論文地址：https://arxiv.org/abs/1710.10903
程式碼地址： https://github.com/PetarV-/GAT
我並沒有完整看過這篇論文，但是在大致瞭解其原理之後就直接看了程式碼= =。
接下來我將從程式碼的整個流程開始講解，首先解析的是不用稀疏矩陣儲存的：
使用的資料集：Cora dataset
Cora資料集簡要介紹：
圖節點數：2708
每個節點的特徵維度：1433
鄰接矩陣：(2708,2708)，關係表示的是論文之間的引用關係
類別數：7
目錄結構：
相關的包：
import numpy as np import pickle as pkl import networkx as nx import scipy.sparse as sp from scipy.sparse.linalg.eigen.arpack import eigsh import sys import time import numpy as np import tensorflow as tf
網路需要的資料是什麼？
在utils檔案下的process.py中：
def load_data(dataset_str): # {'pubmed', 'citeseer', 'cora'} """Load data.""" names = ['x', 'y', 'tx', 'ty', 'allx', 'ally', 'graph'] objects = [] for i in range(len(names)): with open("data/ind.{}.{}".format(dataset_str, names[i]), 'rb') as f: if sys.version_info > (3, 0): objects.append(pkl.load(f, encoding='latin1')) else: objects.append(pkl.load(f)) x, y, tx, ty, allx, ally, graph = tuple(objects) test_idx_reorder = parse_index_file("data/ind.{}.test.index".format(dataset_str)) test_idx_range = np.sort(test_idx_reorder) if dataset_str == 'citeseer': # Fix citeseer dataset (there are some isolated nodes in the graph) # Find isolated nodes, add them as zero-vecs into the right position test_idx_range_full = range(min(test_idx_reorder), max(test_idx_reorder)+1) tx_extended = sp.lil_matrix((len(test_idx_range_full), x.shape[1])) tx_extended[test_idx_range-min(test_idx_range), :] = tx tx = tx_extended ty_extended = np.zeros((len(test_idx_range_full), y.shape[1])) ty_extended[test_idx_range-min(test_idx_range), :] = ty ty = ty_extended features = sp.vstack((allx, tx)).tolil() features[test_idx_reorder, :] = features[test_idx_range, :] adj = nx.adjacency_matrix(nx.from_dict_of_lists(graph)) labels = np.vstack((ally, ty)) labels[test_idx_reorder, :] = labels[test_idx_range, :] idx_test = test_idx_range.tolist() idx_train = range(len(y)) idx_val = range(len(y), len(y)+500) train_mask = sample_mask(idx_train, labels.shape[0]) val_mask = sample_mask(idx_val, labels.shape[0]) test_mask = sample_mask(idx_test, labels.shape[0]) y_train = np.zeros(labels.shape) y_val = np.zeros(labels.shape) y_test = np.zeros(labels.shape) y_train[train_mask, :] = labels[train_mask, :] y_val[val_mask, :] = labels[val_mask, :] y_test[test_mask, :] = labels[test_mask, :] print(adj.shape) print(features.shape) return adj, features, y_train, y_val, y_test, train_mask, val_mask, test_mask
不管怎麼變化，我們需要得到的是：
鄰接矩陣：(2708,2708)，需要注意的是鄰接矩陣是由nx.adjacency_matrix(nx.from_dict_of_lists(graph))獲得，其中graph是引用關係組成的字典，比如：
{ 1:[2,3], 4:[6,7,8], }
還需要注意的是該函式返回的稀疏矩陣的儲存，是一個coo_matrix：
(0, 1) 1 (0, 30) 1 (0, 33) 1 (0, 99) 1
第一個元組表示不為0的值在陣列中出現的(行、列)，第二個值表示具體的值。看一個例子：
三要素:資料、行,列索引、形狀。
特徵：(2708,1433)
訓練集、驗證集、測試集標籤：維度都是(2708,7).。注意：tensorflow中的標籤需要進行one-hot編碼，才能夠使用交叉熵進行計算，而pytorch中是不用的。
掩碼：掩碼的作用就是選取部分節點進行訓練、測試和驗證，具體操作是：
idx_test = test_idx_range.tolist() idx_train = range(len(y)) idx_val = range(len(y), len(y)+500) train_mask = sample_mask(idx_train, labels.shape[0]) val_mask = sample_mask(idx_val, labels.shape[0]) test_mask = sample_mask(idx_test, labels.shape[0])
其中sample_mask函式程式碼，位於process.py中：
