spelling_bee_RNN.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Spelling Bee"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This notebook starts our deep dive (no pun intended) into NLP by introducing sequence-to-sequence learning on Spelling Bee."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "heading_collapsed": true
   },
   "source": [
    "## Data Stuff\n",
    "\n",
    "We take our data set from [The CMU pronouncing dictionary](https://en.wikipedia.org/wiki/CMU_Pronouncing_Dictionary)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/ubuntu/anaconda3/envs/tensorflow/lib/python3.6/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n",
      "  from ._conv import register_converters as _register_converters\n",
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "%matplotlib inline\n",
    "\n",
    "import importlib\n",
    "import utils2; importlib.reload(utils2)\n",
    "from utils2 import *\n",
    "\n",
    "np.set_printoptions(4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "%mkdir data/spellbee"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "--2018-05-20 06:02:27--  https://raw.githubusercontent.com/Alexir/CMUdict/master/cmudict-0.7b\n",
      "Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 151.101.72.133\n",
      "Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|151.101.72.133|:443... connected.\n",
      "HTTP request sent, awaiting response... 200 OK\n",
      "Length: 3865710 (3.7M) [text/plain]\n",
      "Saving to: ‘data/spellbee/cmudict-0.7b’\n",
      "\n",
      "cmudict-0.7b        100%[===================>]   3.69M  4.46MB/s    in 0.8s    \n",
      "\n",
      "2018-05-20 06:02:29 (4.46 MB/s) - ‘data/spellbee/cmudict-0.7b’ saved [3865710/3865710]\n",
      "\n"
     ]
    }
   ],
   "source": [
    "!wget https://raw.githubusercontent.com/Alexir/CMUdict/master/cmudict-0.7b -P data/spellbee/"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "ZYUGANOV  Z Y UW1 G AA0 N AA0 V\r",
      "\r\n",
      "ZYUGANOV(1)  Z UW1 G AA0 N AA0 V\r",
      "\r\n",
      "ZYUGANOV'S  Z Y UW1 G AA0 N AA0 V Z\r",
      "\r\n",
      "ZYUGANOV'S(1)  Z UW1 G AA0 N AA0 V Z\r",
      "\r\n",
      "ZYWICKI  Z IH0 W IH1 K IY0\r",
      "\r\n",
      "{BRACE  B R EY1 S\r",
      "\r\n",
      "{LEFT-BRACE  L EH1 F T B R EY1 S\r",
      "\r\n",
      "{OPEN-BRACE  OW1 P EH0 N B R EY1 S\r",
      "\r\n",
      "}CLOSE-BRACE  K L OW1 Z B R EY1 S\r",
      "\r\n",
      "}RIGHT-BRACE  R AY1 T B R EY1 S\r",
      "\r\n"
     ]
    }
   ],
   "source": [
    "!tail -n 10 data/spellbee/cmudict-0.7b"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "PATH = 'data/spellbee/'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "limit_mem()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "from sklearn.model_selection import train_test_split"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "The CMU pronouncing dictionary consists of sounds/words and their corresponding phonetic description (American pronunciation).\n",
    "\n",
    "The phonetic descriptions are a sequence of phonemes. Note that the vowels end with integers; these indicate where the stress is.\n",
    "\n",
    "Our goal is to learn how to spell these words given the sequence of phonemes."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "The preparation of this data set follows the same pattern we've seen before for NLP tasks.\n",
    "\n",
    "Here we iterate through each line of the file and grab each word/phoneme pair that starts with an uppercase letter. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "line 50 list: ['ABALONE', 'AE2 B AH0 L OW1 N IY0']\n",
      "lines 50 tuple: ('ABALONE', ['AE2', 'B', 'AH0', 'L', 'OW1', 'N', 'IY0'])\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "(('A', ['AH0']), ('ZYWICKI', ['Z', 'IH0', 'W', 'IH1', 'K', 'IY0']))"
      ]
     },
     "execution_count": 21,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "lines = [l.strip().split(\"  \") for l in open(PATH + \"cmudict-0.7b\", encoding=\"latin1\") if re.match(\"^[A-Z]\", l)]\n",
    "print(\"line 50 list:\", lines[50])\n",
    "lines = [(w, ps.split()) for w, ps in lines]\n",
    "print(\"lines 50 tuple:\", lines[50])\n",
    "lines[0], lines[-1]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Next we're going to get a list of the unique phonemes in our vocabulary, as well as add a null \"_\" for zero-padding."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['_', 'AA0', 'AA1', 'AA2', 'AE0']"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "phonemes = [\"_\"] + sorted(set(p for w, ps in lines for p in ps))\n",
    "phonemes[:5]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "70"
      ]
     },
     "execution_count": 24,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "len(phonemes)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Then we create mappings of phonemes and letters to respective indices.\n",
