def __init__(self, config, state_net=None, out_net=None):
super(GNN, self).__init__()
self.config = config
# hyperparameters and general properties
self.convergence_threshold = config.convergence_threshold
self.max_iterations = config.max_iterations
self.n_nodes = config.n_nodes
self.state_dim = config.state_dim
self.label_dim = config.label_dim
self.output_dim = config.output_dim
self.state_transition_hidden_dims = config.state_transition_hidden_dims
self.output_function_hidden_dims = config.output_function_hidden_dims
# node state initialization
self.node_state = torch.zeros(*[self.n_nodes, self.state_dim]).to(self.config.device) # (n,d_n)
self.converged_states = torch.zeros(*[self.n_nodes, self.state_dim]).to(self.config.device)
# state and output transition functions
if state_net is None:
self.state_transition_function = StateTransition(self.state_dim, self.label_dim,
mlp_hidden_dim=self.state_transition_hidden_dims,
activation_function=config.activation)
else:
self.state_transition_function = state_net
if out_net is None:
self.output_function = MLP(self.state_dim, self.output_function_hidden_dims, self.output_dim)
else:
self.output_function = out_net
self.graph_based = self.config.graph_based
def run(args):
train, test = util.get_dataset(args.dataset)
names = ['all-one (standard)', 'linear']
colors = [vz.colors.all_one_lg, vz.colors.linear_lg]
models = [
MLP.MLP(10, cg.uniform, n_units=100),
MLP.MLP(10, cg.linear, n_units=100),
]
comp_ratios = np.linspace(0.1, 1.0, 20)
acc_dict = {}
ratios_dict = {}
for name, model in zip(names, models):
util.load_or_train_model(model, train, test, args)
acc_dict[name] = util.sweep_idp(model, test, comp_ratios, args)
ratios_dict[name] = [100. * cr for cr in comp_ratios]
filename = "MLP_{}".format(args.dataset)
vz.plot(ratios_dict,
acc_dict,
names,
filename,
colors=colors,
folder=args.figure_path,
ext=args.ext,
title='MLP (MNIST)',
xlabel='IDP (%)',
ylabel='Classification Accuracy (%)',
ylim=(85, 100))
def run(args):
train, test = util.get_dataset(args.dataset)
colors = [vz.colors.all_one_lg, vz.colors.linear_lg]
names = ['all-one', 'harmonic', 'linear', 'half_one']
colors = [
vz.colors.all_one_lg, vz.colors.harmonic_lg, vz.colors.linear_lg,
vz.colors.half_one_lg
]
models = [
MLP.MLP(10, cg.uniform),
MLP.MLP(10, cg.harmonic),
MLP.MLP(10, cg.linear),
MLP.MLP(10, cg.uniform_exp),
]
comp_ratios = np.linspace(0.1, 1.0, 20)
acc_dict = {}
ratios_dict = {}
for name, model in zip(names, models):
util.load_or_train_model(model, train, test, args)
acc_dict[name] = util.sweep_idp(model, test, comp_ratios, args)
ratios_dict[name] = [100. * cr for cr in comp_ratios]
filename = "MLP_coef_comparison_{}".format(args.dataset)
vz.plot(ratios_dict,
acc_dict,
names,
filename,
colors=colors,
folder=args.figure_path,
ext=args.ext,
xlabel='IDP (%)',
ylabel='Classification Accuracy (%)',
title='MLP (MNIST)',
legend_loc='lower right',
ylim=(85, 100))
def main():
for seed in range(42,45):
torch.manual_seed(seed)
model = MLP().cuda()
optimizer = optim.SGD(model.parameters(), lr=1e-2, momentum=0)
train_ds = datasets.MNIST('../data', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
]))
train_ds.train_labels = torch.load('./random_labels_mnist.pth').long()
train_loader = torch.utils.data.DataLoader(train_ds, batch_size=64, shuffle=True)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('../data', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=1000, shuffle=True)
# 10 epoches
for epoch in range(1, 150 + 1):
train(model, train_loader, optimizer, epoch)
test(model, test_loader)
torch.save(model.state_dict(), './model_weights/mlp_random_weights_{}.pth'.format(seed))
model.load_state_dict(torch.load('./model_weights/mlp_random_weights_{}.pth'.format(seed)))
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--gpu', '-g', type=int, default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--image', '-i', type=str, default="",
help='pass to input image')
parser.add_argument('--model', '-m', default='my_mnist.model',
help='path to the training model')
parser.add_argument('--unit', '-u', type=int, default=1000,
help='Number of units')
args = parser.parse_args()
model = MLP(args.unit,10)
if args.gpu >= 0:
model.to_gpu(chainer.cuda.get_device_from_id(args.gpu).use())
serializers.load_npz(args.model, model)
try:
img = Image.open(args.image).convert("L").resize((28,28))
except :
print("invalid input")
return
img_array = model.xp.asarray(img,dtype=model.xp.float32).reshape(1,784)
with chainer.using_config('train', False), chainer.no_backprop_mode():
result = model.predict(img_array)
print("predict:", model.xp.argmax(result.data))
class AgentVPG:
"""
Agent
functions:
1) choose_action
input - state of the current environment
output - an action
2) update
input - experience obtained by interacting with the environment
output - losses
"""
def __init__(self, action_space, obs_dim, gamma=1):
self.logits_net = MLP(input_dim=obs_dim, output_dim=len(action_space))
self.act_space = list(action_space)
self.optim = Adam(self.logits_net.parameters(), lr=5e-3)
# make action selection function (outputs int actions, sampled from policy)
def choose_action(self, obs):
with torch.no_grad():
return self._get_policy(obs).sample().item()
def update(self, batch):
obs = torch.as_tensor(batch['obs'], dtype=torch.float32)
act = torch.as_tensor(batch['acts'], dtype=torch.int32)
weights = torch.as_tensor(batch['weights'], dtype=torch.float32)
batch_loss = self._compute_loss(obs, act, weights)
self.optim.zero_grad()
batch_loss.backward()
self.optim.step()
return batch_loss.item()
# make loss function whose gradient, for the right data, is policy gradient
def _compute_loss(self, obs, act, weights):
logp = self._get_policy(obs).log_prob(act)
# print(logp[:10], act[:10])
return -(logp * weights).mean() + (0.2 * logp).mean()
# make function to compute action distribution
def _get_policy(self, obs):
logits = self.logits_net(obs)
return Categorical(logits=logits)
def run(args):
train, test = util.get_dataset(args.dataset)
