我用tensorflow（1.14.1）估计器创建了一个分类器，并在每个步骤中计算了度量。我使用tensorboard来显示指标，并且每个步骤都将其融合。我想在每个时代重置它。
这适用于运行Python（3.7）和tensorflow（1.14.1）的HPC。
def fn_builder(...):
def model_fn(features, labels, mode, params):
# create bert-mlp net
ids = features['ids']
words = features['words']
masks = features['mask']
label_mask = features['label_mask']
# model
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
loss, prd_label = create_model(...)
pre = tf.metrics.precision(labels, prd_label, weights=label_mask)
metrics = {'Precision': pre}
tf.summary.scalar('p', pre[1])
if mode == tf.estimator.ModeKeys.TRAIN:
optimizer = Adam(learning_rate=learning_rate)
train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())
return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op, eval_metric_ops=metrics)
return pi_model_fn
if __name__ == "__main__":
batch_size = 5
epochs = 30
num_eval_steps = 20 / batch_size + 1
num_train_steps = 20 * epochs / batch_size
pi_checkpointing_config = tf.estimator.RunConfig(
model_dir='data/pi',
save_checkpoints_secs=10 * 60, # Save checkpoints every 10min.
keep_checkpoint_max=10, # Retain the 10 most recent checkpoints.
)
pi = tf.estimator.Estimator(
model_fn=fn_builder(...),
config=pi_checkpointing_config)
pi.train(
input_fn=input_fn_creator(...),
max_steps=num_train_steps
)
这是我的：https://tva1.sinaimg.cn/large/006y8mN6ly1g77gj8vie2j30m808ut8y.jpg
