from utils.visualizations import plot_kernels
import matplotlib.pyplot as plt
# from models.ssd_utils import construct_SSD
from models import SSD300
"""
weights_path = '../trained_models/SSD300_weights.hdf5'
model = SSD300(weights_path=weights_path)
model.summary()
kernel_weights = model.get_layer('conv2d_1').get_weights()[0]
plot_kernels(kernel_weights[:, :, 0:1, :])
plot_kernels(kernel_weights[:, :, 1:2, :])
plot_kernels(kernel_weights[:, :, 2:3, :])
plt.show()
"""
weights_path = '../trained_models/VGG16_weights.hdf5'
model = SSD300()
model.load_weights(weights_path, by_name=True)
kernel_weights = model.get_layer('conv2d_1').get_weights()[0]
plot_kernels(kernel_weights[:, :, 0:1, :])
plot_kernels(kernel_weights[:, :, 1:2, :])
plot_kernels(kernel_weights[:, :, 2:3, :])
plt.show()
from image_generator import ImageGenerator from models import SSD300 from utils.prior_box_creator import PriorBoxCreator from utils.prior_box_assigner import PriorBoxAssigner from utils.box_transformer import BoxTransformer from utils.XML_parser import XMLParser from utils.utils import split_data from utils.utils import read_image, resize_image import numpy as np import matplotlib.pyplot as plt image_shape = (300, 300, 3) model = SSD300(image_shape) box_creator = PriorBoxCreator(model) prior_boxes = box_creator.create_boxes() layer_scale, box_arg = 0, 780 box_coordinates = prior_boxes[layer_scale][box_arg, :, :] image_path = '../images/' image_key = '007040.jpg' box_creator.draw_boxes(image_path + image_key, box_coordinates) data_path = '../datasets/VOCdevkit/VOC2007/' ground_truths = XMLParser(data_path + 'Annotations/').get_data() prior_box_manager = PriorBoxAssigner(prior_boxes, ground_truths) assigned_boxes = prior_box_manager.assign_boxes() prior_box_manager.draw_assigned_boxes(image_path, image_shape[0:2], image_key) batch_size = 7 train_keys, validation_keys = split_data(assigned_boxes, training_ratio=.8) assigned_image_generator = ImageGenerator(assigned_boxes, batch_size,
from datasets import DataManager from datasets import get_class_names from preprocessing import get_image_size from models import SSD300 from metrics import compute_average_precision from metrics import compute_precision_and_recall from utils.boxes import create_prior_boxes from utils.boxes import calculate_intersection_over_union from utils.boxes import denormalize_boxes from utils.inference import infer_from_path dataset_name = 'VOC2007' image_prefix = '../datasets/VOCdevkit/VOC2007/JPEGImages/' weights_path = '../trained_models/SSD300_weights.hdf5' model = SSD300(weights_path=weights_path) prior_boxes = create_prior_boxes() input_shape = model.input_shape[1:3] class_threshold = .1 iou_nms_threshold = .45 iou_threshold = .5 num_classes = 21 image_prefix = '../datasets/VOCdevkit/VOC2007/JPEGImages/' with_difficult_objects = False split = 'test' class_names = get_class_names(dataset_name) class_names = class_names[1:] average_precisions = []
import torch
import torchvision
import torch.nn.functional as F
import torchvision.transforms as transforms
from torch.autograd import Variable
from utils.transforms import resize
from datasets import ListDataset
from models import SSD300, SSDBoxCoder
from PIL import Image, ImageDraw
import os
import sys
print('Loading model..')
net = SSD300(num_classes=2)
#print(net)
model = torch.load('./checkpoint/train_synthtext_detection_2.pth')
net.load_state_dict(model['net'])
net.eval()
print('Loading image..')
img = Image.open('/usr/local/share/data/SynthText/160/stage_40_64.jpg')
ow = oh = 300
img = img.resize((ow, oh))
print('Predicting..')
