2018-07-23-Keras-sample_weight.ipynb

2018-07-23-keras-sample_weight.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "name": "2018-07-23-Keras-sample_weight.ipynb",
      "version": "0.3.2",
      "provenance": [],
      "include_colab_link": true
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {
        "id": "view-in-github",
        "colab_type": "text"
      },
      "source": [
        "[View in Colaboratory](https://colab.research.google.com/gist/yuzeh/0d4437f5557489987b7be3038bd5c58e/2018-07-23-keras-sample_weight.ipynb)"
      ]
    },
    {
      "metadata": {
        "id": "rj86XI8Rpd9v",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "import keras\n",
        "import numpy as np"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "EhEsaQjCpgUM",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 85
        },
        "outputId": "85ae3cc2-0791-41c3-f2bb-b2c3bb024a57"
      },
      "cell_type": "code",
      "source": [
        "# I'm trying to figure out how sample_weight works.\n",
        "\n",
        "# - learn model with two heads, y1 := x, y2 := -x\n",
        "\n",
        "input = keras.models.Input((1,))\n",
        "\n",
        "out1 = keras.layers.Dense(1, name = 'y1')(input)\n",
        "out2 = keras.layers.Dense(1, name = 'y2')(input)\n",
        "\n",
        "model = keras.models.Model(inputs = [input], outputs = [out1, out2])\n",
        "\n",
        "xs = np.array([0, 1, 2, 3, 4, 5]).astype(float)\n",
        "y1 = xs\n",
        "y2 = -xs\n",
        "\n",
        "kwargs = dict(\n",
        "    x = xs,\n",
        "    y = dict(y1 = y1, y2 = y2),\n",
        "    sample_weight = None,\n",
        "    batch_size = 3,\n",
        "    epochs = 200,\n",
        "    verbose = 0)\n",
        "\n",
        "model.compile('sgd', 'mse')\n",
        "model.fit(**kwargs)\n",
        "\n",
        "model.get_weights()"
      ],
      "execution_count": 7,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[array([[0.9956291]], dtype=float32),\n",
              " array([0.01543758], dtype=float32),\n",
              " array([[-0.99854046]], dtype=float32),\n",
              " array([-0.00515498], dtype=float32)]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 7
        }
      ]
    },
    {
      "metadata": {
        "id": "LfD-OE6vpy9O",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 85
        },
        "outputId": "b63b34a3-c828-4fd4-8987-fe8c4f9f33a8"
      },
      "cell_type": "code",
      "source": [
        "# Okay this looks about right.\n",
        "\n",
        "# - learn model with two heads, y1 := x, y2 := -x\n",
        "# - add sample_weight to model.fit\n",
        "\n",
        "input = keras.models.Input((1,))\n",
        "\n",
        "out1 = keras.layers.Dense(1, name = 'y1')(input)\n",
        "out2 = keras.layers.Dense(1, name = 'y2')(input)\n",
        "\n",
        "model = keras.models.Model(inputs = [input], outputs = [out1, out2])\n",
        "\n",
        "xs = np.array([0, 1, 2, 3, 4, 5]).astype(float)\n",
        "y1 = xs\n",
        "y2 = -xs\n",
        "\n",
        "kwargs = dict(\n",
        "    x = xs,\n",
        "    y = dict(y1 = y1, y2 = y2),\n",
        "    sample_weight = None,\n",
        "    batch_size = 3,\n",
        "    epochs = 200,\n",
        "    verbose = 0)\n",
        "\n",
        "kwargs['sample_weight'] = dict(\n",
        "    y1 = np.ones(6),\n",
        "    y2 = np.ones(6),\n",
        ")\n",
        "\n",
        "model.compile('sgd', 'mse')\n",
        "model.fit(**kwargs)\n",
        "\n",
        "model.get_weights()"
      ],
      "execution_count": 8,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[array([[1.0023069]], dtype=float32),\n",
              " array([-0.00831777], dtype=float32),\n",
              " array([[-0.99394155]], dtype=float32),\n",
              " array([-0.02184387], dtype=float32)]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 8
        }
      ]
    },
    {
      "metadata": {
        "id": "FDQSVzqOqSyX",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 85
