added ability to save optimizers states which work with K.function,

added custom Adam that can save 'iterations' param

models/ModelBase.py

@@ -1,4 +1,5 @@
 import os
+import json
 import time
 import inspect
 import pickle
@@ -119,6 +120,7 @@ class ModelBase(object):
         
         nnlib.import_all ( nnlib.DeviceConfig(allow_growth=False, **self.device_args) )
         self.device_config = nnlib.active_DeviceConfig
+        self.keras = nnlib.keras
 
         self.onInitialize()
         
@@ -271,26 +273,53 @@ class ModelBase(object):
         }            
         self.model_data_path.write_bytes( pickle.dumps(model_data) )
 
-    def load_weights_safe(self, model_filename_list):
+    def load_weights_safe(self, model_filename_list, optimizer_filename_list=[]):
         for model, filename in model_filename_list:
             filename = self.get_strpath_storage_for_file(filename)
             if Path(filename).exists():
                 model.load_weights(filename)
             
-    def save_weights_safe(self, model_filename_list):
+        if len(optimizer_filename_list) != 0:
+            opt_filename = self.get_strpath_storage_for_file('opt.h5')
+            if Path(opt_filename).exists():
+                h5dict = self.keras.utils.io_utils.H5Dict(opt_filename, mode='r')
+                try:
+                    for x in optimizer_filename_list:
+                        opt, filename = x
+                        if filename in h5dict:                   
+                            opt = opt.__class__.from_config( json.loads(h5dict[filename]) )
+                        x[0] = opt
+                finally:
+                    h5dict.close()
+            
+        return [x[0] for x in optimizer_filename_list]
+        
+    def save_weights_safe(self, model_filename_list, optimizer_filename_list=[]):
         for model, filename in model_filename_list:
             filename = self.get_strpath_storage_for_file(filename)
             model.save_weights( filename + '.tmp' )
 
-        for model, filename in model_filename_list:
+        rename_list = model_filename_list
+        if len(optimizer_filename_list) != 0:            
+            opt_filename = self.get_strpath_storage_for_file('opt.h5')
+            h5dict = self.keras.utils.io_utils.H5Dict(opt_filename + '.tmp', mode='w')
+            try:
+                for opt, filename in optimizer_filename_list:
+                    h5dict[filename] = json.dumps(opt.get_config())
+            finally:
+                h5dict.close()
+                rename_list += [('', 'opt.h5')]
+
+        for _, filename in rename_list:
             filename = self.get_strpath_storage_for_file(filename)        
             source_filename = Path(filename+'.tmp')
+            if source_filename.exists():
                 target_filename = Path(filename)
                 if target_filename.exists():
                     target_filename.unlink()                    
-                
                 source_filename.rename ( str(target_filename) )
 
+        
     def debug_one_epoch(self):
         images = []
         for generator in self.generator_list:        

models/Model_SAE/Model.py

@@ -84,7 +84,6 @@ class SAEModel(ModelBase):
     def onInitialize(self):
         exec(nnlib.import_all(), locals(), globals())
         SAEModel.initialize_nn_functions()
-        
         self.set_vram_batch_requirements({1.5:4})
         
         resolution = self.options['resolution']
@@ -111,6 +110,7 @@ class SAEModel(ModelBase):
         target_dst_ar  = [ Input ( ( bgr_shape[0] // (2**i) ,)*2 + (bgr_shape[-1],) ) for i in range(ms_count-1, -1, -1)]
         target_dstm_ar = [ Input ( ( mask_shape[0] // (2**i) ,)*2 + (mask_shape[-1],) ) for i in range(ms_count-1, -1, -1)]
 
+    
         weights_to_load = []
         if self.options['archi'] == 'liae':
             self.encoder = modelify(SAEModel.LIAEEncFlow(resolution, self.options['lighter_encoder'], ed_ch_dims=ed_ch_dims)  ) (Input(bgr_shape))
@@ -223,7 +223,12 @@ class SAEModel(ModelBase):
                 pred_dst_dstm = self.decoder_dstm(warped_dst_code)
                 pred_src_dstm = self.decoder_srcm(warped_dst_code)
         
