Shearlets #40

kevinmicha · 2021-06-17T08:34:16Z

Now it's possible to add shearlet fixed filters using the alpha-transform package.

zaccharieramzi

Let's make the shearlet addition an option and not something you need to use.

And before merging, we should fix the high-pass bug.

learning_wavelets/models/learnlet_layers.py

zaccharieramzi · 2021-06-17T09:29:02Z

learning_wavelets/models/learnlet_layers.py

                self.kernel_size,
                activation='linear',
                padding='same',
                kernel_initializer='glorot_uniform',
                use_bias=tiling_use_bias,
                kernel_constraint=constraint,
+                trainable = False,
                name=f'{tiling_prefix}_{str(K.get_uid(tiling_prefix))}',


Suggested change

name=f'{tiling_prefix}_{str(K.get_uid(tiling_prefix))}',

name=f'{tiling_prefix}_fixed_{str(K.get_uid(tiling_prefix))}',

@kevinmicha I don't understand why you resolved this

was a mistake. Now really solved

zaccharieramzi · 2021-06-17T09:31:23Z

learning_wavelets/models/learnlet_layers.py

-            details_tiled = self.convs_detail_tiling[i_scale](wav_detail)
+            details_tiled_fixed = self.convs_detail_tiling_fixed[i_scale](wav_detail)
+            details_tiled_train = self.convs_detail_tiling_train[i_scale](wav_detail)
+            details_tiled = Concatenate()([details_tiled_fixed, details_tiled_train])


do not use the layer Concatenate here: it will create a layer at inference time which we want to avoid esp. when running in eager mode.

I think it's best to use tf.concat

same for below

zaccharieramzi · 2021-06-17T09:32:09Z

learning_wavelets/training_scripts/learnlet_training.py

@@ -2,12 +2,16 @@
 import os.path as op
 import time

+import cadmos_lib as cl


Let's make this import optional

learning_wavelets/training_scripts/learnlet_training.py

zaccharieramzi · 2021-06-17T09:33:28Z

learning_wavelets/training_scripts/learnlet_training.py

+        inputs = [tf.zeros((1, 32, 32, 1)), tf.zeros((1, 1))]
+        model(inputs)
+
+    shearlets, _ = cl.get_shearlets(512, 512, 6)


Put the answer in comment if it makes sense

zaccharieramzi · 2021-06-17T09:35:43Z

learning_wavelets/training_scripts/learnlet_training.py

+
+    n_shearlets = np.shape(shearlets)[0]
+    total_size = np.shape(shearlets)[1]
+    half_filters = n_filters//2


Wait I don't get it...
Does it mean that you do not always consider all the shearlet filters?
I think this is a mistake since you can not just take the shearlet filters you want, you need them all.
Surely there is a way to require less shearlet filters though

zaccharieramzi · 2021-06-17T09:37:21Z

learning_wavelets/training_scripts/learnlet_training.py

+    filter_size = analysis_kernel_size
+    crop_min = total_size//2 - filter_size//2
+    crop_max = total_size//2 + filter_size//2 + 1
+    resized_shearlets = np.zeros((5, filter_size, filter_size, 1, half_filters))


Surely here it's not 5 but n_scales or sthg similar.
Same for below the 6,11 shouldn't be hardcoded

changed verbose Co-authored-by: Zaccharie Ramzi <[email protected]>

trainable param adapted to pep8 Co-authored-by: Zaccharie Ramzi <[email protected]>

Co-authored-by: Zaccharie Ramzi <[email protected]>

zaccharieramzi

There is indeed a bug in the shearlets part

zaccharieramzi · 2021-07-23T10:35:49Z

learning_wavelets/models/learnlet_layers.py

                self.kernel_size,
                activation='linear',
                padding='same',
                kernel_initializer='glorot_uniform',
                use_bias=tiling_use_bias,
                kernel_constraint=constraint,
+                trainable = False,
                name=f'{tiling_prefix}_{str(K.get_uid(tiling_prefix))}',


@kevinmicha I don't understand why you resolved this

zaccharieramzi · 2021-07-23T10:36:37Z

learning_wavelets/models/learnlet_layers.py

        if self.mixing_details:
            mixing_prefix = 'details_mixing'
-            self.convs_detail_mixing = [
+            self.convs_detail_mixing_fixed = [


