def calc_potential(exe, params, label_name, noise_precision, prior_precision):
exe.copy_params_from(params)
exe.forward(is_train=False)
ret = 0.0
ret += (nd.norm(
exe.outputs[0] - exe.arg_dict[label_name]).asscalar() ** 2) / 2.0 * noise_precision
for v in params.values():
ret += (nd.norm(v).asscalar() ** 2) / 2.0 * prior_precision
return ret
def calc_potential(exe, params, label_name, noise_precision, prior_precision):
exe.copy_params_from(params)
exe.forward(is_train=False)
ret = 0.0
ret += (nd.norm(exe.outputs[0] - exe.arg_dict[label_name]).asscalar()**
2) / 2.0 * noise_precision
for v in params.values():
ret += (nd.norm(v).asscalar()**2) / 2.0 * prior_precision
return ret
def batched_l2_dist(a, b):
a_squared = nd.power(nd.norm(a, axis=-1), 2)
b_squared = nd.power(nd.norm(b, axis=-1), 2)
squared_res = nd.add(nd.linalg_gemm(
a, nd.transpose(b, axes=(0, 2, 1)), nd.broadcast_axes(nd.expand_dims(b_squared, axis=-2), axis=1, size=a.shape[1]), alpha=-2
), nd.expand_dims(a_squared, axis=-1))
res = nd.sqrt(nd.clip(squared_res, 1e-30, np.finfo(np.float32).max))
return res
def extended_jaccard_dist(x, y, pw=False):
score = dot_dist(x, y, pw)
x = nd.norm(x, ord=2, axis=-1)**2
y = nd.norm(y, ord=2, axis=-1)**2
if pw is False:
x = x.expand_dims(axis=1)
y = y.expand_dims(axis=0)
return score / (x + y - score)
def cosine_dist(x, y, pw=False):
score = dot_dist(x, y, pw)
x = nd.norm(x, ord=2, axis=-1)
y = nd.norm(y, ord=2, axis=-1)
if pw is False:
x = x.expand_dims(axis=1)
y = y.expand_dims(axis=0)
return score / (x * y)
def debug_norm_all(self, debug_gnorm=True):
if debug_gnorm:
for k, v, grad_v in zip(self._param_names,
self._exec_group.param_arrays,
self._exec_group.grad_arrays):
logging.debug("%s: v-norm: %g, g-norm: %g" % (k, nd.norm(
v[0]).asnumpy()[0], nd.norm(grad_v[0]).asnumpy()[0]))
else:
for k, v in zip(self._param_names, self._exec_group.param_arrays):
logging.debug("%s: v-norm: %g" %
(k, nd.norm(v[0]).asnumpy()[0]))
def cal_my_acc(test_files, target_files):
'''
this method is deprecated
:param test_files:
:param target_files:
:return:
'''
mTransform = MTransform()
normalize = transforms.Normalize(mean=0.5, std=0.5)
transform = transforms.Compose([
# transforms.Resize((96, 112)),
transforms.ToTensor(),
normalize,
# mTransform,
])
model = sphere_net.SphereNet20()
model.load_params("log_bn_dy/spherenet.model", ctx=mx.gpu())
correct = 0
total = 0
target_emb = {}
for target_file in target_files:
target_image = transform(nd.array(
Image.open(target_file))).as_in_context(mx.gpu())
target_image = nd.expand_dims(target_image, axis=0)
target_label = ''.join(target_file.split('/')[-1].split('.')[:-1])
target_out = model(target_image)
target_emb[target_label] = target_out
test_emb = {}
for test_file in test_files:
test_image = Image.open(test_file)
test_image = nd.expand_dims(transform(nd.array(test_image)),
axis=0).as_in_context(mx.gpu())
test_label = ''.join(test_file.split('/')[-1].split('.')[:-1])
