[ { "path": ".gitignore", "content": "# Custom\n*.idea\n*.png\n*.pdf\ntmp/\n*.txt\n!requirements.txt\n\n# Byte-compiled / optimized / DLL files\n__pycache__/\n*.py[cod]\n*$py.class\n\n# C extensions\n*.so\n\n# Distribution / packaging\n.Python\nenv/\nbuild/\ndevelop-eggs/\ndist/\ndownloads/\neggs/\n.eggs/\nlib/\nlib64/\nparts/\nsdist/\nvar/\n*.egg-info/\n.installed.cfg\n*.egg\n\n# PyInstaller\n# Usually these files are written by a python script from a template\n# before PyInstaller builds the exe, so as to inject date/other infos into it.\n*.manifest\n*.spec\n\n# Installer logs\npip-log.txt\npip-delete-this-directory.txt\n\n# Unit test / coverage reports\nhtmlcov/\n.tox/\n.coverage\n.coverage.*\n.cache\nnosetests.xml\ncoverage.xml\n*,cover\n.hypothesis/\n\n# Translations\n*.mo\n*.pot\n\n# Django stuff:\n*.log\nlocal_settings.py\n\n# Flask stuff:\ninstance/\n.webassets-cache\n\n# Scrapy stuff:\n.scrapy\n\n# Sphinx documentation\ndocs/_build/\n\n# PyBuilder\ntarget/\n\n# IPython Notebook\n.ipynb_checkpoints\n\n# pyenv\n.python-version\n\n# celery beat schedule file\ncelerybeat-schedule\n\n# dotenv\n.env\n\n# virtualenv\nvenv/\nENV/\n\n# Spyder project settings\n.spyderproject\n\n# Rope project settings\n.ropeproject\n\n*.pickle\n" }, { "path": "LICENCE", "content": "The MIT License\n\nCopyright (c) 2016 Thomas Kipf\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\n\nThe above copyright notice and this permission notice shall be included in\nall copies or substantial portions of the Software.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN\nTHE SOFTWARE." }, { "path": "README.md", "content": "# Graph Convolutional Networks\n\nThis is a TensorFlow implementation of Graph Convolutional Networks for the task of (semi-supervised) classification of nodes in a graph, as described in our paper:\n \nThomas N. Kipf, Max Welling, [Semi-Supervised Classification with Graph Convolutional Networks](http://arxiv.org/abs/1609.02907) (ICLR 2017)\n\nFor a high-level explanation, have a look at our blog post:\n\nThomas Kipf, [Graph Convolutional Networks](http://tkipf.github.io/graph-convolutional-networks/) (2016)\n\n## Installation\n\n```bash\npython setup.py install\n```\n\n## Requirements\n* tensorflow (>0.12)\n* networkx\n\n## Run the demo\n\n```bash\ncd gcn\npython train.py\n```\n\n## Data\n\nIn order to use your own data, you have to provide \n* an N by N adjacency matrix (N is the number of nodes), \n* an N by D feature matrix (D is the number of features per node), and\n* an N by E binary label matrix (E is the number of classes).\n\nHave a look at the `load_data()` function in `utils.py` for an example.\n\nIn this example, we load citation network data (Cora, Citeseer or Pubmed). The original datasets can be found here: http://www.cs.umd.edu/~sen/lbc-proj/LBC.html. In our version (see `data` folder) we use dataset splits provided by https://github.com/kimiyoung/planetoid (Zhilin Yang, William W. Cohen, Ruslan Salakhutdinov, [Revisiting Semi-Supervised Learning with Graph Embeddings](https://arxiv.org/abs/1603.08861), ICML 2016). \n\nYou can specify a dataset as follows:\n\n```bash\npython train.py --dataset citeseer\n```\n\n(or by editing `train.py`)\n\n## Models\n\nYou can choose between the following models: \n* `gcn`: Graph convolutional network (Thomas N. Kipf, Max Welling, [Semi-Supervised Classification with Graph Convolutional Networks](http://arxiv.org/abs/1609.02907), 2016)\n* `gcn_cheby`: Chebyshev polynomial version of graph convolutional network as described in (Michaël Defferrard, Xavier Bresson, Pierre Vandergheynst, [Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering](https://arxiv.org/abs/1606.09375), NIPS 2016)\n* `dense`: Basic multi-layer perceptron that supports sparse inputs\n\n## Graph classification\n\nOur framework also supports batch-wise classification of multiple graph instances (of potentially different size) with an adjacency matrix each. It is best to concatenate respective feature matrices and build a (sparse) block-diagonal matrix where each block corresponds to the adjacency matrix of one graph instance. For pooling (in case of graph-level outputs as opposed to node-level outputs) it is best to specify a simple pooling matrix that collects features from their respective graph instances, as illustrated below:\n\n![graph_classification](https://user-images.githubusercontent.com/7347296/34198790-eb5bec96-e56b-11e7-90d5-157800e042de.png)\n\n\n## Cite\n\nPlease cite our paper if you use this code in your own work:\n\n```\n@inproceedings{kipf2017semi,\n title={Semi-Supervised Classification with Graph Convolutional Networks},\n author={Kipf, Thomas N. and Welling, Max},\n booktitle={International Conference on Learning Representations (ICLR)},\n year={2017}\n}\n```\n" }, { "path": "gcn/__init__.py", "content": "from __future__ import print_function\nfrom __future__ import division\n" }, { "path": "gcn/data/ind.citeseer.test.index", "content": "2488\n2644\n3261\n2804\n3176\n2432\n3310\n2410\n2812\n2520\n2994\n3282\n2680\n2848\n2670\n3005\n2977\n2592\n2967\n2461\n3184\n2852\n2768\n2905\n2851\n3129\n3164\n2438\n2793\n2763\n2528\n2954\n2347\n2640\n3265\n2874\n2446\n2856\n3149\n2374\n3097\n3301\n2664\n2418\n2655\n2464\n2596\n3262\n3278\n2320\n2612\n2614\n2550\n2626\n2772\n3007\n2733\n2516\n2476\n2798\n2561\n2839\n2685\n2391\n2705\n3098\n2754\n3251\n2767\n2630\n2727\n2513\n2701\n3264\n2792\n2821\n3260\n2462\n3307\n2639\n2900\n3060\n2672\n3116\n2731\n3316\n2386\n2425\n2518\n3151\n2586\n2797\n2479\n3117\n2580\n3182\n2459\n2508\n3052\n3230\n3215\n2803\n2969\n2562\n2398\n3325\n2343\n3030\n2414\n2776\n2383\n3173\n2850\n2499\n3312\n2648\n2784\n2898\n3056\n2484\n3179\n3132\n2577\n2563\n2867\n3317\n2355\n3207\n3178\n2968\n3319\n2358\n2764\n3001\n2683\n3271\n2321\n2567