Suggested change

self.convs_detail_mixing_fixed = [

self.convs_detail_mixing = [

I think this part doesn't need to be changed

zaccharieramzi · 2021-07-23T10:38:27Z

learning_wavelets/models/learnlet_layers.py

@@ -203,7 +223,7 @@ def get_config(self):

 class LearnletSynthesis(Layer):
    __name__ = 'learnlet_synthesis'
-    def __init__(self, normalize=True, n_scales=4, n_channels=1, synthesis_use_bias=False, synthesis_norm=False, res=False, kernel_size=5, wav_type='starlet'):
+    def __init__(self, normalize=True, n_scales=5, n_tiling=256, n_channels=1, synthesis_use_bias=False, synthesis_norm=False, res=True, kernel_size=13, wav_type='starlet', n_shearlets=85):


Let's not change the defaults you don't need for backward compatibility's sake

Suggested change

def __init__(self, normalize=True, n_scales=5, n_tiling=256, n_channels=1, synthesis_use_bias=False, synthesis_norm=False, res=True, kernel_size=13, wav_type='starlet', n_shearlets=85):

def __init__(self, normalize=True, n_scales=4, n_tiling=256, n_channels=1, synthesis_use_bias=False, synthesis_norm=False, res=False, kernel_size=5, wav_type='starlet', n_shearlets=85):

zaccharieramzi · 2021-07-23T10:39:12Z

learning_wavelets/models/learnlet_layers.py

+                use_bias=synthesis_use_bias,
+                kernel_constraint=constraint,
+                trainable=False,
+                name=f'{groupping_prefix}_{str(K.get_uid(groupping_prefix))}',


Suggested change

name=f'{groupping_prefix}_{str(K.get_uid(groupping_prefix))}',

name=f'{groupping_prefix}_fixed_{str(K.get_uid(groupping_prefix))}',

now it's fine

zaccharieramzi · 2021-07-23T10:41:01Z

learning_wavelets/models/learnlet_layers.py

+                padding='same',
+                kernel_initializer='glorot_uniform',
+                use_bias=tiling_use_bias,
+                kernel_constraint=constraint, 


This might be a bug: you don't want to use the unit norm constraint on shearlet filters, because they should stay as is.

This constraint might make them evolve during training.

Suggested change

kernel_constraint=constraint,

kernel_constraint=None,

I also removed the constraint for the synthesis shearlet filters

zaccharieramzi · 2021-07-23T10:45:50Z

learning_wavelets/models/learnlet_layers.py

+                image = self.convs_groupping_fixed[i_scale](detail[...,0:self.n_shearlets]) + image
+                image = self.convs_groupping_train[i_scale](detail[...,self.n_shearlets:self.n_tiling]) + image


Suggested change

image = self.convs_groupping_fixed[i_scale](detail[...,0:self.n_shearlets]) + image

image = self.convs_groupping_train[i_scale](detail[...,self.n_shearlets:self.n_tiling]) + image

image = self.convs_groupping_fixed[i_scale](detail[..., :self.n_shearlets]) + image

image = self.convs_groupping_train[i_scale](detail[...,self.n_shearlets:]) + image

We had mentioned this bug actually.

Here the problem is that the total number of details is n_shearlets + n_tiling.
Let's do a quick example, imagine detail = [0, 1, 2, 3, 4, 5], n_shearlets = 2 and n_tiling = 4.
Then we have:

details[0:n_shearlets] == details[0:2] == [0, 1] details[n_shearlets:n_tiling] = details[2:4] = [2, 3]

This also means that you don't need to store n_tiling in this class.

