I am using the github ESRGAN of https://github.com/peteryuX/esrgan-tf2/blob/master/train_esrgan.py in tensorflow, and my inputs and outputs have the same size. I already have low_resolution and high_resolution images.
However, when using the code, the generator multiplies by 4 the image, which I want it to be the exact same size. So I can't re insert the generated image in the discriminator.
For example my batch if of size (8, 128,128, 1), and when passing in the generator, it output (16, 512, 512, 1). The discriminator takes a size of (None, 128, 128,1).
I would like that it doesn't upsample, but I feel that it is only what it does ...
The generator is :
def RRDB_Model(size, channels, cfg_net, gc=32, wd=0., name='RRDB_model'):
"""Residual-in-Residual Dense Block based Model """
nf, nb = cfg_net['nf'], cfg_net['nb']
lrelu_f = functools.partial(LeakyReLU, alpha=0.2)
rrdb_f = functools.partial(ResInResDenseBlock, nf=nf, gc=gc, wd=wd)
conv_f = functools.partial(Conv2D, kernel_size=3, padding='same',
bias_initializer='zeros',
kernel_initializer=_kernel_init(),
kernel_regularizer=_regularizer(wd))
rrdb_truck_f = tf.keras.Sequential(
[rrdb_f(name="RRDB_{}".format(i)) for i in range(nb)],
name='RRDB_trunk')
# extraction
x = inputs = Input([size, size, channels], name='input_image')
fea = conv_f(filters=nf, name='conv_first')(x)
fea_rrdb = rrdb_truck_f(fea)
trunck = conv_f(filters=nf, name='conv_trunk')(fea_rrdb)
fea = fea + trunck
# upsampling
size_fea_h = tf.shape(fea)[1] if size is None else size
size_fea_w = tf.shape(fea)[2] if size is None else size
fea_resize = tf.image.resize(fea, [size_fea_h * 2, size_fea_w * 2],
method='nearest', name='upsample_nn_1')
fea = conv_f(filters=nf, activation=lrelu_f(), name='upconv_1')(fea_resize)
fea_resize = tf.image.resize(fea, [size_fea_h * 4, size_fea_w * 4],
method='nearest', name='upsample_nn_2')
fea = conv_f(filters=nf, activation=lrelu_f(), name='upconv_2')(fea_resize)
fea = conv_f(filters=nf, activation=lrelu_f(), name='conv_hr')(fea)
out = conv_f(filters=channels, name='conv_last')(fea)
return Model(inputs, out, name=name)
EDIT :
I deleted these lines :
fea_resize = tf.image.resize(fea, [size_fea_h * 2, size_fea_w * 2], method='nearest', name='upsample_nn_1')
fea_resize = tf.image.resize(fea, [size_fea_h * 4, size_fea_w * 4],method='nearest',name='upsample_nn_2')
and kept the convolutions layers "fea"
It worked !