# names = ['all-ones,exp', 'all-ones,all', 'linear,exp', 'linear,all']
names = ['linear']
colors = [vz.colors.linear_sm, vz.colors.linear_lg]
models = [
MLP.MLP(10, cg.linear, [(0, 3), (3, 10)], n_units=100),
]
comp_ratios = np.linspace(0.1, 1, 20)
acc_dict = {}
ratios_dict = {}
key_names = []
for name, model in zip(names, models):
util.train_model_profiles(model, train, test, args)
for profile in range(len(model.profiles)):
key = name + '-' + str(profile + 1)
key_names.append(key)
acc_dict[key] = util.sweep_idp(model,
test,
comp_ratios,
args,
profile=profile)
ratios_dict[key] = [100. * cr for cr in comp_ratios]
filename = "MLP_{}_multi".format(args.dataset)
vz.plot(ratios_dict,
acc_dict,
key_names,
filename,
colors=colors,
folder=args.figure_path,
ext=args.ext,
xlabel='IDP (%)',
ylabel='Classification Accuracy (%)',
title='MLP (MNIST)',
ylim=(85, 100))
def __init__(
self,
config,
state_net=None,
out_net=None,
nodewise_lagrangian_flag=True,
dimwise_lagrangian_flag=False,
):
super(LPGNN, self).__init__()
self.config = config
# hyperparameters and general properties
self.n_nodes = config.n_nodes
self.state_dim = config.state_dim
self.label_dim = config.label_dim
self.output_dim = config.output_dim
self.state_transition_hidden_dims = config.state_transition_hidden_dims
self.output_function_hidden_dims = config.output_function_hidden_dims
# TODO add dims of layered LPGNN
self.λ_n = self.n_nodes if nodewise_lagrangian_flag else 1
self.λ_d = self.state_dim if dimwise_lagrangian_flag else 1
self.λ_list = nn.ParameterList()
self.state_variable_list = nn.ParameterList()
self.diffusion_constraint_list = []
self.node_state_list = []
self.state_transition_function_list = nn.ModuleList(
) # to be visible for parameters and gpu
# state constraints
self.state_constraint_function = utils.get_G_function(
descr=self.config.state_constraint_function, eps=self.config.eps)
for l in range(self.config.layers): # loop over layers
# defining lagrangian multipliers
self.λ_list.append(
nn.Parameter(torch.zeros(*[self.λ_n, self.λ_d]),
requires_grad=True)) # (n,1) by default ,
self.state_variable_list.append(
nn.Parameter(torch.zeros(*[self.n_nodes, self.state_dim[l]]),
requires_grad=True)) # (n,d_state)
# adding to lists
# node state initialization
# self.node_state_list.append(
# torch.zeros(*[self.n_nodes, self.state_dim[l]]).to(self.config.device)) # (n,d_n)
# state and output transition functions
if state_net is None:
# torch.manual_seed(self.config.seed)
if l == 0:
input_dim = self.state_dim[
0] + 2 * self.label_dim # arc state computation f(l_v, l_n, x_n)
else:
## f(x_v_l, x_v_l-1, x_n_l-1)
input_dim = self.state_dim[l] + 2 * self.state_dim[
l - 1] # + 2 * self.label_dim ##
output_dim = self.state_dim[l]
self.state_transition_function_list.append(
StateTransition(
input_dim=input_dim,
output_dim=output_dim,
mlp_hidden_dim=self.state_transition_hidden_dims,
activation_function=config.activation))
if state_net is not None: # only once, give as input the list TODO
self.state_transition_function_list = state_net
if out_net is None:
self.output_function = MLP(self.state_dim[-1],
self.output_function_hidden_dims,
self.output_dim)
else:
self.output_function = out_net
self.graph_based = self.config.graph_based
print ('what dataset ?', args.dataset)
sys.exit(1)
key = folder
key+= args.net \
+ '_hdim' + str(args.hdim) \
+ '_Batchsize' + str(args.Batchsize) \
+ '_lr' + str(args.lr) \
+ '_Nsteps' + str(args.Nsteps) \
+ '_epsilon' + str(args.epsilon)
cmd = ['mkdir', '-p', key]
subprocess.check_call(cmd)
if args.net == 'MLP':
net = MLP(dim=dim, hidden_size = args.hdim, use_z2=False)
elif args.net == 'CNN':
net = CNN(L=length, channel=channel, hidden_size = args.hdim, use_z2=False)
elif args.net == 'Simple_MLP':
net = Simple_MLP(dim=dim, hidden_size = args.hdim, use_z2=False)
else:
print ('what network ?', args.net)
sys.exit(1)
model = MongeAmpereFlow(net, args.epsilon, args.Nsteps, device=device, name = key)
model.to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=args.decay)
if args.checkpoint is not None:
try:
load_checkpoint(args.checkpoint, model, optimizer)
class GNN(nn.Module):
def __init__(self, config, state_net=None, out_net=None):
super(GNN, self).__init__()
self.config = config
# hyperparameters and general properties
self.convergence_threshold = config.convergence_threshold
self.max_iterations = config.max_iterations
self.n_nodes = config.n_nodes
self.state_dim = config.state_dim
self.label_dim = config.label_dim
self.output_dim = config.output_dim
self.state_transition_hidden_dims = config.state_transition_hidden_dims
self.output_function_hidden_dims = config.output_function_hidden_dims
# node state initialization
self.node_state = torch.zeros(*[self.n_nodes, self.state_dim]).to(
self.config.device) # (n,d_n)
self.converged_states = torch.zeros(
*[self.n_nodes, self.state_dim]).to(self.config.device)
# state and output transition functions
if state_net is None:
self.state_transition_function = StateTransition(
self.state_dim,
self.label_dim,
mlp_hidden_dim=self.state_transition_hidden_dims,
activation_function=config.activation)
else:
self.state_transition_function = state_net
if out_net is None:
self.output_function = MLP(self.state_dim,
self.output_function_hidden_dims,
self.output_dim)
else:
self.output_function = out_net
self.graph_based = self.config.graph_based
def reset_parameters(self):
self.state_transition_function.mlp.init()
self.output_function.init()
def forward(self,
edges,
agg_matrix,
node_labels,
node_states=None,
graph_agg=None):
n_iterations = 0
# convergence loop
# state initialization
node_states = self.node_state if node_states is None else node_states
while n_iterations < self.max_iterations:
new_state = self.state_transition_function(node_states,
node_labels, edges,
agg_matrix)
n_iterations += 1
# convergence condition
with torch.no_grad():
distance = torch.norm(
input=new_state - node_states,
dim=1) # checked, they are the same (in cuda, some bug)
check_min = distance < self.convergence_threshold