\n2502\n3246\n2715\n3066\n2390\n2381\n3162\n2741\n2498\n2790\n3038\n3321\n2481\n3050\n3161\n3122\n2801\n2957\n3177\n2965\n2621\n3208\n2921\n2802\n2357\n2677\n2519\n2860\n2696\n2368\n3241\n2858\n2419\n2762\n2875\n3222\n3064\n2827\n3044\n2471\n3062\n2982\n2736\n2322\n2709\n2766\n2424\n2602\n2970\n2675\n3299\n2554\n2964\n2597\n2753\n2979\n2523\n2912\n2896\n2317\n3167\n2813\n2482\n2557\n3043\n3244\n2985\n2460\n2363\n3272\n3045\n3192\n2453\n2656\n2834\n2443\n3202\n2926\n2711\n2633\n2384\n2752\n3285\n2817\n2483\n2919\n2924\n2661\n2698\n2361\n2662\n2819\n3143\n2316\n3196\n2739\n2345\n2578\n2822\n3229\n2908\n2917\n2692\n3200\n2324\n2522\n3322\n2697\n3163\n3093\n3233\n2774\n2371\n2835\n2652\n2539\n2843\n3231\n2976\n2429\n2367\n3144\n2564\n3283\n3217\n3035\n2962\n2433\n2415\n2387\n3021\n2595\n2517\n2468\n3061\n2673\n2348\n3027\n2467\n3318\n2959\n3273\n2392\n2779\n2678\n3004\n2634\n2974\n3198\n2342\n2376\n3249\n2868\n2952\n2710\n2838\n2335\n2524\n2650\n3186\n2743\n2545\n2841\n2515\n2505\n3181\n2945\n2738\n2933\n3303\n2611\n3090\n2328\n3010\n3016\n2504\n2936\n3266\n3253\n2840\n3034\n2581\n2344\n2452\n2654\n3199\n3137\n2514\n2394\n2544\n2641\n2613\n2618\n2558\n2593\n2532\n2512\n2975\n3267\n2566\n2951\n3300\n2869\n2629\n2747\n3055\n2831\n3105\n3168\n3100\n2431\n2828\n2684\n3269\n2910\n2865\n2693\n2884\n3228\n2783\n3247\n2770\n3157\n2421\n2382\n2331\n3203\n3240\n2351\n3114\n2986\n2688\n2439\n2996\n3079\n3103\n3296\n2349\n2372\n3096\n2422\n2551\n3069\n2737\n3084\n3304\n3022\n2542\n3204\n2949\n2318\n2450\n3140\n2734\n2881\n2576\n3054\n3089\n3125\n2761\n3136\n3111\n2427\n2466\n3101\n3104\n3259\n2534\n2961\n3191\n3000\n3036\n2356\n2800\n3155\n3224\n2646\n2735\n3020\n2866\n2426\n2448\n3226\n3219\n2749\n3183\n2906\n2360\n2440\n2946\n2313\n2859\n2340\n3008\n2719\n3058\n2653\n3023\n2888\n3243\n2913\n3242\n3067\n2409\n3227\n2380\n2353\n2686\n2971\n2847\n2947\n2857\n3263\n3218\n2861\n3323\n2635\n2966\n2604\n2456\n2832\n2694\n3245\n3119\n2942\n3153\n2894\n2555\n3128\n2703\n2323\n2631\n2732\n2699\n2314\n2590\n3127\n2891\n2873\n2814\n2326\n3026\n3288\n3095\n2706\n2457\n2377\n2620\n2526\n2674\n3190\n2923\n3032\n2334\n3254\n2991\n3277\n2973\n2599\n2658\n2636\n2826\n3148\n2958\n3258\n2990\n3180\n2538\n2748\n2625\n2565\n3011\n3057\n2354\n3158\n2622\n3308\n2983\n2560\n3169\n3059\n2480\n3194\n3291\n3216\n2643\n3172\n2352\n2724\n2485\n2411\n2948\n2445\n2362\n2668\n3275\n3107\n2496\n2529\n2700\n2541\n3028\n2879\n2660\n3324\n2755\n2436\n3048\n2623\n2920\n3040\n2568\n3221\n3003\n3295\n2473\n3232\n3213\n2823\n2897\n2573\n2645\n3018\n3326\n2795\n2915\n3109\n3086\n2463\n3118\n2671\n2909\n2393\n2325\n3029\n2972\n3110\n2870\n3284\n2816\n2647\n2667\n2955\n2333\n2960\n2864\n2893\n2458\n2441\n2359\n2327\n3256\n3099\n3073\n3138\n2511\n2666\n2548\n2364\n2451\n2911\n3237\n3206\n3080\n3279\n2934\n2981\n2878\n3130\n2830\n3091\n2659\n2449\n3152\n2413\n2722\n2796\n3220\n2751\n2935\n3238\n2491\n2730\n2842\n3223\n2492\n3074\n3094\n2833\n2521\n2883\n3315\n2845\n2907\n3083\n2572\n3092\n2903\n2918\n3039\n3286\n2587\n3068\n2338\n3166\n3134\n2455\n2497\n2992\n2775\n2681\n2430\n2932\n2931\n2434\n3154\n3046\n2598\n2366\n3015\n3147\n2944\n2582\n3274\n2987\n2642\n2547\n2420\n2930\n2750\n2417\n2808\n3141\n2997\n2995\n2584\n2312\n3033\n3070\n3065\n2509\n3314\n2396\n2543\n2423\n3170\n2389\n3289\n2728\n2540\n2437\n2486\n2895\n3017\n2853\n2406\n2346\n2877\n2472\n3210\n2637\n2927\n2789\n2330\n3088\n3102\n2616\n3081\n2902\n3205\n3320\n3165\n2984\n3185\n2707\n3255\n2583\n2773\n2742\n3024\n2402\n2718\n2882\n2575\n3281\n2786\n2855\n3014\n2401\n2535\n2687\n2495\n3113\n2609\n2559\n2665\n2530\n3293\n2399\n2605\n2690\n3133\n2799\n2533\n2695\n2713\n2886\n2691\n2549\n3077\n3002\n3049\n3051\n3087\n2444\n3085\n3135\n2702\n3211\n3108\n2501\n2769\n3290\n2465\n3025\n3019\n2385\n2940\n2657\n2610\n2525\n2941\n3078\n2341\n2916\n2956\n2375\n2880\n3009\n2780\n2370\n2925\n2332\n3146\n2315\n2809\n3145\n3106\n2782\n2760\n2493\n2765\n2556\n2890\n2400\n2339\n3201\n2818\n3248\n3280\n2570\n2569\n2937\n3174\n2836\n2708\n2820\n3195\n2617\n3197\n2319\n2744\n2615\n2825\n2603\n2914\n2531\n3193\n2624\n2365\n2810\n3239\n3159\n2537\n2844\n2758\n2938\n3037\n2503\n3297\n2885\n2608\n2494\n2712\n2408\n2901\n2704\n2536\n2373\n2478\n2723\n3076\n2627\n2369\n2669\n3006\n2628\n2788\n3276\n2435\n3139\n3235\n2527\n2571\n2815\n2442\n2892\n2978\n2746\n3150\n2574\n2725\n3188\n2601\n2378\n3075\n2632\n2794\n3270\n3071\n2506\n3126\n3236\n3257\n2824\n2989\n2950\n2428\n2405\n3156\n2447\n2787\n2805\n2720\n2403\n2811\n2329\n2474\n2785\n2350\n2507\n2416\n3112\n2475\n2876\n2585\n2487\n3072\n3082\n2943\n2757\n2388\n2600\n3294\n2756\n3142\n3041\n2594\n2998\n3047\n2379\n2980\n2454\n2862\n3175\n2588\n3031\n3012\n2889\n2500\n2791\n2854\n2619\n2395\n2807\n2740\n2412\n3131\n3013\n2939\n2651\n2490\n2988\n2863\n3225\n2745\n2714\n3160\n3124\n2849\n2676\n2872\n3287\n3189\n2716\n3115\n2928\n2871\n2591\n2717\n2546\n2777\n3298\n2397\n3187\n2726\n2336\n3268\n2477\n2904\n2846\n3121\n2899\n2510\n2806\n2963\n3313\n2679\n3302\n2663\n3053\n2469\n2999\n3311\n2470\n2638\n3120\n3171\n2689\n2922\n2607\n2721\n2993\n2887\n2837\n2929\n2829\n3234\n2649\n2337\n2759\n2778\n2771\n2404\n2589\n3123\n3209\n2729\n3252\n2606\n2579\n2552\n" }, { "path": "gcn/data/ind.cora.test.index", "content": "2692\n2532\n2050\n1715\n2362\n2609\n2622\n1975\n2081\n1767\n2263\n1725\n2588\n2259\n2357\n1998\n2574\n2179\n2291\n2382\n1812\n1751\n2422\n1937\n2631\n2510\n2378\n2589\n2345\n1943\n1850\n2298\n1825\n2035\n2507\n2313\n1906\n1797\n2023\n2159\n2495\n1886\n2122\n2369\n2461\n1925\n2565\n1858\n2234\n2000\n1846\n2318\n1723\n2559\n2258\n1763\n1991\n1922\n2003\n2662\n2250\n2064\n2529\n1888\n2499\n2454\n2320\n2287\n2203\n2018\n2002\n2632\n2554\n2314\n2537\n1760\n2088\n2086\n2218\n2605\n1953\n2403\n1920\n2015\n2335\n2535\n1837\n2009\n1905\n2636\n1942\n2193\n2576\n2373\n