def load_model(modelfile, config=None, args=None):
if config.amp:
from apex import amp
if args.debug: print("Using apex amp")
if modelfile == None:
sys.stderr.write("No model file specified!")
sys.exit(1)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
if args.debug: print("Using device:", device)
if args.arch is not None:
model = network.network(config=config, arch=args.arch).to(device)
else:
model = network.network(config=config).to(device)
if args.debug: print("Loading pretrained weights:", modelfile)
state_dict = torch.load(modelfile)["state_dict"]
if "state_dict" in state_dict:
model.load_state_dict(convert_statedict(state_dict["state_dict"]))
else:
model.load_state_dict(torch.load(modelfile)["state_dict"])
if args.debug: print(model)
model.eval()
torch.set_grad_enabled(False)
optimizer = torch.optim.Adam(model.parameters(), lr=config.lr)
if config.amp:
model, optimizer = amp.initialize(model,
optimizer,
opt_level="O1",
verbosity=0)
return model, device
def main(train_dir, batch_size, num_batches, log_dir, checkpoint_dir=None):
if checkpoint_dir is None:
checkpoint_dir = log_dir
with tf.device('/cpu:0'):
images, labels = build_input('cifar10', 100, 'test')
images_cat_0, labels_cat_0 = build_input_cat_0('cifar10', 100, 'test')
images_cat_1, labels_cat_1 = build_input_cat_1('cifar10', 100, 'test')
predictions, total_loss, labels_cat2 = network(images, labels)
predictions_cat_0, loss_0 = network(images_cat_0, labels_cat_0)
predictions_cat_1, loss_1 = network(images_cat_1, labels_cat_1)
tf.summary.scalar('loss', total_loss)
tf.summary.scalar('loss_0', loss_0)
tf.summary.scalar('loss_1', loss_1)
predictions_cat_0 = tf.argmax(predictions_cat_0, axis=1)
predictions_cat_1 = tf.argmax(predictions_cat_1, axis=1)
predictions = tf.argmax(predictions, 1)
tf.summary.scalar(
'accuracy_cat_0',
slim.metrics.accuracy(predictions_cat_0,
tf.to_int64(labels_cat_0)))
tf.summary.scalar(
'accuracy_cat_1',
slim.metrics.accuracy(predictions_cat_1,
tf.to_int64(labels_cat_1)))
tf.summary.scalar('accuracy_cat2',
slim.metrics.accuracy(predictions, labels_cat2))
# These are streaming metrics which compute the "running" metric,
# e.g running accuracy
metrics_to_values, metrics_to_updates = slim.metrics.aggregate_metric_map(
{
'accuracy_cat2':
slim.metrics.streaming_accuracy(predictions, labels_cat2),
'accuracy_cat_0':
slim.metrics.streaming_accuracy(predictions_cat_0,
labels_cat_0),
'accuracy_cat_1':
slim.metrics.streaming_accuracy(predictions_cat_1,
labels_cat_1),
})
# Define the streaming summaries to write:
for metric_name, metric_value in metrics_to_values.items():
tf.summary.scalar(metric_name, metric_value)
# Evaluate every 30 seconds
slim.evaluation.evaluation_loop('',
checkpoint_dir,
log_dir,
num_evals=num_batches,
eval_op=list(
metrics_to_updates.values()),
summary_op=tf.summary.merge_all(),
eval_interval_secs=60,
max_number_of_evaluations=100000000)
def __init__(self, is_train=False):
self.class_num = 5990
self.character_step = 8
self.character_height = 32
self.character_width = 32
#train
self.train_tfrecords_name = "./make_tfrecords/data.tfrecords"
self.summary_save_path = "./summary/"
self.summary_steps = 10000
self.save_steps = 10
self.save_path = "./save/"
#test:
self.model_path = "./save/sliding_conv.ckpt-10"
if is_train:
self.batch_size = 16
self.with_clip = True
self.network = network(batch_size=self.batch_size,
class_num=self.class_num,
character_height=self.character_height,
character_width=self.character_width,
character_step=self.character_step,
is_train=True)
else:
current_path = os.path.dirname(os.path.abspath(__file__))
self.char_dict = self.create_char_dict(
os.path.join(current_path, "char_std_5990.txt"))
self.batch_size = 1
self.graph = tf.Graph()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=1.0)
self.session = tf.Session(
config=tf.ConfigProto(gpu_options=gpu_options),
graph=self.graph)
with self.session.as_default():
with self.graph.as_default():
self.network = network(
batch_size=self.batch_size,