    "\n",
    "Our letters include the padding element \"_\", but also \"*\" which we'll explain later."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "p2i = dict((v, k) for k, v in enumerate(phonemes))\n",
    "letters = \"_abcdefghijklmnopqrstuvwxyz*\"\n",
    "l2i = dict((v, k) for k, v in enumerate(letters))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Let's create a dictionary mapping words to the sequence of indices corresponding to it's phonemes, and let's do it only for words between 5 and 15 characters long."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "108579"
      ]
     },
     "execution_count": 35,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "maxlen = 15\n",
    "pronounce_dict = { w.lower(): [p2i[p] for p in ps] for w, ps in lines\n",
    "                 if (5 <= len(w) <= maxlen) and re.match(\"^[A-Z]+$\", w) }\n",
    "len(pronounce_dict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Aside on various approaches to python's list comprehension:\n",
    "* the first list is a typical example of a list comprehension subject to a conditional\n",
    "* the second is a list comprehension inside a list comprehension, which returns a list of list\n",
    "* the third is similar to the second, but is read and behaves like a nested loop\n",
    "    * Since there is no inner bracket, there are no lists wrapping the inner loop"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 45,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(['XYZ'], [['x', 'y', 'z'], ['a', 'b', 'c']], ['x', 'y', 'z', 'a', 'b', 'c'])"
      ]
     },
     "execution_count": 45,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "a = [\"xyz\", \"abc\"]\n",
    "# first list comprehension             second list comprehension    third list comprehension\n",
    "[o.upper() for o in a if o[0] == \"x\"], [[p for p in o] for o in a], [p for o in a for p in o]\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Split lines into words, phonemes, convert to indexes (with padding), split into training, validation, test sets. Note we also find the max phoneme sequence length for padding."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "16"
      ]
     },
     "execution_count": 50,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "maxlen_p = max([len(v) for k, v in pronounce_dict.items()])\n",
    "maxlen_p"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 58,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(108579, 16)\n"
     ]
    }
   ],
   "source": [
    "pairs = np.random.permutation(list(pronounce_dict.keys()))\n",
    "n = len(pairs)\n",
    "input_ = np.zeros((n, maxlen_p), np.int32)\n",
    "print(input_.shape)\n",
    "labels_ = np.zeros((n, maxlen), np.int32)\n",
    "\n",
    "for i, k in enumerate(pairs):\n",
    "    for j, p in enumerate(pronounce_dict[k]):\n",
    "        input_[i][j] = p\n",
    "    for j, letter in enumerate(k):\n",
    "        labels_[i][j] = l2i[letter]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 63,
   "metadata": {},
   "outputs": [],
   "source": [
    "go_token = l2i[\"*\"]\n",
    "dec_input_ = np.concatenate([np.ones((n, 1)) * go_token, labels_[:,:-1]], axis=1)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Sklearn's <tt>train_test_split</tt> is an easy way to split data into training and testing sets."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 64,
   "metadata": {},
   "outputs": [],
   "source": [
    "(input_train, input_test, labels_train, labels_test, dec_input_train, dec_input_test\n",
    "    ) = train_test_split(input_, labels_, dec_input_, test_size=0.1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(97721, 16)"
      ]
     },
     "execution_count": 65,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "input_train.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 66,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(97721, 15)"
      ]
     },
     "execution_count": 66,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "labels_train.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 68,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(70, 28)"
      ]
     },
     "execution_count": 68,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "input_vocab_size, output_vocab_size = len(phonemes), len(letters)\n",
    "input_vocab_size, output_vocab_size"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Next we proceed to build our model."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Keras code"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 69,
   "metadata": {},
   "outputs": [],
   "source": [
    "params = { 'verbose': 0, 'callbacks': [TQDMNotebookCallback(leave_inner=True)] }\n",
    "lstm_params = {}"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 70,
   "metadata": {},
   "outputs": [],
   "source": [
