def MBAP(pruning_rate_per_layer, is_first, is_last):
_, mask_1 = utils.prune(L1, pruning_rate_per_layer)
L1_ap = tf.cond(is_first, lambda: L1 * mask_1, lambda: L1)
L2_ap = tf.nn.relu(tf.matmul(L1_ap, W2) + B2)
_, mask_2 = utils.prune(L2_ap, pruning_rate_per_layer)
pruned_L2_ap = tf.cond(is_last, lambda: L2_ap * mask_2, lambda: L2_ap)
model_ap = tf.matmul(pruned_L2_ap, W3) + B3
return model_ap
def add_event():
if request.method == 'GET':
return render_template('forms/add_event.html', event=None)
form = request.form.copy()
utils.prune(form)
utils.format_time(form, 'start')
utils.format_time(form, 'end')
events.add(form)
return redirect(url_for('hub'))
def edit_event():
if request.method == 'GET':
ID = request.args.get('ID')
event = events.get(ID)
utils.reverse_format_time(event, 'start')
utils.reverse_format_time(event, 'end')
return render_template('forms/add_event.html', event=event)
form = request.form.copy()
utils.prune(form)
utils.format_time(form, 'start')
utils.format_time(form, 'end')
events.update(form['ID'], form)
return redirect(url_for('admin'))
def MBFD(pruning_rate_per_layer):
#L1_mb, _ = utils.prune(L1, pruning_rate_per_layer)
L1_mb = L1
L2_mb_prev = tf.matmul(L1_mb, W2)
pruned_L2_mb_prev, _ = utils.prune(L2_mb_prev, pruning_rate_per_layer)
L2_mb = tf.nn.relu(pruned_L2_mb_prev + B2)
model_mb = tf.matmul(L2_mb, W3) + B3
return model_mb
def MFD(pruning_rate_per_layer):
adv_feat_mf_1, adv_feat_mf_2 = compare(is_grad_compare)
#L1_mf = L1 * utils.mask_vec(adjusted_feat_1, pruning_rate_per_layer)
L1_mf = L1
L2_mf_prev = tf.matmul(L1_mf, W2)
_, mask = utils.prune(L2_mf_prev / adv_feat_mf_2, pruning_rate_per_layer)
L2_mf_prev = L2_mf_prev * mask
L2_mf = tf.nn.relu(L2_mf_prev + B2)
model_mf = tf.matmul(L2_mf, W3) + B3
return model_mf
def MBAP(pruning_rate_per_layer, is_first):
_, mask_1 = utils.prune_conv_feature(h_conv1, pruning_rate_per_layer)
h_conv1_ap = tf.cond(is_first, lambda: h_conv1 * mask_1, lambda: h_conv1)
h_pool1_ap = tf.nn.max_pool(h_conv1_ap,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME')
h_conv2_ap_prev = tf.nn.conv2d(h_pool1_ap,
W_conv2,
strides=[1, 1, 1, 1],
padding='SAME')
h_conv2_ap = tf.nn.relu(h_conv2_ap_prev + b_conv2)
h_conv2_ap, _ = utils.prune_conv_feature(h_conv2_ap,
pruning_rate_per_layer)
h_pool2_ap = tf.nn.max_pool(h_conv2_ap,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME')
h_conv3_ap_prev = tf.nn.conv2d(h_pool2_ap,
W_conv3,
strides=[1, 1, 1, 1],
padding='SAME')
h_conv3_ap = tf.nn.relu(h_conv3_ap_prev + b_conv3)
h_conv3_ap, _ = utils.prune_conv_feature(h_conv3_ap,
pruning_rate_per_layer)
h_conv4_ap_prev = tf.nn.conv2d(h_conv3_ap,
W_conv4,
strides=[1, 1, 1, 1],
padding='SAME')
h_conv4_ap = tf.nn.relu(h_conv4_ap_prev + b_conv4)
h_conv4_ap, _ = utils.prune_conv_feature(h_conv4_ap,
pruning_rate_per_layer)
h_conv5_ap_prev = tf.nn.conv2d(h_conv4_ap,
W_conv5,
strides=[1, 1, 1, 1],
padding='SAME')
h_conv5_ap = tf.nn.relu(h_conv5_ap_prev + b_conv5)
h_conv5_ap, _ = utils.prune_conv_feature(h_conv5_ap,
pruning_rate_per_layer)
