def main():
train_data, train_labels = load_data('spam_train.csv', start=1)
test_data, test_labels = load_data('spam_test.csv', start=1)
# train and test without normalization
print "without normalization"
for k in K:
predicted = predict(test_data, train_data, train_labels, k)
error = mse(test_labels, predicted)
print "k=%d accr=%.3f" % (k, 1.0 - error)
print
# train and test with normalization
print "with normalization"
train_data = zscore(train_data)
test_data = zscore(test_data)
for k in K:
predicted = predict(test_data, train_data, train_labels, k)
error = mse(test_labels, predicted)
print "k=%d accr=%.3f" % (k, 1.0 - error)
print
# report labels for first 50 data points
print "labels for first 50 data points"
test_data = test_data[:NUM_SAMPLES]
predicted = np.zeros((len(K), NUM_SAMPLES))
for i, k in enumerate(K):
predicted[i] = predict(test_data, train_data, train_labels, k)
print ' k= %s' % ' '.join(['%3d' % k for k in K])
for i in range(NUM_SAMPLES):
labels = ['yes' if y == 1 else 'no ' for y in predicted[:, i]]
print '%2d: %s' % (i + 1, ' '.join(labels))
def match(ref_block, frame, last_match, R=8, trans=None):
ref = (trans(ref_block)) if trans else ref_block
last_match_top, last_match_left = last_match
min_mad = float('inf')
arg_min_mad = None
mse_ = 0
for i in range(-R, R + 1):
if last_match_top + i < 0 \
or last_match_top + i + BLOCK_SIZE >= frame.shape[0]:
continue
for j in range(-R, R + 1):
if last_match_left + j < 0 \
or last_match_left + j + BLOCK_SIZE >= frame.shape[1]:
continue
block = frame[last_match_top + i:last_match_top + i + BLOCK_SIZE,
last_match_left + j:last_match_left + j + BLOCK_SIZE]
if trans:
m = mad(ref, trans(block))
else:
m = mad(ref_block, block)
if m < min_mad:
min_mad = m
arg_min_mad = (last_match_top + i, last_match_left + j)
mse_ = mse(ref_block, block)
return arg_min_mad, mse_
def curve(train_data, train_labels, test_data, test_labels, lagrange):
avg_errors = np.zeros(train_data.shape[DATA_AXIS])
for trial in range(10):
indices = [i for i in range(train_data.shape[DATA_AXIS])]
shuffle(indices)
for num_samples in range(1, train_data.shape[DATA_AXIS] + 1):
data = train_data[indices[:num_samples]]
labels = train_labels[indices[:num_samples]]
coefs = regress(data, labels, lagrange)
predicted = predict(test_data, coefs)
avg_errors[num_samples - 1] += mse(test_labels, predicted)
return avg_errors / 10
def get_polynomial_log_likelihood(self, ys, tree):
"""Minus ABC distance instead of log p(ys | tree, xs) where xs is
torch.linspace(-10, 10, 100). ABC distance is log(1 + mse).
Args:
ys: torch.tensor of shape [100]
tree: list of lists or string
Returns: -log(1 + mse(ys, eval(tree))); scalar tensor
"""
return -torch.log(1 +
util.mse(ys, util.eval_polynomial(tree, self.xs)))
def plot_regression(name, train_file, test_file):
stdout.write("Drawing plot for data set '%s'... " % name)
stdout.flush()
train_data, train_labels = load_data(train_file, dummy=1.0)
test_data, test_labels = load_data(test_file, dummy=1.0)
lagranges = [lagrange for lagrange in range(151)]
train_errors = []
test_errors = []
log = open('logs/q1/%s.log' % name, 'w')
# for each lagrange regress and calculate error
for lagrange in lagranges:
coefs = regress(train_data, train_labels, lagrange)
predicted = predict(train_data, coefs)
train_error = mse(train_labels, predicted)
train_errors.append(train_error)
predicted = predict(test_data, coefs)
test_error = mse(test_labels, predicted)
test_errors.append(test_error)
message = 'lagrange=%d train_error=%.3f test_error=%.3f\n'
log.write(message % (lagrange, train_error, test_error))
# plot errors as a function of the lagrange
pyplot.figure()
pyplot.xlim(0, 150)
pyplot.title("Data set '%s'" % name)
pyplot.xlabel('Lagrange multiplier')
pyplot.ylabel('Mean squared error')
pyplot.plot(lagranges, train_errors, label="Training")
pyplot.plot(lagranges, test_errors, label="Testing")
pyplot.legend(loc='lower right')
pyplot.savefig('plots/q1/%s.png' % name)
stdout.write("done.\n")
stdout.write("Plot image written to 'plots/q1/%s.png'.\n" % name)
stdout.write("Plot data written to '%s'.\n" % log.name)
stdout.flush()
log.close()
def test_one_epoch(self, loader, epoch):
self.G.eval()
self.D.eval()
test_loss = 0.0
num_examples = 0
imgs = []
pred_labels = []
labels = []
for data in tqdm(loader):
img, label = data
