def predict_autograd(test_data):
checkpoint_path = 'saved_models/qm9_ens_seed60/fold_0/model_0/model.pt'
state = torch.load(checkpoint_path, map_location=lambda storage, loc: storage)
args, loaded_state_dict = state['args'], state['state_dict']
scaler = StandardScaler(state['data_scaler']['means'],
state['data_scaler']['stds']) if state['data_scaler'] is not None else None
for k in ['encoder.encoder.cached_zero_vector', 'encoder.encoder.W_i.weight', 'encoder.encoder.W_h.weight',
'encoder.encoder.W_o.weight', 'encoder.encoder.W_o.bias']:
loaded_state_dict.pop(k, None)
# Build model
model = build_model(args)
model_state_dict = model.state_dict()
# Skip missing parameters and parameters of mismatched size
pretrained_state_dict = {}
for param_name in loaded_state_dict.keys():
if param_name not in model_state_dict:
print(f'Pretrained parameter "{param_name}" cannot be found in model parameters.')
elif model_state_dict[param_name].shape != loaded_state_dict[param_name].shape:
print(f'Pretrained parameter "{param_name}" '
f'of shape {loaded_state_dict[param_name].shape} does not match corresponding '
f'model parameter of shape {model_state_dict[param_name].shape}.')
else:
# print(f'Loading pretrained parameter "{param_name}".')
pretrained_state_dict[param_name] = loaded_state_dict[param_name]
# Load pretrained weights
model_state_dict.update(pretrained_state_dict)
model.load_state_dict(model_state_dict)
model.eval()
test_data = v(torch.from_numpy(test_data).float(), requires_grad=True)
# with torch.no_grad():
model_preds, ale_pred = model(test_data)
ale_pred = torch.exp(ale_pred)
model_preds.backward()
if scaler is not None:
model_preds = scaler.inverse_transform(model_preds.detach())
ale_pred = scaler.inverse_transform_variance(ale_pred.detach())
model_preds = np.array(model_preds.tolist(), dtype=np.float)
ale_pred = np.array(ale_pred.tolist(), dtype=np.float)
grad_rms = torch.sqrt(torch.sum(torch.square(test_data.grad.data))/1000).numpy()
grad_max = torch.max(torch.sqrt(torch.square(test_data.grad.data))).numpy()
return model_preds, ale_pred, grad_rms, grad_max
def val(model,dataloder,device):
model.eval()
confusion_matrix = meter.ConfusionMeter(7)
for ii, (input,label) in enumerate(dataloder):
val_input = v(input)
val_input=val_input.to(device)
label=label.to(device)
score = model(val_input)
confusion_matrix.add(score.data,label.data)
model.train()
cm_value = confusion_matrix.value()
print(cm_value)
correct = 0
for i in range(7):
correct+=cm_value[i][i]
accuracy = 100.*(correct)/(cm_value.sum())
return confusion_matrix, accuracy
def val(args, model, dataloder, device, critertion):
model.eval()
confusion_matrix = meter.ConfusionMeter(cfg.num_classes)
for ii, (input, label) in enumerate(dataloder):
with t.no_grad():
val_input = v(input)
val_input = val_input.to(device)
#label = label.reshape(args.batchsize,6)
label = label.to(device)
score = model(val_input)
val_loss = critertion.forward(score, label.long())
confusion_matrix.add(score.data, label.data)
cm_value = confusion_matrix.value()
print(cm_value)
correct = 0
for i in range(6):
correct += cm_value[i][i]
accuracy = 100. * (correct) / (cm_value.sum())
return confusion_matrix, accuracy, val_loss.item()
import torch as t
from torch.autograd import Variable as v
# simple gradient
a = v(t.FloatTensor([2, 3]), requires_grad=True)
b = a + 3
c = b * b * 3
out = c.mean()
out.backward()
print('*'*10)
print('=====simple gradient======')
print('input')
print(a.data)
print('compute result is')
print(out.data[0])
print('input gradients are')
print(a.grad.data)
# backward on non-scalar output
m = v(t.FloatTensor([[2, 3]]), requires_grad=True)
n = v(t.zeros(1, 2))
n[0, 0] = m[0, 0] ** 2
n[0, 1] = m[0, 1] ** 3
n.backward(t.FloatTensor([[1, 1]]))
print('*'*10)
print('=====non scalar output======')
print('input')
print(m.data)
print('input gradients are')
print(m.grad.data)
def predict_autograd_a(smile, checkpoint_path):
state = torch.load(checkpoint_path,
map_location=lambda storage, loc: storage)
args, loaded_state_dict = state['args'], state['state_dict']
scaler = StandardScaler(
state['data_scaler']['means'], state['data_scaler']
['stds']) if state['data_scaler'] is not None else None
mol_smile = mol2graph_autograd_a(smile, args)
b_num = mol_smile.f_bonds.shape[
0] - 1 # vector number = total - cached_zero_vector(1)
if b_num == 0:
return np.zeros((1, 1)), 0, 0, 0
# print('f_bond', mol_smile.f_bonds)
# Build model
model = build_model_autograd_a(args)
model_state_dict = model.state_dict()
# Skip missing parameters and parameters of mismatched size
pretrained_state_dict = {}
for param_name in loaded_state_dict.keys():
if param_name not in model_state_dict:
print(
f'Pretrained parameter "{param_name}" cannot be found in model parameters.'