node_states = new_state
if check_min.all():
break
states = node_states
self.converged_states = states
if self.graph_based:
states = torch.matmul(graph_agg, node_states)
output = self.output_function(states)
return output, n_iterations
class LPGNN(nn.Module):
def __init__(
self,
config,
state_net=None,
out_net=None,
nodewise_lagrangian_flag=True,
dimwise_lagrangian_flag=False,
):
super(LPGNN, self).__init__()
self.config = config
# hyperparameters and general properties
self.n_nodes = config.n_nodes
self.state_dim = config.state_dim
self.label_dim = config.label_dim
self.output_dim = config.output_dim
self.state_transition_hidden_dims = config.state_transition_hidden_dims
self.output_function_hidden_dims = config.output_function_hidden_dims
# TODO add dims of layered LPGNN
self.λ_n = self.n_nodes if nodewise_lagrangian_flag else 1
self.λ_d = self.state_dim if dimwise_lagrangian_flag else 1
self.λ_list = nn.ParameterList()
self.state_variable_list = nn.ParameterList()
self.diffusion_constraint_list = []
self.node_state_list = []
self.state_transition_function_list = nn.ModuleList(
) # to be visible for parameters and gpu
# state constraints
self.state_constraint_function = utils.get_G_function(
descr=self.config.state_constraint_function, eps=self.config.eps)
for l in range(self.config.layers): # loop over layers
# defining lagrangian multipliers
self.λ_list.append(
nn.Parameter(torch.zeros(*[self.λ_n, self.λ_d]),
requires_grad=True)) # (n,1) by default ,
self.state_variable_list.append(
nn.Parameter(torch.zeros(*[self.n_nodes, self.state_dim[l]]),
requires_grad=True)) # (n,d_state)
# adding to lists
# node state initialization
# self.node_state_list.append(
# torch.zeros(*[self.n_nodes, self.state_dim[l]]).to(self.config.device)) # (n,d_n)
# state and output transition functions
if state_net is None:
# torch.manual_seed(self.config.seed)
if l == 0:
input_dim = self.state_dim[
0] + 2 * self.label_dim # arc state computation f(l_v, l_n, x_n)
else:
## f(x_v_l, x_v_l-1, x_n_l-1)
input_dim = self.state_dim[l] + 2 * self.state_dim[
l - 1] # + 2 * self.label_dim ##
output_dim = self.state_dim[l]
self.state_transition_function_list.append(
StateTransition(
input_dim=input_dim,
output_dim=output_dim,
mlp_hidden_dim=self.state_transition_hidden_dims,
activation_function=config.activation))
if state_net is not None: # only once, give as input the list TODO
self.state_transition_function_list = state_net
if out_net is None:
self.output_function = MLP(self.state_dim[-1],
self.output_function_hidden_dims,
self.output_dim)
else:
self.output_function = out_net
self.graph_based = self.config.graph_based
def reset_parameters(self):
with torch.no_grad():
for l in range(self.config.layers):
self.state_transition_function_list[l].mlp.init()
nn.init.constant_(self.state_variable_list[l], 0)
nn.init.constant_(self.λ_list[l], 0)
self.output_function.init()
def lagrangian_composition(self, new_state_list):
# loss definition TODO add for , add tensorboard
convergence_loss_list = []
for l in range(self.config.layers):
constraint = self.state_constraint_function(
self.state_variable_list[l] - new_state_list[l])
convergence_loss_list.append(
torch.mean(torch.mean(self.λ_list[l] * constraint, -1)))
# self.diffusion_constraint_list.append(torch.mean(torch.mean(self.λ * constraint, -1))) TODO why not working
# convergence_loss = torch.mean(torch.mean(self.λ * constraint, -1))
return torch.sum(torch.stack(convergence_loss_list), dim=0)
# if use_energy_constraint:
# total_loss += lpgnn.energy_loss(lpgnn.state_variable, new_state)
def forward(self,
edges,
agg_matrix,
node_labels,
node_states=None,
graph_agg=None):
# convergence loop
# state initialization
node_states = self.state_variable_list[
0] if node_states is None else node_states #
new_state_list = []
for l in range(self.config.layers):
if l == 0:
new_layer_state = self.state_transition_function_list[l](
node_states, node_labels, edges, agg_matrix, l)
new_state_list.append(new_layer_state)
else:
new_layer_state = self.state_transition_function_list[l](
self.state_variable_list[l], new_layer_state, edges,
agg_matrix, l)
new_state_list.append(new_layer_state)
new_state = new_layer_state
output = self.output_function(new_state)
return new_state_list, output
class GNN(nn.Module):
def __init__(self, config, state_net=None, out_net=None):
super(GNN, self).__init__()
self.config = config
# hyperparameters and general properties
self.convergence_threshold = config.convergence_threshold
self.max_iterations = config.max_iterations
self.n_nodes = config.n_nodes
self.state_dim = config.state_dim
self.label_dim = config.label_dim
self.output_dim = config.output_dim
self.state_transition_hidden_dims = config.state_transition_hidden_dims
self.output_function_hidden_dims = config.output_function_hidden_dims
# node state initialization
self.node_state = torch.zeros(*[self.n_nodes, self.state_dim]).to(
self.config.device) # (n,d_n)
self.converged_states = torch.zeros(
*[self.n_nodes, self.state_dim]).to(self.config.device)
# state and output transition functions
if state_net is None:
self.state_transition_function = StateTransition(
self.state_dim,
self.label_dim,
mlp_hidden_dim=self.state_transition_hidden_dims,
activation_function=config.activation)
# self.state_transition_function = GINTransition(self.state_dim, self.label_dim,
# mlp_hidden_dim=self.state_transition_hidden_dims,
# activation_function=config.activation)
# self.state_transition_function = GINPreTransition(self.state_dim, self.label_dim,
# mlp_hidden_dim=self.state_transition_hidden_dims,
# activation_function=config.activation)
else:
self.state_transition_function = state_net
if out_net is None:
self.output_function = MLP(self.state_dim,
self.output_function_hidden_dims,
self.output_dim)
else:
self.output_function = out_net
self.graph_based = self.config.graph_based
def reset_parameters(self):
self.state_transition_function.mlp.init()
self.output_function.init()
def forward(self,
edges,
agg_matrix,
node_labels,
node_states=None,
graph_agg=None):
