288 lines
11 KiB
Plaintext
288 lines
11 KiB
Plaintext
{
|
||
"cells": [
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"这是Python，TensorFlow和Keras教程系列的深度学习基础知识的第4部分。\n",
|
||
"\n",
|
||
"在这一部分，我们将讨论的是TensorBoard。TensorBoard是一个方便的应用程序，允许您在浏览器中查看模型或模型的各个方面。我们将TensorBoard与Keras一起使用的方式是通过Keras回调。实际上有很多Keras回调，你可以自己制作。"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 1,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stderr",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Using TensorFlow backend.\n"
|
||
]
|
||
}
|
||
],
|
||
"source": [
|
||
"from tensorflow.keras.callbacks import TensorBoard"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 2,
|
||
"metadata": {
|
||
"collapsed": true
|
||
},
|
||
"outputs": [],
|
||
"source": [
|
||
"#创建TensorBoard回调对象\n",
|
||
"NAME = \"Cats-vs-dogs-CNN\"\n",
|
||
"\n",
|
||
"tensorboard = TensorBoard(log_dir=\"logs/{}\".format(NAME))"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"最终，你会希望获得更多的自定义NAME，但现在这样做。因此，这将保存模型的训练数据logs/NAME，然后由TensorBoard读取。\n",
|
||
"\n",
|
||
"最后，我们可以通过将它添加到.fit方法中来将此回调添加到我们的模型中，例如：\n",
|
||
"```python\n",
|
||
"model.fit(X, y,\n",
|
||
" batch_size=32,\n",
|
||
" epochs=3,\n",
|
||
" validation_split=0.3,\n",
|
||
" callbacks=[tensorboard])\n",
|
||
"```\n",
|
||
"请注意，这callbacks是一个列表。您也可以将其他回调传递到此列表中。我们的模型还没有定义，所以现在让我们把它们放在一起："
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 4,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Train on 17462 samples, validate on 7484 samples\n",
|
||
"Epoch 1/3\n",
|
||
"17462/17462 [==============================] - 44s 3ms/step - loss: 0.6992 - acc: 0.5480 - val_loss: 0.6900 - val_acc: 0.5274\n",
|
||
"Epoch 2/3\n",
|
||
"17462/17462 [==============================] - 41s 2ms/step - loss: 0.6754 - acc: 0.5782 - val_loss: 0.6685 - val_acc: 0.5885\n",
|
||
"Epoch 3/3\n",
|
||
"17462/17462 [==============================] - 41s 2ms/step - loss: 0.6377 - acc: 0.6483 - val_loss: 0.6217 - val_acc: 0.6625\n"
|
||
]
|
||
},
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"<keras.callbacks.History at 0x7ff86d691c18>"
|
||
]
|
||
},
|
||
"execution_count": 4,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"import tensorflow as tf\n",
|
||
"from tensorflow.keras.datasets import cifar10\n",
|
||
"from tensorflow.keras.preprocessing.image import ImageDataGenerator\n",
|
||
"from tensorflow.keras.models import Sequential\n",
|
||
"from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten\n",
|
||
"from tensorflow.keras.layers import Conv2D, MaxPooling2D\n",
|
||
"from tensorflow.keras.callbacks import TensorBoard\n",
|
||
"# more info on callbakcs: https://keras.io/callbacks/ model saver is cool too.\n",
|
||
"import pickle\n",
|
||
"import time\n",
|
||
"\n",
|
||
"NAME = \"Cats-vs-dogs-CNN\"\n",
|
||
"\n",
|
||
"pickle_in = open(\"../datasets/X.pickle\",\"rb\")\n",
|
||
"X = pickle.load(pickle_in)\n",
|
||
"\n",
|
||
"pickle_in = open(\"../datasets/y.pickle\",\"rb\")\n",
|
||
"y = pickle.load(pickle_in)\n",
|
||
"\n",
|
||