test_out = model(test_image)
max_s = mx.nd.zeros(1, ctx=mx.gpu())
max_label = ''
sims = {}
for target_label, target_out in target_emb.items():
similarity = nd.sum(test_out * target_out) / \
(nd.norm(test_out) * nd.norm(target_out))
sims[target_label] = similarity.asscalar()
if max_s < similarity:
max_s = similarity
max_label = target_label
if ''.join(max_label.split('_')[:-1]) == ''.join(
test_label.split('_')[:-1]):
correct += 1
else:
print test_label, max_s.asscalar(), max_label
total += 1
test_emb[test_label] = test_out
# print correct, total, float(correct)/total
return float(correct) / total, test_emb, target_emb
def f(a):
b = a * 2
print('a', a)
print('nd.norm(a).asscalar()', nd.norm(a).asscalar())
print('nd.norm(b).asscalar()', nd.norm(b).asscalar())
while nd.norm(b).asscalar() < 1000:
b = b * 2
if nd.sum(b).asscalar() > 0:
c = b
else:
c = 100 * b
return c
def _realize_parameters(sym,
params,
graph,
inputs_ext,
target_bits={},
params_sim={}):
logger = logging.getLogger('log.calib.realize.parameters')
name = sym.attr('name')
attr = sym.list_attr()
if 'precision' not in attr or name in inputs_ext:
return sym, params
target_bit = int(attr['precision'])
data = params[name]
params[name] = sim.int_realize(data, target_bit, logger=logger)
# calculate error
error = params[name].astype('float32') - data
error_rate = error / data
if nd.sum(error).asscalar() == 0:
rate = 0
else:
rate = nd.norm(error_rate).asscalar() / np.product(data.shape)
if rate > 0.001:
logger.warn("realize parameter %-60s avg error=%10.9f shape=%s", name,
rate, data.shape)
else:
logger.debug("realize parameter %-60s avg error=%10.9f shape=%s", name,
rate, data.shape)
return sym, params
def _realize_parameters(sym, params, graph, inputs_ext, precs):
logger = logging.getLogger('log.realize.parameters')
name, op_name = sym.attr('name'), sym.attr('op_name')
attr = sym.list_attr()
if op_name != 'null':
return sym, params
if name in inputs_ext:
attr['precision'] = str(precs[name][out_key])
return mx.sym.var(name, attr=attr), params
prec = precs[name][out_key]
data = params[name]
params[name] = sim.int_realize(data, prec, logger=logger)
# calculate error
error = params[name].astype('float32') - data
if nd.sum(error).asscalar() == 0:
rate = 0
else:
rate = nd.norm(error / data).asscalar() / np.product(data.shape)
if rate > 0.001:
logger.warn("realize parameter %-60s avg error=%10.9f shape=%s", name,
rate, data.shape)
else:
logger.debug("realize parameter %-60s avg error=%10.9f shape=%s", name,
rate, data.shape)
attr['precision'] = str(prec)
node = mx.sym.var(name, attr=attr)
return node, params
def infer(self, head_emb, rel_emb, tail_emb):
head_emb = head_emb.expand_dims(axis=1)
rel_emb = rel_emb.expand_dims(axis=0)
score = (head_emb + rel_emb).expand_dims(
axis=2) - tail_emb.expand_dims(axis=0).expand_dims(axis=0)
return self.gamma - nd.norm(score, ord=self.dist_ord, axis=-1)
def my_loss(data, nc, ns, nq):
