Checkpoint

.gitignore

@@ -2,4 +2,5 @@
 /code/data/models/*
 /code/data/full/*
 /code/data/reduced/*
+/code/data/notebook/*
 /papers/*

code/TaskB.ipynb → code/HAR_system.ipynb

@@ -4,15 +4,26 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# Task B - single subject - model ?\n",
+    "# HAR system - Lincetto Riccardo, Drago Matteo\n",
+    "This notebook runs:\n",
+    "- Classification with null class (One Shot classification);\n",
+    "- Binary classification for activity detection (Two Steps - detection);\n",
+    "- Classification without null class (Two Steps - classification);\n",
+    "- Cascade of the last to methods.\n",
+    "\n",
+    "The operations performed here are very similar to those execute in 'main.py', with the exception that here the program is executed for specified user and model.\n",
     "\n",
     "## Notebook setup\n",
     "This first cell contains the parameters that can be tuned for code execution:\n",
     "- subject: select the subject on which to test the model, between [1,4];\n",
-    "- label: index of feature column to be selected to perform activity detection, between [0,6]. The default value for task B is 6;\n",
-    "- folder: directory name where '.mat' files are stored;\n",
+    "- task: choose \"A\" for locomotion classification or \"B\" for gesture recognition;\n",
+    "- model_name: choose between \"Convolutional\", \"Convolutional1DRecurrent\", \"Convolutional2DRecurrent\" and \"ConvolutionalDeepRecurrent\";\n",
+    "- data_folder: directory name where '.mat' files are stored;\n",
     "- window_size: parameter that sets the length of temporal windows on which to perform the convolution;\n",
-    "- stride: step length to chose the next window."
+    "- stride: step length to chose the next window;\n",
+    "- GPU: boolean flag indicatin wheter GPU is present on the machine that executes the code;\n",
+    "- epochs: number of complete sweeps of the data signals during training;\n",
+    "- batch_size: number of forward propagations in the networks between consecutives backpropagations."
    ]
   },
   {
@@ -22,10 +33,21 @@
    "outputs": [],
    "source": [
     "subject = 1\n",
-    "label = 0\n",
-    "folder = \"../data/full/\"\n",
+    "task = \"A\"\n",
+    "model_name = \"Convolutional\"\n",
+    "data_folder = \"./data/full/\"\n",
     "window_size = 15\n",
-    "stride = 5"
+    "stride = 5\n",
+    "GPU = True\n",
+    "epochs = 10\n",
+    "batch_size = 32"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Here the useful functions are imported."
    ]
   },
   {
@@ -47,21 +69,22 @@
     "import preprocessing\n",
     "import models\n",
     "import utils\n",
+    "import os\n",
     "import numpy as np\n",
-    "import seaborn as sns\n",
-    "import matplotlib.pyplot as plt\n",
     "from sklearn.metrics import classification_report, f1_score, confusion_matrix\n",
     "from keras.models import load_model\n",
     "from keras.optimizers import Adam\n",
-    "from keras.callbacks import ModelCheckpoint\n",
-    "from keras.utils import to_categorical"
+    "from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau\n",
+    "from keras.utils import to_categorical\n",
+    "import seaborn as sns\n",
+    "import matplotlib.pyplot as plt"
    ]
   },
   {
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "In the following cell, we make use of some functions of Keras which have been removed, but of which the code is still available at https://github.com/keras-team/keras/commit/a56b1a55182acf061b1eb2e2c86b48193a0e88f7. These are used to evaulate the f1 score during training on batches of data: this is only an approximation though, which is the reason why they have been removed."
+    "Differently from 'main.py', all results saved from this notebook are going to be stored in a dedicated folder: './data/notebook/'. For proper execution, this folder needs first to be created."
    ]
   },
   {
@@ -70,50 +93,104 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "import os\n",
     "if not(os.path.exists(\"./data\")):\n",
-    "    os.mkdir(\"./data\")"
+    "    os.mkdir(\"./data\")\n",
+    "if not(os.path.exists(\"./data/notebook\")):\n",
+    "    os.mkdir(\"./data/notebook\")"
    ]
   },
   {
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# One-shot classification\n",
-    "Here classification is performed with null class.\n",
-    "### Preprocessing"
+    "If task A is selected, calssifications in the following notebook are based on the labels of column 0; if instead it's task B, column 6 labels are used."
    ]
   },
   {
    "cell_type": "code",
    "execution_count": 4,
    "metadata": {},
-   "outputs": [],
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Task A uses labels column 0\n"
+     ]
+    }
+   ],
    "source": [
-    "X_train, Y_train, X_test, Y_test, n_features, n_classes, class_weights = preprocessing.loadData(subject=subject,\n",
-    "                                                                                                label=label,\n",
-    "                                                                                                folder=folder,\n",
-    "                                                                                                window_size=window_size,\n",
-    "                                                                                                stride=stride,\n",