-        self.load_weights_safe(weights_to_load)
+        self.src_dst_opt, \
+        self.src_dst_mask_opt = self.load_weights_safe(
+                                       weights_to_load, 
+                                       [ [Adam(lr=5e-5, beta_1=0.5, beta_2=0.999), 'src_dst_opt'],
+                                         [Adam(lr=5e-5, beta_1=0.5, beta_2=0.999), 'src_dst_mask_opt']
+                                       ])
 
         pred_src_src, pred_dst_dst, pred_src_dst, = [ [x] if type(x) != list else x for x in [pred_src_src, pred_dst_dst, pred_src_dst, ] ]
         
@@ -262,9 +267,6 @@ class SAEModel(ModelBase):
         psd_target_dst_anti_masked_ar = [ pred_src_dst_sigm_ar[i]*target_dstm_anti_sigm_ar[i]  for i in range(len(pred_src_dst_sigm_ar))]
         
         if self.is_training_mode:
-            def optimizer():
-                return Adam(lr=5e-5, beta_1=0.5, beta_2=0.999)
-   
             if self.options['archi'] == 'liae':          
                 src_dst_loss_train_weights = self.encoder.trainable_weights + self.inter_B.trainable_weights + self.inter_AB.trainable_weights + self.decoder.trainable_weights
                 if self.options['learn_mask']:
@@ -307,7 +309,7 @@ class SAEModel(ModelBase):
             feed += target_dst_ar[::-1]
             feed += target_dstm_ar[::-1]
     
-            self.src_dst_train = K.function (feed,[src_loss,dst_loss], optimizer().get_updates(src_loss+dst_loss, src_dst_loss_train_weights) )
+            self.src_dst_train = K.function (feed,[src_loss,dst_loss], self.src_dst_opt.get_updates(src_loss+dst_loss, src_dst_loss_train_weights) )
 
             if self.options['learn_mask']:
                 src_mask_loss = sum([ K.mean(K.square(target_srcm_ar[-1]-pred_src_srcm[-1])) for i in range(len(target_srcm_ar)) ])
@@ -317,7 +319,7 @@ class SAEModel(ModelBase):
                 feed += target_srcm_ar[::-1]
                 feed += target_dstm_ar[::-1]            
                 
-                self.src_dst_mask_train = K.function (feed,[src_mask_loss, dst_mask_loss], optimizer().get_updates(src_mask_loss+dst_mask_loss, src_dst_mask_loss_train_weights) )
+                self.src_dst_mask_train = K.function (feed,[src_mask_loss, dst_mask_loss], self.src_dst_mask_opt.get_updates(src_mask_loss+dst_mask_loss, src_dst_mask_loss_train_weights) )
 
             if self.options['learn_mask']:
                 self.AE_view = K.function ([warped_src, warped_dst], [pred_src_src[-1], pred_dst_dst[-1], pred_src_dst[-1], pred_src_dstm[-1]])
@@ -353,8 +355,12 @@ class SAEModel(ModelBase):
                 ])
     #override
     def onSave(self):
+        opt_ar = [ [self.src_dst_opt,      'src_dst_opt'],
+                   [self.src_dst_mask_opt, 'src_dst_mask_opt']
+                 ]
+        ar = []
         if self.options['archi'] == 'liae':
-            ar = [[self.encoder, 'encoder.h5'],
+            ar += [[self.encoder, 'encoder.h5'],
                   [self.inter_B, 'inter_B.h5'],
                   [self.inter_AB, 'inter_AB.h5'],
                   [self.decoder, 'decoder.h5']
@@ -362,7 +368,7 @@ class SAEModel(ModelBase):
             if self.options['learn_mask']:
                  ar += [ [self.decoderm, 'decoderm.h5'] ]
         elif self.options['archi'] == 'df' or self.options['archi'] == 'vg':
-           ar = [[self.encoder, 'encoder.h5'],
+           ar += [[self.encoder, 'encoder.h5'],
                  [self.decoder_src, 'decoder_src.h5'],
                  [self.decoder_dst, 'decoder_dst.h5']
                 ]
@@ -370,7 +376,7 @@ class SAEModel(ModelBase):
                 ar += [ [self.decoder_srcm, 'decoder_srcm.h5'],
                         [self.decoder_dstm, 'decoder_dstm.h5'] ]
                
