iui-group-l-name-zensiert/1-first-project/Abgabe.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5046da60",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "\n",
    "glob_path = '/opt/iui-datarelease2-sose2021/*/split_letters_csv/*'\n",
    "\n",
    "pickle_file = 'data.pickle'\n",
    "\n",
    "create_new = True\n",
    "checkpoint_path = \"training_1/cp.ckpt\"\n",
    "checkpoint_dir = os.path.dirname(checkpoint_path)\n",
    "\n",
    "# divisor for neuron count step downs (hard to describe), e.g. dense_step = 3: layer1=900, layer2 = 300, layer3 = 100, layer4 = 33...\n",
    "dense_steps = 2\n",
    "# amount of dense/dropout layers\n",
    "layer_count = 3\n",
    "# how much to drop\n",
    "drop_count = 0.1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "52925985",
   "metadata": {},
   "outputs": [],
   "source": [
    "from glob import glob\n",
    "import pandas as pd\n",
    "from tqdm import tqdm\n",
    "\n",
    "def dl_from_blob(filename) -> list:\n",
    "    all_data = []\n",
    "    \n",
    "    for path in tqdm(glob(filename)):\n",
    "        path = path\n",
    "        df = pd.read_csv(path, ';')\n",
    "        u = path.split('/')[3]\n",
    "        l = ''.join(filter(lambda x: x.isalpha(), path.split('/')[5]))[0] \n",
    "        d = {\n",
    "            'file': path,\n",
    "            'data': df,\n",
    "            'user': u,\n",
    "            'label': l\n",
    "        }\n",
    "        all_data.append(d)\n",
    "    return all_data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3dc191e7",
   "metadata": {},
   "outputs": [],
   "source": [
    "def save_pickle(f, structure):\n",
    "    _p = open(f, 'wb')\n",
    "    pickle.dump(structure, _p)\n",
    "    _p.close()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "858d807e",
   "metadata": {},
   "outputs": [],
   "source": [
    "import pickle\n",
    "\n",
    "def load_pickles(f) -> list:\n",
    "    _p = open(pickle_file, 'rb')\n",
    "    _d = pickle.load(_p)\n",
    "    _p.close()\n",
    "    \n",
    "    return _d"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "0885776a",
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "def load_data() -> list:\n",
    "    if os.path.isfile(pickle_file):\n",
    "        print(f'{pickle_file} found...')\n",
    "        return load_pickles(pickle_file)\n",
    "    print(f'Didn\\'t find {pickle_file}...')\n",
    "    all_data = dl_from_blob(glob_path)\n",
    "    print(f'Creating {pickle_file}...')\n",
    "    save_pickle(pickle_file, all_data)\n",
    "    return all_data\n",
    "\n",
    "print(\"Loading data...\")\n",
    "data = load_data()\n",
    "# plot_pd(data[0]['data'], False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4ec7f36a",
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "\n",
    "def plot_pd(data, force=True):\n",
    "    fig, axs = plt.subplots(5, 3, figsize=(3*3, 3*5))\n",
    "    axs[0][0].plot(data['Acc1 X'])\n",
    "    axs[0][1].plot(data['Acc1 Y'])\n",
    "    axs[0][2].plot(data['Acc1 Z'])\n",
    "    axs[1][0].plot(data['Acc2 X'])\n",
    "    axs[1][1].plot(data['Acc2 Y'])\n",
    "    axs[1][2].plot(data['Acc2 Z'])\n",
    "    axs[2][0].plot(data['Gyro X'])\n",
    "    axs[2][1].plot(data['Gyro Y'])\n",
    "    axs[2][2].plot(data['Gyro Z'])\n",
    "    axs[3][0].plot(data['Mag X'])\n",
    "    axs[3][1].plot(data['Mag Y'])\n",
    "    axs[3][2].plot(data['Mag Z'])\n",
    "    axs[4][0].plot(data['Time'])\n",
    "\n",
    "    if force:\n",
    "        for a in axs:\n",
    "            for b in a:\n",
    "                b.plot(data['Force'])\n",
    "    else:\n",
    "        axs[4][1].plot(data['Force'])\n",
    "\n",
    "def plot_np(data, force=True):\n",
    "    fig, axs = plt.subplots(5, 3, figsize=(3*3, 3*5))\n",
    "    axs[0][0].plot(data[0])\n",
    "    axs[0][1].plot(data[1])\n",
    "    axs[0][2].plot(data[2])\n",
    "    axs[1][0].plot(data[3])\n",
    "    axs[1][1].plot(data[4])\n",
    "    axs[1][2].plot(data[5])\n",
