Merge remote-tracking branch 'iperov/master'

DFLIMG/DFLJPG.py

@@ -281,6 +281,13 @@ class DFLJPG(object):
     def has_xseg_mask(self):
         return self.dfl_dict.get('xseg_mask',None) is not None
 
+    def get_xseg_mask_compressed(self):
+        mask_buf = self.dfl_dict.get('xseg_mask',None)
+        if mask_buf is None:
+            return None
+
+        return mask_buf
+        
     def get_xseg_mask(self):
         mask_buf = self.dfl_dict.get('xseg_mask',None)
         if mask_buf is None:
@@ -301,7 +308,7 @@ class DFLJPG(object):
         mask_a = imagelib.normalize_channels(mask_a, 1)
         img_data = np.clip( mask_a*255, 0, 255 ).astype(np.uint8)
 
-        data_max_len = 4096
+        data_max_len = 8192
 
         ret, buf = cv2.imencode('.png', img_data)
 

README.md

@@ -29,11 +29,11 @@ More than 95% of deepfake videos are created with DeepFaceLab.
 
 DeepFaceLab is used by such popular youtube channels as
 
-|![](doc/youtube_icon.png) [Ctrl Shift Face](https://www.youtube.com/channel/UCKpH0CKltc73e4wh0_pgL3g)|![](doc/youtube_icon.png) [VFXChris Ume](https://www.youtube.com/channel/UCGf4OlX_aTt8DlrgiH3jN3g/videos)|
+|![](doc/youtube_icon.png) [VFXChris Ume](https://www.youtube.com/channel/UCGf4OlX_aTt8DlrgiH3jN3g/videos)|![](doc/youtube_icon.png) [Sham00k](https://www.youtube.com/channel/UCZXbWcv7fSZFTAZV4beckyw/videos)|
 |---|---|
 
-|![](doc/youtube_icon.png) [Sham00k](https://www.youtube.com/channel/UCZXbWcv7fSZFTAZV4beckyw/videos)|![](doc/youtube_icon.png) [Collider videos](https://www.youtube.com/watch?v=A91P2qtPT54&list=PLayt6616lBclvOprvrC8qKGCO-mAhPRux)|![](doc/youtube_icon.png) [iFake](https://www.youtube.com/channel/UCC0lK2Zo2BMXX-k1Ks0r7dg/videos)|
+|![](doc/youtube_icon.png) [Collider videos](https://www.youtube.com/watch?v=A91P2qtPT54&list=PLayt6616lBclvOprvrC8qKGCO-mAhPRux)|![](doc/youtube_icon.png) [iFake](https://www.youtube.com/channel/UCC0lK2Zo2BMXX-k1Ks0r7dg/videos)|
-|---|---|---|
+|---|---|
 
 |![](doc/youtube_icon.png) [NextFace](https://www.youtube.com/channel/UCFh3gL0a8BS21g-DHvXZEeQ/videos)|![](doc/youtube_icon.png) [Futuring Machine](https://www.youtube.com/channel/UCC5BbFxqLQgfnWPhprmQLVg)|![](doc/youtube_icon.png) [RepresentUS](https://www.youtube.com/channel/UCRzgK52MmetD9aG8pDOID3g)|
 |---|---|---|
@@ -317,7 +317,7 @@ QQ 951138799
 </td></tr>
 
 <tr><td colspan=2 align="center">
-<a href="https://www.paypal.com/cgi-bin/webscr?cmd=_donations&business=lepersorium@gmail.com&lc=US&no_note=0&item_name=Support+DeepFaceLab&cn=&curency_code=USD&bn=PP-DonationsBF:btn_donateCC_LG.gif:NonHosted">Donate via Paypal</a>
+<a href="https://www.paypal.com/paypalme/DeepFaceLab">Donate via Paypal</a>
 </td></tr>
 
 <tr><td colspan=2 align="center">

core/leras/nn.py

@@ -70,19 +70,23 @@ class nn():
                         first_run = True
                     os.environ['CUDA_CACHE_PATH'] = str(compute_cache_path)
 
-            os.environ['CUDA_​CACHE_​MAXSIZE'] = '536870912' #512Mb (32mb default)
+            #nvcuda.dll ignores this param : os.environ['CUDA_​CACHE_​MAXSIZE'] = '536870912' #512Mb (32mb default)
             os.environ['TF_MIN_GPU_MULTIPROCESSOR_COUNT'] = '2'
-            os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' # tf log errors only
+            os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' # tf log errors only
 
             if first_run:
                 io.log_info("Caching GPU kernels...")
 