-    "                                                                                                make_binary=False,\n",
-    "                                                                                                null_class=True,\n",
-    "                                                                                                print_info=False)"
+    "if task == \"A\":\n",
+    "    label = 0\n",
+    "elif task == \"B\":\n",
+    "    label = 6\n",
+    "else:\n",
+    "    print(\"Error: invalid task.\")\n",
+    "print(\"Task\", task, \"uses labels column\", label)"
    ]
   },
   {
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "Y_train and Y_test contain the correct labels for each signals window. Y_test in particular will be used to evaluate predictions for both this (one-shot) and the two-steps models. For this reason it is here saved with a different name, to avoid having it being overwritten later."
+    "## Classification with null class: One Shot classification\n",
+    "Here classification is performed considering inactivity as a class, alongside with the others. In the case of locomotion classification (task A), this becomes a 5-class problem, while in the case of gesture recognition (task B) the classes become 18. In the following cell are perfomed in order:\n",
+    "- preprocessing;\n",
+    "- model selection;\n",
+    "- model compilation;\n",
+    "- training.\n",
+    "\n",
+    "Note that in case \"Convolutional2DRecurrent\" is the model selected, then the preprocessed data need to be reshaped, adding one dimension; this is automatically done by the code."
    ]
   },
   {
    "cell_type": "code",
-   "execution_count": 5,
+   "execution_count": 6,
    "metadata": {},
    "outputs": [],
    "source": [
-    "Y_test_true = Y_test"
+    "# preprocessing\n",
+    "X_train, Y_train, X_test, Y_test, n_features, n_classes, class_weights = preprocessing.loadData(subject=subject,\n",
+    "                                                                                                label=label,\n",
+    "                                                                                                folder=data_folder,\n",
+    "                                                                                                window_size=window_size,\n",
+    "                                                                                                stride=stride,\n",
+    "                                                                                                make_binary=False,\n",
+    "                                                                                                null_class=True,\n",
+    "                                                                                                print_info=False)\n",
+    "\n",
+    "# model selection\n",
+    "if model_name == \"Convolutional\":\n",
+    "    model = models.Convolutional((window_size, n_features), n_classes, print_info=False)\n",
+    "elif model_name == \"Convolutional1DRecurrent\":\n",
+    "    model = models.Convolutional1DRecurrent((window_size, n_features), n_classes, GPU=GPU, print_info=False)\n",
+    "elif model_name == \"Convolutional2DRecurrent\":\n",
+    "    model = models.Convolutional2DRecurrent((window_size, n_features, 1), n_classes, GPU=GPU, print_info=False)\n",
+    "    # reshaping for 2D convolutional model\n",
+    "    X_train = X_train.reshape(X_train.shape[0], window_size, n_features, 1)\n",
+    "    X_test = X_test.reshape(X_test.shape[0], window_size, n_features, 1)\n",
+    "elif model_name == \"ConvolutionalDeepRecurrent\":\n",
+    "    model = models.ConvolutionalDeepRecurrent((window_size, n_features), n_classes, GPU=GPU, print_info=False)\n",
+    "else:\n",
+    "    print(\"Model not found.\")\n",
+    "    break\n",
+    "\n",
+    "# model compilation\n",
+    "model.compile(optimizer = Adam(lr=0.001), loss = \"categorical_crossentropy\", metrics = [\"accuracy\"])\n",
+    "save_model_name = task + \"_\" + model_name + \"_OS_\" + str(s)\n",
+    "filepath = './data/notebook/'+save_model_name+'.hdf5'\n",
+    "print(\"Model:\", save_model_name, \"\\nLocation:\", filepath, \"\\n\")\n",
+    "\n",
+    "# training\n",
+    "checkpointer = ModelCheckpoint(filepath=filepath, verbose=1, save_best_only=True)\n",
+    "lr_reducer = ReduceLROnPlateau(factor=0.1, patience=5, min_lr=0.00001, verbose=1)\n",
+    "model.fit(x = X_train, \n",
+    "        y = to_categorical(Y_train), \n",
+    "        epochs = epochs, \n",
+    "        batch_size = batch_size,\n",
+    "        verbose = 1,\n",
+    "        validation_data=(X_test, to_categorical(Y_test)),\n",
+    "        callbacks=[checkpointer, lr_reducer])"
    ]
   },
   {

code/main.py

@@ -13,7 +13,7 @@
 subject = [1,2,3,4]
 task = "A"    # choose between "A" or "B"
 model_names = ["Convolutional", "Convolutional1DRecurrent", "Convolutional2DRecurrent", "ConvolutionalDeepRecurrent"]
-data_folder = "../data/full/"
+data_folder = "./data/full/"
 window_size = 15
 stride = 5
 GPU = True

code/main_multiuser.py

@@ -13,7 +13,7 @@
 subject = [23]
 task = "A"    # choose between "A" or "B"
 model_names = ["Convolutional", "Convolutional1DRecurrent", "Convolutional2DRecurrent", "ConvolutionalDeepRecurrent"]
-data_folder = "../data/full/"
+data_folder = "./data/full/"
 window_size = 15
 stride = 5
 GPU = True