rnn_memo.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import warnings\n",
    "warnings.simplefilter(action='ignore', category=FutureWarning)\n",
    "import keras"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "# data数\n",
    "N_DATA = 1234\n",
    "# 特徴数\n",
    "N_FEATURE = 122\n",
    "# 時刻幅\n",
    "TIME_STEP = 100\n",
    "# class数\n",
    "N_CLASS = 2"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((1234, 100, 122), (1234, 2))"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import numpy as np\n",
    "from keras.utils import to_categorical\n",
    "X_train = np.random.random(size=(N_DATA, TIME_STEP, N_FEATURE))\n",
    "y_train = to_categorical(np.random.randint(N_CLASS, size=N_DATA), N_CLASS)\n",
    "X_train.shape, y_train.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "simple_rnn_1 (SimpleRNN)     (None, 100)               22300     \n",
      "_________________________________________________________________\n",
      "dense_1 (Dense)              (None, 2)                 202       \n",
      "=================================================================\n",
      "Total params: 22,502\n",
      "Trainable params: 22,502\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from keras.models import Sequential\n",
    "from keras.layers import Dense, SimpleRNN\n",
    "from keras.optimizers import Adam\n",
    "\n",
    "# 時刻幅に対して1つのラベル\n",
    "model = Sequential()\n",
    "model.add(SimpleRNN(100, input_shape=(TIME_STEP, N_FEATURE)))\n",
    "model.add(Dense(N_CLASS, activation='softmax'))\n",
    "model.compile(\n",
    "    loss='categorical_crossentropy',\n",
    "    optimizer=Adam(lr=1e-3),\n",
    "    metrics=['accuracy']\n",
    ")\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      "1234/1234 [==============================] - 4s 3ms/step - loss: 0.7240 - acc: 0.4968\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f4a905bcfd0>"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model.fit(\n",
    "    X_train, y_train,\n",
    "    batch_size=32,\n",
    "    epochs=1,\n",
    "    verbose=1\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1234, 122)"
      ]
     },
     "execution_count": 6,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 元データからの変換\n",
    "X_data = np.arange(N_DATA*N_FEATURE).reshape(N_DATA, N_FEATURE)\n",
    "X_data.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(12, 100, 122)"
      ]
     },
     "execution_count": 7,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# オーバラップ無しで分割\n",
    "np.array(list(zip(*[iter(X_data)]*TIME_STEP))).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1134, 100, 122)"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 1時刻ずつずらして分割\n",
    "def sliding_window(a, step):\n",
    "    return np.lib.stride_tricks.as_strided(a, shape=(a.shape[0]-step, step, a.shape[-1]), strides=a.strides + (a.strides[-1]*step,))\n",
    "\n",
    "sliding_window(X_data, TIME_STEP).shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1234, 100, 122)"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from keras.preprocessing.sequence import pad_sequences\n",
    "\n",
    "# 長さが異なる場合\n",
    "sequences = [\n",
    "    np.random.random(size=(i+1, N_FEATURE)) for i in range(N_DATA)\n",
    "]\n",
    "pad_sequences(\n",
    "    sequences,\n",
    "    maxlen=TIME_STEP,  # 揃えたい長さ\n",
    "    dtype=np.float32,\n",
    "    value=0.0,  # 0埋め\n",
    "    padding='pre',  # maxlenより短い場合に、前をvalueで埋める(後の場合は'post')\n",
    "    truncating='pre',  # maxlenより長い場合に、前を切り捨てる(後の場合は'post')\n",
    ").shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((1234, 100, 122), (1234, 100, 2))"
      ]
     },
     "execution_count": 10,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "X_train2 = np.random.random(size=(N_DATA, TIME_STEP, N_FEATURE))\n",
    "y_train2 = to_categorical(np.random.randint(N_CLASS, size=(N_DATA, TIME_STEP)), N_CLASS)\n",
    "X_train2.shape, y_train2.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "simple_rnn_2 (SimpleRNN)     (None, 100, 100)          22300     \n",
      "_________________________________________________________________\n",
      "time_distributed_1 (TimeDist (None, 100, 2)            202       \n",
      "=================================================================\n",
      "Total params: 22,502\n",
      "Trainable params: 22,502\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from keras.layers import TimeDistributed\n",
    "\n",
    "# 時刻に対して1つのラベル\n",