this has been discussed

kevinmicha added 8 commits April 26, 2021 19:09

Update unet.py with exact reconstruction

bd4ea22

Back to previous version

baa8d4f

Merge branch 'zaccharieramzi:master' into master

4499a95

modified analysis class to allow fixed filters

9d88a3a

including shearlets and exact_recon parameter

fadd9af

Added extra dot

774e972

adapted tiling part

ee1c145

fixed indentation for tiling func

42b7674

zaccharieramzi requested changes Jun 17, 2021

View reviewed changes

kevinmicha and others added 19 commits June 17, 2021 11:40

Update learning_wavelets/training_scripts/learnlet_training.py

3b8dc98

changed verbose Co-authored-by: Zaccharie Ramzi <[email protected]>

Update learning_wavelets/models/learnlet_layers.py

3e5652d

trainable param adapted to pep8 Co-authored-by: Zaccharie Ramzi <[email protected]>

changed layer name

e8f1649

Co-authored-by: Zaccharie Ramzi <[email protected]>

back to original version

678b328

including synthesis filters part

0e0eb29

optional import: cadmos

f7bc4a0

removed wrong indent

beedb0b

used kernel's keywords

6f249cf

added spaces back

a0589a7

added untrainable layers

3139c2b

fixed white space

9a923d3

not prepared for exact recon True

8ff2877

changed n_tiling. Doesn't make sense 3 as default

6f8c3f0

removed last change

5af5fd2

changed some default values for testing matters

3001d00

more def values changed

a61f1ef

removing exact recon test from this branch. No sense and tests fail

07ee9b1

exact recon unet test is back

5d928be

Removed the part that tests exact recon. No sense

19b021a

kevinmicha requested a review from zaccharieramzi July 23, 2021 09:31

zaccharieramzi requested changes Jul 23, 2021

View reviewed changes

added fixed keyword in names

7e061ee

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Shearlets #40

Shearlets #40

kevinmicha commented Jun 17, 2021

zaccharieramzi left a comment •

edited

Loading

zaccharieramzi Jun 17, 2021

zaccharieramzi Jul 23, 2021

kevinmicha Jul 23, 2021

zaccharieramzi Jun 17, 2021

zaccharieramzi Jun 17, 2021

kevinmicha Jul 23, 2021

zaccharieramzi Jun 17, 2021

kevinmicha Jul 23, 2021

zaccharieramzi Jun 17, 2021

zaccharieramzi Jun 17, 2021

kevinmicha Jul 23, 2021

zaccharieramzi Jun 17, 2021

kevinmicha Jul 23, 2021

zaccharieramzi Jun 17, 2021

kevinmicha Jul 23, 2021

zaccharieramzi left a comment

zaccharieramzi Jul 23, 2021

zaccharieramzi Jul 23, 2021

kevinmicha Jul 23, 2021

zaccharieramzi Jul 23, 2021

zaccharieramzi Jul 23, 2021

kevinmicha Jul 23, 2021

zaccharieramzi Jul 23, 2021

zaccharieramzi Jul 23, 2021

kevinmicha Jul 23, 2021

kevinmicha Jul 23, 2021 •

edited

Loading

zaccharieramzi Jul 23, 2021

kevinmicha Jul 23, 2021

	name=f'{tiling_prefix}_{str(K.get_uid(tiling_prefix))}',
	name=f'{tiling_prefix}_fixed_{str(K.get_uid(tiling_prefix))}',

	self.convs_detail_mixing_fixed = [
	self.convs_detail_mixing = [

	def __init__(self, normalize=True, n_scales=5, n_tiling=256, n_channels=1, synthesis_use_bias=False, synthesis_norm=False, res=True, kernel_size=13, wav_type='starlet', n_shearlets=85):
	def __init__(self, normalize=True, n_scales=4, n_tiling=256, n_channels=1, synthesis_use_bias=False, synthesis_norm=False, res=False, kernel_size=5, wav_type='starlet', n_shearlets=85):

	name=f'{groupping_prefix}_{str(K.get_uid(groupping_prefix))}',
	name=f'{groupping_prefix}_fixed_{str(K.get_uid(groupping_prefix))}',

		image = self.convs_groupping_fixed[i_scale](detail[...,0:self.n_shearlets]) + image
		image = self.convs_groupping_train[i_scale](detail[...,self.n_shearlets:self.n_tiling]) + image

Shearlets #40

Are you sure you want to change the base?

Shearlets #40

Conversation

kevinmicha commented Jun 17, 2021

zaccharieramzi left a comment • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zaccharieramzi left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kevinmicha Jul 23, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zaccharieramzi left a comment •

edited

Loading

kevinmicha Jul 23, 2021 •

edited

Loading