标签：图神经网络，GAT，注意力机制，tensorflow
GraphSAGE+GNN换成GAT的本质分析
在GraphSAGE+GNN的实现中，对邻居节点采用某种方式聚合计算（例如求向量均值），再和中心节点拼接的方式，GraphSAGE固定每层采样的个数，GNN固定层数，模型学习的就是每一层邻居聚合之后的W以及中心节点向量的W，以及最后一个分类的全连接。将GNN换为GAT之后，还是采用邻居聚合计算的方式（带上中心节点一起加权求和），此时要多学习一个权重，比如中心节点采样了10个邻居，要学习11个权重参数，如果有8个注意力机制，就要学习88个权重参数，这个权重参数不是直接定义学习的，而是通过一个共享的W和两个单独a运算得到，模型实际上学习迭代的是W和a1，a2，每套注意力机制下W和a1，a2共享，W，a1，a2在训练之后保存在tensorflow网络结构中，新的节点在预测的时候拿到计算权重再加权求和得到下一层的表征。这三个参数的训练结果是由中心节点和他的一组邻居节点的特征向量决定的，直接应用在新的预测中心节点和他的一组邻居节点上，且这个和邻居节点采样的排列顺序无关，跟着特征向量走。
数据准备预处理
使用CORA数据集，数据预处理如下，主要是sample函数，会进行一个两层采样。将前1000个作为训练，1000~1708为验证，1708到最后为测试
import os
import pickle
import numpy as np
import scipy.sparse as sp
BASIC_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
def data_split():
names = ['x', 'y', 'tx', 'ty', 'allx', 'ally', 'graph']
objects = []
for i in range(len(names)):
with open(os.path.join(BASIC_PATH, "./data/ind.cora.{}".format(names[i])), 'rb') as f:
objects.append(pickle.load(f, encoding='latin1'))
x, y, tx, ty, allx, ally, graph = tuple(objects)
test_idx_reorder = [int(x.strip()) for x in open(os.path.join(BASIC_PATH, "./data/ind.cora.test.index"), "r").readlines()]
test_idx_range = np.sort(test_idx_reorder)
# 测试索引位置修正
features = sp.vstack((allx, tx)).tolil()
features[test_idx_reorder, :] = features[test_idx_range, :]
labels = np.vstack((ally, ty))
labels[test_idx_reorder, :] = labels[test_idx_range, :]
# 训练[:1000],验证[1000:1708],测试[1708:]
train_nodes = list(range(1000))
train_y = labels[train_nodes]
val_nodes = list(range(1000, 1708))
val_y = labels[val_nodes]
test_nodes = list(range(1708, 2708))
test_y = labels[test_nodes]
return train_nodes, train_y, val_nodes, val_y, test_nodes, test_y, graph, features
def sample(nodes, neighbour_list, k=2, num_supports=None):
if num_supports is None:
num_supports = [10, 25]
assert len(num_supports) == k, "num_supports长度必须和k阶相等"
layer_neighbours = {}
for i in range(k):
neighbours = []
num_support = num_supports[i]
for node in nodes:
one_neighbour = neighbour_list[node]
if len(one_neighbour) >= num_support:
neighbours.append(np.random.choice(neighbour_list[node], num_support, replace=False).tolist())
else:
neighbours.append(np.random.choice(neighbour_list[node], num_support, replace=True).tolist())
layer_neighbours[k - i] = neighbours
nodes = sum(neighbours, [])
return layer_neighbours
def get_nodes_features(nodes, features_embedding, std=True):
embedding = features_embedding[nodes]
if std:
embedding = embedding / embedding.sum(axis=1)
return embedding
if __name__ == '__main__':
train_nodes, train_y, val_nodes, val_y, test_nodes, test_y, graph, features = data_split()
pickle.dump((train_nodes, train_y), open(os.path.join(BASIC_PATH, "./data/train.pkl"), "wb"))
pickle.dump((val_nodes, val_y), open(os.path.join(BASIC_PATH, "./data/val.pkl"), "wb"))
pickle.dump((test_nodes, test_y), open(os.path.join(BASIC_PATH, "./data/test.pkl"), "wb"))
pickle.dump(graph, open(os.path.join(BASIC_PATH, "./data/graph.pkl"), "wb"))
pickle.dump(features, open(os.path.join(BASIC_PATH, "./data/features.pkl"), "wb"))
模型部分
模型部分直接上代码，模型中写死了只有2层的GAT，tensor的链路参考GAT源码阅读和tensor链路
import numpy as np
import tensorflow as tf
def glorot(shape, name=None):
init_range = np.sqrt(6.0 / (shape[0] + shape[1]))
initial = tf.random_uniform(shape, minval=-init_range, maxval=init_range, dtype=tf.float32)
return tf.Variable(initial, name=name)