h_conv5_ap_flat = tf.reshape(h_conv5_ap, [-1, 8 * 8 * 128])
h_fc1_ap_prev = tf.matmul(h_conv5_ap_flat, W_fc1)
h_fc1_ap = tf.nn.relu(h_fc1_ap_prev + b_fc1)
h_fc1_ap, _ = utils.prune(h_fc1_ap, pruning_rate_per_layer)
model_ap = tf.matmul(h_fc1_ap, W_fc2) + b_fc2
return model_ap
def prune(splited):
# Prune split with empty sentence
old2new = dict()
new_splited = list()
for (idx, span) in enumerate(splited):
span['text'] = utils.prune(span['text'])
if span['text'] == '':
continue
old2new[idx] = len(new_splited)
new_splited.append(span)
# Update ifobj after prune
for item in new_splited:
if item['ifobj'] is not None:
item['ifobj'] = [old2new[obj] for obj in item['ifobj'] if obj in old2new]
return new_splited
def __init__(self,
base_dir,
batch_size,
mode=1,
cls=1,
prune=None,
de_norm=False):
TRAIN = 1
TEST = 2
self.base_dir = base_dir
self.batch_size = batch_size
ds_dir = os.path.join(self.base_dir, 'dataset/class_{}'.format(cls))
if mode == TRAIN:
self.x = utils.pickle_load(ds_dir + '/imgs_train.pkl')
self.y = utils.pickle_load(ds_dir + '/marks_train.pkl')
elif mode == TEST:
self.x = utils.pickle_load(ds_dir + '/imgs_test.pkl')
self.y = utils.pickle_load(ds_dir + '/marks_test.pkl')
else:
raise ("Invalid option, should be one {} or {}".format(
TRAIN, TEST))
if de_norm:
self.x = utils.de_norm(self.x)
self.y = utils.de_norm(self.y)
self.labels = np.array(
[1 if np.sum(mask) > 0 else 0 for mask in self.y])
if prune is not None:
self.x, self.y, self.labels = utils.prune(self.x, self.y,
self.labels, prune)
self.x = utils.norm(self.x)
self.y = utils.norm(self.y)
self.classes = np.unique(self.labels)
self.per_class_ids = {}
ids = np.array(range(len(self.x)))
for c in self.classes:
self.per_class_ids[c] = ids[self.labels == c]
print(Counter(self.labels))
def plot_jct_cdf(logfile_paths,
labels,
v100s,
p100s,
k80s,
max_input_job_rate,
policies,
min_job_id,
max_job_id,
partition=True,
finish_time_fairness=False,
output_directory=None):
from utils import get_jcts, prune
lambdas = list(set([x[5] for x in logfile_paths]))
lambdas.sort(reverse=True)
print(policies)
for l in lambdas:
handles_in_legend = []
labels_in_legend = []
input_job_rate = 3600.0 / l
if input_job_rate > max_input_job_rate:
continue
print("Input job rate: %.2f" % input_job_rate)
plt.figure(figsize=(8, 3))
if partition:
axes = [
plt.subplot2grid((1, 2), (0, 0), rowspan=1),
plt.subplot2grid((1, 2), (0, 1), rowspan=1),
]
titles = ["Short jobs", "Long jobs"]
else:
axes = [
plt.subplot2grid((1, 1), (0, 0), rowspan=1),
]
titles = [None]
if not finish_time_fairness:
axes.append(axes[0].inset_axes([0.4, 0.2, 0.5, 0.6]))
titles.append(None)
if finish_time_fairness:
relevant_logfile_paths = list(
reversed(
prune(logfile_paths,
v100s,
p100s,
k80s,
"isolated",
seed=0)))
relevant_logfile_paths = [
x for x in relevant_logfile_paths if x[0] == l
]
if len(relevant_logfile_paths) != 1:
continue
isolated_jcts = get_jcts(relevant_logfile_paths[0][1],
seed=0,
min_job_id=min_job_id,
max_job_id=max_job_id)
isolated_jcts.sort(key=lambda x: x[1])
linestyles = ['--', '-.', ':', '--', '-.']