1873\n2463\n2509\n1954\n2656\n2455\n2494\n2295\n2114\n2561\n2176\n2275\n2635\n2442\n2704\n2127\n2085\n2214\n2487\n1739\n2543\n1783\n2485\n2262\n2472\n2326\n1738\n2170\n2100\n2384\n2152\n2647\n2693\n2376\n1775\n1726\n2476\n2195\n1773\n1793\n2194\n2581\n1854\n2524\n1945\n1781\n1987\n2599\n1744\n2225\n2300\n1928\n2042\n2202\n1958\n1816\n1916\n2679\n2190\n1733\n2034\n2643\n2177\n1883\n1917\n1996\n2491\n2268\n2231\n2471\n1919\n1909\n2012\n2522\n1865\n2466\n2469\n2087\n2584\n2563\n1924\n2143\n1736\n1966\n2533\n2490\n2630\n1973\n2568\n1978\n2664\n2633\n2312\n2178\n1754\n2307\n2480\n1960\n1742\n1962\n2160\n2070\n2553\n2433\n1768\n2659\n2379\n2271\n1776\n2153\n1877\n2027\n2028\n2155\n2196\n2483\n2026\n2158\n2407\n1821\n2131\n2676\n2277\n2489\n2424\n1963\n1808\n1859\n2597\n2548\n2368\n1817\n2405\n2413\n2603\n2350\n2118\n2329\n1969\n2577\n2475\n2467\n2425\n1769\n2092\n2044\n2586\n2608\n1983\n2109\n2649\n1964\n2144\n1902\n2411\n2508\n2360\n1721\n2005\n2014\n2308\n2646\n1949\n1830\n2212\n2596\n1832\n1735\n1866\n2695\n1941\n2546\n2498\n2686\n2665\n1784\n2613\n1970\n2021\n2211\n2516\n2185\n2479\n2699\n2150\n1990\n2063\n2075\n1979\n2094\n1787\n2571\n2690\n1926\n2341\n2566\n1957\n1709\n1955\n2570\n2387\n1811\n2025\n2447\n2696\n2052\n2366\n1857\n2273\n2245\n2672\n2133\n2421\n1929\n2125\n2319\n2641\n2167\n2418\n1765\n1761\n1828\n2188\n1972\n1997\n2419\n2289\n2296\n2587\n2051\n2440\n2053\n2191\n1923\n2164\n1861\n2339\n2333\n2523\n2670\n2121\n1921\n1724\n2253\n2374\n1940\n2545\n2301\n2244\n2156\n1849\n2551\n2011\n2279\n2572\n1757\n2400\n2569\n2072\n2526\n2173\n2069\n2036\n1819\n1734\n1880\n2137\n2408\n2226\n2604\n1771\n2698\n2187\n2060\n1756\n2201\n2066\n2439\n1844\n1772\n2383\n2398\n1708\n1992\n1959\n1794\n2426\n2702\n2444\n1944\n1829\n2660\n2497\n2607\n2343\n1730\n2624\n1790\n1935\n1967\n2401\n2255\n2355\n2348\n1931\n2183\n2161\n2701\n1948\n2501\n2192\n2404\n2209\n2331\n1810\n2363\n2334\n1887\n2393\n2557\n1719\n1732\n1986\n2037\n2056\n1867\n2126\n1932\n2117\n1807\n1801\n1743\n2041\n1843\n2388\n2221\n1833\n2677\n1778\n2661\n2306\n2394\n2106\n2430\n2371\n2606\n2353\n2269\n2317\n2645\n2372\n2550\n2043\n1968\n2165\n2310\n1985\n2446\n1982\n2377\n2207\n1818\n1913\n1766\n1722\n1894\n2020\n1881\n2621\n2409\n2261\n2458\n2096\n1712\n2594\n2293\n2048\n2359\n1839\n2392\n2254\n1911\n2101\n2367\n1889\n1753\n2555\n2246\n2264\n2010\n2336\n2651\n2017\n2140\n1842\n2019\n1890\n2525\n2134\n2492\n2652\n2040\n2145\n2575\n2166\n1999\n2434\n1711\n2276\n2450\n2389\n2669\n2595\n1814\n2039\n2502\n1896\n2168\n2344\n2637\n2031\n1977\n2380\n1936\n2047\n2460\n2102\n1745\n2650\n2046\n2514\n1980\n2352\n2113\n1713\n2058\n2558\n1718\n1864\n1876\n2338\n1879\n1891\n2186\n2451\n2181\n2638\n2644\n2103\n2591\n2266\n2468\n1869\n2582\n2674\n2361\n2462\n1748\n2215\n2615\n2236\n2248\n2493\n2342\n2449\n2274\n1824\n1852\n1870\n2441\n2356\n1835\n2694\n2602\n2685\n1893\n2544\n2536\n1994\n1853\n1838\n1786\n1930\n2539\n1892\n2265\n2618\n2486\n2583\n2061\n1796\n1806\n2084\n1933\n2095\n2136\n2078\n1884\n2438\n2286\n2138\n1750\n2184\n1799\n2278\n2410\n2642\n2435\n1956\n2399\n1774\n2129\n1898\n1823\n1938\n2299\n1862\n2420\n2673\n1984\n2204\n1717\n2074\n2213\n2436\n2297\n2592\n2667\n2703\n2511\n1779\n1782\n2625\n2365\n2315\n2381\n1788\n1714\n2302\n1927\n2325\n2506\n2169\n2328\n2629\n2128\n2655\n2282\n2073\n2395\n2247\n2521\n2260\n1868\n1988\n2324\n2705\n2541\n1731\n2681\n2707\n2465\n1785\n2149\n2045\n2505\n2611\n2217\n2180\n1904\n2453\n2484\n1871\n2309\n2349\n2482\n2004\n1965\n2406\n2162\n1805\n2654\n2007\n1947\n1981\n2112\n2141\n1720\n1758\n2080\n2330\n2030\n2432\n2089\n2547\n1820\n1815\n2675\n1840\n2658\n2370\n2251\n1908\n2029\n2068\n2513\n2549\n2267\n2580\n2327\n2351\n2111\n2022\n2321\n2614\n2252\n2104\n1822\n2552\n2243\n1798\n2396\n2663\n2564\n2148\n2562\n2684\n2001\n2151\n2706\n2240\n2474\n2303\n2634\n2680\n2055\n2090\n2503\n2347\n2402\n2238\n1950\n2054\n2016\n1872\n2233\n1710\n2032\n2540\n2628\n1795\n2616\n1903\n2531\n2567\n1946\n1897\n2222\n2227\n2627\n1856\n2464\n2241\n2481\n2130\n2311\n2083\n2223\n2284\n2235\n2097\n1752\n2515\n2527\n2385\n2189\n2283\n2182\n2079\n2375\n2174\n2437\n1993\n2517\n2443\n2224\n2648\n2171\n2290\n2542\n2038\n1855\n1831\n1759\n1848\n2445\n1827\n2429\n2205\n2598\n2657\n1728\n2065\n1918\n2427\n2573\n2620\n2292\n1777\n2008\n1875\n2288\n2256\n2033\n2470\n2585\n2610\n2082\n2230\n1915\n1847\n2337\n2512\n2386\n2006\n2653\n2346\n1951\n2110\n2639\n2520\n1939\n2683\n2139\n2220\n1910\n2237\n1900\n1836\n2197\n1716\n1860\n2077\n2519\n2538\n2323\n1914\n1971\n1845\n2132\n1802\n1907\n2640\n2496\n2281\n2198\n2416\n2285\n1755\n2431\n2071\n2249\n2123\n1727\n2459\n2304\n2199\n1791\n1809\n1780\n2210\n2417\n1874\n1878\n2116\n1961\n1863\n2579\n2477\n2228\n2332\n2578\n2457\n2024\n1934\n2316\n1841\n1764\n1737\n2322\n2239\n2294\n1729\n2488\n1974\n2473\n2098\n2612\n1834\n2340\n2423\n2175\n2280\n2617\n2208\n2560\n1741\n2600\n2059\n1747\n2242\n2700\n2232\n2057\n2147\n2682\n1792\n1826\n2120\n1895\n2364\n2163\n1851\n2391\n2414\n2452\n1803\n1989\n2623\n2200\n2528\n2415\n1804\n2146\n2619\n2687\n1762\n2172\n2270\n2678\n2593\n2448\n1882\n2257\n2500\n1899\n2478\n2412\n2107\n1746\n2428\n2115\n1800\n1901\n2397\n2530\n1912\n2108\n2206\n2091\n1740\n2219\n1976\n2099\n2142\n2671\n2668\n2216\n2272\n2229\n2666\n2456\n2534\n2697\n2688\n2062\n2691\n2689\n2154\n2590\n2626\n2390\n1813\n2067\n1952\n2518\n2358\n1789\n2076\n2049\n2119\n2013\n2124\n2556\n2105\n2093\n1885\n2305\n2354\n2135\n2601\n1770\n1995\n2504\n1749\n2157\n" }, { "path": "gcn/data/ind.pubmed.test.index", "content": "18747\n19392\n19181\n18843\n19221\n18962\n19560\n19097\n18966\n19014\n18756\n19313\n19000\n19569\n19359\n18854\n18970\n19073\n19661\n19180\n19377\n18750\n19401\n18788\n19224\n19447\n19017\n19241\n18890\n18908\n18965\n19001\n18849\n19641\n18852\n19222\n19172\n18762\n19156\n19162\n18856\n18763\n19318\