# optimizer, otherwise I'd recommend the commented-out Adam optimizer. #adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) sgd = SGD(lr=0.001, momentum=0.9, decay=0.0, nesterov=False) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=sgd, loss=ssd_loss.compute_loss) #Set up the data generators for the training # 1: Instantiate two `DataGenerator` objects: One for training, one for validation. # Optional: If you have enough memory, consider loading the images into memory for the reasons explained above. train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) # 2: Parse the image and label lists for the training and validation datasets. This can take a while. # TODO: Set the paths to the datasets here. # The directories that contain the images. VOC_2007_images_dir = '../../datasets/VOCdevkit/VOC2007/JPEGImages/' # The directories that contain the annotations. VOC_2007_annotations_dir = '../../datasets/VOCdevkit/VOC2007/Annotations/' # The paths to the image sets. VOC_2007_train_image_set_filename = '../../datasets/VOCdevkit/VOC2007/ImageSets/Main/train.txt'
# 加载之前保存的模型或者创建新模型 K.clear_session() # Clear previous models from memory. if os.path.exists(model_saved): # We need to create an SSDLoss object in order to pass that to the model loader. ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model = load_model(model_saved, custom_objects={ 'AnchorBoxes': AnchorBoxes, 'compute_loss': ssd_loss.compute_loss }) else: logger.error(f'无法加载训练好的模型:{model_saved}') val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=dataset_prefix + 'traffic_val.h5') val_dataset_size = val_dataset.get_dataset_size() logger.info("Number of images in the validation dataset:\t{:>6}".format( val_dataset_size)) predictor_sizes = [ model.get_layer('classes4').output_shape[1:3], model.get_layer('classes5').output_shape[1:3], model.get_layer('classes6').output_shape[1:3], model.get_layer('classes7').output_shape[1:3] ] ssd_input_encoder = SSDInputEncoder(img_height=img_height, img_width=img_width, n_classes=n_classes,
def run(train_dir, valid_dir, set_dir, model_dir): # train_dir = arguments.train_dir # valid_dir = arguments.valid_dir train_dataset_dir = train_dir train_annot_dir = train_dir + '/annot/' train_set = train_dir + '/img_set.txt' valid_dataset_dir = valid_dir valid_annot_dir = valid_dir + '/annot/' valid_set = valid_dir + '/valid_set.txt' # Set Training and Validation dataset paths batch_size = 16 print('Using batch size of: {}'.format(batch_size)) #model_path = 'COCO_512.h5' model_path = model_dir # model_path = 'saved_model.h5' # Needs to know classes and order to map to integers classes = ['background', 'car', 'bus', 'truck'] # Set required parameters for training of SSD img_height = 512 img_width = 512 img_channels = 3 # Colour image mean_color = [123, 117, 104] # DO NOT CHANGE swap_channels = [2, 1, 0] # Original SSD used BGR n_classes = 3 # 80 for COCO scales_coco = [0.04, 0.1, 0.26, 0.42, 0.58, 0.74, 0.9, 1.06] scales = scales_coco aspect_ratios = [[1.0, 2.0, 0.5], [1.0, 2.0, 0.5, 3.0, 1.0 / 3.0], [1.0, 2.0, 0.5, 3.0, 1.0 / 3.0], [1.0, 2.0, 0.5, 3.0, 1.0 / 3.0], [1.0, 2.0, 0.5, 3.0, 1.0 / 3.0], [1.0, 2.0, 0.5], [1.0, 2.0, 0.5]] two_boxes_for_ar1 = True steps = [8, 16, 32, 64, 128, 256, 512] offsets = [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5] clip_boxes = False variances = [0.1, 0.1, 0.2, 0.2] normalize_coords = True K.clear_session() model = ssd_512(image_size=(img_height, img_width, img_channels), n_classes=n_classes, mode='training', l2_regularization=0.0005, scales=scales, aspect_ratios_per_layer=aspect_ratios, two_boxes_for_ar1=two_boxes_for_ar1, steps=steps, offsets=offsets, clip_boxes=clip_boxes, variances=variances, normalize_coords=normalize_coords, subtract_mean=mean_color, swap_channels=swap_channels) model.load_weights(model_path, by_name=True) sgd = SGD(lr=0.001, momentum=0.9, decay=0.0, nesterov=False) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=sgd, loss=ssd_loss.compute_loss) # model = load_model(model_path, custom_objects={'AnchorBoxes': AnchorBoxes, # 'L2Normalization': L2Normalization, # 'compute_loss': ssd_loss.compute_loss}) # Create Data Generators for train and valid sets train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) valid_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) train_dataset.parse_xml(images_dirs=[train_dataset_dir], image_set_filenames=[train_set], annotations_dirs=[train_annot_dir], classes=classes, include_classes='all', exclude_truncated=False, exclude_difficult=False, ret=False) valid_dataset.parse_xml(images_dirs=[valid_dataset_dir], image_set_filenames=[valid_set], annotations_dirs=[valid_annot_dir], classes=classes, include_classes='all', exclude_truncated=False, exclude_difficult=False, ret=False) # Will speed up trainig but requires more memory # Can comment out to avoid memory requirements ''' train_dataset.create_hdf5_dataset(file_path='dataset_pascal_voc_07+12_trainval.h5', resize=False, variable_image_size=True, verbose=True) valid_dataset.create_hdf5_dataset(file_path='dataset_pascal_voc_07_test.h5', resize=False, variable_image_size=True, verbose=True) ''' ssd_data_augmentation = SSDDataAugmentation(img_height=img_height, img_width=img_width, background=mean_color) convert_to_3_channels = ConvertTo3Channels() resize = Resize(height=img_height, width=img_width) predictor_sizes = [ model.get_layer('conv4_3_norm_mbox_conf').output_shape[1:3], model.get_layer('fc7_mbox_conf').output_shape[1:3], model.get_layer('conv6_2_mbox_conf').output_shape[1:3], model.get_layer('conv7_2_mbox_conf').output_shape[1:3], model.get_layer('conv8_2_mbox_conf').output_shape[1:3], model.get_layer('conv9_2_mbox_conf').output_shape[1:3], model.get_layer('conv10_2_mbox_conf').output_shape[1:3] ] ssd_input_encoder = SSDInputEncoder(img_height=img_height, img_width=img_width, n_classes=n_classes, predictor_sizes=predictor_sizes, scales=scales, aspect_ratios_per_layer=aspect_ratios, two_boxes_for_ar1=two_boxes_for_ar1, steps=steps, offsets=offsets, clip_boxes=clip_boxes, variances=variances, matching_type='multi', pos_iou_threshold=0.5, neg_iou_limit=0.5, normalize_coords=normalize_coords) train_generator = train_dataset.generate( batch_size=batch_size, shuffle=True, transformations=[ssd_data_augmentation], label_encoder=ssd_input_encoder, returns={'processed_images', 'encoded_labels'}, keep_images_without_gt=False) val_generator = valid_dataset.generate( batch_size=batch_size, shuffle=False, transformations=[convert_to_3_channels, resize], label_encoder=ssd_input_encoder, returns={'processed_images', 'encoded_labels'}, keep_images_without_gt=False) # Get the number of samples in the training and validations datasets. train_dataset_size = train_dataset.get_dataset_size() valid_dataset_size = valid_dataset.get_dataset_size() print("Number of images in the training dataset:\t{:>6}".format( train_dataset_size)) print("Number of images in the validation dataset:\t{:>6}".format( valid_dataset_size)) model_checkpoint = ModelCheckpoint( filepath= 'ssd_epoch-{epoch:02d}_loss-{loss:.4f}_val_loss-{val_loss:.4f}.h5', monitor='val_loss', verbose=1, save_best_only=True, save_weights_only=False, mode='auto', period=1) #csv_logger = CSVLogger(filename='ssd512_training_log.csv', # separator=',', # append=True) learning_rate_scheduler = LearningRateScheduler(schedule=lr_schedule, verbose=1) terminate_on_nan = TerminateOnNaN() callbacks = [ model_checkpoint, csv_logger, learning_rate_scheduler, terminate_on_nan ] #callbacks = [learning_rate_scheduler, # terminate_on_nan] initial_epoch = 0 final_epoch = 150 # 150 steps_per_epoch = math.ceil(119 / batch_size) # ceil(num_samples/batch_size) # Training history = model.fit_generator(generator=train_generator, steps_per_epoch=steps_per_epoch, epochs=final_epoch, callbacks=callbacks, validation_data=val_generator, validation_steps=math.ceil( valid_dataset_size / batch_size), initial_epoch=initial_epoch) # Save final trained model model.save('trained.h5') # Make predictions predict_generator = valid_dataset.generate( batch_size=1, shuffle=True, transformations=[convert_to_3_channels, resize], label_encoder=None, returns={ 'processed_images', 'filenames', 'inverse_transform', 'original_images', 'original_labels' }, keep_images_without_gt=False) batch_images, batch_filenames, batch_inverse_transforms, batch_original_images, batch_original_labels = next( predict_generator) i = 0 # Which batch item to look at print("Image:", batch_filenames[i]) print() print("Ground truth boxes:\n") print(np.array(batch_original_labels[i])) y_pred = model.predict(batch_images) y_pred_decoded = decode_detections(y_pred, confidence_thresh=0.2, iou_threshold=0.4, top_k=200, normalize_coords=normalize_coords, img_height=img_height, img_width=img_width) y_pred_decoded_inv = apply_inverse_transforms(y_pred_decoded, batch_inverse_transforms) np.set_printoptions(precision=2, suppress=True, linewidth=90) print("Predicted boxes:\n") print(' class conf xmin ymin xmax ymax') print(y_pred_decoded_inv[i]) # Set the colors for the bounding boxes colors = plt.cm.hsv(np.linspace(0, 1, n_classes + 1)).tolist() # classes = ['background', 'car', 'bus', 'truck', 'motorbike'] # Already set at start plt.figure(figsize=(20, 12)) plt.imshow(batch_original_images[i]) current_axis = plt.gca() for box in batch_original_labels[i]: xmin = box[1] ymin = box[2] xmax = box[3] ymax = box[4] label = '{}'.format(classes[int(box[0])]) current_axis.add_patch( plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin, color='green', fill=False, linewidth=2)) current_axis.text(xmin, ymin, label, size='x-large', color='white', bbox={ 'facecolor': 'green', 'alpha': 1.0 }) for box in y_pred_decoded_inv[i]: xmin = box[2] ymin = box[3] xmax = box[4] ymax = box[5] color = colors[int(box[0])] label = '{}: {:.2f}'.format(classes[int(box[0])], box[1]) current_axis.add_patch( plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin, color=color, fill=False, linewidth=2)) current_axis.text(xmin, ymin, label, size='x-large', color='white', bbox={ 'facecolor': color, 'alpha': 1.0 }) plt.show() return