img = img.to(self.device)
label = label.to(self.device)
pred_label = self.predict(img)
loss = self.criterion(pred_label, label)
batch_size = img.size(0)
test_loss += loss.item() * batch_size
num_examples += batch_size
imgs.append(img.cpu().numpy())
labels.append(label.cpu().numpy())
pred_labels.append(pred_label.detach().cpu().numpy())
img = np.concatenate(imgs, axis=0)
label = np.concatenate(labels, axis=0)
pred_label = np.concatenate(pred_labels, axis=0)
log = {
'loss': test_loss / num_examples,
'img': img,
'label': label,
'pred_label': pred_label,
'pp_r2': pp_r2(pred_label, label),
'mse': mse(pred_label, label),
'rmse': rmse(pred_label, label),
'mae': mae(pred_label, label),
'pp_mse': pp_mse(pred_label, label).tolist(),
'pp_rmse': pp_rmse(pred_label, label).tolist(),
'pp_mae': pp_mae(pred_label, label).tolist(),
}
log['avg_r2'] = np.mean(log['pp_r2'])
self.logger.write(log, epoch=epoch, stage='test')
if test_loss < self.best_test_loss:
self.best_test_loss = test_loss
self.save(os.path.join(self.exp_path, 'models', 'model.best.t7'))
return log
def cross_validate(name, file):
data, labels = load_data(file, dummy=1.0)
log = open('logs/q3/%s.log' % name, 'w')
stdout.write("Evaluating data set '%s'..." % name)
stdout.flush()
# split the data into folds
indices = [i for i in range(data.shape[DATA_AXIS])]
shuffle(indices)
fold_size = ceil(float(data.shape[DATA_AXIS]) / NUM_FOLDS)
# evaluate each lagrange
best_error = maxint
for lagrange in range(0, 151):
avg_error = 0.0
# try each fold average errors
for i in range(NUM_FOLDS):
low = int(i * fold_size)
high = int((i + 1) * fold_size)
train_indices = indices[:low] + indices[high:]
test_indices = indices[low:high]
coefs = regress(data[train_indices], labels[train_indices],
lagrange)
predicted = predict(data[test_indices], coefs)
error = mse(labels[test_indices], predicted)
avg_error += error / NUM_FOLDS
message = 'lagrange=%d fold=%d error=%.3f\n'
log.write(message % (lagrange, i, error))
# update best error and lagrange if result is better
if avg_error < best_error:
best_error = avg_error
best_lagrange = lagrange
# report the results
stdout.write('done.\n')
stdout.write('Best Lagrange value is %d.\n' % best_lagrange)
stdout.write('Best error is %.3f.\n' % best_error)
stdout.write("Logs written to '%s'.\n" % log.name)
stdout.flush()
log.close()
def train_one_epoch(self, loader, epoch):
self.G.train()
self.scheduler.step()
train_loss = 0.0
num_examples = 0
pred_labels = []
labels = []
for data in tqdm(loader):
img, label = data
img = img.to(self.device)
label = label.to(self.device)
self.opt.zero_grad()
pred_label = self.predict(img)
pred_labels.append(pred_label.detach().cpu().numpy())
labels.append(label.cpu().numpy())
loss = self.criterion(pred_label*self.param_scale, label*self.param_scale) \
+ torch.sum((self.psf*(pred_label-label))**2)
loss.backward()
self.opt.step()
batch_size = img.size(0)
train_loss += loss.item() * batch_size
num_examples += batch_size
pred_label = np.concatenate(pred_labels, axis=0)
label = np.concatenate(labels, axis=0)
log = {
'loss': train_loss / num_examples,
'pp_r2': pp_r2(pred_label, label),
'mse': mse(pred_label, label),
'rmse': rmse(pred_label, label),
'mae': mae(pred_label, label),
'pp_mse': pp_mse(pred_label, label).tolist(),
'pp_rmse': pp_rmse(pred_label, label).tolist(),
'pp_mae': pp_mae(pred_label, label).tolist(),
}
log['avg_r2'] = np.mean(log['pp_r2'])
self.logger.write(log, epoch=epoch)
self.save(os.path.join(self.exp_path, 'models', 'model.%d.t7' % epoch))
return log
def cv(X, y, c=1, wts=None, nfolds=10):
"""
Runs nfold cross-validation on the input data set. Uses ridge regression
as the training algorithm.
Parameters
----------
X: array of shape = [n_samples, n_features]
Input examples
y: array of shape = [n_samples]
labels
c: positive real number
regularization parameter
wts: array of shape = [n_samples]
example weights
nfolds: scalar
no. of folds in cross-validation
Returns
-------
average mean squared error (cross-validation error)
"""
kf = StratifiedKFold(sign(y), n_folds=nfolds)
err = []
for tr_ids, te_ids in kf:
model = Ridge(alpha=c)
if wts is not None:
model.fit(X[tr_ids], y[tr_ids], sample_weight=wts[tr_ids])
else:
model.fit(X[tr_ids], y[tr_ids])
f = model.predict(X[te_ids])
err.append(util.mse(f, y[te_ids]))
return mean(err)