class_num=self.class_num,
character_height=self.character_height,
character_width=self.character_width,
character_step=self.character_step,
is_train=False)
self.decoded, self.log_prob = tf.nn.ctc_greedy_decoder(
self.network["outputs"],
self.network["seq_len"],
merge_repeated=True)
init = tf.global_variables_initializer()
self.session.run(init)
saver = tf.train.Saver(tf.global_variables(),
max_to_keep=100)
saver.restore(self.session, self.model_path)
def __init__(self, state_size, action_size, lr=5e-4):
self.action_size = action_size
self.local_network = network(state_size, action_size)
self.target_network = network(state_size, action_size)
self.optimizer = opt.Adam(self.local_network.parameters(), lr=lr)
self.memory = ReplayBuffer()
self.eps = 1.0
self.decay = 0.99
self.eps_min = 0.01
self.update_every = 4
self.t_step = 1
self.gamma = 0.99
self.TAU = 1e-3
def train(model_name):
#Get the lables from the .env file
labels = env('LABELS')
size = int(env('IMG_SIZE'))
#create the train data (format the training images)
train_data = create_train_data()
tf.reset_default_graph()
convnet = network(size, labels)
model = tflearn.DNN(convnet, tensorboard_dir='log')
#if the model already exists, load it so we are not training from scratch
if os.path.exists('{}.meta'.format(model_name)):
model.load(model_name)
print('model loaded!')
X = np.array([i[0] for i in train_data]).reshape(-1, size, size, 1)
Y = [i[1] for i in train_data]
model.fit(X, Y, n_epoch=50)
#save the model in the models folder
model.save('../models/' + model_name)
print("here")
print(type(model))
print("here")
return model
def main(train_dir, batch_size, num_batches, log_dir, checkpoint_dir=None):
if checkpoint_dir is None:
checkpoint_dir = log_dir
images, labels = build_input('cifar10', 100, 'test')
predictions, total_loss = network(images, labels)
tf.summary.scalar('loss', total_loss)
predictions = tf.to_int32(tf.argmax(predictions, 1))
tf.summary.scalar('accuracy', slim.metrics.accuracy(predictions, labels))
# These are streaming metrics which compute the "running" metric,
# e.g running accuracy
metrics_to_values, metrics_to_updates = slim.metrics.aggregate_metric_map({
'accuracy':
slim.metrics.streaming_accuracy(predictions, labels),
'streaming_mse':
slim.metrics.streaming_mean_squared_error(predictions, labels),
})
# Define the streaming summaries to write:
for metric_name, metric_value in metrics_to_values.items():
tf.summary.scalar(metric_name, metric_value)
# Evaluate every 30 seconds
slim.evaluation.evaluation_loop('',
checkpoint_dir,
log_dir,
num_evals=num_batches,
eval_op=list(metrics_to_updates.values()),
summary_op=tf.summary.merge_all(),
eval_interval_secs=20,
max_number_of_evaluations=100000000)
def main():
data = CIFAR10(paths)
data_loader = datautil.DataLoader(dataset=data,
batch_size=512,
num_workers=8,
shuffle=True)
epoch = 0
net = network().to(device)
criterion = nn.CrossEntropyLoss().to(device)
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
running_loss = 0
for ep in tqdm(range(epoch, 50)):
for i, (image, label) in tqdm(enumerate(data_loader)):
optimizer.zero_grad()
# print (repr(image))
output = net(image.to(device, dtype=torch.float32))
# print(label)
loss = criterion(output, label.to(device))
loss.backward()
optimizer.step()
# running_loss += loss.item()
if i % 2000 == 0:
print(loss)
pass
torch.save(net.state_dict(), './model/model1.pt')
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--epochs', default=100, type=int, help='epoch number')
parser.add_argument('--start_epoch', default=0, type=int, help='start epoch number')
parser.add_argument('-b', '--batch_size', default=4, type=int, help='mini-batch size')
parser.add_argument('--lr', '--learning_rate', default=1e-4, type=float, help='initial learning rate')
parser.add_argument('--weight-decay', default=0.0, type=float, help='weight decay')
parser.add_argument('-c', '--continue', dest='continue_path', type=str, required=False)
parser.add_argument('--exp_name', default=config.exp_name, type=str, required=False)