    "dim = 240"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Without attention"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 77,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Modified codes for Keras 2. Original codes were for Keras 1.\n",
    "# If not, will receive \"UserWarning: Update your `LSTM` call to the Keras 2 API\"\n",
    "# See Keras 2 release notes: https://github.com/keras-team/keras/wiki/Keras-2.0-release-notes#recurrent-layers\n",
    "# dropout_U -> recurrent_dropout\n",
    "# dropout_W -> dropout\n",
    "# consume_less -> implementation. String values have been replaced with integers: implementation 0 (default), 1 or 2.\n",
    "def get_rnn(return_sequences=True):\n",
    "    return LSTM(dim, recurrent_dropout=0.1, dropout=0.1,\n",
    "               implementation=2, return_sequences=return_sequences)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The model has three parts:\n",
    "* We first pass list of phonemes through an embedding function to get a list of phoneme embeddings. Our goal is to turn this sequence of embeddings into a single distributed representation that captures what our phonemes say.\n",
    "* Turning a sequence into a representation can be done using an RNN. This approach is useful because RNN's are able to keep track of state and memory, which is obviously important in forming a complete understanding of a pronunciation.\n",
    "    * <tt>BiDirectional</tt> passes the original sequence through an RNN, and the reversed sequence through a different RNN and concatenates the results. This allows us to look forward and backwards.\n",
    "    * We do this because in language things that happen later often influence what came before (i.e. in Spanish, \"el chico, la chica\" means the boy, the girl; the word for \"the\" is determined by the gender of the subject, which comes after).\n",
    "* Finally, we arrive at a vector representation of the sequence which captures everything we need to spell it. We feed this vector into more RNN's, which are trying to generate the labels. After this, we make a classification for what each letter is in the output sequence.\n",
    "    * We use <tt>RepeatVector</tt> to help our RNN remember at each point what the original word is that it's trying to translate.\n",
    "    \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 78,
   "metadata": {},
   "outputs": [],
   "source": [
    "inp = Input((maxlen_p,)) # (16,)\n",
    "x = Embedding(input_vocab_size, 120)(inp)\n",
    "\n",
    "x = Bidirectional(get_rnn())(x)\n",
    "x = get_rnn(False)(x)\n",
    "\n",
    "x = RepeatVector(maxlen)(x)\n",
    "x = get_rnn()(x)\n",
    "x = get_rnn()(x)\n",
    "x = TimeDistributed(Dense(output_vocab_size, activation=\"softmax\"))(x)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can refer to the parts of the model before and after <tt>get_rnn(False)</tt> returns a vector as the encoder and decoder. The encoder has taken a sequence of embeddings and encoded it into a numerical vector that completely describes it's input, while the decoder transforms that vector into a new sequence.\n",
    "\n",
    "Now we can fit our model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 79,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = Model(inp, x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 80,
   "metadata": {},
   "outputs": [],
   "source": [
    "model.compile(Adam(), 'sparse_categorical_crossentropy', metrics=['acc'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 83,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "e2d99c3fe26042d899091b94a6685646",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Training', max=3), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "d972ca782dee4cfc9f33c8ebab844f3b",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Epoch 0', max=97721), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "b223f66ef028447faf21cf3930b45667",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Epoch 1', max=97721), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "e345b2685f4d43da83d68995265ad628",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "HBox(children=(IntProgress(value=0, description='Epoch 2', max=97721), HTML(value='')))"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "hist = model.fit(input_train, np.expand_dims(labels_train, -1),\n",
    "                validation_data=[input_test, np.expand_dims(labels_test, -1)],\n",
    "                batch_size=64, **params, epochs=3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 89,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[1.3975943435641542, 0.8072535572383127, 0.5976672048703986]"
      ]
     },
     "execution_count": 89,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "hist.history['val_loss']"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "To evaluate, we don't want to know what percentage of letters are correct but what percentage of words are."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 90,
