nbs/keras_raw-xception-149.ipynb

keras_raw-xception-149.ipynb

{
  "cells": [
    {
      "metadata": {},
      "cell_type": "markdown",
      "source": "## Start"
    },
    {
      "metadata": {
        "collapsed": true,
        "trusted": true
      },
      "cell_type": "code",
      "source": "%reload_ext autoreload\n%autoreload 2\n%matplotlib inline",
      "execution_count": 1,
      "outputs": []
    },
    {
      "metadata": {
        "trusted": true
      },
      "cell_type": "code",
      "source": "from imports import *\nfrom keras.applications import xception",
      "execution_count": 2,
      "outputs": [
        {
          "output_type": "stream",
          "text": "Using TensorFlow backend.\n",
          "name": "stderr"
        }
      ]
    },
    {
      "metadata": {
        "collapsed": true,
        "trusted": true
      },
      "cell_type": "code",
      "source": "bs=64; sz=149; lr=2e-3\npath = \"/data/jhoward/fast/dogscats/\"",
      "execution_count": 3,
      "outputs": []
    },
    {
      "metadata": {
        "collapsed": true,
        "trusted": true
      },
      "cell_type": "code",
      "source": "gen = image.ImageDataGenerator(preprocessing_function=xception.preprocess_input)",
      "execution_count": 4,
      "outputs": []
    },
    {
      "metadata": {},
      "cell_type": "markdown",
      "source": "## 1. Fine-tune last layer of full network"
    },
    {
      "metadata": {
        "trusted": true,
        "collapsed": true
      },
      "cell_type": "code",
      "source": "K.set_learning_phase(0)",
      "execution_count": 5,
      "outputs": []
    },
    {
      "metadata": {
        "trusted": true,
        "collapsed": true
      },
      "cell_type": "code",
      "source": "mn=Xception(include_top=False, input_shape=(sz,sz,3), pooling='avg')\nfor l in mn.layers[:-1]: l.trainable=False",
      "execution_count": 6,
      "outputs": []
    },
    {
      "metadata": {
        "trusted": true
      },
      "cell_type": "code",
      "source": "outp = Dense(2, activation='softmax')(mn.output)\nm = Model(mn.input, outp)\nK.set_learning_phase(0)\nm.compile(SGD(lr, momentum=0.9), 'categorical_crossentropy', metrics=['accuracy'])",
      "execution_count": 7,
      "outputs": []
    },
    {
      "metadata": {
        "scrolled": true,
        "trusted": true
      },
      "cell_type": "code",
      "source": "trn_batches = gen.flow_from_directory(f'{path}train', (sz,sz), batch_size=bs)\nval_batches = gen.flow_from_directory(f'{path}valid', (sz,sz), batch_size=bs, shuffle=False)\nnb_trn = math.ceil(trn_batches.n/bs)\nnb_val = math.ceil(val_batches.n/bs)",
      "execution_count": 8,
      "outputs": [
        {
          "output_type": "stream",
          "text": "Found 23000 images belonging to 2 classes.\nFound 2000 images belonging to 2 classes.\n",
          "name": "stdout"
        }
      ]
    },
    {
      "metadata": {
        "scrolled": false,
        "trusted": true
      },
      "cell_type": "code",
      "source": "K.set_learning_phase(0)\nm.fit_generator(trn_batches, nb_trn, workers=1, epochs=2,\n                 validation_data=val_batches, validation_steps=nb_val)",
      "execution_count": 9,
      "outputs": [
        {
          "output_type": "stream",
          "text": "Epoch 1/2\n360/360 [==============================] - 68s - loss: 0.3416 - acc: 0.8640 - val_loss: 0.2901 - val_acc: 0.8880\nEpoch 2/2\n360/360 [==============================] - 67s - loss: 0.2644 - acc: 0.8900 - val_loss: 0.2543 - val_acc: 0.8945\n",
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "execution_count": 9,
          "data": {
