Graph Convolutional Networks for Text Classification原码解读[pytorch]

前言

啊，之前看锅tensorflow的原码，也记了点Graph Convolutional Networks for Text Classification原码解读[tensorflow]

项目地址

https://github.com/iworldtong/text_gcn.pytorch

环境配置

就在tensorflow那个版本的环境下，补装了1.7.1+cu101的pytorch

代码解析

remove_words.py

就是原版的代码

build_graph.py

也是原版的代码
好奇:看到也用了scipy的csr_matrix函数，难道pytorch也有类似的矩阵运算？

train.py

一些维度&一些输出

adj (61603, 61603)
features (61603, 300)
y_train (61603, 20)
y_val (61603, 20)
y_test (61603, 20)
train_mask (61603,)
val_mask (61603,)
test_mask (61603,)
train_size 11314
test_size 7532
tm_train_mask torch.Size([61603, 20])
t_support[0] torch.Size([61603, 61603])
pre_sup torch.Size([61603, 200])
support0 torch.Size([61603, 61603])
out torch.Size([61603, 200])
logits * tm_train_mask[0] ： tensor([ 0.0521, -0.0080,  0.2177, -0.1337,  0.1672, -0.0428,  0.0664, -0.1221,
         0.0376,  0.0709, -0.3589,  0.2038,  0.0118, -0.1365, -0.2384, -0.1432,
         0.0838,  0.1781,  0.2771,  0.1930], grad_fn=)
t_y_train[0]：tensor([0., 0., 0., 0., 0., 0., 0., 0., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0.], dtype=torch.float64)
torch.max(t_y_train, 1)[0]：tensor(8)

features = preprocess_features(features)

# return sparse_to_tuple(features)
return features.A

删除了sparse_to_tuple函数

def sparse_to_tuple(sparse_mx):
    """Convert sparse matrix to tuple representation."""
    def to_tuple(mx):
        if not sp.isspmatrix_coo(mx):
            mx = mx.tocoo()
        coords = np.vstack((mx.row, mx.col)).transpose()
        values = mx.data
        shape = mx.shape
        return coords, values, shape

    if isinstance(sparse_mx, list):
        for i in range(len(sparse_mx)):
            sparse_mx[i] = to_tuple(sparse_mx[i])
    else:
        sparse_mx = to_tuple(sparse_mx)

    return sparse_mx

preprocess_adj(adj)

同样是这样

# return sparse_to_tuple(adj_normalized)
return adj_normalized.A

tm_train_mask = torch.transpose(torch.unsqueeze(t_train_mask, 0), 1, 0).repeat(1, y_train.shape[1])

把原来是(real_train_size+valid_size+vocab_size+test_size,)的向量转成了一个(real_train_size+valid_size+vocab_size+test_size, 标签个数)的tensor

训练

啊，略了。

评估

from sklearn import metrics
print_log("Test Precision, Recall and F1-Score...")
print_log(metrics.classification_report(test_labels, test_pred, digits=4))
print_log("Macro average Test Precision, Recall and F1-Score...")
print_log(metrics.precision_recall_fscore_support(test_labels, test_pred, average='macro'))
print_log("Micro average Test Precision, Recall and F1-Score...")
print_log(metrics.precision_recall_fscore_support(test_labels, test_pred, average='micro'))

构建自己的数据集-wiki80

选择了80/10/10的wiki_727K玩玩：

adj (7002, 7002)
features (7002, 300)
y_train (7002, 2)
y_val (7002, 2)
y_test (7002, 2)
train_mask (7002,)
val_mask (7002,)
test_mask (7002,)
train_size 3927
test_size 352

构建自己的数据集-wiki800

adj (109772, 109772)
features (109772, 300)
y_train (109772, 2)
y_val (109772, 2)
y_test (109772, 2)
train_mask (109772,)
val_mask (109772,)
test_mask (109772,)
train_size 83497
test_size 4564

构建自己的数据集-wiki8000