A Tensorflow convolutional autoencoder sample with pretrained models. I trined my example to learn basic emboss effect samples.
Encoder model with 3 convolutional layers to extract features vector
Model: "EncoderModel" Input and 3 convolutional layers
_________________________________________________________________
Layer (type) Output Shape Param #
_________________________________________________________________
InputLayer (InputLayer) [(None, 256, 256, 3)] 0
_________________________________________________________________
conv2d (Conv2D) (None, 256, 256, 16) 448
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 128, 128, 16) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 128, 128, 16) 2320
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 64, 64, 16) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 64, 64, 16) 2320
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 32, 32, 16) 0
_________________________________________________________________
Total params: 5,088
Trainable params: 5,088
Non-trainable params: 0
Decoder Model to reverse and upscale latent space
_________________________________________________________________
Model: "DecoderModel"
_________________________________________________________________
Layer (type) Output Shape Param #
_________________________________________________________________
LatentSpace (InputLayer) [(None, 32, 32, 16)] 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 32, 32, 16) 2320
_________________________________________________________________
up_sampling2d (UpSampling2D) (None, 64, 64, 16) 0
_________________________________________________________________
conv2d_4 (Conv2D) (None, 64, 64, 16) 2320
_________________________________________________________________
up_sampling2d_1 (UpSampling2 (None, 128, 128, 16) 0
_________________________________________________________________
conv2d_5 (Conv2D) (None, 128, 128, 16) 2320
_________________________________________________________________
up_sampling2d_2 (UpSampling2 (None, 256, 256, 16) 0
_________________________________________________________________
OutputLayer (Conv2D) (None, 256, 256, 3) 435
_________________________________________________________________
Total params: 7,395
Trainable params: 7,395
Non-trainable params: 0
Final autoencoder model that connects encoder and decoder
_________________________________________________________________
Model: "Autoencoder"
_________________________________________________________________
Layer (type) Output Shape Param #
_________________________________________________________________
InputLayer (InputLayer) [(None, 256, 256, 3)] 0
_________________________________________________________________
EncoderModel (Model) (None, 32, 32, 16) 5088
_________________________________________________________________
DecoderModel (Model) (None, 256, 256, 3) 7395
_________________________________________________________________
Total params: 12,483
Trainable params: 12,483
Non-trainable params: 0
I trained my instance of autoEncoder.h5 model to learn emboss effect. Donwloaded flower pictures from https://www.robots.ox.ac.uk/~vgg/data/flowers/102/index.html. Resized these pictures by 256x256 and stored them in img/x folder. Applied emboss effect on each image and stored them in img/y folder with same file names (emboss.py).
| Original | Embossed |
 |
 |
Runned trainer.py and trained models for 8189 image instances over 10 epochs and here is the final result.
| Actual | Target | Predicted |
 |
 |
 |