parser.add_argument('--valid', action='store_true')
parser.add_argument('--style_loss', action='store_true')
args = parser.parse_args()
print(args)
config.exp_name = args.exp_name
config.make_dir()
save_args(args, config.log_dir)
net = network()
vgg = vgg_for_style_transfer()
net = torch.nn.DataParallel(net).cuda()
vgg = torch.nn.DataParallel(vgg).cuda()
sess = Session(config, net=net)
train_loader = get_dataloaders(os.path.join(config.data_dir, 'train.json'),
batch_size=args.batch_size, shuffle=True)
valid_loader = get_dataloaders(os.path.join(config.data_dir, 'val.json'),
batch_size=args.batch_size, shuffle=True)
if args.continue_path and os.path.exists(args.continue_path):
sess.load_checkpoint(args.continue_path)
clock = sess.clock
tb_writer = sess.tb_writer
criterion = nn.L1Loss().cuda()
optimizer = optim.Adam(sess.net.parameters(), args.lr, weight_decay=args.weight_decay)
scheduler = ReduceLROnPlateau(optimizer, 'min', factor=0.5, patience=10, verbose=True)
for e in range(args.epochs):
train_model(train_loader, sess.net, vgg,
criterion, optimizer, clock.epoch, tb_writer)
valid_out = valid_model(valid_loader, sess.net, vgg,
criterion, optimizer, clock.epoch, tb_writer)
tb_writer.add_scalar('train/learning_rate', optimizer.param_groups[-1]['lr'], clock.epoch)
scheduler.step(valid_out['epoch_loss'])
if valid_out['epoch_loss'] < sess.best_val_loss:
sess.best_val_loss = valid_out['epoch_loss']
sess.save_checkpoint('best_model.pth.tar')
if clock.epoch % 10 == 0:
sess.save_checkpoint('epoch{}.pth.tar'.format(clock.epoch))
sess.save_checkpoint('latest.pth.tar')
clock.tock()
def main():
tf.reset_default_graph()
input_node = tf.placeholder(tf.float32,
shape=(24, 24,
3)) #这个是你送入网络的图片大小，如果你是其他的大小自行修改
input_node = tf.expand_dims(input_node, 0)
flow = network(input_node)
flow = tf.cast(flow, tf.uint8, 'out') #设置输出类型以及输出的接口名字，为了之后的调用pb的时候使用
saver = tf.train.Saver()
with tf.Session() as sess:
saver.restore(sess, model_path)
#保存图
tf.train.write_graph(sess.graph_def, 'output_model/pb_model',
'model.pb')
#把图和参数结构一起
freeze_graph.freeze_graph('output_model/pb_model/model.pb', '', False,
model_path, 'out', 'save/restore_all',
'save/Const:0',
'output_model/pb_model/frozen_model.pb',
False, "")
print("done")
def test(model_name):
labels = env('LABELS')
size = int(env('IMG_SIZE'))
#format the images that are uploaded
test_upload = process_test_data(env('TEST_UPLOAD'))
convnet = network(size, labels)
model = tflearn.DNN(convnet, tensorboard_dir='log')
#load the model for testing
model.load('models/' + model_name)
for num, data in enumerate(test_upload):
img_num = data[1]
img_data = data[0]
orig = img_data
data = img_data.reshape(size, size, 1)
model_out = model.predict([data])[0]
print(model_out)
print(np.argmax(model_out))
str_label = labels[np.argmax(model_out)]
#output as a text
print(str_label)
def __init__(self, sess, s_size, a_size, scope, queues, trainer):
self.queue = queues[0]
self.param_queue = queues[1]
self.replaymemory = ReplayMemory(100000)
self.sess = sess
self.learner_net = network(s_size, a_size, scope, 20)
self.q = self.learner_net.q
self.Q = self.learner_net.Q
self.actions_q = tf.placeholder(shape=[None, a_size, N],
dtype=tf.float32)
self.q_target = tf.placeholder(shape=[None, N], dtype=tf.float32)
self.ISWeights = tf.placeholder(shape=[None, N], dtype=tf.float32)
self.q_actiona = tf.multiply(self.q, self.actions_q)
self.q_action = tf.reduce_sum(self.q_actiona, axis=1)
self.u = tf.abs(self.q_target - self.q_action)
self.loss = tf.reduce_mean(
tf.reduce_sum(tf.square(self.u) * self.ISWeights, axis=1))
self.local_vars = self.learner_net.local_vars #tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope)
self.gradients = tf.gradients(self.loss, self.local_vars)
#grads,self.grad_norms = tf.clip_by_norm(self.gradients,40.0)
self.apply_grads = trainer.apply_gradients(
zip(self.gradients, self.local_vars))
self.sess.run(tf.global_variables_initializer())
def test_input(self):