# 2: Load the trained weights into the model. weights_path = "C:/Users/t_tor/Unsynced/ssd_weights/VGG_VOC0712Plus_SSD_300x300_ft_iter_160000.h5" model.load_weights(weights_path, by_name=True) # 3: Compile the model so that Keras won't complain the next time you load it. adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=adam, loss=ssd_loss.compute_loss) dataset = DataGenerator() # The directories that contain the images. images = "C:/Users/t_tor/Unsynced/complete_dataset/jpeg/" # The directories that contain the annotations. annotations = "C:/Users/t_tor/Unsynced/complete_dataset/annotations/" # The paths to the image sets. imagenet_images = "C:/Users/t_tor/Unsynced/complete_dataset/imagenet_images.txt" custom_images = "C:/Users/t_tor/Unsynced/complete_dataset/custom_images.txt" coco_images = "C:/Users/t_tor/Unsynced/complete_dataset/coco_images.txt" voc07test_images = "C:/Users/t_tor/Unsynced/complete_dataset/voc07_test.txt" # The XML parser needs to now what object class names to look for and in which order to map them to integers. #classes = ['background', 'boat']
def load_VOC_IMG_generators(self,model): print('Making VOC image generators') datadir = self.datas['DATA_PATH'] train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) test_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) images_dir = os.path.join(datadir,'Images') annotations_dir = os.path.join(datadir,'Annotations') train_image_set_filename = os.path.join(datadir,'ImageSets','train.txt') val_image_set_filename = os.path.join(datadir,'ImageSets','val.txt') test_image_set_filename = os.path.join(datadir,'ImageSets','test.txt') generator_options = self.datas['GENERATOR'] train_dataset.parse_xml(images_dirs=[images_dir], image_set_filenames=[train_image_set_filename], annotations_dirs=[annotations_dir], classes=self.datas['CLASSES'], include_classes='all', exclude_truncated=False, exclude_difficult=False, ret=False) val_dataset.parse_xml(images_dirs=[images_dir], image_set_filenames=[val_image_set_filename], annotations_dirs=[annotations_dir], classes=self.datas['CLASSES'], include_classes='all', exclude_truncated=False, exclude_difficult=False, ret=False) test_dataset.parse_xml(images_dirs=[images_dir], image_set_filenames=[test_image_set_filename], annotations_dirs=[annotations_dir], classes=self.datas['CLASSES'], include_classes='all', exclude_truncated=False, exclude_difficult=False, ret=False) convert_to_3_channels = ConvertTo3Channels() target_size = generator_options['TARGET_SIZE'] resize = Resize(height=target_size[0], width=target_size[1]) predictor_sizes = [model.get_layer('conv4_3_norm_mbox_conf').output_shape[1:3], model.get_layer('fc7_mbox_conf').output_shape[1:3], model.get_layer('conv6_2_mbox_conf').output_shape[1:3], model.get_layer('conv7_2_mbox_conf').output_shape[1:3], model.get_layer('conv8_2_mbox_conf').output_shape[1:3], model.get_layer('conv9_2_mbox_conf').output_shape[1:3]] scales_pascal = [0.1, 0.2, 0.37, 0.54, 0.71, 0.88, 1.05] # The anchor box scaling factors used in the original SSD300 for the Pascal VOC datasets scales = scales_pascal aspect_ratios = [[1.0, 2.0, 0.5], [1.0, 2.0, 0.5, 3.0, 1.0/3.0], [1.0, 2.0, 0.5, 3.0, 1.0/3.0], [1.0, 2.0, 0.5, 3.0, 1.0/3.0], [1.0, 2.0, 0.5], [1.0, 2.0, 0.5]] # The anchor box aspect ratios used in the original SSD300; the order matters steps = [8, 16, 32, 64, 100, 300] # The space between two adjacent anchor box center points for each predictor layer. two_boxes_for_ar1 = True mean_color=[123,117,104] #TODO : add this as a parameter offsets = [0.5, 0.5, 0.5, 0.5, 0.5, 0.5] clip_boxes=False variances=[0.1, 0.1, 0.2, 0.2] normalize_coords=True ssd_input_encoder = SSDInputEncoder(img_height = target_size[0], img_width = target_size[1], n_classes = 20, #TODO : handle subsampling predictor_sizes=predictor_sizes, scales=scales, aspect_ratios_per_layer=aspect_ratios, two_boxes_for_ar1=two_boxes_for_ar1, steps=steps, offsets=offsets, clip_boxes=clip_boxes, variances=variances, matching_type='multi', pos_iou_threshold=0.5, neg_iou_limit=0.5, normalize_coords=normalize_coords ) train_generator = train_dataset.generate(batch_size=generator_options['BATCH_SIZE'], shuffle=True, transformations=[convert_to_3_channels, resize], label_encoder=ssd_input_encoder, returns={'processed_images', 'encoded_labels'}, keep_images_without_gt=False) val_generator = val_dataset.generate(batch_size=generator_options['BATCH_SIZE'], shuffle=True, transformations=[convert_to_3_channels, resize], label_encoder=ssd_input_encoder, returns={'processed_images', 'encoded_labels'}, keep_images_without_gt=False) test_generator = test_dataset.generate(batch_size=generator_options['BATCH_SIZE'], shuffle=True, transformations=[convert_to_3_channels, resize], label_encoder=ssd_input_encoder, returns={'processed_images', 'encoded_labels'}, keep_images_without_gt=False) return [train_generator,train_dataset.get_dataset_size()],[val_generator,val_dataset.get_dataset_size()],[test_generator,train_dataset.get_dataset_size()]
def data_generator_func(config: Dict): """Data Generator for training data and validation data Parameters ---------- config : Dict Config yaml/json containing all parameter Returns ------- train_dataset, val_dataset """ # Init DataGenerator start_data = timer() train_dataset = DataGenerator(load_images_into_memory=config['training'] ['train_load_images_into_memory'], hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=config['training'] ['validation_load_images_into_memory'], hdf5_dataset_path=None) if config['training']['train_load_images_into_memory'] is not False: print("[INFO]... You have chosen to load data into memory") else: print( "[WARNING]... You have chosen not to load data into memory. It will still work but will be much slower" ) train_img_dir = config['training']['train_img_dir'] val_img_dir = config['training']['val_img_dir'] train_annotation_dir = config['training']['train_annotation_dir'] val_annotation_dir = config['training']['val_annotation_dir'] train_image_set_filename = config['training']['train_image_set_filename'] val_image_set_filename = config['training']['val_image_set_filename'] classes = config['training']['classes'] if config['training']['annotation_type'] == 'xml': train_dataset.parse_xml(images_dirs=[train_img_dir], image_set_filenames=[train_image_set_filename], annotations_dirs=[train_annotation_dir], classes=classes, include_classes='all', exclude_truncated=False, exclude_difficult=False, ret=False) val_dataset.parse_xml(images_dirs=[val_img_dir], image_set_filenames=[val_image_set_filename], annotations_dirs=[val_annotation_dir], classes=classes, include_classes='all', exclude_truncated=False, exclude_difficult=True, ret=False) if config['training']['annotation_type'] == 'csv': train_dataset.parse_csv(images_dir=train_img_dir, labels_filename=train_annotation_dir, input_format=[ 'image_name', 'xmin', 'xmax', 'ymin', 'ymax', 'class_id' ], include_classes='all') val_dataset.parse_csv(images_dir=val_img_dir, labels_filename=val_annotation_dir, input_format=[ 'image_name', 'xmin', 'xmax', 'ymin', 'ymax', 'class_id' ], include_classes='all') end_data = timer() print( f"[INFO]...Time taken by Data loading/transformation Job is {(end_data - start_data)/60:.2f} min(s)" ) return train_dataset, val_dataset
# The `DataGenerator` itself is fairly generic. I doesn't contain any data augmentation or bounding box encoding logic. Instead, you pass a list of image transformations and an encoder for the bounding boxes in the `transformations` and `label_encoder` arguments of the data generator's `generate()` method, and the data generator will then apply those given transformations and the encoding to the data. Everything here is preset already, but if you'd like to learn more about the data generator and its data augmentation capabilities, take a look at the detailed tutorial in [this](https://github.com/pierluigiferrari/data_generator_object_detection_2d) repository. # # The data augmentation settings defined further down reproduce the data augmentation pipeline of the original SSD training. The training generator receives an object `ssd_data_augmentation`, which is a transformation object that is itself composed of a whole chain of transformations that replicate the data augmentation procedure used to train the original Caffe implementation. The validation generator receives an object `resize`, which simply resizes the input images. # # An `SSDInputEncoder` object, `ssd_input_encoder`, is passed to both the training and validation generators. As explained above, it matches the ground truth labels to the model's anchor boxes and encodes the box coordinates into the format that the model needs. # # In order to train the model on a dataset other than Pascal VOC, either choose `DataGenerator`'s appropriate parser method that corresponds to your data format, or, if `DataGenerator` does not provide a suitable parser for your data format, you can write an additional parser and add it. Out of the box, `DataGenerator` can handle datasets that use the Pascal VOC format (use `parse_xml()`), the MS COCO format (use `parse_json()`) and a wide range of CSV formats (use `parse_csv()`). # In[ ]: # 1: Instantiate two `DataGenerator` objects: One for training, one for validation. # Optional: If you have enough memory, consider loading the images into memory for the reasons explained above. train_dataset = DataGenerator( load_images_into_memory=False, hdf5_dataset_path='dataset_udacity_traffic_train.h5') val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path='dataset_udacity_traffic_val.h5') # 2: Parse the image and label lists for the training and validation datasets. This can take a while. # TODO: Set the paths to the datasets here. # The directories that contain the images. # VOC_2007_images_dir = '../../datasets/VOCdevkit/VOC2007/JPEGImages/' # VOC_2012_images_dir = '../../datasets/VOCdevkit/VOC2012/JPEGImages/' # The directories that contain the annotations. # VOC_2007_annotations_dir = '../../datasets/VOCdevkit/VOC2007/Annotations/' # VOC_2012_annotations_dir = '../../datasets/VOCdevkit/VOC2012/Annotations/'
print("Weights file loaded:", weights_path) # 3: Instantiate an Adam optimizer and the SSD loss function and compile the model. adam = Adam(lr=0.00001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=adam, loss=ssd_loss.compute_loss) # 1: Instantiate two `DataGenerator` objects: One for training, one for validation. # Optional: If you have enough memory, consider loading the images into memory for the reasons explained above. train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) # 2: Parse the image and label lists for the training and validation datasets. This can take a while. # TODO: Set the paths to the datasets here. # The directories that contain the images. images_train_dir = '/home/2017018/rblin01/BD_QCAV/train_polar/PARAM_POLAR/RetinaNet_Stokes' images_test_dir = '/home/2017018/rblin01/BD_QCAV/test_polar/PARAM_POLAR/RetinaNet_Stokes' # The directories that contain the annotations. annotations_train_dir = '/home/2017018/rblin01/BD_QCAV/train_polar/LABELS' annotations_test_dir = '/home/2017018/rblin01/BD_QCAV/test_polar/LABELS' # The paths to the image sets.
