Update after review

Author: VictorBebnev

Graph-Convolutional-Networks-Model-Relations-In-Data/README.md

@@ -1,4 +1,4 @@
 # Training
 For training run [jupyter notebook](graph_convolutional_networks_model_relations_in_data.ipynb)
 
-_Note: we use dataset from_ [_this post_](../PyTorch-Multi-Label-Image-Classification-Image-Tagging/README.md)
+_Note: we use the dataset from [this post](../PyTorch-Multi-Label-Image-Classification-Image-Tagging/README.md), so please refer to it for a detailed description and instructions._

Graph-Convolutional-Networks-Model-Relations-In-Data/graph_convolutional_networks_model_relations_in_data.ipynb

@@ -18,32 +18,33 @@
    "metadata": {},
    "outputs": [],
    "source": [
+    "import itertools\n",
+    "import json\n",
+    "import math\n",
     "import os\n",
+    "import random\n",
+    "import tarfile\n",
     "import time\n",
+    "import urllib.request\n",
+    "import zipfile\n",
+    "from shutil import copyfile\n",
+    "\n",
     "import numpy as np\n",
-    "from PIL import Image\n",
-    "from torch.utils.data.dataset import Dataset\n",
-    "from tqdm import tqdm\n",
-    "from torchvision import transforms\n",
-    "from torchvision import models\n",
+    "import requests\n",
     "import torch\n",
-    "from torch.utils.tensorboard import SummaryWriter\n",
-    "from sklearn.metrics import precision_score, recall_score, f1_score\n",
-    "from torch import nn\n",
-    "from torch.utils.data.dataloader import DataLoader\n",
+    "from PIL import Image\n",
     "from matplotlib import pyplot as plt\n",
     "from numpy import printoptions\n",
-    "import requests\n",
-    "import tarfile\n",
-    "import random\n",
-    "import json\n",
-    "from shutil import copyfile\n",
-    "import zipfile\n",
     "from sklearn.manifold import TSNE\n",
-    "import itertools\n",
+    "from sklearn.metrics import precision_score, recall_score, f1_score\n",
+    "from torch import nn\n",
     "from torch.nn import Parameter\n",
-    "import math\n",
-    "import urllib.request"
+    "from torch.utils.data.dataloader import DataLoader\n",
+    "from torch.utils.data.dataset import Dataset\n",
+    "from torch.utils.tensorboard import SummaryWriter\n",
+    "from torchvision import models\n",
+    "from torchvision import transforms\n",
+    "from tqdm import tqdm"
    ]
   },
   {
@@ -102,10 +103,9 @@
     "    with tarfile.open(path_to_tar_file) as tar_ref:\n",
     "        tar_ref.extractall(os.path.dirname(img_folder))\n",
     "    os.remove(path_to_tar_file)\n",
-    "# Also, copy our pre-processed annotations to the dataset folder. \n",
-    "# Note: you can find script for generating such annotations in attachments\n",
-    "copyfile('../PyTorch-Multi-Label-Image-Classification:-Image-Tagging/nus_wide/small_test.json', os.path.join(img_folder, 'small_test.json'))\n",
-    "copyfile('../PyTorch-Multi-Label-Image-Classification:-Image-Tagging/nus_wide/small_train.json', os.path.join(img_folder, 'small_train.json'))"
+    "# Also, copy our pre-processed annotations to the dataset folder.\n",
+    "copyfile('../PyTorch-Multi-Label-Image-Classification-Image-Tagging/nus_wide/small_test.json', os.path.join(img_folder, 'small_test.json'))\n",
+    "copyfile('../PyTorch-Multi-Label-Image-Classification-Image-Tagging/nus_wide/small_train.json', os.path.join(img_folder, 'small_train.json'))"
    ]
   },
   {
@@ -116,7 +116,6 @@
    },
    "outputs": [],
    "source": [
-    "# Download Glove model trained on wikipedia.\n",
     "# We want to represent our label names as vectors in order to use them as features further.\n",
     "# To do that we decided to use GloVe model (https://nlp.stanford.edu/projects/glove/).\n",
     "# Let's download GloVe model trained on a Wikipedia Text Corpus.\n",
@@ -162,13 +161,13 @@
     "               'nighttime', 'boats', 'mountain', 'tree', 'snow', 'beach', 'vehicle', 'rocks',\n",
     "               'reflection', 'sunset', 'road', 'flowers', 'ocean', 'lake', 'window', 'plants',\n",
     "               'buildings', 'grass', 'water', 'animal', 'person', 'clouds', 'sky']\n",
-    "vectorised_labels = [embeddings_dict[label].tolist() for label in small_labels]\n",
+    "vectorized_labels = [embeddings_dict[label].tolist() for label in small_labels]\n",
     "\n",
     "# Save them for further use.\n",
     "word_2_vec_path = 'word_2_vec_glow_classes.json'\n",
     "with open(word_2_vec_path, 'w') as fp:\n",
     "    json.dump({\n",
-    "        'vect_labels': vectorised_labels,\n",
+    "        'vect_labels': vectorized_labels,\n",
     "    }, fp, indent=3)\n"
    ]
   },