            "text/plain": "<keras.callbacks.History at 0x7fb74a4c9940>"
          },
          "metadata": {}
        }
      ]
    },
    {
      "metadata": {},
      "cell_type": "markdown",
      "source": "## 2. Pre-compute output of penultimate layer and train single layer net"
    },
    {
      "metadata": {
        "collapsed": true,
        "trusted": true
      },
      "cell_type": "code",
      "source": "mn=Xception(include_top=False, input_shape=(sz,sz,3), pooling='avg')",
      "execution_count": 10,
      "outputs": []
    },
    {
      "metadata": {},
      "cell_type": "markdown",
      "source": "Precompute pooling output:"
    },
    {
      "metadata": {
        "trusted": true
      },
      "cell_type": "code",
      "source": "fix_batches = gen.flow_from_directory(f'{path}train', (sz,sz), batch_size=bs, shuffle=False)\nval_batches = gen.flow_from_directory(f'{path}valid', (sz,sz), batch_size=bs, shuffle=False)\n\ntrn_acts = mn.predict_generator(generator=fix_batches, verbose=1, \n                                steps=nb_trn, workers=1)\nval_acts = mn.predict_generator(generator=val_batches, verbose=1,\n                                steps=nb_val, workers=1)",
      "execution_count": 11,
      "outputs": [
        {
          "output_type": "stream",
          "text": "Found 23000 images belonging to 2 classes.\nFound 2000 images belonging to 2 classes.\n31/32 [============================>.] - ETA: 0s",
          "name": "stdout"
        }
      ]
    },
    {
      "metadata": {},
      "cell_type": "markdown",
      "source": "Train single layer:"
    },
    {
      "metadata": {
        "collapsed": true,
        "trusted": true
      },
      "cell_type": "code",
      "source": "inp = Input(batch_shape=mn.output_shape)\noutp = Dense(1, activation='sigmoid')(inp)\nfc = Model(inp, outp)\nfc.compile(SGD(lr, momentum=0.9), 'binary_crossentropy', metrics=['accuracy'])",
      "execution_count": 12,
      "outputs": []
    },
    {
      "metadata": {
        "trusted": true
      },
      "cell_type": "code",
      "source": "fc.fit(trn_acts, fix_batches.classes, bs, 2, validation_data=(val_acts, val_batches.classes))",
      "execution_count": 13,
      "outputs": [
        {
          "output_type": "stream",
          "text": "Train on 23000 samples, validate on 2000 samples\nEpoch 1/2\n23000/23000 [==============================] - 1s - loss: 0.1184 - acc: 0.9540 - val_loss: 0.0836 - val_acc: 0.9690\nEpoch 2/2\n23000/23000 [==============================] - 0s - loss: 0.0802 - acc: 0.9693 - val_loss: 0.0774 - val_acc: 0.9710\n",
          "name": "stdout"
        },
        {
          "output_type": "execute_result",
          "execution_count": 13,
          "data": {
            "text/plain": "<keras.callbacks.History at 0x7fb6d58fce10>"
          },
          "metadata": {}
        }
      ]
    },
    {
      "metadata": {
        "collapsed": true,
        "trusted": true
      },
      "cell_type": "code",
      "source": "",
      "execution_count": null,
      "outputs": []
    }
  ],
  "metadata": {
    "kernelspec": {
      "name": "conda-root-py",
      "display_name": "Python [conda root]",
      "language": "python"
    },
    "language_info": {
      "name": "python",
      "version": "3.6.2",
      "mimetype": "text/x-python",
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "pygments_lexer": "ipython3",
      "nbconvert_exporter": "python",
      "file_extension": ".py"
    },
    "gist": {
      "id": "",
      "data": {
        "description": "nbs/keras_raw-xception-149.ipynb",
        "public": true
      }
    }
  },
  "nbformat": 4,
  "nbformat_minor": 2
}