    "    axs[2][0].plot(data[6])\n",
    "    axs[2][1].plot(data[7])\n",
    "    axs[2][2].plot(data[8])\n",
    "    axs[3][0].plot(data[9])\n",
    "    axs[3][1].plot(data[10])\n",
    "    axs[3][2].plot(data[11])\n",
    "    axs[4][0].plot(data[13])\n",
    "\n",
    "    if force:\n",
    "        for a in axs:\n",
    "            for b in a:\n",
    "                b.plot(data[12])\n",
    "    else:\n",
    "        axs[4][1].plot(data[12])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "631fedbc",
   "metadata": {},
   "outputs": [],
   "source": [
    "def mill_drop(entry):\n",
    "    #drop millis on single\n",
    "    data_wo_mill = entry['data'].drop(labels='Millis', axis=1, inplace=False)\n",
    "    drop_entry = entry\n",
    "    drop_entry['data'] = data_wo_mill.reset_index(drop=True)\n",
    "    \n",
    "    return drop_entry"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "c86572fb",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "\n",
    "def cut_force(drop_entry):\n",
    "    # force trans\n",
    "    shorten_entry = drop_entry\n",
    "    shorten_data = shorten_entry['data']\n",
    "    sf_entry = shorten_data['Force']\n",
    "    leeway = 10\n",
    "    \n",
    "    try:\n",
    "        thresh = 70\n",
    "        temps_over_T = np.where(sf_entry > thresh)[0]\n",
    "        shorten_data = shorten_data[max(temps_over_T.min()-leeway,0):min(len(sf_entry)-1,temps_over_T.max()+leeway)]\n",
    "    except:\n",
    "        thresold = 0.05\n",
    "        thresh = sf_entry.max()*thresold\n",
    "        temps_over_T = np.where(sf_entry > thresh)[0]\n",
    "        shorten_data = shorten_data[max(temps_over_T.min()-leeway,0):min(len(sf_entry)-1,temps_over_T.max()+leeway)]\n",
    "    \n",
    "    shorten_entry['data'] = shorten_data.reset_index(drop=True)\n",
    "    return shorten_entry"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "875a8179",
   "metadata": {},
   "outputs": [],
   "source": [
    "def norm_force(shorten_entry, flist):\n",
    "    fnorm_entry = shorten_entry\n",
    "    u = fnorm_entry['user']\n",
    "    d = fnorm_entry['data']\n",
    "    \n",
    "    \n",
    "    d['Force'] = ((d['Force'] - flist[u].mean())/flist[u].std())\n",
    "    \n",
    "    fnorm_entry['data'] = fnorm_entry['data'].reset_index(drop=True)\n",
    "    return fnorm_entry"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "084feef0",
   "metadata": {},
   "outputs": [],
   "source": [
    "def time_trans(fnorm_entry):\n",
    "    #timetrans\n",
    "    time_entry = fnorm_entry\n",
    "    \n",
    "    time_entry['data']['Time'] = fnorm_entry['data']['Time']-fnorm_entry['data']['Time'][0]\n",
    "    \n",
    "    time_entry['data'] = time_entry['data'].reset_index(drop=True)\n",
    "\n",
    "    return time_entry"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ab674796",
   "metadata": {},
   "outputs": [],
   "source": [
    "def norm(time_entry):\n",
    "    # normalize\n",
    "    norm_entry = time_entry\n",
    "    \n",
    "    norm_entry['data']['Acc1 X'] = norm_entry['data']['Acc1 X'] / 32768\n",
    "    norm_entry['data']['Acc1 Y'] = norm_entry['data']['Acc1 Y'] / 32768\n",
    "    norm_entry['data']['Acc1 Z'] = norm_entry['data']['Acc1 Z'] / 32768\n",
    "    norm_entry['data']['Acc2 X'] = norm_entry['data']['Acc2 X'] / 8192\n",
    "    norm_entry['data']['Acc2 Y'] = norm_entry['data']['Acc2 Y'] / 8192\n",
    "    norm_entry['data']['Acc2 Z'] = norm_entry['data']['Acc2 Z'] / 8192\n",
    "    norm_entry['data']['Gyro X'] = norm_entry['data']['Gyro X'] / 32768\n",
    "    norm_entry['data']['Gyro Y'] = norm_entry['data']['Gyro Y'] / 32768\n",
    "    norm_entry['data']['Gyro Z'] = norm_entry['data']['Gyro Z'] / 32768\n",
    "    norm_entry['data']['Mag X']  = norm_entry['data']['Mag X']  / 8192\n",
    "    norm_entry['data']['Mag Y']  = norm_entry['data']['Mag Y']  / 8192\n",
    "    norm_entry['data']['Mag Z']  = norm_entry['data']['Mag Z']  / 8192\n",
    "    \n",
