def compute_matching(current_pool_list,
curr_type_weights,
e_weights_by_type,
gamma=0.000001):
# current_pool_list should have lhs and rhs, get them both as tensors
lhs_current_elems = torch.tensor([x[0] for x in current_pool_list.lhs])
rhs_current_elems = torch.tensor([x[0] for x in current_pool_list.rhs])
l_n = lhs_current_elems.shape[0]
r_n = rhs_current_elems.shape[0]
A, b = make_matching_matrix(l_n, r_n)
A = torch.from_numpy(A).float()
b = torch.from_numpy(b).float()
# should take lhs and rhs
e_weights = weight_matrix(current_pool_list.lhs, current_pool_list.rhs,
e_weights_by_type).view(l_n, r_n)
jitter_e_weights = e_weights + 1e-4 * jitter_matrix(l_n, r_n)
# e_weights = torch.rand(n,n)
model_params_quad = make_gurobi_model(A.detach().numpy(),
b.detach().numpy(), None, None,
gamma * np.eye(A.shape[1]))
func = QPFunction(verbose=False,
solver=QPSolvers.GUROBI,
model_params=model_params_quad)
Q_mat = gamma * torch.eye(A.shape[1])
curr_elem_weights = type_weight_matrix(lhs_current_elems,
rhs_current_elems,
curr_type_weights).view(l_n, r_n)
modified_edge_weights = jitter_e_weights - 0.5 * (curr_elem_weights)
# may need some negative signs
resulting_match = func(Q_mat, -modified_edge_weights.view(-1), A, b,
torch.Tensor(), torch.Tensor()).view(l_n, r_n)
return resulting_match, e_weights
def test(self, X, y, relaxation=False):
Q = self.Q
G = self.G
h = self.h
n_train = 1
n_items = self.n_items
epochs = self.epochs
net = self.net
capacity = self.capacity
model = self.model
weights = self.weights
model.eval()
X_tensor = torch.tensor(X, dtype=torch.float)
y_pred = model(X_tensor).detach().numpy().squeeze()
n_knapsacks = X.shape[0] // n_items
regret_list = []
cf_list = []
loss_list = []
time = 0
model_params_quad = make_gurobi_model(G.detach().numpy(),
h.detach().numpy(), None, None,
Q.detach().numpy())
for i in range(n_knapsacks):
n_start = n_items * i
n_stop = n_start + n_items
regret, cf = regret_knapsack([y[n_start:n_stop]],
[y_pred[n_start:n_stop]],
weights=weights,
cap=[self.capacity],
relaxation=relaxation)
regret_list.append(regret)
cf_list.append(cf)
z = torch.tensor(y[n_start:n_stop], dtype=torch.float)
X_tensor = torch.tensor(X[n_start:n_stop, :], dtype=torch.float)
c_true = -z
c_pred = -(model(X_tensor))
solver = QPFunction(verbose=False,
solver=QPSolvers.GUROBI,
model_params=model_params_quad)
x = solver(Q.expand(n_train, *Q.shape), c_pred.squeeze(),
G.expand(n_train, *G.shape),
h.expand(n_train, *h.shape), torch.Tensor(),
torch.Tensor())
time += solver.Runtime()
loss_list.append((x.squeeze() * c_true).mean().item())
model.train()
if not relaxation:
tn, fp, fn, tp = np.sum(np.stack(cf_list), axis=0).ravel()
accuracy = (tn + tp) / (tn + fp + fn + tp)
else:
accuracy = None
return np.median(regret_list), mse(
y, y_pred), accuracy, np.median(loss_list), time
def train_fwbw(self,
data,
init_criterion,
init_opti,
sol_noispool,
epoch=None):
indices = list(range(len(data)))
train_dl = DataLoader(data, batch_size=self.bsizeCOP, shuffle=True)
optimizer = init_opti
sign = -1 if self.maximize else 1 # QPf minimizes
ii = 0
A, b, G, h = self.get_qpt_mat(**self.param)
G_trch = torch.from_numpy(G if G is not None else np.array([])).float()
h_trch = torch.from_numpy(h if h is not None else np.array([])).float()
A_trch = torch.from_numpy(A if A is not None else np.array([])).float()
b_trch = torch.from_numpy(b if b is not None else np.array([])).float()
Q_trch = (self.tau) * torch.eye(G.shape[1])
model_params_quad = make_gurobi_model(G, h, A, b,
Q_trch.detach().numpy())
for xb, yb, solb in train_dl:
optimizer.zero_grad()
for i in range(len(yb)):
y_pred = self.model(xb[i]).squeeze()
_, _, G, h = self.get_mat(mb[i], **self.param)
G_trch = torch.from_numpy(G).float()
h_trch = torch.from_numpy(h).float()
sol = IPOfunc(A=A_trch,
