some housekeeping

.gitignore

@@ -5,11 +5,13 @@ __pycache__/
 # IPython checkpoints
 .ipynb_checkpoints/
 
-# Datasets
+# Local configuration
+.env
+.python-version
+
+# Data
 data/
 
-# Tensorflow summaries
+# Tensorflow summaries & model parameters
 summaries/
-
-# Model parameters
 checkpoints/

README.md

@@ -74,6 +74,33 @@ setting.
 
 [usage]: http://nbviewer.jupyter.org/github/mdeff/cnn_graph/blob/outputs/usage.ipynb
 
+## Repository organization
+
+* The models (the introduced model and some reference models) are contained in [models.py](models.py).
+* Various side functions are implemented in [graph.py](graph.py), [coarsening.py](coarsening.py) and [utils.py](utils.py).
+* We did experiments on three datasets: MNIST ([notebook](mnist.ipynb)), 20NEWS ([notebook](20news.ipynb)) and RCV1 ([notebook](rcv1.ipynb)).
+* TensorBoard summaries are saved in the `summaries` folder.
+* Model parameters are saved in the `checkpoints` folder.
+* Data is placed in the `data` folder.
+	* [MNIST](http://yann.lecun.com/exdb/mnist/) is downloaded automatically.
+	* [20NEWS](http://qwone.com/~jason/20Newsgroups/) (`20news-bydate.tar.gz`) is downloaded automatically.
+	* [RCV1](http://trec.nist.gov/data/reuters/reuters.html) should be downloaded manually and placed in TODO.
+	* [pre-trained word2vec embeddings](https://code.google.com/archive/p/word2vec/) (`GoogleNews-vectors-negative300.bin.gz`).
+	* Wikipedia graph and activations are available here. Please cite .. if you use it.
+* The [trials](trials) folder contains various small experiences in the form of IPython notebooks.
+	1. [Learning graph filters][trial1]: first experiments on learning
+	   synthesized graph filters through observations of filtered and source
+	   graph signals. The Chebychev and Lanczos methods as well as optimization
+	   methods are compared there.
+	2. [Classification][trial2]: learning filters who extract good features for
+	   classification.
+	3. [TensorFlow][trial3]: first experience with TensorFlow.
+	4. [Coarsening][trial4]: implementation of the Graclus coarsening algorithm
+	   and comparison with a previous matlab implementation.
+* A [makefile](makefile) who runs every notebook as a sanity check. It only runs the code, there is no check on the results.
+
+[trial1]: h
+
 ## License & co
 
 The code in this repository is released under the terms of the [MIT license](LICENSE.txt).

