Donkeycar_Keras-Speedup.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os\n",
    "import sys\n",
    "import time\n",
    "import json\n",
    "import datetime\n",
    "import random\n",
    "import tarfile\n",
    "\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "from PIL import Image\n",
    "PATH = \"../tubs/tub5\" #Shadow"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1.8.0\n"
     ]
    }
   ],
   "source": [
    "#make sure tensorflow version is the same as one installed on donkeycar (1.8.0)\n",
    "import tensorflow as tf\n",
    "print(tf.__version__)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "['/device:GPU:0']"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "from tensorflow.python.client import device_lib\n",
    "\n",
    "def get_available_gpus():\n",
    "    local_device_protos = device_lib.list_local_devices()\n",
    "    return [x.name for x in local_device_protos if x.device_type == 'GPU']\n",
    "get_available_gpus()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "#create df from path\n",
    "def df_from_jsons(PATH, max_items=None):\n",
    "    \n",
    "    #Get json files from path\n",
    "    json_files = [file for file in os.listdir(PATH) if file.endswith('json') ]\n",
    "    \n",
    "    df = pd.DataFrame()\n",
    "    for i, j in enumerate(json_files[1:max_items]):\n",
    "        with open(f'{PATH}/{j}') as f:\n",
    "            data = json.load(f)\n",
    "\n",
    "            userThrottle = data[\"user/throttle\"]\n",
    "            userAngle = data[\"user/angle\"]\n",
    "            camImage_path = data[\"cam/image_array\"]\n",
    "            index = int(camImage_path[:-21])\n",
    "            camImage_path =  PATH + \"/\" + camImage_path\n",
    "            df_row = pd.DataFrame([[userThrottle, userAngle, camImage_path]], index = [index], \n",
    "                                  columns = [\"user/throttle\", \"user/angle\", \"cam/image_array\"])\n",
    "\n",
    "            df = df.append(df_row)\n",
    "    df = df.sort_index()\n",
    "    print(\"Loaded {} json files\".format(len(json_files)))\n",
    "    return df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "#df = df_from_jsons(PATH, max_items=None)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "#df.to_feather('tmp/df_wPath')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "C:\\Users\\Jason\\Anaconda3\\envs\\tf-gpu\\lib\\site-packages\\pandas\\io\\feather_format.py:112: FutureWarning: `nthreads` argument is deprecated, pass `use_threads` instead\n",
      "  return feather.read_dataframe(path, nthreads=nthreads)\n"
     ]
    }
   ],
   "source": [
    "df = pd.read_feather('tmp/df_wPATH')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(38574, 3)"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "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>user/throttle</th>\n",
       "      <th>user/angle</th>\n",
       "      <th>cam/image_array</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>0</th>\n",
       "      <td>0.08</td>\n",
       "      <td>0.01</td>\n",
       "      <td>../tubs/tub5/0_cam-image_array_.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1</th>\n",
       "      <td>0.12</td>\n",
       "      <td>-0.03</td>\n",
       "      <td>../tubs/tub5/1_cam-image_array_.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2</th>\n",
       "      <td>0.12</td>\n",
       "      <td>-0.03</td>\n",
       "      <td>../tubs/tub5/2_cam-image_array_.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>3</th>\n",
       "      <td>0.14</td>\n",
       "      <td>-0.06</td>\n",
       "      <td>../tubs/tub5/3_cam-image_array_.jpg</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>4</th>\n",
       "      <td>0.14</td>\n",
       "      <td>-0.06</td>\n",
       "      <td>../tubs/tub5/4_cam-image_array_.jpg</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "   user/throttle  user/angle                      cam/image_array\n",
       "0           0.08        0.01  ../tubs/tub5/0_cam-image_array_.jpg\n",
       "1           0.12       -0.03  ../tubs/tub5/1_cam-image_array_.jpg\n",
       "2           0.12       -0.03  ../tubs/tub5/2_cam-image_array_.jpg\n",
       "3           0.14       -0.06  ../tubs/tub5/3_cam-image_array_.jpg\n",
       "4           0.14       -0.06  ../tubs/tub5/4_cam-image_array_.jpg"
      ]
     },