b=b_trch,
G=G_trch,
h=h_trch,
pc=True,
max_iter=self.max_iter,
thr=self.thr,
method=self.method,
damping=self.damping)(sign * y_pred)
loss = -(sol * yb[i]).mean()
loss.backward()
optimizer.step()
def train_fwbw(self,
data,
init_criterion,
init_opti,
sol_noispool,
epoch=None):
indices = list(range(len(data)))
train_dl = DataLoader(data, batch_size=self.bsizeCOP, shuffle=True)
optimizer = init_opti
sign = -1 if self.maximize else 1 # QPf minimizes
ii = 0
A, b, G, h = self.get_qpt_mat(**self.param)
G_trch = torch.from_numpy(G if G is not None else np.array([])).float()
h_trch = torch.from_numpy(h if h is not None else np.array([])).float()
A_trch = torch.from_numpy(A if A is not None else np.array([])).float()
b_trch = torch.from_numpy(b if b is not None else np.array([])).float()
Q_trch = (self.tau) * torch.eye(G.shape[1])
model_params_quad = make_gurobi_model(G, h, A, b,
Q_trch.detach().numpy())
for xb, yb, solb in train_dl:
optimizer.zero_grad()
for i in range(len(yb)):
y_pred = self.model(xb[i]).squeeze()
sol = QPFunction(verbose=False,
solver=QPSolvers.GUROBI,
model_params=model_params_quad)(
Q_trch.expand(1, *Q_trch.shape),
sign * y_pred,
G_trch.expand(1, *G_trch.shape),
h_trch.expand(1, *h_trch.shape),
A_trch.expand(1, *A_trch.shape),
b_trch.expand(1, *b_trch.shape))
ii += 1
loss = -(sol * yb[i]).mean()
loss.backward()
optimizer.step()
def fit(self, economic_data, properties_data):
logging.info("QPTL")
train_df = MyCustomDataset(economic_data, properties_data)
grad_list = []
for e in range(self.epochs):
total_loss = 0
# for i in range(30):
# logging.info("EPOCH Starts")
# train_prop = properties_data.sample(n = 279,random_state =i)
# valid_prop = properties_data.loc[~properties_data.index.isin(train_prop.index)]
# train_sl = train_prop.Sl.unique().tolist()
# valid_sl = valid_prop.Sl.unique().tolist()
# train_prop = train_prop.sort_values(['Sl'],ascending=[True])
# valid_prop = valid_prop.sort_values(['Sl'],ascending=[True])
# train_econ = economic_data[economic_data.Sl.isin(train_sl)]
# valid_econ = economic_data[economic_data.Sl.isin(valid_sl)]
# train_econ = train_econ.sort_values(['Sl','Lag'],ascending=[True,False])
# valid_econ = valid_econ.sort_values(['Sl','Lag'],ascending=[True,False])
# train_df = MyCustomDataset(train_econ,train_prop)
train_dl = data_utils.DataLoader(train_df,
batch_size=self.batch_size,
shuffle=False)
for x_f, x_c, x_t, y, cst in train_dl:
self.optimizer.zero_grad()
h = torch.zeros(
(self.num_layers, x_t.shape[0], self.hidden_size),
dtype=torch.float)
c = torch.zeros(
(self.num_layers, x_t.shape[0], self.hidden_size),
dtype=torch.float)
op, states = self.model(x_f, x_c, x_t, (h, c))
h, c = states
G, h, A, b = make_matrix_qp(cst.detach().numpy(),
x_t.shape[0] * self.budget)
Q = torch.eye(x_t.shape[0]) / self.tau
model_params_quad = make_gurobi_model(
G.detach().numpy(),
h.detach().numpy(),
A.detach().numpy(),
b.detach().numpy(), self.tau * np.eye(x_t.shape[0]))
x = QPFunction(
verbose=False,
solver=QPSolvers.GUROBI,
model_params=model_params_quad)(Q.expand(1, *Q.shape), -op,
G.expand(1, *G.shape),
h.expand(1, *h.shape),
A.expand(1, *A.shape),
b.expand(1, *b.shape))
loss = -(x * y).mean()
op.retain_grad()
loss.backward()
# op_grad = copy.deepcopy(op.grad)
# grad_dict = {}
# grad_dict['epoch'] = e
# grad_dict['subepoch'] = i
# for l in range(len(op_grad)):
# grad_dict['qpt_cgrad'] = op_grad[l].item()
# grad_dict['prediction'] = op[l].item()
# grad_dict['true'] = y[l].item()
# grad_list.append(copy.deepcopy(grad_dict))
self.optimizer.step()
total_loss += loss.item()
logging.info("EPOCH Ends")
def fit(self,
X,
y,
X_validation=None,
y_validation=None,
X_test=None,
y_test=None):
# if validation true validation and tets data should be provided
tau = self.tau
start = time.time()
validation_time = 0