    "    norm_entry['data'] = norm_entry['data'].reset_index(drop=True)\n",
    "    \n",
    "    return norm_entry"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "cb1866fc",
   "metadata": {},
   "outputs": [],
   "source": [
    "def preproc(d):\n",
    "    flist = {}  \n",
    "    d_res = []\n",
    "    for e in data:\n",
    "        if e['user'] not in flist:\n",
    "            flist[e['user']] = e['data']['Force']\n",
    "        else:\n",
    "            flist[e['user']] = flist[e['user']].append(e['data']['Force'])\n",
    "    \n",
    "    for e in tqdm(data):\n",
    "        d_res.append(preproc_entry(e, flist))\n",
    "    return d_res\n",
    "    \n",
    "def preproc_entry(entry, flist):\n",
    "    drop_entry = mill_drop(entry)\n",
    "#     plot_pd(drop_entry['data'])\n",
    "#     \n",
    "    shorten_entry = cut_force(drop_entry)\n",
    "#     plot_pd(shorten_entry['data'])\n",
    "#     \n",
    "    fnorm_entry = norm_force(shorten_entry, flist)\n",
    "#     plot_pd(fnorm_entry['data'])\n",
    "#     \n",
    "    time_entry = time_trans(shorten_entry)\n",
    "#     plot_pd(time_entry['data'])\n",
    "#     \n",
    "    norm_entry = norm(time_entry)\n",
    "#     plot_pd(norm_entry['data'], False)\n",
    "    return norm_entry\n",
    "\n",
    "print(\"Preprocessing...\")\n",
    "pdata = preproc(data)\n",
    "# plot_pd(pdata[0]['data'], False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "12ec5a1a",
   "metadata": {},
   "outputs": [],
   "source": [
    "def throw(pdata):\n",
    "    llist = pd.Series([len(x['data']) for x in pdata])\n",
    "    threshold = int(llist.quantile(threshold_p))\n",
    "    longdex = np.where(llist <= threshold)[0]\n",
    "    return np.array(pdata)[longdex]\n",
    "\n",
    "llist = pd.Series([len(x['data']) for x in pdata])\n",
    "threshold_p = 0.75\n",
    "threshold = int(llist.quantile(threshold_p))\n",
    "\n",
    "print(\"Truncating...\")\n",
    "tpdata = throw(pdata)\n",
    "# plot_pd(tpdata[0]['data'], False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "44aaa509",
   "metadata": {},
   "outputs": [],
   "source": [
    "from tensorflow.keras.preprocessing.sequence import pad_sequences\n",
    "# ltpdata = []\n",
    "def elong(tpdata):\n",
    "    for x in tqdm(tpdata):\n",
    "        y = x['data'].to_numpy().T\n",
    "        x['data'] = pad_sequences(y, dtype=float, padding='post', maxlen=threshold)\n",
    "    return tpdata\n",
    "\n",
    "print(\"Padding...\")\n",
    "ltpdata = elong(tpdata)\n",
    "# plot_np(ltpdata[0]['data'], False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "594a3d76",
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "from tensorflow.keras.regularizers import l2\n",
    "from tensorflow.keras.models import Sequential\n",
    "from tensorflow.keras.layers import Dense, Flatten, BatchNormalization, Dropout\n",
    "from tensorflow.keras.callbacks import ModelCheckpoint, ReduceLROnPlateau\n",
    "from tensorflow.keras.optimizers import Adam\n",
    "\n",
    "def build_model(shape, classes):\n",
    "    model = Sequential()\n",
    "    \n",
    "    ncount = shape[0]*shape[1]\n",
    "    \n",
    "    model.add(Flatten(input_shape=shape, name='flatten'))\n",
    "    \n",
    "    model.add(Dropout(drop_count, name=f'dropout_{drop_count*100}'))\n",
    "    model.add(BatchNormalization(name='batchNorm'))\n",
    "    \n",
    "    for i in range(1,layer_count+1):\n",
    "        neurons = int(ncount/pow(dense_steps,i))\n",
    "        if neurons <= classes:\n",
    "            break\n",
    "        model.add(Dropout(drop_count*i, name=f'HiddenDropout_{drop_count*i*100:.0f}'))\n",
    "        model.add(Dense(neurons, activation='relu', \n",
    "                        kernel_regularizer=l2(0.001), name=f'Hidden_{i}')\n",
    "                 )\n",
    "    \n",
    "    model.add(Dense(classes, activation='softmax', name='Output'))\n",
    "    \n",
    "    model.compile(\n",
    "        optimizer=Adam(),\n",