def zeros(shape, name=None):
initial = tf.zeros(shape, dtype=tf.float32)
return tf.Variable(initial, name=name)
class GraphSageGAT(object):
def __init__(self, num_class, feature_size, num_supports_1=10, num_supports_2=10, learning_rate=0.01,
weight_decay=0.01, decay_learning_rate=0.99):
# 中心节点
self.input_self = tf.placeholder(tf.float32, [None, feature_size], name="input_self")
# 1跳
self.input_neigh_1 = tf.placeholder(tf.float32, [None, num_supports_1, feature_size], name="input_neigh_1")
# 2跳
self.input_neigh_2 = tf.placeholder(tf.float32, [None, num_supports_1, num_supports_2, feature_size], name="input_neigh_2")
# label
self.input_y = tf.placeholder(tf.int64, [None, num_class])
self.feature_size = feature_size
self.num_supports_1 = num_supports_1
self.num_supports_2 = num_supports_2
self.weight_decay = weight_decay
self.learning_rate = learning_rate
self.decay_learning_rate = decay_learning_rate
self.w_dropout_keep_prob = tf.placeholder(tf.float32, name="w_dropout_keep_prob")
self.e_dropout_keep_prob = tf.placeholder(tf.float32, name="e_dropout_keep_prob")
self.global_step = tf.Variable(0, name="global_step", trainable=False)
def build(self, n_heads_1=8, n_heads_2=8, w_size_1=8, w_size_2=7):
"""
模型网络构建，指定第一层聚合注意力个数和第二层个数
:param w_size_2:
:param w_size_1:
:param n_heads_1:
:param n_heads_2:
:return:
"""
# 第一层
one_att_out_one_hop = []
one_att_out_two_hop = []
for i in range(n_heads_1):
# [512, 8]
one_att_out_one_hop.append(self.attn_head(self.input_self, self.input_neigh_1, w_size=w_size_1,
w_drop=self.w_dropout_keep_prob, eij_drop=self.e_dropout_keep_prob,
input_size=self.feature_size))
# [5120, 8]
one_att_out_two_hop.append(self.attn_head(
tf.reshape(self.input_neigh_1, [-1, self.feature_size]),
tf.reshape(self.input_neigh_2, [-1, self.num_supports_2, self.feature_size]), w_size=w_size_1,
w_drop=self.w_dropout_keep_prob, eij_drop=self.e_dropout_keep_prob, input_size=self.feature_size))
h_1_one_hop = tf.concat(one_att_out_one_hop, axis=-1) # [512, 64]
h_1_two_hop = tf.concat(one_att_out_two_hop, axis=-1) # [5120, 64]
# 第二层
two_att_out = []
for i in range(n_heads_2):
# [512, 7]
two_att_out.append(self.attn_head(h_1_one_hop, tf.reshape(h_1_two_hop, [-1, self.num_supports_1, w_size_1 * n_heads_1]),
w_size=w_size_2, w_drop=self.w_dropout_keep_prob, eij_drop=self.e_dropout_keep_prob,
input_size=w_size_1 * n_heads_1))
# [512, 7]
output = tf.add_n(two_att_out) / n_heads_2
# softmax
with tf.name_scope("softmax_out"):
self.probs = tf.nn.softmax(output, dim=1, name="probs")
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.arg_max(self.probs, 1), tf.arg_max(self.input_y, 1)), dtype=tf.float32))
# loss
with tf.name_scope('loss'):
self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=output, labels=self.input_y))
# 增加l2 loss
vars = tf.trainable_variables()
loss_l2 = tf.add_n([tf.nn.l2_loss(v) for v in vars if v.name not in ['bias', 'gamma', 'b', 'g', 'beta']]) * self.weight_decay
self.loss += loss_l2
# optimizer
with tf.name_scope("optimizer"):
if self.decay_learning_rate:
learning_rate = tf.train.exponential_decay(self.learning_rate, self.global_step, 100, self.decay_learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
self.train_step = optimizer.minimize(self.loss, global_step=self.global_step)
with tf.name_scope("summaries"):
tf.summary.scalar("loss", self.loss)
tf.summary.scalar("accuracy", self.accuracy)