     "execution_count": 9,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "df.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "X_keys = ['cam/image_array']\n",
    "y_keys = ['user/angle', 'user/throttle']\n",
    "new_model_path = 'models/pilot_jn'\n",
    "bs = 256\n",
    "train_frac = 0.8"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [],
   "source": [
    "def to_bin(a):\n",
    "    arr = np.zeros(15)\n",
    "    a = a + 1\n",
    "    b = round(a/(2/14))\n",
    "    arr[int(b)] = 1\n",
    "    return arr"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "def read_record(record_dict, path = PATH):\n",
    "    data = {}\n",
    "    for key, val in record_dict.items():\n",
    "        if key == 'cam/image_array':\n",
    "            img = Image.open(f'{val}')\n",
    "            val = np.array(img)\n",
    "        \n",
    "        data[key] = val\n",
    "    return data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<PIL.Image.Image image mode=RGB size=160x120 at 0x1EA03E4D1D0>"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "#Testing\n",
    "df_t = df.sample(n=1).to_dict(orient='record')[0]\n",
    "record = read_record(df_t)\n",
    "data = record['cam/image_array']\n",
    "pic = Image.fromarray(data, 'RGB')\n",
    "pic"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_record_gen(record_transform=None, shuffle=True, df=None):\n",
    "    while True:\n",
    "        for _ in df.iterrows():\n",
    "            if shuffle:\n",
    "                record_dict = df.sample(n=1).to_dict(orient='record')[0]\n",
    "            \n",
    "            record_dict = read_record(record_dict)\n",
    "            \n",
    "            if record_transform:\n",
    "                record_dict = record_transform(record_dict)\n",
    "            \n",
    "            yield record_dict"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# record_gen = get_record_gen(record_transform=None, shuffle=True, df=df)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "# next(record_gen)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_batch_gen(keys=None, batch_size=128, record_transform=None, shuffle=True, df=None):\n",
    "    \n",
    "    record_gen = get_record_gen(record_transform=record_transform, shuffle=shuffle, df=df)\n",
    "    \n",
    "    while True:\n",
    "        record_list = [next(record_gen) for _ in range(batch_size)]\n",
    "        \n",
    "        batch_arrays = {}\n",
    "        for i, k in enumerate(keys):\n",
    "            arr = np.array([r[k] for r in record_list])\n",
    "            batch_arrays[k] = arr\n",
    "        yield batch_arrays"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_train_gen(X_keys, Y_keys, batch_size=128, record_transform=None, df=None): #\n",
    "    batch_gen = get_batch_gen(X_keys + Y_keys, batch_size=batch_size,\n",
    "                                  record_transform=record_transform, df=df)\n",
    "    while True:\n",
    "        batch = next(batch_gen)\n",
    "        X = [batch[k] for k in X_keys]\n",
    "        Y = [batch[k] for k in Y_keys] \n",
    "        yield X, Y"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "def get_train_val_gen(df, X_keys, Y_keys, batch_size = 128, train_frac=0.8, #\n",
    "                      train_record_transform=None, val_record_transform=None): #\n",
    "    train_df = df.sample(frac=train_frac, random_state=200) #\n",
    "    val_df = df.drop(train_df.index) #\n",
    "    \n",
    "    train_gen = get_train_gen(X_keys=X_keys, Y_keys=Y_keys, batch_size=batch_size,\n",
    "                             record_transform=train_record_transform, df=train_df)\n",
    "    val_gen = get_train_gen(X_keys=X_keys, Y_keys=Y_keys, batch_size=batch_size,\n",
    "                           record_transform=val_record_transform, df=val_df)\n",
    "    return train_gen, val_gen\n",
    "    "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "def record_transform(record):\n",
    "    record['user/angle'] = to_bin(record['user/angle'])\n",
    "    return record"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "train_gen, val_gen = get_train_val_gen(df, X_keys, y_keys,\n",
    "                                               train_record_transform=record_transform, #\n",
    "                                               val_record_transform=record_transform, #\n",
    "                                               batch_size=bs, #\n",