    "        loss=\"categorical_crossentropy\", \n",
    "        metrics=[\"acc\"],\n",
    "    )\n",
    "    \n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e63d8c6a",
   "metadata": {},
   "outputs": [],
   "source": [
    "checkpoint_file = './goat.weights'\n",
    "\n",
    "def train(X_train, y_train, X_test, y_test):\n",
    "    model = build_model(X_train[0].shape, 52)\n",
    "    \n",
    "    model.summary()\n",
    "    \n",
    "    # Create a callback that saves the model's weights\n",
    "    model_checkpoint = ModelCheckpoint(filepath=checkpoint_path, monitor='loss', \n",
    "\t\t\tsave_best_only=True)\n",
    "    \n",
    "    history = model.fit(X_train, y_train, \n",
    "                  epochs=100,\n",
    "                  batch_size=32,\n",
    "                  shuffle=True,\n",
    "                  validation_data=(X_test, y_test),\n",
    "                verbose=2,\n",
    "            callbacks=[model_checkpoint]\n",
    "             )\n",
    "    \n",
    "    \n",
    "    model.load_weights(checkpoint_path)\n",
    "    print(\"Evaluate on test data\")\n",
    "    return model, history"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8ba620e1",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'  # this is required\n",
    "os.environ['CUDA_VISIBLE_DEVICES'] = '0'         # set to '0' for GPU0, '1' for GPU1 or '2' for GPU2. Check \"gpustat\" in a terminal."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "4449465c",
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn.preprocessing import LabelEncoder, LabelBinarizer\n",
    "\n",
    "X = np.array([x['data'] for x in ltpdata])\n",
    "y = np.array([x['label'] for x in ltpdata])\n",
    "\n",
    "lb = LabelBinarizer()\n",
    "y_tran = lb.fit_transform(y)\n",
    "\n",
    "X_train, X_test, y_train, y_test = train_test_split(X, y_tran, test_size=0.2, random_state=177013)\n",
    "\n",
    "X_train=X_train.reshape(X_train.shape[0],X_train.shape[1],X_train.shape[2])\n",
    "X_test=X_test.reshape(X_test.shape[0],X_test.shape[1],X_test.shape[2])\n",
    "\n",
    "train_shape = X_train[0].shape\n",
    "classes = y_train[0].shape[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f36d7904",
   "metadata": {
    "tags": []
   },
   "outputs": [],
   "source": [
    "%%time\n",
    "if not os.path.isdir(checkpoint_dir) or create_new:\n",
    "    tf.keras.backend.clear_session()\n",
    "    model, history = train(np.array(X), np.array(y_tran), np.array(X_test), np.array(y_test))\n",
    "else:\n",
    "    print(\"Loaded weights...\")\n",
    "    model = build_model(X_train[0].shape, 52)\n",
    "    model.load_weights(checkpoint_path)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "45d3110f",
   "metadata": {},
   "source": [
    "# Evaluation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "e05ce2ac",
   "metadata": {},
   "outputs": [],
   "source": [
    "ptest = [lb.classes_[e] for e in np.argmax(model.predict(X_test), axis=-1)]\n",
    "ltest = lb.inverse_transform(y_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8b2f0353",
   "metadata": {},
   "outputs": [],
   "source": [
    "%%time\n",
    "\n",
    "from sklearn.metrics import confusion_matrix\n",
    "import seaborn as sn\n",
    "\n",
    "from sklearn.metrics import classification_report\n",
    "\n",
    "set_digits = sorted(list(set(ltest)))\n",
    "\n",
    "test_cm = confusion_matrix(ltest, ptest, labels=list(set_digits), normalize='true')\n",
    "\n",
    "df_cm = pd.DataFrame(test_cm, index=set_digits, columns=set_digits)\n",
    "plt.figure(figsize = (20,14))\n",
    "sn_plot = sn.heatmap(df_cm, cmap=\"Greys\")\n",
    "plt.ylabel(\"True Label\")\n",
    "plt.xlabel(\"Predicted Label\")\n",
    "plt.show()\n",
    "\n",
    "print(classification_report(ltest, ptest, zero_division=0))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "5121b883",
   "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": "3.8.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}