TensorFlow图神经网络报错：SparseTensor无法转为Tensor求解决

报错信息 #

TypeError: Failed to convert object of type 
<class 'TensorFlow.python.framework.sparse_tensor.SparseTensor'> 
to Tensor. 
Contents: SparseTensor(indices=Tensor("DeserializeSparse:0", 
shape=(None, 2), dtype=int64), 
values=Tensor("DeserializeSparse:1", shape=(None,), dtype=float32), 
dense_shape=Tensor("stack:0", shape=(2,), dtype=int64)). 
Consider casting elements to a supported type.

环境信息 #

• Python 3.6
• TensorFlow 2.6.2
• Keras 2.6.0

运行代码 #

import numpy as np
import tensorflow
from tensorflow.python.keras.callbacks import ModelCheckpoint
from tensorflow.python.keras.layers import Lambda
from tensorflow.python.keras.models import Model
from tensorflow.python.keras.optimizer_v1 import adam
from tensorflow.python.keras.optimizers import adam_v2
import tensorflow as tf
import sklearn
from sklearn.feature_extraction.text import CountVectorizer

from gcn import GCN
from utils import preprocess_adj, plot_embeddings, load_data_v1

if __name__ == "__main__":
    physical_devices = tf.config.experimental.list_physical_devices('GPU')

    if len(physical_devices) > 0:
        for k in range(len(physical_devices)):
            tf.config.experimental.set_memory_growth(physical_devices[k], True)
            print('memory growth:', tf.config.experimental.get_memory_growth(physical_devices[k]))
    else:
        print("Not enough GPU hardware devices available")


    FEATURE_LESS = False
    print("Num GPUs Available: ", len(tf.config.list_physical_devices('GPU')))

    A, features, y_train, y_val, y_test, train_mask, val_mask, test_mask = load_data_v1(
        'cora')

    A = preprocess_adj(A)

    features /= features.sum(axis=1, ).reshape(-1, 1)

    if FEATURE_LESS:
        X = np.arange(A.shape[-1])

        feature_dim = A.shape[-1]
    else:
        X = features

        feature_dim = X.shape[-1]
    model_input = [X, A]

    # Compile model
    model = GCN(A.shape[-1], feature_dim, 16, y_train.shape[1], dropout_rate=0.5, l2_reg=2.5e-4,
                feature_less=FEATURE_LESS, )
    model.compile(optimizer=adam_v2.Adam(0.01), loss='categorical_crossentropy',
                  weighted_metrics=['categorical_crossentropy', 'acc'])

    NB_EPOCH = 200
    PATIENCE = 200  # early stopping patience

    val_data = (model_input, y_val, val_mask)
    mc_callback = ModelCheckpoint('./best_model.h5',
                                  monitor='val_weighted_categorical_crossentropy',
                                  save_best_only=True,
                                  save_weights_only=True)

    # train
    print("start training")
    model.fit(model_input, y_train, sample_weight=train_mask, validation_data=val_data, batch_size=A.shape[0], epochs=NB_EPOCH, shuffle=False, verbose=2,
              callbacks=[mc_callback])
    # test
    model.load_weights('./best_model.h5')
    eval_results = model.evaluate(
        model_input, y_test, sample_weight=test_mask, batch_size=A.shape[0])
    print('Done.\n'
          'Test loss: {}\n'
          'Test weighted_loss: {}\n'
          'Test accuracy: {}'.format(*eval_results))

    embedding_model = Model(model.input, outputs=Lambda(lambda x: model.layers[-1].output)(model.input))
    embedding_weights = embedding_model.predict(model_input, batch_size=A.shape[0])
    y = np.genfromtxt("{}{}.content".format('../data/cora/', 'cora'), dtype=np.dtype(str))[:, -1]
    plot_embeddings(embedding_weights, np.arange(A.shape[0]), y)

解决方案 #

报错核心是preprocess_adj函数返回了SparseTensor，但Keras模型无法直接将其转换为普通Tensor，以下是两种可行解决思路：

思路1：转换为稠密Tensor（小数据集适用） #

在预处理邻接矩阵后，添加代码将稀疏矩阵转为稠密矩阵：

A = preprocess_adj(A)
# 新增：将SparseTensor转为稠密矩阵
A = tf.sparse.to_dense(A).numpy()

注意：数据集过大时（节点数超10000），稠密矩阵会占用大量内存，不建议使用此方法。

思路2：修改GCN模型适配SparseTensor（大数据集推荐） #

保留稀疏矩阵以节省内存，需要调整GCN层实现：

• 前向传播中用tf.sparse.sparse_dense_matmul()替代普通tf.matmul，处理稀疏邻接矩阵与特征矩阵的乘法
• 模型输入层明确声明支持SparseTensor，用tf.keras.Input(..., sparse=True)定义输入

额外优化建议 #

• 检查load_data_v1和preprocess_adj函数，确保返回格式适配TensorFlow 2.x（很多老GCN代码基于TF1.x编写，需做版本兼容）
• 统一使用tf.keras模块，避免混用tensorflow.python.keras，减少兼容性问题，例如将from tensorflow.python.keras.callbacks import ModelCheckpoint改为from tensorflow.keras.callbacks import ModelCheckpoint

内容的提问来源于stack exchange，提问作者LUKAS