def train(model_conf:ModelConf, train_set, validation_set=None):
"""
train model for fashion mnist
:param model_conf:
:param train_set:
:param validation_set:
:return:
"""
model_conf.print_conf()
images_placeholder = tf.placeholder(dtype=tf.float32, shape=[None, model_conf.HEIGHT*model_conf.WIDTH])
labels_placeholder = tf.placeholder(dtype=tf.int64, shape=[None])
model = DenseNet(model_conf, is_training=True, images=images_placeholder, labels=labels_placeholder)
model.build_graph()
summary_op = tf.summary.merge_all()
saver = tf.train.Saver(tf.all_variables())
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
summary_writer = tf.summary.FileWriter(model_conf.SUMMARY_DIR, graph=sess.graph)
for step in range(model_conf.NUM_STEPS):
train_images, train_labels = train_set.next_batch()
feed_dict = {images_placeholder: train_images,
labels_placeholder: train_labels
}
start_time = time.time()
[loss_value, acc_value, _, lr_value] = sess.run([model.loss, model.acc, model.train_op, model.learning_rate],
feed_dict=feed_dict)
duration = time.time() - start_time
if step % 10 == 0:
_print_log('training', step, loss_value, acc_value, lr_value, duration)
if step % 100 == 0:
summary_str = sess.run(summary_op, feed_dict=feed_dict)
summary_writer.add_summary(summary_str, step)
if step % 500 == 0:
validation_images = validation_set.images
validation_labels = validation_set.labels
feed_dict = {
images_placeholder: validation_images,
labels_placeholder: validation_labels
}
[loss_value, acc_value] = sess.run([model.loss, model.acc], feed_dict=feed_dict)
logger.info('(validation)loss: %f, acc: %f' % (loss_value, acc_value))
if (step != 0 and step % 10000 == 0) or step + 1 == model_conf.NUM_STEPS:
checkpoint_path = os.path.join(model_conf.MODEL_SAVER_DIR, 'model.ckpt')
saver.save(sess, checkpoint_path, step)
def test(model_conf: ModelConf, test_set):
"""
test the performance of trained model
:param model_conf:
:param test_set:
:return:
"""
images_placeholder = tf.placeholder(dtype=tf.float32, shape=[None, model_conf.HEIGHT*model_conf.WIDTH])
labels_placeholder = tf.placeholder(dtype=tf.int64, shape=[model_conf.BATCH_SIZE])
model = DenseNet(model_conf, is_training=False, images=images_placeholder, labels=labels_placeholder)
model.build_graph()
saver = tf.train.Saver()
sess = tf.Session()
try:
ckpt_state = tf.train.get_checkpoint_state(model_conf.MODEL_SAVER_DIR)
except tf.errors.OutOfRangeError as e:
logger.info('Can not restore checkpoint %s', e)
return
if not (ckpt_state and ckpt_state.model_checkpoint_path):
logger.info('No model to eval yet at %s', model_conf.MODEL_SAVER_DIR)
return
logger.info('Loading checkpoint %s', ckpt_state.model_checkpoint_path)
saver.restore(sess, ckpt_state.model_checkpoint_path)
num_step = int(test_set.num_examples/model_conf.BATCH_SIZE)+1
num_correct = 0
for step in range(num_step):
test_images, test_labels = test_set.next_batch()
feed_dict = {images_placeholder: test_images, labels_placeholder:test_labels}
[loss, prediction, acc] = sess.run([model.loss, model.prediction, model.acc], feed_dict=feed_dict)
num_correct += np.sum(np.argmax(prediction, axis=1) == test_labels)
logger.info("(Test)batch %d, loss=%f, acc=%f" % (step, loss, acc))
precision = num_correct*1.0/(num_step*model_conf.BATCH_SIZE)
logger.info("Total precision: %f" % precision)
args.bc_mode = True
model_params = vars(args)
if not args.train and not args.test:
print("You should train or test your network. Please check params.")
exit()
# some default params dataset/architecture related
train_params = get_train_params_by_name(args.dataset)
print("Params:")
for k, v in model_params.items():
print("\t%s: %s" % (k, v))
print("Train params:")
for k, v in train_params.items():
print("\t%s: %s" % (k, v))
print("Prepare training data...")
data_provider = get_data_provider_by_name(args.dataset, train_params)
print("Initialize the model..")
