Encoder (Dense connection without flatten) & Decoder input does not have reshape

[anugrahasinha]

@VeeranjaneyuluToka

Please let me know if my understanding is correct here or not?

Encoder - Dense connection without Flatten

class ConvAutoEncoder(HyperModel):
    def __init__(self, input_shape, is_deeper=False):
--snip--
    """ encoder """
    def encoder(self, input_img, hp):

        conv1 = tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same')(input_img) #224 x 224 x 32
        conv1 = tf.keras.layers.BatchNormalization()(conv1)
        conv1 = tf.keras.layers.Conv2D(32, (3, 3), activation='relu', padding='same')(conv1)
--snip--
        conv6 = tf.keras.layers.Conv2D(512, (3, 3), activation='relu', padding='same')(conv5)
        conv6 = tf.keras.layers.BatchNormalization()(conv6)
        conv6 = tf.keras.layers.Conv2D(512, (3, 3), activation='relu', padding='same')(conv6)
        conv6 = tf.keras.layers.BatchNormalization()(conv6)

        # Comment (asinha) : conv6 output would be (14,14,512)
        # However, we do not flatten this before building a dense below this
        # would this be correct?

        dense = tf.keras.layers.Dense(16, activation='relu', name='encoder')(conv6)

Decoder : Input does not reshape the encoded input before applying a Conv layer

class ConvAutoEncoder(HyperModel):
    def __init__(self, input_shape, is_deeper=False):
--snip--
    """ decoder """
    def decoder(self, dense, hp):

        # Comment (asinha) : If we are getting a dense input, which I am believing a vector if 16 (as 
        # per last layer at encoder), needs to be reshaped before we apply a CONV2D(256)
        # Also I do not understand why do we have #56x56x128 added ahead of conv5?

        conv5 = tf.keras.layers.Conv2D(256, (3, 3), activation='relu', padding='same')(dense) #56 x 56 x 128
        conv5 = tf.keras.layers.BatchNormalization()(conv5)
        conv5 = tf.keras.layers.Conv2D(256, (3, 3), activation='relu', padding='same')(conv5)
        conv5 = tf.keras.layers.BatchNormalization()(conv5)

[VeeranjaneyuluToka]

Really do not matter looks like though i do not have answers what advantage it has with or without flatten. There are some cases where implementation do not use Flatten but did not state any reason behind that for example YOLO (source https://datascience.stackexchange.com/questions/27135/should-there-be-a-flat-layer-in-between-the-conv-layers-and-dense-layer-in-yolo). Here is the summary of my network around dense layer

conv2d_9 (Conv2D) (None, 28, 28, 512) 2359808

---

batch_normalization_9 (Batch (None, 28, 28, 512) 2048

---

max_pooling2d_3 (MaxPooling2 (None, 14, 14, 512) 0

---

conv2d_10 (Conv2D) (None, 14, 14, 512) 2359808

---

batch_normalization_10 (Batc (None, 14, 14, 512) 2048

---

conv2d_11 (Conv2D) (None, 14, 14, 512) 2359808

---

batch_normalization_11 (Batc (None, 14, 14, 512) 2048

---

encoder (Dense) (None, 14, 14, 16) 8208

---

conv2d_12 (Conv2D) (None, 14, 14, 256) 37120

---

batch_normalization_12 (Batc (None, 14, 14, 256) 1024

---

conv2d_13 (Conv2D) (None, 14, 14, 256) 590080

---

batch_normalization_13 (Batc (None, 14, 14, 256) 1024

---

conv2d_14 (Conv2D) (None, 14, 14, 128) 295040

---

batch_normalization_14 (Batc (None, 14, 14, 128) 512

---

conv2d_15 (Conv2D) (None, 14, 14, 128) 147584

---

batch_normalization_15 (Batc (None, 14, 14, 128) 512

---

up_sampling2d (UpSampling2D) (None, 28, 28, 128) 0

---

conv2d_16 (Conv2D) (None, 28, 28, 64) 73792

However i have experimented with Flatten() and reshape() and i do not see any change using them in accuracy but there is some difference in no.of parameters. However i am yet to conclude on this and post my observations on this point as soon as i find.

Thank you very much for point out this.