self.summary_op = tf.summary.merge_all()
def attn_head(self, center_tensor, neigh_tensor, w_size, w_drop, eij_drop, input_size):
"""
注意力机制，输入中心节点特征向量和采样的一度邻居特征向量，输出注意力机制后的中心节点聚合特征向量
:param input_size:
:param center_tensor: 中心节点特征向量
:param neigh_tensor: 邻居节点特征向量
:param w_size: W向量转化之后的维度
:param w_drop: Wh的dropout
:param eij_drop: eij的dropout
:return: 下一层的中心节点特征向量
"""
# w共享
w = glorot([input_size, w_size]) # [1433, 8]
wh_center = tf.matmul(center_tensor, w) # [512, 1433] => [512, 8]
wh_neigh = tf.matmul(neigh_tensor, w) # [512, 10, 1433] => [512, 10, 8]
# concat
wh_center_neigh = tf.concat([wh_neigh, tf.expand_dims(wh_center, 1)], axis=1) # [512, 10, 8] => [512, 11, 8]
# a单独两个
a_center = glorot([w_size, 1]) # [8, 1]
a_neigh = glorot([w_size, 1]) # [8, 1]
wha_center = tf.contrib.layers.bias_add(tf.matmul(wh_center, a_center)) # [512, 8] => [512, 1]
wha_neigh = tf.contrib.layers.bias_add(tf.matmul(wh_center_neigh, a_neigh)) # [512, 11, 8] => [512, 11, 1]
# wha+wha
wha = tf.squeeze(wha_neigh + tf.reshape(wha_center, [-1, 1, 1]), axis=2) # [512, 11]
# eij
eij = tf.nn.softmax(tf.nn.leaky_relu(wha)) # [512, 11]
# dropout
eij = tf.nn.dropout(eij, eij_drop) # [512, 11]
wh_center_neigh = tf.nn.dropout(wh_center_neigh, w_drop) # [512, 11, 8]
# 加权求和
new_center_tensor = tf.matmul(tf.expand_dims(eij, axis=1), wh_center_neigh)
new_center_tensor = tf.contrib.layers.bias_add(new_center_tensor) # [512, 1, 8]
new_center_tensor = tf.squeeze(tf.nn.elu(new_center_tensor), axis=1) # [512, 8]
return new_center_tensor
训练部分
训练部分通过实例化GAT对象的build方法构图，指定了第一层8个注意力，第二层8个注意力，第一层的维度转化向量到8维，第二层输出7维直接映射到label
import sys
import os
import pickle
import shutil
import random
import time
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
DATA_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
ROOT_PATH = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
import numpy as np
import tensorflow as tf
from tensorflow.python.saved_model import tag_constants
from model import GraphSageGAT
from utils.config import get_string
from preprocessing import sample, get_nodes_features
(train_nodes, train_y) = pickle.load(open(os.path.join(DATA_PATH, get_string("train_data_path")), "rb"))
(val_nodes, val_y) = pickle.load(open(os.path.join(DATA_PATH, get_string("val_data_path")), "rb"))
(test_nodes, test_y) = pickle.load(open(os.path.join(DATA_PATH, get_string("test_data_path")), "rb"))
neighbour_list = pickle.load(open(os.path.join(DATA_PATH, get_string("neighbour_data_path")), "rb"))
nodes_features = pickle.load(open(os.path.join(DATA_PATH, get_string("feature_data_path")), "rb"))
features_size = nodes_features.shape[1]
def get_batch(epoches, batch_size, nodes, labels, neighbours, features, layer1_supports=10, layer2_supports=5):
for epoch in range(epoches):
tmp = list(zip(nodes, labels))
random.shuffle(tmp)
nodes, labels = zip(*tmp)
for batch in range(0, len(nodes), batch_size):
if batch + batch_size < len(nodes):
batch_nodes = nodes[batch: (batch + batch_size)]
batch_labels = labels[batch: (batch + batch_size)]
else:
batch_nodes = nodes[batch: len(nodes)]
batch_labels = labels[batch: len(nodes)]
# 得到训练集的1跳2跳
layer_neighbours = sample(batch_nodes, neighbours, num_supports=[layer2_supports, layer1_supports])
# 所有节点的embedding
input_x = get_nodes_features(list(batch_nodes), features)
input_x_1 = get_nodes_features(sum(layer_neighbours[2], []), features)
