qqwweee · qqwweee · Jun 10, 2018 · Jun 9, 2018 · Jun 10, 2018
diff --git a/README.md b/README.md
@@ -6,6 +6,7 @@
 
 A Keras implementation of YOLOv3 (Tensorflow backend) inspired by [allanzelener/YAD2K](https://github.com/allanzelener/YAD2K).
 
+
 ---
 
 ## Quick Start
@@ -70,3 +71,5 @@ If you want to use original pretrained weights for YOLOv3:
 5. Always load pretrained weights and freeze layers in the first stage of training. Or try Darknet training. It's OK if there is a mismatch warning.
 
 6. The training strategy is for reference only. Adjust it according to your dataset and your goal. And add further strategy if needed.
+
+7. For speeding up the training process with frozen layers train_bottleneck.py can be used. It will compute the bottleneck features of the frozen model first and then only trains the last layers. This makes training on CPU possible in a reasonable time. See [this](https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html) for more information on bottleneck features.
diff --git a/train_bottleneck.py b/train_bottleneck.py
@@ -0,0 +1,222 @@
+"""
+Retrain the YOLO model for your own dataset.
+"""
+import os
+import numpy as np
+import keras.backend as K
+from keras.layers import Input, Lambda
+from keras.models import Model
+from keras.optimizers import Adam
+from keras.callbacks import TensorBoard, ModelCheckpoint, ReduceLROnPlateau, EarlyStopping
+
+from yolo3.model import preprocess_true_boxes, yolo_body, tiny_yolo_body, yolo_loss
+from yolo3.utils import get_random_data
+
+
+def _main():
+    annotation_path = 'train.txt'
+    log_dir = 'logs/000/'
+    classes_path = 'model_data/coco_classes.txt'
+    anchors_path = 'model_data/yolo_anchors.txt'
+    class_names = get_classes(classes_path)
+    num_classes = len(class_names)
+    anchors = get_anchors(anchors_path)
+
+    input_shape = (416,416) # multiple of 32, hw
+
+    model, bottleneck_model, last_layer_model = create_model(input_shape, anchors, num_classes,
+            freeze_body=2, weights_path='model_data/yolo_weights.h5') # make sure you know what you freeze
+
+    logging = TensorBoard(log_dir=log_dir)
+    checkpoint = ModelCheckpoint(log_dir + 'ep{epoch:03d}-loss{loss:.3f}-val_loss{val_loss:.3f}.h5',
+        monitor='val_loss', save_weights_only=True, save_best_only=True, period=3)
+    reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.1, patience=3, verbose=1)
+    early_stopping = EarlyStopping(monitor='val_loss', min_delta=0, patience=10, verbose=1)
+
+    val_split = 0.1
+    with open(annotation_path) as f:
+        lines = f.readlines()
+    np.random.seed(10101)
+    np.random.shuffle(lines)
+    np.random.seed(None)
+    num_val = int(len(lines)*val_split)
+    num_train = len(lines) - num_val
+
+    # Train with frozen layers first, to get a stable loss.
+    # Adjust num epochs to your dataset. This step is enough to obtain a not bad model.
+    if True:
+        # perform bottleneck training
+        if not os.path.isfile("bottlenecks.npz"):
+            print("calculating bottlenecks")
+            batch_size=8
+            bottlenecks=bottleneck_model.predict_generator(data_generator_wrapper(lines, batch_size, input_shape, anchors, num_classes, random=False, verbose=True),
+             steps=(len(lines)//batch_size)+1, max_queue_size=1)
+            np.savez("bottlenecks.npz", bot0=bottlenecks[0], bot1=bottlenecks[1], bot2=bottlenecks[2])
+
+        # load bottleneck features from file
+        dict_bot=np.load("bottlenecks.npz")
+        bottlenecks_train=[dict_bot["bot0"][:num_train], dict_bot["bot1"][:num_train], dict_bot["bot2"][:num_train]]
+        bottlenecks_val=[dict_bot["bot0"][num_train:], dict_bot["bot1"][num_train:], dict_bot["bot2"][num_train:]]
+
+        # train last layers with fixed bottleneck features
+        batch_size=8
+        print("Training last layers with bottleneck features")
+        print('with {} samples, val on {} samples and batch size {}.'.format(num_train, num_val, batch_size))
+        last_layer_model.compile(optimizer='adam', loss={'yolo_loss': lambda y_true, y_pred: y_pred})
+        last_layer_model.fit_generator(bottleneck_generator(lines[:num_train], batch_size, input_shape, anchors, num_classes, bottlenecks_train),
+                steps_per_epoch=max(1, num_train//batch_size),
+                validation_data=bottleneck_generator(lines[num_train:], batch_size, input_shape, anchors, num_classes, bottlenecks_val),