data = data.astype('float64')
cls_data = nd.reshape(data[0:nc * ns], (nc, ns, -1))
cls_center = nd.mean(cls_data, axis=1) + 1e-10
data_center_dis = nd.norm(data[nc * ns:].expand_dims(axis=1) -
cls_center.expand_dims(axis=0),
axis=2)**2
weight = nd.zeros((nc * nq, nc), ctx=data.context, dtype='float64')
for i in range(0, nc):
weight[i * nq:i * nq + nq, i] = 1
weight2 = 1 - weight
temp1 = nd.log_softmax(-data_center_dis, axis=1)
temp2 = nd.sum(temp1, axis=1)
temp3 = nd.sum(-temp2)
label = nd.argmin(data_center_dis, axis=1)
return temp3 / (nc * nq), label
loss1 = nd.sum(data_center_dis * weight)
temp = nd.sum(nd.exp(-data_center_dis), axis=1)
loss2 = nd.sum(nd.log(temp))
if loss1 is np.nan or loss2 is np.nan:
raise StopIteration
return (loss1 + loss2) / (nc * nq), label
def proto_loss(embedding, nc, ns, nq):
embedding = embedding.astype('float64');
cls_data = nd.reshape(embedding[0:nc*ns], (nc, ns, -1)); cls_data.attach_grad()
cls_center = nd.mean(cls_data, axis=1);
data_center_dis = nd.norm(embedding[nc*ns:].expand_dims(axis=1) - cls_center.expand_dims(axis=0),
axis=2) ** 2
# print(nd.max(data_center_dis).asscalar())
weight = nd.zeros((nc*nq, nc), ctx=embedding.context, dtype='float64')
pick_vec = nd.zeros((nc*nq), ctx=embedding.context)
for i in range(0, nc):
weight[i*nq:i*nq+nq, i] = 1
pick_vec[i*nq:i*nq+nq] = i
"""
temp = nd.SoftmaxOutput(-data_center_dis, label)
temp = nd.log(temp) * weight
temp = nd.sum(-temp, axis=1)
predict = nd.argmin(data_center_dis, axis=1)
return -temp * nd.log(temp), predict
"""
temp1 = nd.log_softmax(-data_center_dis, axis=1);
temp2 = nd.pick(temp1, index=pick_vec, axis=1);
temp3 = nd.sum(-temp2);
label = nd.argmin(data_center_dis, axis=1)
return temp3 / (nc * nq), label
def _get_opt(out, lambd):
absmax = out.abs().max().asscalar()
if lambd is None:
return absmax
mean = nd.mean(out).asscalar()
sqrt_n = math.sqrt(np.product(out.shape))
std = nd.norm(out - mean).asscalar() / sqrt_n
alpha = abs(mean) + lambd * std
# pos_out = nd.abs(out)
# pos_mean = nd.mean(pos_out).asscalar()
# pos_std = nd.norm(pos_out - pos_mean).asscalar() / sqrt_n
# pos_alpha = abs(pos_mean) + lambd * pos_std
opt = absmax
if alpha < 0.95 * absmax:
print("mean, std = [", mean, std, "]", "alpha=", alpha, "absmax=",
absmax)
opt = alpha
# if opt > 30:
# print ("mean, std = [", mean, std, "]", "alpha=", alpha,
# "absmax=", absmax)
# print ("ABS mean, std = [", pos_mean, pos_std, "]",
# "alpha=", pos_alpha, "absmax=", absmax)
return opt
def compute_retrospective_loss(self, observed_arr, encoded_arr,
decoded_arr, re_encoded_arr):
'''
Compute retrospective loss.
Returns:
The tuple data.
- `np.ndarray` of delta.
- `np.ndarray` of losses of each batch.
- float of loss of all batch.
'''
if self.__output_neuron_count == self.__hidden_neuron_count:
target_arr = nd.broadcast_sub(
encoded_arr, nd.expand_dims(observed_arr.mean(axis=2), axis=2))
summary_delta_arr = nd.sqrt(nd.power(decoded_arr - target_arr, 2))
else:
# For each batch, draw a samples from the Uniform distribution.
if self.__output_neuron_count > self.__hidden_neuron_count:
all_dim_arr = np.arange(self.__output_neuron_count)
np.random.shuffle(all_dim_arr)
choiced_dim_arr = all_dim_arr[:self.__hidden_neuron_count]
target_arr = nd.broadcast_sub(
encoded_arr,
nd.expand_dims(observed_arr[:, :,
choiced_dim_arr].mean(axis=2),
axis=2))
summary_delta_arr = nd.sqrt(
nd.power(decoded_arr[:, :, choiced_dim_arr] - target_arr,
2))
else:
all_dim_arr = np.arange(self.__hidden_neuron_count)
np.random.shuffle(all_dim_arr)
choiced_dim_arr = all_dim_arr[:self.__output_neuron_count]
target_arr = nd.broadcast_sub(
encoded_arr[:, :, choiced_dim_arr],
nd.expand_dims(observed_arr.mean(axis=2), axis=2))
summary_delta_arr = nd.sqrt(
nd.power(decoded_arr - target_arr, 2))
match_delta_arr = None
for i in range(self.__batch_size):
arr = nd.sqrt(
nd.power(encoded_arr[i, -1] - re_encoded_arr[i, -1], 2))
if match_delta_arr is None:
match_delta_arr = nd.expand_dims(arr, axis=0)
else:
match_delta_arr = nd.concat(match_delta_arr,
nd.expand_dims(arr, axis=0),
dim=0)
delta_arr = summary_delta_arr + nd.expand_dims(
self.__retrospective_lambda * match_delta_arr, axis=1)
v = nd.norm(delta_arr)
if v > self.__grad_clip_threshold:
delta_arr = delta_arr * self.__grad_clip_threshold / v
loss = nd.mean(delta_arr, axis=0, exclude=True)
return loss
def edge_func(self, edges):
head = edges.src['emb']
tail = edges.dst['emb']
rel = edges.data['emb']
score = head + rel - tail
return {
'score': self.gamma - nd.norm(score, ord=self.dist_ord, axis=-1)
}
def fn(heads, relations, tails, num_chunks, chunk_size,
neg_sample_size):
hidden_dim = heads.shape[1]
heads = heads + relations
heads = heads.reshape(num_chunks, chunk_size, 1, hidden_dim)
tails = tails.reshape(num_chunks, 1, neg_sample_size,
hidden_dim)
return gamma - nd.norm(heads - tails, ord=1, axis=-1)
def fn(heads, relations, tails, num_chunks, chunk_size,
neg_sample_size):
relations = relations.reshape(num_chunks, -1,
self.relation_dim)
heads = heads - relations
heads = heads.reshape(num_chunks, -1, 1, self.relation_dim)
score = heads - tails
return gamma - nd.norm(score, ord=1, axis=-1)
def SGLD(sym, X, Y, X_test, Y_test, total_iter_num,
data_inputs=None,
learning_rate=None,
lr_scheduler=None, prior_precision=1,
out_grad_f=None,
initializer=None,
minibatch_size=100, thin_interval=100, burn_in_iter_num=1000, task='classification',
dev=mx.gpu()):
if out_grad_f is None:
label_key = list(set(data_inputs.keys()) - set(['data']))[0]
exe, params, params_grad, _ = get_executor(sym, dev, data_inputs, initializer)
optimizer = mx.optimizer.create('sgld', learning_rate=learning_rate,
rescale_grad=X.shape[0] / minibatch_size,
lr_scheduler=lr_scheduler,
wd=prior_precision)
updater = mx.optimizer.get_updater(optimizer)
sample_pool = []
start = time.time()
for i in xrange(total_iter_num):
indices = numpy.random.randint(X.shape[0], size=minibatch_size)
X_batch = X[indices]
Y_batch = Y[indices]
exe.arg_dict['data'][:] = X_batch
if out_grad_f is None:
exe.arg_dict[label_key][:] = Y_batch
exe.forward(is_train=True)
exe.backward()
else:
exe.forward(is_train=True)
exe.backward(out_grad_f(exe.outputs, nd.array(Y_batch, ctx=dev)))
for k in params:
updater(k, params_grad[k], params[k])
print k, nd.norm(params_grad[k]).asnumpy()
if i < burn_in_iter_num:
continue
else:
if 0 == (i - burn_in_iter_num) % thin_interval:
if optimizer.lr_scheduler is not None:
lr = optimizer.lr_scheduler(optimizer.num_update)
else:
lr = learning_rate
sample_pool.append([lr, copy_param(exe)])
if (i + 1) % 100000 == 0:
end = time.time()
if task == 'classification':
print "Current Iter Num: %d" % (i + 1), "Time Spent: %f" % (end - start)
test_correct, test_total, test_acc = \
sample_test_acc(exe, sample_pool=sample_pool, X=X_test, Y=Y_test, label_num=10,
minibatch_size=minibatch_size)
print "Test %d/%d=%f" % (test_correct, test_total, test_acc)
else:
print "Current Iter Num: %d" % (i + 1), "Time Spent: %f" % (end - start), "MSE:",
print sample_test_regression(exe=exe, sample_pool=sample_pool,
X=X_test,
Y=Y_test, minibatch_size=minibatch_size,
save_path='regression_SGLD.txt')
start = time.time()
return exe, sample_pool
def f(a):
b = a * 2
while nd.norm(b).asscalar() < 1000:
b = b * 2
if nd.sum(b).asscalar() > 0:
c = b
else:
c = 100 * b
return c
def calculate_norm(x, y):
assert x.shape == y.shape
ndims = np.product(x.shape)
x = nd.reshape(x, shape=(ndims, ))
y = nd.reshape(y, shape=(ndims, ))
res = x - y
nx = nd.norm(x)
ny = nd.norm(y)
nr = nd.norm(res)
print("saving...")