test_time = 0
# if validation true validation and tets data should be provided
X = X[:, 1:]
validation = (X_validation is not None) and (y_validation is not None)
test = (X_test is not None) and (y_test is not None)
if self.doScale:
self.scaler = preprocessing.StandardScaler().fit(X)
X = self.scaler.transform(X)
if validation:
start_validation = time.time()
X_validation = X_validation[:, 1:]
if self.doScale:
X_validation = self.scaler.transform(X_validation)
end_validation = time.time()
validation_time += end_validation - start_validation
if test:
start_test = time.time()
X_test = X_test[:, 1:]
if self.doScale:
X_test = self.scaler.transform(X_test)
end_test = time.time()
test_time += end_test - start_test
validation_relax = self.validation_relax
test_relax = self.test_relax
n_items = self.n_items
epochs = self.epochs
net = self.net
capacity = self.capacity
weights = self.weights
hyperparams = self.hyperparams
#Q= torch.diagflat(torch.ones(n_items)/tau)
Q = torch.eye(n_items) / tau
#G = torch.cat((torch.from_numpy(weights).float(), torch.diagflat(torch.ones(n_items)),
# torch.diagflat(torch.ones(n_items)*-1)), 0)
#h = torch.cat((torch.tensor([capacity],dtype=torch.float),torch.ones(n_items),torch.zeros(n_items)))
G = torch.from_numpy(weights).float()
h = torch.tensor([capacity], dtype=torch.float)
self.Q = Q
self.G = G
self.h = h
model = net(X.shape[1], 1)
self.model = model
#optimizer = torch.optim.Adam(model.parameters(),**hyperparams)
optimizer = self.optimizer(model.parameters(), **hyperparams)
model_params_quad = make_gurobi_model(G.detach().numpy(),
h.detach().numpy(), None, None,
Q.detach().numpy())
n_knapsacks = X.shape[0] // n_items
loss_list = []
accuracy_list = []
regret_list = []
subepoch_list = []
subepoch = 0
logger = []
test_list = []
n_train = 1
shuffled_batches = [i for i in range(n_knapsacks)]
for e in range(epochs):
np.random.shuffle(shuffled_batches)
logging.info('Epoch %d' % e)
for i in range(n_knapsacks):
n_start = n_items * shuffled_batches[i]
n_stop = n_start + n_items
z = torch.tensor(y[n_start:n_stop], dtype=torch.float)
X_tensor = torch.tensor(X[n_start:n_stop, :],
dtype=torch.float)
c_true = -z
c_pred = -(model(X_tensor))
solver = QPFunction(verbose=False,
solver=QPSolvers.GUROBI,
model_params=model_params_quad)
x = solver(Q.expand(n_train, *Q.shape), c_pred.squeeze(),
G.expand(n_train, *G.shape),
h.expand(n_train, *h.shape), torch.Tensor(),
torch.Tensor())
self.model_time += solver.Runtime()
loss = (x.squeeze() * c_true).mean()
optimizer.zero_grad()
loss.backward()
optimizer.step()
subepoch += 1
if i % 20 == 0:
if self.verbose:
dict_validation = {}
logging.info('Validation starts\n ')
#train_result = self.test(X,y)
if validation:
start_validation = time.time()
validation_result = self.test(
X_validation,
y_validation,
relaxation=validation_relax)
self.model_time += validation_result[4]
dict_validation[
'validation_regret'] = validation_result[0]
dict_validation[
'validation_mse'] = validation_result[1]
dict_validation[
'validation_accuracy'] = validation_result[2]
dict_validation[
'validation_loss'] = validation_result[3]
end_validation = time.time()
validation_time += end_validation - start_validation
if test:
start_test = time.time()
test_result = self.test(X_test,
y_test,
relaxation=test_relax)
self.model_time += validation_result[4]
dict_validation['test_regret'] = test_result[0]
dict_validation['test_mse'] = test_result[1]
dict_validation['test_accuracy'] = test_result[2]
dict_validation['test_loss'] = test_result[3]
end_test = time.time()
test_time += end_test - start_test
dict_validation['subepoch'] = subepoch
dict_validation['Runtime'] = self.model_time
dict_validation['time'] = time.time() - start
test_list.append(dict_validation)
logging.info(
