optimization of multi-regression CNN architecture

Question

I use T4 GPU on Google Colab to train a model for multi regression. I use to training 30k RGB images 256x256, 5k to validate, 7k to evaluate. The model has 11 outputs [0-1] range (the sum of outputs [y1, y2 ... y11] is always equal 1, eg. 0 + 0 + 0 + 0.5 + 0.3 + 0.2 .... = 1). After the datasets preparing and the model compiling GPU VRAM utilization = 2.4/15 GB. During fitting VRAM utilization = 14.2/15 GB. It is near to max.

• Could be there a problem with training performance? Why?
• Are there any big mistakes in my architecture or hiperparams?
• Is there a better way to load this data to training?

I'm a CNN newbie - any advice welcome.

Images are loaded by this piece of code from my Google Drive

def preproc_image_and_label(x, y):
  img = tf.io.read_file(x)
  img = tf.io.decode_jpeg(img, channels=3)
  img.set_shape((256, 256, 3))
  img = tf.cast(img, dtype=tf.float32)
  img = img / 255.
  return img, y


tf.data.Dataset.from_tensor_slices((tf.constant(glob("drive/MyDrive/path/to/img/train/")), tf.constant(y_train.values))).map(preproc_image_and_label).batch(32).prefetch(1)

The model is compiled with:

layers = [
    Conv2D(filters=128, kernel_size=5, activation="relu", input_shape=(256, 256, 3), padding="same"),
    MaxPooling2D(),
    Conv2D(filters=256, kernel_size=4, activation="relu", padding="same"),
    Conv2D(filters=256, kernel_size=4, activation="relu", padding="same"),
    MaxPooling2D(),
    Conv2D(filters=512, kernel_size=3, activation="relu", padding="same"),
    Conv2D(filters=512, kernel_size=3, activation="relu", padding="same"),
    MaxPooling2D(),
    Flatten(),
    Dense(256, activation="relu"),
    Dense(128, activation="relu"),
    Dense(11, activation="softmax"),
]

model.compile(
    optimizer=keras.optimizers.Adam(learning_rate=0.001),
    loss=keras.losses.MeanSquaredError(),
    metrics=[keras.metrics.MeanSquaredError()]
  )