n_iterations = 0
# convergence loop
# state initialization
node_states = self.node_state if node_states is None else node_states
# while n_iterations < self.max_iterations:
# with torch.no_grad(): # without memory consumption
# new_state = self.state_transition_function(node_states, node_labels, edges, agg_matrix)
# n_iterations += 1
# # convergence condition
#
# # if torch.dist(node_states, new_state) < self.convergence_threshold: # maybe uses broadcst?
# # break
# # with torch.no_grad():
# # distance = torch.sqrt(torch.sum((new_state - node_states) ** 2, 1) + 1e-20)
# distance = torch.norm(input=new_state - node_states,
# dim=1) # checked, they are the same (in cuda, some bug)
# #
# # diff =torch.norm(input=new_state - node_states, dim=1) - torch.sqrt(torch.sum((new_state - node_states) ** 2, 1) )
#
# check_min = distance < self.convergence_threshold
# node_states = new_state
#
# if check_min.all():
# break
# node_states = self.state_transition_function(node_states, node_labels, edges, agg_matrix) # one more to propagate gradient only on last
while n_iterations < self.max_iterations:
new_state = self.state_transition_function(node_states,
node_labels, edges,
agg_matrix)
n_iterations += 1
# convergence condition
with torch.no_grad():
# distance = torch.sqrt(torch.sum((new_state - node_states) ** 2, 1) + 1e-20)
distance = torch.norm(
input=new_state - node_states,
dim=1) # checked, they are the same (in cuda, some bug)
check_min = distance < self.convergence_threshold
node_states = new_state
if check_min.all():
break
states = node_states
self.converged_states = states
if self.graph_based:
states = torch.matmul(graph_agg, node_states)
output = self.output_function(states)
return output, n_iterations
iny = data[1]
data = make_blobs(n_features=2,centers=2,n_samples=10)
inX = data[0]
iny = data[1]
for _ in xrange(50):
X, tX, y, ty = train_test_split(inX,iny,test_size=0.33)
clf1 = LogisticRegression()
clf2 = LogisticRegressionNet(optimize='fmin_bfgs')
#clf2 = LogisticRegressionBin(optimize='fmin_bfgs')
clf3 = MLP(optimize='fmin_bfgs')
# X = [[0],[2]]
# y = [0,1]
# params = [0,1]
#
# X = np.asarray(X)
# y = np.asarray(y)
# params = np.asarray(params)
clf1.fit(X, y)
# X = np.asarray([[3, 5],[3,5],[3,5]])
# params = np.asarray([2, 1])
# y = np.asarray([1,1,1])
import torch
from net import MLP,CNN #
from torchvision import datasets, transforms
from sklearn.metrics import multilabel_confusion_matrix
#
test_loader = torch.utils.data.DataLoader(
datasets.FashionMNIST('./fashionmnist_data/', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=1, shuffle=True)
model=MLP()
device=torch.device('cpu')
model=model.to(device)
model.load_state_dict(torch.load('output/MLP.pt'))
model.eval()
pres=[]
labels=[]
i=0
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
pres.append(pred[0][0].item())
labels.append(target[0].item())
mcm = multilabel_confusion_matrix(labels, pres)#mcm
print(mcm)
#
import torch
import cv2
from PIL import Image
from net import MLP,CNN #
from torchvision import datasets, transforms
#
model=MLP()#这里可选CNN(),要看你前面训练的是哪个
device=torch.device('cpu')#用cpu进行推理
model=model.to(device)
model.load_state_dict(torch.load('output/MLP.pt'))
model.eval()#这一步很重要，这是告诉模型我们要验证，而不是训练
#--------以上就是推理之前模型的导入--------
print("-------加载模型成功----------")
class_dic={0:"T恤",1:"裤子",2:"套头衫",3:"连衣裙",4:"外套",5:"凉鞋",6:"衬衫",7:"运动鞋",8:"包",9:"靴子"}
data_transforms = transforms.Compose([
#transforms.ToTensor() convert a PIL image to tensor (HWC) in range [0,255] to a
#torch.Tensor(CHW)in the range [0.0,1.0]
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])
img=Image.open('test_image.jpg')#用于推理的图片
image=data_transforms(img)#预处理，转成tensor同时正则化
image=image.unsqueeze(0)#[1,28,28]->[1,1,28,28]
output = model(image.to(device))
pred = output.argmax(dim=1, keepdim=True)#
cls=pred.item()#输出在0~10之间代表10个类别
print("分类结果:",class_dic[cls])
correct = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.cuda(), target.cuda().long()
output = model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item()
pred = output.max(1, keepdim=True)[1]
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
return test_loss
Z_ = []
for i in trange(X.shape[0]):
Z_ += [[]]
for j in trange(Y.shape[0]):
convex_hull_weights = sum_weights([
multiply_weights(weight_dict_1, X[i, j]),
multiply_weights(weight_dict_2, Y[i, j]),
multiply_weights(weight_dict_3, Z[i, j])
])
net = MLP().cuda()
net.load_state_dict(convex_hull_weights)
Z_[i].append(test(net, test_loader))
np.save('./plots/X_mnist_random', X)
np.save('./plots/Y_mnist_random', Y)
np.save('./plots/Z_mnist_random', Z_)
pred = output.max(1, keepdim=True)[1]
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
return test_loss
global_vals = []
for i in trange(3):
# natural
weight_dict = torch.load('model_weights/mlp_weights_{}.pth'.format(i),
map_location='cpu')
# random
# weight_dict = torch.load('model_weights/mlp_random_weights_{}.pth'.format(i+42),
# map_location='cpu')
net = MLP().cuda()
net.load_state_dict(weight_dict)
I_w = test(net, test_loader)
vals = []
for tick in trange(20):
weight_dict_delta, delta = deepcopy(weight_dict),\
deepcopy(weight_dict)
norm = 0
for key in list(weight_dict_delta.keys())[-2:]:
delta[key] = torch.randn(delta[key].size())
norm += delta[key].norm().pow(2)
norm = norm.pow(0.5)
I_w_delta, r = I_w, 0.