def train_VOC(config): ''' Train the given configuration ; the configuration must be constructed according to the utility script found in utils/generateconfig.py. Arguments: config : the configuration of the model to use ; should already be loaded ''' ################################### ### PATHS AND PARAMETERS ################################## datadir = config.DATA_DIR local_dir = config.ROOT_FOLDER img_shape = config.IMG_SHAPE classes = config.CLASSES checkpoint_output = os.path.join(local_dir, 'models', config.CHECKPOINT_NAME) model_output = os.path.join(local_dir, 'models', config.MODEL_NAME) img_height = img_shape[0] # Height of the model input images img_width = img_shape[1] # Width of the model input images img_channels = img_shape[ 2] # Number of color channels of the model input images mean_color = [ 123, 117, 104 ] # The per-channel mean of the images in the dataset. Do not change this value if you're using any of the pre-trained weights. swap_channels = [ 2, 1, 0 ] # The color channel order in the original SSD is BGR, so we'll have the model reverse the color channel order of the input images. n_classes = 20 # Number of positive classes, e.g. 20 for Pascal VOC, 80 for MS COCO scales_pascal = [ 0.1, 0.2, 0.37, 0.54, 0.71, 0.88, 1.05 ] # The anchor box scaling factors used in the original SSD300 for the Pascal VOC datasets scales = scales_pascal aspect_ratios = [ [1.0, 2.0, 0.5], [1.0, 2.0, 0.5, 3.0, 1.0 / 3.0], [1.0, 2.0, 0.5, 3.0, 1.0 / 3.0], [1.0, 2.0, 0.5, 3.0, 1.0 / 3.0 ], [1.0, 2.0, 0.5 ], [1.0, 2.0, 0.5] ] # The anchor box aspect ratios used in the original SSD300; the order matters two_boxes_for_ar1 = True steps = [ 8, 16, 32, 64, 100, 300 ] # The space between two adjacent anchor box center points for each predictor layer. offsets = [ 0.5, 0.5, 0.5, 0.5, 0.5, 0.5 ] # The offsets of the first anchor box center points from the top and left borders of the image as a fraction of the step size for each predictor layer. clip_boxes = False # Whether or not to clip the anchor boxes to lie entirely within the image boundaries variances = [ 0.1, 0.1, 0.2, 0.2 ] # The variances by which the encoded target coordinates are divided as in the original implementation normalize_coords = True batch_size = config.BATCH_SIZE # Change the batch size if you like, or if you run into GPU memory issues. ################################### ### BUILDING MODEL ################################## K.clear_session() # Clear previous models from memory. model = ssd_300(image_size=(img_height, img_width, img_channels), n_classes=n_classes, mode='training', l2_regularization=0.0005, scales=scales, aspect_ratios_per_layer=aspect_ratios, two_boxes_for_ar1=two_boxes_for_ar1, steps=steps, offsets=offsets, clip_boxes=clip_boxes, variances=variances, normalize_coords=normalize_coords, subtract_mean=mean_color, swap_channels=swap_channels) weights_path = os.path.join(local_dir, 'weights', 'VGG_VOC0712_SSD_300x300_iter_120000.h5') model.load_weights(weights_path, by_name=True) #adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) sgd = SGD(lr=0.001, momentum=0.9, decay=0.0, nesterov=False) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=sgd, loss=ssd_loss.compute_loss) ################################### ### LOADING DATAS ################################## train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) images_dir = os.path.join(datadir, 'Images') annotations_dir = os.path.join(datadir, 'Annotations') trainval_image_set_filename = os.path.join(datadir, 'ImageSets', 'train.txt') test_image_set_filename = os.path.join(datadir, 'ImageSets', 'val.txt') # The XML parser needs to now what object class names to look for and in which order to map them to integers. # train_dataset.parse_xml(images_dirs=[images_dir], image_set_filenames=[trainval_image_set_filename], annotations_dirs=[annotations_dir], classes=classes, include_classes='all', exclude_truncated=False, exclude_difficult=False, ret=False) val_dataset.parse_xml(images_dirs=[images_dir], image_set_filenames=[test_image_set_filename], annotations_dirs=[annotations_dir], classes=classes, include_classes='all', exclude_truncated=False, exclude_difficult=True, ret=False) train_dataset.create_hdf5_dataset(file_path='flowers_train.h5', resize=False, variable_image_size=True, verbose=True) val_dataset.create_hdf5_dataset(file_path='flowers_val.h5', resize=False, variable_image_size=True, verbose=True) ssd_data_augmentation = SSDDataAugmentation(img_height=img_height, img_width=img_width, background=mean_color) convert_to_3_channels = ConvertTo3Channels() resize = Resize(height=img_height, width=img_width) # The encoder constructor needs the spatial dimensions of the model's predictor layers to create the anchor boxes. predictor_sizes = [ model.get_layer('conv4_3_norm_mbox_conf').output_shape[1:3], model.get_layer('fc7_mbox_conf').output_shape[1:3], model.get_layer('conv6_2_mbox_conf').output_shape[1:3], model.get_layer('conv7_2_mbox_conf').output_shape[1:3], model.get_layer('conv8_2_mbox_conf').output_shape[1:3], model.get_layer('conv9_2_mbox_conf').output_shape[1:3] ] ssd_input_encoder = SSDInputEncoder(img_height=img_height, img_width=img_width, n_classes=n_classes, predictor_sizes=predictor_sizes, scales=scales, aspect_ratios_per_layer=aspect_ratios, two_boxes_for_ar1=two_boxes_for_ar1, steps=steps, offsets=offsets, clip_boxes=clip_boxes, variances=variances, matching_type='multi', pos_iou_threshold=0.5, neg_iou_limit=0.5, normalize_coords=normalize_coords) train_generator = train_dataset.generate( batch_size=batch_size, shuffle=True, transformations=[ssd_data_augmentation], label_encoder=ssd_input_encoder, returns={'processed_images', 'encoded_labels'}, keep_images_without_gt=False) val_generator = val_dataset.generate( batch_size=batch_size, shuffle=False, transformations=[convert_to_3_channels, resize], label_encoder=ssd_input_encoder, returns={'processed_images', 