    "                                               train_frac=train_frac) #"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "from tensorflow.python.keras.layers import Input\n",
    "from tensorflow.python.keras.models import Model, load_model\n",
    "from tensorflow.python.keras.layers import Convolution2D, MaxPooling2D, BatchNormalization\n",
    "from tensorflow.python.keras.layers import Dropout, Flatten, Dense, Cropping2D, Lambda\n",
    "from tensorflow.python.keras.callbacks import ModelCheckpoint, EarlyStopping"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "def default_categorical():\n",
    "    img_in = Input(shape=(120, 160, 3),\n",
    "                   name='img_in')  # First layer, input layer, Shape comes from camera.py resolution, RGB\n",
    "    x = img_in\n",
    "    x = Convolution2D(24, (5, 5), strides=(2, 2), activation='relu')(x)\n",
    "    x = BatchNormalization()(x)\n",
    "    x = Dropout(0.2)(x)\n",
    "    x = Convolution2D(32, (4, 4), strides=(2, 2), activation='relu')(x)\n",
    "    x = MaxPooling2D(pool_size=(2, 2))(x)\n",
    "    x = BatchNormalization()(x)\n",
    "    x = Dropout(0.3)(x)\n",
    "    x = Convolution2D(64, (3, 3), strides=(1, 1), activation='relu')(x)\n",
    "    x = MaxPooling2D(pool_size=(2, 2))(x)\n",
    "    x = BatchNormalization()(x)\n",
    "    x = Dropout(0.5)(x)\n",
    "    x = Convolution2D(64, (3, 3), strides=(1, 1), activation='relu')(x)\n",
    "    x = MaxPooling2D(pool_size=(2, 2))(x)\n",
    "    x = BatchNormalization()(x)\n",
    "    x = Dropout(0.5)(x)\n",
    "\n",
    "    x = Flatten(name='flattened')(x)\n",
    "    x = Dense(50, activation='relu')(x)\n",
    "    x = Dropout(.5)(x)\n",
    "    \n",
    "    angle_out = Dense(15, activation='softmax', name='angle_out')(x)\n",
    "    throttle_out = Dense(1, activation='relu', name='throttle_out')(x)\n",
    "\n",
    "    model = Model(inputs=[img_in], outputs=[angle_out, throttle_out])\n",
    "    model.compile(optimizer='adam',\n",
    "                  loss={'angle_out': 'categorical_crossentropy',\n",
    "                        'throttle_out': 'mean_absolute_error'},\n",
    "                  loss_weights={'angle_out': 0.9, 'throttle_out': .01})\n",
    "\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "img_in (InputLayer)             (None, 120, 160, 3)  0                                            \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_4 (Conv2D)               (None, 58, 78, 24)   1824        img_in[0][0]                     \n",
      "__________________________________________________________________________________________________\n",
      "batch_normalization_4 (BatchNor (None, 58, 78, 24)   96          conv2d_4[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "dropout_5 (Dropout)             (None, 58, 78, 24)   0           batch_normalization_4[0][0]      \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_5 (Conv2D)               (None, 28, 38, 32)   12320       dropout_5[0][0]                  \n",
      "__________________________________________________________________________________________________\n",
      "max_pooling2d_4 (MaxPooling2D)  (None, 14, 19, 32)   0           conv2d_5[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "batch_normalization_5 (BatchNor (None, 14, 19, 32)   128         max_pooling2d_4[0][0]            \n",
      "__________________________________________________________________________________________________\n",
      "dropout_6 (Dropout)             (None, 14, 19, 32)   0           batch_normalization_5[0][0]      \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_6 (Conv2D)               (None, 12, 17, 64)   18496       dropout_6[0][0]                  \n",
      "__________________________________________________________________________________________________\n",
      "max_pooling2d_5 (MaxPooling2D)  (None, 6, 8, 64)     0           conv2d_6[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "batch_normalization_6 (BatchNor (None, 6, 8, 64)     256         max_pooling2d_5[0][0]            \n",
      "__________________________________________________________________________________________________\n",
      "dropout_7 (Dropout)             (None, 6, 8, 64)     0           batch_normalization_6[0][0]      \n",
      "__________________________________________________________________________________________________\n",