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--batchsize', '-b', type=int, default=100,
help='Number of images in each mini-batch')
parser.add_argument('--epoch', '-e', type=int, default=20,
help='Number of sweeps over the mini_cifar to train')
parser.add_argument('--frequency', '-f', type=int, default=-1,
help='Frequency of taking a snapshot')
parser.add_argument('--gpu', '-g', type=int, default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--out', '-o', default='result',
help='Directory to output the result')
parser.add_argument('--resume', '-r', default='',
help='Resume the training from snapshot')
parser.add_argument('--unit', '-u', type=int, default=1000,
help='Number of units')
args = parser.parse_args()
print('GPU: {}'.format(args.gpu))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
# Set up a neural network to train
# Classifier reports softmax cross entropy loss and accuracy at every
# iteration, which will be used by the PrintReport extension below.
model = MLP(args.unit, 10)
if args.gpu >= 0:
# Make a specified GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu() # Copy the model to the GPU
# Setup an optimizer
optimizer = chainer.optimizers.Adam()
optimizer.setup(model)
# Load the MNIST mini_cifar
train, test = chainer.datasets.get_mnist()
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test, args.batchsize,
repeat=False, shuffle=False)
# Set up a trainer
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), out=args.out)
# Evaluate the model with the test mini_cifar for each epoch
trainer.extend(extensions.Evaluator(test_iter, model, device=args.gpu))
trainer.extend(extensions.dump_graph('main/loss'))
frequency = args.epoch if args.frequency == -1 else max(1, args.frequency)
trainer.extend(extensions.snapshot(), trigger=(frequency, 'epoch'))
trainer.extend(extensions.snapshot_object(model, 'model_{.updater.epoch}'),
trigger=(frequency, 'epoch'))
trainer.extend(extensions.LogReport())
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(['main/loss', 'validation/main/loss'],
'epoch', file_name='loss.png'))
trainer.extend(
extensions.PlotReport(
['main/accuracy', 'validation/main/accuracy'],
'epoch', file_name='accuracy.png'))
trainer.extend(extensions.PrintReport(
['epoch', 'main/loss', 'validation/main/loss',
'main/accuracy', 'validation/main/accuracy', 'elapsed_time']))
trainer.extend(extensions.ProgressBar())
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
trainer.run()
def __init__(self, action_space, obs_dim, gamma=1):
self.logits_net = MLP(input_dim=obs_dim, output_dim=len(action_space))
self.act_space = list(action_space)
self.optim = Adam(self.logits_net.parameters(), lr=5e-3)
def main():
db = sqlite3.connect('race.db')
c = db.cursor()
for i in range(10, 18):
sql = "select * from inputdata where headCount = " + str(
i) + " and race_id <= 27605 order by race_id,order_of_finish;"
c.execute(sql)
inputline = [] # [0,1,...,2]
inputdata = [] # 完成inputデータ
inputdataall = []
count = 0
label = []
labels = []
printflag = 1
for row in c:
row = list(row)
if (isinstance(row[47], int) == False):
row[47] = 18
if (isinstance(row[48], int) == False):
row[48] = 18
if (count % i == 0):
noneflag = 0
for j in range(53):
if (row[j] == None):
noneflag = 1
inputline.append(row[3])
inputline.append(row[35])
try:
inputline.append(row[46] / row[38])
except:
inputline.append(0)
inputline.append(row[39])
inputline.append(row[41])
inputline.append(row[45])
inputline.append(row[47])
inputline.append(row[48])
inputline.append(row[49])
inputdata.append(inputline)
inputline = []
label.append(row[2])
## if (count % i == 0):
## label.append(0)
## wintime = row[53]
## else:
## label.append(row[53] - wintime)
if (count % i == i - 1):
# inputline.insert(0, label)
if (noneflag == 0):
# dmean = np.array(inputdata).mean(axis=0, keepdims=True)
# dstd = np.std(inputdata, axis=0, keepdims=True)
# inputdata = (inputdata - dmean) / dstd
inputdata = scipy.stats.zscore(inputdata)
#分散0の処理
inputdata[np.isnan(inputdata)] = 0
inputdataall.extend(inputdata)
# lmean = np.mean(np.array(label),keepdims=True)
# lstd = np.std(label,keepdims=True)
horcenum = np.array([row[1]] * len(label))
labelnp = np.array(label) / horcenum
## labelnp = np.array(label)
labels.extend(labelnp)
inputdata = []
label = []
count = count + 1
inputdataall2 = np.empty((len(inputdataall), len(inputdataall[0])))
inputdataall2[:] = inputdataall
inputdataall = inputdataall2
# print(inputdata2)
# print(inputdata)
# X = inputdata[:, 1:].astype(np.float32)
if (i == 10):
allX = np.array(inputdataall, dtype='float32')
Y = np.array(labels, dtype='float32')
# le = LabelEncoder()
# allY = le.fit_transform(Y).astype(np.float32)
allY = Y.astype(np.float32)
else:
X = np.array(inputdataall, dtype='float32')
Y = np.array(labels, dtype='float32')
# le = LabelEncoder()
# Y = le.fit_transform(Y).astype(np.float32)
Y = Y.astype(np.float32)
allX = np.vstack((allX, X))
allY = np.hstack((allY, Y))