   "metadata": {},
   "outputs": [],
   "source": [
    "def eval_keras(input):\n",
    "    preds = model.predict(input, batch_size=128)\n",
    "    predict = np.argmax(preds, axis=2)\n",
    "    return (np.mean([all(real == p) for real, p in zip(labels_test, predict)]), predict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The accuracy isn't great."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 91,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.09743967581506723"
      ]
     },
     "execution_count": 91,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "acc, preds = eval_keras(input_test)\n",
    "acc"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 92,
   "metadata": {},
   "outputs": [],
   "source": [
    "def print_examples(preds):\n",
    "    print(\"pronounciation\".ljust(40), \"real spelling\".ljust(17),\n",
    "          \"model spelling\".ljust(17), \"is correct\")\n",
    "    \n",
    "    for index in range(20):\n",
    "        ps = \"-\".join([phonemes[p] for p in input_test[index]])\n",
    "        real = [letters[l] for l in labels_test[index]]\n",
    "        predict = [letters[l] for l in preds[index]]\n",
    "        print (ps.split(\"-_\")[0].ljust(40), \"\".join(real).split(\"_\")[0].ljust(17),\n",
    "            \"\".join(predict).split(\"_\")[0].ljust(17), str(real == predict))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "We can see that sometimes the mistakes are completely reasonable, occasionally they're totally off. This tends to happen with the longer words that have large phoneme sequences.\n",
    "\n",
    "That's understandable; we'd expect larger sequences to lose more information in an encoding."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 93,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "pronounciation                           real spelling     model spelling    is correct\n",
      "K-W-IH1-L-M-Z                            quilmes           quells            False\n",
      "V-EH1-T-ER0                              vetter            vetter            True\n",
      "EH0-Z-EH1-L                              ezell             essll             False\n",
      "P-EH1-SH-K                               peschke           pesck             False\n",
      "B-AA1-R-AH0-K-AA2-T                      barakat           baraccat          False\n",
      "AA0-R-K-EY1-D-Z                          arcades           arcodds           False\n",
      "S-AH0-V-AE1-N-AH0                        savanna           savana            False\n",
      "AH0-SH-AA1-JH-IY0                        ashaji            ashagi            False\n",
      "V-R-IY1-L-AH0-N-D                        vreeland          vrellnnd          False\n",
      "B-R-IH1-NG-K-S                           brinks            brinks            True\n",
      "P-ER1-L-Z                                pearls            perlss            False\n",
      "JH-IY2-AH0-K-AO1-N-D-AA2                 giaconda          giacooda          False\n",
      "S-L-AE1-K-T                              slacked           slackd            False\n",
      "HH-IH1-L-AH0-N-B-R-AE2-N-D               hillenbrand       hillenarnd        False\n",
      "K-ER0-OW1-ZH-AH0-N                       corrosion         carouson          False\n",
      "S-T-AH1-T-ER0-IH0-NG                     stuttering        stetteing         False\n",
      "G-AA1-R-B-AH0-L-D                        garbled           garblld           False\n",
      "P-R-AH0-S-EH1-SH-AH0-N-AH0-L             processional      proseciccal       False\n",
      "K-AH1-Z-N-EH2-T-S                        kuznets           kuenitt           False\n",
      "AH0-S-P-ER1-ZH-AH0-N                     aspersion         osprrssnn         False\n"
     ]
    }
   ],
   "source": [
    "print_examples(preds)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "heading_collapsed": true
   },
   "source": [
    "### Attention model"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "This graph demonstrates the accuracy decay for a neural translation task. With an encoding/decoding technique, larger input sequences result in less accuracy.\n",
    "\n",
    "<img src=\"https://smerity.com/media/images/articles/2016/bahdanau_attn.png\" width=\"600\">\n",
    "\n",
    "This can be mitigated using an attentional model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 105,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "import attention_wrapper; importlib.reload(attention_wrapper)\n",