def sample_mask(idx, l): """Create mask.""" mask = np.zeros(l) mask[idx] = 1 return np.array(mask, dtype=np.bool)
也就是先根據l的大小生成一個形狀相同的陣列，然後根據Idx的值將指定行的值變為True。請看以下例子：
如何對初始資料進行處理？
adj, features, y_train, y_val, y_test, train_mask, val_mask, test_mask = process.load_data(dataset) features, spars = process.preprocess_features(features) nb_nodes = features.shape[0] #節點數 ft_size = features.shape[1] #特徵維度 nb_classes = y_train.shape[1] #類別數 adj = adj.todense() features = features[np.newaxis] adj = adj[np.newaxis] y_train = y_train[np.newaxis] y_val = y_val[np.newaxis] y_test = y_test[np.newaxis] train_mask = train_mask[np.newaxis] val_mask = val_mask[np.newaxis] test_mask = test_mask[np.newaxis] biases = process.adj_to_bias(adj, [nb_nodes], nhood=1)
使用process.preprocess_features(features)對特徵進行處理：
def preprocess_features(features): """Row-normalize feature matrix and convert to tuple representation""" rowsum = np.array(features.sum(1)) r_inv = np.power(rowsum, -1).flatten() r_inv[np.isinf(r_inv)] = 0. r_mat_inv = sp.diags(r_inv) features = r_mat_inv.dot(features) return features.todense(), sparse_to_tuple(features)
我們獲得的features是由：sp.vstack((allx, tx)).tolil()得到的，也就是一個lil_matrix。lil_matrix基於行連線儲存的稀疏矩陣，看以下例子：
回到上面對特徵進行處理的函式，具體步驟為：
- 對每一行求和
- 計算每個值的倒數，然後展開成一維陣列
- 將inf型別的值置為0
- 生成一個對角矩陣，對角的值為一維陣列的值，其餘的為0
- 最後與特徵進行np.matmul
- 使用todense()函式lil_matrix還原為原始矩陣
接下來我們同樣將adj轉換為原始的矩陣，並且使用np.newaxis新增一個維度：
features；(2708,1433) ->(1,2708,1433)
adj：(2708,2708) -> (1,2708,2708)
y_train：(2708,7) -> (1,2708,7)
y_val：(2708,7) -> (1,2708,7)
y_test：(2708,7) -> (1,2708,7)
train_mask：(2708,.) -> (1,2708)
val_mask：(2708,.) -> (1,2708)
test_mask：(2708,.) -> (1,2708)
最後是：biases = process.adj_to_bias(adj, [nb_nodes], nhood=1)
看一下adj_to_bias(adj, [nb_nodes], nhood=1)函式：
def adj_to_bias(adj, sizes, nhood=1): nb_graphs = adj.shape[0] #1 mt = np.empty(adj.shape) #(1，2708,2708) for g in range(nb_graphs): #range(1) mt[g] = np.eye(adj.shape[1]) #mt[0] = np.eye(2708) -> (2708,2708) for _ in range(nhood): #range(1) mt[g] = np.matmul(mt[g], (adj[g] + np.eye(adj.shape[1]))) #np.matmul((2708,2708),((2708,2708)+(2708,2708)))->(2708,2708) for i in range(sizes[g]): #range(2078) for j in range(sizes[g]): #range(2078) if mt[g][i][j] > 0.0: mt[g][i][j] = 1.0 return -1e9 * (1.0 - mt)
np.eye()：生成一個對角矩陣，對角的值為1，其餘地方為0。
這裡要注意np.matmul的用法，看一個例子：
對應步驟：
- 生成一個空(1,2708,2708)的矩陣mt
- 將mt[0]，也就是(2708,2708)變為對角值為1的對角矩陣
- 該對角矩陣和(adj[g] + np.eye(adj.shape[1]))進行np.matmul
- 然後將mt中大於0的元素置為1
- 最後返回-1e9*(1-mt)，形狀為(1,2708,2708)
可以看出該函式的作用就是為不相鄰的兩個節點間新增一個偏置項。
GAT模型是怎麼建立的？
多頭注意力：
在models資料夾下的gat.py中：
import numpy as np import tensorflow as tf from utils import layers from models.base_gattn import BaseGAttN class GAT(BaseGAttN): def inference(inputs, nb_classes, nb_nodes, training, attn_drop, ffd_drop, bias_mat, hid_units, n_heads, activation=tf.nn.elu, residual=False): attns = [] for _ in range(n_heads[0]): attns.append(layers.attn_head(inputs, bias_mat=bias_mat, out_sz=hid_units[0], activation=activation, in_drop=ffd_drop, coef_drop=attn_drop, residual=False)) h_1 = tf.concat(attns, axis=-1) for i in range(1, len(hid_units)): h_old = h_1 attns = [] for _ in range(n_heads[i]): attns.append(layers.attn_head(h_1, bias_mat=bias_mat, out_sz=hid_units[i], activation=activation, in_drop=ffd_drop, coef_drop=attn_drop, residual=residual)) h_1 = tf.concat(attns, axis=-1) out = [] for i in range(n_heads[-1]): out.append(layers.attn_head(h_1, bias_mat=bias_mat, out_sz=nb_classes, activation=lambda x: x, in_drop=ffd_drop, coef_drop=attn_drop, residual=False)) logits = tf.add_n(out) / n_heads[-1] return logits