# print(X)
# print(X[0])
# print("-------")
# print(Y[0].dtype)
# print(Y[0])
# print(Y[0])
# Y=Y[:, None]
# threshold = np.int32(len(inputdata) / 10 * 9)
# train = np.array(inputdata[0:threshold],dtype=np.float32)
# test = np.array(inputdata[threshold:],dtype=np.float32)
# train = np.array(inputdata[0:threshold])
# train = train.astype(np.float32)
# test = np.array(inputdata[threshold:])
# test = test.astype(np.float32)
train, test = datasets.split_dataset_random(
datasets.TupleDataset(allX, allY), int(inputdataall.shape[0] * .7))
#全てtrainにぶちこむ
train, test2 = datasets.split_dataset_random(
datasets.TupleDataset(allX, allY),
int(inputdataall.shape[0] * .999999))
parser = argparse.ArgumentParser(description='Chainer example: RACE')
parser.add_argument('--batchsize',
'-b',
type=int,
default=100,
help='Number of images in each mini-batch')
parser.add_argument('--epoch',
'-e',
type=int,
default=20,
help='Number of sweeps over the mini_cifar to train')
parser.add_argument('--frequency',
'-f',
type=int,
default=-1,
help='Frequency of taking a snapshot')
parser.add_argument('--gpu',
'-g',
type=int,
default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--out',
'-o',
default='result',
help='Directory to output the result')
parser.add_argument('--resume',
'-r',
default='',
help='Resume the training from snapshot')
parser.add_argument('--unit',
'-u',
type=int,
default=1000,
help='Number of units')
args = parser.parse_args()
print('GPU: {}'.format(args.gpu))
print('# unit: {}'.format(args.unit))
print('# Minibatch-size: {}'.format(args.batchsize))
print('# epoch: {}'.format(args.epoch))
print('')
# Set up a neural network to train
# Classifier reports softmax cross entropy loss and accuracy at every
# iteration, which will be used by the PrintReport extension below.
model = MLP(args.unit, 1)
if args.gpu >= 0:
# Make a specified GPU current
chainer.cuda.get_device_from_id(args.gpu).use()
model.to_gpu() # Copy the model to the GPU
# Setup an optimizer
optimizer = chainer.optimizers.Adam(weight_decay_rate=0.01)
optimizer.setup(model)
# Load the MNIST mini_cifar
# train, test = chainer.datasets.get_mnist()
train_iter = chainer.iterators.SerialIterator(train, args.batchsize)
test_iter = chainer.iterators.SerialIterator(test,
args.batchsize,
repeat=False,
shuffle=False)
# Set up a trainer
updater = training.StandardUpdater(train_iter, optimizer, device=args.gpu)
trainer = training.Trainer(updater, (args.epoch, 'epoch'), out=args.out)
# Evaluate the model with the test mini_cifar for each epoch
trainer.extend(extensions.Evaluator(test_iter, model, device=args.gpu))
trainer.extend(extensions.dump_graph('main/loss'))
frequency = args.epoch if args.frequency == -1 else max(1, args.frequency)
trainer.extend(extensions.snapshot(), trigger=(frequency, 'epoch'))
trainer.extend(extensions.snapshot_object(model, 'model_{.updater.epoch}'),
trigger=(frequency, 'epoch'))
trainer.extend(extensions.LogReport())
if extensions.PlotReport.available():
trainer.extend(
extensions.PlotReport(['main/loss', 'validation/main/loss'],
'epoch',
file_name='loss.png'))
trainer.extend(
extensions.PlotReport(
['main/accuracy', 'validation/main/accuracy'],
'epoch',
file_name='accuracy.png'))
trainer.extend(
extensions.PrintReport([
'epoch', 'main/loss', 'validation/main/loss', 'main/accuracy',
'validation/main/accuracy', 'elapsed_time'
]))
trainer.extend(extensions.ProgressBar())
if args.resume:
chainer.serializers.load_npz(args.resume, trainer)
trainer.run()
def main():
# parser是训练和测试的一些参数设置，如果default里面有数值，则默认用它，
# 要修改可以修改default，也可以在命令行输入
parser = argparse.ArgumentParser(description='PyTorch MNIST Example')
parser.add_argument('--model', default='CNN',#这里选择你要训练的模型
help='CNN or MLP')
parser.add_argument('--batch-size', type=int, default=128, metavar='N',
help='input batch size for training (default: 64)')
parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
help='input batch size for testing (default: 1000)')
parser.add_argument('--epochs', type=int, default=10, metavar='N',
help='number of epochs to train (default: 10)')
parser.add_argument('--lr', type=float, default=0.01, metavar='LR',
help='learning rate (default: 0.01)')
parser.add_argument('--momentum', type=float, default=0.5, metavar='M',
help='SGD momentum (default: 0.5)')
parser.add_argument('--no-cuda', action='store_true', default=False,
help='disables CUDA training')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 1)')
parser.add_argument('--log-interval', type=int, default=50, metavar='N',
help='how many batches to wait before logging training status')
parser.add_argument('--save-model', action='store_true', default=True,
help='For Saving the current Model')
parser.add_argument('--save_dir', default='output/',#模型保存路径
help='dir saved models')
args = parser.parse_args()
#torch.cuda.is_available()会判断电脑是否有可用的GPU,没有则用cpu训练
use_cuda = not args.no_cuda and torch.cuda.is_available()
print(use_cuda)
torch.manual_seed(args.seed)
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
train_loader = torch.utils.data.DataLoader(
datasets.FashionMNIST('./fashionmnist_data/', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(
datasets.FashionMNIST('./fashionmnist_data/', train=False, transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=args.test_batch_size, shuffle=True, **kwargs)
writer=SummaryWriter()#用于记录训练和测试的信息:loss,acc等
if args.model=='CNN':