"Epoch %d::subepoch %d Total time %d, validation time %d & test time %d"
% (e, i, time.time() - start, validation_time,
test_time))
#print('Epoch[{}/{}], loss(train):{:.2f} '.format(e+1, i, loss.item() ))
if self.plotting:
subepoch_list.append(subepoch)
reg, loss, acc = self.test(X, y)
loss_list.append(loss)
regret_list.append(reg)
accuracy_list.append(acc)
if self.plotting:
fig, (ax1, ax2, ax3) = plt.subplots(3, 1, figsize=(6, 6))
ax1.plot(subepoch_list, regret_list)
ax1.set_ylabel('Regret')
ax2.plot(subepoch_list, loss_list)
ax2.set_yscale('log')
ax2.set_ylabel('Loss')
ax3.plot(subepoch_list, accuracy_list)
ax3.set_xlabel('Sub Epochs')
ax3.set_ylabel('Accuracy')
plt.savefig(self.figname)
if self.verbose:
dd = defaultdict(list)
for d in test_list:
for key, value in d.items():
dd[key].append(value)
df = pd.DataFrame.from_dict(dd)
return df
def fit(self,
X,
y,
X_validation=None,
y_validation=None,
X_test=None,
y_test=None):
def make_model_matrix(nbMachines, nbTasks, nbResources, MC, U, D, E, L,
P, idle, up, down, q, **h):
# nbMachines: number of machine
# nbTasks: number of task
# nb resources: number of resources
# MC[m][r] resource capacity of machine m for resource r
# U[f][r] resource use of task f for resource r
# D[f] duration of tasks f
# E[f] earliest start of task f
# L[f] latest end of task f
# P[f] power use of tasks f
# idle[m] idle cost of server m
# up[m] startup cost of server m
# down[m] shut-down cost of server m
# q time resolution
# timelimit in seconds
Machines = range(nbMachines)
Tasks = range(nbTasks)
Resources = range(nbResources)
N = 1440 // q
### G and h
G = torch.zeros((nbMachines * N, nbTasks * nbMachines * N))
h = torch.zeros(nbMachines * N)
F = torch.zeros((N, nbTasks * nbMachines * N))
for m in Machines:
for t in range(N):
h[m * N + t] = MC[m][0]
for f in Tasks:
c_index = (f * nbMachines + m) * N
G[t + m * N,
(c_index + max(0, t - D[f] + 1)):(c_index +
(t + 1))] = 1
F[t, (c_index + max(0, t - D[f] + 1)):(c_index +
(t + 1))] = P[f]
### A and b
A1 = torch.zeros((nbTasks, nbTasks * nbMachines * N))
A2 = torch.zeros((nbTasks, nbTasks * nbMachines * N))
A3 = torch.zeros((nbTasks, nbTasks * nbMachines * N))
for f in Tasks:
A1[f, (f * N * nbMachines):((f + 1) * N * nbMachines)] = 1
A2[f, (f * N * nbMachines):(f * N * nbMachines + E[f])] = 1
A3[f, (f * N * nbMachines + L[f] - D[f] + 1):((f + 1) * N *
nbMachines)] = 1
b = torch.cat((torch.ones(nbTasks), torch.zeros(2 * nbTasks)))
A = torch.cat((A1, A2, A3))
return A, b, G, h, torch.transpose(F, 0, 1)
############################
logging.info('Model Training Starts\n')
tau = self.tau
start = time.time()
validation_time = 0
test_time = 0
# if validation true validation and tets data should be provided
X = X[:, 1:]
validation = (X_validation is not None) and (y_validation is not None)
test = (X_test is not None) and (y_test is not None)
if self.doScale:
self.scaler = preprocessing.StandardScaler().fit(X)
X = self.scaler.transform(X)
if validation:
start_validation = time.time()
X_validation = X_validation[:, 1:]
if self.doScale:
X_validation = self.scaler.transform(X_validation)
end_validation = time.time()
validation_time += end_validation - start_validation
if test:
start_test = time.time()
X_test = X_test[:, 1:]
if self.doScale:
X_test = self.scaler.transform(X_test)
end_test = time.time()
test_time += end_test - start_test
n_items = self.n_items
epochs = self.epochs
net = self.net
param = self.param
hyperparams = self.hyperparams
validation_relax = self.validation_relax
test_relax = self.test_relax
if validation:
start_validation = time.time()
solver_validation = Gurobi_ICON(relax=validation_relax,
reset=True,
presolve=True,
**param)
solver_validation.make_model()
self.solver_validation = solver_validation
end_validation = time.time()
validation_time += end_validation - start_validation