"X = X/255.0\n",
|
||
"\n",
|
||
"model = Sequential()\n",
|
||
"\n",
|
||
"model.add(Conv2D(256, (3, 3), input_shape=X.shape[1:]))\n",
|
||
"model.add(Activation('relu'))\n",
|
||
"model.add(MaxPooling2D(pool_size=(2, 2)))\n",
|
||
"\n",
|
||
"model.add(Conv2D(256, (3, 3)))\n",
|
||
"model.add(Activation('relu'))\n",
|
||
"model.add(MaxPooling2D(pool_size=(2, 2)))\n",
|
||
"\n",
|
||
"model.add(Flatten()) # this converts our 3D feature maps to 1D feature vectors\n",
|
||
"model.add(Dense(64))\n",
|
||
"\n",
|
||
"model.add(Dense(1))\n",
|
||
"model.add(Activation('sigmoid'))\n",
|
||
"\n",
|
||
"tensorboard = TensorBoard(log_dir=\"logs/{}\".format(NAME))\n",
|
||
"\n",
|
||
"model.compile(loss='binary_crossentropy',\n",
|
||
" optimizer='adam',\n",
|
||
" metrics=['accuracy'],\n",
|
||
" )\n",
|
||
"\n",
|
||
"model.fit(X, y,\n",
|
||
" batch_size=32,\n",
|
||
" epochs=3,\n",
|
||
" validation_split=0.3,\n",
|
||
" callbacks=[tensorboard])"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"运行此之后，您应该有一个名为的新目录logs。我们现在可以使用tensorboard从这个目录中可视化初始结果。打开控制台，切换到工作目录，然后键入：tensorboard --logdir=logs/。您应该看到一个通知：TensorBoard 1.10.0 at http://H-PC:6006 (Press CTRL+C to quit)“h-pc”是您机器的名称。打开浏览器并前往此地址。你应该看到类似的东西：\n",
|
||
"<img src = \"https://pythonprogramming.net/static/images/machine-learning/tensorboard-basic.png\">\n",
|
||
"现在我们可以看到我们的模型随着时间的推移。让我们改变模型中的一些东西。首先，我们从未在密集层中添加激活。另外，让我们尝试整体较小的模型："
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": 5,
|
||
"metadata": {},
|
||
"outputs": [
|
||
{
|
||
"name": "stdout",
|
||
"output_type": "stream",
|
||
"text": [
|
||
"Train on 17462 samples, validate on 7484 samples\n",
|
||
"Epoch 1/10\n",
|
||
"17462/17462 [==============================] - 11s 604us/step - loss: 0.6033 - acc: 0.6652 - val_loss: 0.5298 - val_acc: 0.7320\n",
|
||
"Epoch 2/10\n",
|
||
"17462/17462 [==============================] - 11s 646us/step - loss: 0.4859 - acc: 0.7659 - val_loss: 0.4723 - val_acc: 0.7763\n",
|
||
"Epoch 3/10\n",
|
||
"17462/17462 [==============================] - 11s 641us/step - loss: 0.4270 - acc: 0.8045 - val_loss: 0.4603 - val_acc: 0.7803\n",
|
||
"Epoch 4/10\n",
|
||
"17462/17462 [==============================] - 12s 699us/step - loss: 0.3675 - acc: 0.8347 - val_loss: 0.4476 - val_acc: 0.7929\n",
|
||
"Epoch 5/10\n",
|
||
"17462/17462 [==============================] - 12s 707us/step - loss: 0.3012 - acc: 0.8694 - val_loss: 0.4854 - val_acc: 0.7797\n",
|
||
"Epoch 6/10\n",
|
||
"17462/17462 [==============================] - 12s 705us/step - loss: 0.2165 - acc: 0.9118 - val_loss: 0.5450 - val_acc: 0.7865\n",
|
||
"Epoch 7/10\n",
|
||
"17462/17462 [==============================] - 12s 712us/step - loss: 0.1332 - acc: 0.9510 - val_loss: 0.6512 - val_acc: 0.7821\n",
|
||
"Epoch 8/10\n",
|
||
"17462/17462 [==============================] - 12s 705us/step - loss: 0.0764 - acc: 0.9743 - val_loss: 0.7487 - val_acc: 0.7809\n",
|
||
"Epoch 9/10\n",
|
||
"17462/17462 [==============================] - 12s 713us/step - loss: 0.0389 - acc: 0.9887 - val_loss: 0.9041 - val_acc: 0.7743\n",
