Lots of fixes for transfer learning

Author: vanpelt

keras-transfer/config-defaults.yaml

@@ -2,77 +2,111 @@
 import numpy as np
 from keras.preprocessing.image import ImageDataGenerator
 from keras.models import Sequential
+from keras.callbacks import Callback
 from keras.layers import Dropout, Flatten, Dense
-from keras import applications
+from keras.applications.vgg16 import VGG16, preprocess_input
+from dogcat_data import generators, get_nb_files
+import os
+import sys
+import wandb
+from wandb.keras import WandbCallback
 
-# dimensions of our images.
-img_width, img_height = 150, 150
+wandb.init()
+config = wandb.config
 
-top_model_weights_path = 'bottleneck_fc_model.h5'
-train_data_dir = 'dogcat-data/train'
-validation_data_dir = 'dogcat-data/validation'
-nb_train_samples = 2000
-nb_validation_samples = 2000
-epochs = 50
-batch_size = 10 # should be divisible nb_samples
+# dimensions of our images.
+config.img_width = 224
+config.img_height = 224
+config.epochs = 50
+config.batch_size = 40 
 
+top_model_weights_path = 'bottleneck.h5'
+train_dir = 'dogcat-data/train'
+validation_dir = 'dogcat-data/validation'
+nb_train_samples = 1000
+nb_validation_samples = 1000
 
 def save_bottlebeck_features():
-    datagen = ImageDataGenerator(rescale=1. / 255)
-
+    if os.path.exists('bottleneck_features_train.npy') and (len(sys.argv) == 1 or sys.argv[1] != "--force"):
+        print("Using saved features, pass --force to save new features")
+        return
+    datagen = ImageDataGenerator(preprocessing_function=preprocess_input)
+    train_generator = datagen.flow_from_directory(
+            train_dir,
+            target_size=(config.img_width, config.img_height),
+            batch_size=config.batch_size,
+            class_mode="binary")
+
+    val_generator = datagen.flow_from_directory(
+            validation_dir,
+            target_size=(config.img_width, config.img_height),
+            batch_size=config.batch_size,
+            class_mode="binary")
+    
     # build the VGG16 network
-    model = applications.VGG16(include_top=False, weights='imagenet')
-
-    generator = datagen.flow_from_directory(
-        train_data_dir,
-        target_size=(img_width, img_height),
-        batch_size=batch_size,
-        class_mode=None,
-        shuffle=False)
-    bottleneck_features_train = model.predict_generator(
-        generator, nb_train_samples // batch_size)
-    print(bottleneck_features_train.shape)
-    np.save(open('bottleneck_features_train.npy', 'wb'),
-            bottleneck_features_train)
-
-    generator = datagen.flow_from_directory(
-        validation_data_dir,
-        target_size=(img_width, img_height),
-        batch_size=batch_size,
-        class_mode=None,
-        shuffle=False)
-    bottleneck_features_validation = model.predict_generator(
-        generator, nb_validation_samples // batch_size)
-    np.save(open('bottleneck_features_validation.npy', 'wb'),
-            bottleneck_features_validation)
+    model = VGG16(include_top=False, weights='imagenet')
+    
+    print("Predicting bottleneck training features")
+    training_labels = []
+    training_features = []
+    for batch in range(5): #nb_train_samples // config.batch_size):
+        data, labels = next(train_generator)
+        training_labels.append(labels)
+        training_features.append(model.predict(data))
+    training_labels = np.concatenate(training_labels)
+    training_features = np.concatenate(training_features)
+    np.savez(open('bottleneck_features_train.npy', 'wb'),
+            features=training_features, labels=training_labels)
+    
+    print("Predicting bottleneck validation features")
+    validation_labels = []
+    validation_features = []
+    validation_data = []
+    for batch in range(nb_validation_samples // config.batch_size):
+        data, labels = next(val_generator)
+        validation_features.append(model.predict(data))
+        validation_labels.append(labels)
+        validation_data.append(data)
+    validation_labels = np.concatenate(validation_labels)
+    validation_features = np.concatenate(validation_features)
+    validation_data = np.concatenate(validation_data)
+    np.savez(open('bottleneck_features_validation.npy', 'wb'),
+            features=training_features, labels=training_labels, data=validation_data)
 
 
 def train_top_model():
-    train_data = np.load(open('bottleneck_features_train.npy', 'rb'))
-    train_labels = np.array(
-        [0] * int(nb_train_samples / 2) + [1] * int(nb_train_samples / 2))
-
-    validation_data = np.load(open('bottleneck_features_validation.npy', 'rb'))
-    validation_labels = np.array(
-        [0] * int(nb_validation_samples / 2) + [1] * int(nb_validation_samples / 2))
+    train = np.load(open('bottleneck_features_train.npy', 'rb'))
+    X_train, y_train = (train['features'], train['labels'])
+    test = np.load(open('bottleneck_features_validation.npy', 'rb'))
+    X_test, y_test, val_data = (test['features'], test['labels'], test['data'])
 
-    print(validation_data.shape)
-    print(validation_labels.shape)
-    #train_data.reshape(192,2048)
-    #validation_data.reshape(192,2048)
     model = Sequential()
-    model.add(Flatten(input_shape=train_data.shape[1:]))
+    model.add(Flatten(input_shape=X_train[0].shape))
     model.add(Dense(256, activation='relu'))
     model.add(Dropout(0.5))
     model.add(Dense(1, activation='sigmoid'))
 
     model.compile(optimizer='rmsprop',
-                  loss='binary_crossentropy', metrics=['accuracy'])
-
-    model.fit(train_data, train_labels,
-              epochs=epochs,
-              batch_size=batch_size,
-              validation_data=(validation_data, validation_labels))
+                  loss='binary_crossentropy', metrics=['binary_accuracy'])
+    
+    class Images(Callback):
+        def on_epoch_end(self, epoch, logs):
+            base_model = VGG16(include_top=False, weights='imagenet')
+            indices = np.random.randint(val_data.shape[0], size=36)
+            test_data = val_data[indices]
+            features = base_model.predict(np.array([preprocess_input(data) for data in test_data]))
+            pred_data = model.predict(features)
+            wandb.log({
+                  "examples": [
+                        wandb.Image(test_data[i], caption="cat" if pred_data[i] < 0.5 else "dog")
+                        for i, data in enumerate(test_data)]
+            }, commit=False)
+
+    model.fit(X_train, y_train,
+              epochs=config.epochs,
+              batch_size=config.batch_size,
+              validation_data=(X_test, y_test),
+              callbacks=[Images(), WandbCallback(save_model=False)])
     model.save_weights(top_model_weights_path)