nips2016/mnist.ipynb → experiments/1_mnist.ipynb

@@ -3,9 +3,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "%load_ext autoreload\n",
@@ -25,9 +23,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "flags = tf.app.flags\n",
@@ -54,9 +50,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "def grid_graph(m, corners=False):\n",
@@ -96,9 +90,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "from tensorflow.examples.tutorials.mnist import input_data\n",
@@ -116,7 +108,7 @@
     "val_data = coarsening.perm_data(val_data, perm)\n",
     "test_data = coarsening.perm_data(test_data, perm)\n",
     "print('Execution time: {:.2f}s'.format(time.process_time() - t_start))\n",
-    "del perm"
+    "#del perm"
    ]
   },
   {
@@ -129,9 +121,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": true
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "#model = fc1()\n",
@@ -157,9 +147,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "common = {}\n",
@@ -178,9 +166,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "if True:\n",
@@ -203,9 +189,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": true
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "# Common hyper-parameters for networks with one convolutional layer.\n",
@@ -223,9 +207,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "if True:\n",
@@ -234,33 +216,31 @@
     "    params['dir_name'] += name\n",
     "    params['filter'] = 'fourier'\n",
     "    params['K'] = [L[0].shape[0]]\n",
-    "    model_perf.test(models.cgcnn(L, **params), name, params,\n",
+    "    model_f = models.cgcnn(L, **params)\n",
+    "    model_perf.test(model_f, name, params,\n",
     "                    train_data, train_labels, val_data, val_labels, test_data, test_labels)"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "if True:\n",
     "    name = 'sgconv_softmax'\n",
     "    params = common.copy()\n",
     "    params['dir_name'] += name\n",
     "    params['filter'] = 'spline'\n",
-    "    model_perf.test(models.cgcnn(L, **params), name, params,\n",
+    "    model_s = models.cgcnn(L, **params)\n",
+    "    model_perf.test(model_s, name, params,\n",
     "                    train_data, train_labels, val_data, val_labels, test_data, test_labels)"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "# With 'chebyshev2' and 'b2relu', it corresponds to cgcnn2_2(L[0], F=10, K=20).\n",
@@ -271,16 +251,15 @@
     "    params['filter'] = 'chebyshev5'\n",
     "#    params['filter'] = 'chebyshev2'\n",
     "#    params['brelu'] = 'b2relu'\n",
-    "    model_perf.test(models.cgcnn(L, **params), name, params,\n",
+    "    model_c = models.cgcnn(L, **params)\n",
+    "    model_perf.test(model_c, name, params,\n",
     "                    train_data, train_labels, val_data, val_labels, test_data, test_labels)"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": true
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "# Common hyper-parameters for LeNet5-like networks.\n",
@@ -298,9 +277,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "# Architecture of TF MNIST conv model (LeNet-5-like).\n",
@@ -319,9 +296,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "if True:\n",
@@ -336,9 +311,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "if True:\n",
@@ -353,9 +326,7 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": false
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "model_perf.show()"
@@ -364,16 +335,114 @@
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "collapsed": true
-   },
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def plot_filters(coeffs):\n",
+    "    fig = plt.figure(figsize=(15,5))\n",
+    "    ax = fig.add_subplot(1,1,1)\n",
+    "    for coeff in coeffs:\n",
+    "        c = eval(coeff)\n",
+    "        label = '{}: L={:1.2e}, |dL|={:1.2e}'.format(coeff, L(c), np.linalg.norm(dL(X,Y,c)))\n",
+    "        ax.plot(lamb, c, '.-', label=label)\n",
+    "#        np.testing.assert_allclose(np.linalg.norm(c)**2, E, rtol=1e-2)\n",
+    "    ax.set_xlim(lamb[0], lamb[-1])\n",
+    "    ax.set_title('Filter coefficients, M={}, N={}, eps={}'.format(M, N, eps))\n",
+    "    ax.set_xlabel('frequency')\n",
+    "    ax.set_ylabel('amplitude')\n",
+    "    ax.legend(loc='best')\n",
+    "\n",
+    "import matplotlib.pyplot as plt\n",
+    "import numpy as np"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
    "outputs": [],
    "source": [
     "if False:\n",
     "    grid_params = {}\n",
     "    data = (train_data, train_labels, val_data, val_labels, test_data, test_labels)\n",
     "    utils.grid_search(params, grid_params, *data, model=lambda x: models.cgcnn(L,**x))"
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Filter visualization (experimental)\n",
+    "\n",
+    "**Disclaimer**: left as is, not sure if it works. To be checked before usage."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "a = np.array([1,2,3])\n",
+    "print(a[[2,0,1]][:2])\n",
+    "\n",
+    "print(mnist.train.labels[0])\n",
+    "plt.imshow(mnist.train.images[0].reshape(28,28))\n",
+    "\n",
+    "a = np.random.permutation(range(len(perm)))\n",
+    "b = a[perm]\n",
+    "c = b[idx]\n",
+    "assert np.all(a == c)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "model = model_f\n",
+    "\n",
+    "sess = tf.Session(graph=model.graph)\n",
+    "filename = tf.train.latest_checkpoint(os.path.join('..', 'checkpoints', model.dir_name))\n",
+    "model.op_saver.restore(sess, filename)\n",
+    "var = model.graph.get_tensor_by_name('conv1/weights' + ':0')\n",
+    "val = sess.run(var)\n",
+    "sess.close()\n",
+    "\n",
+    "lamb, U = graph.fourier(L[0])\n",
+    "\n",
+    "#filters = model_f.get_var('conv1/weights')\n",
+    "filters = val\n",
+    "#filt_fourier = filt\n",
+    "\n",
+    "i = 6\n",
+    "\n",
+    "print(filters.shape)\n",
+    "plt.figure(figsize=(15,5))\n",
+    "plt.plot(lamb, filters[:,i,0])\n",
+    "\n",
+    "print(lamb[0], lamb[-1])\n",
+    "\n",
+    "filt = U.dot(filters[:,i,0])\n",
+    "plt.figure(figsize=(15,5))\n",
+    "plt.plot(filt)\n",
+    "\n",
+    "print(len(lamb)-28**2)\n",
+    "indices = np.array(perm) >= 28**2\n",
+    "\n",
+    "print(train_data[0,indices])\n",
+    "\n",
+    "idx = np.argsort(perm)\n",
+    "filt = filt[idx]\n",
+    "plt.figure(figsize=(15,5))\n",
+    "plt.plot(train_data[0,perm])\n",
+    "\n",
+    "plt.figure(figsize=(15,5))\n",
+    "img = train_data[0,idx][:28**2].reshape(28,28)\n",
+    "plt.imshow(train_data[0,idx][:28**2].reshape(28,28))\n",
+    "assert np.allclose(train_data[0,idx][:28**2].reshape(28,28), mnist.train.images[0].reshape(28,28))"
+   ]
   }
  ],
  "metadata": {