      "conv2d_7 (Conv2D)               (None, 4, 6, 64)     36928       dropout_7[0][0]                  \n",
      "__________________________________________________________________________________________________\n",
      "max_pooling2d_6 (MaxPooling2D)  (None, 2, 3, 64)     0           conv2d_7[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "batch_normalization_7 (BatchNor (None, 2, 3, 64)     256         max_pooling2d_6[0][0]            \n",
      "__________________________________________________________________________________________________\n",
      "dropout_8 (Dropout)             (None, 2, 3, 64)     0           batch_normalization_7[0][0]      \n",
      "__________________________________________________________________________________________________\n",
      "flattened (Flatten)             (None, 384)          0           dropout_8[0][0]                  \n",
      "__________________________________________________________________________________________________\n",
      "dense_2 (Dense)                 (None, 50)           19250       flattened[0][0]                  \n",
      "__________________________________________________________________________________________________\n",
      "dropout_9 (Dropout)             (None, 50)           0           dense_2[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "angle_out (Dense)               (None, 15)           765         dropout_9[0][0]                  \n",
      "__________________________________________________________________________________________________\n",
      "throttle_out (Dense)            (None, 1)            51          dropout_9[0][0]                  \n",
      "==================================================================================================\n",
      "Total params: 90,370\n",
      "Trainable params: 90,002\n",
      "Non-trainable params: 368\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "model = default_categorical()\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "def train(train_gen, val_gen, saved_model_path, epochs=100, steps=100, train_split=0.8,\n",
    "         verbose=1, min_delta=0.0005, patience=5, use_early_stop=True):\n",
    "    \n",
    "#     checkpoint to save model\n",
    "    save_best = ModelCheckpoint(saved_model_path, monitor='val_loss', verbose=verbose, save_best_only=True, mode='min')\n",
    "    \n",
    "    #stop training if stop improving\n",
    "    early_stop = EarlyStopping(monitor='val_loss', min_delta=min_delta, patience=patience, verbose=verbose, mode='auto')\n",
    "    \n",
    "    callbacks_list = [save_best]\n",
    "    \n",
    "    if use_early_stop:\n",
    "        callbacks_list.append(early_stop)\n",
    "        \n",
    "    hist = model.fit_generator(train_gen, steps_per_epoch=steps, epochs=epochs, verbose=1, validation_data=val_gen,\n",
    "                                callbacks=callbacks_list, validation_steps=steps * (1.0 - train_frac)//train_frac)\n",
    "    return hist"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "120"
      ]
     },
     "execution_count": 26,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "total_records = len(df)\n",
    "total_train = int(total_records * train_frac)\n",
    "total_val = total_records - total_train\n",
    "steps_per_epoch = total_train // bs\n",
    "steps_per_epoch"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 2.2083 - angle_out_loss: 2.4483 - throttle_out_loss: 0.4795\n",
      "Epoch 00001: val_loss improved from inf to 1.81939, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 91s 755ms/step - loss: 2.2054 - angle_out_loss: 2.4451 - throttle_out_loss: 0.4787 - val_loss: 1.8194 - val_angle_out_loss: 2.0182 - val_throttle_out_loss: 0.3039\n",
      "Epoch 2/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.7334 - angle_out_loss: 1.9223 - throttle_out_loss: 0.3361\n",
      "Epoch 00002: val_loss improved from 1.81939 to 1.42053, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 86s 716ms/step - loss: 1.7317 - angle_out_loss: 1.9204 - throttle_out_loss: 0.3361 - val_loss: 1.4205 - val_angle_out_loss: 1.5748 - val_throttle_out_loss: 0.3198\n",
      "Epoch 3/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.5850 - angle_out_loss: 1.7576 - throttle_out_loss: 0.3222\n",