相關引數：
hid_units=[8]
n_heads=[8，1]
每個注意力頭的輸出為：(1,2708,8)，h_1將每個注意頭(共8個)的輸出進行拼接得到：(1,2708,64)。
根據len(hid_units)=1，因此第2個for迴圈是沒有被執行的，然後再進行一次多頭注意力(1個注意力頭)得到的輸出out就是：(1,2708,7)
每一個注意力頭：attn_head
接下來是重頭，具體的注意力的實現，論文的核心也就是在這裡了。
在utils資料夾下的layers.py中：
def attn_head(seq, out_sz, bias_mat, activation, in_drop=0.0, coef_drop=0.0, residual=False): with tf.name_scope('my_attn'): if in_drop != 0.0: seq = tf.nn.dropout(seq, 1.0 - in_drop) seq_fts = tf.layers.conv1d(seq, out_sz, 1, use_bias=False) # simplest self-attention possible f_1 = tf.layers.conv1d(seq_fts, 1, 1) f_2 = tf.layers.conv1d(seq_fts, 1, 1) logits = f_1 + tf.transpose(f_2, [0, 2, 1]) coefs = tf.nn.softmax(tf.nn.leaky_relu(logits) + bias_mat) if coef_drop != 0.0: coefs = tf.nn.dropout(coefs, 1.0 - coef_drop) if in_drop != 0.0: seq_fts = tf.nn.dropout(seq_fts, 1.0 - in_drop) vals = tf.matmul(coefs, seq_fts) ret = tf.contrib.layers.bias_add(vals) # residual connection if residual: if seq.shape[-1] != ret.shape[-1]: ret = ret + conv1d(seq, ret.shape[-1], 1) # activation else: ret = ret + seq return activation(ret) # activation
有關於dropout和殘差連線就不細講了，一個是為了緩解過擬合問題，一個是為了緩解梯度消失問題。
步驟如下：
- 輸入features：(1,2708,1433)
- 第一步對特徵features卷積，得到seq_fts：(1,2708,8)
- 第二步分別對seq_fts進行卷積，得到f_1：(1,2708,1)，f_2：(1,2708,1)
- 第三步將f_2調整形狀為:(1,1,2708)，利用廣播機制與f_1相加得到logits：(1,2708,2708)
- 第四步加上偏置項biases_mat，我們注意到偏置項的值很小，是為了過濾掉和該節點不相鄰的節點，因為考慮的是某節點和其鄰居節點之間的注意力。並使用leakyrelu進行啟用，最後經過softmax生成注意力得分coefs：(1,2708,2708)
- 第五步將注意力得分與seq_fts進行np.matmul得到帶注意力得分的vals：(1,2708,8)
- 第六步：ret = tf.contrib.layers.bias_add(vals)，這裡我執行時是報錯的，說沒有什麼TPU之類的，我將其註釋掉了才執行成功（tensorflow1.14）
- 最後將ret進行啟用：tf.nn.elu
那麼為什麼要這麼做呢？
主要是如何對注意力進行建模，實際上f_1+f_2也就是(1,2708,1) * (1,1,2708)，這裡就計算了每個節點和其它節點之間的得分。
其它的一些相關？
GAT類繼承了BasicAttn，我們看下BasicAttn類：
class BaseGAttN: def loss(logits, labels, nb_classes, class_weights): sample_wts = tf.reduce_sum(tf.multiply(tf.one_hot(labels, nb_classes), class_weights), axis=-1) xentropy = tf.multiply(tf.nn.sparse_softmax_cross_entropy_with_logits( labels=labels, logits=logits), sample_wts) return tf.reduce_mean(xentropy, name='xentropy_mean') def training(loss, lr, l2_coef): # weight decay vars = tf.trainable_variables() lossL2 = tf.add_n([tf.nn.l2_loss(v) for v in vars if v.name not in ['bias', 'gamma', 'b', 'g', 'beta']]) * l2_coef # optimizer opt = tf.train.AdamOptimizer(learning_rate=lr) # training op train_op = opt.minimize(loss+lossL2) return train_op def preshape(logits, labels, nb_classes): new_sh_lab = [-1] new_sh_log = [-1, nb_classes] log_resh = tf.reshape(logits, new_sh_log) lab_resh = tf.reshape(labels, new_sh_lab) return log_resh, lab_resh def confmat(logits, labels): preds = tf.argmax(logits, axis=1) return