-        self.save_weights_safe(ar)
+        self.save_weights_safe(ar, opt_ar)
        
     
     #override

nnlib/nnlib.py

@@ -71,17 +71,18 @@ ZeroPadding2D = keras.layers.ZeroPadding2D
 RandomNormal = keras.initializers.RandomNormal
 Model = keras.models.Model
 
-Adam = keras.optimizers.Adam
+#Adam = keras.optimizers.Adam
+Adam = nnlib.Adam
 
 modelify = nnlib.modelify
 gaussian_blur = nnlib.gaussian_blur
 style_loss = nnlib.style_loss
 dssim = nnlib.dssim
 
-
 PixelShuffler = nnlib.PixelShuffler
 SubpixelUpscaler = nnlib.SubpixelUpscaler
 Scale = nnlib.Scale
+
 #ReflectionPadding2D = nnlib.ReflectionPadding2D
 #AddUniformNoise = nnlib.AddUniformNoise
 """
@@ -433,6 +434,77 @@ NLayerDiscriminator = nnlib.NLayerDiscriminator
                 return dict(list(base_config.items()) + list(config.items()))
         nnlib.Scale = Scale
 
+        class Adam(keras.optimizers.Optimizer):
+            def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
+                         epsilon=None, decay=0., amsgrad=False, iterations=0, **kwargs):
+                super(Adam, self).__init__(**kwargs)
+                with K.name_scope(self.__class__.__name__):
+                    self.iterations = K.variable(iterations, dtype='int64', name='iterations')
+                    self.lr = K.variable(lr, name='lr')
+                    self.beta_1 = K.variable(beta_1, name='beta_1')
+                    self.beta_2 = K.variable(beta_2, name='beta_2')
+                    self.decay = K.variable(decay, name='decay')
+                if epsilon is None:
+                    epsilon = K.epsilon()
+                self.epsilon = epsilon
+                self.initial_decay = decay
+                self.amsgrad = amsgrad
+
+            @keras.legacy.interfaces.legacy_get_updates_support
+            def get_updates(self, loss, params):
+                grads = self.get_gradients(loss, params)
+                self.updates = [K.update_add(self.iterations, 1)]
+
+                lr = self.lr
+                if self.initial_decay > 0:
+                    lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
+                                                              K.dtype(self.decay))))
+
+                t = K.cast(self.iterations, K.floatx()) + 1
+                lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
+                             (1. - K.pow(self.beta_1, t)))
+
+                ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
+                vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
+                if self.amsgrad:
+                    vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
+                else:
+                    vhats = [K.zeros(1) for _ in params]
+                self.weights = [self.iterations] + ms + vs + vhats
+
+                for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
+                    m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
+                    v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
+                    if self.amsgrad:
+                        vhat_t = K.maximum(vhat, v_t)
+                        p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
+                        self.updates.append(K.update(vhat, vhat_t))
+                    else:
+                        p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
+
+                    self.updates.append(K.update(m, m_t))
+                    self.updates.append(K.update(v, v_t))
+                    new_p = p_t
+
+                    # Apply constraints.
+                    if getattr(p, 'constraint', None) is not None:
+                        new_p = p.constraint(new_p)
+
+                    self.updates.append(K.update(p, new_p))
+                return self.updates
+
+            def get_config(self):
+                config = {'iterations': int(K.get_value(self.iterations)),
+                          'lr': float(K.get_value(self.lr)),
+                          'beta_1': float(K.get_value(self.beta_1)),
+                          'beta_2': float(K.get_value(self.beta_2)),
+                          'decay': float(K.get_value(self.decay)),
+                          'epsilon': self.epsilon,
+                          'amsgrad': self.amsgrad}
+                base_config = super(Adam, self).get_config()
+                return dict(list(base_config.items()) + list(config.items()))
+        nnlib.Adam = Adam
+
         '''        
         not implemented in plaidML
         class ReflectionPadding2D(keras.layers.Layer):