      "Epoch 00003: val_loss improved from 1.42053 to 1.37798, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 87s 727ms/step - loss: 1.5849 - angle_out_loss: 1.7574 - throttle_out_loss: 0.3222 - val_loss: 1.3780 - val_angle_out_loss: 1.5275 - val_throttle_out_loss: 0.3197\n",
      "Epoch 4/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.5024 - angle_out_loss: 1.6657 - throttle_out_loss: 0.3217\n",
      "Epoch 00004: val_loss improved from 1.37798 to 1.37575, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 88s 732ms/step - loss: 1.5021 - angle_out_loss: 1.6655 - throttle_out_loss: 0.3216 - val_loss: 1.3757 - val_angle_out_loss: 1.5251 - val_throttle_out_loss: 0.3193\n",
      "Epoch 5/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.4502 - angle_out_loss: 1.6077 - throttle_out_loss: 0.3227\n",
      "Epoch 00005: val_loss improved from 1.37575 to 1.31645, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 86s 721ms/step - loss: 1.4505 - angle_out_loss: 1.6081 - throttle_out_loss: 0.3227 - val_loss: 1.3164 - val_angle_out_loss: 1.4592 - val_throttle_out_loss: 0.3193\n",
      "Epoch 6/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.4360 - angle_out_loss: 1.5919 - throttle_out_loss: 0.3213\n",
      "Epoch 00006: val_loss did not improve\n",
      "120/120 [==============================] - 86s 720ms/step - loss: 1.4358 - angle_out_loss: 1.5918 - throttle_out_loss: 0.3213 - val_loss: 1.3227 - val_angle_out_loss: 1.4661 - val_throttle_out_loss: 0.3191\n",
      "Epoch 7/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.4129 - angle_out_loss: 1.5663 - throttle_out_loss: 0.3208\n",
      "Epoch 00007: val_loss improved from 1.31645 to 1.29638, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 86s 721ms/step - loss: 1.4126 - angle_out_loss: 1.5660 - throttle_out_loss: 0.3208 - val_loss: 1.2964 - val_angle_out_loss: 1.4369 - val_throttle_out_loss: 0.3190\n",
      "Epoch 8/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3989 - angle_out_loss: 1.5508 - throttle_out_loss: 0.3202\n",
      "Epoch 00008: val_loss improved from 1.29638 to 1.28934, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 85s 711ms/step - loss: 1.3986 - angle_out_loss: 1.5504 - throttle_out_loss: 0.3202 - val_loss: 1.2893 - val_angle_out_loss: 1.4290 - val_throttle_out_loss: 0.3201\n",
      "Epoch 9/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3744 - angle_out_loss: 1.5236 - throttle_out_loss: 0.3202\n",
      "Epoch 00009: val_loss did not improve\n",
      "120/120 [==============================] - 86s 720ms/step - loss: 1.3733 - angle_out_loss: 1.5224 - throttle_out_loss: 0.3203 - val_loss: 1.2996 - val_angle_out_loss: 1.4404 - val_throttle_out_loss: 0.3181\n",
      "Epoch 10/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3606 - angle_out_loss: 1.5082 - throttle_out_loss: 0.3202\n",
      "Epoch 00010: val_loss improved from 1.28934 to 1.27320, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 86s 718ms/step - loss: 1.3602 - angle_out_loss: 1.5077 - throttle_out_loss: 0.3202 - val_loss: 1.2732 - val_angle_out_loss: 1.4111 - val_throttle_out_loss: 0.3197\n",
      "Epoch 11/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3606 - angle_out_loss: 1.5082 - throttle_out_loss: 0.3201\n",
      "Epoch 00011: val_loss improved from 1.27320 to 1.25804, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 85s 712ms/step - loss: 1.3607 - angle_out_loss: 1.5083 - throttle_out_loss: 0.3201 - val_loss: 1.2580 - val_angle_out_loss: 1.3943 - val_throttle_out_loss: 0.3189\n",
      "Epoch 12/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3391 - angle_out_loss: 1.4844 - throttle_out_loss: 0.3198\n",
      "Epoch 00012: val_loss improved from 1.25804 to 1.24591, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 86s 713ms/step - loss: 1.3387 - angle_out_loss: 1.4839 - throttle_out_loss: 0.3198 - val_loss: 1.2459 - val_angle_out_loss: 1.3808 - val_throttle_out_loss: 0.3196\n",
      "Epoch 13/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3320 - angle_out_loss: 1.4764 - throttle_out_loss: 0.3196\n",
      "Epoch 00013: val_loss did not improve\n",
      "120/120 [==============================] - 83s 692ms/step - loss: 1.3314 - angle_out_loss: 1.4758 - throttle_out_loss: 0.3195 - val_loss: 1.2916 - val_angle_out_loss: 1.4315 - val_throttle_out_loss: 0.3196\n",
      "Epoch 14/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3284 - angle_out_loss: 1.4725 - throttle_out_loss: 0.3199\n",
      "Epoch 00014: val_loss improved from 1.24591 to 1.21110, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 84s 698ms/step - loss: 1.3275 - angle_out_loss: 1.4714 - throttle_out_loss: 0.3200 - val_loss: 1.2111 - val_angle_out_loss: 1.3421 - val_throttle_out_loss: 0.3200\n",
      "Epoch 15/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3216 - angle_out_loss: 1.4649 - throttle_out_loss: 0.3201\n",
      "Epoch 00015: val_loss did not improve\n",
      "120/120 [==============================] - 84s 697ms/step - loss: 1.3212 - angle_out_loss: 1.4644 - throttle_out_loss: 0.3201 - val_loss: 1.2174 - val_angle_out_loss: 1.3491 - val_throttle_out_loss: 0.3202\n",
      "Epoch 16/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3151 - angle_out_loss: 1.4577 - throttle_out_loss: 0.3196\n",
      "Epoch 00016: val_loss improved from 1.21110 to 1.20129, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 83s 689ms/step - loss: 1.3147 - angle_out_loss: 1.4572 - throttle_out_loss: 0.3196 - val_loss: 1.2013 - val_angle_out_loss: 1.3312 - val_throttle_out_loss: 0.3189\n",
      "Epoch 17/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3108 - angle_out_loss: 1.4529 - throttle_out_loss: 0.3191\n",
      "Epoch 00017: val_loss did not improve\n",
      "120/120 [==============================] - 83s 695ms/step - loss: 1.3103 - angle_out_loss: 1.4524 - throttle_out_loss: 0.3191 - val_loss: 1.2773 - val_angle_out_loss: 1.4157 - val_throttle_out_loss: 0.3192\n",
      "Epoch 18/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.3052 - angle_out_loss: 1.4467 - throttle_out_loss: 0.3194\n",
      "Epoch 00018: val_loss did not improve\n",
      "120/120 [==============================] - 83s 692ms/step - loss: 1.3048 - angle_out_loss: 1.4463 - throttle_out_loss: 0.3193 - val_loss: 1.2179 - val_angle_out_loss: 1.3497 - val_throttle_out_loss: 0.3190\n",
      "Epoch 19/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2917 - angle_out_loss: 1.4316 - throttle_out_loss: 0.3196\n",
      "Epoch 00019: val_loss did not improve\n",
      "120/120 [==============================] - 84s 697ms/step - loss: 1.2915 - angle_out_loss: 1.4314 - throttle_out_loss: 0.3196 - val_loss: 1.2891 - val_angle_out_loss: 1.4288 - val_throttle_out_loss: 0.3193\n",
      "Epoch 20/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2902 - angle_out_loss: 1.4300 - throttle_out_loss: 0.3196\n",
      "Epoch 00020: val_loss did not improve\n",
      "120/120 [==============================] - 85s 711ms/step - loss: 1.2908 - angle_out_loss: 1.4307 - throttle_out_loss: 0.3196 - val_loss: 1.2217 - val_angle_out_loss: 1.3539 - val_throttle_out_loss: 0.3182\n",
      "Epoch 21/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2893 - angle_out_loss: 1.4291 - throttle_out_loss: 0.3181\n",
      "Epoch 00021: val_loss improved from 1.20129 to 1.18588, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 84s 700ms/step - loss: 1.2893 - angle_out_loss: 1.4290 - throttle_out_loss: 0.3182 - val_loss: 1.1859 - val_angle_out_loss: 1.3141 - val_throttle_out_loss: 0.3188\n",
      "Epoch 22/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2752 - angle_out_loss: 1.4134 - throttle_out_loss: 0.3178\n",
      "Epoch 00022: val_loss did not improve\n",
      "120/120 [==============================] - 85s 711ms/step - loss: 1.2763 - angle_out_loss: 1.4146 - throttle_out_loss: 0.3178 - val_loss: 1.2013 - val_angle_out_loss: 1.3313 - val_throttle_out_loss: 0.3193\n",
      "Epoch 23/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2734 - angle_out_loss: 1.4114 - throttle_out_loss: 0.3152\n",
      "Epoch 00023: val_loss improved from 1.18588 to 1.17485, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 86s 716ms/step - loss: 1.2739 - angle_out_loss: 1.4120 - throttle_out_loss: 0.3152 - val_loss: 1.1748 - val_angle_out_loss: 1.3018 - val_throttle_out_loss: 0.3185\n",
      "Epoch 24/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2684 - angle_out_loss: 1.4059 - throttle_out_loss: 0.3124\n",
      "Epoch 00024: val_loss did not improve\n",
      "120/120 [==============================] - 86s 720ms/step - loss: 1.2675 - angle_out_loss: 1.4049 - throttle_out_loss: 0.3124 - val_loss: 1.1903 - val_angle_out_loss: 1.3190 - val_throttle_out_loss: 0.3202\n",
      "Epoch 25/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2623 - angle_out_loss: 1.3993 - throttle_out_loss: 0.2956\n",
