{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "%matplotlib inline" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "\npynet: self supervised clustering\n=================================\n\nCredit: A Grigis\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "import os\nimport sys\nif \"CI_MODE\" in os.environ:\n sys.exit()\n\n# Imports\nimport collections\nimport logging\nimport pynet\nfrom pynet.metrics import SKMetrics\nfrom pynet.datasets import DataManager\nfrom pynet.interfaces import DeepLearningInterface\nfrom pynet.interfaces import DeepClusterClassifier\nfrom pynet.models import BrainNetCNN\nfrom pynet.utils import setup_logging\nfrom pynet.plotting import Board, update_board\nfrom pynet.models.deepcluster import update_pseudo_labels\nimport torch\nimport torch.nn as nn\nfrom torch.utils.data.sampler import Sampler\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport pandas as pd\nfrom sklearn.cluster import MiniBatchKMeans, KMeans\nfrom sklearn.metrics import classification_report\nfrom sklearn.metrics import roc_curve, auc\ntry:\n import faiss\nexcept:\n pass\n\n\n# Global Parameters\nOUTDIR = \"/tmp/graph_connectome\"\nBATCH_SIZE = 5\nN_EPOCHS = 20\nN_CLUSTERS = 2\nN_SAMPLES = 40\nAVOID_EMPTY_CLUSTERS = False\nUNIFORM_SAMPLING = True\nsetup_logging(level=\"info\")\n\n\n# Load the data\ndata = []\nlabels = []\nfor idx in range(N_CLUSTERS):\n\n x_train = np.ones((N_SAMPLES, 1, 90, 90)) * idx\n x_train += (np.random.rand(*x_train.shape) - 0.5) * 0.01\n y_train = np.asarray([idx] * N_SAMPLES)\n print(\"sub data: x {0} - y {1}\".format(x_train.shape, y_train.shape))\n data.append(x_train)\n labels.extend([idx] * len(x_train))\n\n # Show example noisy training data that have the signatures applied.\n # It's not obvious to the human eye the subtle differences, but the cross\n # row and column above perturbed the below matrices with the y weights.\n # Show in the title how much each signature is weighted by.\n plt.figure(figsize=(16, 4))\n for idx in range(3):\n plt.subplot(1, 3, idx + 1)\n plt.imshow(np.squeeze(x_train[idx]), interpolation=\"None\")\n plt.colorbar()\n plt.title(y_train[idx])\n\ndata = np.concatenate(data, axis=0)\nlabels = np.asarray(labels)\nprint(\"dataset: x {0} - y {1}\".format(data.shape, labels.shape))\n\n\nclass UniformLabelSampler(Sampler):\n \"\"\" Samples elements uniformely accross pseudo labels.\n \"\"\"\n def __init__(self, data_array):\n \"\"\" Init class.\n\n Parameters\n ----------\n data_array: ArrayDataset\n the train data array that contains the pseudo labels.\n \"\"\"\n self.n_samples = len(data_array)\n self.data_array = data_array\n self.indexes = self.generate_indexes_epoch()\n\n def generate_indexes_epoch(self):\n \"\"\" Generate sampling indexes.\n \"\"\"\n labels = self.data_array.labels\n clusters_to_images = self.get_clusters(labels)\n n_non_empty_clusters = len(clusters_to_images)\n size_per_pseudolabel = int(self.n_samples / n_non_empty_clusters) + 1\n res = np.array([])\n for name, cluster_indexes in clusters_to_images.items():\n indexes = np.random.choice(\n cluster_indexes, size_per_pseudolabel,\n replace=(len(cluster_indexes) <= size_per_pseudolabel))\n res = np.concatenate((res, indexes))\n np.random.shuffle(res)\n res = list(res.astype(\"int\"))\n if len(res) >= self.n_samples:\n return res[:self.n_samples]\n res += res[: (self.n_samples - len(res))]\n return res\n\n def get_clusters(self, labels):\n \"\"\" Get indexes associated to each cluster.