损失减少，但模型无法覆盖

Question

我尝试使用张量流实现简单的点击率预测模型。但是，训练损失减少了，但准确性保持不变。 （实际上，它等于正样本或负样本的比率，似乎该模型预测所有样本的比率为0或1）。我该怎么办才能解决这个问题？

def parse_exmp(serial_exmp):
    click = fc.numeric_column("click", dtype=tf.float32)
    fea_columns = [click]
    fea_columns += my_feature_columns
    feature_spec=tf.feature_column.make_parse_example_spec(fea_columns)
    feats = tf.parse_single_example(serial_exmp, features=feature_spec)
    click = feats.pop('click')
    return feats, {'ctr': tf.to_float(click)}


def train_input_fn(filenames, batch_size, shuffle_buffer_size):
    dataset = tf.data.TFRecordDataset(filenames)
    dataset = dataset.map(parse_exmp, num_parallel_calls=5)
    dataset = dataset.repeat().batch(batch_size).prefetch(1000)
    return dataset


def eval_input_fn(filename, batch_size):
    dataset = tf.data.TFRecordDataset(filename)
    dataset = dataset.map(parse_exmp, num_parallel_calls=5)
    # Shuffle, repeat, and batch the examples.
    dataset = dataset.batch(batch_size)
    # Return the read end of the pipeline.
    return dataset


def build_model(features, mode, params):
    net = fc.input_layer(features, params['feature_columns'])
    # Build the hidden layers, sized according to the 'hidden_units' param.

    for units in params['hidden_units']:
    net = tf.layers.dense(net, units=units, kernel_regularizer=None, 
                              bias_regularizer=None, activation=tf.nn.relu)
    if 'dropout_rate' in params and params['dropout_rate'] > 0.0:
        net = tf.layers.dropout(net, params['dropout_rate'], training=(mode == tf.estimator.ModeKeys.TRAIN))

    # Compute logits
    logits = tf.layers.dense(net, 1, activation=None)
    return logits

def my_model(features, labels, mode, params):
    ctr_logits = build_model(features, mode, params)
    ctr_predictions = tf.sigmoid(ctr_logits, name="CTR")

    if mode == tf.estimator.ModeKeys.PREDICT:
        predictions = ctr_predictions
        export_outputs = {
            'prediction': tf.estimator.export.PredictOutput(predictions)
        }
        return tf.estimator.EstimatorSpec(mode, predictions=predictions, export_outputs=export_outputs)

    loss = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(labels=labels['ctr'], logits=ctr_logits), name="loss")

    if mode == tf.estimator.ModeKeys.TRAIN:
        optimizer = tf.train.AdagradOptimizer(learning_rate=params['learning_rate'])
        train_op = optimizer.minimize(loss, global_step=tf.train.get_global_step())
        return tf.estimator.EstimatorSpec(mode=mode, loss=loss, train_op=train_op)

    ctr_accuracy = tf.metrics.accuracy(labels=labels['ctr'], predictions=tf.to_float(tf.greater_equal(ctr_predictions, 0.5)))
    ctr_auc = tf.metrics.auc(labels['ctr'], ctr_predictions)

    metrics = {'ctr_accuracy': ctr_accuracy, 'ctr_auc': ctr_auc}
    tf.summary.scalar('ctr_accuracy', ctr_accuracy[1])
    tf.summary.scalar('ctr_auc', ctr_auc[1])

    if mode == tf.estimator.ModeKeys.EVAL:
        return tf.estimator.EstimatorSpec(mode, loss=loss, eval_metric_ops=metrics)


create_feature_columns()
classifier = tf.estimator.Estimator(
    model_fn=my_model,
    params={
            'feature_columns': my_feature_columns,
            'hidden_units': hidden_units.split(','),
            'learning_rate': learning_rate,
            'dropout_rate': dropout_rate
    },
    config=tf.estimator.RunConfig(model_dir=model_dir, save_checkpoints_steps=save_checkpoints_steps)
)
batch_size = batch_size


train_spec = tf.estimator.TrainSpec(
    input_fn=lambda: train_input_fn(train_files, batch_size, shuffle_buffer_size),
    max_steps=train_steps
)
input_fn_for_eval = lambda: eval_input_fn(eval_files, batch_size)
eval_spec = tf.estimator.EvalSpec(input_fn=input_fn_for_eval, throttle_secs = 60, steps=None)

这是日志：

INFO:tensorflow:Running training and evaluation locally (non-distributed).
INFO:tensorflow:Start train and evaluate loop. The evaluate will happen after 60 secs (eval_spec.throttle_secs) or training is finished.
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Saving checkpoints for 1 into ./model_dir/model.ckpt.
INFO:tensorflow:loss = 8575.396, step = 1
INFO:tensorflow:Saving checkpoints for 45 into ./model_dir/model.ckpt.
INFO:tensorflow:Loss for final step: 92.291855.
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Starting evaluation at 2019-06-10-07:05:03
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from ./model_dir/model.ckpt-45
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Finished evaluation at 2019-06-10-07:05:17
INFO:tensorflow:Saving dict for global step 45: ctr_accuracy = 0.436, ctr_auc = 0.5, global_step = 45, loss = 6467.209
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from ./model_dir/model.ckpt-45
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Saving checkpoints for 46 into ./model_dir/model.ckpt.
INFO:tensorflow:loss = 1430.6135, step = 46
INFO:tensorflow:Saving checkpoints for 87 into ./model_dir/model.ckpt.
INFO:tensorflow:Loss for final step: 469.1917.
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Starting evaluation at 2019-06-10-07:06:44
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from ./model_dir/model.ckpt-87
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Finished evaluation at 2019-06-10-07:06:58
INFO:tensorflow:Saving dict for global step 87: ctr_accuracy = 0.436, ctr_auc = 0.5, global_step = 87, loss = 3759.593
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Create CheckpointSaverHook.
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from ./model_dir/model.ckpt-87
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Saving checkpoints for 88 into ./model_dir/model.ckpt.
INFO:tensorflow:loss = 938.0892, step = 88
INFO:tensorflow:Saving checkpoints for 131 into ./model_dir/model.ckpt.
INFO:tensorflow:Loss for final step: 3.707542e-12.
INFO:tensorflow:Calling model_fn.
INFO:tensorflow:Done calling model_fn.
INFO:tensorflow:Starting evaluation at 2019-06-10-07:08:26
INFO:tensorflow:Graph was finalized.
INFO:tensorflow:Restoring parameters from ./model_dir/model.ckpt-131
INFO:tensorflow:Running local_init_op.
INFO:tensorflow:Done running local_init_op.
INFO:tensorflow:Finished evaluation at 2019-06-10-07:08:39
INFO:tensorflow:Saving dict for global step 131: ctr_accuracy = 0.436, ctr_auc = 0.5, global_step = 131, loss = 2633.0947