batch_size = 1
# Mock input pipeline.
mock_imgs = np.ones([batch_size, 32, 32, 1], dtype=np.float32)
mock_edges = np.ones([batch_size, 32, 32, 1], dtype=np.float32)
mock_interp = np.ones([batch_size, 32, 32, 1], dtype=np.float32)
t_image = tf.placeholder(tf.float32,
shape=(None, None, None, 1),
name='t_input')
t_interpolated = tf.placeholder(tf.float32,
shape=(None, None, None, 1),
name='t_input')
t_edges = tf.placeholder(tf.float32,
shape=(None, None, None, 1),
name='t_input')
d_flg = tf.placeholder(tf.bool, name='is_train')
net_g = network(t_image, t_edges, is_train=d_flg)
sess = tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
result = sess.run(net_g,
feed_dict={
t_image: mock_imgs,
t_interpolated: mock_interp,
t_edges: mock_edges,
d_flg: False
})
print(result.shape)
def test(args, shared_model, inputs, labels, log):
start_time = time.time()
local_model = model.network(args,
class_num=args.class_num,
input_dim=inputs.shape[1])
local_model.load_state_dict(shared_model.state_dict())
if args.gpu:
local_model = local_model.cuda()
outputs = local_model(inputs)
predict_labels = torch.argmax(outputs, dim=1)
if args.gpu:
predict_labels = predict_labels.cpu()
predict_labels = predict_labels.numpy()
cnt = 0
if args.gpu:
labels = labels.cpu()
labels = labels.numpy()
for i in range(labels.shape[0]):
if predict_labels[i] == labels[i]:
cnt += 1
accuracy = cnt / labels.shape[0]
f1 = f1_score(list(labels), list(predict_labels), average='macro')
log.info('Test time ' + \
time.strftime("%Hh %Mm %Ss", time.gmtime(time.time() - start_time)) + \
',f1: %0.4f\t accuracy: %0.4f' % (f1, 100* accuracy))
return accuracy, f1
def build_program(main_prog, startup_prog, args):
image_shape = [int(m) for m in args.image_shape.split(",")]
with fluid.program_guard(main_prog, startup_prog):
with fluid.unique_name.guard():
image = fluid.data(
name="image", shape=[None] + image_shape, dtype="float32")
label = fluid.data(name="label", shape=[None, 1], dtype="int64")
genotype = eval("genotypes.%s" % args.arch)
drop_path_prob = ''
drop_path_mask = ''
do_drop_path = False
logits, logits_aux = network(
x=image,
is_train=False,
c_in=args.init_channels,
num_classes=args.class_num,
layers=args.layers,
auxiliary=False,
genotype=genotype,
do_drop_path=do_drop_path,
drop_prob=drop_path_prob,
drop_path_mask=drop_path_mask,
args=args,
name='model')
top1 = fluid.layers.accuracy(input=logits, label=label, k=1)
top5 = fluid.layers.accuracy(input=logits, label=label, k=5)
loss = fluid.layers.reduce_mean(
fluid.layers.softmax_with_cross_entropy(logits, label))
outs = [loss, top1, top5]
return outs
def load_check_point(train_id, path=None):
"""
"""
if path is None:
path = os.getcwd()
param_fn = os.path.join(path, str(train_id) + '.param.npz')
config_fn = os.path.join(path, str(train_id) + '.nnconfig.gz')
params = joblib.load(config_fn)
mdl, net = build(network(params), params)
layers = L.get_all_layers(L.ConcatLayer(net.values(), axis=1))
if os.path.exists(param_fn):
try:
print('Loadong pre-trained weight...')
with np.load(param_fn) as f:
param_values = [f['arr_%d' % i] for i in range(len(f.files))]
L.set_all_param_values(layers, param_values)
except Exception as e:
print(e)
print('Cannot load parameters!')
else:
print('Cannot find parameters!')