model = DenseNet(data_provider=data_provider, **model_params)
if args.train:
print("Data provider train images: ", data_provider.train.num_examples)
model.train_all_epochs(train_params)
if args.test:
if not args.train:
model.load_model()
print("Data provider test images: ", data_provider.test.num_examples)
print("Testing...")
loss, accuracy = model.test(data_provider.test, batch_size=200)
print("mean cross_entropy: %f, mean accuracy: %f" % (loss, accuracy))
print('Run config:')
for k, v in run_config.get_config().items():
print('\t%s: %s' % (k, v))
print('Network config:')
for k, v in net_config.items():
print('\t%s: %s' % (k, v))
print('Prepare training data...')
data_provider = get_data_provider_by_name(run_config.dataset,
run_config.get_config())
# set net config
net_config = DenseNetConfig().set_standard_dense_net(
data_provider=data_provider, **net_config)
print('Initialize the model...')
model = DenseNet(args.path, data_provider, run_config, net_config)
# save configs
if args.save_config:
model.save_config(args.path)
if args.load_model: model.load_model()
if args.test:
# test
print('Data provider test images: ', data_provider.test.num_examples)
print('Testing...')
loss, accuracy = model.test(data_provider.test, batch_size=200)
print('mean cross_entropy: %f, mean accuracy: %f' % (loss, accuracy))
json.dump({
'test_loss': '%s' % loss,
'test_acc': '%s' % accuracy
if not args.train and not args.test:
print("You should train or test your network. Please check params.")
exit()
# some default params dataset/architecture related
train_params = train_params_set
print("Params:")
for k, v in model_params.items():
print("\t%s: %s" % (k, v))
print("Train params:")
for k, v in train_params.items():
print("\t%s: %s" % (k, v))
print("Prepare training data...")
data_provider = DataProvider()
print("Initialize the model..")
model = DenseNet(data_provider=data_provider, **model_params)
if args.train:
# print("Data provider train images: ", np.shape(data_provider.train_imgs))
print("Data provider train images: ")
model.train_all_epochs(train_params)
if args.test:
# model.load_model()
# print("Data provider test images: ", np.shape(data_provider.test_imgs))
print("Testing...")
model.test()
# model.test(data_provider, batch_size=10)
# loss, accuracy = model.test(data_provider, batch_size=200)
# print("mean cross_entropy: %f, mean accuracy: %f" % (loss, accuracy))
def extract_feature_batch():
args = parse_opts()
model_params = vars(args)
batch_size = args.batch_size
print("Initialize the model..")
# fake data_provider
DataProvider = collections.namedtuple('DataProvider', ['data_shape', 'n_classes'])
data_provider = DataProvider(data_shape=(img_size, img_size, 1), n_classes=10)
model = DenseNet(data_provider=data_provider, **model_params)
end_points = model.end_points
# for key, value in end_points.iteritems():
# print(key, value.get_shape().as_list())
# restore model
model.saver.restore(model.sess, args.model_path)
print("Successfully load model from model path: %s" % args.model_path)
video_names = [x for x in os.listdir(args.face_dir)]
video_names.sort()
avg_num_imgs = 0
for vid, video_name in enumerate(video_names):
video_dir = os.path.join(args.face_dir, video_name)
img_paths = os.listdir(video_dir)
if len(img_paths) == 0:
continue
output_subdir = os.path.join(args.outft_dir, video_name)
if os.path.exists(output_subdir):
continue
else:
os.makedirs(output_subdir)
img_paths.sort(key=lambda x:int(x.split('.')[0]))
avg_num_imgs += len(img_paths)
imgs = []
for img_path in img_paths:
img_path = os.path.join(video_dir, img_path)
img = cv2.imread(img_path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = cv2.resize(img, (img_size, img_size))
imgs.append(img)
imgs = (np.array(imgs, np.float32) - images_mean) / images_std
imgs = np.expand_dims(imgs, 3)
# pool4.shape=(batch_size, 4, 4, 256)
# fc5.shape=fc6.shape=(batch_size, 1, 1, 512)
# prob.shape=(batch_size, num_classes)
fcs, probs = [], []
for i in xrange(0, imgs.shape[0], batch_size):
feed_dict = {
model.images: imgs[i: i + batch_size],
model.is_training: False
}
fc, prob = model.sess.run(
[end_points['fc'], end_points['preds']],
feed_dict=feed_dict)
# prev_last_pools.extend(prev_last_pool)
fcs.extend(fc)
probs.extend(prob)
# prev_last_pools = np.array(prev_last_pools, np.float32)
fcs = np.array(fcs, np.float32)
probs = np.array(probs, np.float32)
# with open(os.path.join(output_subdir, 'pool.npy'), 'wb') as f:
# np.save(f, prev_last_pools)
with open(os.path.join(output_subdir, 'fc.npy'), 'wb') as f:
np.save(f, fcs)
with open(os.path.join(output_subdir, 'prob.npy'), 'wb') as f:
np.save(f, probs)
print(vid, video_name, len(img_paths),
fcs.shape, probs.shape)
avg_num_imgs /= float(len(video_names))
print('average faces per video', avg_num_imgs)
torch_device = torch.device("cuda")
# load images that you want to visualize
log_path = save_path + '/TP.txt'