\n \"\"\"\n tally = collections.defaultdict(list)\n for idx, item in enumerate(labels):\n tally[item].append(idx)\n return tally\n\n def __iter__(self):\n return iter(self.indexes)\n\n def __len__(self):\n return len(self.indexes)\n\n\n# Create data manager\nif UNIFORM_SAMPLING:\n sampler = UniformLabelSampler\nelse:\n sampler = \"random\"\nmanager = DataManager.from_numpy(\n train_inputs=data, train_labels=np.zeros(labels.shape),\n batch_size=BATCH_SIZE, sampler=sampler)\n\n\nclass FKmeans(object):\n def __init__(self, n_clusters):\n self.n_clusters = n_clusters\n\n def fit(self, data):\n n_data, d = data.shape\n self.clus = faiss.Kmeans(d, self.n_clusters)\n self.clus.seed = np.random.randint(1234)\n self.clus.niter = 20\n self.clus.max_points_per_centroid = 10000000\n self.clus.train(data)\n\n def predict(self, data):\n _, I = self.clus.index.search(data, 1)\n losses = self.clus.obj\n print(\"k-means loss evolution: {0}\".format(losses))\n return np.asarray([int(n[0]) for n in I])\n\n\ndef my_loss(x, y):\n criterion = nn.CrossEntropyLoss()\n print(\" x: {0} - {1}\".format(x.shape, x.dtype))\n print(torch.argmax(x, dim=1))\n print(\" y: {0} - {1}\".format(y.shape, y.dtype))\n print(y)\n return criterion(x, y)\n\n\n# Create model\ntrain_loader = manager.get_dataloader(train=True, fold_index=0).train\nif AVOID_EMPTY_CLUSTERS:\n kmeans = FKmeans(n_clusters=N_CLUSTERS)\nelse:\n kmeans = KMeans(\n n_clusters=N_CLUSTERS,\n random_state=None,\n # verbose=100,\n max_iter=20)\nnet = BrainNetCNN(\n input_shape=(90, 90),\n in_channels=1,\n num_classes=N_CLUSTERS,\n nb_e2e=32,\n nb_e2n=64,\n nb_n2g=30,\n dropout=0,\n leaky_alpha=0.1,\n twice_e2e=False,\n dense_sml=False)\nnet_params = pynet.NetParameters(\n network=net,\n clustering=kmeans,\n data_loader=train_loader,\n n_batchs=10,\n pca_dim=6,\n assignment_logfile=None,\n use_cuda=False)\nmodel = DeepClusterClassifier(\n net_params,\n optimizer_name=\"SGD\",\n learning_rate=0.001,\n momentum=0.9,\n weight_decay=10**-5,\n # loss=my_loss)\n loss_name=\"CrossEntropyLoss\")\nmodel.board = Board(port=8097, host=\"http://localhost\", env=\"deepcluster\")\nmodel.add_observer(\"before_epoch\", update_pseudo_labels)\nmodel.add_observer(\"after_epoch\", update_board)\n\n\n# Train model\ntest_history, train_history = model.training(\n manager=manager,\n nb_epochs=N_EPOCHS,\n checkpointdir=None,\n fold_index=0,\n scheduler=None,\n with_validation=False)\n\n\n# Test model\nmanager = DataManager.from_numpy(\n test_inputs=data, test_labels=labels, batch_size=BATCH_SIZE)\ntest_model = DeepLearningInterface(\n model=model.model.network,\n optimizer_name=\"SGD\",\n learning_rate=0.01,\n momentum=0.9,\n weight_decay=10**-5,\n loss_name=\"CrossEntropyLoss\")\ny_pred, X, y_true, loss, values = test_model.testing(\n manager=manager,\n with_logit=True,\n # logit_function=\"sigmoid\",\n predict=False)\nprint(y_pred.shape, X.shape, y_true.shape)\n\n\n# Inspect results\nresult = pd.DataFrame.from_dict(collections.OrderedDict([\n (\"pred\", (np.argmax(y_pred, axis=1)).astype(int)),\n (\"truth\", y_true.squeeze()),\n (\"prob_0\", y_pred[:, 0].squeeze()),\n (\"prob_1\", y_pred[:, 1].squeeze())]))\nprint(result)\ny_pred = np.argmax(y_pred, axis=1)\nfig, ax = plt.subplots()\ncmap = plt.get_cmap(\"Blues\")\ncm = SKMetrics(\"confusion_matrix\", with_logit=False)(y_pred, y_true)\nsns.heatmap(cm, cmap=cmap, annot=True, fmt=\"g\", ax=ax)\nax.set_xlabel(\"predicted values\")\nax.set_ylabel(\"actual values\")\nprint(classification_report(y_true, y_pred >= 0.4))\nfpr, tpr, _ = roc_curve(y_true, y_pred)\nroc_auc = auc(fpr, tpr)\nplt.figure()\nplt.plot(fpr, tpr, color=\"darkorange\", lw=2,\n label=\"ROC curve (area = %0.2f)\" % roc_auc)\nplt.plot([0, 1], [0, 1], color=\"navy\", lw=2, linestyle=\"--\")\nplt.xlim([0.0, 1.0])\nplt.ylim([0.0, 1.05])\nplt.xlabel(\"False Positive Rate\")\nplt.ylabel(\"True Positive Rate\")\nplt.title(\"Receiver operating characteristic\")\nplt.legend(loc=\"lower right\")\n\nplt.show()" ] } ], "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": "3.6.12" } }, "nbformat": 4, "nbformat_minor": 0 }