generalize autoencoder models + execution on floyd

Makefile

@@ -69,6 +69,20 @@ exec_clustering:
 	$(PYTHON_INTERPRETER) -m iss.exec.clustering
 
 
+#################################################################################
+# FLOYDHUB                                                                      #
+#################################################################################
+
+floyd_training:
+	floyd run --task training
+
+floyd_training_prod:
+	floyd run --task training_prod
+
+floyd_retraining:
+	floyd run --task retraining
+
+
 #################################################################################
 # Self Documenting Commands                                                     #
 #################################################################################

iss/exec/training.py

@@ -1,4 +1,5 @@
 import os
+import click
 
 from iss.init_config import CONFIG
 from iss.models.DataLoader import ImageDataGeneratorWrapper
@@ -8,29 +9,45 @@ from iss.models import SimpleConvAutoEncoder
 from iss.models import VarAutoEncoder
 from iss.models import VarConvAutoEncoder
 
-## Variables globales
-_MODEL_TYPE = 'simple_conv'
-_LOAD_NAME = None
-_LOAD = False
 
-## Data loader
-data_loader = ImageDataGeneratorWrapper(CONFIG, model = _MODEL_TYPE)
+@click.command()
+@click.option('--model-type', default='simple_conv', show_default=True, type=str)
+@click.option('--load', default=False, is_flag=True)
+@click.option('--load-name', default=None, show_default=True, type=str)
+def main(model_type, load, load_name):
+
+    ## Variables globales
+    _MODEL_TYPE = model_type
+    _LOAD_NAME = load_name
+    _LOAD = load
+
+    ## Data loader
+    data_loader = ImageDataGeneratorWrapper(CONFIG, model = _MODEL_TYPE)
+
+    ## Model
+    if _MODEL_TYPE in ['simple_conv']:
+        model = SimpleConvAutoEncoder(CONFIG.get('models')[_MODEL_TYPE])
+    elif _MODEL_TYPE in ['simple']:
+        model = SimpleAutoEncoder(CONFIG.get('models')[_MODEL_TYPE])
+    else:
+        raise Exception
 
-## Model
-if _MODEL_TYPE in ['simple_conv']:
-    model = SimpleConvAutoEncoder(CONFIG.get('models')[_MODEL_TYPE])
     if _LOAD:
         model.load(which = _LOAD_NAME)
+        
     model.encoder_model.summary()
     model.decoder_model.summary()
+    model.model.summary()
 
-model.model.summary()
+    ## Entraineur
+    trainer = ModelTrainer(model, data_loader, CONFIG.get('models')[_MODEL_TYPE], callbacks=[])
 
-## Entraineur
-trainer = ModelTrainer(model, data_loader, CONFIG.get('models')[_MODEL_TYPE], callbacks=[])
+    ## Entrainement
+    try:
+        trainer.train()
+    except KeyboardInterrupt:
+        trainer.model.save()
 
-## Entrainement
-try:
-    trainer.train()
-except KeyboardInterrupt:
-    trainer.model.save()
+
+if __name__ == '__main__':
+    main()

iss/models/Callbacks.py

@@ -42,7 +42,6 @@ class TensorboardCallback(Callback):
         super(TensorboardCallback, self).__init__()
 
     def on_epoch_end(self, epoch, logs=None):
-        print(logs)
         image_summaries = []
         
         for input_pict in self.data_loader.next()[0][:self.limit_image]:
@@ -75,3 +74,10 @@ class TensorboardCallback(Callback):
         self.writer.flush()
 
 
+class FloydhubTrainigMetricsCallback(Callback):
+    """FloydHub Training Metric Integration"""
+    def on_epoch_end(self, epoch, logs=None):
+        """Print Training Metrics"""
+        print('{{"metric": "loss", "value": {}, "epoch": {}}}'.format(logs.get('loss'), epoch))
+        print('{{"metric": "val_loss", "value": {}, "epoch": {}}}'.format(logs.get('val_loss'), epoch))
+

iss/models/ModelTrainer.py

@@ -1,7 +1,7 @@
 # -*- coding: utf-8 -*-
 
 from keras.callbacks import ModelCheckpoint, CSVLogger
-from iss.models.Callbacks import DisplayPictureCallback, TensorboardCallback
+from iss.models.Callbacks import DisplayPictureCallback, TensorboardCallback, FloydhubTrainigMetricsCallback
 from iss.tools.tools import Tools
 import keras
 
@@ -98,5 +98,7 @@ class ModelTrainer:
 				data_loader = self.data_loader.get_train_generator()
 			)])
 
+		if 'floyd' in config['callbacks']:
+			self.callbacks.extend([FloydhubTrainigMetricsCallback()])
 
 		return self

iss/models/SimpleAutoEncoder.py

@@ -17,6 +17,7 @@ class SimpleAutoEncoder(AbstractAutoEncoderModel):
 
 			self.activation = config['activation']
 			self.input_shape = (config['input_height'], config['input_width'], config['input_channel'])
+			self.latent_shape = config['latent_shape']
 			self.lr = config['learning_rate']
 			self.build_model()
 
@@ -25,25 +26,25 @@ class SimpleAutoEncoder(AbstractAutoEncoderModel):
 
 			picture = Input(shape = input_shape)
 
+			# encoded network
 			x = Flatten()(picture)
 			layer_1 = Dense(1000, activation = 'relu', name = 'enc_1')(x)
 			layer_2 = Dense(100, activation = 'relu', name = 'enc_2')(layer_1)
-			layer_3 = Dense(50, activation = 'relu', name = 'enc_3')(layer_2)
-			layer_4 = Dense(100, activation = 'relu', name = 'dec_1')(layer_3)
+			encoded = Dense(self.latent_shape, activation = 'relu', name = 'enc_3')(layer_2)
+			self.encoder_model = Model(picture, encoded, name = "encoder")
+
+			#  decoded netword
+			latent_input = Input(shape = (self.latent_shape,))
+			layer_4 = Dense(100, activation = 'relu', name = 'dec_1')(latent_input)
 			layer_5 = Dense(1000, activation = 'relu', name = 'dec_2')(layer_4)
 
-			# encoded network
-			# x = Conv2D(1, (3, 3), activation = 'relu', padding = 'same', name = 'enc_conv_1')(picture)
-			# encoded = MaxPooling2D((2, 2))(x)
-
-			# decoded network
-			# x = Conv2D(1, (3, 3), activation = 'relu', padding = 'same', name = 'dec_conv_1')(encoded)
-			# x = UpSampling2D((2, 2))(x)
-			# x = Flatten()(x)
 			x = Dense(np.prod(input_shape), activation = self.activation)(layer_5)
 			decoded = Reshape((input_shape))(x)
 
-			self.model = Model(picture, decoded)
+			self.decoder_model = Model(latent_input, decoded, name = "decoder")
+			
+			picture_dec = self.decoder_model(self.encoder_model(picture))
+			self.model = Model(picture, picture_dec)
 			
 			# optimizer = Adadelta(lr = self.lr, rho = 0.95, epsilon = None, decay = 0.0)
 			optimizer = Adam(lr = 0.001, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False)