# For example, TP.txt file contains name of image files manully selected from
# the test set, which are classified as true positive images.
# ex) 20080169_1_1-0-74-16.13539388
f = open(log_path, 'r')
lines = f.readlines()
lines = np.array(lines)
lines = np.char.rstrip(lines)
model = DenseNet(growthRate=12,
depth=40,
reduction=0.5,
bottleneck=True,
nClasses=2)
model = nn.DataParallel(model).to(torch_device)
model.load_state_dict(ckpoint['network'])
# by changing 'hooks', you can choose the layer which you want to visualize
# to check name of layers, use model.module to print out the whole structure of the neural net
grad_cam = GradCam(model=model.module,
hooks=["dense2"],
device=torch_device)
test_path = "/home/intern/osteo_classification/data/test/"
test_loader = loader(test_path,
1,
transform=None,
sampler='',
print("You should train or test your network. Please check arguments.")
exit()
# Get model params (the arguments) and train params (depend on dataset).
model_params = vars(args)
train_params = get_train_params_by_name(args.dataset)
print("\nModel parameters (specified as arguments):")
for k, v in model_params.items():
print("\t%s: %s" % (k, v))
print("Train parameters (depend on specified dataset):")
for k, v in train_params.items():
print("\t%s: %s" % (k, v))
# Train and/or test the specified model.
print("\nPrepare training data...")
data_provider = get_data_provider_by_name(args.dataset, train_params)
print("Initialize the model...")
model = DenseNet(data_provider=data_provider, **model_params)
if args.train:
print("Data provider train images: ", data_provider.train.num_examples)
model.train_all_epochs(train_params)
if args.test:
if not args.train:
model.load_model()
print("Data provider test images: ", data_provider.test.num_examples)
print("Testing...")
loss, accuracy = model.test(data_provider.test, batch_size=200)
model.print_pertinent_features(loss, accuracy, -1, True)
print("mean cross_entropy: %f, mean accuracy: %f" %
(loss[-1], accuracy))
print("You should train or test your network. Please check params.")
exit()
# some default params dataset/architecture related
train_params = get_train_params_by_name(args.dataset)
print("Params:")
for k, v in model_params.items():
print("\t%s: %s" % (k, v))
print("Train params:")
for k, v in train_params.items():
print("\t%s: %s" % (k, v))
print("Prepare training data...")
data_provider = get_data_provider_by_name(args.dataset, train_params)
print("Initialize the model..")
model = DenseNet(data_provider=data_provider, **model_params)
model_path = 'saves/%s_growth_rate=%d_depth=%d_dataset_%s/' % (
args.model_type, args.growth_rate, args.depth, args.dataset)
print(model_path)
if args.train:
current_step = model.load_model(model_path)
print("Data provider train images: ", data_provider.train.num_examples)
model.train_all_epochs(train_params, args.epoch, current_step)
if args.test:
if not args.train:
model.load_model(model_path)
print("Data provider test images: ", data_provider.test.num_examples)
print("Testing...")