n18826\n19712\n19192\n19695\n19030\n19523\n19249\n19079\n19232\n19455\n18743\n18800\n19071\n18885\n19593\n19394\n19390\n18832\n19445\n18838\n19632\n19548\n19546\n18825\n19498\n19266\n19117\n19595\n19252\n18730\n18913\n18809\n19452\n19520\n19274\n19555\n19388\n18919\n19099\n19637\n19403\n18720\n19526\n18905\n19451\n19408\n18923\n18794\n19322\n19431\n18912\n18841\n19239\n19125\n19258\n19565\n18898\n19482\n19029\n18778\n19096\n19684\n19552\n18765\n19361\n19171\n19367\n19623\n19402\n19327\n19118\n18888\n18726\n19510\n18831\n19490\n19576\n19050\n18729\n18896\n19246\n19012\n18862\n18873\n19193\n19693\n19474\n18953\n19115\n19182\n19269\n19116\n18837\n18872\n19007\n19212\n18798\n19102\n18772\n19660\n19511\n18914\n18886\n19672\n19360\n19213\n18810\n19420\n19512\n18719\n19432\n19350\n19127\n18782\n19587\n18924\n19488\n18781\n19340\n19190\n19383\n19094\n18835\n19487\n19230\n18791\n18882\n18937\n18928\n18755\n18802\n19516\n18795\n18786\n19273\n19349\n19398\n19626\n19130\n19351\n19489\n19446\n18959\n19025\n18792\n18878\n19304\n19629\n19061\n18785\n19194\n19179\n19210\n19417\n19583\n19415\n19443\n18739\n19662\n18904\n18910\n18901\n18960\n18722\n18827\n19290\n18842\n19389\n19344\n18961\n19098\n19147\n19334\n19358\n18829\n18984\n18931\n18742\n19320\n19111\n19196\n18887\n18991\n19469\n18990\n18876\n19261\n19270\n19522\n19088\n19284\n19646\n19493\n19225\n19615\n19449\n19043\n19674\n19391\n18918\n19155\n19110\n18815\n19131\n18834\n19715\n19603\n19688\n19133\n19053\n19166\n19066\n18893\n18757\n19582\n19282\n19257\n18869\n19467\n18954\n19371\n19151\n19462\n19598\n19653\n19187\n19624\n19564\n19534\n19581\n19478\n18985\n18746\n19342\n18777\n19696\n18824\n19138\n18728\n19643\n19199\n18731\n19168\n18948\n19216\n19697\n19347\n18808\n18725\n19134\n18847\n18828\n18996\n19106\n19485\n18917\n18911\n18776\n19203\n19158\n18895\n19165\n19382\n18780\n18836\n19373\n19659\n18947\n19375\n19299\n18761\n19366\n18754\n19248\n19416\n19658\n19638\n19034\n19281\n18844\n18922\n19491\n19272\n19341\n19068\n19332\n19559\n19293\n18804\n18933\n18935\n19405\n18936\n18945\n18943\n18818\n18797\n19570\n19464\n19428\n19093\n19433\n18986\n19161\n19255\n19157\n19046\n19292\n19434\n19298\n18724\n19410\n19694\n19214\n19640\n19189\n18963\n19218\n19585\n19041\n19550\n19123\n19620\n19376\n19561\n18944\n19706\n19056\n19283\n18741\n19319\n19144\n19542\n18821\n19404\n19080\n19303\n18793\n19306\n19678\n19435\n19519\n19566\n19278\n18946\n19536\n19020\n19057\n19198\n19333\n19649\n19699\n19399\n19654\n19136\n19465\n19321\n19577\n18907\n19665\n19386\n19596\n19247\n19473\n19568\n19355\n18925\n19586\n18982\n19616\n19495\n19612\n19023\n19438\n18817\n19692\n19295\n19414\n19676\n19472\n19107\n19062\n19035\n18883\n19409\n19052\n19606\n19091\n19651\n19475\n19413\n18796\n19369\n19639\n19701\n19461\n19645\n19251\n19063\n19679\n19545\n19081\n19363\n18995\n19549\n18790\n18855\n18833\n18899\n19395\n18717\n19647\n18768\n19103\n19245\n18819\n18779\n19656\n19076\n18745\n18971\n19197\n19711\n19074\n19128\n19466\n19139\n19309\n19324\n18814\n19092\n19627\n19060\n18806\n18929\n18737\n18942\n18906\n18858\n19456\n19253\n19716\n19104\n19667\n19574\n18903\n19237\n18864\n19556\n19364\n18952\n19008\n19323\n19700\n19170\n19267\n19345\n19238\n18909\n18892\n19109\n19704\n18902\n19275\n19680\n18723\n19242\n19112\n19169\n18956\n19343\n19650\n19541\n19698\n19521\n19087\n18976\n19038\n18775\n18968\n19671\n19412\n19407\n19573\n19027\n18813\n19357\n19460\n19673\n19481\n19036\n19614\n18787\n19195\n18732\n18884\n19613\n19657\n19575\n19226\n19589\n19234\n19617\n19707\n19484\n18740\n19424\n18784\n19419\n19159\n18865\n19105\n19315\n19480\n19664\n19378\n18803\n19605\n18870\n19042\n19426\n18848\n19223\n19509\n19532\n18752\n19691\n18718\n19209\n19362\n19090\n19492\n19567\n19687\n19018\n18830\n19530\n19554\n19119\n19442\n19558\n19527\n19427\n19291\n19543\n19422\n19142\n18897\n18950\n19425\n19002\n19588\n18978\n19551\n18930\n18736\n19101\n19215\n19150\n19263\n18949\n18974\n18759\n19335\n19200\n19129\n19328\n19437\n18988\n19429\n19368\n19406\n19049\n18811\n19296\n19256\n19385\n19602\n18770\n19337\n19580\n19476\n19045\n19132\n19089\n19120\n19265\n19483\n18767\n19227\n18934\n19069\n18820\n19006\n19459\n18927\n19037\n19280\n19441\n18823\n19015\n19114\n19618\n18957\n19176\n18853\n19648\n19201\n19444\n19279\n18751\n19302\n19505\n18733\n19601\n19533\n18863\n19708\n19387\n19346\n19152\n19206\n18851\n19338\n19681\n19380\n19055\n18766\n19085\n19591\n19547\n18958\n19146\n18840\n19051\n19021\n19207\n19235\n19086\n18979\n19300\n18939\n19100\n19619\n19287\n18980\n19277\n19326\n19108\n18920\n19625\n19374\n19078\n18734\n19634\n19339\n18877\n19423\n19652\n19683\n19044\n18983\n19330\n19529\n19714\n19468\n19075\n19540\n18839\n19022\n19286\n19537\n19175\n19463\n19167\n19705\n19562\n19244\n19486\n19611\n18801\n19178\n19590\n18846\n19450\n19205\n19381\n18941\n19670\n19185\n19504\n19633\n18997\n19113\n19397\n19636\n19709\n19289\n19264\n19353\n19584\n19126\n18938\n19669\n18964\n19276\n18774\n19173\n19231\n18973\n18769\n19064\n19040\n19668\n18738\n19082\n19655\n19236\n19352\n19609\n19628\n18951\n19384\n19122\n18875\n18992\n18753\n19379\n19254\n19301\n19506\n19135\n19010\n19682\n19400\n19579\n19316\n19553\n19208\n19635\n19644\n18891\n19024\n18989\n19250\n18850\n19317\n18915\n19607\n18799\n18881\n19479\n19031\n19365\n19164\n18744\n18760\n19502\n19058\n19517\n18735\n19448\n19243\n19453\n19285\n18857\n19439\n19016\n18975\n19503\n18998\n18981\n19186\n18994\n19240\n19631\n19070\n19174\n18900\n19065\n19220\n19229\n18880\n19308\n19372\n19496\n18771\n19325\n19538\n19033\n18874\n19077\n19211\n18764\n19458\n19571\n19121\n19019\n19059\n19497\n18969\n19666\n19297\n19219\n19622\n19184\n18977\n19702\n19539\n19329\n19095\n19675\n18972\n19514\n19703\n19188\n18866\n18812\n19314\n18822\n18845\n19494\n19411\n18916\n19686\n18967\n19294\n19143\n19