+                validation_steps=max(1, num_val//batch_size),
+                epochs=30,
+                initial_epoch=0, max_queue_size=1)
+        model.save_weights(log_dir + 'trained_weights_stage_0.h5')
+
+        # train last layers with random augmented data
+        model.compile(optimizer=Adam(lr=1e-3), loss={
+            # use custom yolo_loss Lambda layer.
+            'yolo_loss': lambda y_true, y_pred: y_pred})
+        batch_size = 16
+        print('Train on {} samples, val on {} samples, with batch size {}.'.format(num_train, num_val, batch_size))
+        model.fit_generator(data_generator_wrapper(lines[:num_train], batch_size, input_shape, anchors, num_classes),
+                steps_per_epoch=max(1, num_train//batch_size),
+                validation_data=data_generator_wrapper(lines[num_train:], batch_size, input_shape, anchors, num_classes),
+                validation_steps=max(1, num_val//batch_size),
+                epochs=50,
+                initial_epoch=0,
+                callbacks=[logging, checkpoint])
+        model.save_weights(log_dir + 'trained_weights_stage_1.h5')
+
+    # Unfreeze and continue training, to fine-tune.
+    # Train longer if the result is not good.
+    if True:
+        for i in range(len(model.layers)):
+            model.layers[i].trainable = True
+        model.compile(optimizer=Adam(lr=1e-4), loss={'yolo_loss': lambda y_true, y_pred: y_pred}) # recompile to apply the change
+        print('Unfreeze all of the layers.')
+
+        batch_size = 4 # note that more GPU memory is required after unfreezing the body
+        print('Train on {} samples, val on {} samples, with batch size {}.'.format(num_train, num_val, batch_size))
+        model.fit_generator(data_generator_wrapper(lines[:num_train], batch_size, input_shape, anchors, num_classes),
+            steps_per_epoch=max(1, num_train//batch_size),
+            validation_data=data_generator_wrapper(lines[num_train:], batch_size, input_shape, anchors, num_classes),
+            validation_steps=max(1, num_val//batch_size),
+            epochs=100,
+            initial_epoch=50,
+            callbacks=[logging, checkpoint, reduce_lr, early_stopping])
+        model.save_weights(log_dir + 'trained_weights_final.h5')
+
+    # Further training if needed.
+
+
+def get_classes(classes_path):
+    '''loads the classes'''
+    with open(classes_path) as f:
+        class_names = f.readlines()
+    class_names = [c.strip() for c in class_names]
+    return class_names
+
+def get_anchors(anchors_path):
+    '''loads the anchors from a file'''
+    with open(anchors_path) as f:
+        anchors = f.readline()
+    anchors = [float(x) for x in anchors.split(',')]
+    return np.array(anchors).reshape(-1, 2)
+
+
+def create_model(input_shape, anchors, num_classes, load_pretrained=True, freeze_body=2,
+            weights_path='model_data/yolo_weights.h5'):
+    '''create the training model'''
+    K.clear_session() # get a new session
+    image_input = Input(shape=(None, None, 3))
+    h, w = input_shape
+    num_anchors = len(anchors)
+
+    y_true = [Input(shape=(h//{0:32, 1:16, 2:8}[l], w//{0:32, 1:16, 2:8}[l], \
+        num_anchors//3, num_classes+5)) for l in range(3)]
+
+    model_body = yolo_body(image_input, num_anchors//3, num_classes)
+    print('Create YOLOv3 model with {} anchors and {} classes.'.format(num_anchors, num_classes))
+
+    if load_pretrained:
+        model_body.load_weights(weights_path, by_name=True, skip_mismatch=True)
+        print('Load weights {}.'.format(weights_path))
+        if freeze_body in [1, 2]:
+            # Freeze darknet53 body or freeze all but 3 output layers.
+            num = (185, len(model_body.layers)-3)[freeze_body-1]
+            for i in range(num): model_body.layers[i].trainable = False
+            print('Freeze the first {} layers of total {} layers.'.format(num, len(model_body.layers)))
+
+    # get output of second last layers and create bottleneck model of it
+    out1=model_body.layers[246].output
+    out2=model_body.layers[247].output
+    out3=model_body.layers[248].output
+    bottleneck_model = Model([model_body.input, *y_true], [out1, out2, out3])
+
+    # create last layer model of last layers from yolo model
+    in0 = Input(shape=bottleneck_model.output[0].shape[1:].as_list()) 
