def main():
model = config.get('config', 'model')
cachedir = utils.get_cachedir(config)
with open(os.path.join(cachedir, 'names'), 'r') as f:
names = [line.strip() for line in f]
width, height = np.array(utils.get_downsampling(config)) * 13
anchors = pd.read_csv(os.path.expanduser(os.path.expandvars(config.get(model, 'anchors'))), sep='\t').values
func = getattr(inference, config.get(model, 'inference'))
with tf.Session() as sess:
image = tf.placeholder(tf.float32, [1, height, width, 3], name='image')
func(image, len(names), len(anchors))
tf.contrib.framework.get_or_create_global_step()
tf.global_variables_initializer().run()
prog = re.compile(r'[_\w\d]+\/conv(\d*)\/(weights|biases|(BatchNorm\/(gamma|beta|moving_mean|moving_variance)))$')
variables = [(prog.match(v.op.name).group(1), v) for v in tf.global_variables() if prog.match(v.op.name)]
variables = sorted([[int(k) if k else -1, [v for _, v in g]] for k, g in itertools.groupby(variables, operator.itemgetter(0))], key=operator.itemgetter(0))
assert variables[0][0] == -1
variables[0][0] = len(variables) - 1
variables.insert(len(variables), variables.pop(0))
with tf.name_scope('assign'):
with open(os.path.expanduser(os.path.expandvars(args.file)), 'rb') as f:
major, minor, revision, seen = struct.unpack('4i', f.read(16))
tf.logging.info('major=%d, minor=%d, revision=%d, seen=%d' % (major, minor, revision, seen))
for i, layer in variables:
tf.logging.info('processing layer %d' % i)
total = 0
for suffix in ['biases', 'beta', 'gamma', 'moving_mean', 'moving_variance', 'weights']:
try:
v = next(filter(lambda v: v.op.name.endswith(suffix), layer))
except StopIteration:
continue
shape = v.get_shape().as_list()
cnt = np.multiply.reduce(shape)
total += cnt
tf.logging.info('%s: %s=%d' % (v.op.name, str(shape), cnt))
p = struct.unpack('%df' % cnt, f.read(4 * cnt))
if suffix == 'weights':
ksize1, ksize2, channels_in, channels_out = shape
p = np.reshape(p, [channels_out, channels_in, ksize1, ksize2]) # Darknet format
p = np.transpose(p, [2, 3, 1, 0]) # TensorFlow format (ksize1, ksize2, channels_in, channels_out)
sess.run(v.assign(p))
tf.logging.info('%d parameters assigned' % total)
remaining = os.fstat(f.fileno()).st_size - f.tell()
transpose(sess, layer, len(anchors))
saver = tf.train.Saver()
logdir = utils.get_logdir(config)
if args.delete:
tf.logging.warn('delete logging directory: ' + logdir)
shutil.rmtree(logdir, ignore_errors=True)
os.makedirs(logdir, exist_ok=True)
model_path = os.path.join(logdir, 'model.ckpt')
tf.logging.info('save model into ' + model_path)
saver.save(sess, model_path)
if args.summary:
path = os.path.join(logdir, args.logname)
summary_writer = tf.summary.FileWriter(path)
summary_writer.add_graph(sess.graph)
tf.logging.info('tensorboard --logdir ' + logdir)
if remaining > 0:
tf.logging.warn('%d bytes remaining' % remaining)
def get_instance():
env = utils.get_env()
for k, v in env.__dict__.items():
print(f'{k}: {v}')
n_states = env.n_states
n_actions = env.n_actions
agent = DQN(n_states, n_actions)
logdir = utils.get_logdir()
return env, agent, logdir
def get_instance(diffs):
default_conf = utils.get_config('default_config.yaml')
conf = deepcopy(default_conf)
recursive_merge(conf, diffs)
env = utils.get_env(**conf['env'])
# for k in diffs.keys():
# print(k, ':', env.__getattribute__(k))
n_states = env.n_states
n_actions = env.n_actions
agent = DQN(n_states, n_actions, **conf['agent'])
# different to pa_main
logdir = utils.get_logdir(conf, default_conf)
return env, agent, logdir
def train():
for time in range(5):
logx.initialize(get_logdir("../runs"),
tensorboard=True,
coolname=False)
model.load_state_dict(
torch.load("..\\runs\exp10\last_checkpoint_ep0.pth")