    "from attention_wrapper import Attention"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "The attentional model doesn't encode into a single vector, but rather a sequence of vectors. The decoder then at every point is passing through this sequence. For example, after the bi-directional RNN we have 16 vectors corresponding to each phoneme's output state. Each output state describes how each phoneme relates between the other phonemes before and after it. After going through more RNN's, our goal is to transform this sequence into a vector of length 15 so we can classify into characters. \n",
    "\n",
    "A smart way to take a weighted average of the 16 vectors for each of the 15 outputs, where each set of weights is unique to the output. For example, if character 1 only needs information from the first phoneme vector, that weight might be 1 and the others 0; if it needed information from the 1st and 2nd equally, those two might be 0.5 each.\n",
    "\n",
    "The weights for combining all the input states to produce specific outputs can be learned using an attentional model; we update the weights using SGD, and train it jointly with the encoder/decoder. Once we have the outputs, we can classify the character using softmax as usual."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Notice below we do not have an RNN that returns a flat vector as we did before; we have a sequence of vectors as desired. We can then pass a sequence of encoded states into the our custom <tt>Attention</tt> model.\n",
    "\n",
    "This attention model also uses a technique called teacher forcing; in addition to passing the encoded hidden state, we also pass the correct answer for the previous time period. We give this information to the model because it makes it easier to train. In the beginning of training, the model will get most things wrong, and if your earlier character predictions are wrong then your later ones will likely be as well. Teacher forcing allows the model to still learn how to predict later characters, even if the earlier characters were all wrong."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 106,
   "metadata": {
    "hidden": true,
    "scrolled": false
   },
   "outputs": [
    {
     "ename": "TypeError",
     "evalue": "add_weight() got multiple values for argument 'shape'",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mTypeError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[0;32m<ipython-input-106-6adcb8c19efe>\u001b[0m in \u001b[0;36m<module>\u001b[0;34m()\u001b[0m\n\u001b[1;32m      8\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mget_rnn\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m      9\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mget_rnn\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 10\u001b[0;31m \u001b[0mx\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mAttention\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mget_rnn\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;36m3\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0memb_dec\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     11\u001b[0m \u001b[0mx\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mTimeDistributed\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mDense\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0moutput_vocab_size\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mactivation\u001b[0m\u001b[0;34m=\u001b[0m\u001b[0;34m'softmax'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mx\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/anaconda3/envs/tensorflow/lib/python3.6/site-packages/keras/engine/topology.py\u001b[0m in \u001b[0;36m__call__\u001b[0;34m(self, inputs, **kwargs)\u001b[0m\n\u001b[1;32m    592\u001b[0m                     \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbuild\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0minput_shapes\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;36m0\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    593\u001b[0m                 \u001b[0;32melse\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m--> 594\u001b[0;31m                     \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbuild\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0minput_shapes\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m    595\u001b[0m                 \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mbuilt\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0;32mTrue\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m    596\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/development/fastai-dl2-2017/attention_wrapper.py\u001b[0m in \u001b[0;36mbuild\u001b[0;34m(self, input_shape)\u001b[0m\n\u001b[1;32m     35\u001b[0m         \u001b[0minit\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0ml0\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mkernel_initializer\u001b[0m \u001b[0;31m#l0.init. migrated to Keras 2\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     36\u001b[0m         \u001b[0mout_dim\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0ml0\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0munits\u001b[0m \u001b[0;31m#l0.output_dim. migrated to Keras 2\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 37\u001b[0;31m         \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mW1\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mw\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0menc_shape\u001b[0m\u001b[0;34m[\u001b[0m\u001b[0;34m-\u001b[0m\u001b[0;36m1\u001b[0m\u001b[0;34m]\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mnb_dims\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minit\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'{}_W1'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     38\u001b[0m         \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mW2\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mw\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mout_dim\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mnb_dims\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minit\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'{}_W2'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     39\u001b[0m         \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mb2\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mw\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mnb_dims\u001b[0m\u001b[0;34m,\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mzero\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m'{}_b2'\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/development/fastai-dl2-2017/attention_wrapper.py\u001b[0m in \u001b[0;36mw\u001b[0;34m(self, dims, init, name)\u001b[0m\n\u001b[1;32m     19\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     20\u001b[0m     \u001b[0;32mdef\u001b[0m \u001b[0mw\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mdims\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minit\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mname\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m:\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0;32m---> 21\u001b[0;31m         \u001b[0;32mreturn\u001b[0m \u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0madd_weight\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mdims\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0minit\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0mname\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mformat\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0mself\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0mname\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     22\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     23\u001b[0m \u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;32m~/anaconda3/envs/tensorflow/lib/python3.6/site-packages/keras/legacy/interfaces.py\u001b[0m in \u001b[0;36mwrapper\u001b[0;34m(*args, **kwargs)\u001b[0m\n\u001b[1;32m     89\u001b[0m                 warnings.warn('Update your `' + object_name +\n\u001b[1;32m     90\u001b[0m                               '` call to the Keras 2 API: ' + signature, stacklevel=2)\n\u001b[0;32m---> 91\u001b[0;31m             \u001b[0;32mreturn\u001b[0m \u001b[0mfunc\u001b[0m\u001b[0;34m(\u001b[0m\u001b[0;34m*\u001b[0m\u001b[0margs\u001b[0m\u001b[0;34m,\u001b[0m \u001b[0;34m**\u001b[0m\u001b[0mkwargs\u001b[0m\u001b[0;34m)\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[0m\u001b[1;32m     92\u001b[0m         \u001b[0mwrapper\u001b[0m\u001b[0;34m.\u001b[0m\u001b[0m_original_function\u001b[0m \u001b[0;34m=\u001b[0m \u001b[0mfunc\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n\u001b[1;32m     93\u001b[0m         \u001b[0;32mreturn\u001b[0m \u001b[0mwrapper\u001b[0m\u001b[0;34m\u001b[0m\u001b[0m\n",
      "\u001b[0;31mTypeError\u001b[0m: add_weight() got multiple values for argument 'shape'"
     ]
    }
   ],
   "source": [
    "inp = Input((maxlen_p,))\n",
    "inp_dec = Input((maxlen,))\n",
    "emb_dec = Embedding(output_vocab_size, 120)(inp_dec)\n",
    "emb_dec = Dense(dim)(emb_dec)\n",
    "\n",
    "x = Embedding(input_vocab_size, 120)(inp)\n",
    "x = Bidirectional(get_rnn())(x)\n",
    "x = get_rnn()(x)\n",
    "x = get_rnn()(x)\n",
    "x = Attention(get_rnn, 3)([x, emb_dec])\n",
    "x = TimeDistributed(Dense(output_vocab_size, activation='softmax'))(x)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "We can now train, passing in the decoder inputs as well for teacher forcing."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 67,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "model = Model([inp, inp_dec], x)\n",
    "model.compile(Adam(), 'sparse_categorical_crossentropy', metrics=['acc'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 68,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n",
      "\n",
      "\n"
     ]
    }
   ],
   "source": [
    "hist=model.fit([input_train, dec_input_train], np.expand_dims(labels_train,-1), \n",