for i, policy in enumerate(policies):
relevant_logfile_paths = list(
reversed(
prune(logfile_paths, v100s, p100s, k80s, policy, seed=0)))
relevant_logfile_paths = [
x for x in relevant_logfile_paths if x[0] == l
]
if len(relevant_logfile_paths) != 1:
continue
jcts = get_jcts(relevant_logfile_paths[0][1],
seed=0,
min_job_id=min_job_id,
max_job_id=max_job_id)
jcts.sort(key=lambda x: x[1])
partition_point = int(len(jcts) * 0.8)
if finish_time_fairness:
jcts = [x[0] / y[0] for (x, y) in zip(jcts, isolated_jcts)]
else:
jcts = [x[0] for x in jcts]
print("%s: %.2f" % (policy, np.mean(jcts)))
if partition:
jcts = np.split(np.array(jcts), [partition_point])
else:
jcts = [np.array(jcts), np.array(jcts)]
for j, (ax, jcts_segment) in enumerate(zip(axes, jcts)):
jcts_segment.sort()
percentiles = [(k + 1) / len(jcts_segment)
for k in range(len(jcts_segment))]
if "Gavel" in labels[policy]:
handle = ax.plot(jcts_segment,
percentiles,
color=current_palette[i],
linestyle='-',
linewidth=3)
else:
handle = ax.plot(jcts_segment,
percentiles,
color=current_palette[i],
linestyle=linestyles[i])
if j == 0:
handles_in_legend.append(handle[0])
labels_in_legend.append(labels[policy])
for i, (ax, title) in enumerate(zip(axes, titles)):
if finish_time_fairness:
if partition:
ax.set_xlabel("FTF" + "\n" + title)
else:
ax.set_xlabel("FTF")
ax.set_xlim([0, 4])
ax.set_xticks([0, 1, 2, 3, 4])
ax.set_yticks([0, 0.2, 0.4, 0.6, 0.8, 1.0])
else:
if partition:
ax.set_xlabel("JCT (hrs)" + "\n" + title)
else:
if i == 0:
ax.set_xlabel("JCT (hrs)")
if not partition:
if i == 0:
ax.set_xlim([0, 500])
ax.set_yticks([0, 0.2, 0.4, 0.6, 0.8, 1.0])
else:
ax.set_xlim([0, 25])
ax.set_xticks([0, 5, 10, 15, 20, 25])
ax.set_yticks([0, 0.33, 0.67, 1.0])
if i == 0:
ax.set_ylabel("Fraction of jobs")
if partition:
if i > 0:
ax.set_yticklabels(["", "", "", "", "", ""])
ax.set_ylim([0, 1.0])
sns.despine()
if not partition and not finish_time_fairness:
axes[0].indicate_inset_zoom(axes[1], linewidth=3)
leg = plt.figlegend(handles=handles_in_legend,
labels=labels_in_legend,
ncol=3,
frameon=False,
loc='upper center')
bb = leg.get_bbox_to_anchor().inverse_transformed(axes[0].transAxes)
bb.y0 += 0.22
bb.y1 += 0.22
leg.set_bbox_to_anchor(bb, transform=axes[0].transAxes)
if output_directory is not None:
output_filename = os.path.join(
output_directory,
"input_job_rate=%d.pdf" % (input_job_rate * 10))
with PdfPages(output_filename) as pdf:
pdf.savefig(bbox_inches='tight')
plt.show()
def plot_metric_vs_inverse_lambda_different_metric_fns(logfile_paths,
labels,
v100s,
p100s,
k80s,
policies,
metric_fns,
metric_fn_labels,
metric_label,
xmin=0,
xmax=None,
ymin=0,
ymax=None,
output_filename=None,
verbose=False):
from utils import prune
plt.figure(figsize=(8, 3))
ax = plt.subplot2grid((1, 1), (0, 0), colspan=1)
data = {"input_job_rate": [], "metric": [], "seed": [], "policy": []}
print(policies)
for policy in policies:
relevant_logfile_paths = list(
reversed(prune(logfile_paths, v100s, p100s, k80s, policy)))