@@ -182,17 +181,9 @@
     "# It would be hard to visualize vectors with 300 values, but luckly we have t-SNE for that.\n",
     "# This function builds a t-SNE model(https://www.learnopencv.com/t-sne-for-feature-visualization/) \n",
     "# for label embeddings and visualizes them.\n",
-    "def tsne_plot(model):\n",
-    "    labels = []\n",
-    "    tokens = []\n",
-    "\n",
-    "    for word in model:\n",
-    "        tokens.append(model[word])\n",
-    "        labels.append(word)\n",
-    "    \n",
+    "def tsne_plot(tokens, labels):\n",
     "    tsne_model = TSNE(perplexity=2, n_components=2, init='pca', n_iter=25000, random_state=2020, n_jobs=4)\n",
     "    new_values = tsne_model.fit_transform(tokens)\n",
-    "\n",
     "    x = []\n",
     "    y = []\n",
     "    for value in new_values:\n",
@@ -201,7 +192,7 @@
     "        \n",
     "    plt.figure(figsize=(13, 13)) \n",
     "    for i in range(len(x)):\n",
-    "        plt.scatter(x[i],y[i])\n",
+    "        plt.scatter(x[i], y[i])\n",
     "        plt.annotate(labels[i],\n",
     "                     xy=(x[i], y[i]),\n",
     "                     xytext=(5, 2),\n",
@@ -218,8 +209,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# Now we can draw t-SNE visualisation.\n",
-    "tsne_plot(dict(zip(small_labels, vectorised_labels)))"
+    "# Now we can draw t-SNE visualization.\n",
+    "tsne_plot(vectorized_labels, small_labels)"
    ]
   },
   {
@@ -322,7 +313,8 @@
     "adj_matrix_path = 'adjacency_matrix.json'\n",
     "# Count all labels.\n",
     "nums = np.sum(np.array(dataset_train.annos), axis=0)\n",
-    "adj = np.zeros((len(small_labels), len(small_labels)), dtype=int)\n",
+    "label_len = len(small_labels)\n",
+    "adj = np.zeros((label_len, label_len), dtype=int)\n",
     "# Now iterate over the whole training set and consider all pairs of labels in sample annotation.\n",
     "for sample in dataset_train.annos:\n",
     "    sample_idx = np.argwhere(sample > 0)[:, 0]\n",
@@ -381,7 +373,7 @@
     "               + str(self.in_features) + ' -> ' \\\n",
     "               + str(self.out_features) + ')'\n",
     "\n",
-    "# Create adjacency matrix from probabilities.\n",
+    "# Create adjacency matrix from statistics.\n",
     "def gen_A(num_classes, t, p, adj_data):\n",
     "    adj = np.array(adj_data['adj']).astype(np.float32)\n",
     "    nums = np.array(adj_data['nums']).astype(np.float32)\n",
@@ -393,7 +385,7 @@
     "    adj = adj + np.identity(num_classes, np.int)\n",
     "    return adj\n",
     "\n",
-    "# Apply adjacency matrix renormalisation.\n",
+    "# Apply adjacency matrix re-normalization.\n",
     "def gen_adj(A):\n",
     "    D = torch.pow(A.sum(1).float(), -0.5)\n",
     "    D = torch.diag(D).type_as(A)\n",
@@ -413,7 +405,7 @@
     "        self.gc1 = GraphConvolution(in_channel, 1024)\n",
     "        self.gc2 = GraphConvolution(1024, 2048)\n",
     "        self.relu = nn.LeakyReLU(0.2)\n",
-    "        # Load probability data for adjacency matrix\n",
+    "        # Load data for adjacency matrix\n",
     "        with open(adj_path) as fp:\n",
     "            adj_data = json.load(fp)\n",
     "        # Compute adjacency matrix\n",
@@ -474,7 +466,7 @@
     "batch_size = 32\n",
     "save_freq = 1 # Save checkpoint frequency (epochs)\n",
     "test_freq = 200 # Test model frequency (iterations)\n",
-    "max_epoch_number = 50 # Number of epochs for training \n",
+    "max_epoch_number = 35 # Number of epochs for training \n",
     "# Note: on the small subset of data overfitting happens after 30-35 epochs.\n",
     "\n",
     "mean = [0.485, 0.456, 0.406]\n",
@@ -527,7 +519,7 @@
     "    transforms.ColorJitter(),\n",
     "    transforms.RandomAffine(degrees=20, translate=(0.2, 0.2), scale=(0.5, 1.5),\n",
     "                            shear=None, resample=False, \n",
-    "                            fillcolor=tuple(np.array(np.array(mean)*255).astype(int).tolist())),\n",
+    "                            fillcolor=tuple(np.array(np.array(mean) * 255).astype(int).tolist())),\n",
     "    transforms.ToTensor(),\n",
     "    transforms.Normalize(mean, std)\n",
     "])"
@@ -650,7 +642,7 @@
    "source": [
     "# Run inference on the test data.\n",
     "model.eval()\n",
-    "for sample_id in [1,2,3,4,6]:\n",
+    "for sample_id in [1, 2, 3, 4, 6]:\n",
     "    test_img, test_labels, gcn_input  = test_dataset[sample_id]\n",
     "    test_img_path = os.path.join(img_folder, test_dataset.imgs[sample_id])\n",
     "    with torch.no_grad():\n",
@@ -689,4 +681,4 @@
  },
  "nbformat": 4,
  "nbformat_minor": 4
-}
+}