204\n18805\n19689\n19233\n18758\n18748\n19011\n19685\n19336\n19608\n19454\n19124\n18868\n18807\n19544\n19621\n19228\n19154\n19141\n19145\n19153\n18860\n19163\n19393\n19268\n19160\n19305\n19259\n19471\n19524\n18783\n19396\n18894\n19430\n19690\n19348\n19597\n19592\n19677\n18889\n19331\n18773\n19137\n19009\n18932\n19599\n18816\n19054\n19067\n19477\n19191\n18921\n18940\n19578\n19183\n19004\n19072\n19710\n19005\n19610\n18955\n19457\n19148\n18859\n18993\n19642\n19047\n19418\n19535\n19600\n19312\n19039\n19028\n18879\n19003\n19026\n19013\n19149\n19177\n19217\n18987\n19354\n19525\n19202\n19084\n19032\n18749\n18867\n19048\n18999\n19260\n19630\n18727\n19356\n19083\n18926\n18789\n19370\n18861\n19311\n19557\n19531\n19436\n19140\n19310\n19501\n18721\n19604\n19713\n19262\n19563\n19507\n19440\n19572\n19513\n19515\n19518\n19421\n19470\n19499\n19663\n19508\n18871\n19528\n19500\n19307\n19288\n19594\n19271\n" }, { "path": "gcn/inits.py", "content": "import tensorflow as tf\nimport numpy as np\n\n\ndef uniform(shape, scale=0.05, name=None):\n \"\"\"Uniform init.\"\"\"\n initial = tf.random_uniform(shape, minval=-scale, maxval=scale, dtype=tf.float32)\n return tf.Variable(initial, name=name)\n\n\ndef glorot(shape, name=None):\n \"\"\"Glorot & Bengio (AISTATS 2010) init.\"\"\"\n init_range = np.sqrt(6.0/(shape[0]+shape[1]))\n initial = tf.random_uniform(shape, minval=-init_range, maxval=init_range, dtype=tf.float32)\n return tf.Variable(initial, name=name)\n\n\ndef zeros(shape, name=None):\n \"\"\"All zeros.\"\"\"\n initial = tf.zeros(shape, dtype=tf.float32)\n return tf.Variable(initial, name=name)\n\n\ndef ones(shape, name=None):\n \"\"\"All ones.\"\"\"\n initial = tf.ones(shape, dtype=tf.float32)\n return tf.Variable(initial, name=name)" }, { "path": "gcn/layers.py", "content": "from gcn.inits import *\nimport tensorflow as tf\n\nflags = tf.app.flags\nFLAGS = flags.FLAGS\n\n# global unique layer ID dictionary for layer name assignment\n_LAYER_UIDS = {}\n\n\ndef get_layer_uid(layer_name=''):\n \"\"\"Helper function, assigns unique layer IDs.\"\"\"\n if layer_name not in _LAYER_UIDS:\n _LAYER_UIDS[layer_name] = 1\n return 1\n else:\n _LAYER_UIDS[layer_name] += 1\n return _LAYER_UIDS[layer_name]\n\n\ndef sparse_dropout(x, keep_prob, noise_shape):\n \"\"\"Dropout for sparse tensors.\"\"\"\n random_tensor = keep_prob\n random_tensor += tf.random_uniform(noise_shape)\n dropout_mask = tf.cast(tf.floor(random_tensor), dtype=tf.bool)\n pre_out = tf.sparse_retain(x, dropout_mask)\n return pre_out * (1./keep_prob)\n\n\ndef dot(x, y, sparse=False):\n \"\"\"Wrapper for tf.matmul (sparse vs dense).\"\"\"\n if sparse:\n res = tf.sparse_tensor_dense_matmul(x, y)\n else:\n res = tf.matmul(x, y)\n return res\n\n\nclass Layer(object):\n \"\"\"Base layer class. Defines basic API for all layer objects.\n Implementation inspired by keras (http://keras.io).\n\n # Properties\n name: String, defines the variable scope of the layer.\n logging: Boolean, switches Tensorflow histogram logging on/off\n\n # Methods\n _call(inputs): Defines computation graph of layer\n (i.e. takes input, returns output)\n __call__(inputs): Wrapper for _call()\n _log_vars(): Log all variables\n \"\"\"\n\n def __init__(self, **kwargs):\n allowed_kwargs = {'name', 'logging'}\n for kwarg in kwargs.keys():\n assert kwarg in allowed_kwargs, 'Invalid keyword argument: ' + kwarg\n name = kwargs.get('name')\n if not name:\n layer = self.__class__.__name__.lower()\n name = layer + '_' + str(get_layer_uid(layer))\n self.name = name\n self.vars = {}\n logging = kwargs.get('logging', False)\n self.logging = logging\n self.sparse_inputs = False\n\n def _call(self, inputs):\n return inputs\n\n def __call__(self, inputs):\n with tf.name_scope(self.name):\n if self.logging and not self.sparse_inputs:\n tf.summary.histogram(self.name + '/inputs', inputs)\n outputs = self._call(inputs)\n if self.logging:\n tf.summary.histogram(self.name + '/outputs', outputs)\n return outputs\n\n def _log_vars(self):\n for var in self.vars:\n tf.summary.histogram(self.name + '/vars/' + var, self.vars[var])\n\n\nclass Dense(Layer):\n \"\"\"Dense layer.\"\"\"\n def __init__(self, input_dim, output_dim, placeholders, dropout=0., sparse_inputs=False,\n act=tf.nn.relu, bias=False, featureless=False, **kwargs):\n super(Dense, self).__init__(**kwargs)\n\n if dropout:\n self.dropout = placeholders['dropout']\n else:\n self.dropout = 0.\n\n self.act = act\n self.sparse_inputs = sparse_inputs\n self.featureless = featureless\n self.bias = bias\n\n # helper variable for sparse dropout\n self.num_features_nonzero = placeholders['num_features_nonzero']\n\n with tf.variable_scope(self.name + '_vars'):\n self.vars['weights'] = glorot([input_dim, output_dim],\n name='weights')\n if self.bias:\n self.vars['bias'] = zeros([output_dim], name='bias')\n\n if self.logging:\n self._log_vars()\n\n def _call(self, inputs):\n x = inputs\n\n # dropout\n if self.sparse_inputs:\n x = sparse_dropout(x, 1-self.dropout, self.num_features_nonzero)\n else:\n x = tf.nn.dropout(x, 1-self.dropout)\n\n # transform\n output = dot(x, self.vars['weights'], sparse=self.sparse_inputs)\n\n # bias\n if self.bias:\n output += self.vars['bias']\n\n return self.act(output)\n\n\nclass GraphConvolution(Layer):\n \"\"\"Graph convolution layer.\"\"\"\n def __init__(self, input_dim, output_dim, placeholders, dropout=0.,\n sparse_inputs=False, act=tf.nn.relu, bias=False,\n featureless=False, **kwargs):\n super(GraphConvolution, self).__init__(**kwargs)\n\n if dropout:\n self.dropout = placeholders['dropout']\n else:\n self.dropout = 0.