-            import tensorflow as tf
+            #import tensorflow as tf
-            nn.tf = tf
+            import tensorflow.compat.v1 as tf
             
             import logging
             # Disable tensorflow warnings
-            logging.getLogger('tensorflow').setLevel(logging.ERROR)
+            tf_logger = logging.getLogger('tensorflow')
+            tf_logger.setLevel(logging.ERROR)
+            
+            tf.disable_v2_behavior()
+            nn.tf = tf
 
             # Initialize framework
             import core.leras.ops

core/leras/ops/__init__.py

@@ -333,7 +333,17 @@ def depth_to_space(x, size):
         x = tf.reshape(x, (-1, oh, ow, oc, ))
         return x
     else:
+        
+        b,c,h,w = x.shape.as_list()
+        oh, ow = h * size, w * size
+        oc = c // (size * size)
+
+        x = tf.reshape(x, (-1, size, size, oc, h, w, ) )
+        x = tf.transpose(x, (0, 3, 4, 1, 5, 2))
+        x = tf.reshape(x, (-1, oc, oh, ow))
+        return x
         return tf.depth_to_space(x, size, data_format=nn.data_format)
+        
 nn.depth_to_space = depth_to_space
 
 def rgb_to_lab(srgb):

core/leras/optimizers/RMSprop.py

@@ -10,20 +10,22 @@ class RMSprop(nn.OptimizerBase):
             raise ValueError('name must be defined.')
 
         self.lr_dropout = lr_dropout
+        self.lr = lr
+        self.rho = rho
+        self.epsilon = epsilon
+        
         self.clipnorm = clipnorm
 
         with tf.device('/CPU:0') :
             with tf.variable_scope(self.name):
-                self.lr = tf.Variable (lr, name="lr")
+                
-                self.rho = tf.Variable (rho, name="rho")
-                self.epsilon = tf.Variable (epsilon, name="epsilon")
                 self.iterations = tf.Variable(0, dtype=tf.int64, name='iters')
 
         self.accumulators_dict = {}
         self.lr_rnds_dict = {}
 
     def get_weights(self):
-        return [self.lr, self.rho, self.epsilon, self.iterations] + list(self.accumulators_dict.values())
+        return [self.iterations] + list(self.accumulators_dict.values())
 
     def initialize_variables(self, trainable_weights, vars_on_cpu=True, lr_dropout_on_cpu=False):
         # Initialize here all trainable variables used in training
@@ -53,13 +55,11 @@ class RMSprop(nn.OptimizerBase):
 
             a = self.accumulators_dict[ v.name ]
 
-            rho = tf.cast(self.rho, a.dtype)
+            new_a = self.rho * a + (1. - self.rho) * tf.square(g)
-            new_a = rho * a + (1. - rho) * tf.square(g)
 
-            lr = tf.cast(self.lr, a.dtype)
+            lr = tf.constant(self.lr, g.dtype)
-            epsilon = tf.cast(self.epsilon, a.dtype)
 
-            v_diff = - lr * g / (tf.sqrt(new_a) + epsilon)
+            v_diff = - lr * g / (tf.sqrt(new_a) + self.epsilon)
             if self.lr_dropout != 1.0:
                 lr_rnd = self.lr_rnds_dict[v.name]
                 v_diff *= lr_rnd

core/mplib/MPSharedList.py

@@ -114,7 +114,7 @@ class ArrayFillerSubprocessor(Subprocessor):
     def __init__(self, sh_b, data_list ):
         self.sh_b = sh_b        
         self.data_list = data_list
-        super().__init__('ArrayFillerSubprocessor', ArrayFillerSubprocessor.Cli, 60)
+        super().__init__('ArrayFillerSubprocessor', ArrayFillerSubprocessor.Cli, 60, io_loop_sleep_time=0.001)
 