)
elif model_state_dict[param_name].shape != loaded_state_dict[
param_name].shape:
print(
f'Pretrained parameter "{param_name}" '
f'of shape {loaded_state_dict[param_name].shape} does not match corresponding '
f'model parameter of shape {model_state_dict[param_name].shape}.'
)
else:
# print(f'Loading pretrained parameter "{param_name}".')
pretrained_state_dict[param_name] = loaded_state_dict[param_name]
# Load pretrained weights
model_state_dict.update(pretrained_state_dict)
model.load_state_dict(model_state_dict)
model.eval()
# load bond_vec
mol_graph = mol2graph_autograd_a(smile, args)
f_atoms, f_bonds, a2b, b2a, b2revb, a_scope, b_scope, smiles_batch = mol_graph.get_components(
)
f_bonds = v(f_bonds, requires_grad=True)
model_preds, ale_pred = model(smile, f_bonds)
ale_pred = torch.exp(ale_pred)
model_preds.backward()
atom_grad = []
for i in torch.square(f_bonds.grad.data):
atom_grad.append((torch.sum(i).numpy() / len(i))**(1 / 2))
atom_max = max(atom_grad)
atom_rms = sum(atom_grad[1:]) / len(atom_grad[1:])
if scaler is not None:
model_preds = scaler.inverse_transform(model_preds.detach())
ale_pred = scaler.inverse_transform_variance(ale_pred.detach())
model_preds = np.array(model_preds.tolist(), dtype=np.float)
ale_pred = np.array(ale_pred.tolist(), dtype=np.float)
return model_preds, ale_pred, atom_rms, atom_max
import torch as t
from torch.autograd import Variable as v
# simple gradient
a = v(t.FloatTensor([2, 3]), requires_grad=True)
b = a + 3
c = b * b * 3
out = c.mean()
out.backward()
print('*'*20)
print('------- simple gradient -------')
print('input')
print(a.data)
print('compute result is')
print(out.item())
print('input gradients are')
# As c = (3*(a+3)^2)/2
# so, d(c)/d(a) = 3*(a+3)
# => a.grad.data = [3*(2+3), 3*(3+3)] = [15, 18]
print(a.grad.data)
# backward on non-scalar output
m = v(t.FloatTensor([[2, 3]]), requires_grad=True)
n = v(t.zeros(1, 2))
n[0, 0] = m[0, 0] ** 2
n[0, 1] = m[0, 1] ** 3
n.backward(t.FloatTensor([[1, 1]]))
print('*' * 20)
print('------- non scalar output -------')
print('input')
model = make_model(args.model,num_classes=6,pretrained=args.pretrained,input_size = (cfg.IMAGE_SIZE,cfg.IMAGE_SIZE))
device = t.device("cuda" if t.cuda.is_available() else "cpu")
model = nn.DataParallel(model)
model = model.to(device)
model.load_state_dict(t.load(resume_model))
model.eval()
confusion_matrix = meter.ConfusionMeter(6)
fpr = dict()
tpr = dict()
roc_auc = dict()
thresholds = dict()
labels = np.zeros((cfg.BATCHSIZE,6))
scores = np.zeros((cfg.BATCHSIZE,6))
for ii, (input,label) in enumerate(test_dataloader):
with t.no_grad():
val_input = v(input)
val_input=val_input.to(device)
label=label.to(device)
score = model(val_input)
confusion_matrix.add(score.data,label.data)
label = label_binarize(label.cpu().detach().numpy(),classes=[0,1,2,3,4,5])
score = score.cpu().detach().numpy()
labels = np.concatenate((labels,label),axis=0)
scores = np.concatenate((scores,score),axis=0)
for i in range(6):
fpr[i], tpr[i], _ = roc_curve(labels[:,i],scores[:,i])
roc_auc[i] = auc(fpr[i],tpr[i])
# Compute macro-average ROC curve and ROC area（方法一）
# First aggregate all false positive rates
all_fpr = np.unique(np.concatenate([fpr[i] for i in range(6)]))
import torch as t
from torch.autograd import Variable as v
from matplotlib import pyplot as plt
t.manual_seed(1000)
def get_fake_data(batch_size=8):
x = t.rand(batch_size, 1) * 20
y = x * 2 + (1 + t.rand(batch_size, 1)) * 3
return x, y
w = v(t.rand(1, 1), requires_grad=True)
b = v(t.zeros(1, 1), requires_grad=True)
lr = 0.001
for ii in range(1000):
x, y = get_fake_data()
x, y = v(x), v(y)
y_pred = x.mm(w) + b.expand_as(y)
loss = 0.5 * (y_pred - y)**2
loss = loss.sum()
loss.backward()
w.data.sub_(lr * w.grad.data)
b.data.sub_(lr * b.grad.data)
w.grad.data.zero_()
b.grad.data.zero_()