for metric_fn_label, metric_fn in zip(metric_fn_labels, metric_fns):
lambdas = [x[0] for x in relevant_logfile_paths]
input_job_rates = [3600.0 / x for x in lambdas]
metrics = [metric_fn(x[1]) for x in relevant_logfile_paths]
seeds = [x[2] for x in relevant_logfile_paths]
policies = [
labels[policy] + " (%s)" % metric_fn_label
for i in range(len(metrics))
]
import pandas as pd
data["input_job_rate"] += input_job_rates
data["metric"] += metrics
data["seed"] += seeds
data["policy"] += policies
if verbose:
import pandas as pd
df = pd.DataFrame(data)
print(df.groupby(["policy", "input_job_rate"]).mean())
sns.lineplot(x='input_job_rate',
y='metric',
style='policy',
hue='policy',
data=data,
ci='sd',
markers=True)
ax.set_xlabel("Input job rate (jobs/hr)")
ax.set_ylabel(metric_label)
ax.set_xlim([xmin, xmax])
ax.set_ylim([ymin, ymax])
sns.despine()
leg = plt.legend(frameon=False)
bb = leg.get_bbox_to_anchor().inverse_transformed(ax.transAxes)
bb.y0 += 0.22
bb.y1 += 0.22
leg.set_bbox_to_anchor(bb, transform=ax.transAxes)
if output_filename is not None:
with PdfPages(output_filename) as pdf:
pdf.savefig(bbox_inches='tight')
plt.show()
def plot_metric_vs_inverse_lambda_different_mechanisms(all_logfile_paths,
labels,
label_modifiers,
v100s,
p100s,
k80s,
policies,
metric_fn,
metric_label,
xmax=None,
ymax=None,
output_filename=None,
extrapolate=False):
from utils import prune
plt.figure(figsize=(4.5, 3))
ax = plt.subplot2grid((1, 1), (0, 0), colspan=1)
data = {"input_job_rate": [], "metric": [], "seed": [], "policy": []}
print(policies)
for policy in policies:
for logfile_paths, label_modifier in zip(all_logfile_paths,
label_modifiers):
relevant_logfile_paths = list(
reversed(prune(logfile_paths, v100s, p100s, k80s, policy)))
label = labels[policy] + label_modifier
lambdas = []
metrics = []
seeds = []
for x in relevant_logfile_paths:
metric = metric_fn(x[1])
if metric is not None:
lambdas.append(x[0])
metrics.append(metric)
seeds.append(x[2])
input_job_rates = [3600.0 / x for x in lambdas]
policies = [label for i in range(len(metrics))]
data["input_job_rate"] += input_job_rates
data["metric"] += metrics
data["seed"] += seeds
data["policy"] += policies
if len(input_job_rates) > 0 and extrapolate:
data["input_job_rate"] += [max(input_job_rates) + 0.2]
data["metric"] += [105.0]
data["seed"] += [0]
data["policy"] += [label]
sns.lineplot(x='input_job_rate',
y='metric',
style='policy',
hue='policy',
data=data,
ci='sd',
markers=True)
ax.set_xlabel("Input job rate (jobs/hr)")
ax.set_ylabel(metric_label)
ax.set_xlim([0, xmax])
ax.set_ylim([0, ymax])
sns.despine()
leg = plt.legend(loc='upper left', frameon=False)
bb = leg.get_bbox_to_anchor().inverse_transformed(ax.transAxes)
bb.y0 += 0.1
bb.y1 += 0.1
leg.set_bbox_to_anchor(bb, transform=ax.transAxes)
if output_filename is not None:
with PdfPages(output_filename) as pdf:
pdf.savefig(bbox_inches='tight')
plt.show()
def pruning(self, num_keep):
with torch.no_grad():
self.mask_normal = prune(self.alphas_normal, num_keep,