     #override
     def process_info_generator(self):
@@ -124,7 +124,7 @@ class ArrayFillerSubprocessor(Subprocessor):
     #override
     def get_data(self, host_dict):
         if len(self.data_list) > 0:
-            return self.data_list.pop(0)
+            return self.data_list.pop(-1)
 
         return None
 

project.code-workspace

@@ -1,50 +0,0 @@
-{
-	"folders": [
-		{
-			"path": "."
-		}
-	],
-	"settings": {
-        "workbench.colorTheme": "Visual Studio Light",
-        "diffEditor.ignoreTrimWhitespace": true,
-        "workbench.sideBar.location": "right",
-        "breadcrumbs.enabled": false,
-        "editor.renderWhitespace": "none",
-        "editor.minimap.enabled": false,
-        "workbench.activityBar.visible": true,
-        "window.menuBarVisibility": "default",
-        "editor.fastScrollSensitivity": 10,
-        "editor.mouseWheelScrollSensitivity": 2,
-        "window.zoomLevel": 0,
-        "extensions.ignoreRecommendations": true,
-        
-        "python.linting.pylintEnabled": false,
-        "python.linting.enabled": false,
-        "python.linting.pylamaEnabled": false,
-        "python.linting.pydocstyleEnabled": false,
-        "python.pythonPath": "${env:PYTHON_EXECUTABLE}",
-		"workbench.editor.tabCloseButton": "off",
-		"workbench.editor.tabSizing": "shrink",
-		"workbench.editor.highlightModifiedTabs": true,
-		"editor.mouseWheelScrollSensitivity": 3,
-		"editor.folding": false,
-		"editor.glyphMargin": false,
-		"files.exclude": {
-			"**/__pycache__": true,
-			"**/.github": true,
-			"**/.vscode": true,
-			"**/*.dat": true,
-			"**/*.h5": true,
-            "**/*.npy": true
-		},
-		"editor.quickSuggestions": {
-			"other": false,
-			"comments": false,
-			"strings": false
-			},
-			"editor.trimAutoWhitespace": false,
-			"python.linting.pylintArgs": [
-				"--disable=import-error"
-			]
-	}
-}

requirements-colab.txt

@@ -1,9 +1,9 @@
 tqdm
-numpy==1.17.0
+numpy==1.19.3
 h5py==2.9.0
 opencv-python==4.1.0.25
 ffmpeg-python==0.1.17
 scikit-image==0.14.2
 scipy==1.4.1
 colorama
-tensorflow-gpu==1.13.2
+tensorflow-gpu==2.3.1

requirements-cuda.txt

@@ -1,11 +1,10 @@
 tqdm
-numpy==1.17.0
+numpy==1.19.3
 h5py==2.9.0
 opencv-python==4.1.0.25
 ffmpeg-python==0.1.17
 scikit-image==0.14.2
 scipy==1.4.1
 colorama
-labelme==4.2.9
+tensorflow-gpu==2.4.0rc1
-tensorflow-gpu==1.13.2
 pyqt5

samplelib/Sample.py

@@ -79,6 +79,9 @@ class Sample(object):
 
         self._filename_offset_size = None
 
+    def has_xseg_mask(self):
+        return self.xseg_mask is not None or self.xseg_mask_compressed is not None
+        
     def get_xseg_mask(self):
         if self.xseg_mask_compressed is not None:
             xseg_mask = cv2.imdecode(self.xseg_mask_compressed, cv2.IMREAD_UNCHANGED)

samplelib/SampleGeneratorFaceXSeg.py

@@ -26,11 +26,14 @@ class SampleGeneratorFaceXSeg(SampleGeneratorBase):
         samples = sum([ SampleLoader.load (SampleType.FACE, path) for path in paths ]  )
         seg_sample_idxs = SegmentedSampleFilterSubprocessor(samples).run()
 