+    in1 = Input(shape=bottleneck_model.output[1].shape[1:].as_list())
+    in2 = Input(shape=bottleneck_model.output[2].shape[1:].as_list())
+    last_out0=model_body.layers[249](in0)
+    last_out1=model_body.layers[250](in1)
+    last_out2=model_body.layers[251](in2)
+    model_last=Model(inputs=[in0, in1, in2], outputs=[last_out0, last_out1, last_out2])
+    model_loss_last =Lambda(yolo_loss, output_shape=(1,), name='yolo_loss',
+        arguments={'anchors': anchors, 'num_classes': num_classes, 'ignore_thresh': 0.5})(
+        [*model_last.output, *y_true])
+    last_layer_model = Model([in0,in1,in2, *y_true], model_loss_last)
+
+
+    model_loss = Lambda(yolo_loss, output_shape=(1,), name='yolo_loss',
+        arguments={'anchors': anchors, 'num_classes': num_classes, 'ignore_thresh': 0.5})(
+        [*model_body.output, *y_true])
+    model = Model([model_body.input, *y_true], model_loss)
+
+    return model, bottleneck_model, last_layer_model
+
+def data_generator(annotation_lines, batch_size, input_shape, anchors, num_classes, random=True, verbose=False):
+    '''data generator for fit_generator'''
+    n = len(annotation_lines)
+    i = 0
+    while True:
+        image_data = []
+        box_data = []
+        for b in range(batch_size):
+            if i==0 and random:
+                np.random.shuffle(annotation_lines)
+            image, box = get_random_data(annotation_lines[i], input_shape, random=random)
+            image_data.append(image)
+            box_data.append(box)
+            i = (i+1) % n
+        image_data = np.array(image_data)
+        if verbose:
+            print("Progress: ",i,"/",n)
+        box_data = np.array(box_data)
+        y_true = preprocess_true_boxes(box_data, input_shape, anchors, num_classes)
+        yield [image_data, *y_true], np.zeros(batch_size)
+
+def data_generator_wrapper(annotation_lines, batch_size, input_shape, anchors, num_classes, random=True, verbose=False):
+    n = len(annotation_lines)
+    if n==0 or batch_size<=0: return None
+    return data_generator(annotation_lines, batch_size, input_shape, anchors, num_classes, random, verbose)
+
+def bottleneck_generator(annotation_lines, batch_size, input_shape, anchors, num_classes, bottlenecks):
+    n = len(annotation_lines)
+    i = 0
+    while True:
+        box_data = []
+        b0=np.zeros((batch_size,bottlenecks[0].shape[1],bottlenecks[0].shape[2],bottlenecks[0].shape[3]))
+        b1=np.zeros((batch_size,bottlenecks[1].shape[1],bottlenecks[1].shape[2],bottlenecks[1].shape[3]))
+        b2=np.zeros((batch_size,bottlenecks[2].shape[1],bottlenecks[2].shape[2],bottlenecks[2].shape[3]))
+        for b in range(batch_size):
+            _, box = get_random_data(annotation_lines[i], input_shape, random=False, proc_img=False)
+            box_data.append(box)
+            b0[b]=bottlenecks[0][i]
+            b1[b]=bottlenecks[1][i]
+            b2[b]=bottlenecks[2][i]
+            i = (i+1) % n
+        box_data = np.array(box_data)
+        y_true = preprocess_true_boxes(box_data, input_shape, anchors, num_classes)
+        yield [b0, b1, b2, *y_true], np.zeros(batch_size)
+
+if __name__ == '__main__':
+    _main()
diff --git a/yolo3/utils.py b/yolo3/utils.py
@@ -33,7 +33,7 @@ def letterbox_image(image, size):
 def rand(a=0, b=1):
     return np.random.rand()*(b-a) + a
 
-def get_random_data(annotation_line, input_shape, random=True, max_boxes=20, jitter=.3, hue=.1, sat=1.5, val=1.5):
+def get_random_data(annotation_line, input_shape, random=True, max_boxes=20, jitter=.3, hue=.1, sat=1.5, val=1.5, proc_img=True):
     '''random preprocessing for real-time data augmentation'''
     line = annotation_line.split()
     image = Image.open(line[0])
@@ -48,10 +48,12 @@ def get_random_data(annotation_line, input_shape, random=True, max_boxes=20, jit
         nh = int(ih*scale)
         dx = (w-nw)//2
         dy = (h-nh)//2
-        image = image.resize((nw,nh), Image.BICUBIC)
-        new_image = Image.new('RGB', (w,h), (128,128,128))
-        new_image.paste(image, (dx, dy))
-        image_data = np.array(new_image)/255.
+        image_data=0
+        if proc_img:
+            image = image.resize((nw,nh), Image.BICUBIC)
+            new_image = Image.new('RGB', (w,h), (128,128,128))
+            new_image.paste(image, (dx, dy))
+            image_data = np.array(new_image)/255.
 
         # correct boxes
         box_data = np.zeros((max_boxes,5))