Skip to content

Instantly share code, notes, and snippets.

@NTT123

NTT123/sonnet-mnist.ipynb

Last active December 13, 2018 18:23
Show Gist options
  • Save NTT123/9823e517b691094c64472d0f2baf6ae1 to your computer and use it in GitHub Desktop.
Save NTT123/9823e517b691094c64472d0f2baf6ae1 to your computer and use it in GitHub Desktop.
Download ZIP
Sonnet MNIST
Display the source blob
Display the rendered blob
Raw
sonnet-mnist.ipynb
{
"nbformat": 4,
"nbformat_minor": 0,
"metadata": {
"colab": {
"name": "Sonnet MNIST",
"version": "0.3.2",
"provenance": [],
"collapsed_sections": [
"gDIpKEqMFJsM",
"GxUq9wH5RTZL"
],
"include_colab_link": true
},
"kernelspec": {
"name": "python3",
"display_name": "Python 3"
},
"accelerator": "GPU"
},
"cells": [
{
"cell_type": "markdown",
"metadata": {
"id": "view-in-github",
"colab_type": "text"
},
"source": [
"<a href=\"https://colab.research.google.com/gist/NTT123/9823e517b691094c64472d0f2baf6ae1/sonnet-mnist.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>"
]
},
{
"metadata": {
"id": "542hNYow-N-S",
"colab_type": "text"
},
"cell_type": "markdown",
"source": [
"## Deepmind Sonnet + MNIST"
]
},
{
"metadata": {
"id": "HwfC79RJxqxf",
"colab_type": "toc"
},
"cell_type": "markdown",
"source": [
">>[Deepmind Sonnet + MNIST](#updateTitle=true&folderId=1vS4NPhLsF-s0dhVeKeGK54KqT7dstcOZ&scrollTo=542hNYow-N-S)\n",
"\n",
">>>[Model](#updateTitle=true&folderId=1vS4NPhLsF-s0dhVeKeGK54KqT7dstcOZ&scrollTo=nCn-Q0RsEjJV)\n",
"\n",
">>>[Tensorboard](#updateTitle=true&folderId=1vS4NPhLsF-s0dhVeKeGK54KqT7dstcOZ&scrollTo=ZJbtnfA2cq4M)\n",
"\n",
">>>[Dataloader](#updateTitle=true&folderId=1vS4NPhLsF-s0dhVeKeGK54KqT7dstcOZ&scrollTo=gDIpKEqMFJsM)\n",
"\n",
">>>[Define Loss and Accuracy](#updateTitle=true&folderId=1vS4NPhLsF-s0dhVeKeGK54KqT7dstcOZ&scrollTo=GxUq9wH5RTZL)\n",
"\n",
">>>[Training](#updateTitle=true&folderId=1vS4NPhLsF-s0dhVeKeGK54KqT7dstcOZ&scrollTo=2VTv6CcwRYx9)\n",
"\n"
]
},
{
"metadata": {
"id": "nCn-Q0RsEjJV",
"colab_type": "text"
},
"cell_type": "markdown",
"source": [
"### Model"
]
},
{
"metadata": {
"id": "DEX1uUX29-Jv",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"import tensorflow as tf\n",
"import sonnet as snt"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "PLMy4zgWaAnF",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"class Module(snt.AbstractModule):\n",
" \n",
" def __init__(self, *args, **kwargs):\n",
" super().__init__( *args, **kwargs)\n",
" \n",
" self.is_training = True\n",
" \n",
" def eval(self):\n",
" self.is_training = False\n",
" \n",
" def train(self):\n",
" self.is_training = True"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "9B8Rm6mX-Sd1",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"class MLP(Module):\n",
" def __init__(self):\n",
" super().__init__()\n",
" \n",
" with self._enter_variable_scope():\n",
" self.linear1 = snt.Linear(128)\n",
" self.linear2 = snt.Linear(256)\n",
" self.linear3 = snt.Linear(10)\n",
" \n",
" def _build(self, x):\n",
" x = tf.layers.flatten(x)\n",
" x = self.linear1(x)\n",
" x = tf.nn.relu(x)\n",
" x = tf.nn.dropout(x, 0.9) if self.is_training else x\n",
" x = self.linear2(x)\n",
" x = tf.nn.relu(x)\n",
" x = tf.nn.dropout(x, 0.9) if self.is_training else x\n",
" x = self.linear3(x)\n",
" return x"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "ZJbtnfA2cq4M",
"colab_type": "text"
},
"cell_type": "markdown",
"source": [
"### Tensorboard"
]
},
{
"metadata": {
"id": "cUfT26OlBNTN",
"colab_type": "code",
"outputId": "6e80a2a4-a26b-4a18-bcd5-8252376224d7",
"colab": {
"base_uri": "https://localhost:8080/",
"height": 102
}
},
"cell_type": "code",
"source": [
"!pip install tensorboardcolab\n",
"from tensorboardcolab import *\n",
"\n",
"x = tf.placeholder(tf.float32, shape=(1, 28,28))\n",
"net = MLP()\n",
"y = net(x)\n",
"tbc=TensorBoardColab(graph_path=\"./log\")\n",
