tricks in keras sequential model

Question

I have a Keras CNN model and I'm trying to recreate the same model using tensorflow v1 API. But I noticed even though the model structures are the same, their performance are very different, Keras model performance is always better. So I wonder if there is any well known tricks or optimizations enabled by default in Keras Sequential model.

The dataset I used is https://www.kaggle.com/trolukovich/food11-image-dataset. This is my Keras model and the parameters here are just something arbitrary.

model = models.Sequential()
model.add(layers.Conv2D(4, (8, 8), 
          padding='SAME',
          activation='relu', 
          input_shape=(128, 128, 3), 
          kernel_initializer='random_normal'))
model.add(layers.MaxPooling2D((2, 2),  padding='SAME'))
model.add(layers.Conv2D(8, (8, 8), 
          activation='relu', 
          padding='SAME', 
          kernel_initializer='random_normal'))
model.add(layers.MaxPooling2D((2, 2), padding='SAME'))
model.add(layers.Flatten())
model.add(layers.Dense(16, activation='relu', kernel_initializer='random_normal'))
model.add(layers.Dense(11, kernel_initializer='random_normal'))

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

And this is the model in tensorflow v1 API

def assemble_graph(g):
  with g.as_default():
    # code for input
    image_paths = tfv1.placeholder("string", [None], "image_path")
    labels = tfv1.placeholder("int32", [None], "y")

    image_paths_ts = tf.convert_to_tensor(image_paths, dtype=tf.string)
    labels_ts = tf.one_hot(labels, 11, name="y_onehot", dtype=tf.int32)
    dataset = tf.data.Dataset.from_tensor_slices((image_paths_ts, labels_ts))
    def map_fn(path, label):
      image = tf.image.decode_jpeg(tf.io.read_file(path))
      image = tf.image.resize(image, [128, 128])
      image = tfv1.to_float(image) / 255.0
      return image, label
  
    dataset = dataset.map(map_fn).shuffle(5000).batch(128)
    iter = tfv1.data.make_initializable_iterator(dataset)
    x, y = iter.get_next()

    # model
    weights = {
        'W_conv1': tf.Variable(tfv1.random_normal([8, 8, 3, 4])),
        'W_conv2': tf.Variable(tfv1.random_normal([8, 8, 4, 8])),
        'W_fc': tf.Variable(tfv1.random_normal([32*32*8, 16])),
        'out': tf.Variable(tfv1.random_normal([16, 11]))
    }

    biases = { 
        'b_conv1': tf.Variable(tfv1.zeros([4])),
        'b_conv2': tf.Variable(tfv1.zeros([8])),
        'b_fc': tf.Variable(tfv1.zeros([16])),
        'out': tf.Variable(tfv1.zeros([11])) #for the output neurons
    }
    layer1 = maxpool2d(tfv1.nn.relu(conv2d(x, weights['W_conv1']) + biases['b_conv1']), 2)
    layer2 = maxpool2d(tfv1.nn.relu(conv2d(layer1, weights['W_conv2']) + biases['b_conv2']), 2)
    layer2_flat = tf.reshape(layer2, [-1, 32*32*8])
    layer3 = tfv1.nn.relu(tf.matmul(layer2_flat, weights['W_fc'])+biases['b_fc'])
    logits = tf.matmul(layer3, weights['out'])+biases['out']

    cross_entropy = tfv1.reduce_mean(tfv1.nn.softmax_cross_entropy_with_logits_v2(labels=y, logits=logits))
    optimizer = tfv1.train.AdamOptimizer().minimize(cross_entropy)

    return optimizer, cross_entropy

One issue I found with my tfv1 implementation was that logit numbers are huge in the first a couple of iteration, which caused big loss numbers. After I applies batch norm the issue got fixed. But I didn't see big loss numbers in the Keras implementation, so this made my wonder if there is any hidden tricks that Keras Sequential model used.

Thanks.