与Keras相比，训练MLP时准确性低

如何解决与Keras相比，训练MLP时准确性低

我正在尝试使用Covertype数据集（https://archive.ics.uci.edu/ml/datasets/covertype）来构建没有Keras的MLP，与使用Keras（0.85预测）相比，我的准确性非常低。如何提高训练数据的准确性？

数据集已被拆分，并进行了相应的热编码

X_train.shape,y_train.shape,X_test.shape,y_test.shape
((406703,54),(406703,7),(174309,7))

X = tf.placeholder(tf.float32,[None,54])
y_true = tf.placeholder(tf.float32,7])

def mlp(x):
    w1 = tf.Variable(tf.random_uniform([54,18]))
    b1 = tf.Variable(tf.zeros([18]))
    h1 = tf.nn.relu(tf.matmul(x,w1) + b1)
    # output layer
    w3 = tf.Variable(tf.random_uniform([18,7]))
    b3 = tf.Variable(tf.zeros([7]))
    logits= tf.matmul(h1,w3) + b3
    
    return logits

logits = mlp(X)

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits,labels=y_true))

training_op = tf.train.AdamOptimizer(learning_rate=0.01).minimize(cost)

# initialize
init = tf.global_variables_initializer()

# train hyperparameters
epoch_cnt = 30
batch_size = 1000
iteration = len(X_train) // batch_size

# Start training
with tf.Session() as sess:
    # Run the initializer
    sess.run(init)
    for epoch in range(epoch_cnt):
        avg_loss = 0.
        start = 0; end = batch_size
        
        for i in range(iteration):
            _,loss = sess.run([training_op,cost],feed_dict={X: X_train[start: end],y_true: y_train[start: end]})
            start += batch_size; end += batch_size
            # Compute average loss
            avg_loss += loss / iteration
            
        # Validate model
        preds = tf.nn.softmax(logits)  # Apply softmax to logits
        correct_prediction = tf.equal(tf.argmax(preds,1),tf.argmax(y_true,1))
        # Calculate accuracy
        accuracy = tf.reduce_mean(tf.cast(correct_prediction,"float"))
        cur_val_acc = accuracy.eval({X: X_test,y_true: y_test})
        #cur_val_acc = accuracy.eval({X: X_train,y_true: y_train})
        print("epoch: "+str(epoch)+",validation accuracy: " + str(cur_val_acc) +',loss: '+str(avg_loss))
    
    # Test model
    preds = tf.nn.softmax(logits)  # Apply softmax to logits
    correct_prediction = tf.equal(tf.argmax(preds,1))
    # Calculate accuracy
    accuracy = tf.reduce_mean(tf.cast(correct_prediction,"float"))
    print("[Test Accuracy] :",accuracy.eval({X: X_test,y_true: y_test}))

我的结果如下：

epoch: 0,validation accuracy: 0.03530512,loss: 1143.4326061579973
epoch: 1,validation accuracy: 0.48712343,loss: 173.77011137162083
epoch: 2,validation accuracy: 0.48746192,loss: 1.329379230957901
epoch: 3,validation accuracy: 0.48755372,loss: 1.2196576229016765
epoch: 4,validation accuracy: 0.4875709,loss: 1.2079436527978022
epoch: 5,validation accuracy: 0.4875652,loss: 1.2053790902122492
epoch: 6,validation accuracy: 0.48757666,loss: 1.2046689584924664
epoch: 7,loss: 1.2044714586855156
epoch: 8,validation accuracy: 0.4875824,loss: 1.2044040254991633
epoch: 9,loss: 1.2043715470603535
epoch: 10,loss: 1.2043672442216014
epoch: 11,loss: 1.204392935732021
epoch: 12,loss: 1.2043428919805674
epoch: 13,loss: 1.2043236044093304
epoch: 14,loss: 1.204309070235108
epoch: 15,loss: 1.2043108996279133
epoch: 16,loss: 1.2043230445881186
epoch: 17,loss: 1.2043319362504727
epoch: 18,loss: 1.2043379787814439
epoch: 19,loss: 1.2043409186777816
epoch: 20,loss: 1.2043399413304368
epoch: 21,loss: 1.2043371798810107
epoch: 22,loss: 1.204332315752953
epoch: 23,loss: 1.2043318411179365
epoch: 24,loss: 1.2043327070368917
epoch: 25,loss: 1.2043332901641064
epoch: 26,loss: 1.2043333211776064
epoch: 27,loss: 1.2043338380570483
epoch: 28,loss: 1.2043338463517836
epoch: 29,loss: 1.2043339615972168
[Test Accuracy] : 0.4875824