self.mask_normal)
self.mask_reduce = prune(self.alphas_reduce, num_keep,
self.mask_reduce)
def pruning(self, num_keep): self._mask = prune(self._arch_parameters, num_keep, self._mask)
def main(data, target, args):
# Names for various stuff
model_name = 'model_{0}_{1}_{2}.lp'.format(args.data, args.kernel, args.function)
param_name = 'model_{0}_{1}_{2}.prm'.format(args.data, args.kernel, args.function)
solution_name = 'solution_{0}_{1}_{2}.sol'.format(args.data, args.kernel, args.function)
ultrametric_name = 'ultrametric_{0}_{1}_{2}'.format(args.data, args.kernel, args.function)
var_name = 'var_{0}_{1}_{2}_{3}.pkl'.format(args.data, args.data, args.kernel, args.function)
obj_name = 'obj_{0}_{1}_{2}_{3}.pkl'.format(args.data, args.data, args.kernel, args.function)
laminar_name = 'laminar_{0}_{1}_{2}.pkl'.format(args.data, args.kernel, args.function)
tree_name = 'lp_tree_{0}_{1}_{2}_{3}.pdf'.format(args.data, args.kernel, args.function, args.eps)
err_dict_name = 'error_{0}_{1}_{2}_{3}.txt'.format(args.data, args.kernel, args.function, args.eps)
# Test other hierarchical clustering algorithms
one_target = map(lambda x: x + 1, target)
k = args.prune
y = pdist(data, metric='euclidean')
Z = []
Z.append(hac.linkage(y, method='single'))
Z.append(hac.linkage(y, method='complete'))
Z.append(hac.linkage(y, method='average'))
ward = hac.linkage(data, method='ward')
Z.append(ward)
errors = []
while Z:
x = Z.pop(0)
pred = hac.fcluster(x, k, 'maxclust')
# print('pred = ', pred)
err = utils.error(list(pred), one_target)
errors.append(err)
# K means
clf = KMeans(k)
pred = clf.fit_predict(data)
pred = map(lambda x: x + 1, pred)
err = utils.error(list(pred), one_target)
errors.append(err)
# print('kmeans = ', pred)
error_dict = {'single linkage': errors[0], 'complete linkage': errors[1], 'average linkage': errors[2], 'ward': errors[3]}
error_dict['kmeans'] = errors[4]
print(error_dict)
# initialize model
if args.function == 'linear':
m = init_model(data, args.kernel, args.triangle, utils.linear)
if args.function == 'quadratic':
m = init_model(data, args.kernel, args.triangle, utils.quadratic)
if args.function == 'cubic':
m = init_model(data, args.kernel, args.triangle, utils.cubic)
if args.function == 'logarithm':
m = init_model(data, args.kernel, args.triangle, utils.logarithm)
if args.function == 'exponential':
m = init_model(data, args.kernel, args.triangle, utils.exponential)
m._n = data.shape[0]
# Check if reading solution from file
if args.solution:
print('Reading LP solution from ', args.solution)
solution_dict = read_solution(m, args.solution)
else:
start = time.time()
print('Optimizing over model')
m.optimize()
flag = args.triangle
while flag and time.time() - start < args.time:
print("Time_diff = {}".format(time.time() - start))
m.optimize()
# Feed solution to separation oracle
flag = separation_oracle(m, args.triangle)
end = time.time()
print('Total time to optimize = {0}'.format(end - start))
print('Writing solution to ', solution_name)