'encoded_labels'}, keep_images_without_gt=False) # Get the number of samples in the training and validations datasets. train_dataset_size = train_dataset.get_dataset_size() val_dataset_size = val_dataset.get_dataset_size() print("Number of images in the training dataset:\t{:>6}".format( train_dataset_size)) print("Number of images in the validation dataset:\t{:>6}".format( val_dataset_size)) ################################### ### PREPARE TRAINING ################################## def lr_schedule(epoch): if epoch < 80: return 0.001 elif epoch < 100: return 0.0001 else: return 0.00001 model_checkpoint = ModelCheckpoint(filepath=checkpoint_output, monitor='val_loss', verbose=1, save_best_only=True, save_weights_only=False, mode='auto', period=1) early_stopping = EarlyStopping(monitor='val_loss', min_delta=0.0, patience=10, verbose=1) learning_rate_scheduler = LearningRateScheduler(schedule=lr_schedule, verbose=1) terminate_on_nan = TerminateOnNaN() callbacks = [ model_checkpoint, learning_rate_scheduler, terminate_on_nan, early_stopping ] ################################### ### TRAINING ################################## epochs = config.EPOCHS steps_per_epoch = ceil(train_dataset_size / batch_size) model.summary() history = model.fit_generator(generator=train_generator, steps_per_epoch=steps_per_epoch, epochs=epochs, callbacks=callbacks, validation_data=val_generator, validation_steps=ceil(val_dataset_size / batch_size)) model.save(model_output)
# The `DataGenerator` itself is fairly generic. I doesn't contain any data augmentation or bounding box encoding logic. Instead, you pass a list of image transformations and an encoder for the bounding boxes in the `transformations` and `label_encoder` arguments of the data generator's `generate()` method, and the data generator will then apply those given transformations and the encoding to the data. Everything here is preset already, but if you'd like to learn more about the data generator and its data augmentation capabilities, take a look at the detailed tutorial in [this](https://github.com/pierluigiferrari/data_generator_object_detection_2d) repository. # # The data augmentation settings defined further down reproduce the data augmentation pipeline of the original SSD training. The training generator receives an object `ssd_data_augmentation`, which is a transformation object that is itself composed of a whole chain of transformations that replicate the data augmentation procedure used to train the original Caffe implementation. The validation generator receives an object `resize`, which simply resizes the input images. # # An `SSDInputEncoder` object, `ssd_input_encoder`, is passed to both the training and validation generators. As explained above, it matches the ground truth labels to the model's anchor boxes and encodes the box coordinates into the format that the model needs. # # In order to train the model on a dataset other than Pascal VOC, either choose `DataGenerator`'s appropriate parser method that corresponds to your data format, or, if `DataGenerator` does not provide a suitable parser for your data format, you can write an additional parser and add it. Out of the box, `DataGenerator` can handle datasets that use the Pascal VOC format (use `parse_xml()`), the MS COCO format (use `parse_json()`) and a wide range of CSV formats (use `parse_csv()`). # 1: Instantiate two `DataGenerator` objects: One for training, one for validation. # Optional: If you have enough memory, consider loading the images into memory for the reasons explained above. # train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) # val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path='dataset_pascal_voc_07+12_trainval.h5') val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path='dataset_pascal_voc_07_test.h5') # 2: Parse the image and label lists for the training and validation datasets. This can take a while. # TODO: Set the paths to the datasets here. # The directories that contain the images. VOC_2007_images_dir = '../../datasets/VOCdevkit/VOC2007/JPEGImages/' VOC_2012_images_dir = '../../datasets/VOCdevkit/VOC2012/JPEGImages/' # The directories that contain the annotations. VOC_2007_annotations_dir = '../../datasets/VOCdevkit/VOC2007/Annotations/' VOC_2012_annotations_dir = '../../datasets/VOCdevkit/VOC2012/Annotations/' # The paths to the image sets.
# # The image processing chain defined further down in the object named `data_augmentation_chain` is just one possibility of what a data augmentation pipeline for unform-size images could look like. Feel free to put together other image processing chains, you can use the `DataAugmentationConstantInputSize` class as a template. Or you could use the original SSD data augmentation pipeline by instantiting an `SSDDataAugmentation` object and passing that to the generator instead. This procedure is not exactly efficient, but it evidently produces good results on multiple datasets. # # An `SSDInputEncoder` object, `ssd_input_encoder`, is passed to both the training and validation generators. As explained above, it matches the ground truth labels to the model's anchor boxes and encodes the box coordinates into the format that the model needs. # ### Note: # # The example setup below was used to train SSD7 on two road traffic datasets released by [Udacity](https://github.com/udacity/self-driving-car/tree/master/annotations) with around 20,000 images in total and 5 object classes (car, truck, pedestrian, bicyclist, traffic light), although the vast majority of the objects are cars. The original datasets have a constant image size of 1200x1920 RGB. I consolidated the two datasets, removed a few bad samples (although there are probably many more), and resized the images to 300x480 RGB, i.e. to one sixteenth of the original image size. In case you'd like to train a model on the same dataset, you can download the consolidated and resized dataset I used [here](https://drive.google.com/open?id=1tfBFavijh4UTG4cGqIKwhcklLXUDuY0D) (about 900 MB). # In[4]: # 1: Instantiate two `DataGenerator` objects: One for training, one for validation. # Optional: If you have enough memory, consider loading the images into memory for the reasons explained above. train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) # 2: Parse the image and label lists for the training and validation datasets. # TODO: Set the paths to your dataset here. # Images images_dir = '../gray_images/' # Ground truth train_labels_filename = '../train.csv' val_labels_filename = '../val.csv' train_dataset.parse_csv(
8, 16, 32, 64, 100, 300 ] # The space between two adjacent anchor box center points for each predictor layer. offsets = [ 0.5, 0.5, 0.5, 0.5, 0.5, 0.5 ] # The offsets of the first anchor box center points from the top and left borders of the image as a fraction of the step size for each predictor layer. clip_boxes = False # Whether or not to clip the anchor boxes to lie entirely within the image boundaries variances = [ 0.1, 0.1, 0.2, 0.2 ] # The variances by which the encoded target coordinates are divided as in the original implementation normalize_coords = True # 1: Instantiate two `DataGenerator` objects: One for training, one for validation. # Optional: If you have enough memory, consider loading the images into memory for the reasons explained above. train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) # 2: Parse the image and label lists for the training and validation datasets. This can take a while. # TODO: Set the paths to the datasets here. dataset_dir = 'D:/Develop/data/raccoon_dataset/' # The directories that contain the images. VOC_2007_images_dir = 'D:/Develop/data/VOCdevkit/VOC2007/JPEGImages/' VOC_2007_images_dir = 'D:/Develop/data/raccoon_dataset/images/' VOC_2007_images_dir = dataset_dir + 'images/' #VOC_2012_images_dir = 'D:/Develop/data/VOCdevkit/VOC2012/JPEGImages/'
# Make datagen images_dir = '../FLIR_ADAS/validation/PreviewData' annotation_csv = '../validation.csv' input_format = ["image_name", "class_id", "xmin", "xmax", "ymin", "ymax"] # order of columns in csv #dataset = DataGenerator() #dataset.parse_csv( # images_dir, # annotation_csv, # input_format, # verbose=True # ) dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path='./validation.h5') evaluator = Evaluator(model=model, n_classes=n_classes, data_generator=dataset, model_mode=model_mode) results = evaluator(img_height=img_height, img_width=img_width, batch_size=32, data_generator_mode='resize', round_confidences=False, matching_iou_threshold=0.5, border_pixels='include', sorting_algorithm='quicksort', average_precision_mode='sample',
def main(): model_mode = 'inference' K.clear_session() # Clear previous models from memory. model = build_model(image_size=(Config.img_height, Config.img_width, Config.img_channels), n_classes=Config.n_classes, mode=model_mode, l2_regularization=Config.l2_regularization, scales=Config.scales, aspect_ratios_per_layer=Config.steps, two_boxes_for_ar1=True, steps=Config.steps, offsets=Config.offsets, clip_boxes=False, variances=Config.variances, normalize_coords=Config.normalize_coords, subtract_mean=Config.intensity_mean, swap_channels=[2, 1, 0], confidence_thresh=0.01, iou_threshold=0.45, top_k=200, nms_max_output_size=400) # 2: Load the trained weights into the model. weights_path = os.getcwd() + '/weights/' + args.model_name + ".h5" model.load_weights(weights_path, by_name=True) adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=adam, loss=ssd_loss.compute_loss) test_dataset = DataGenerator(load_images_into_memory=True, hdf5_dataset_path=os.getcwd() + "/data/" + args.dataset + '/polyp_test.h5') test_dataset_size = test_dataset.get_dataset_size() print("Number of images in the test dataset:\t{:>6}".format(test_dataset_size)) classes = ['background', 'polyp'] evaluator = Evaluator(model=model, n_classes=Config.n_classes, data_generator=test_dataset, model_mode=model_mode) results = evaluator(img_height=Config.img_height, img_width=Config.img_width, batch_size=args.batch_size, data_generator_mode='resize', round_confidences=False, matching_iou_threshold=0.5, border_pixels='include', sorting_algorithm='quicksort', average_precision_mode='sample', num_recall_points=11, ignore_neutral_boxes=True, return_precisions=True, return_recalls=True, return_average_precisions=True, verbose=True) mean_average_precision, average_precisions, precisions, recalls = results for i in range(1, len(average_precisions)): print("{:<14}{:<6}{}".format(classes[i], 'AP', round(average_precisions[i], 3))) print("{:<14}{:<6}{}".format('', 'mAP', round(mean_average_precision, 3))) m = max((Config.n_classes + 1) // 2, 2) n = 2 fig, cells = plt.subplots(m, n, figsize=(n * 8, m * 8)) val = 0 for i in range(m): for j in range(n): if n * i + j + 1 > Config.n_classes: break cells[i, j].plot(recalls[n * i + j + 1], precisions[n * i + j + 1], color='blue', linewidth=1.0) cells[i, j].set_xlabel('recall', fontsize=14) cells[i, j].set_ylabel('precision', fontsize=14) cells[i, j].grid(True) cells[i, j].set_xticks(np.linspace(0, 1, 11)) cells[i, j].set_yticks(np.linspace(0, 1, 11)) cells[i, j].set_title("{}, AP: {:.3f}".format(classes[n * i + j + 1], average_precisions[n * i + j + 1]), fontsize=16) image = plt.gcf() # plt.show() plt.draw() image.savefig(os.getcwd() + "/test_out/test_" + str(val) + ".png", dpi=100) val += 1
############################################################ # 4. Draw the predicted boxes onto the image ############################################################ classes = [ 'background', 'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat', 'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person', 'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor' ] num_classes = len(classes) - 1 images_dir = '/home/adam/.keras/datasets/VOCdevkit' # Ground truth test_hdf5_path = osp.join(images_dir, '07_test.h5') test_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=test_hdf5_path) model_mode = 'training' evaluator = Evaluator(model=model, n_classes=num_classes, data_generator=test_dataset, model_mode=model_mode) results = evaluator(img_height=image_height, img_width=image_width, batch_size=4, data_generator_mode='resize', round_confidences=False, matching_iou_threshold=0.5, border_pixels='include', sorting_algorithm='quicksort', average_precision_mode='sample',