f = "/home/ryt/data/cmp_"
names = ["nx", "ny", "nr"]
objs = [nx, ny, nr]
for obj in objs:
print(type(obj), obj.shape)
for i in range(3):
nd.save(f + names[i], objs[i])
print('success')
def norm_clipping(params_grad, threshold):
assert isinstance(params_grad, dict)
norm_val = numpy.sqrt(sum([nd.norm(grad).asnumpy()[0]**2 for grad in params_grad.values()]))
# print('grad norm: %g' % norm_val)
ratio = 1.0
if norm_val > threshold:
ratio = threshold / norm_val
for grad in params_grad.values():
grad *= ratio
return norm_val
def norm_clipping(params_grad, threshold):
assert isinstance(params_grad, dict)
norm_val = numpy.sqrt(sum([nd.norm(grad).asnumpy()[0]**2 for grad in params_grad.values()]))
# print('grad norm: %g' % norm_val)
ratio = 1.0
if norm_val > threshold:
ratio = threshold / norm_val
for grad in params_grad.values():
grad *= ratio
return norm_val
def f(a):
b = a *2
#b的L2范数的标量
while nd.norm(b).asscalar() < 1000:
b = b *2
#b的轴上和的标量
if nd.sum(b).asscalar() > 0:
c = b
else:
c = 100 * b
return c
def _get_opt(out, lambd):
absmax = out.abs().max().asscalar()
if lambd is None:
return absmax
mean = nd.mean(out).asscalar()
std = nd.norm(out - mean).asscalar() / math.sqrt(np.product(out.shape))
alpha = abs(mean) + lambd * std
if alpha < 0.95 * absmax:
print("[", mean, std, "]", alpha, absmax)
return alpha
return absmax
def nd_global_norm(t_list):
"""Computes the global norm of multiple tensors.
Given a tuple or list of tensors t_list, this operation returns the global norm of the elements
in all tensors in t_list. The global norm is computed as:
``global_norm = sqrt(sum([l2norm(t)**2 for t in t_list]))``
Any entries in t_list that are of type None are ignored.