m.write(solution_name)
print('Saving model to ', model_name)
m.write(model_name)
solution_dict = get_solution_dict(solution_name)
# print('Triangle inequality satisfied: ', check_triangle_constraints(m))
# print('Spreading constraints satisfied: ', check_spreading_constraints(m))
# Get ultrametric from LP
print('Rounding LP')
if args.function == 'linear':
d = get_ultrametric_from_lp(m, solution_dict, args.eps, utils.linear)
utils.inverse_ultrametric(d, utils.inverse_linear)
elif args.function == 'quadratic':
d = get_ultrametric_from_lp(m, solution_dict, args.eps, utils.quadratic)
utils.inverse_ultrametric(d, utils.inverse_quadratic)
elif args.function == 'cubic':
d = get_ultrametric_from_lp(m, solution_dict, args.eps, utils.cubic)
utils.inverse_ultrametric(d, utils.inverse_cubic)
elif args.function == 'exponential':
d = get_ultrametric_from_lp(m, solution_dict, args.eps, utils.exponential)
utils.inverse_ultrametric(d, utils.inverse_exponential)
elif args.function == 'logarithm':
d = get_ultrametric_from_lp(m, solution_dict, args.eps, utils.logarithm)
utils.inverse_ultrametric(d, utils.inverse_logarithm)
print('d = ', d)
cost = utils.get_cost(m, d)
print('Cost of hierarchy: ', cost)
print('Check ultrametric: ', utils.check_ultrametric(d))
# print(d)
total_obj = utils.get_total(m)
print('Total objective = ', total_obj)
print('Scaled cost = ', cost/total_obj)
utils.complete_ultrametric(d)
print('Building laminar list')
L = utils.build_laminar_list(d)
# print('Laminar list = ', L)
print('Check laminar: ', utils.test_laminar(L))
labels = [1]*m._n
pruned = utils.prune(L, one_target, k, labels)
print('Error on pruning: ', pruned[0])
error_dict['lp rounding'] = pruned[0]
with open(err_dict_name, 'wb') as f:
f.write(str(error_dict))
# Build and draw the hierarchy
G = utils.build_hierarchy(d)
print('Drawing tree to ', tree_name)
utils.draw(G, target, m._n, tree_name)
def plot_metric_vs_inverse_lambda(logfile_paths,
labels,
v100s,
p100s,
k80s,
policies,
metric_fn,
metric_label,
xmax=None,
ymax=None,
output_filename=None,
extrapolate=False,
verbose=False):
from utils import prune
plt.figure(figsize=(8, 3))
ax = plt.subplot2grid((1, 1), (0, 0), colspan=1)
data = {"input_job_rate": [], "metric": [], "seed": [], "policy": []}
print(policies)
for policy in policies:
relevant_logfile_paths = list(
reversed(prune(logfile_paths, v100s, p100s, k80s, policy)))
lambdas = [x[0] for x in relevant_logfile_paths]
input_job_rates = [3600.0 / x for x in lambdas]
metrics = [metric_fn(x[1]) for x in relevant_logfile_paths]
seeds = [x[2] for x in relevant_logfile_paths]
policies = [labels[policy] for i in range(len(metrics))]
data["input_job_rate"] += input_job_rates
data["metric"] += metrics
data["seed"] += seeds
data["policy"] += policies
if len(input_job_rates) > 0 and extrapolate:
data["input_job_rate"] += [max(input_job_rates) + 0.4]
data["metric"] += [105.0]
data["seed"] += [0]
data["policy"] += [labels[policy]]
if verbose:
df = pd.DataFrame(data)
