我正在使用Tensorflow学习深度学习和机器学习，我需要做出一个预测，即输入是直方图（每天），我以序号形式介绍日期（736608是2017-10-06）该输入所需的输出为8334.71。问题是经过一番迭代之后，损失又让我难忘了。
模型：
def create_model(self):
input = tf.placeholder(dtype=tf.float32, name='input', shape=(None, 1))
self.model = tf.layers.dense(inputs=input, units=1, name='output')
def get_training_data_from_elastic_result(self, file_path):
result = {
'data': np.array([], dtype=np.float32),
'label': np.array([], dtype=np.float32)
}
input_json = json.load(open(file_path, 'r'))
bucket_values = input_json['aggregations']['id']['date']['histogram']['buckets']
for bucket in bucket_values:
current_ordinal_date = datetime.fromtimestamp(float(bucket['key'] / 1000)).toordinal()
result['data'] = np.append(arr=result['data'], values=current_ordinal_date)
result['label'] = np.append(arr=result['label'], values=round(float(bucket['relevance_sum']['value']), 2))
"""Reshape to
[
[737051.],
[737052.],
[737053.]
]
[
[3108.11],
[1499.51],
[1327.35]
]
"""
result['data'] = result['data'].reshape(-1, 1)
result['label'] = result['label'].reshape(-1, 1)
return result
def train(self, input_data, input_labels, learning_rate=0.1, num_steps=1000, display_steps=10):
real_output = tf.placeholder(dtype=tf.float32, name='real_output', shape=(None, 1))
loss = tf.losses.mean_squared_error(labels=real_output, predictions=self.model)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(loss=loss)
for step in range(num_steps):
_, error_value = self.session.run([optimizer, loss], feed_dict={'input:0': input_data, 'real_output:0': input_labels})
if (step + 1) % display_steps == 0:
self.MODEL_LOGGER.info("Epoch: {epoch} Cost {cost}".format(
epoch=(step + 1),
cost=error_value
))
输出
INFO:model.model:Epoch: 1 Cost=1.06697175859e+12
INFO:model.model:Epoch: 2 Cost=1.25770053337e+34
INFO:model.model:Epoch: 3 Cost=inf
INFO:model.model:Epoch: 4 Cost=inf
INFO:model.model:Epoch: 5 Cost=inf
INFO:model.model:Epoch: 6 Cost=nan
INFO:model.model:Epoch: 7 Cost=nan
谢谢：）