I was working with segmentation using u-net and MobileNet.
While I trained with input size 256*256 it had an output with Val loss: 0.044 (In this time dense layer was 256, 128, 64, 32 with a learning rate of 0.0001).
But I needed to modify input for deployment to 128*128(see below code). But then the val_loss increased drastically around 0.052 which is not acceptable.
EPOCHS = 50
BATCH_SIZE = 8
LR = 0.0001
BUFFER_SIZE = 1000
VAL_SUBSPLITS = 5
SPLIT=0.3
SMOOTH = 1e-15
IMG_WIDTH=128
IMG_HEIGHT=128
# Expansive path
decode_1 = decoder_block(bridge_layer, encode_4, 256)
decode_2 = decoder_block(decode_1, encode_3, 128)
decode_3 = decoder_block(decode_2, encode_2, 64)
decode_4 = decoder_block(decode_3, encode_1, 32)
What should I take care of?