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225))
])
num_classes = len(class_names)
exit(0)
val_data_manager = CSVDataManager(val_dataset, val_split, class_names, False)
val_data = val_data_manager.load_data()
# generator
prior_boxes = to_point_form(create_prior_boxes())
train_sequencer = SequenceManager(train_data, 'train', prior_boxes,
batch_size, box_scale_factors, num_classes)
val_sequencer = SequenceManager(val_data, 'val', prior_boxes,
batch_size, box_scale_factors, num_classes)
# model
multibox_loss = MultiboxLoss(num_classes, negative_positive_ratio, batch_size)
model = SSD300(image_shape, num_classes, weights_path)
model.compile(optimizer, loss=multibox_loss.compute_loss)
# callbacks
if not os.path.exists(model_path):
os.makedirs(model_path)
checkpoint = ModelCheckpoint(save_path, verbose=1, period=1)
log = CSVLogger(model_path + 'SSD_scratch.log')
learning_rate_manager = LearningRateManager(learning_rate, decay, step_epochs)
learning_rate_schedule = LearningRateScheduler(learning_rate_manager.schedule)
tbCallBack = TensorBoard(log_dir='./Graph/', histogram_freq=0, batch_size=batch_size, write_graph=True)
callbacks = [checkpoint, log, learning_rate_schedule, tbCallBack]
# model fit
xmin = int(round(top_xmin[i] * img.shape[1]))
ymin = int(round(top_ymin[i] * img.shape[0]))
xmax = int(round(top_xmax[i] * img.shape[1]))
ymax = int(round(top_ymax[i] * img.shape[0]))
score = top_conf[i]
label = int(top_label_indices[i])
#label_name = voc_classes[label - 1]
#label_name = self.class_names[label - 1]
label_name = self.class_names[label]
display_txt = '{:0.2f}, {}'.format(score, label_name)
coords = (xmin, ymin), xmax-xmin+1, ymax-ymin+1
color = colors[label]
currentAxis.add_patch(plt.Rectangle(*coords, fill=False, edgecolor=color, linewidth=2))
currentAxis.text(xmin, ymin, display_txt, bbox={'facecolor':color, 'alpha':0.5})
plt.show()
if __name__ == "__main__":
#weights_path = '../trained_models/ssd300_weights.34-1.54.hdf5'
weights_path = '../trained_models/ssd300_weights.17-1.51.hdf5'
#weights_path = '../trained_models/SSD300_COCO_weights.08-0.35.hdf5'
model = SSD300(num_classes=21)
model.load_weights(weights_path)
tester = Tester(model)
tester.test_prior_boxes()
tester.test_XML_parser()
tester.test_image_loading()
tester.test_model()
dataset_manager = DataManager(datasets, splits, class_names, difficult_boxes)
train_data = dataset_manager.load_data()
val_data = test_data = DataManager('VOC2007', 'test').load_data()
class_names = dataset_manager.class_names
num_classes = len(class_names)
# generator
prior_boxes = to_point_form(create_prior_boxes())
generator = ImageGenerator(train_data, val_data, prior_boxes, batch_size,
box_scale_factors, num_classes)
# model
multibox_loss = MultiboxLoss(num_classes, negative_positive_ratio, batch_size)
model = SSD300(image_shape, num_classes, weights_path,
frozen_layers, randomize_top)
model.compile(optimizer, loss=multibox_loss.compute_loss)
# callbacks
if not os.path.exists(model_path):
os.makedirs(model_path)
checkpoint = ModelCheckpoint(save_path, verbose=1, period=1)
log = CSVLogger(model_path + 'SSD_scratch.log')
learning_rate_manager = LearningRateManager(gamma_decay, learning_rate)
learning_rate_schedule = LearningRateScheduler(learning_rate_manager.schedule)
plateau = ReduceLROnPlateau('val_loss', factor=0.9, patience=1, verbose=1)
early_stop = EarlyStopping('val_loss', min_delta=1e-4, patience=14, verbose=1)
callbacks = [checkpoint, log, learning_rate_schedule]
# model fit
plt.show()
def denormalize_box(self, box_coordinates, image_shape):
x_min = box_coordinates[:, 0]
y_min = box_coordinates[:, 1]
x_max = box_coordinates[:, 2]
y_max = box_coordinates[:, 3]
original_image_width, original_image_height = image_shape
x_min = x_min * original_image_width
y_min = y_min * original_image_height
x_max = x_max * original_image_width
y_max = y_max * original_image_height
return [x_min, y_min, x_max, y_max]
if __name__ == '__main__':
import sys
sys.path.insert(0, '../')
from models import SSD300
from prior_box_creator import PriorBoxCreator
from XML_parser import XMLParser
model = SSD300((300, 300, 3))
box_creator = PriorBoxCreator(model)
prior_boxes = box_creator.create_boxes()
data_path = '../../datasets/VOCdevkit/VOC2007/'
ground_truths = XMLParser(data_path + 'Annotations/').get_data()
prior_box_manager = PriorBoxAssigner(prior_boxes, ground_truths)
assigned_boxes = prior_box_manager.assign_boxes()
prior_box_manager.draw_assigned_boxes(data_path + 'JPEGImages/')
scheduled_epochs = [155, 195, 235]
negative_positive_ratio = 3
base_weights_path = '../trained_models/VGG16_weights.hdf5'
# data
class_names = get_class_names('VOC2007')
val_dataset, val_split = 'VOC2007', 'test'
train_datasets, train_splits = ['VOC2007', 'VOC2012'], ['trainval', 'trainval']
train_data_manager = DataManager(train_datasets, train_splits, class_names)
train_data = train_data_manager.load_data()
num_classes = len(class_names)
val_data_manager = DataManager(val_dataset, val_split, class_names)
val_data = val_data_manager.load_data()
# model
model = SSD300(num_classes=num_classes)
model.load_weights(base_weights_path, by_name=True)
prior_boxes = to_point_form(create_prior_boxes())
multibox_loss = MultiboxLoss(num_classes, negative_positive_ratio, alpha_loss)
optimizer = SGD(learning_rate, momentum, weight_decay)
model.compile(optimizer, loss=multibox_loss.compute_loss)
data_generator = DataGenerator(train_data, prior_boxes, batch_size,
num_classes, val_data)
# callbacks
model_path = '../trained_models/' + model_name + '/'
save_path = model_path + 'weights.{epoch:02d}-{val_loss:.2f}.hdf5'
if not os.path.exists(model_path):
os.makedirs(model_path)
log = CSVLogger(model_path + model_name + '.log')
checkpoint = ModelCheckpoint(save_path, verbose=1, save_weights_only=True)
def main():
"""
Training.
"""
global start_epoch, label_map, epoch, checkpoint, decay_lr_at
# Initialize model or load checkpoint
if checkpoint is None:
start_epoch = 0
model = SSD300(n_classes=n_classes)
# Initialize the optimizer, with twice the default learning rate for biases, as in the original Caffe repo
biases = list()
not_biases = list()
for param_name, param in model.named_parameters():
if param.requires_grad:
if param_name.endswith('.bias'):
biases.append(param)
else:
not_biases.append(param)
optimizer = torch.optim.SGD(params=[{
'params': biases,
'lr': 2 * lr
}, {
'params': not_biases
}],
lr=lr,
momentum=momentum,
weight_decay=weight_decay)
else:
checkpoint = torch.load(checkpoint)
start_epoch = checkpoint['epoch'] + 1
print('\nLoaded checkpoint from epoch %d.\n' % start_epoch)
model = checkpoint['model']
optimizer = checkpoint['optimizer']
# Move to default device
model = model.to(device)
criterion = MultiBoxLoss(priors_cxcy=model.priors_cxcy).to(device)
# Custom dataloaders
train_dataset = PascalVOCDataset(data_folder,
split='train',
keep_difficult=keep_difficult)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=batch_size,
shuffle=True,
collate_fn=train_dataset.collate_fn,
num_workers=workers,
pin_memory=True) # note that we're passing the collate function here
# Calculate total number of epochs to train and the epochs to decay learning rate at (i.e. convert iterations to epochs)
# To convert iterations to epochs, divide iterations by the number of iterations per epoch
# The paper trains for 120,000 iterations with a batch size of 32, decays after 80,000 and 100,000 iterations
epochs = iterations // (len(train_dataset) // 32)
decay_lr_at = [it // (len(train_dataset) // 32) for it in decay_lr_at]
# Epochs
for epoch in range(start_epoch, epochs):
# Decay learning rate at particular epochs
if epoch in decay_lr_at:
adjust_learning_rate(optimizer, decay_lr_to)
# One epoch's training
train(train_loader=train_loader,
model=model,
criterion=criterion,
optimizer=optimizer,
epoch=epoch)
# Save checkpoint
save_checkpoint(epoch, model, optimizer)
dataset_manager = DataManager(['VOC2007', 'VOC2012'], ['trainval', 'trainval'])
train_data = dataset_manager.load_data()
val_data = test_data = DataManager('VOC2007', 'test').load_data()
class_names = dataset_manager.class_names
num_classes = len(class_names)
prior_boxes = create_prior_boxes()
generator = ImageGenerator(train_data, val_data, prior_boxes, batch_size)
weights_path = '../trained_models/SSD300_weights.hdf5'
frozen_layers = ['input_1', 'conv1_1', 'conv1_2', 'pool1',
'conv2_1', 'conv2_2', 'pool2',
'conv3_1', 'conv3_2', 'conv3_3', 'pool3']
model = SSD300(image_shape, num_classes, weights_path, frozen_layers, True)
multibox_loss = MultiboxLoss(num_classes, neg_pos_ratio=2.0).compute_loss
model.compile(Adam(lr=3e-4), loss=multibox_loss)
# callbacks
model_path = '../trained_models/SSD_scratch/'
if not os.path.exists(model_path):
os.makedirs(model_path)
save_path = model_path + 'weights.{epoch:02d}-{val_loss:.2f}.hdf5'
checkpoint = ModelCheckpoint(save_path, verbose=1, save_best_only=True)
plateau = ReduceLROnPlateau(factor=0.5, patience=10, verbose=1)
early_stop = EarlyStopping(min_delta=1e-3, patience=25, verbose=1)
log = CSVLogger(model_path + 'SSD_scratch.log')
from models import SSD300
from utils.prior_box_creator_paper import PriorBoxCreator
from utils.prior_box_manager import PriorBoxManager
from utils.XML_parser import XMLParser
from utils.utils import split_data
from utils.utils import scheduler
# constants
batch_size = 7
num_epochs = 15
num_classes = 21
root_prefix = '../datasets/VOCdevkit/VOC2007/'
ground_data_prefix = root_prefix + 'Annotations/'
image_prefix = root_prefix + 'JPEGImages/'
image_shape = (224, 224, 3)
model = SSD300(image_shape, num_classes)
model.load_weights('../trained_models/weights_SSD300.hdf5', by_name=True)
freeze = [
'input_1', 'conv1_1', 'conv1_2', 'pool1', 'conv2_1', 'conv2_2', 'pool2',
'conv3_1', 'conv3_2', 'conv3_3', 'pool3'
]
for layer in model.layers:
if layer.name in freeze:
layer.trainable = False
multibox_loss = MultiboxLoss(num_classes, neg_pos_ratio=2.0).compute_loss
model.compile(optimizer=Adam(lr=3e-4), loss=multibox_loss, metrics=['acc'])
box_creator = PriorBoxCreator(model)
weights_path = '../trained_models/ssd_300_VOC0712.pth'
vgg_weights = torch.load(weights_path)
biases = []
weights = []
for name, torch_weights in vgg_weights.items():
print(name)
if 'bias' in name:
biases.append(torch_weights.numpy())
if 'weight' in name:
print(torch_weights.numpy().shape)
conv_weights = torch_weights.numpy()
conv_weights = np.rollaxis(conv_weights, 0, 4)
conv_weights = np.rollaxis(conv_weights, 0, 3)
weights.append(conv_weights)
vgg_weights = list(zip(weights, biases))
base_model = SSD300(return_base=True)
pytorch_layer_arg = 0
for layer in base_model.layers[1:]:
conv_weights = vgg_weights[pytorch_layer_arg][0]
bias_weights = vgg_weights[pytorch_layer_arg][1]
if ('conv2d' in layer.name) or ('branch' in layer.name):
print(layer.name)
print('pre-trained_weigths:', conv_weights.shape)
print('model weights:', layer.get_weights()[0].shape)
layer.set_weights([conv_weights, bias_weights])
pytorch_layer_arg = pytorch_layer_arg + 1
base_model.save_weights('../trained_models/SSD300_weights.hdf5')
def evaluate_detections(box_list, output_dir, dataset):
write_voc_results_file(box_list, dataset)
do_python_eval(output_dir)
if __name__ == '__main__':
# load net
num_classes = len(VOC_CLASSES) + 1 # +1 background
"""
net = build_ssd('test', 300, num_classes) # initialize SSD
net.load_state_dict(torch.load(args.trained_model))
net.eval()
"""
weights_path = '../trained_models/SSD300_weights.hdf5'
net = SSD300(weights_path=weights_path)
print('Finished loading model!')
# load data
dataset = VOCDetection(args.voc_root, [('2007', set_type)],
BaseTransform(300, dataset_mean),
AnnotationTransform())
# evaluation
cuda = False
test_net(args.save_folder,
net,
cuda,
dataset,
BaseTransform(size=300, mean=dataset_mean),
args.top_k,
300,
thresh=args.confidence_threshold)
from videotest import VideoTest
import sys
sys.path.append("..")
from models import SSD300 as SSD
input_shape = (300, 300, 3)
# Change this if you run with other classes than VOC
class_names = [
"background", "aeroplane", "bicycle", "bird", "boat", "bottle", "bus",
"car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike",
"person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"
]
NUM_CLASSES = len(class_names)
model = SSD(input_shape, num_classes=NUM_CLASSES)
# Change this path if you want to use your own trained weights
model.load_weights('../../trained_models/weights_SSD300.hdf5')
vid_test = VideoTest(class_names, model, input_shape)
# To test on webcam 0, remove the parameter (or change it to another number
# to test on that webcam)
#vid_test.run('path/to/your/video.mkv')
vid_test.run(0)