eileen-code4fun

from sklearn.decomposition import PCA

def project(words, phrases, model):
    embeddings = []
    associated_words = []
    pca = PCA(n_components=2)
    for word in words:
        associated_words.append(word)
        embeddings.append(model[word])
        for pphrase in phrases[word]:

eileen-code4fun

def ngram_paraphrase(words, phrases, ngram, factor=2):
   sentences = []
   for i in range(0, len(words)-ngram):
       phrase = ''
       for j in range(0, ngram):
           phrase += words[i+j] + ' '
       phrase = phrase.strip()
       if not phrase in phrases:
           continue
       for p in list(phrases[phrase])[0:factor]:

eileen-code4fun

def nn_sentence_aug(sentence, factor=2, model):
   sentences = []
   words = sentence.split()
   for i in range(len(words)):
       ps = model.most_similar(words[i], topn=factor)
       for p in ps:
           new_words = words[0:i] + [p[0]] + words[i+1:]
           sentences.append(' '.join(new_words))
   return sentences

eileen-code4fun

import gensim
import gensim.downloader
model = gensim.downloader.load('glove-twitter-50')

eileen-code4fun

from dgl.nn.tensorflow import GATConv

class GAT(tf.keras.Model):
    def __init__(self, feat_dim, hidden_dim, class_num):
        super(GAT, self).__init__()
        self.num_heads = 8
        self.h1 = GATConv(in_feats=feat_dim, out_feats=hidden_dim, feat_drop=0.5, num_heads=self.num_heads, allow_zero_in_degree=True)
        self.h2 = GATConv(in_feats=hidden_dim*self.num_heads, out_feats=class_num, feat_drop=0.5, num_heads=1, allow_zero_in_degree=True)
    
    def call(self, g, features):

eileen-code4fun

from dgl.nn.tensorflow import GINConv

class GIN(tf.keras.Model):
    def mlp(feat_dim, hidden_dim, out_dim):
        m = tf.keras.models.Sequential()
        m.add(tf.keras.layers.Input(shape=(feat_dim,)))
        m.add(tf.keras.layers.Dense(hidden_dim, activation='relu'))
        m.add(tf.keras.layers.Dropout(0.5))
        m.add(tf.keras.layers.Dense(out_dim))
        return m

eileen-code4fun

from dgl.nn.tensorflow import SAGEConv

class SAGE(tf.keras.Model):
    def __init__(self, feat_dim, hidden_dim, class_num):
        super(SAGE, self).__init__()
        self.h1 = SAGEConv(in_feats=feat_dim, out_feats=hidden_dim, aggregator_type='pool', feat_drop=0.5, activation=tf.nn.relu)
        self.h2 = SAGEConv(in_feats=feat_dim, out_feats=class_num, aggregator_type='pool')
    
    def call(self, g, features):
        h = features

eileen-code4fun

from dgl.nn.tensorflow import GraphConv

class GCN(tf.keras.Model):
    def __init__(self, feat_dim, hidden_dim, class_num):
        super(GCN, self).__init__()
        self.h1 = GraphConv(feat_dim, hidden_dim, activation=tf.nn.relu, allow_zero_in_degree=True)
        self.dropout = tf.keras.layers.Dropout(0.5)
        self.h2 = GraphConv(hidden_dim, class_num, allow_zero_in_degree=True)
    
    def call(self, g, features):

eileen-code4fun

import dgl
import tensorflow as tf

def eval(model, g, node_ids_to_eval):
    logits = model(g, g.ndata['feat'], training=False)
    logits = tf.gather(logits, node_ids_to_eval)
    labels = tf.gather(g.ndata['label'], node_ids_to_eval)
    indices = tf.math.argmax(logits, axis=1)
    acc = tf.reduce_mean(tf.cast(indices == labels, dtype=tf.float32))
    return acc.numpy().item()

eileen-code4fun

for epoch in range(20):
    with tf.GradientTape() as tape:
        logits = g_model(g.ndata['feat'])
        # Loss is only calculated on the training set.
        loss = loss_fcn(g.ndata['label'][g.ndata['train_mask']], logits[g.ndata['train_mask']])
        grads = tape.gradient(loss, g_model.trainable_weights)
        opt.apply_gradients(zip(grads, g_model.trainable_weights))