    "          validation_data=[[input_test, dec_input_test], np.expand_dims(labels_test,-1)], \n",
    "          batch_size=64, **parms, nb_epoch=3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "hidden": true,
    "scrolled": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[1.1220142268966762,\n",
       " 0.8048337527904541,\n",
       " 0.5265462693931019,\n",
       " 0.31194811385601234,\n",
       " 0.23548489567190725,\n",
       " 0.22613589595439379,\n",
       " 0.19347934096944011,\n",
       " 0.19403484622484754,\n",
       " 0.18044625614216103,\n",
       " 0.17990882804400168]"
      ]
     },
     "execution_count": 25,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "hist.history['val_loss']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 998,
   "metadata": {
    "collapsed": true,
    "hidden": true
   },
   "outputs": [],
   "source": [
    "K.set_value(model.optimizer.lr, 1e-4)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 999,
   "metadata": {
    "hidden": true,
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "\n"
     ]
    }
   ],
   "source": [
    "hist=model.fit([input_train, dec_input_train], np.expand_dims(labels_train,-1), \n",
    "          validation_data=[[input_test, dec_input_test], np.expand_dims(labels_test,-1)], \n",
    "          batch_size=64, **parms, nb_epoch=5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1000,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "array([ 0.1591,  0.1563,  0.1532,  0.1517,  0.1499])"
      ]
     },
     "execution_count": 1000,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.array(hist.history['val_loss'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1001,
   "metadata": {
    "collapsed": true,
    "hidden": true
   },
   "outputs": [],
   "source": [
    "def eval_keras():\n",
    "    preds = model.predict([input_test, dec_input_test], batch_size=128)\n",
    "    predict = np.argmax(preds, axis = 2)\n",
    "    return (np.mean([all(real==p) for real, p in zip(labels_test, predict)]), predict)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "Better accuracy!"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 895,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.51134154244977315"
      ]
     },
     "execution_count": 895,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "acc, preds = eval_keras(); acc"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "hidden": true
   },
   "source": [
    "This model is certainly performing better with longer words. The mistakes it's making are reasonable, and it even succesfully formed the word \"partisanship\"."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 896,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "pronunciation                            real spelling     model spelling    is correct\n",
      "D-AY0-AE1-F-AH0-N-IH0-S                  diaphanous        diaphoneus        False\n",
      "T-IH1-R-IY0                              teary             tiary             False\n",
      "M-IH1-T-AH0-N                            mittan            mitton            False\n",
      "P-AA1-R-N-AH0-S                          parness           parnass           False\n",
      "SH-UW1-ER0-M-AH0-N                       schuermann        schuerman         False\n",
      "P-AA1-R-T-AH0-Z-AH0-N-SH-IH2-P           partisanship      partisanship      True\n",
      "AE1-SH-IH0-Z                             ashes             ashes             True\n",
      "S-P-IY1-K                                speak             speek             False\n",
      "K-AO1-R-B-IH0-T                          corbitt           corbit            False\n",
      "B-IH1-L-D-ER0-Z                          builders          biilders          False\n",
      "EH0-S-P-R-IY1                            esprit            espree            False\n",
      "EH0-NG-G-EH0-L-B-EH1-R-T-AH0             engelberta        engelberta        True\n",
      "AE0-N-T-AY2-P-ER0-S-AH0-N-EH1-L          antipersonnel     antipersoneal     False\n",
      "W-IH1-M-Z                                whims             whimm             False\n",
      "AO2-L-T-AH0-G-EH1-DH-ER0                 altogether        altagether        False\n",
      "AE1-K-S-AH0-L-Z                          axles             axees             False\n",
      "D-UW1-S                                  doose             doose             True\n",
      "B-R-AH0-Z-IH1-L-Y-AH0-N                  brazilian         brazilian         True\n",
      "G-R-EH1-N-Z                              grenz             grenz             True\n",
      "B-AH1-K-HH-AE2-N-T-S                     buckhantz         buckhants         False\n"
     ]
    }
   ],
   "source": [
    "print(\"pronunciation\".ljust(40), \"real spelling\".ljust(17), \n",
    "      \"model spelling\".ljust(17), \"is correct\")\n",
    "\n",