      "Epoch 00025: val_loss improved from 1.17485 to 1.16595, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 86s 716ms/step - loss: 1.2618 - angle_out_loss: 1.3988 - throttle_out_loss: 0.2955 - val_loss: 1.1660 - val_angle_out_loss: 1.2920 - val_throttle_out_loss: 0.3193\n",
      "Epoch 26/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2663 - angle_out_loss: 1.4042 - throttle_out_loss: 0.2529\n",
      "Epoch 00026: val_loss did not improve\n",
      "120/120 [==============================] - 85s 706ms/step - loss: 1.2664 - angle_out_loss: 1.4043 - throttle_out_loss: 0.2526 - val_loss: 1.2868 - val_angle_out_loss: 1.4272 - val_throttle_out_loss: 0.2387\n",
      "Epoch 27/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2559 - angle_out_loss: 1.3930 - throttle_out_loss: 0.2217\n",
      "Epoch 00027: val_loss improved from 1.16595 to 1.13405, saving model to models/pilot_jn\n",
      "120/120 [==============================] - 84s 700ms/step - loss: 1.2557 - angle_out_loss: 1.3927 - throttle_out_loss: 0.2215 - val_loss: 1.1340 - val_angle_out_loss: 1.2574 - val_throttle_out_loss: 0.2354\n",
      "Epoch 28/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2529 - angle_out_loss: 1.3905 - throttle_out_loss: 0.1502\n",
      "Epoch 00028: val_loss did not improve\n",
      "120/120 [==============================] - 87s 722ms/step - loss: 1.2528 - angle_out_loss: 1.3903 - throttle_out_loss: 0.1501 - val_loss: 1.1861 - val_angle_out_loss: 1.3160 - val_throttle_out_loss: 0.1723\n",
      "Epoch 29/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2461 - angle_out_loss: 1.3834 - throttle_out_loss: 0.0975\n",
      "Epoch 00029: val_loss did not improve\n",
      "120/120 [==============================] - 85s 709ms/step - loss: 1.2460 - angle_out_loss: 1.3834 - throttle_out_loss: 0.0970 - val_loss: 1.1952 - val_angle_out_loss: 1.3275 - val_throttle_out_loss: 0.0410\n",
      "Epoch 30/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2442 - angle_out_loss: 1.3820 - throttle_out_loss: 0.0380\n",
      "Epoch 00030: val_loss did not improve\n",
      "120/120 [==============================] - 85s 709ms/step - loss: 1.2447 - angle_out_loss: 1.3826 - throttle_out_loss: 0.0381 - val_loss: 1.1807 - val_angle_out_loss: 1.3115 - val_throttle_out_loss: 0.0361\n",
      "Epoch 31/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2405 - angle_out_loss: 1.3779 - throttle_out_loss: 0.0364\n",
      "Epoch 00031: val_loss did not improve\n",
      "120/120 [==============================] - 86s 716ms/step - loss: 1.2397 - angle_out_loss: 1.3771 - throttle_out_loss: 0.0364 - val_loss: 1.1688 - val_angle_out_loss: 1.2982 - val_throttle_out_loss: 0.0357\n",
      "Epoch 32/100\n",
      "119/120 [============================>.] - ETA: 0s - loss: 1.2349 - angle_out_loss: 1.3717 - throttle_out_loss: 0.0366\n",
      "Epoch 00032: val_loss did not improve\n",
      "120/120 [==============================] - 83s 690ms/step - loss: 1.2345 - angle_out_loss: 1.3712 - throttle_out_loss: 0.0366 - val_loss: 1.1383 - val_angle_out_loss: 1.2644 - val_throttle_out_loss: 0.0352\n",
      "Epoch 00032: early stopping\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tensorflow.python.keras._impl.keras.callbacks.History at 0x1ea5bccf7b8>"
      ]
     },
     "execution_count": 27,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "train(train_gen, val_gen, saved_model_path=new_model_path, steps=steps_per_epoch, train_split=0.8) #change train_split to 0.9 for large data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'\\nRecords\\n\\nOriginal architecture with 79k samples\\n498/498 [==============================] - 215s 432ms/step - loss: 0.4393 - angle_out_loss: 0.4875 \\n- throttle_out_loss: 0.0551 - val_loss: 0.8894 - val_angle_out_loss: 0.9876 - val_throttle_out_loss: 0.0586\\nEpoch 00028: early stopping\\n\\nOriginal architecture with 79k samples and 0.2 and 0.3 dropout at the end\\n498/498 [==============================] - 239s 480ms/step - loss: 0.5324 - angle_out_loss: 0.5910 - \\nthrottle_out_loss: 0.0585 - val_loss: 0.9268 - val_angle_out_loss: 1.0292 - val_throttle_out_loss: 0.0530\\nEpoch 00030: early stopping\\n\\nOriginal architecture with 79k samples and 0.3 and 0.5 dropout at the end\\n\\n'"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "\"\"\"\n",
    "Records\n",
    "\n",
    "Original architecture with 79k samples\n",