|
||
"Epoch 10/10\n",
|
||
"17462/17462 [==============================] - 12s 708us/step - loss: 0.0287 - acc: 0.9921 - val_loss: 1.0411 - val_acc: 0.7702\n"
|
||
]
|
||
},
|
||
{
|
||
"data": {
|
||
"text/plain": [
|
||
"<keras.callbacks.History at 0x7ff86073ec50>"
|
||
]
|
||
},
|
||
"execution_count": 5,
|
||
"metadata": {},
|
||
"output_type": "execute_result"
|
||
}
|
||
],
|
||
"source": [
|
||
"from tensorflow.keras.models import Sequential\n",
|
||
"from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten\n",
|
||
"from tensorflow.keras.layers import Conv2D, MaxPooling2D\n",
|
||
"from tensorflow.keras.callbacks import TensorBoard\n",
|
||
"# more info on callbakcs: https://keras.io/callbacks/ model saver is cool too.\n",
|
||
"import pickle\n",
|
||
"import time\n",
|
||
"\n",
|
||
"NAME = \"Cats-vs-dogs-64x2-CNN\"\n",
|
||
"\n",
|
||
"pickle_in = open(\"../datasets/X.pickle\",\"rb\")\n",
|
||
"X = pickle.load(pickle_in)\n",
|
||
"\n",
|
||
"pickle_in = open(\"../datasets/y.pickle\",\"rb\")\n",
|
||
"y = pickle.load(pickle_in)\n",
|
||
"\n",
|
||
"X = X/255.0\n",
|
||
"\n",
|
||
"model = Sequential()\n",
|
||
"\n",
|
||
"model.add(Conv2D(64, (3, 3), input_shape=X.shape[1:]))\n",
|
||
"model.add(Activation('relu'))\n",
|
||
"model.add(MaxPooling2D(pool_size=(2, 2)))\n",
|
||
"\n",
|
||
"model.add(Conv2D(64, (3, 3)))\n",
|
||
"model.add(Activation('relu'))\n",
|
||
"model.add(MaxPooling2D(pool_size=(2, 2)))\n",
|
||
"\n",
|
||
"model.add(Flatten()) # this converts our 3D feature maps to 1D feature vectors\n",
|
||
"model.add(Dense(64))\n",
|
||
"model.add(Activation('relu'))\n",
|
||
"\n",
|
||
"model.add(Dense(1))\n",
|
||
"model.add(Activation('sigmoid'))\n",
|
||
"\n",
|
||
"tensorboard = TensorBoard(log_dir=\"logs/{}\".format(NAME))\n",
|
||
"\n",
|
||
"model.compile(loss='binary_crossentropy',\n",
|
||
" optimizer='adam',\n",
|
||
" metrics=['accuracy'],\n",
|
||
" )\n",
|
||
"\n",
|
||
"model.fit(X, y,\n",
|
||
" batch_size=32,\n",
|
||
" epochs=10,\n",
|
||
" validation_split=0.3,\n",
|
||
" callbacks=[tensorboard])"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "markdown",
|
||
"metadata": {},
|
||
"source": [
|
||
"除此之外，我还改名为NAME = \"Cats-vs-dogs-64x2-CNN\"。不要忘记这样做，否则你会偶然附加到你以前的型号的日志，它看起来不太好。我们现在检查TensorBoard：\n",
|
||
"<img src = \"https://pythonprogramming.net/static/images/machine-learning/second-model-tensorboard.png\">\n",
|
||
"看起来更好！但是，您可能会立即注意到验证丢失的形状。损失是衡量错误的标准，看起来很明显，在我们的第四个时代之后，事情开始变得糟糕。\n",
|
||
"\n",
|
||
"有趣的是，我们的验证准确性仍然持续，但我想它最终会开始下降。更可能的是，第一件遭受的事情确实是你的验证损失。这应该提醒你，你几乎肯定会开始过度适应。这种情况发生的原因是该模型不断尝试减少样本损失。\n",
|
||
"\n",
|
||
"在某些时候，模型不是学习关于实际数据的一般事物，而是开始只记忆输入数据。如果你继续这样做，是的，样本中的“准确性”会上升，但你的样本，以及你试图为模型提供的任何新数据可能会表现得很差。"
|
||
]
|
||
},
|
||
{
|
||
"cell_type": "code",
|
||
"execution_count": null,
|
||
"metadata": {
|
||
"collapsed": true
|
||
},
|
||
"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": "3.6.2"
|
||
}
|
||
},
|
||
"nbformat": 4,
|
||
"nbformat_minor": 2
|
||
}
|