model = CNN().to(device)#CNN() or MLP
if args.model=='MLP':
model = MLP().to(device)#CNN() or MLP
optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum) #optimizer存储了所有parameters的引用，每个parameter都包含gradient
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[12, 24], gamma=0.1) #学习率按区间更新
model.train()
log_loss=0
log_acc=0
for epoch in range(1, args.epochs + 1):
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = model(data)
loss = F.nll_loss(output, target) # negative log likelihood loss(nll_loss), sum up batch cross entropy
loss.backward()
optimizer.step() # 根据parameter的梯度更新parameter的值
# 这里设置每args.log_interval个间隔打印一次训练信息，同时进行一次验证，并且将验证(测试)的准确率存入writer
if batch_idx % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
epoch, batch_idx * len(data), len(train_loader.dataset),
100. * batch_idx / len(train_loader), loss.item()))
#下面是模型验证过程
model.eval()
test_loss = 0
correct = 0
with torch.no_grad(): # 无需计算梯度
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss
pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(test_loader.dataset)
writer.add_scalars('loss', {'train_loss':loss,'val_loss':test_loss},global_step=log_acc)
writer.add_scalar('val_accuracy', correct / len(test_loader.dataset), global_step=log_acc)
log_acc += 1
print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
test_loss, correct, len(test_loader.dataset),
100. * correct / len(test_loader.dataset)))
model.train()
if (args.save_model):#保存训练好的模型
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
torch.save(model.state_dict(), os.path.join(args.save_dir,args.model+".pt"))
writer.add_graph(model, (data,))# 将模型结构保存成图，跟踪数据流动
writer.close()
if steps % 100 == 0:
print("%i/%i" % (steps, max_test_steps))
if steps >= max_test_steps:
break
print('total_reward:', totalr)
def predict(self, X):
#print("\t", X.shape)
y_pred = self.sess.run(self.net.pred, feed_dict={self.net.X: X})
return y_pred
def init_all_variables(self):
if not self.init_variables:
self.init_variables = True
self.sess.run(tf.initialize_all_variables())
def save(self, ):
pass
if __name__ == '__main__':
#dagger = DAgger()
# now I find the best num of the total of epoch is around 50
env = gym.make('Hopper-v1')
X_dim = env.observation_space.shape[0]
y_dim = env.action_space.shape[0]
net = MLP(X_dim, y_dim, [50, 100])
dagger = DAgger(env, net)
dagger.train_dagger(num_dagger=2)
dagger.test()
def main():
parser = argparse.ArgumentParser(description='Chainer example: MNIST')
parser.add_argument('--gpu',
'-g',
type=int,
default=-1,
help='GPU ID (negative value indicates CPU)')
parser.add_argument('--image',
'-i',
type=str,
default="",
help='pass to input image')
parser.add_argument('--model',
'-m',
default='my_mnist.model',
help='path to the training model')
parser.add_argument('--unit',
'-u',
type=int,
default=1000,
help='Number of units')
args = parser.parse_args()
model = MLP(args.unit, 1)
if args.gpu >= 0:
model.to_gpu(chainer.cuda.get_device_from_id(args.gpu).use())
serializers.load_npz(args.model, model)
# try:
# img = Image.open(args.image).convert("L").resize((28,28))
# except :
# print("invalid input")
# return
# img_array = model.xp.asarray(img,dtype=model.xp.float32).reshape(1,784)
## df = pd.read_csv('test1.csv')
db = sqlite3.connect('race.db')
c = db.cursor()
win = []
lose = []
quinella = []
place = []
none_flag = 0
for race_id in range(27606, 34440):
# for race_id, sdf in df.groupby('race_id'):
if (race_id % 100 == 1):
print("finished ", race_id - 1)
df = pd.read_sql("select horse_number,age,winrate,eps,odds,weight,preOOF,pre2OOF,preLastPhase, "\
"payoff_quinella,payoff_place, race_id "\
"from (select "\
"inputdata.race_id, "\
"inputdata.order_of_finish, "\
"inputdata.horse_number horse_number, "\
"age, "\
"case when enterTimes != 0 then winRun/enterTimes else 0 end as winrate, "\
"eps, "\
"odds, "\
"weight, "\
"preOOF, "\
"pre2OOF, "\
"preLastPhase, "\
"pay1.payoff payoff_quinella, "\
"pay2.payoff payoff_place "\
"from inputdata "\
"inner join payoff pay1 "\
" on pay1.ticket_type = 3 and pay1.race_id = inputdata.race_id "\
"left join payoff pay2"\
" on pay2.ticket_type = 1"\
" and pay2.horse_number = inputdata.horse_number"\
" and pay2.race_id = inputdata.race_id"\
") as a "\
"where a.race_id = "+str(race_id)+" "\
"order by a.race_id,order_of_finish;", db)
# img_array=df.values.reshape(1,-1)[~np.isnan(df.values.reshape(1,-1))]
# img_array = model.xp.asarray(img_array, dtype=model.xp.float32).reshape(1,-1)
arr = df.values
#あいてるid用
# if(len(arr)==0):
# continue
for i in range(len(arr)):
if ((isinstance(arr[i][6], int)
or isinstance(arr[i][6], float)) == False):
arr[i][6] = 18
if ((isinstance(arr[i][7], int)
or isinstance(arr[i][7], float)) == False):
arr[i][7] = 18
arr = np.array(arr, dtype=float)
#None処理
for i in range(len(arr)):
if (np.isnan(arr[i][10])):
arr[i][10] = 0
copy_arr = arr
winner = arr[0][0]
second = arr[1][0]
winner_odds = arr[0][4]
quinella_odds = arr[0][9]
#none,nanがあるならとばす。
for i in range(len(arr)):
for j in range(len(arr[0])):
if arr[i][j] is None:
none_flag = 1
elif (math.isnan(float(arr[i][j]))):
none_flag = 1
if (none_flag):
none_flag = 0
continue
arr = arr.astype(np.float32)
arr = scipy.stats.zscore(arr)