\n\n self.act = act\n self.support = placeholders['support']\n self.sparse_inputs = sparse_inputs\n self.featureless = featureless\n self.bias = bias\n\n # helper variable for sparse dropout\n self.num_features_nonzero = placeholders['num_features_nonzero']\n\n with tf.variable_scope(self.name + '_vars'):\n for i in range(len(self.support)):\n self.vars['weights_' + str(i)] = glorot([input_dim, output_dim],\n name='weights_' + str(i))\n if self.bias:\n self.vars['bias'] = zeros([output_dim], name='bias')\n\n if self.logging:\n self._log_vars()\n\n def _call(self, inputs):\n x = inputs\n\n # dropout\n if self.sparse_inputs:\n x = sparse_dropout(x, 1-self.dropout, self.num_features_nonzero)\n else:\n x = tf.nn.dropout(x, 1-self.dropout)\n\n # convolve\n supports = list()\n for i in range(len(self.support)):\n if not self.featureless:\n pre_sup = dot(x, self.vars['weights_' + str(i)],\n sparse=self.sparse_inputs)\n else:\n pre_sup = self.vars['weights_' + str(i)]\n support = dot(self.support[i], pre_sup, sparse=True)\n supports.append(support)\n output = tf.add_n(supports)\n\n # bias\n if self.bias:\n output += self.vars['bias']\n\n return self.act(output)\n" }, { "path": "gcn/metrics.py", "content": "import tensorflow as tf\n\n\ndef masked_softmax_cross_entropy(preds, labels, mask):\n \"\"\"Softmax cross-entropy loss with masking.\"\"\"\n loss = tf.nn.softmax_cross_entropy_with_logits(logits=preds, labels=labels)\n mask = tf.cast(mask, dtype=tf.float32)\n mask /= tf.reduce_mean(mask)\n loss *= mask\n return tf.reduce_mean(loss)\n\n\ndef masked_accuracy(preds, labels, mask):\n \"\"\"Accuracy with masking.\"\"\"\n correct_prediction = tf.equal(tf.argmax(preds, 1), tf.argmax(labels, 1))\n accuracy_all = tf.cast(correct_prediction, tf.float32)\n mask = tf.cast(mask, dtype=tf.float32)\n mask /= tf.reduce_mean(mask)\n accuracy_all *= mask\n return tf.reduce_mean(accuracy_all)\n" }, { "path": "gcn/models.py", "content": "from gcn.layers import *\nfrom gcn.metrics import *\n\nflags = tf.app.flags\nFLAGS = flags.FLAGS\n\n\nclass Model(object):\n def __init__(self, **kwargs):\n allowed_kwargs = {'name', 'logging'}\n for kwarg in kwargs.keys():\n assert kwarg in allowed_kwargs, 'Invalid keyword argument: ' + kwarg\n name = kwargs.get('name')\n if not name:\n name = self.__class__.__name__.lower()\n self.name = name\n\n logging = kwargs.get('logging', False)\n self.logging = logging\n\n self.vars = {}\n self.placeholders = {}\n\n self.layers = []\n self.activations = []\n\n self.inputs = None\n self.outputs = None\n\n self.loss = 0\n self.accuracy = 0\n self.optimizer = None\n self.opt_op = None\n\n def _build(self):\n raise NotImplementedError\n\n def build(self):\n \"\"\" Wrapper for _build() \"\"\"\n with tf.variable_scope(self.name):\n self._build()\n\n # Build sequential layer model\n self.activations.append(self.inputs)\n for layer in self.layers:\n hidden = layer(self.activations[-1])\n self.activations.append(hidden)\n self.outputs = self.activations[-1]\n\n # Store model variables for easy access\n variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=self.name)\n self.vars = {var.name: var for var in variables}\n\n # Build metrics\n self._loss()\n self._accuracy()\n\n self.opt_op = self.optimizer.minimize(self.loss)\n\n def predict(self):\n pass\n\n def _loss(self):\n raise NotImplementedError\n\n def _accuracy(self):\n raise NotImplementedError\n\n def save(self, sess=None):\n if not sess:\n raise AttributeError(\"TensorFlow session not provided.\")\n saver = tf.train.Saver(self.vars)\n save_path = saver.save(sess, \"tmp/%s.ckpt\" % self.name)\n print(\"Model saved in file: %s\" % save_path)\n\n def load(self, sess=None):\n if not sess:\n raise AttributeError(\"TensorFlow session not provided.\")\n saver = tf.train.Saver(self.vars)\n save_path = \"tmp/%s.ckpt\" % self.name\n saver.restore(sess, save_path)\n print(\"Model restored from file: %s\" % save_path)\n\n\nclass MLP(Model):\n def __init__(self, placeholders, input_dim, **kwargs):\n super(MLP, self).__init__(**kwargs)\n\n self.inputs = placeholders['features']\n self.input_dim = input_dim\n # self.input_dim = self.inputs.get_shape().as_list()[1] # To be supported in future Tensorflow versions\n self.output_dim = placeholders['labels'].get_shape().as_list()[1]\n self.placeholders = placeholders\n\n self.optimizer = tf.train.AdamOptimizer(learning_rate=FLAGS.learning_rate)\n\n self.build()\n\n def _loss(self):\n # Weight decay loss\n for var in self.layers[0].vars.values():\n self.loss += FLAGS.weight_decay * tf.nn.l2_loss(var)\n\n # Cross entropy error\n self.loss += masked_softmax_cross_entropy(self.outputs, self.placeholders['labels'],\n self.placeholders['labels_mask'])\n\n def _accuracy(self):\n self.accuracy = masked_accuracy(self.outputs, self.placeholders['labels'],\n self.placeholders['labels_mask'])\n\n def _build(self):\n self.layers.append(Dense(input_dim=self.input_dim,\n output_dim=FLAGS.hidden1,\n placeholders=self.placeholders,\n act=tf.nn.relu,\n dropout=True,\n sparse_inputs=True,\n logging=self.logging))\n\n self.layers.append(Dense(input_dim=FLAGS.hidden1,\n output_dim=self.output_dim,\n placeholders=self.placeholders,\n act=lambda x: x,\n dropout=True,\n logging=self.logging))\n\n def predict(self):\n return tf.nn.softmax(self.outputs)\n\n\nclass GCN(Model):\n def __init__(self, placeholders, input_dim, **kwargs):\n super(GCN, self).