# TODO: Set the path to the weights you want to load. weights_path = '/home/ubuntu/Result/ssd_face/ssd_face_initial_weights.h5' model.load_weights(weights_path, by_name=True) # 3: Instantiate an optimizer and the SSD loss function and compile the model. sgd = SGD(lr=0.001, momentum=0.9, decay=0.0005) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=sgd, loss=ssd_loss.compute_loss) # 1: Instantiate two `DataGenerator` objects: One for training, one for validation. # Optional: If you have enough memory, consider loading the images into memory for the reasons explained above. train_dataset = DataGenerator(load_images_into_memory=True, hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=True, hdf5_dataset_path=None) # 2: Parse the image and label lists for the training and validation datasets. # TODO: Set the paths to your dataset here. # Images images_dir = '/home/ubuntu/Datasets/WiderFace/' # Ground truth train_labels_filename = '/home/ubuntu/Datasets/WiderFace/labels_train.csv' val_labels_filename = '/home/ubuntu/Datasets/WiderFace/labels_val.csv' train_dataset.parse_csv(
# Load the weights that we've just created via sub-sampling. model.load_weights(weights_path, by_name=True) print("Weights file loaded:", weights_path) # 3: Instantiate an Adam optimizer and the SSD loss function and compile the model. adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=adam, loss=ssd_loss.compute_loss) #--- train_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) images_dir = '/home/student/Documents/GitHub/darknet/data/obj' train_labels_filename = 'onsite-images-export.csv' val_labels_filename = 'onsite-images-valid.csv' train_dataset.parse_csv( images_dir=images_dir, labels_filename=train_labels_filename, input_format=[ 'image_name', 'xmin', 'ymin', 'xmax', 'ymax', 'class_id' ], # This is the order of the first six columns in the CSV file that contains the labels for your dataset. If your labels are in XML format, maybe the XML parser will be helpful, check the documentation. include_classes='all')
def main(): # create dataset dataset = DataGenerator() dataset.parse_xml(images_dirs=[dataset_images_dir], image_set_filenames=[test_image_set_filename], annotations_dirs=[dataset_annotations_dir], classes=classes, include_classes='all', exclude_truncated=False, exclude_difficult=False, ret=False) # create model model = ssd_512(image_size=(img_height, img_width, img_channels), n_classes=n_classes, mode=model_mode, l2_regularization=0.0005, scales=scales, aspect_ratios_per_layer=aspect_ratios, two_boxes_for_ar1=two_boxes_for_ar1, steps=steps, offsets=offsets, clip_boxes=clip_boxes, variances=variances, normalize_coords=normalize_coords, subtract_mean=mean_color, swap_channels=swap_channels) # load weights and compile it model.load_weights(weights_path, by_name=True) adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0) ssd_loss = SSDLoss(neg_pos_ratio=3, alpha=1.0) model.compile(optimizer=adam, loss=ssd_loss.compute_loss) evaluator = Evaluator(model=model, n_classes=n_classes, data_generator=dataset, model_mode=model_mode) results = evaluator(img_height=img_height, img_width=img_width, batch_size=8, data_generator_mode='resize', round_confidences=False, matching_iou_threshold=0.5, border_pixels='include', sorting_algorithm='quicksort', average_precision_mode='sample', num_recall_points=11, ignore_neutral_boxes=True, return_precisions=True, return_recalls=True, return_average_precisions=True, verbose=True) mean_average_precision, average_precisions, precisions, recalls = results for i in range(1, len(average_precisions)): print("{:<14}{:<6}{}".format(classes[i], 'AP', round(average_precisions[i], 3))) print() print("{:<14}{:<6}{}".format('', 'mAP', round(mean_average_precision, 3))) m = max((n_classes + 1) // 2, 2) n = 2 fig, cells = plt.subplots(m, n, figsize=(n * 8, m * 8)) for i in range(m): for j in range(n): if n * i + j + 1 > n_classes: break cells[i, j].plot(recalls[n * i + j + 1], precisions[n * i + j + 1], color='blue', linewidth=1.0) cells[i, j].set_xlabel('recall', fontsize=14) cells[i, j].set_ylabel('precision', fontsize=14) cells[i, j].grid(True) cells[i, j].set_xticks(np.linspace(0, 1, 6)) cells[i, j].set_yticks(np.linspace(0, 1, 6)) cells[i, j].set_xlim(0.0, 1.0) cells[i, j].set_ylim(0.0, 1.0) cells[i, j].set_title("{}, AP: {:.3f}".format( classes[n * i + j + 1], average_precisions[n * i + j + 1]), fontsize=16) if not os.path.isdir("evaluate_result"): os.makedirs("evaluate_result") plt.savefig('evaluate_result/ssd512_face_detection.png')
train = json.load( open("/home/jsearcy/Desktop/ML Data Sets/malaria/training.json")) test = json.load(open("/home/jsearcy/Desktop/ML Data Sets/malaria/test.json")) write_new_json( train, '/home/jsearcy/UO Data Science Dropbox/Public Datasets/BBBC041o/train.json', folder='train') write_new_json( test, '/home/jsearcy/UO Data Science Dropbox/Public Datasets/BBBC041o/test.json', folder='test') oa.aosefoeji train_dataset = DataGenerator(load_images_into_memory=True, hdf5_dataset_path=None) train_dataset.parse_json(images_dirs=images_dir, annotations_filenames=["output_train.json"], ground_truth_available=True, include_classes=use_list) val_dataset = DataGenerator(load_images_into_memory=False, hdf5_dataset_path=None) val_dataset.parse_json(images_dirs=images_dir, annotations_filenames=["output_test.json"], ground_truth_available=True, include_classes=use_list) val_dataset.create_hdf5_dataset(file_path='val_dataset.h5', resize=(1200, 1600),