SAEHD:

Maximum resolution is increased to 640.

‘hd’ archi is removed. ‘hd’ was experimental archi created to remove subpixel shake, but ‘lr_dropout’ and ‘disable random warping’ do that better.

‘uhd’ is renamed to ‘-u’
dfuhd and liaeuhd will be automatically renamed to df-u and liae-u in existing models.

Added new experimental archi (key -d) which doubles the resolution using the same computation cost.
It is mean same configs will be x2 faster, or for example you can set 448 resolution and it will train as 224.
Strongly recommended not to train from scratch and use pretrained models.

New archi naming:
'df' keeps more identity-preserved face.
'liae' can fix overly different face shapes.
'-u' increased likeness of the face.
'-d' (experimental) doubling the resolution using the same computation cost
Examples: df, liae, df-d, df-ud, liae-ud, ...

Improved GAN training (GAN_power option).  It was used for dst model, but actually we don’t need it for dst.
Instead, a second src GAN model with x2 smaller patch size was added, so the overall quality for hi-res models should be higher.

Added option ‘Uniform yaw distribution of samples (y/n)’:
	Helps to fix blurry side faces due to small amount of them in the faceset.

Quick96:
	Now based on df-ud archi and 20% faster.

XSeg trainer:
	Improved sample generator.
Now it randomly adds the background from other samples.
Result is reduced chance of random mask noise on the area outside the face.
Now you can specify ‘batch_size’ in range 2-16.

Reduced size of samples with applied XSeg mask. Thus size of packed samples with applied xseg mask is also reduced.

samplelib/SampleGeneratorFace.py

@@ -6,11 +6,13 @@ import cv2
 import numpy as np
 
 from core import mplib
+from core.interact import interact as io
 from core.joblib import SubprocessGenerator, ThisThreadGenerator
 from facelib import LandmarksProcessor
 from samplelib import (SampleGeneratorBase, SampleLoader, SampleProcessor,
                        SampleType)
 
+
 '''
 arg
 output_sample_types = [
@@ -23,15 +25,15 @@ class SampleGeneratorFace(SampleGeneratorBase):
                         random_ct_samples_path=None,
                         sample_process_options=SampleProcessor.Options(),
                         output_sample_types=[],
-                        add_sample_idx=False,
+                        uniform_yaw_distribution=False,
                         generators_count=4,
-                        raise_on_no_data=True,
+                        raise_on_no_data=True,                        
                         **kwargs):
 
         super().__init__(debug, batch_size)
+        self.initialized = False
         self.sample_process_options = sample_process_options
         self.output_sample_types = output_sample_types
-        self.add_sample_idx = add_sample_idx
         
         if self.debug:
             self.generators_count = 1
@@ -40,15 +42,40 @@ class SampleGeneratorFace(SampleGeneratorBase):
 
         samples = SampleLoader.load (SampleType.FACE, samples_path)
         self.samples_len = len(samples)
-
-        self.initialized = False
+        
         if self.samples_len == 0:
             if raise_on_no_data:
                 raise ValueError('No training data provided.')
             else:
                 return
+                
+        if uniform_yaw_distribution:
+            samples_pyr = [ ( idx, sample.get_pitch_yaw_roll() ) for idx, sample in enumerate(samples) ]
+            
+            grads = 128
+            #instead of math.pi / 2, using -1.2,+1.2 because actually maximum yaw for 2DFAN landmarks are -1.2+1.2
+            grads_space = np.linspace (-1.2, 1.2,grads)
 
-        index_host = mplib.IndexHost(self.samples_len)
+            yaws_sample_list = [None]*grads
+            for g in io.progress_bar_generator ( range(grads), "Sort by yaw"):
+                yaw = grads_space[g]
+                next_yaw = grads_space[g+1] if g < grads-1 else yaw
+
+                yaw_samples = []
+                for idx, pyr in samples_pyr:
+                    s_yaw = -pyr[1]
+                    if (g == 0          and s_yaw < next_yaw) or \
+                    (g < grads-1     and s_yaw >= yaw and s_yaw < next_yaw) or \
+                    (g == grads-1    and s_yaw >= yaw):
+                        yaw_samples += [ idx ]
+                if len(yaw_samples) > 0:
+                    yaws_sample_list[g] = yaw_samples
+            
+            yaws_sample_list = [ y for y in yaws_sample_list if y is not None ]
+            
+            index_host = mplib.Index2DHost( yaws_sample_list )
+        else:
+            index_host = mplib.IndexHost(self.samples_len)
 
         if random_ct_samples_path is not None:
             ct_samples = SampleLoader.load (SampleType.FACE, random_ct_samples_path)
@@ -110,14 +137,8 @@ class SampleGeneratorFace(SampleGeneratorBase):
 
                 if batches is None:
                     batches = [ [] for _ in range(len(x)) ]
-                    if self.add_sample_idx:
-                        batches += [ [] ]
-                        i_sample_idx = len(batches)-1
 
                 for i in range(len(x)):
                     batches[i].append ( x[i] )
 
-                if self.add_sample_idx:
-                    batches[i_sample_idx].append (sample_idx)
-
             yield [ np.array(batch) for batch in batches]