if test:
start_test = time.time()
solver_test = Gurobi_ICON(**param, reset=True, presolve=True)
solver_test.make_model()
self.solver_test = solver_test
end_test = time.time()
test_time += end_test - start_test
#sol_train = self.solution_func(y)
if self.validation:
if validation:
sol_validation = self.solution_func(
y_validation, solver=self.solver_validation)
if test:
sol_test = self.solution_func(y_test, solver=self.solver_test)
A, b, G, h, F = make_model_matrix(**param)
#Q= torch.diagflat(torch.ones(F.shape[0])/tau)
#print(F.shape)
Q = torch.eye(F.shape[0]) / tau
self.Q = Q
self.G = G
self.h = h
self.A = A
self.b = b
self.F = F
model = net(X.shape[1], 1)
#optimizer = torch.optim.Adam(model.parameters(),**hyperparams)
optimizer = self.optimizer(model.parameters(), **hyperparams)
model_params_quad = make_gurobi_model(G.detach().numpy(),
h.detach().numpy(),
A.detach().numpy(),
b.detach().numpy(),
Q.detach().numpy())
self.gurobi_model = model_params_quad
n_knapsacks = X.shape[0] // n_items
loss_list = []
regret_list = []
subepoch = 0
logger = []
test_list = []
n_train = 1
for e in range(epochs):
logging.info('Epoch %d' % e)
subepoch_list = [j for j in range(n_knapsacks)]
random.shuffle(subepoch_list)
for i in range(n_knapsacks):
n_start = n_items * subepoch_list[i]
n_stop = n_start + n_items
c_true = torch.mm(
F,
torch.tensor(y[n_start:n_stop],
dtype=torch.float).unsqueeze(1))
X_tensor = torch.tensor(X[n_start:n_stop, :],
dtype=torch.float)
c_pred = (model(X_tensor))
c_pred = torch.mm(F, model(X_tensor))
#logging.info('Call to qp function starts' )
solver = QPFunction(verbose=False,
solver=QPSolvers.GUROBI,
model_params=model_params_quad)
x = solver(Q.expand(n_train, *Q.shape), c_pred.squeeze(),
G.expand(n_train, *G.shape),
h.expand(n_train, *h.shape),
A.expand(n_train, *A.shape),
b.expand(n_train, *b.shape))
#logging.info('Call to qp function ends' )
self.model_time += solver.Runtime()
loss = (x.squeeze() * c_true.squeeze()).mean()
optimizer.zero_grad()
#print(loss)
loss.backward()
optimizer.step()
self.model = model
subepoch += 1
if self.model_save:
if i % 10 == 0:
logging.info("Model saving:%d-%d\n " % (e, i))
torch.save(
self.model.state_dict(),
str(self.model_name + "_Epoch" + str(e) + "_" +
str(i) + ".pth"))
if i % 50 == 0:
if self.verbose:
#train_result = self.test(X,y,sol_train)
dict_validation = {}
logging.info('Validation starts\n ')
if validation:
start_validation = time.time()
validation_result = self.test(
X_validation,
y_validation,
sol_validation,
solver=self.solver_validation)
logging.info('Validation on test data starts\n ')
self.model_time += validation_result[3]
dict_validation[
'validation_regret'] = validation_result[0]
dict_validation[
'validation_mse'] = validation_result[1]
dict_validation[
'validation_loss'] = validation_result[2]
end_validation = time.time()
validation_time += end_validation - start_validation
if test:
start_test = time.time()
test_result = self.test(X_test,
y_test,
sol_test,
solver=self.solver_test)
logging.info('Validation Ends \n ')
dict_validation['test_regret'] = test_result[0]
dict_validation['test_mse'] = test_result[1]
dict_validation['test_loss'] = test_result[2]
end_test = time.time()
test_time += end_test - start_test
dict_validation['subepoch'] = subepoch
dict_validation['Runtime'] = self.model_time
dict_validation['time'] = time.time() - start
test_list.append(dict_validation)
logging.info(
"Epoch %d::subepoch %d Total time %d, validation time %d & test time %d"
% (e, i, time.time() - start, validation_time,
test_time))
print('Epoch[%d::%d], loss(train):%.2f at %s' %
(e + 1, i, loss, datetime.datetime.now()))
if self.verbose:
dd = defaultdict(list)
for d in test_list:
for key, value in d.items():
dd[key].append(value)
df = pd.DataFrame.from_dict(dd)
return df