input_x_2 = get_nodes_features(sum(layer_neighbours[1], []), features)
yield [epoch, input_x, input_x_1, input_x_2, batch_labels]
def train_main():
tf.reset_default_graph()
model = GraphSageGAT(num_class=7, feature_size=features_size,
num_supports_1=int(get_string("layer2_supports")),
num_supports_2=int(get_string("layer1_supports")),
decay_learning_rate=float(get_string("decay_learning_rate")),
learning_rate=float(get_string("learning_rate")),
weight_decay=float(get_string("weight_decay")))
model.build(n_heads_1=8, n_heads_2=8, w_size_1=8, w_size_2=7)
saver = tf.train.Saver(tf.global_variables(), max_to_keep=1)
with tf.Session() as sess:
init_op = tf.group(tf.global_variables_initializer())
sess.run(init_op)
shutil.rmtree(os.path.join(ROOT_PATH, "./summary"), ignore_errors=True)
writer = tf.summary.FileWriter(os.path.join(ROOT_PATH, "./summary"), sess.graph)
batches = get_batch(int(get_string("epoches")), int(get_string("batch_size")), train_nodes, train_y,
neighbour_list, nodes_features, layer1_supports=int(get_string("layer1_supports")),
layer2_supports=int(get_string("layer2_supports")))
# 验证数据
layer_neighbours = sample(val_nodes, neighbour_list,
num_supports=[int(get_string("layer2_supports")), int(get_string("layer1_supports"))])
val_input_x = get_nodes_features(val_nodes, nodes_features)
val_input_x_1 = get_nodes_features(sum(layer_neighbours[2], []), nodes_features)
val_input_x_2 = get_nodes_features(sum(layer_neighbours[1], []), nodes_features)
val_feed_dict = {model.input_self: val_input_x,
model.input_neigh_1: val_input_x_1.A.reshape(-1, int(get_string("layer1_supports")),
features_size),
model.input_neigh_2: val_input_x_2.A.reshape(-1, int(get_string("layer1_supports")),
int(get_string("layer2_supports")),
features_size),
model.input_y: val_y,
model.w_dropout_keep_prob: 1,
model.e_dropout_keep_prob: 1}
val_loss_list = []
for batch in batches:
epoch, input_x, input_x_1, input_x_2, input_y = batch
feed_dict = {model.input_self: input_x,
model.input_neigh_1: input_x_1.A.reshape(-1, int(get_string("layer1_supports")),
features_size),
model.input_neigh_2: input_x_2.A.reshape(-1, int(get_string("layer1_supports")),
int(get_string("layer2_supports")), features_size),
model.input_y: input_y,
model.w_dropout_keep_prob: float(get_string("dropout_keep_prob")),
model.e_dropout_keep_prob: float(get_string("dropout_keep_prob"))}
_, step, loss_val, acc_val, merged = sess.run(
[model.train_step, model.global_step, model.loss, model.accuracy, model.summary_op],
feed_dict=feed_dict)
writer.add_summary(merged, step)
if step % 5 == 0:
print("epoch:", epoch + 1, "step:", step, "loss:", loss_val, "accuracy:", acc_val)
if step % 50 == 0:
loss_val, acc_val = sess.run([model.loss, model.accuracy], feed_dict=val_feed_dict)
print("{:-^30}".format("evaluation"))
print("[evaluation]", "loss:", loss_val, "accuracy:", acc_val)
# 计算当前loss相比之前的最有loss下降多少
diff = (loss_val - min(val_loss_list)) if len(val_loss_list) else 0
val_loss_list.append(loss_val)
print("本轮loss比之前最小loss{}：{}, 当前最小loss: {}"
.format("上升" if diff > 0 else "下降", abs(diff), min(val_loss_list)))
if diff < 0:
saver.save(sess, os.path.join(ROOT_PATH, get_string("checkpoint_path")))
print("[save checkpoint]")
print("-" * 40)
if early_stop(val_loss_list, windows=int(get_string("early_stop_windows"))):
print("{:-^30}".format("early stop!"))
