transform(trainPredict)\n",
    "testPredict = scaler.inverse_transform(testPredict)\n",
    "# shift train predictions for plotting\n",
    "trainPredictPlot = np.empty_like(dataset)\n",
    "trainPredictPlot[:, :] = np.nan\n",
    "trainPredictPlot[look_back:len(trainPredict)+look_back, :] = trainPredict\n",
    "# shift test predictions for plotting\n",
    "testPredictPlot = np.empty_like(dataset)\n",
    "testPredictPlot[:, :] = np.nan\n",
    "testPredictPlot[len(trainPredict)+(look_back*2)+1:len(dataset)-1, :] = testPredict\n",
    "# plot baseline and predictions\n",
    "plt.plot(scaler.inverse_transform(dataset))\n",
    "plt.plot(trainPredictPlot)\n",
    "plt.plot(testPredictPlot)\n",
    "plt.show()"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}