Parameters
----------
t_list: list or tuple
The NDArray list
Returns
-------
ret: NDArray
The global norm. The shape of the NDArray will be (1,)
Examples
--------
>>> x = mx.nd.ones((2, 3))
>>> y = mx.nd.ones((5, 6))
>>> z = mx.nd.ones((4, 2, 3))
>>> print(nd_global_norm([x, y, z]).asscalar())
7.74597
>>> xnone = None
>>> ret = nd_global_norm([x, y, z, xnone])
>>> print(ret.asscalar())
7.74597
"""
ret = None
for arr in t_list:
if arr is not None:
if ret is None:
ret = nd.square(nd.norm(arr))
else:
ret += nd.square(nd.norm(arr))
ret = nd.sqrt(ret)
return ret
def f(a):
b = a * 2
i = 0
while nd.norm(b).asscalar() < 1000:
i += 1
print(i)
b = b * 2
if nd.sum(b).asscalar() > 0:
c = b
else:
print('100')
c = 100 * b
return c
def norm_clipping(params_grad, threshold):
norm_val = 0.0
for i in range(len(params_grad[0])):
norm_val += np.sqrt(
sum([nd.norm(grads[i]).asnumpy()[0]**2 for grads in params_grad]))
norm_val /= float(len(params_grad[0]))
if norm_val > threshold:
ratio = threshold / float(norm_val)
for grads in params_grad:
for grad in grads:
grad[:] *= ratio
return norm_val
def predict(net, data_loader, ctx):
label = []
acc = 0
for data, cls_id in data_loader:
data = data.as_in_context(ctx)
out = net(data)
min_dis = math.inf
p_key = None
for key in net.cls_center:
cur_dis = nd.norm(net.cls_center[key] - out)
if cur_dis.asscalar() < min_dis:
min_dis = cur_dis.asscalar()
p_key = key
if p_key == cls_id.asscalar():
acc += 1
label.append(p_key)
return label, acc / len(label)
def get_opt(self, raw_ft, out, **kwargs):
logger = kwargs.get("logger", logging.getLogger("optimize"))
hist_ft = kwargs.get("hist_ft", None)
name = kwargs.get("name", _NULL_NAME)
if isinstance(raw_ft, AFeature):
# hyperparameter 'lambd' for fine tuning
absmax = raw_ft.get()
if self.lambd is not None:
mean = nd.mean(out).asscalar()
sqrt_n = math.sqrt(np.product(out.shape))
std = nd.norm(out - mean).asscalar() / sqrt_n
alpha = abs(mean) + self.lambd * std
absmax = alpha if alpha < 0.95 * absmax else absmax
if hist_ft is None:
p = logger.debug if absmax < 30 else logger.warn
p("collect symbol %-40s, out_shape=%-20s, opt: (%s)", name,
out.shape, absmax)
opt = AFeature(absmax)
else:
opt = AFeature(max(habsmax, hist_ft.get()))
elif isinstance(raw_ft, MMFeature):
minv, maxv = raw_ft.get()
if hist_ft is None:
opt = MMFeature(minv, maxv)
else:
hminv, hmaxv = hist_ft.get()
opt = MMFeature(min(minv, hminv), max(maxv, hmaxv))
elif isinstance(raw_ft, ALFeature):
if hist_ft is None:
opt = raw_ft
else:
absmax_list = raw_ft.get()
habsmax_list = raw_ft.get()
nabsmax_list = [ \
max(absmax_list[i], habsmax_list[i]) \
for i in range(len(absmax_list))
]
opt = ALFeature(nabsmax_list)
else:
raise TypeError("Unsupported feature type: %s for HVOptimizor",
type(raw_ft))
return opt
def get_global_norm_val(self):
"""Get the overall gradient norm ||W||_2
Parameters
----------
net : mx.mod.Module
Returns
-------
norm_val : float
"""
assert self.binded and self.params_initialized
#TODO The code in the following will cause the estimated norm to be different for multiple gpus
norm_val = 0.0
for i in range(len(self._exec_group.grad_arrays[0])):
norm_val += np.sqrt(
sum([
nd.norm(grads[i]).asnumpy()[0]**2
for grads in self._exec_group.grad_arrays
]))
norm_val /= float(len(self._exec_group.grad_arrays[0]))
return norm_val
def l1_dist(x, y, pw=False):
if pw is False:
x = x.expand_dims(axis=1)
y = y.expand_dims(axis=0)
return -nd.norm(x-y, ord=1, axis=-1)
def gradient_penalty(gradient):
gradient = gradient.reshape(gradient.shape[0], -1)
gradient_norm = nd.norm(gradient, ord=2, axis=1)
penalty = nd.mean((gradient_norm - 1) ** 2)
return penalty