break
def early_stop(loss_list, windows=5):
if len(loss_list) <= windows:
return False
latest_loss = loss_list[-windows:]
previous_loss = loss_list[:-windows]
min_previous_loss = min(previous_loss)
min_latest_loss = min(latest_loss)
if min_latest_loss > min_previous_loss:
return True
return False
运行train_main如下
epoch: 52 step: 415 loss: 0.54380906 accuracy: 0.921875
epoch: 53 step: 420 loss: 0.50763553 accuracy: 0.9296875
epoch: 54 step: 425 loss: 0.50306284 accuracy: 0.953125
epoch: 54 step: 430 loss: 0.5158588 accuracy: 0.953125
epoch: 55 step: 435 loss: 0.51461077 accuracy: 0.953125
epoch: 55 step: 440 loss: 0.48844844 accuracy: 0.96153843
epoch: 56 step: 445 loss: 0.4956718 accuracy: 0.953125
epoch: 57 step: 450 loss: 0.48547933 accuracy: 0.9609375
----------evaluation----------
[evaluation] loss: 0.66410494 accuracy: 0.8926554
本轮loss比之前最小loss上升：0.007818341255187988, 当前最小loss: 0.6562865972518921
----------------------------------------
---------early stop!----------
训练集softmax7分类的准确率0.96，验证有0.89
测试部分
测试引入测试样本，直接去读ckpt预测对比测试集的label
def test_main():
layer_neighbours = sample(test_nodes, neighbour_list,
num_supports=[int(get_string("layer2_supports")), int(get_string("layer1_supports"))])
test_input_x = get_nodes_features(test_nodes, nodes_features)
test_input_x_1 = get_nodes_features(sum(layer_neighbours[2], []), nodes_features)
test_input_x_2 = get_nodes_features(sum(layer_neighbours[1], []), nodes_features)
tf.reset_default_graph()
with tf.Session() as sess:
last_ckpt = tf.train.latest_checkpoint(
os.path.join(ROOT_PATH, "/".join(get_string("checkpoint_path").split("/")[:-1])))
print("读取ckpt: {}".format(last_ckpt))
saver = tf.train.import_meta_graph("{}.meta".format(last_ckpt))
saver.restore(sess, last_ckpt)
graph = tf.get_default_graph()
# get tensor
input_self = graph.get_tensor_by_name("input_self:0")
input_neigh_1 = graph.get_tensor_by_name("input_neigh_1:0")
input_neigh_2 = graph.get_tensor_by_name("input_neigh_2:0")
w_dropout_keep_prob = graph.get_tensor_by_name("w_dropout_keep_prob:0")
e_dropout_keep_prob = graph.get_tensor_by_name("e_dropout_keep_prob:0")
pred = graph.get_tensor_by_name("softmax_out/probs:0")
prediction = sess.run(pred, feed_dict={input_self: test_input_x,
input_neigh_1: test_input_x_1.A.reshape(-1, int(get_string("layer1_supports")), features_size),
input_neigh_2: test_input_x_2.A.reshape(-1, int(get_string("layer1_supports")), int(get_string("layer2_supports")), features_size),
w_dropout_keep_prob: 1.0,
e_dropout_keep_prob: 1.0})
hit = np.equal(np.argmax(prediction, axis=1), np.argmax(test_y, axis=1))
accuracy = hit.sum() / len(hit)
print("[test]:", accuracy)
运行结果
读取ckpt: /home/myproject/GRAPHSAGE_GAT_CORA/./ckpt/ckpt
[test]: 0.85
tensorboard查看训练标量
切刀项目工程根目录下执行
tensorboard --logdir ../summary
tensorboard
模型保存为pb
将模型保存为pb文件
def save_pb():
# 模型保存
pb_num = str(int(time.time()))
pb_path = os.path.join(ROOT_PATH, get_string("pb_path"), pb_num)
shutil.rmtree(pb_path, ignore_errors=True)
tf.reset_default_graph()
with tf.Session() as sess:
last_ckpt = tf.train.latest_checkpoint(
os.path.join(ROOT_PATH, "/".join(get_string("checkpoint_path").split("/")[:-1])))