['state_dict']) # warmup
dataset_train = TrainDataset(
'../' + cfg.root_folder +
'/five_fold/train_kfold_{}.csv'.format(time),
'../' + cfg.root_folder + '/train/', train_transform)
train_loader = DataLoader(dataset_train,
batch_size=cfg.bs,
shuffle=True)
test_data = TrainDataset(
'../' + cfg.root_folder +
'/five_fold/test_kfold_{}.csv'.format(time),
'../' + cfg.root_folder + '/train/',
)
test_load = DataLoader(test_data, batch_size=cfg.bs, shuffle=False)
# train
for epoch in range(cfg.epoch):
loss_epoch = 0
total = 0
correct = 0
for i, (x, y) in enumerate(train_loader, 1):
x, y = x.to(device), y.to(device)
y_hat = model(x)
# 计算正确率
total += x.size(0)
_, predict = torch.max(y_hat.data, dim=1)
correct += (predict == y).sum().item()
# 损失
loss = criterion(y_hat, y)
loss_epoch += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
# 过程可视化
if i % 30 == 0:
print(
'epoch:%d, enumerate:%d, loss_avg:%f, now_acc:%f' %
(epoch, i, loss_epoch / i, correct / total))
# epoch matric 可视化
train_loss = loss_epoch / i
train_acc = (correct / total) * 100
logx.metric('train', {'loss': train_loss, 'acc': train_acc}, epoch)
# valid
# 开发集正确率
correct = 0
total = 0
val_loss = 0
with torch.no_grad():
for i, (img, label) in enumerate(test_load, 1):
img, label = img.to(device), label.to(device)
output = model(img)
loss = criterion(output, label)
val_loss += loss.cpu().item()
_, predicted = torch.max(output.data, dim=1) # 最大值，位置
total += img.size(0)
correct += (predicted == label).sum().item()
val_acc = (100 * correct / total)
val_loss /= i
logx.metric('val', {'loss': val_loss, 'acc': val_acc}, epoch)
# epoch lossand other metric
print(
'epoch over; train_loss:%f, val_loss:%f, train_acc=%f, val_acc:%f'
% (train_loss, val_loss, train_acc, val_acc))
logx.save_model({
'state_dict': model.state_dict(),
'epoch': epoch
},
val_acc,
higher_better=True,
epoch=epoch,
delete_old=True)
scheduler.step()
def main():
model = config.get('config', 'model')
cachedir = utils.get_cachedir(config)
with open(os.path.join(cachedir, 'names'), 'r') as f:
names = [line.strip() for line in f]
width, height = np.array(utils.get_downsampling(config)) * 13
anchors = pd.read_csv(os.path.expanduser(
os.path.expandvars(config.get(model, 'anchors'))),
sep='\t').values
func = getattr(inference, config.get(model, 'inference'))
with tf.Session() as sess:
image = tf.placeholder(tf.float32, [1, height, width, 3], name='image')
func(image, len(names), len(anchors))
tf.contrib.framework.get_or_create_global_step()
tf.global_variables_initializer().run()
prog = re.compile(
r'[_\w\d]+\/conv(\d*)\/(weights|biases|(BatchNorm\/(gamma|beta|moving_mean|moving_variance)))$'
)
variables = [(prog.match(v.op.name).group(1), v)
for v in tf.global_variables() if prog.match(v.op.name)]
variables = sorted(
[[int(k) if k else -1, [v for _, v in g]]
for k, g in itertools.groupby(variables, operator.itemgetter(0))],
key=operator.itemgetter(0))
assert variables[0][0] == -1
variables[0][0] = len(variables) - 1
variables.insert(len(variables), variables.pop(0))
with tf.name_scope('assign'):
with open(os.path.expanduser(os.path.expandvars(args.file)),
'rb') as f:
major, minor, revision, seen = struct.unpack('4i', f.read(16))
tf.logging.info('major=%d, minor=%d, revision=%d, seen=%d' %
(major, minor, revision, seen))
for i, layer in variables:
tf.logging.info('processing layer %d' % i)
total = 0
for suffix in [
'biases', 'beta', 'gamma', 'moving_mean',
'moving_variance', 'weights'
]:
try:
v = next(
filter(lambda v: v.op.name.endswith(suffix),
layer))
except StopIteration:
continue
shape = v.get_shape().as_list()
cnt = np.multiply.reduce(shape)
total += cnt
tf.logging.info('%s: %s=%d' %
(v.op.name, str(shape), cnt))