loss, accuracy = model.test(data_provider.test, batch_size=200)
print("mean cross_entropy: %f, mean accuracy: %f" % (loss, accuracy))
def main(unused_argv):
# Load training and eval data
# export_cifar('/backups/datasets/cifar-10-python.tar.gz', '/backups/work/CIFAR10')
imagenet = ImageNet('/backups/work/ILSVRC2017/ILSVRC',
shuffle=True,
normalize=True,
augment=False,
one_hot=False,
batch_size=128)
# imagenet = CIFAR('/backups/work/CIFAR10',
# shuffle=True, normalize=True, augment=True, one_hot=False, batch_size=32)
densenet = DenseNet(num_classes=imagenet.num_classes,
growth_rate=12,
bc_mode=True,
block_config=(6, 12, 24, 16),
dropout_rate=0.2,
reduction=0.5,
weight_decay=1e-4,
nesterov_momentum=0.9)
def train_input_fn(learning_rate):
dataset = imagenet.train_set
dataset = dataset.repeat(1)
dataset = dataset.skip(imagenet.train_set_size % imagenet.batch_size)
iterator = dataset.make_one_shot_iterator()
features, labels = iterator.get_next()
return {'images': features, 'learning_rate': learning_rate}, labels
def eval_input_fn():
dataset = imagenet.val_set
dataset = dataset.repeat(1)
iterator = dataset.make_one_shot_iterator()
features, labels = iterator.get_next()
return {'images': features}, labels
# Create the Estimator
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
config = tf.estimator.RunConfig().replace(session_config=sess_config)
classifier = tf.estimator.Estimator(
model_fn=densenet.imagenet_model_fn,
model_dir="/backups/work/logs/imagenet_model1",
# params={'image_shape': imagenet.image_shape},
config=config)
# Set up logging
# tensors_to_log = {"accuracy": "Accuracy/train_accuracy"}
# logging_hook = tf.train.LoggingTensorHook(
# tensors=tensors_to_log, every_n_iter=100)
# debug_hook = tfdbg.LocalCLIDebugHook()
# debug_hook.add_tensor_filter("has_inf_or_nan", tfdbg.has_inf_or_nan)
# # Train the model
for i in range(30):
# Train the model
classifier.train(input_fn=lambda: train_input_fn(learning_rate=0.1))
# Evaluate the model and print results
eval_results = classifier.evaluate(input_fn=eval_input_fn)
print(eval_results)
for i in range(30):
# Train the model
classifier.train(input_fn=lambda: train_input_fn(learning_rate=0.01))
# Evaluate the model and print results
eval_results = classifier.evaluate(input_fn=eval_input_fn)
print(eval_results)
for i in range(30):
# Train the model
classifier.train(input_fn=lambda: train_input_fn(learning_rate=0.001))
# Evaluate the model and print results
eval_results = classifier.evaluate(input_fn=eval_input_fn)
print(eval_results)
for i in range(10):
# Train the model
classifier.train(input_fn=lambda: train_input_fn(learning_rate=0.0001))
# Evaluate the model and print results
eval_results = classifier.evaluate(input_fn=eval_input_fn)
print(eval_results)
# some default params dataset/architecture related
train_params = get_train_params_by_name(args.dataset)
print("Params:")
for k, v in model_params.items():
print("\t%s: %s" % (k, v))
print("Train params:")
for k, v in train_params.items():
print("\t%s: %s" % (k, v))
print("Prepare training data...")
data_provider = get_data_provider_by_name(args.dataset, train_params)
print("Initialize the model..")
model = DenseNet(for_test_only=False,
init_variables=None,
init_global=None,
bottleneck_output_size=None,
first_output_features=None,
data_provider=data_provider,
**model_params)
if args.train:
print("Data provider train images: ", data_provider.train.num_examples)
model.train_all_epochs(train_params)
if args.comprese:
if not args.train:
model.load_model()
old_varaible = model.get_trainable_variables_value()
old_param_num = model.total_parameters
print("Commpresing the network")
comprese_model = CompreseDenseNet(model, args.clusster_num)
all_new_comprese_kernels, all_new_bottleneck_kernels, all_new_batch_norm, all_new_transion_kernels, all_new_batch_norm_for_transion, new_W, new_transion_to_class_batch_norm = comprese_model.comprese(
)
drop_last=True,
late_fusion=arg.late_fuse)
# the default loss function is the cross entropy loss
class_loss = nn.CrossEntropyLoss()
if arg.late_fuse:
net = LateFuse_DenseNet(growthRate=arg.growthRate,
depth=arg.depth,
reduction=0.5,
bottleneck=True,
nClasses=2)
net = nn.DataParallel(net).to(torch_device)
model = LateFuseClassifyTrainer(arg,
net,
torch_device,
class_loss=class_loss)
else:
net = DenseNet(growthRate=arg.growthRate,
depth=arg.depth,
reduction=0.5,
bottleneck=True,
nClasses=2)
net = nn.DataParallel(net).to(torch_device)
model = ClassifyTrainer(arg, ent, torch_device, class_loss=class_loss)
if arg.test is False:
model.train(train_loader, val_loader)
model.test(test_loader, val_loader)