        },
        "outputId": "972e1b08-3a22-410f-e93d-645f68076182"
      },
      "cell_type": "code",
      "source": [
        "# So far so good. This is where things get weird:\n",
        "\n",
        "# - learn model with two heads, y1 := x, y2 := -x\n",
        "# - add sample_weight to model.fit\n",
        "# - drop half of the training examples by setting sample_weight = 0 for each head.\n",
        "\n",
        "input = keras.models.Input((1,))\n",
        "\n",
        "out1 = keras.layers.Dense(1, name = 'y1')(input)\n",
        "out2 = keras.layers.Dense(1, name = 'y2')(input)\n",
        "\n",
        "model = keras.models.Model(inputs = [input], outputs = [out1, out2])\n",
        "\n",
        "xs = np.array([0, 1, 2, 3, 4, 5]).astype(float)\n",
        "y1 = xs\n",
        "y2 = -xs\n",
        "\n",
        "kwargs = dict(\n",
        "    x = xs,\n",
        "    y = dict(y1 = y1, y2 = y2),\n",
        "    sample_weight = None,\n",
        "    batch_size = 3,\n",
        "    epochs = 200,\n",
        "    verbose = 0)\n",
        "\n",
        "kwargs['sample_weight'] = dict(\n",
        "    y1 = np.ones(6),\n",
        "    y2 = np.ones(6),\n",
        ")\n",
        "\n",
        "kwargs['sample_weight']['y1'][[0, 2, 4]] = 0.\n",
        "kwargs['sample_weight']['y2'][[1, 3, 5]] = 0.\n",
        "\n",
        "model.compile('sgd', 'mse')\n",
        "model.fit(**kwargs)\n",
        "\n",
        "model.get_weights()"
      ],
      "execution_count": 9,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[array([[nan]], dtype=float32),\n",
              " array([nan], dtype=float32),\n",
              " array([[nan]], dtype=float32),\n",
              " array([nan], dtype=float32)]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 9
        }
      ]
    },
    {
      "metadata": {
        "id": "wZhjjJ0ArdNI",
        "colab_type": "code",
        "colab": {
          "base_uri": "https://localhost:8080/",
          "height": 34
        },
        "outputId": "34c2f974-68d5-491b-d5b5-57ced93cd96f"
      },
      "cell_type": "code",
      "source": [
        "# WTF???\n",
        "\n",
        "# Okay, how about just one head instead of two??\n",
        "\n",
        "input = keras.models.Input((1,))\n",
        "\n",
        "out1 = keras.layers.Dense(1, name = 'y1')(input)\n",
        "\n",
        "model = keras.models.Model(inputs = [input], outputs = [out1])\n",
        "\n",
        "xs = np.array([0, 1, 2, 3, 4, 5]).astype(float)\n",
        "y1 = xs\n",
        "\n",
        "kwargs = dict(\n",
        "    x = xs,\n",
        "    y = dict(y1 = y1),\n",
        "    sample_weight = None,\n",
        "    batch_size = 3,\n",
        "    epochs = 200,\n",
        "    verbose = 0)\n",
        "\n",
        "kwargs['sample_weight'] = dict(\n",
        "    y1 = np.ones(6),\n",
        ")\n",
        "\n",
        "kwargs['sample_weight']['y1'][[0, 2, 4]] = 0.\n",
        "\n",
        "model.compile('sgd', 'mse')\n",
        "model.fit(**kwargs)\n",
        "\n",
        "model.get_weights()"
      ],
      "execution_count": 10,
      "outputs": [
        {
          "output_type": "execute_result",
          "data": {
            "text/plain": [
              "[array([[nan]], dtype=float32), array([nan], dtype=float32)]"
            ]
          },
          "metadata": {
            "tags": []
          },
          "execution_count": 10
        }
      ]
    },
    {
      "metadata": {
        "id": "1QW2RXWgrtX7",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        "# At this point I'm pretty sure I don't understand how sample_weight is supposed to be used.\n",
        "# Someone please help me!"
      ],
      "execution_count": 0,
      "outputs": []
    },
    {
      "metadata": {
        "id": "z3tW0Fl9r2SX",
        "colab_type": "code",
        "colab": {}
      },
      "cell_type": "code",
      "source": [
        ""
      ],
      "execution_count": 0,
      "outputs": []
    }
  ]
}