__init__(**kwargs)\n\n self.inputs = placeholders['features']\n self.input_dim = input_dim\n # self.input_dim = self.inputs.get_shape().as_list()[1] # To be supported in future Tensorflow versions\n self.output_dim = placeholders['labels'].get_shape().as_list()[1]\n self.placeholders = placeholders\n\n self.optimizer = tf.train.AdamOptimizer(learning_rate=FLAGS.learning_rate)\n\n self.build()\n\n def _loss(self):\n # Weight decay loss\n for var in self.layers[0].vars.values():\n self.loss += FLAGS.weight_decay * tf.nn.l2_loss(var)\n\n # Cross entropy error\n self.loss += masked_softmax_cross_entropy(self.outputs, self.placeholders['labels'],\n self.placeholders['labels_mask'])\n\n def _accuracy(self):\n self.accuracy = masked_accuracy(self.outputs, self.placeholders['labels'],\n self.placeholders['labels_mask'])\n\n def _build(self):\n\n self.layers.append(GraphConvolution(input_dim=self.input_dim,\n output_dim=FLAGS.hidden1,\n placeholders=self.placeholders,\n act=tf.nn.relu,\n dropout=True,\n sparse_inputs=True,\n logging=self.logging))\n\n self.layers.append(GraphConvolution(input_dim=FLAGS.hidden1,\n output_dim=self.output_dim,\n placeholders=self.placeholders,\n act=lambda x: x,\n dropout=True,\n logging=self.logging))\n\n def predict(self):\n return tf.nn.softmax(self.outputs)\n" }, { "path": "gcn/train.py", "content": "from __future__ import division\nfrom __future__ import print_function\n\nimport time\nimport tensorflow as tf\n\nfrom gcn.utils import *\nfrom gcn.models import GCN, MLP\n\n# Set random seed\nseed = 123\nnp.random.seed(seed)\ntf.set_random_seed(seed)\n\n# Settings\nflags = tf.app.flags\nFLAGS = flags.FLAGS\nflags.DEFINE_string('dataset', 'cora', 'Dataset string.') # 'cora', 'citeseer', 'pubmed'\nflags.DEFINE_string('model', 'gcn', 'Model string.') # 'gcn', 'gcn_cheby', 'dense'\nflags.DEFINE_float('learning_rate', 0.01, 'Initial learning rate.')\nflags.DEFINE_integer('epochs', 200, 'Number of epochs to train.')\nflags.DEFINE_integer('hidden1', 16, 'Number of units in hidden layer 1.')\nflags.DEFINE_float('dropout', 0.5, 'Dropout rate (1 - keep probability).')\nflags.DEFINE_float('weight_decay', 5e-4, 'Weight for L2 loss on embedding matrix.')\nflags.DEFINE_integer('early_stopping', 10, 'Tolerance for early stopping (# of epochs).')\nflags.DEFINE_integer('max_degree', 3, 'Maximum Chebyshev polynomial degree.')\n\n# Load data\nadj, features, y_train, y_val, y_test, train_mask, val_mask, test_mask = load_data(FLAGS.dataset)\n\n# Some preprocessing\nfeatures = preprocess_features(features)\nif FLAGS.model == 'gcn':\n support = [preprocess_adj(adj)]\n num_supports = 1\n model_func = GCN\nelif FLAGS.model == 'gcn_cheby':\n support = chebyshev_polynomials(adj, FLAGS.max_degree)\n num_supports = 1 + FLAGS.max_degree\n model_func = GCN\nelif FLAGS.model == 'dense':\n support = [preprocess_adj(adj)] # Not used\n num_supports = 1\n model_func = MLP\nelse:\n raise ValueError('Invalid argument for model: ' + str(FLAGS.model))\n\n# Define placeholders\nplaceholders = {\n 'support': [tf.sparse_placeholder(tf.float32) for _ in range(num_supports)],\n 'features': tf.sparse_placeholder(tf.float32, shape=tf.constant(features[2], dtype=tf.int64)),\n 'labels': tf.placeholder(tf.float32, shape=(None, y_train.shape[1])),\n 'labels_mask': tf.placeholder(tf.int32),\n 'dropout': tf.placeholder_with_default(0., shape=()),\n 'num_features_nonzero': tf.placeholder(tf.int32) # helper variable for sparse dropout\n}\n\n# Create model\nmodel = model_func(placeholders, input_dim=features[2][1], logging=True)\n\n# Initialize session\nsess = tf.Session()\n\n\n# Define model evaluation function\ndef evaluate(features, support, labels, mask, placeholders):\n t_test = time.time()\n feed_dict_val = construct_feed_dict(features, support, labels, mask, placeholders)\n outs_val = sess.run([model.loss, model.accuracy], feed_dict=feed_dict_val)\n return outs_val[0], outs_val[1], (time.time() - t_test)\n\n\n# Init variables\nsess.run(tf.global_variables_initializer())\n\ncost_val = []\n\n# Train model\nfor epoch in range(FLAGS.epochs):\n\n t = time.time()\n # Construct feed dictionary\n feed_dict = construct_feed_dict(features, support, y_train, train_mask, placeholders)\n feed_dict.update({placeholders['dropout']: FLAGS.dropout})\n\n # Training step\n outs = sess.run([model.opt_op, model.loss, model.accuracy], feed_dict=feed_dict)\n\n # Validation\n cost, acc, duration = evaluate(features, support, y_val, val_mask, placeholders)\n cost_val.append(cost)\n\n # Print results\n print(\"Epoch:\", '%04d' % (epoch + 1), \"train_loss=\", \"{:.5f}\".format(outs[1]),\n \"train_acc=\", \"{:.5f}\".format(outs[2]), \"val_loss=\", \"{:.5f}\".format(cost),\n \"val_acc=\", \"{:.5f}\".format(acc), \"time=\", \"{:.5f}\".format(time.time() - t))\n\n if epoch > FLAGS.early_stopping and cost_val[-1] > np.mean(cost_val[-(FLAGS.early_stopping+1):-1]):\n print(\"Early stopping...\")\n break\n\nprint(\"Optimization Finished!\")\n\n# Testing\ntest_cost, test_acc, test_duration = evaluate(features, support, y_test, test_mask, placeholders)\nprint(\"Test set results:\", \"cost=\", \"{:.5f}\".format(test_cost),\n \"accuracy=\", \"{:.5f}\".format(test_acc), \"time=\", \"{:.5f}\".format(test_duration))\n" }, { "path": "gcn/utils.py", "content": "import numpy as np\nimport pickle as pkl\nimport networkx as nx\nimport scipy.sparse as sp\nfrom scipy.sparse.linalg.eigen.arpack import eigsh\nimport sys\n\n\ndef parse_index_file(filename):\n \"\"\"Parse index file.\"\"\"\n index = []\n for line in open(filename):\n index.append(int(line.strip()))\n return index\n\n\ndef sample_mask(idx, l):\n \"\"\"Create mask.\"\"\"\n mask = np.zeros(l)\n mask[idx] = 1\n return np.array(mask, dtype=np.bool)\n\n\ndef load_data(dataset_str):\n \"\"\"\n Loads input data from gcn/data directory\n\n ind.dataset_str.x => the feature vectors of the training instances as scipy.sparse.csr.csr_matrix object;\n ind.dataset_str.tx => the feature vectors of the test instances as scipy.sparse.csr.csr_matrix object;\n ind.dataset_str.allx => the feature vectors of both labeled and unlabeled training instances\n (a superset of ind.dataset_str.x) as scipy.sparse.csr.csr_matrix object;\n ind.dataset_str.y => the one-hot labels of the labeled training instances as numpy.ndarray object;\n ind.dataset_str.ty => the one-hot labels of the test instances as numpy.ndarray object;\n ind.dataset_str.ally => the labels for instances in ind.dataset_str.allx as numpy.ndarray object;\n ind.dataset_str.graph => a dict in the format {index: [index_of_neighbor_nodes]} as collections.defaultdict\n object;\n ind.dataset_str.test.index => the indices of test instances in graph, for the inductive setting as list object.