#分散0の処理
arr[np.isnan(arr)] = 0
res = []
for i in range(len(arr)):
img_array = arr[i][0:9]
img_array = model.xp.asarray(img_array,
dtype=model.xp.float32).reshape(
1, -1)
with chainer.using_config('train',
False), chainer.no_backprop_mode():
result = model.predict(img_array)
res.append(result.data[0])
# print("predict:", model.xp.argmax(result.data))
# arg_sorted = model.xp.argsort(result.data)
# arg_sorted = arg_sorted [:, ::-1]
# print(arg_sorted[:, :3])
x = np.array(res).reshape((1, -1))[0]
# 一着がほかより抜けている時のみ買う
if ((x[np.argsort(x)[1]] - x[np.argsort(x)[0]]) < 0.001):
continue
# for i in range(len(x)):
# print (np.argsort(x)[i]+1,"-", x[np.argsort(x)[i]])
# 一二着がほかより抜けている時のみ買う
# if ((x[np.argsort(x)[2]] - x[np.argsort(x)[1]]) < 0.001):
# continue
# 狙うオッズが微妙ならとばす。
continue_flag = 0
for j in range(len(copy_arr)):
if (copy_arr[j][0] == np.argsort(x)[0] + 1):
if (copy_arr[j][4] >= 50 or copy_arr[j][4] < 2):
continue_flag = 1
if (continue_flag == 1):
continue
if (np.argsort(x)[0] + 1 == winner):
win.append(winner_odds)
# print(race_id,np.argsort(x)[0]+1,winner_odds)
else:
win.append(0)
for j in range(len(copy_arr)):
if (copy_arr[j][0] == np.argsort(x)[0] + 1):
lose.append(copy_arr[j][4])
if (((np.argsort(x)[0] + 1 == winner) and
(np.argsort(x)[1] + 1 == second))
or ((np.argsort(x)[0] + 1 == second) and
(np.argsort(x)[1] + 1 == winner))):
quinella.append(quinella_odds)
else:
quinella.append(0)
for i in range(len(arr)):
if (np.argsort(x)[0] + 1 == copy_arr[i][0]):
place.append(copy_arr[i][10])
print(win)
print(lose)
print(place)
print(quinella)
print("単勝")
print("回収率 = ",
sum(win) / len(win) * 100, " 的中率 = ",
(1 - win.count(0) / len(win)) * 100)
print("\n複勝")
print("回収率 = ",
sum(place) / len(place), " 的中率 = ",
(1 - place.count(0) / len(place)) * 100)
print("\n馬連")
print("回収率 = ",
sum(quinella) / len(quinella), " 的中率 = ",
(1 - quinella.count(0) / len(quinella)) * 100)
def main():
# ----- 根据data来读取不同的数据和不同的loss、metrics -----
if config.args.data == 'brca':
rna = RnaData.predicted_data(config.brca_cli, config.brca_rna,
{'PAM50Call_RNAseq': 'pam50'})
rna.transform(tf.LabelMapper(config.brca_label_mapper))
out_shape = len(config.brca_label_mapper)
criterion = nn.CrossEntropyLoss()
scorings = (mm.Loss(), mm.Accuracy(), mm.BalancedAccuracy(),
mm.F1Score(average='macro'), mm.Precision(average='macro'),
mm.Recall(average='macro'), mm.ROCAUC(average='macro'))
elif config.args.data == 'survival':
if os.path.exists('./DATA/temp_pan.pth'):
rna = RnaData.load('./DATA/temp_pan.pth')
else:
rna = RnaData.survival_data(config.pan_cli, config.pan_rna,
'_OS_IND', '_OS')
out_shape = 1
if config.args.loss_type == 'cox':
criterion = NegativeLogLikelihood()
elif config.args.loss_type == 'svm':
criterion = SvmLoss(rank_ratio=config.args.svm_rankratio)
scorings = (mm.Loss(), mm.CIndex())
rna.transform(tf.ZeroFilterCol(0.8))
rna.transform(tf.MeanFilterCol(1))
rna.transform(tf.StdFilterCol(0.5))
norm = tf.Normalization()
rna.transform(norm)
# ----- 构建网络和优化器 -----
inpt_shape = rna.X.shape[1]
if config.args.net_type == 'mlp':
net = MLP(inpt_shape, out_shape, config.args.hidden_num,
config.args.block_num).cuda()
elif config.args.net_type == 'atten':
net = SelfAttentionNet(inpt_shape, out_shape, config.args.hidden_num,
config.args.bottle_num, config.args.block_num,
config.args.no_res, config.act,
config.args.no_head, config.args.no_bottle,
config.args.no_atten,
config.args.dropout_rate).cuda()
elif config.args.net_type == 'resnet':
net = ResidualNet(inpt_shape, out_shape, config.args.hidden_num,
config.args.bottle_num,
config.args.block_num).cuda()
# ----- 训练网络，cross validation -----
split_iterator = rna.split_cv(config.args.test_size,
config.args.cross_valid)
train_hists = []
test_hists = []
for split_index, (train_rna, test_rna) in enumerate(split_iterator):
print('##### save: %s, split: %d #####' %
(config.args.save, split_index))
# 从train中再分出一部分用作验证集，决定停止
train_rna, valid_rna = train_rna.split(0.1)
dats = {
'train': train_rna.to_torchdat(),
'valid': valid_rna.to_torchdat(),
}
dataloaders = {
k: data.DataLoader(v, batch_size=config.args.batch_size)
for k, v in dats.items()
}
test_dataloader = data.DataLoader(test_rna.to_torchdat(),
batch_size=config.args.batch_size)
# 网络训练前都进行一次参数重置，避免之前的训练的影响
net.reset_parameters()
# train
optimizer = optim.Adamax(net.parameters(),
lr=config.args.learning_rate)
lrs = config.lrs(optimizer)
net, hist = train(
net,
criterion,
optimizer,
dataloaders,
epoch=config.args.epoch,
metrics=scorings,
l2=config.args.l2,
standard_metric_index=config.args.standard_metric_index,
scheduler=lrs)
# test
test_res = evaluate(net, criterion, test_dataloader, metrics=scorings)
# 将多次训练的结果保存到一个df中
hist = pd.DataFrame(hist)
hist['split_index'] = split_index
train_hists.append(hist)
# 保存多次test的结果
test_res['split_index'] = split_index
test_hists.append(test_res)
# 每个split训练的模型保存为一个文件
torch.save(net.state_dict(),
os.path.join(config.save_dir, 'model%d.pth' % split_index))
# 保存train的结果
train_hists = pd.concat(train_hists)
train_hists.to_csv(os.path.join(config.save_dir, 'train.csv'))
# 保存test的结果
test_hists = pd.DataFrame(test_hists)
test_hists.to_csv(os.path.join(config.save_dir, 'test.csv'))