    "for index in range(20):\n",
    "    ps = \"-\".join([phonemes[p] for p in input_test[index]]) \n",
    "    real = [letters[l] for l in labels_test[index]] \n",
    "    predict = [letters[l] for l in preds[index]]\n",
    "    print (ps.split(\"-_\")[0].ljust(40), \"\".join(real).split(\"_\")[0].ljust(17),\n",
    "        \"\".join(predict).split(\"_\")[0].ljust(17), str(real == predict))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "collapsed": true,
    "heading_collapsed": true
   },
   "source": [
    "## Test code for the attention layer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 301,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "nb_samples, nb_time, input_dim, output_dim = (64, 4, 32, 48)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 302,
   "metadata": {
    "collapsed": true,
    "hidden": true
   },
   "outputs": [],
   "source": [
    "x = tf.placeholder(np.float32, (nb_samples, nb_time, input_dim))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 303,
   "metadata": {
    "collapsed": true,
    "hidden": true
   },
   "outputs": [],
   "source": [
    "xr = K.reshape(x,(-1,nb_time,1,input_dim))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 304,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([Dimension(32), Dimension(32)])"
      ]
     },
     "execution_count": 304,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "W1 = tf.placeholder(np.float32, (input_dim, input_dim)); W1.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 305,
   "metadata": {
    "collapsed": true,
    "hidden": true
   },
   "outputs": [],
   "source": [
    "W1r = K.reshape(W1, (1, input_dim, input_dim))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 306,
   "metadata": {
    "collapsed": true,
    "hidden": true
   },
   "outputs": [],
   "source": [
    "W1r2 = K.reshape(W1, (1, 1, input_dim, input_dim))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 307,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([Dimension(64), Dimension(4), Dimension(32)])"
      ]
     },
     "execution_count": 307,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "xW1 = K.conv1d(x,W1r,border_mode='same'); xW1.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 308,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([Dimension(64), Dimension(4), Dimension(1), Dimension(32)])"
      ]
     },
     "execution_count": 308,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "xW12 = K.conv2d(xr,W1r2,border_mode='same'); xW12.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 251,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "xW2 = K.dot(x, W1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 245,
   "metadata": {
    "collapsed": true,
    "hidden": true
   },
   "outputs": [],
   "source": [
    "x1 = np.random.normal(size=(nb_samples, nb_time, input_dim))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 246,
   "metadata": {
    "collapsed": true,
    "hidden": true
   },
   "outputs": [],
   "source": [
    "w1 = np.random.normal(size=(input_dim, input_dim))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 248,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "res = sess.run(xW1, {x:x1, W1:w1})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 252,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "res2 = sess.run(xW2, {x:x1, W1:w1})"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 253,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "True"
      ]
     },
     "execution_count": 253,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "np.allclose(res, res2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 283,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([Dimension(48), Dimension(32)])"
      ]
     },
     "execution_count": 283,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "W2 = tf.placeholder(np.float32, (output_dim, input_dim)); W2.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 295,
   "metadata": {
    "hidden": true
   },
   "outputs": [],
   "source": [
    "h = tf.placeholder(np.float32, (nb_samples, output_dim))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 296,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([Dimension(64), Dimension(32)])"
      ]
     },
     "execution_count": 296,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "hW2 = K.dot(h,W2); hW2.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 297,
   "metadata": {
    "hidden": true
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "TensorShape([Dimension(64), Dimension(1), Dimension(1), Dimension(32)])"
      ]
     },
     "execution_count": 297,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "hW2 = K.reshape(hW2,(-1,1,1,input_dim)); hW2.shape"
   ]
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "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": "3.6.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}