return net, mdl, params
def main():
args = get_arguments()
num_classes = dataset_settings[args.dataset]['num_classes']
input_size = dataset_settings[args.dataset]['input_size']
label = dataset_settings[args.dataset]['label']
model = network(num_classes=num_classes, pretrained=None).cuda()
model = nn.DataParallel(model)
state_dict = torch.load(args.restore_weight)
model.load_state_dict(state_dict)
model.eval()
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.406, 0.456, 0.485], std=[0.225, 0.224, 0.229])
])
dataset = SCHPDataset(root=args.input, input_size=input_size, transform=transform)
dataloader = DataLoader(dataset)
if not os.path.exists(args.output):
os.makedirs(args.output)
palette = get_palette(num_classes)
with torch.no_grad():
for idx, batch in enumerate(dataloader):
image, meta = batch
img_name = meta['name'][0]
c = meta['center'].numpy()[0]
s = meta['scale'].numpy()[0]
w = meta['width'].numpy()[0]
h = meta['height'].numpy()[0]
output = model(image.cuda())
upsample = torch.nn.Upsample(size=input_size, mode='bilinear', align_corners=True)
upsample_output = upsample(output)
upsample_output = upsample_output.squeeze()
upsample_output = upsample_output.permute(1, 2, 0) #CHW -> HWC
logits_result = transform_logits(upsample_output.data.cpu().numpy(), c, s, w, h, input_size=input_size)
parsing_result = np.argmax(logits_result, axis=2)
parsing_result_path = os.path.join(args.output, img_name[:-4]+'.png')
# plt.imshow(parsing_result)
# plt.show()
output_img = Image.fromarray(np.asarray(parsing_result, dtype=np.uint8))
output_img.putpalette(palette)
output_img.save(parsing_result_path)
if args.logits:
#logits_result_path = os.path.join(args.output, img_name[:-4] + '.npy')
np_result_path = os.path.join(args.output, img_name[:-4] + '.npy')
#np.save(logits_result_path, logits_result)
np.save(np_result_path, parsing_result)
return
def main(_):
config = tf.ConfigProto(device_count={'GPU': 0})
with tf.Session(config=config) as sess:
model = network(sess, args)
if args.phase == 'train':
model.train(args)
else:
model.test(args)
def export():
tf.logging.set_verbosity(tf.logging.INFO)
nn = network()
inp = tf.placeholder(tf.float32, [None], name=INPUT_TENSOR_NAME)
inp = tf.reshape(inp, [-1, CROPPED_IMAGE_SIZE, CROPPED_IMAGE_SIZE, 3])
td = dict(input=inp)
nn.export_savedmodel(
EXPORT_FOLDER,
tf.estimator.export.build_raw_serving_input_receiver_fn(td))
def act(state, checkpoint):
state = torch.from_numpy(state).float().unsqueeze(0)
model = network(37, 4)
model.load_state_dict(torch.load(checkpoint))
model.eval()
with torch.no_grad():
action = model(state)
action = np.argmax(action.numpy())
return action
def main(train_dir, batch_size, num_batches, log_dir):
images, labels = build_input('cifar10', 100, 'train')
predictions, total_loss = network(images, labels)
tf.summary.scalar('loss', total_loss)
optimizer = tf.train.GradientDescentOptimizer(0.1)
train_op = slim.learning.create_train_op(total_loss, optimizer, summarize_gradients=True)
slim.learning.train(train_op, log_dir, save_summaries_secs=20, save_interval_secs=20)
def __init__(self, sess, env, name, s_size, a_size, queues):
self.name = "worker_" + str(name)
self.number = name
self.sess = sess
self.Actor_net = network(s_size, a_size, self.name)
self.env = env
self.queue = queues[0]
self.param_queue = queues[1]
self.sess.run(tf.global_variables_initializer())
def act(state, checkpoint, num_agents):
actions = []
for i in range(num_agents):
state_tensor = torch.from_numpy(state[i]).float()
model = network(33, 4)
model.load_state_dict(torch.load(checkpoint))
model.eval()
with torch.no_grad():
action = model(state_tensor).data.numpy()
actions.append(action)
return actions
def save_weights_for_DeepSVDD(self, model, dataloader):
"""Initialize Deep SVDD weights using the encoder weights of the pretrained autoencoder."""