p = struct.unpack('%df' % cnt, f.read(4 * cnt))
if suffix == 'weights':
ksize1, ksize2, channels_in, channels_out = shape
p = np.reshape(
p, [channels_out, channels_in, ksize1, ksize2
]) # Darknet format
p = np.transpose(
p, [2, 3, 1, 0]
) # TensorFlow format (ksize1, ksize2, channels_in, channels_out)
sess.run(v.assign(p))
tf.logging.info('%d parameters assigned' % total)
remaining = os.fstat(f.fileno()).st_size - f.tell()
transpose(sess, layer, len(anchors))
saver = tf.train.Saver()
logdir = utils.get_logdir(config)
if args.delete:
tf.logging.warn('delete logging directory: ' + logdir)
shutil.rmtree(logdir, ignore_errors=True)
os.makedirs(logdir, exist_ok=True)
model_path = os.path.join(logdir, 'model.ckpt')
tf.logging.info('save model into ' + model_path)
saver.save(sess, model_path)
if args.summary:
path = os.path.join(logdir, args.logname)
summary_writer = tf.summary.FileWriter(path)
summary_writer.add_graph(sess.graph)
tf.logging.info('tensorboard --logdir ' + logdir)
if remaining > 0:
tf.logging.warn('%d bytes remaining' % remaining)
def run(args):
start_epoch = 1
best_loss = 1e+9
# logs
args.logdir = get_logdir(args)
logger = get_logger(os.path.join(args.logdir, 'main.log'))
logger.info(args)
writer = SummaryWriter(args.logdir)
# data
train_set = MovingMNIST(root='./data', train=True, download=True)
valid_set = MovingMNIST(root='./data',
train=False,
download=True,
split=args.test_size)
train_loader = DataLoader(train_set,
batch_size=args.batch_size,
num_workers=args.n_workers,
shuffle=True)
valid_loader = DataLoader(valid_set,
batch_size=args.batch_size,
num_workers=args.n_workers,
shuffle=False)
# network
model = models.__dict__[args.model](args=args)
model = nn.DataParallel(model)
args.device = torch.device(
'cuda:0' if torch.cuda.is_available() else 'cpu')
model = model.to(args.device)
# training
criterion = get_loss_fn(args)
optimizer = get_optimizer(model, args)
scheduler = get_scheduler(optimizer, args)
if args.resume:
if os.path.isfile(args.resume):
checkpoint = torch.load(args.resume)
start_epoch = checkpoint['epoch'] + 1
best_loss = checkpoint['best/{}'.format(args.loss)]
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
logger.info('Loaded checkpoint {} (epoch {})'.format(
args.resume, start_epoch - 1))
else:
raise IOError('No such file {}'.format(args.resume))
for epoch_i in range(start_epoch, args.epochs + 1):
training = train(train_loader,
model,
criterion,
optimizer,
logger=logger,
args=args)
validation = validate(valid_loader,
model,
criterion,
logger=logger,
args=args)
writer.add_scalar('Train/{}'.format(args.loss), training[args.loss],
epoch_i)
writer.add_scalar('Valid/{}'.format(args.loss), validation[args.loss],
epoch_i)
writer.add_image('Train/Predict', _get_images(training['output'],
args), epoch_i)
writer.add_image('Train/Target', _get_images(training['target'], args),
epoch_i)
writer.add_image('Valid/Predict',
_get_images(validation['output'], args), epoch_i)
writer.add_image('Valid/Target', _get_images(validation['target'],
args), epoch_i)
message = '[{}] Epoch {} Train/{} {:.4f} Valid/{} {:.4f} '
message = message.format(
args.expid,
epoch_i,
args.loss,
training[args.loss],
args.loss,
validation[args.loss],
)
is_best = validation[args.loss] < best_loss
if is_best:
best_loss = validation[args.loss]
message += '(Best)'
save_checkpoint(
{
'epoch': epoch_i,
'state_dict': model.state_dict(),
'valid/{}'.format(args.loss): validation[args.loss],
'best/{}'.format(args.loss): best_loss,
'optimizer': optimizer.state_dict(),
}, is_best, args.logdir)
if scheduler is not None:
scheduler.step(epoch=epoch_i)
logger.debug('Scheduler stepped.')
for param_group in optimizer.param_groups:
logger.debug(param_group['lr'])
logger.info(message)