Predicting point sequence in image

My training set is a set of images (either 3 channel or 1 ofc i use only one type of channel). And the labels are a sequence of points in a specific order that i want to predict from the images.

I am using a model inspired by the image captioning example on the tensorflow website. This is the also the approach that this paper takes https://arxiv.org/pdf/1901.03781.pdf

class CNN_Encoder(tf.keras.Model):
    # Since you have already extracted the features and dumped it using pickle
    # This encoder passes those features through a Fully connected layer
    def __init__(self, embedding_dim):
        super(CNN_Encoder, self).__init__()
        self.fc = tf.keras.layers.Dense(embedding_dim)

    def call(self, x):
        x = self.fc(x)
        x = tf.nn.relu(x)
        return x

class RNN_Decoder(tf.keras.Model):
    def __init__(self, embedding_dim, units, output_dim):
        super(RNN_Decoder, self).__init__()
        self.units = units


        self.gru = tf.keras.layers.GRU(self.units,
                                       return_sequences=True,
                                       return_state=True,
                                       recurrent_initializer='glorot_uniform')
        self.fc1 = tf.keras.layers.Dense(self.units)
        self.fc2 = tf.keras.layers.Dense(output_dim)

    def call(self, x, features, hidden):


        x = tf.concat((features, x), axis=-1)
        output, state = self.gru(x)
        x = self.fc1(state)
        x = self.fc2(x)
        return x

    def reset_state(self, batch_size):
        return tf.zeros((batch_size, self.units))

@tf.function
def train_step(img_tensor, target):
    loss = 0


    hidden = decoder.reset_state(batch_size=target.shape[0])
    dec_input = tf.expand_dims([[0., 0.]] * target.shape[0], 1)
    with tf.GradientTape() as tape:

        features = encoder(img_tensor)
        for i in (range(1, target.shape[1])):
            predictions = decoder(dec_input, features, hidden)
            loss += loss_function(target[:, i], predictions)

            # using teacher forcing
            dec_input = tf.expand_dims(target[:, i], 1)
    total_loss = (loss / int(target.shape[1]))
    trainable_variables = encoder.trainable_variables + decoder.trainable_variables
    gradients = tape.gradient(loss, trainable_variables)
    optimizer.apply_gradients(zip(gradients, trainable_variables))
    return loss, total_loss

EPOCHS = 20
batch_size = 8
for epoch in tqdm(range(start_epoch, EPOCHS)):
    start = time.time()
    total_loss = 0

    for (batch, (img_tensor, target)) in enumerate((data_generator(preds_t, labels_t))):
        img_tensor = img_tensor.reshape((-1, 1, 128*128))
        batch_loss, t_loss = train_step(img_tensor, target)
        total_loss += t_loss

        if batch % 100 == 0:
            print ('Epoch {} Batch {} Loss {:.4f}'.format(
              epoch + 1, batch, batch_loss.numpy() / int(target.shape[1])))
        if batch == 10000:

            break
    # storing the epoch end loss value to plot later
    #loss_plot.append(total_loss / num_steps)

    if epoch % 5 == 0:
        ckpt_manager.save()

    print ('Epoch {} Loss {:.6f}'.format(epoch + 1,
                                         total_loss/num_steps))
    print ('Time taken for 1 epoch {} secn'.format(time.time() - start))

For the features vector. I am extracting the last layer of a unet. So each image has a size 1x128x128. I reshape it to be 1x1x128*128. Which i then pass through a fully connected layer. The shape then becomes 1x1x256

My labels i want to predict are image coordinates so (x, y). The input to the gru layer is the concatenated 1x1x256 , 1x1x2 (t-1 coordinates). Which i then further pass through a 2 layer fc layer with output dimension 2 for the 2 coordinates. I have removed attention for now to get a simpler model. I normalize my images. I pad the coordinate sequences with 0,0 for the start -1, -1 for the end and -2,-2 for the regular padding to get uniform sequence length of 350×2.

The network doesnt seem to learn much. I just get a few points scattered diagonally across the image. The biggest difference i see with the image captioning model is that the words can be converted to embeddings and then you have a 128 image features 128 word features being concatenated and fed into the lstm. In my case the sequence information is just 1 entry. Could that be the reason that the network is not learning much.

If someone has any insights into what i should change that would be great