c = self.set_c(model, dataloader)
net = network(self.args.latent_dim).to(self.device)
state_dict = model.state_dict()
net.load_state_dict(state_dict, strict=False)
torch.save(
{
'center': c.cpu().data.numpy().tolist(),
'net_dict': net.state_dict()
}, 'weights/pretrained_parameters.pth')
def ex(net, learning_rate, split, epsilon, beta, dur,
n_epochs, targets, batch_sz, shuffle, data_path,
overlap_chunk=True, kernel_multiplier=1, train_id=None):
""""""
if train_id is None:
train_id = uuid.uuid4()
logger = tblog.Logger('runs/{}'.format(train_id))
params = {
'network': net,
'data_fn': os.path.join(data_path, 'train.h5'),
'scaler': SCALER_FN,
'split_fn': split,
'learning_rate': learning_rate,
'epsilon': epsilon,
'beta': beta,
'verbose': True,
'n_epochs': n_epochs,
'batch_sz': batch_sz,
'dur': dur, # frames
'overlap_chunk': True if overlap_chunk else False,
'kernel_multiplier': kernel_multiplier,
'report_every': 100,
'class_weight': False,
'prepare_submission': False,
'iter': 0
}
params['targets'] = targets
with h5py.File(params['data_fn']) as hf:
# load split info
split = joblib.load(params['split_fn'])
params.update(
{'split': {k: map(int, v) for k, v in split.iteritems()}})
# load class weight if needed
if params['class_weight']:
params.update(
{'class_weight': get_class_weight(hf['y'][:])}
)
mdl, net = build(network(params), params)
train(mdl, hf, params, shuffle, logger)
save_check_point(net, params, train_id, path='results/')
f1, ll = evaluate(mdl, hf, params)
if params['prepare_submission']:
# predict test dataset and prepare submission
test(mdl, hf, train_id,
os.path.join(data_path, 'test.h5'), params)
return train_id, f1, ll
def main(train_dir, batch_size, num_batches, log_dir):
images, labels = build_input('cifar10', 100, 'train')
images_cat_0, labels_cat_0 = build_input_cat_0('cifar10', 100, 'test')
images_cat_1, labels_cat_1 = build_input_cat_1('cifar10', 100, 'test')
predictions, total_loss, labels_cat2 = network(images, labels)
predictions_cat_0, loss_0, labels_cat_0_ = network(images_cat_0,
labels_cat_0)
predictions_cat_1, loss_1, labels_cat_1_ = network(images_cat_1,
labels_cat_1)
report()
tf.summary.scalar('loss', total_loss)
tf.summary.scalar('loss_0', loss_0)
tf.summary.scalar('loss_1', loss_1)
predictions = tf.argmax(predictions, axis=1)
predictions_cat_0 = tf.argmax(predictions_cat_0, axis=1)
predictions_cat_1 = tf.argmax(predictions_cat_1, axis=1)
tf.summary.scalar(
'accuracy', slim.metrics.accuracy(predictions,
tf.to_int64(labels_cat2)))
tf.summary.scalar(
'accuracy_cat_0',
slim.metrics.accuracy(predictions_cat_0, tf.to_int64(labels_cat_0_)))
tf.summary.scalar(
'accuracy_cat_1',
slim.metrics.accuracy(predictions_cat_1, tf.to_int64(labels_cat_1_)))
optimizer = tf.train.GradientDescentOptimizer(0.1)
train_op = slim.learning.create_train_op(total_loss,
optimizer,
summarize_gradients=True)
slim.learning.train(train_op,
log_dir,
save_summaries_secs=20,
save_interval_secs=20)
def build_program(main_prog, startup_prog, is_train, args):
image_shape = [int(m) for m in args.image_shape.split(",")]
with fluid.program_guard(main_prog, startup_prog):
with fluid.unique_name.guard():
image = fluid.data(name="image",
shape=[None] + image_shape,
dtype="float32")
label = fluid.data(name="label", shape=[None, 1], dtype="int64")
data_loader = fluid.io.DataLoader.from_generator(
feed_list=[image, label],
capacity=64,
use_double_buffer=True,
iterable=True)
genotype = eval("genotypes.%s" % args.arch)
logits, logits_aux = network(x=image,
is_train=is_train,
c_in=args.init_channels,
num_classes=args.class_num,
layers=args.layers,
auxiliary=args.auxiliary,
genotype=genotype,
name='model')
top1 = fluid.layers.accuracy(input=logits, label=label, k=1)
top5 = fluid.layers.accuracy(input=logits, label=label, k=5)