"filewriter = tf.summary.FileWriter(\"./log/1\", net.graph)"
],
"execution_count": 0,
"outputs": [
{
"output_type": "stream",
"text": [
"Requirement already satisfied: tensorboardcolab in /usr/local/lib/python3.6/dist-packages (0.0.20)\n"
],
"name": "stdout"
},
{
"output_type": "stream",
"text": [
"Using TensorFlow backend.\n"
],
"name": "stderr"
},
{
"output_type": "stream",
"text": [
"Wait for 8 seconds...\n",
"TensorBoard link:\n",
"http://3e5efe18.ngrok.io\n"
],
"name": "stdout"
}
]
},
{
"metadata": {
"id": "gDIpKEqMFJsM",
"colab_type": "text"
},
"cell_type": "markdown",
"source": [
"### Dataloader"
]
},
{
"metadata": {
"id": "nsB_-B0nBYoH",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"tf.reset_default_graph()"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "pG2Zm8fgEXZP",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"mnist = tf.keras.datasets.mnist\n",
"(x_train, y_train),(x_test, y_test) = mnist.load_data()\n",
"x_train, x_test = x_train / 255.0, x_test / 255.0"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "lWDLlcEKKQLB",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"mnist_ds = tf.data.Dataset.from_tensor_slices((x_train,y_train))\n",
"mnist_ds = mnist_ds.repeat().batch(32, drop_remainder=True).prefetch(2)\n",
"mnist_ds_test = tf.data.Dataset.from_tensor_slices((x_test,y_test))\n",
"mnist_ds_test = mnist_ds_test.repeat().batch(128, drop_remainder=True).prefetch(2)"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "GxUq9wH5RTZL",
"colab_type": "text"
},
"cell_type": "markdown",
"source": [
"### Define Loss and Accuracy"
]
},
{
"metadata": {
"id": "sx0Ipn6FHLuV",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"it = mnist_ds.make_one_shot_iterator()\n",
"x, target = it.get_next()\n",
"x = tf.to_float(x)\n",
"target = tf.to_int32(target)\n",
"net = MLP()\n",
"y = net(x)"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "woVQ5044H_n4",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=target, logits=y)\n",
"loss = tf.reduce_mean(loss)\n",
"loss_summary = tf.summary.scalar(\"loss\", loss)"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "X4UN_Y5sh8y5",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"it_test = mnist_ds_test.make_one_shot_iterator()\n",
"x, target = it_test.get_next()\n",
"x = tf.to_float(x)\n",
"target = tf.to_int32(target)\n",
"net.eval()\n",
"y = net(x)\n",
"net.train()\n",
"y = tf.argmax(y, axis=1, output_type=tf.int32)\n",
"acc, acc_ops = tf.metrics.accuracy(labels=target, predictions=y)\n",
"with tf.control_dependencies([acc, acc_ops]):\n",
" acc_summary = tf.summary.scalar(\"accuracy\", acc)"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "2VTv6CcwRYx9",
"colab_type": "text"
},
"cell_type": "markdown",
"source": [
"### Training"
]
},
{
"metadata": {
"id": "aO9rP0u1IRtL",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"optimizer = tf.train.AdamOptimizer(1e-4)\n",
"train_ops = optimizer.minimize(loss, var_list=net.get_all_variables())"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "Vt3n57d3Mgjc",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"sess = tf.Session()\n",
"sess.run(tf.global_variables_initializer())\n",
"sess.run(tf.local_variables_initializer())"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "E5cFK5FWZCt8",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"step = 0"
],
"execution_count": 0,
"outputs": []
},
{
"metadata": {
"id": "399d4l7xMmVf",
"colab_type": "code",
"colab": {}
},
"cell_type": "code",
"source": [
"for step in range(step, 1_000_000):\n",
" _, ls = sess.run([train_ops, loss_summary])\n",
" filewriter.add_summary(ls, step)\n",
" \n",
" if step % 100 == 0:\n",
" accsum = sess.run(acc_summary)\n",
" filewriter.add_summary(accsum, step)"
],
"execution_count": 0,
"outputs": []
}
]
}
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment