def load_my_model(arch_file=None, weight_file=None):
if arch_file is None:
model = resnet_model()
with open(arch_file, 'r') as json_file:
architecture = json.load(json_file)
model = model_from_json(json.dumps(architecture))
# model = model_from_json(arch_file)
model.load_weights(weight_file)
return model
def make_model(args, gids=None):
"""
Initialize ResNet-50 model.
"""
if 'softmax' in args.loss_type:
if args.dataset == 'market1501':
num_classes = 751
elif args.dataset == 'duke':
num_classes = 702
else:
raise NotImplementedError
model = resnet_model(num_classes=num_classes, include_top=True,
remove_downsample=args.remove_downsample)
else:
model = resnet_model(remove_downsample=args.remove_downsample)
if gids is not None:
model = model.cuda(gids[0])
if len(gids) > 1:
model = DataParallel(model, gids)
return model
def mnist_model(image, labels, mesh):
"""The model.
Args:
image: tf.Tensor with shape [batch, 28*28]
labels: a tf.Tensor with shape [batch] and dtype tf.int32
mesh: a mtf.Mesh
Returns:
logits: a mtf.Tensor with shape [batch, 10]
loss: a mtf.Tensor with shape []
"""
batch_dim = mtf.Dimension("batch", FLAGS.batch_size)
rows_dim = mtf.Dimension("rows_size", image_height)
cols_dim = mtf.Dimension("cols_size", image_width)
channel_dim = mtf.Dimension("image_channel", num_channels)
classes_dim = mtf.Dimension(name='classesnum', size=classesnum)
x = mtf.import_tf_tensor(
mesh,
tf.reshape(
image,
[FLAGS.batch_size, image_height, image_width, num_channels]),
mtf.Shape([batch_dim, rows_dim, cols_dim, channel_dim]))
# x = mtf.transpose(x, [batch_dim, rows_dim, cols_dim, channel_dim])
# print(x.shape)
logits = resnet_model(x, classes_dim=classes_dim, depth=depth)
logits = mtf.cast(logits, dtype=tf.float32)
if labels is None:
loss = None
else:
labels = mtf.import_tf_tensor(mesh,
tf.reshape(labels, [FLAGS.batch_size]),
mtf.Shape([batch_dim]))
loss = mtf.layers.softmax_cross_entropy_with_logits(
logits, mtf.one_hot(labels, classes_dim), classes_dim)
loss = mtf.reduce_mean(loss)
return logits, loss
input_path = os.path.join(DATA_DIR,'*','train-*.tfrecord')
mode = 'validation'
else:
input_path = os.path.join(DATA_DIR, 'test', 'test.tfrecord')
mode='test'
# Creating data iterator based on tensorflow Dataset api
test_iter, test_data = data.extract_image(input_path, BATCH_SIZE, SEED, mode=mode)
input_shape = (None, 224, 224, 3)
label_shape = (None, 6)
# Loading base model whose embedding layer will be extracted
if base_model == 'resnet_50':
_, x, _, is_train, features = model.resnet_model(input_shape, label_shape)
feature_size = 2048
elif base_model == 'inception_resnet':
_, x, _, is_train, features = model.inception_resnet_model(input_shape, label_shape)
feature_size = 1536
else:
AssertionError('Unknown base model present')
sess = tf.Session()
# Loading trained model checkpoint
saver = utils.load_checkpoint(os.path.join(model_dir,'model'), sess, 'model')
saver_unloaded = utils.load_checkpoint(os.path.join(model_dir,'model_unsaved'), sess, 'model_unsaved')
# Extracting feature dictionary for the provided dataset
def main(gid=None,
dataset=None,
dataset_root=None,
which=None,
exp_dir=None,
verbose=False):
"""
Configs
"""
GPU_ID = 0 # gpu id or 'None'
BATCH_SIZE = 32 # batch size when extracting query and gallery features
IMG_SIZE = (256, 128)
DATASET = 'market1501' # market1501, duke
WHICH = 'last' # which model to load
EXP_DIR = './exp/dmml/market1501'
NORMALIZE_FEATURE = True # whether to normalize features in evaluation
NUM_WORKERS = 8
if gid is not None:
GPU_ID = gid
if dataset is not None:
DATASET = dataset
if which is not None:
WHICH = which
if exp_dir is not None:
EXP_DIR = exp_dir
"""
Datasets
"""
if dataset_root is None:
if DATASET == 'market1501':
dataset_root = './datasets/market1501'
elif DATASET == 'duke':
dataset_root = './datasets/duke'
else:
raise NotImplementedError
print('Generating dataset...')
eval_transform = transforms.Compose([
transforms.Resize(IMG_SIZE, interpolation=3),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])
if DATASET == 'market1501':
datasets = {
x: Market1501(dataset_root, transform=eval_transform, split=x)
for x in ['gallery', 'query']
}
num_classes = 751
elif DATASET == 'duke':
datasets = {
x: DukeMTMC_reID(dataset_root, transform=eval_transform, split=x)
for x in ['gallery', 'query']
}
num_classes = 702
dataloaders = {
x: torch.utils.data.DataLoader(datasets[x],
batch_size=BATCH_SIZE,
shuffle=False,
num_workers=NUM_WORKERS)
for x in ['gallery', 'query']
}
print('Done.')
"""
Model
"""
print('Restoring model...')
### You may need to modify the arguments of the model according to your training settings.
model = resnet_model(remove_downsample=True)
# model = resnet_model(num_classes=num_classes, include_top=False, remove_downsample=False)
model.load_state_dict(torch.load('{}/model_{}.pth'.format(EXP_DIR, WHICH)),
strict=False)
if GPU_ID is not None:
model.cuda(GPU_ID)
model.eval()
print('Done.')
"""
Test
"""
print('Getting image ID...')
gallery_cam, gallery_label = get_id(datasets['gallery'].imgs,
dataset=DATASET)
query_cam, query_label = get_id(datasets['query'].imgs, dataset=DATASET)
print('Done.')
# Extract feature
print('Extracting gallery feature...')
gallery_feature, g_images = extract_feature(
model,
dataloaders['gallery'],
normalize_feature=NORMALIZE_FEATURE,
gid=GPU_ID,
verbose=verbose)
print('Done.')
print('Extracting query feature...')
query_feature, q_images = extract_feature(
model,
dataloaders['query'],
normalize_feature=NORMALIZE_FEATURE,
gid=GPU_ID,
verbose=verbose)
print('Done.')
query_cam = np.array(query_cam)
query_label = np.array(query_label)
gallery_cam = np.array(gallery_cam)
gallery_label = np.array(gallery_label)
# Evaluate
print('Evaluating...')
CMC = torch.IntTensor(len(gallery_label)).zero_()
ap = 0.0
for i in range(len(query_label)):
ap_tmp, CMC_tmp = evaluate(query_feature[i], query_label[i],
query_cam[i], gallery_feature,
gallery_label, gallery_cam)
if CMC_tmp[0] == -1:
continue
CMC = CMC + CMC_tmp
ap += ap_tmp
CMC = CMC.float()
CMC = CMC / len(query_label) # average CMC
print('Done.')
print('Rank-1: {:.6f} Rank-5: {:.6f} Rank-10: {:.6f} mAP: {:.6f}'.format(
CMC[0].item(), CMC[4].item(), CMC[9].item(), ap / len(query_label)))
return ap / len(query_label), CMC
DATA_DIR = ARGS.data_dir
if ARGS.model_type == 'base':
base_model = config['pretrained_model']
input_path = os.path.join(DATA_DIR, 'test', 'test.tfrecord')
test_iter, test_data = data.extract_image(input_path,
BATCH_SIZE,
SEED,
mode='test')
input_shape = (None, 224, 224, 3)
label_shape = (None, 6)
if base_model == 'resnet_50':
preds, x, y, is_train, _ = model.resnet_model(input_shape, label_shape)
elif base_model == 'inception_resnet':
preds, x, y, is_train, _ = model.inception_resnet_model(
input_shape, label_shape)
else:
AssertionError('Unknown base model present')
sess = tf.Session()
saver = utils.load_checkpoint(os.path.join(model_dir, 'model'), sess,
'model')
saver_unloaded = utils.load_checkpoint(
os.path.join(model_dir, 'model_unsaved'), sess, 'model_unsaved')
data_pl = [x, is_train]
def main(argv):
# turn off log message
tf.compat.v1.logging.set_verbosity(tf.compat.v1.logging.FATAL)
# check folder
if not (os.path.exists(os.path.join(FLAGS.outdir, 'tensorboard'))):
os.makedirs(os.path.join(FLAGS.outdir, 'tensorboard'))
if not (os.path.exists(os.path.join(FLAGS.outdir, 'model'))):
os.makedirs(os.path.join(FLAGS.outdir, 'model'))
# load train data(cifar10, class: 10)
if FLAGS.problem == 'cifar10':
(x_train, y_train), (x_test,
y_test) = tf.keras.datasets.cifar10.load_data()
num_class = 10
# load train data(cifar100, class: 100)
if FLAGS.problem == 'cifar100':
(x_train, y_train), (x_test,
y_test) = tf.keras.datasets.cifar100.load_data(
label_mode='fine')
num_class = 100
# preprocess
train_gen = tf.keras.preprocessing.image.ImageDataGenerator(
rescale=1.0 / 255,
horizontal_flip=True,
width_shift_range=4.0 / 32.0,
height_shift_range=4.0 / 32.0)
test_gen = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1.0 /
255)
# initializer
init_op = tf.group(tf.initializers.global_variables(),
tf.initializers.local_variables())
with tf.Session(config=utils.config(index=FLAGS.gpu_index)) as sess:
if FLAGS.is_octconv:
network = octconv_resnet50
else:
network = normal_resnet50
# set network
kwargs = {
'sess': sess,
'outdir': FLAGS.outdir,
'input_size': FLAGS.image_size,
'alpha': FLAGS.alpha,
'network': network,
'num_class': num_class,
'is_training': True,
'learning_rate': 1e-3
}
Model = resnet_model(**kwargs)
utils.cal_parameter()
# prepare tensorboard
writer_train = tf.summary.FileWriter(
os.path.join(FLAGS.outdir, 'tensorboard', 'train'), sess.graph)
value_loss = tf.Variable(0.0)
tf.summary.scalar("train_loss", value_loss)
merge_op = tf.summary.merge_all()
# initialize
sess.run(init_op)
tbar = tqdm(range(FLAGS.epoch), ascii=True)
epoch_loss = []
for i in tbar:
train_data_shuffled = train_gen.flow(x_train,
y_train,
FLAGS.batch_size,
shuffle=True)
# one epoch
for iter in range(x_train.shape[0] // FLAGS.batch_size):
train_data_batch = next(train_data_shuffled)
label = tf.keras.utils.to_categorical(train_data_batch[1],
num_classes=num_class)
# training
train_loss = Model.update(train_data_batch[0], label)
epoch_loss.append(np.mean(train_loss))
s = "epoch:{}, step:{}, Loss: {:.4f}".format(
i, iter, np.mean(epoch_loss))
tbar.set_description(s)
summary_train_loss = sess.run(merge_op,
{value_loss: np.mean(epoch_loss)})
writer_train.add_summary(summary_train_loss, i)
epoch_loss.clear()
# save model
Model.save_model(i)
train_data_gen = image_generator.flow_from_directory(directory=str(data_path),
batch_size=BATCH_SIZE,
shuffle=True,
target_size=(IMG_HEIGHT,
IMG_WIDTH),
classes=list(CLASS_NAMES))
test_data_gen = image_generator.flow_from_directory(
directory=str(data_path_validation),
batch_size=BATCH_SIZE,
shuffle=False,
target_size=(IMG_HEIGHT, IMG_WIDTH),
classes=list(CLASS_NAMES))
IMG_SHAPE = (IMG_HEIGHT, IMG_WIDTH, 3)
model = resnet_model(IMG_SHAPE)
loss_object = tf.keras.losses.CategoricalCrossentropy()
optimizer = tf.keras.optimizers.Adam(learning_rate=0.00001)
writer = tf.summary.create_file_writer(
"/tmp/mylogs/test_classification_dataaugmentation_{}".format(timestamp))
@tf.function
def train_step(images, labels, step):
with tf.GradientTape() as tape:
predictions = model(images)
loss = loss_object(labels, predictions)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
eq = tf.equal(tf.argmax(labels, -1), tf.argmax(predictions, -1))
def main(argv):
# turn off log message
tf.logging.set_verbosity(tf.logging.WARN)
# check folder
if not (os.path.exists(os.path.join(FLAGS.outdir, 'tensorboard'))):
os.makedirs(os.path.join(FLAGS.outdir, 'tensorboard'))
# load train data(cifar10, class: 10)
if FLAGS.problem == 'cifar10':
(x_train, y_train), (x_test,
y_test) = tf.keras.datasets.cifar10.load_data()
num_class = 10
# load train data(cifar100, class: 100)
if FLAGS.problem == 'cifar100':
(x_train, y_train), (x_test,
y_test) = tf.keras.datasets.cifar100.load_data(
label_mode='fine')
num_class = 100
# preprocess
test_gen = tf.keras.preprocessing.image.ImageDataGenerator(rescale=1.0 /
255)
# initializer
init_op = tf.group(tf.initializers.global_variables(),
tf.initializers.local_variables())
with tf.Session(config=utils.config(index=FLAGS.gpu_index)) as sess:
if FLAGS.is_octconv:
network = octconv_resnet50
else:
network = normal_resnet50
# set network
kwargs = {
'sess': sess,
'outdir': FLAGS.outdir,
'input_size': FLAGS.image_size,
'alpha': FLAGS.alpha,
'network': network,
'num_class': num_class,
'is_training': False,
'learning_rate': 1e-4
}
Model = resnet_model(**kwargs)
utils.cal_parameter()
# prepare tensorboard
writer_test = tf.summary.FileWriter(
os.path.join(FLAGS.outdir, 'tensorboard', 'test'))
value_acc = tf.Variable(0.0)
tf.summary.scalar("test_accuracy", value_acc)
merge_op = tf.summary.merge_all()
# initialize
sess.run(init_op)
tbar = tqdm(range(FLAGS.epoch), ascii=True)
epoch_acc = []
for i in tbar:
test_data = test_gen.flow(x_test,
y_test,
FLAGS.batch_size,
shuffle=False)
# one epoch
Model.restore_model(FLAGS.model_path + '/model_{}'.format(i))
for iter in range(x_test.shape[0] // FLAGS.batch_size):
train_data_batch = next(test_data)
label = tf.keras.utils.to_categorical(train_data_batch[1],
num_classes=num_class)
test_acc = Model.test(train_data_batch[0], label)
epoch_acc.append(np.mean(test_acc))
s = "epoch:{}, acc: {:.4f}".format(i, np.mean(epoch_acc))
tbar.set_description(s)
summary_test_acc = sess.run(merge_op,
{value_acc: np.mean(epoch_acc)})
writer_test.add_summary(summary_test_acc, i)
epoch_acc.clear()
train_input_path = os.path.join(ARGS.data_dir, 'train', 'train-*.tfrecord')
val_input_path = os.path.join(ARGS.data_dir, 'validation', 'train-*.tfrecord')
SEED = 88
input_shape = (None, 224, 224, 3)
label_shape = (None, 6)
# Creating save directory where model checkpoints and
# tensorboard files will be stored
save_dir = utils.create_save_dir(ARGS, EXPERIMENT_TIMESTAMP, 'base_model')
# loading resnet-50 or inception resnet computation
# graph and adding loss and optimizer to the graph
if ARGS.pretrained_model == 'resnet_50':
outputs, inputs, labels, is_train, _ = model.resnet_model(
input_shape, label_shape)
if ARGS.pretrained_model == 'inception_resnet':
outputs, inputs, labels, is_train, _ = model.inception_resnet_model(
input_shape, label_shape)
loss = model.build_loss(labels, outputs, loss_name=ARGS.loss)
prediction = tf.cast(tf.greater(tf.sigmoid(outputs), 0.5), tf.float32)
correct_prediction = tf.equal(labels[:, -1], prediction[:, -1])
correct_images = tf.boolean_mask(inputs, correct_prediction)
wrong_images = tf.boolean_mask(inputs, tf.logical_not(correct_prediction))
train_loss = tf.summary.scalar('train_loss', loss)
validation_loss = tf.summary.scalar('val_loss', loss)
# val_correct_img = tf.summary.image('correct_imgs', correct_images, max_outputs=5)
# val_wrong_img = tf.summary.image('wrong_imgs', wrong_images, max_outputs=5)
def main(argv):
# check folder
if not (os.path.exists(os.path.join(FLAGS.outdir, 'tensorboard'))):
os.makedirs(os.path.join(FLAGS.outdir, 'tensorboard'))
if not (os.path.exists(os.path.join(FLAGS.outdir, 'model'))):
os.makedirs(os.path.join(FLAGS.outdir, 'model'))
# get file list
train_data_list = glob.glob(FLAGS.indir + '/*')
# shuffle list
random.shuffle(train_data_list)
# load train data
train_set = tf.data.Dataset.list_files(train_data_list)
train_set = train_set.apply(
tf.contrib.data.parallel_interleave(
lambda x: tf.data.TFRecordDataset(x), cycle_length=os.cpu_count()))
train_set = train_set.map(utils._parse_function,
num_parallel_calls=os.cpu_count())
train_set = train_set.shuffle(buffer_size=FLAGS.shuffle_buffer_size)
train_set = train_set.repeat()
train_set = train_set.batch(FLAGS.batch_size)
train_set = train_set.prefetch(1)
train_iter = train_set.make_one_shot_iterator()
train_data = train_iter.get_next()
# step of each epoch
if FLAGS.num_data % FLAGS.batch_size == 0:
step_of_epoch = FLAGS.num_data / FLAGS.batch_size
else:
step_of_epoch = FLAGS.num_data // FLAGS.batch_size + 1
# initializer
init_op = tf.group(tf.initializers.global_variables(),
tf.initializers.local_variables())
with tf.Session(config=utils.config(index=FLAGS.gpu_index)) as sess:
if FLAGS.is_octconv:
network = octconv_resnet50
else:
network = normal_resnet50
# set network
kwargs = {
'sess': sess,
'outdir': FLAGS.outdir,
'input_size': FLAGS.image_size,
'alpha': FLAGS.alpha,
'network': network,
'is_training': True,
'learning_rate': 1e-4
}
Model = resnet_model(**kwargs)
utils.cal_parameter()
# prepare tensorboard
writer_train = tf.summary.FileWriter(
os.path.join(FLAGS.outdir, 'tensorboard', 'train'), sess.graph)
value_loss = tf.Variable(0.0)
tf.summary.scalar("train_loss", value_loss)
merge_op = tf.summary.merge_all()
# initialize
sess.run(init_op)
tbar = tqdm(range(FLAGS.epoch), ascii=True)
epoch_loss = []
for i in tbar:
# one epoch
for step in range(step_of_epoch):
train_data_batch, train_label_batch = sess.run(train_data)
train_loss = Model.update(train_data_batch, train_label_batch)
epoch_loss.append(train_loss)
s = "epoch:{}, step:{}, Loss: {:.4f}".format(
i, step, np.mean(epoch_loss))
tbar.set_description(s)
summary_train_loss = sess.run(merge_op,
{value_loss: np.mean(epoch_loss)})
writer_train.add_summary(summary_train_loss, i)
epoch_loss.clear()
# save model
Model.save_model(i)
from tensorflow.keras.preprocessing import image
from model import resnet_model, activation_model
import numpy as np
import matplotlib.pyplot as plt
import scipy as sp
import glob
if __name__ == '__main__':
# image size is default of resnet
image_size = (224, 244, 3)
# image directory
imgfiles = glob.glob('./*.png')
# output probs (Dense) ,(None, 1000)
resnet = resnet_model(1000)
# this model is resnet till last conv, not Denses
# output post_relu (Activation) (None, 7, 7, 2048)
activation_layer = resnet.get_layer('post_relu')
model = activation_model(resnet)
# getting last layer weights
final = resnet.get_layer('predictions')
w = final.get_weights()
# input processing
x = np.expand_dims(image.img_to_array(
image.load_img(imgfiles[2], target_size=(224, 224, 3))),
axis=0)
x = preprocess_input(x)