\n\n All objects above must be saved using python pickle module.\n\n :param dataset_str: Dataset name\n :return: All data input files loaded (as well the training/test data).\n \"\"\"\n names = ['x', 'y', 'tx', 'ty', 'allx', 'ally', 'graph']\n objects = []\n for i in range(len(names)):\n with open(\"data/ind.{}.{}\".format(dataset_str, names[i]), 'rb') as f:\n if sys.version_info > (3, 0):\n objects.append(pkl.load(f, encoding='latin1'))\n else:\n objects.append(pkl.load(f))\n\n x, y, tx, ty, allx, ally, graph = tuple(objects)\n test_idx_reorder = parse_index_file(\"data/ind.{}.test.index\".format(dataset_str))\n test_idx_range = np.sort(test_idx_reorder)\n\n if dataset_str == 'citeseer':\n # Fix citeseer dataset (there are some isolated nodes in the graph)\n # Find isolated nodes, add them as zero-vecs into the right position\n test_idx_range_full = range(min(test_idx_reorder), max(test_idx_reorder)+1)\n tx_extended = sp.lil_matrix((len(test_idx_range_full), x.shape[1]))\n tx_extended[test_idx_range-min(test_idx_range), :] = tx\n tx = tx_extended\n ty_extended = np.zeros((len(test_idx_range_full), y.shape[1]))\n ty_extended[test_idx_range-min(test_idx_range), :] = ty\n ty = ty_extended\n\n features = sp.vstack((allx, tx)).tolil()\n features[test_idx_reorder, :] = features[test_idx_range, :]\n adj = nx.adjacency_matrix(nx.from_dict_of_lists(graph))\n\n labels = np.vstack((ally, ty))\n labels[test_idx_reorder, :] = labels[test_idx_range, :]\n\n idx_test = test_idx_range.tolist()\n idx_train = range(len(y))\n idx_val = range(len(y), len(y)+500)\n\n train_mask = sample_mask(idx_train, labels.shape[0])\n val_mask = sample_mask(idx_val, labels.shape[0])\n test_mask = sample_mask(idx_test, labels.shape[0])\n\n y_train = np.zeros(labels.shape)\n y_val = np.zeros(labels.shape)\n y_test = np.zeros(labels.shape)\n y_train[train_mask, :] = labels[train_mask, :]\n y_val[val_mask, :] = labels[val_mask, :]\n y_test[test_mask, :] = labels[test_mask, :]\n\n return adj, features, y_train, y_val, y_test, train_mask, val_mask, test_mask\n\n\ndef sparse_to_tuple(sparse_mx):\n \"\"\"Convert sparse matrix to tuple representation.\"\"\"\n def to_tuple(mx):\n if not sp.isspmatrix_coo(mx):\n mx = mx.tocoo()\n coords = np.vstack((mx.row, mx.col)).transpose()\n values = mx.data\n shape = mx.shape\n return coords, values, shape\n\n if isinstance(sparse_mx, list):\n for i in range(len(sparse_mx)):\n sparse_mx[i] = to_tuple(sparse_mx[i])\n else:\n sparse_mx = to_tuple(sparse_mx)\n\n return sparse_mx\n\n\ndef preprocess_features(features):\n \"\"\"Row-normalize feature matrix and convert to tuple representation\"\"\"\n rowsum = np.array(features.sum(1))\n r_inv = np.power(rowsum, -1).flatten()\n r_inv[np.isinf(r_inv)] = 0.\n r_mat_inv = sp.diags(r_inv)\n features = r_mat_inv.dot(features)\n return sparse_to_tuple(features)\n\n\ndef normalize_adj(adj):\n \"\"\"Symmetrically normalize adjacency matrix.\"\"\"\n adj = sp.coo_matrix(adj)\n rowsum = np.array(adj.sum(1))\n d_inv_sqrt = np.power(rowsum, -0.5).flatten()\n d_inv_sqrt[np.isinf(d_inv_sqrt)] = 0.\n d_mat_inv_sqrt = sp.diags(d_inv_sqrt)\n return adj.dot(d_mat_inv_sqrt).transpose().dot(d_mat_inv_sqrt).tocoo()\n\n\ndef preprocess_adj(adj):\n \"\"\"Preprocessing of adjacency matrix for simple GCN model and conversion to tuple representation.\"\"\"\n adj_normalized = normalize_adj(adj + sp.eye(adj.shape[0]))\n return sparse_to_tuple(adj_normalized)\n\n\ndef construct_feed_dict(features, support, labels, labels_mask, placeholders):\n \"\"\"Construct feed dictionary.\"\"\"\n feed_dict = dict()\n feed_dict.update({placeholders['labels']: labels})\n feed_dict.update({placeholders['labels_mask']: labels_mask})\n feed_dict.update({placeholders['features']: features})\n feed_dict.update({placeholders['support'][i]: support[i] for i in range(len(support))})\n feed_dict.update({placeholders['num_features_nonzero']: features[1].shape})\n return feed_dict\n\n\ndef chebyshev_polynomials(adj, k):\n \"\"\"Calculate Chebyshev polynomials up to order k. Return a list of sparse matrices (tuple representation).\"\"\"\n print(\"Calculating Chebyshev polynomials up to order {}...\".format(k))\n\n adj_normalized = normalize_adj(adj)\n laplacian = sp.eye(adj.shape[0]) - adj_normalized\n largest_eigval, _ = eigsh(laplacian, 1, which='LM')\n scaled_laplacian = (2. / largest_eigval[0]) * laplacian - sp.eye(adj.shape[0])\n\n t_k = list()\n t_k.append(sp.eye(adj.shape[0]))\n t_k.append(scaled_laplacian)\n\n def chebyshev_recurrence(t_k_minus_one, t_k_minus_two, scaled_lap):\n s_lap = sp.csr_matrix(scaled_lap, copy=True)\n return 2 * s_lap.dot(t_k_minus_one) - t_k_minus_two\n\n for i in range(2, k+1):\n t_k.append(chebyshev_recurrence(t_k[-1], t_k[-2], scaled_laplacian))\n\n return sparse_to_tuple(t_k)\n" }, { "path": "requirements.txt", "content": "networkx==2.2\nscipy==1.1.0\nsetuptools==40.6.3\nnumpy==1.15.4\ntensorflow==1.15.4\n" }, { "path": "setup.py", "content": "from setuptools import setup\nfrom setuptools import find_packages\n\nsetup(name='gcn',\n version='1.0',\n description='Graph Convolutional Networks in Tensorflow',\n author='Thomas Kipf',\n author_email='thomas.kipf@gmail.com',\n url='https://tkipf.github.io',\n download_url='https://github.com/tkipf/gcn',\n license='MIT',\n install_requires=['numpy>=1.15.4',\n 'tensorflow>=1.15.2,<2.0',\n 'networkx>=2.2',\n 'scipy>=1.1.0'\n ],\n package_data={'gcn': ['README.md']},\n packages=find_packages())\n" } ]