loss = fluid.layers.reduce_mean(
cross_entropy_label_smooth(logits, label, args.label_smooth))
if is_train:
if args.auxiliary:
loss_aux = fluid.layers.reduce_mean(
cross_entropy_label_smooth(logits_aux, label,
args.label_smooth))
loss = loss + args.auxiliary_weight * loss_aux
step_per_epoch = int(args.trainset_num / args.batch_size)
learning_rate = fluid.layers.exponential_decay(
args.learning_rate,
step_per_epoch,
args.decay_rate,
staircase=True)
clip = fluid.clip.GradientClipByGlobalNorm(
clip_norm=args.grad_clip)
optimizer = fluid.optimizer.MomentumOptimizer(
learning_rate,
args.momentum,
regularization=fluid.regularizer.L2DecayRegularizer(
args.weight_decay),
grad_clip=clip)
optimizer.minimize(loss)
outs = [loss, top1, top5, learning_rate]
else:
outs = [loss, top1, top5]
return outs, data_loader
def test(test_dir, checkpoint_dir='./checkpoint_res50/'):
# predict the result
test_images = os.listdir(test_dir)
features = tf.placeholder(
"float32",
shape=[None, CROP_SIZE, CROP_SIZE, IMAGE_CHANNEL],
name="features")
labels = tf.placeholder("float32", [None], name="labels")
is_training = tf.placeholder("bool")
one_hot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=80)
net = model.network(NUM_CLASS, model_name, use_batch_norm=use_bn)
logits = net.inference(inputs=features,
is_training=is_training,
reuse=False)
values, indices = tf.nn.top_k(logits, 3)
with tf.Session() as sess:
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
if ckpt and ckpt.model_checkpoint_path:
print('Restore the model from checkpoint %s' %
ckpt.model_checkpoint_path)
# Restores from checkpoint
saver.restore(sess, ckpt.model_checkpoint_path)
start_step = int(
ckpt.model_checkpoint_path.split('/')[-1].split('-')[-1])
else:
raise Exception('no checkpoint find')
result = []
for test_image in test_images:
temp_dict = {}
x = scene_input.img_resize(os.path.join(test_dir, test_image),
CROP_SIZE)
indices_eval = sess.run(indices,
feed_dict={
features: np.expand_dims(x, axis=0),
is_training: False
})
predictions = np.squeeze(indices_eval)
temp_dict['image_id'] = test_image
temp_dict['label_id'] = predictions.tolist()
result.append(temp_dict)
print('image %s is %d,%d,%d' %
(test_image, predictions[0], predictions[1], predictions[2]))
with open('submit.json', 'w') as f:
json.dump(result, f)
print('write result json, num is %d' % len(result))
def make_optimal_network(deep_load=False, load_black=True):
_, routes = load_airport_and_route(deep_load=deep_load)
netx = from_edgelist(routes)
N = number_of_nodes(netx)
net = network(N, graph=netx)
optimal = optimal_distribution(deep_load=deep_load)
R = B = [balls_per_node] * N
if load_black:
B = optimal
else:
R = optimal
net.set_initial_distribution(R, B)
def show_results(network, features, labels, examples=4):
figure, axis = plt.subplots(nrows=4, ncols=4, figsize=(15, 4 * examples))
dim1 = features.shape[0]
for row in range(examples):
img_idx = np.random.randint(dim1)
image_array = network(
mx.nd.array(features[img_idx:img_idx + 1],
ctx=context).astype('float32')).squeeze(0).asnumpy()
axis[row][0].imshow(
np.transpose(features[img_idx], (1, 2, 0))[:, :, 0])
axis[row][1].imshow(np.transpose(image_array, (1, 2, 0))[:, :, 0])
axis[row][2].imshow(image_array.argmax(0))
axis[row][3].imshow(np.transpose(labels[img_idx], (1, 2, 0))[:, :, 0])
plt.show()
def test_build_graph(self):
###========================== DEFINE MODEL ============================###
t_image = tf.placeholder(tf.float32,
shape=(None, 32, 32, 1),
name='t_input')
t_interpolated = tf.placeholder(tf.float32,
shape=(None, 32, 32, 1),
name='t_input')
t_edges = tf.placeholder(tf.float32,
shape=(None, 32, 32, 1),
name='t_input')
d_flg = tf.placeholder(tf.bool, name='is_train')
net_g = network(t_image, t_edges, t_interpolated, is_train=d_flg)
print('out', net_g.shape)