def train_lstm():
pred, _ = lstm(pm.batch_size) # 调用的构建的lstm变量
# 损失函数 平均平方误差(MSE)
loss = tf.reduce_mean(
tf.square(tf.reshape(pred, [-1]) - tf.reshape(Y, [-1])))
# 实现梯度下降算法的优化器，优化损失函数
train_op = tf.train.AdamOptimizer(pm.lr).minimize(loss)
# 保存和恢复模型的方法；方法返回checkpoint文件的路径。可以直接传给restore() 进行调用
saver = tf.train.Saver(tf.global_variables())
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# 重复训练10000次
for i in range(5):
step = 0
start = 0
end = start + pm.batch_size
while (end < len(train_x)):
_, loss_ = sess.run([train_op, loss],
feed_dict={
X: train_x[start:end],
Y: train_y[start:end]
})
start += pm.batch_size
end = start + pm.batch_size
# 每10步保存一次参数
if step % 10 == 0:
print(i, step, loss_)
print("保存模型：", saver.save(sess, save_path))
step += 1
word_to_index["the"],
logfile=logpath,
is_verbose=is_verbose)
log("The word of index 20 is:",
index_to_word[20],
logfile=logpath,
is_verbose=is_verbose)
# lstm = LSTM(batch_size, embedding_size, vocab_size, hidden_size, max_size)
x = tf.placeholder(tf.int32, (batch_size, max_size - 1), name="x")
label = tf.placeholder(tf.int32, (batch_size, max_size - 1), name="label")
teacher_forcing = tf.placeholder(tf.bool, (), name="teacher_forcing")
word_embeddings, output, softmax_output = lstm(x, label, vocab_size,
hidden_size, max_size,
batch_size, embedding_size,
teacher_forcing)
with tf.Session() as sess:
onehot = tf.argmax(softmax_output, 2)
with tf.variable_scope("optimizer", reuse=tf.AUTO_REUSE):
optimizer, loss, cross_entropy_out, weights = optimize(
output, label, learning_rate)
perplexity = tf.exp(cross_entropy_out)
tf.summary.scalar('loss', loss)
tf.summary.scalar('perplexity', perplexity)
"""Now let's execute the graph in the session.
We ge a data batch with `dataloader.get_batch(batch_size)`. This fetches a batch of word sequences.
def __init__(self):
self.name = "predictionApi"
self.lstm = lstm()
self.Company = Configuration().GetData()['CompanyList']
load_folder = False #'/home/osvald/Projects/Diagnostics/Sepsis/Models/lstm/32_2x64_l01_p25'
offset = 0
n = 40336
n = 20000
split = 0.9
ind = list(range(n))
div = int(n * split)
partition = dict([])
partition['train'] = list(ind[:div])
partition['validation'] = list(ind[div:n])
model = lstm(embedding=args.embedding,
hidden_size=args.hidden,
fcl=args.fcl,
num_layers=args.layers,
batch_size=batch_size,
fcl_out=bool(args.fcl),
embed=bool(args.embedding),
droprate=args.drop).to(args.device)
criterion = nn.BCEWithLogitsLoss(pos_weight=torch.DoubleTensor([1.8224]).to(
args.device),
reduction='none')
optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
verbose=False,
factor=args.gamma,
patience=args.patience)
if load_model:
model.load_state_dict(torch.load(load_model))
train_losses = np.concatenate(