print("读取ckpt: {}".format(last_ckpt))
saver = tf.train.import_meta_graph("{}.meta".format(last_ckpt))
saver.restore(sess, last_ckpt)
graph = tf.get_default_graph()
# get tensor
input_self = graph.get_tensor_by_name("input_self:0")
input_neigh_1 = graph.get_tensor_by_name("input_neigh_1:0")
input_neigh_2 = graph.get_tensor_by_name("input_neigh_2:0")
w_dropout_keep_prob = graph.get_tensor_by_name("w_dropout_keep_prob:0")
e_dropout_keep_prob = graph.get_tensor_by_name("e_dropout_keep_prob:0")
pred = graph.get_tensor_by_name("softmax_out/probs:0")
builder = tf.saved_model.builder.SavedModelBuilder(pb_path)
inputs = {'input_self': tf.saved_model.utils.build_tensor_info(input_self),
'input_neigh_1': tf.saved_model.utils.build_tensor_info(input_neigh_1),
'input_neigh_2': tf.saved_model.utils.build_tensor_info(input_neigh_2),
'w_dropout_keep_prob': tf.saved_model.utils.build_tensor_info(w_dropout_keep_prob),
'e_dropout_keep_prob': tf.saved_model.utils.build_tensor_info(e_dropout_keep_prob)
}
outputs = {'output': tf.saved_model.utils.build_tensor_info(pred)}
signature = tf.saved_model.signature_def_utils.build_signature_def(
inputs=inputs,
outputs=outputs,
method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)
builder.add_meta_graph_and_variables(sess, [tag_constants.SERVING], {'my_signature': signature})
builder.save()
print("pb文件保存完成:", pb_num)
运行结果
读取ckpt: /home/myproject/GRAPHSAGE_GAT_CORA/./ckpt/ckpt
pb文件保存完成: 1657025573
模型server API服务
先启一个docker tensorflow-model-server服务
#!/bin/bash
docker run --rm \
-p 13713:8501 \
-v /home/myproject/GRAPHSAGE_GAT_CORA/tfserving/:/models/graphsage_gat_cora/ \
-e MODEL_NAME=graphsage_gat_cora \
--name graphsage_gat_api \
tensorflow/serving \
--enable_batching=true
requests请求测试，请求原数据中2555节点的分类
import os
import pickle
import sys
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
import requests
from preprocessing import sample, get_nodes_features
from utils.config import get_string
DATA_PATH = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
neighbour_list = pickle.load(open(os.path.join(DATA_PATH, get_string("neighbour_data_path")), "rb"))
nodes_features = pickle.load(open(os.path.join(DATA_PATH, get_string("feature_data_path")), "rb"))
if __name__ == '__main__':
nodes = [2555]
layer_neighbours = sample(nodes, neighbour_list,
num_supports=[int(get_string("layer2_supports")), int(get_string("layer1_supports"))])
test_input_x = get_nodes_features(nodes, nodes_features)
test_input_x_1 = get_nodes_features(sum(layer_neighbours[2], []), nodes_features)
test_input_x_2 = get_nodes_features(sum(layer_neighbours[1], []), nodes_features)
res = requests.post("http://127.0.0.1:13713/v1/models/graphsage_gat_cora:predict", json={"instances": [{
"input_self": test_input_x.A[0].tolist(),
"input_neigh_1": test_input_x_1.A.reshape(-1, 10, 1433)[0].tolist(),
"input_neigh_2": test_input_x_2.A.reshape(-1, 10, 10, 1433)[0].tolist(),
"w_dropout_keep_prob": 1.0,
"e_dropout_keep_prob": 1.0
}], "signature_name": "my_signature"})
print(res.json())
预测结果如下，结果是第四分类
{'predictions': [[0.00618362846, 0.00481465, 0.00280520879, 0.973845184, 0.00697943475, 0.00364588178, 0.00172607298]]}
网友评论