-        seg_samples_len = len(seg_sample_idxs)
+        if len(seg_sample_idxs) == 0:
-        if seg_samples_len == 0:
+            seg_sample_idxs = SegmentedSampleFilterSubprocessor(samples, count_xseg_mask=True).run()
+            if len(seg_sample_idxs) == 0:
                 raise Exception(f"No segmented faces found.")
             else:
-            io.log_info(f"Using {seg_samples_len} segmented samples.")
+                io.log_info(f"Using {len(seg_sample_idxs)} xseg labeled samples.")                
+        else:
+            io.log_info(f"Using {len(seg_sample_idxs)} segmented samples.")
 
         if self.debug:
             self.generators_count = 1
@@ -80,8 +83,16 @@ class SampleGeneratorFaceXSeg(SampleGeneratorBase):
         def gen_img_mask(sample):
             img = sample.load_bgr()
             h,w,c = img.shape
+            
+            if sample.seg_ie_polys.has_polys():
                 mask = np.zeros ((h,w,1), dtype=np.float32)
                 sample.seg_ie_polys.overlay_mask(mask)
+            elif sample.has_xseg_mask():                
+                mask = sample.get_xseg_mask()
+                mask[mask < 0.5] = 0.0
+                mask[mask >= 0.5] = 1.0
+            else:
+                raise Exception(f'no mask in sample {sample.filename}')
 
             if face_type == sample.face_type:
                 if w != resolution:
@@ -158,9 +169,10 @@ class SampleGeneratorFaceXSeg(SampleGeneratorBase):
 
 class SegmentedSampleFilterSubprocessor(Subprocessor):
     #override
-    def __init__(self, samples ):
+    def __init__(self, samples, count_xseg_mask=False ):
         self.samples = samples
         self.samples_len = len(self.samples)
+        self.count_xseg_mask = count_xseg_mask
 
         self.idxs = [*range(self.samples_len)]
         self.result = []
@@ -169,7 +181,7 @@ class SegmentedSampleFilterSubprocessor(Subprocessor):
     #override
     def process_info_generator(self):
         for i in range(multiprocessing.cpu_count()):
-            yield 'CPU%d' % (i), {}, {'samples':self.samples}
+            yield 'CPU%d' % (i), {}, {'samples':self.samples, 'count_xseg_mask':self.count_xseg_mask}
 
     #override
     def on_clients_initialized(self):
@@ -203,6 +215,10 @@ class SegmentedSampleFilterSubprocessor(Subprocessor):
         #overridable optional
         def on_initialize(self, client_dict):
             self.samples = client_dict['samples']
+            self.count_xseg_mask = client_dict['count_xseg_mask']
 
         def process_data(self, idx):
+            if self.count_xseg_mask:
+                return idx, self.samples[idx].has_xseg_mask()
+            else:
                 return idx, self.samples[idx].seg_ie_polys.get_pts_count() != 0

samplelib/SampleLoader.py

@@ -81,7 +81,7 @@ class SampleLoader:
               shape,
               landmarks,
               seg_ie_polys,
-              xseg_mask,
+              xseg_mask_compressed,
               eyebrows_expand_mod,
               source_filename ) = data
               
@@ -91,7 +91,7 @@ class SampleLoader:
                                         shape=shape,
                                         landmarks=landmarks,
                                         seg_ie_polys=seg_ie_polys,
-                                        xseg_mask=xseg_mask,
+                                        xseg_mask_compressed=xseg_mask_compressed,
                                         eyebrows_expand_mod=eyebrows_expand_mod,
                                         source_filename=source_filename,
                                     ))
@@ -163,7 +163,7 @@ class FaceSamplesLoaderSubprocessor(Subprocessor):
                         dflimg.get_shape(),
                         dflimg.get_landmarks(),
                         dflimg.get_seg_ie_polys(),
-                        dflimg.get_xseg_mask(),
+                        dflimg.get_xseg_mask_compressed(),
                         dflimg.get_eyebrows_expand_mod(),
                         dflimg.get_source_filename() )