Why pure tensorflow autoencoder cannot converge, when Keras one fits good - tensorflow

Question

I'm newbie at machine learning(and at stackoverflow too). I want to ask for help.
I have two implementations of same two-layer autoencoder for mnist.
First one fits good:
import tensorflow as tf, numpy as np
def in_pics(pics):
return np.array(pics, np.float32)/255.
def out_pics(pics):
npformed = np.array(pics, np.float32)
samples = np.shape(npformed)[0]
return np.reshape(npformed, (samples, 784))/255.
(train_data, test_data) = tf.keras.datasets.mnist.load_data()
train_dataset = tf.data.Dataset.from_tensor_slices((in_pics(train_data[0]), out_pics(train_data[0]))).batch(100)
test_dataset = tf.data.Dataset.from_tensor_slices((in_pics(test_data[0]), out_pics(test_data[0]))).batch(100)
model = tf.keras.Sequential(
[
tf.keras.layers.InputLayer(input_shape=(28,28)),
tf.keras.layers.Reshape(target_shape=(784,)),
tf.keras.layers.Dense(128, activation="sigmoid", kernel_initializer=tf.keras.initializers.truncated_normal(stddev=0.1),
bias_initializer =tf.keras.initializers.truncated_normal(stddev=0.1)),
tf.keras.layers.Dense(784, activation="sigmoid", kernel_initializer=tf.keras.initializers.truncated_normal(stddev=0.1),
bias_initializer =tf.keras.initializers.truncated_normal(stddev=0.1)),
]
)
model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=True), optimizer="adam")
model.fit(x = train_dataset, epochs=20, verbose=1, validation_data=test_dataset)
Keras result
And I have the same model, same learning procedure, but written in terms of pure tensorflow(almost):
import tensorflow as tf, numpy as np
def prep_pics(pics):
npformed = np.array(pics, np.float32)
samples = np.shape(npformed)[0]
return np.reshape(npformed, (samples, 784))/255.
batch_size, num_batches = 100, 12000
(train_data, test_data) = tf.keras.datasets.mnist.load_data()
(train_pics_prepared, test_pics_prepared) = (prep_pics(train_data[0]), prep_pics(test_data[0]))
train_dataset = tf.data.Dataset.from_tensor_slices((train_pics_prepared, train_pics_prepared))
test_dataset = tf.data.Dataset.from_tensor_slices((test_pics_prepared, test_pics_prepared)).batch(batch_size)
encoder_W = tf.Variable(tf.keras.initializers.truncated_normal(stddev=0.1)(shape = [784,128], dtype = tf.float32))
encoder_b = tf.Variable(tf.keras.initializers.truncated_normal(stddev=0.1)(shape = [128],dtype = tf.float32))
decoder_W = tf.Variable(tf.keras.initializers.truncated_normal(stddev=0.1)(shape = [128,784], dtype = tf.float32))
decoder_b = tf.Variable(tf.keras.initializers.truncated_normal(stddev=0.1)(shape = [784],dtype = tf.float32))
optimizer = tf.keras.optimizers.Adam()
bin_cross = tf.keras.losses.BinaryCrossentropy(from_logits=True)
#tf.function
def trainstep(X_pack, Y_pack):
with tf.GradientTape() as tape:
tape.watch([encoder_W, encoder_b, decoder_W, decoder_b])
encoded = tf.nn.sigmoid(tf.matmul(X_pack, encoder_W) + encoder_b)
decoded = tf.nn.sigmoid(tf.matmul(encoded, decoder_W) + decoder_b)
loss = bin_cross(decoded, Y_pack)
gradients = tape.gradient(loss, [encoder_W, encoder_b, decoder_W, decoder_b])
optimizer.apply_gradients(zip(gradients, [encoder_W, encoder_b, decoder_W, decoder_b]))
num_samples = len(train_data[0])
epochs = num_batches//(num_samples//batch_size)
for i in range(epochs):
print(f"Epoch num {i+1}:")
dataset = train_dataset.shuffle(num_samples).batch(batch_size)
for x,y in dataset:
trainstep(x,y)
Loss is never becomes lower than 0.64(First one had 0.07 in the end of education).
Tensorflow result.
What I tried:
1.) Change optimizer to Adagrad and even to SGD.
2.) Use other loss - sigmoid_cross_entropy_with_logits.
3.) Learn more. Especially for SGD.
Cannot find mistake or typo almost for week. Please, help!