    "498/498 [==============================] - 215s 432ms/step - loss: 0.4393 - angle_out_loss: 0.4875 \n",
    "- throttle_out_loss: 0.0551 - val_loss: 0.8894 - val_angle_out_loss: 0.9876 - val_throttle_out_loss: 0.0586\n",
    "Epoch 00028: early stopping\n",
    "\n",
    "Original architecture with 79k samples and 0.2 and 0.3 dropout at the end\n",
    "498/498 [==============================] - 239s 480ms/step - loss: 0.5324 - angle_out_loss: 0.5910 - \n",
    "throttle_out_loss: 0.0585 - val_loss: 0.9268 - val_angle_out_loss: 1.0292 - val_throttle_out_loss: 0.0530\n",
    "Epoch 00030: early stopping\n",
    "\n",
    "Original architecture with 79k samples and 0.3 and 0.5 dropout at the end\n",
    "\n",
    "\"\"\"\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 175,
   "metadata": {},
   "outputs": [],
   "source": [
    "#check against original\n",
    "def unbin(arr):\n",
    "    index = np.argmax(arr)\n",
    "    return index/14*2-1\n",
    "#     return index"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 174,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "0.0"
      ]
     },
     "execution_count": 174,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "unbin([0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 176,
   "metadata": {},
   "outputs": [],
   "source": [
    "def run(img_arr):\n",
    "    img_arr = img_arr.reshape((1,) + img_arr.shape)\n",
    "    angle_binned, throttle = model.predict(img_arr)\n",
    "    angle_unbinned = unbin(angle_binned[0])\n",
    "    return angle_unbinned, throttle[0][0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 194,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=160x120 at 0x2119C0EA780>"
      ]
     },
     "execution_count": 194,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "img_name = df['cam/image_array'][150]\n",
    "img = Image.open(f'{PATH}/{img_name}')\n",
    "img"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 195,
   "metadata": {},
   "outputs": [],
   "source": [
    "val = np.array(img)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 196,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(120, 160, 3)"
      ]
     },
     "execution_count": 196,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "val.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 198,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(-0.4285714285714286, 0.31011686)"
      ]
     },
     "execution_count": 198,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "#prediction\n",
    "run(val)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 199,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(-0.42, 0.27)"
      ]
     },
     "execution_count": 199,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "#compare to actual data\n",
    "df['user/angle'][150], df['user/throttle'][150]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "\"\"\"\n",
    "Record\n",
    ">>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n",
    "def default_categorical():\n",
    "    img_in = Input(shape=(120, 160, 3),\n",
    "                   name='img_in')  # First layer, input layer, Shape comes from camera.py resolution, RGB\n",
    "    x = img_in\n",
    "    x = Convolution2D(24, (5, 5), strides=(2, 2), activation='relu')(x)\n",
    "    x = MaxPooling2D(pool_size=(2, 2))(x)\n",
    "    x = Convolution2D(32, (4, 4), strides=(2, 2), activation='relu')(x)\n",
    "    x = MaxPooling2D(pool_size=(2, 2))(x)\n",
    "    x = Dropout(0.3)(x)\n",
    "    x = Convolution2D(64, (3, 3), strides=(1, 1), activation='relu')(x)\n",
    "    x = MaxPooling2D(pool_size=(2, 2))(x)\n",
    "    x = Dropout(0.5)(x)\n",
    "\n",
    "    x = Flatten(name='flattened')(x)\n",
    "    x = Dense(30, activation='relu')(x)\n",
    "    x = Dropout(0.5)(x)\n",
    "    \n",
    "    angle_out = Dense(15, activation='softmax', name='angle_out')(x)\n",
    "    throttle_out = Dense(1, activation='relu', name='throttle_out')(x)\n",
    "\n",
    "    model = Model(inputs=[img_in], outputs=[angle_out, throttle_out])\n",
    "    model.compile(optimizer='adam',\n",
    "                  loss={'angle_out': 'categorical_crossentropy',\n",
    "                        'throttle_out': 'mean_absolute_error'},\n",
    "                  loss_weights={'angle_out': 0.9, 'throttle_out': .01})\n",
    "\n",
    "    return model\n",
    "\"\"\"\n"
   ]
  }
 ],
 "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
}