grouped_df = df.groupby(["policy", "input_job_rate", "seed"])
for name_of_the_group, group in grouped_df:
print(name_of_the_group)
print(group.mean())
sns.lineplot(x='input_job_rate',
y='metric',
style='policy',
hue='policy',
data=data,
ci='sd',
markers=True)
ax.set_xlabel("Input job rate (jobs/hr)")
ax.set_ylabel(metric_label)
ax.set_xlim([0, xmax])
ax.set_ylim([0, ymax])
sns.despine()
leg = plt.legend(loc='upper left', frameon=False)
bb = leg.get_bbox_to_anchor().inverse_transformed(ax.transAxes)
bb.y0 += 0.22
bb.y1 += 0.22
leg.set_bbox_to_anchor(bb, transform=ax.transAxes)
if output_filename is not None:
with PdfPages(output_filename) as pdf:
pdf.savefig(bbox_inches='tight')
plt.show()
def __init__(
self,
data_src,
batch_size=5,
dataset='MNIST',
rst=64,
prune_classes=None,
k_shot=5,
):
self.batch_size = batch_size
self.data_src = data_src
if dataset == 'chest':
if self.data_src == self.TEST:
x, y = utils.load_test_data(rst)
self.dataset_x = x
self.dataset_y = y
else:
x, y = utils.load_train_data(rst)
self.dataset_x = x
self.dataset_y = y
elif dataset == 'flowers':
x, y = utils.pickle_load(BASE_DIR +
'/dataset/flowers/imgs_labels.pkl')
to_train_classes = self.to_train_classes
to_test_classes = self.to_test_classes
if self.data_src == self.TEST:
to_keep = np.array(
[i for i, l in enumerate(y) if l in to_test_classes])
x, y = x[to_keep], y[to_keep]
self.dataset_x = x
# TODO start from 0
self.dataset_y = y
else:
to_keep = np.array(
[i for i, l in enumerate(y) if l in to_train_classes])
x, y = x[to_keep], y[to_keep]
self.dataset_x = x
# TODO start from 0
self.dataset_y = y
else: # multi chest
x, y = self._load_data(rst)
x = utils.denormalize(x)
to_train_classes = self.to_train_classes
to_test_classes = self.to_test_classes
if self.data_src == self.TEST:
to_keep = []
counter = {12: 0, 13: 0, 14: 0}
for i, l in enumerate(y):
if l not in to_train_classes and counter[l] < k_shot:
to_keep.append(i)
counter[l] += 1
to_keep = np.array(to_keep)
if len(to_keep) > 0:
x, y = x[to_keep], y[to_keep]
self.dataset_x = x
self.dataset_y = np.array([l for l in y])
else:
to_keep = np.array(
[i for i, l in enumerate(y) if l in to_train_classes])
x, y = x[to_keep], y[to_keep]
self.dataset_x = x
self.dataset_y = np.array([l for l in y])
# Normalize between -1 and 1
self.dataset_x = utils.normalize(self.dataset_x)
print(self.dataset_x.shape[0], self.dataset_y.shape[0])
assert (self.dataset_x.shape[0] == self.dataset_y.shape[0])
# Compute per class instance count.
classes = np.unique(self.dataset_y)
self.classes = classes
per_class_count = list()
for c in classes:
per_class_count.append(np.sum(np.array(self.dataset_y == c)))
if prune_classes:
self.dataset_x, self.dataset_y = utils.prune(
self.dataset_x, self.dataset_y, prune_classes)
# Recount after pruning
per_class_count = list()
for c in classes:
per_class_count.append(np.sum(np.array(self.dataset_y == c)))
self.per_class_count = per_class_count
# List of labels
self.label_table = [str(c) for c in range(len(self.classes))]