    "model2 = Sequential()\n",
    "model2.add(SimpleRNN(100, return_sequences=True, input_shape=(TIME_STEP, N_FEATURE)))\n",
    "model2.add(TimeDistributed(Dense(N_CLASS, activation='softmax')))\n",
    "model2.compile(\n",
    "    loss='categorical_crossentropy',\n",
    "    optimizer=Adam(lr=1e-3),\n",
    "    metrics=['accuracy']\n",
    ")\n",
    "model2.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      "1234/1234 [==============================] - 5s 4ms/step - loss: 0.7272 - acc: 0.5005\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f4a992134a8>"
      ]
     },
     "execution_count": 12,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model2.fit(\n",
    "    X_train2, y_train2,\n",
    "    batch_size=32,\n",
    "    epochs=1,\n",
    "    verbose=1\n",
    ")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "* 個人的には`SimpleRNN`よりも`LSTM`がおすすめ\n",
    "* `SimpleRNN`は長期の依存関係が学習できない\n",
    "* `return_sequences=False (default)`の場合、RNNの最終出力になるので最終時刻に影響される\n",
    "* kerasで可変長はできないので、RNNの時間幅とpad_sequencesのmaxlenをデータの最大長より大きく設定\n",
    "* 特徴方向で特徴化する場合は`TimeDistributed`\n",
    "* 時間と特徴から特徴する場合には`Flatten`"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "lstm_1 (LSTM)                (None, 100)               89200     \n",
      "_________________________________________________________________\n",
      "dense_3 (Dense)              (None, 2)                 202       \n",
      "=================================================================\n",
      "Total params: 89,402\n",
      "Trainable params: 89,402\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from keras.layers import LSTM\n",
    "\n",
    "# 時刻幅に対して1つのラベル(LSTMに置き換え)\n",
    "model3 = Sequential()\n",
    "model3.add(LSTM(100, input_shape=(TIME_STEP, N_FEATURE)))\n",
    "model3.add(Dense(N_CLASS, activation='softmax'))\n",
    "model3.compile(\n",
    "    loss='categorical_crossentropy',\n",
    "    optimizer=Adam(lr=1e-3),\n",
    "    metrics=['accuracy']\n",
    ")\n",
    "model3.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      "1234/1234 [==============================] - 12s 9ms/step - loss: 0.7087 - acc: 0.4878\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f4a9008c080>"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model3.fit(\n",
    "    X_train, y_train,\n",
    "    batch_size=32,\n",
    "    epochs=1,\n",
    "    verbose=1\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "lstm_2 (LSTM)                (None, 100, 122)          119560    \n",
      "_________________________________________________________________\n",
      "time_distributed_2 (TimeDist (None, 100, 10)           1230      \n",
      "_________________________________________________________________\n",
      "flatten_1 (Flatten)          (None, 1000)              0         \n",
      "_________________________________________________________________\n",
      "dense_5 (Dense)              (None, 10)                10010     \n",
      "_________________________________________________________________\n",
      "dense_6 (Dense)              (None, 2)                 22        \n",
      "=================================================================\n",
      "Total params: 130,822\n",
      "Trainable params: 130,822\n",
      "Non-trainable params: 0\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from keras.layers import Flatten\n",
    "\n",
    "model4 = Sequential()\n",
    "# 各時刻における時間と特徴を考慮した122の特徴量を出力(特徴に時間依存を反映)\n",
    "model4.add(LSTM(N_FEATURE, return_sequences=True, input_shape=(TIME_STEP, N_FEATURE)))\n",
    "model4.add(TimeDistributed(Dense(10)))  # 特徴方向のみ圧縮\n",
    "model4.add(Flatten())\n",
    "model4.add(Dense(10))  # 時間と特徴で圧縮\n",
    "model4.add(Dense(N_CLASS, activation='softmax'))\n",
    "model4.compile(\n",
    "    loss='categorical_crossentropy',\n",
    "    optimizer=Adam(lr=1e-3),\n",
    "    metrics=['accuracy']\n",
    ")\n",
    "model4.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/1\n",
      "1234/1234 [==============================] - 16s 13ms/step - loss: 0.7298 - acc: 0.4846\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x7f4a99213438>"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "model4.fit(\n",
    "    X_train, y_train,\n",
    "    batch_size=32,\n",
    "    epochs=1,\n",
    "    verbose=1\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "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": "3.6.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}