tf.confusion_matrix(labels, preds) ########################## # Adapted from tkipf/gcn # ########################## def masked_softmax_cross_entropy(logits, labels, mask): """Softmax cross-entropy loss with masking.""" loss = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels) mask = tf.cast(mask, dtype=tf.float32) mask /= tf.reduce_mean(mask) loss *= mask return tf.reduce_mean(loss) def masked_sigmoid_cross_entropy(logits, labels, mask): """Softmax cross-entropy loss with masking.""" labels = tf.cast(labels, dtype=tf.float32) loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=logits, labels=labels) loss=tf.reduce_mean(loss,axis=1) mask = tf.cast(mask, dtype=tf.float32) mask /= tf.reduce_mean(mask) loss *= mask return tf.reduce_mean(loss) def masked_accuracy(logits, labels, mask): """Accuracy with masking.""" correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(labels, 1)) accuracy_all = tf.cast(correct_prediction, tf.float32) mask = tf.cast(mask, dtype=tf.float32) mask /= tf.reduce_mean(mask) accuracy_all *= mask return tf.reduce_mean(accuracy_all) def micro_f1(logits, labels, mask): """Accuracy with masking.""" predicted = tf.round(tf.nn.sigmoid(logits)) # Use integers to avoid any nasty FP behaviour predicted = tf.cast(predicted, dtype=tf.int32) labels = tf.cast(labels, dtype=tf.int32) mask = tf.cast(mask, dtype=tf.int32) # expand the mask so that broadcasting works ([nb_nodes, 1]) mask = tf.expand_dims(mask, -1) # Count true positives, true negatives, false positives and false negatives. tp = tf.count_nonzero(predicted * labels * mask) tn = tf.count_nonzero((predicted - 1) * (labels - 1) * mask) fp = tf.count_nonzero(predicted * (labels - 1) * mask) fn = tf.count_nonzero((predicted - 1) * labels * mask) # Calculate accuracy, precision, recall and F1 score. precision = tp / (tp + fp) recall = tp / (tp + fn) fmeasure = (2 * precision * recall) / (precision + recall) fmeasure = tf.cast(fmeasure, tf.float32) return fmeasure
這裡面定義了訓練迴圈，損失計算，模型評價指標等等。
最後是計算圖以及session：
with tf.Graph().as_default(): with tf.name_scope('input'): ftr_in = tf.placeholder(dtype=tf.float32, shape=(batch_size, nb_nodes, ft_size)) bias_in = tf.placeholder(dtype=tf.float32, shape=(batch_size, nb_nodes, nb_nodes)) lbl_in = tf.placeholder(dtype=tf.int32, shape=(batch_size, nb_nodes, nb_classes)) msk_in = tf.placeholder(dtype=tf.int32, shape=(batch_size, nb_nodes)) attn_drop = tf.placeholder(dtype=tf.float32, shape=()) ffd_drop = tf.placeholder(dtype=tf.float32, shape=()) is_train = tf.placeholder(dtype=tf.bool, shape=()) logits = model.inference(ftr_in, nb_classes, nb_nodes, is_train, attn_drop, ffd_drop, bias_mat=bias_in, hid_units=hid_units, n_heads=n_heads, residual=residual, activation=nonlinearity) log_resh = tf.reshape(logits, [-1, nb_classes]) lab_resh = tf.reshape(lbl_in, [-1, nb_classes]) msk_resh = tf.reshape(msk_in, [-1]) loss = model.masked_softmax_cross_entropy(log_resh, lab_resh, msk_resh) accuracy = model.masked_accuracy(log_resh, lab_resh, msk_resh) train_op = model.training(loss, lr, l2_coef) saver = tf.train.Saver() init_op = tf.group(tf.global_variables_initializer(), tf.local_variables_initializer()) vlss_mn = np.inf vacc_mx = 0.0 curr_step = 0 with tf.Session() as sess: sess.run(init_op) train_loss_avg = 0 train_acc_avg = 0 val_loss_avg = 0 val_acc_avg = 0 for epoch in range(nb_epochs): tr_step = 0 tr_size = features.shape[0] while tr_step * batch_size < tr_size: _, loss_value_tr, acc_tr = sess.run([train_op, loss, accuracy], feed_dict={ ftr_in: features[tr_step*batch_size:(tr_step+1)*batch_size], bias_in: biases[tr_step*batch_size:(tr_step+1)*batch_size], lbl_in: y_train[tr_step*batch_size:(tr_step+1)*batch_size], msk_in: train_mask[tr_step*batch_size:(tr_step+1)*batch_size], is_train: True, attn_drop: 0.6, ffd_drop: 0.6}) train_loss_avg += loss_value_tr train_acc_avg += acc_tr tr_step += 1 vl_step = 0 vl_size = features.shape[0] while vl_step * batch_size < vl_size: loss_value_vl, acc_vl = sess.run([loss, accuracy], feed_dict={ ftr_in: features[vl_step*batch_size:(vl_step+1)*batch_size], bias_in: biases[vl_step*batch_size:(vl_step+1)*batch_size], lbl_in: y_val[vl_step*batch_size:(vl_step+1)*batch_size], msk_in: val_mask[vl_step*batch_size:(vl_step+1)*batch_size], is_train: False, attn_drop: 0.0, ffd_drop: 0.0}) val_loss_avg += loss_value_vl val_acc_avg += acc_vl vl_step += 1 print('Training: loss = %.5f, acc = %.5f | Val: loss = %.5f, acc = %.5f' % (train_loss_avg/tr_step, train_acc_avg/tr_step, val_loss_avg/vl_step, val_acc_avg/vl_step)) if val_acc_avg/vl_step >= vacc_mx or val_loss_avg/vl_step <= vlss_mn: if val_acc_avg/vl_step >= vacc_mx and val_loss_avg/vl_step <= vlss_mn: vacc_early_model = val_acc_avg/vl_step vlss_early_model = val_loss_avg/vl_step saver.save(sess, checkpt_file) vacc_mx = np.max((val_acc_avg/vl_step, vacc_mx)) vlss_mn = np.min((val_loss_avg/vl_step, vlss_mn)) curr_step = 0 else: curr_step += 1 if curr_step == patience: print('Early stop! Min loss: ', vlss_mn, ', Max accuracy: ', vacc_mx) print('Early stop model validation loss: ', vlss_early_model, ', accuracy: ', vacc_early_model) break train_loss_avg = 0 train_acc_avg = 0 val_loss_avg = 0 val_acc_avg = 0 saver.restore(sess, checkpt_file) ts_size = features.shape[0] ts_step = 0 ts_loss = 0.0 ts_acc = 0.0 while ts_step * batch_size < ts_size: loss_value_ts, acc_ts = sess.run([loss, accuracy], feed_dict={ ftr_in: features[ts_step*batch_size:(ts_step+1)*batch_size], bias_in: biases[ts_step*batch_size:(ts_step+1)*batch_size], lbl_in: y_test[ts_step*batch_size:(ts_step+1)*batch_size], msk_in: test_mask[ts_step*batch_size:(ts_step+1)*batch_size], is_train: False, attn_drop: 0.0, ffd_drop: 0.0}) ts_loss += loss_value_ts ts_acc += acc_ts ts_step += 1 print('Test loss:', ts_loss/ts_step, '; Test accuracy:', ts_acc/ts_step) sess.close()
這裡需要注意的是，雖然目前tensorflow出了2.x版本之後就簡化了很多，但是之前1.x版本還是需要了解的。以上程式碼就是1.x版本，1.x版本的tensorflow的一般分為以下步驟：
（1）利用類似python的上下文管理器定義計算圖；
（2）為計算圖中的一些變數定義佔位符；
（3）定義損失、模型、評價指標、訓練迴圈、初始化引數、儲存模型等；
（4）將訓練流程放入到session中，然後在相關的位置喂入實際的資料；
總結：
最後還是做個總結吧。能看到這裡相信你對GAT的整個脈絡已經瞭解的很清楚了，順帶還了解了其它的一些庫的用法。如何才能真正的去弄懂一篇論文，每個人都有其自己的方式，包括去看原論文，去網上找相應的部落格，其實在有大致瞭解之後去看原始碼不失為一種好的辦法。對於程式碼量比較少而言，可以像我這樣一步一步的去剖析每行程式碼的作用，但是對於龐大的程式碼而言，還是要根據功能點劃分程式碼塊，針對於不同功能的程式碼塊花不同的精力去研究。
由於我的水平有限，其中難免有理解不正確的地方，歡迎指出。