# Preload all the labels.
self.labels = self.dataset_y[:]
# per class ids
self.per_class_ids = dict()
ids = np.array(range(len(self.dataset_x)))
for c in classes:
self.per_class_ids[c] = ids[self.labels == c]
def main(data, target, args):
model_name = 'model_{0}_{1}_{2}.lp'.format(args.data, args.kernel,
args.function)
param_name = 'model_{0}_{1}_{2}.prm'.format(args.data, args.kernel,
args.function)
solution_name = 'solution_{0}_{1}_{2}.sol'.format(args.data, args.kernel,
args.function)
ultrametric_name = 'ultrametric_{0}_{1}_{2}'.format(
args.data, args.kernel, args.function)
var_name = 'var_{0}_{1}_{2}_{3}.pkl'.format(args.data, args.data,
args.kernel, args.function)
obj_name = 'obj_{0}_{1}_{2}_{3}.pkl'.format(args.data, args.data,
args.kernel, args.function)
laminar_name = 'laminar_{0}_{1}_{2}.pkl'.format(args.data, args.kernel,
args.function)
tree_name = 'ip_tree_{0}_{1}_{2}.pdf'.format(args.data, args.kernel,
args.function)
if args.kernel == 'cosine':
y = pdist(data, metric='cosine')
# Make condensed distance matrix into redundant form
similarity = 1 - y
similarity = squareform(similarity)
if args.kernel == 'gaussian':
y = pdist(data, metric='sqeuclidean')
s = 1
y = 1 - np.exp(-(y**2) / (2 * s**2))
# Make condensed distance matrix into redundant form
similarity = 1 - y
similarity = squareform(similarity)
if args.kernel == 'sqeuclidean':
y = pdist(data, metric='sqeuclidean')
similarity = -y
similarity = squareform(similarity)
if args.function == 'linear':
m = init_model(data, similarity, target, utils.linear)
elif args.function == 'quadratic':
m = init_model(data, similarity, target, utils.quadratic)
elif args.function == 'cubic':
m = init_model(data, similarity, target, utils.cubic)
elif args.function == 'exponential':
m = init_model(data, similarity, target, utils.exponential)
elif args.function == 'logarithm':
m = init_model(data, similarity, target, utils.logarithm)
else:
exit(0)
print('Saving model')
m.write(model_name)
# Use concurrent optimization
m.params.method = 3
# Limit memory
m.params.NodeFileStart = 10
# Limit number of threads
m.params.Threads = args.num_threads
# Set MIP Focus
m.params.MIPFocus = 3
# Tune parameters
print('Tuning parameters')
m.params.tuneResults = 1
m.tune()
if m.tuneResultCount > 0:
m.getTuneResult(0)
# Set MIP Gap
m.params.MIPGap = 0.01
print('Saving model parameters')
m.write(param_name)
print('Saving objective functions')
with open(obj_name, 'wb') as f:
pickle.dump(m._obj, f)
print('Optimizing over model')
m._n = data.shape[0]
m.optimize(callback_function)
if m.status == GRB.Status.OPTIMAL:
# Write solution
m.write(solution_name)
print('Check binary triangle for solution: ', check_binary_triangle(m))
# Get ultrametric
if args.function == 'linear':
d = get_ultrametric(m, utils.linear)
utils.inverse_ultrametric(d, utils.inverse_linear)
elif args.function == 'quadratic':
d = get_ultrametric(m, utils.quadratic)
utils.inverse_ultrametric(d, utils.inverse_quadratic)
elif args.function == 'cubic':
d = get_ultrametric(m, utils.cubic)
utils.inverse_ultrametric(d, utils.inverse_cubic)
elif args.function == 'exponential':
d = get_ultrametric(m, utils.exponential)
utils.inverse_ultrametric(d, utils.inverse_exponential)
elif args.function == 'logarithm':
d = get_ultrametric(m, utils.logarithm)
utils.inverse_ultrametric(d, utils.inverse_logarithm)
print('d = ', d)
print('Check ultrametric: ', utils.check_ultrametric(d))
cost = utils.get_cost(m, d)
print('Cost of hierarchy = ', cost)
total_obj = utils.get_total(m)
print('Total cost = ', total_obj)
print('Scaled cost = ', cost / total_obj)
# Complete ultrametric
utils.complete_ultrametric(d)
# Build laminar list from d
print('building laminar list')
L = utils.build_laminar_list(d)
print('L = ', L)
print('Check laminar: ', utils.test_laminar(L))
labels = [1] * m._n
one_target = map(lambda x: x + 1, target)
# Prune laminar list
pruned = utils.prune(L, one_target, args.prune, labels)
print('Error on pruning: ', pruned[0])
with open(ultrametric_name, 'wb') as f:
pickle.dump(d, f)
# Build hierarchy
print('Building hierarchy')
G = utils.build_hierarchy(d)
# Draw hierarchy
print('Drawing hierarchy to ', tree_name)
utils.draw(G, target, m._n, tree_name)
elif m.status == GRB.Status.INFEASIBLE:
# Compute IIS, for debugging purposes
m.computeIIS()
m.write('infeasible.ilp')