def batch_train(batch_sz, epoch, X_train_t, Y_train_t, model, opt, cost_fn):
train_cost = 0
batch_data_, batch_targets_ = \
batch.get_vars(batch_sz, X_train_t, Y_train_t)
size = batch_sz
for i in range(0, X_train_t.size(0), batch_sz):
# Deal with potentially incomplete (last) batch
if i + batch_sz > X_train_t.size(0):
size = X_train_t.size(0) - i
batch_data_, batch_targets_ = \
batch.get_vars(size, X_train_t, Y_train_t)
batch_data_.data[:] = X_train_t[i:i+size]
batch_targets_.data[:] = Y_train_t[i:i+size]
opt.zero_grad()
preds = model(batch_data_)
batch_cost = cost_fn(preds, batch_targets_)
batch_cost.backward()
opt.step()
## Keep running average of loss
train_cost += (batch_cost - train_cost) * size / (i + size)
print('Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.4f}'.format(
epoch, i+size, X_train_t.size(0),
float(i+size)/X_train_t.size(0)*100,
batch_cost.data[0]))
return train_cost
def batch_train(batch_sz, epoch, X_train_t, Y_train_t, Y_train_int_t,
model, cost_fn_news, nll, opt, lam):
train_cost_agg = 0
train_nll_agg = 0
batch_data_, batch_targets_ = \
batch.get_vars(batch_sz, X_train_t, Y_train_t)
_, batch_targets_int_ = \
batch.get_vars_scalar_out(batch_sz, X_train_t, Y_train_int_t)
size = batch_sz
for i in range(0, X_train_t.size(0), batch_sz):
# Deal with potentially incomplete (last) batch
if i + batch_sz > X_train_t.size(0):
size = X_train_t.size(0) - i
batch_data_, batch_targets_ = batch.get_vars(
size, X_train_t, Y_train_t)
_, batch_targets_int_ = batch.get_vars_scalar_out(
size, X_train_t, Y_train_int_t)
batch_data_.data[:] = X_train_t[i:i+size]
batch_targets_.data[:] = Y_train_t[i:i+size]
batch_targets_int_.data[:] = Y_train_int_t[i:i+size]
opt.zero_grad()
preds = model(batch_data_)
train_cost = cost_fn_news(preds, batch_targets_)
train_nll = nll(preds, batch_targets_int_)
(train_cost + lam * train_nll).backward()
opt.step()
# Keep running average of losses
train_cost_agg += \
(train_cost - train_cost_agg) * batch_sz / (i + batch_sz)
train_nll_agg += \
(train_nll - train_nll_agg) * batch_sz / (i + batch_sz)
print('Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.4f}'.format(
epoch, i+batch_sz, X_train_t.size(0),
float(i+batch_sz)/X_train_t.size(0)*100,
train_cost.data[0]))
return train_cost_agg, train_nll_agg