def compute_minEig(SKS, eps=1e-16):
mult = 1.0
minEig = torch.lobpcg(SKS.expand(1, -1, -1),
k=1,
largest=False,
method="ortho")[0]
# minEig = min(0, eigh(SKS, eigvals_only=True, subset_by_index=[0,0])) - eps
return minEig
def get_top_k_eigenvalues(matrix_, num_eigs):
try:
return torch.lobpcg(matrix_.expand(1, -1, -1),
k=num_eigs,
largest=True,
method="ortho")[0].type(torch.complex64)
except:
return torch.sort(torch.symeig(matrix_, eigenvectors=False)[0],
descending=True)[0][0:num_eigs].type(torch.complex64)
def max_frequency(self, lap_type: str = "sym"):
lap = self.L(lap_type)
if self._max_fs is not None:
max_fs = self._max_fs
else:
max_fs = torch.lobpcg(lap.to_torch_sparse_coo_tensor(), k=1, largest=True)[
0
].item()
if self.cache:
self._max_fs = max_fs
return max_fs
def forward(self, M, k, tol):
r"""
:param M: square symmetric matrix :math:`N \times N`
:param k: desired rank (must be smaller than :math:`N`)
:type M: torch.tensor
:type k: int
:return: eigenvalues D, leading k eigenvectors U
:rtype: torch.tensor, torch.tensor
**Note:** `depends on scipy`
Return leading k-eigenpairs of a matrix M, where M is symmetric
:math:`M=M^T`, by computing the symmetric decomposition :math:`M= UDU^T`
up to rank k. Partial eigendecomposition is done through LOBPCG method.
"""
# (optional) verify hermicity
M_asymm_norm = torch.norm(M - M.t())
assert M_asymm_norm / torch.abs(M).max() < 1.0e-8, "M is not symmetric"
# X (tensor, optional) - the input tensor of size (∗,m,n) where k <= n <= m.
# When specified, it is used as initial approximation of eigenvectors.
# X must be a dense tensor.
# iK (tensor, optional) - the input tensor of size (∗,m,m). When specified, it will be used
# as preconditioner.
# tol (float, optional) - residual tolerance for stopping criterion. Default is
# feps ** 0.5 where feps is smallest non-zero floating-point
# number of the given input tensor A data type.
# niter (int, optional) - maximum number of iterations. When reached, the iteration
# process is hard-stopped and the current approximation of eigenpairs
# is returned. For infinite iteration but until convergence criteria is met,
# use -1.
# tracker (callable, optional) - a function for tracing the iteration process. When specified,
# it is called at each iteration step with LOBPCG instance as an argument.
# The LOBPCG instance holds the full state of the iteration process in
# the following attributes:
# iparams, fparams, bparams - dictionaries of integer, float, and boolean valued input parameters,
# respectively
# ivars, fvars, bvars, tvars - dictionaries of integer, float, boolean, and Tensor valued
# iteration variables, respectively.
# A, B, iK - input Tensor arguments.
# E, X, S, R - iteration Tensor variables.
# ortho_fparams, ortho_bparams (ortho_iparams,) - various parameters to LOBPCG algorithm when
# using (default) method=”ortho”.
D, U= torch.lobpcg(M, k=k, \
X=None, n=None, iK=None, largest=True, tol=tol, niter=None, \
tracker=None, ortho_iparams=None, ortho_fparams=None, ortho_bparams=None)
self.save_for_backward(D, U)
return D, U
def forward(ctx, A):
#n = int(len(S) ** 0.5)
#A = S.reshape(n, n)
print('A=', A)
n = len(A)
k = n
if 1:
e, v = T.symeig(-A, eigenvectors=True)
e = -e[:k]
v = v[:, :k]
else:
e, v = T.lobpcg(A, k=k)
r = T.cat((T.flatten(v), e), 0)
ctx.save_for_backward(e, v, A)
print('r=', r)
return e, v
def eigen_solve(A, mode):
"""solve for eigenpairs using a specified method"""
if mode == 'arpack_eigsh':
return arpack.eigsh(A, largest=True, tol=1e-4)
elif mode == 'arpack_eigsh_mkl':
return arpack.eigsh_mkl(A, largest=True, tol_dps=4)
elif mode == 'torch_eigh':
return torch.linalg.eigh(A)
elif mode == 'torch_lobpcg':
return torch.lobpcg(A, k=1, largest=True, tol=1e-4)
elif mode == 'scipy_eigsh':
# For some reason scipy's eigsh requires slightly smaller tolerance
# (1e-5 vs 1e-4) to reach equiavelent accuracy
return splinalg.eigsh(A.numpy(), k=1, which="LA", tol=1e-5)
elif mode == 'scipy_lobpcg':
X = A.new_empty(A.size(0), 1).normal_()
return splinalg.lobpcg(A.numpy(), X.numpy(), largest=True, tol=1e-4)
else:
raise ValueError
def evaluate_with_lambda(net: torch.nn.Module,
dataloader: torch.utils.data.DataLoader, criterion):
net.eval()
avg_loss = AverageMeter()
avg_acc = AverageMeter()
avg_acc5 = AverageMeter()
device = next(net.parameters()).device
num_data = 0
total_loss = 0
K_mat = torch.zeros(num_parameters(net),
num_parameters(net),
device=device)
for data, label in dataloader:
data, label = data.to(device), label.to(device)
output = net(data)
loss = criterion(output, label)
batch_size = data.size(0)
num_data += batch_size
total_loss = (total_loss + loss * batch_size) / num_data
total_loss.backward()
g = extract_grad_vec(net)
for data, label in dataloader:
net.zero_grad()
data, label = data.to(device), label.to(device)
output = net(data)
loss = criterion(output, label)
loss.backward()
g_i = extract_grad_vec(net)
d = (g_i - g).to(device)
K_mat.addr_(d, d)
# K_mat += d.ger(d) / len(dataloader)
prec1, prec5 = accuracy(output.data, label.data, topk=(1, 5))
avg_loss.update(loss.item(), data.size(0))
avg_acc.update(prec1.item())
avg_acc5.update(prec5.item())
s, v = torch.lobpcg(K_mat, k=1)
return avg_loss.avg, avg_acc.avg, avg_acc5.avg, s.item()
def compute_label_stats(data,
targets=None,
indices=None,
classnames=None,
online=True,
batch_size=100,
to_tensor=True,
eigen_correction=False,
eigen_correction_scale=1.0,
nworkers=0,
diagonal_cov=False,
embedding=None,
device=None,
dtype=torch.FloatTensor):
"""
Computes mean/covariance of examples grouped by label. Data can be passed as
a pytorch dataset or a dataloader. Uses dataloader to avoid loading all
classes at once.
Arguments:
data (pytorch Dataset or Dataloader): data to compute stats on
targets (Tensor, optional): If provided, will use this target array to
avoid re-extracting targets.
indices (array-like, optional): If provided, filtering is based on these
indices (useful if e.g. dataloader has subsampler)
eigen_correction (bool, optional): If ``True``, will shift the covariance
matrix's diagonal by :attr:`eigen_correction_scale` to ensure PSD'ness.
eigen_correction_scale (numeric, optional): Magnitude of eigenvalue
correction (used only if :attr:`eigen_correction` is True)
Returns:
M (dict): Dictionary with sample means (Tensors) indexed by target class
S (dict): Dictionary with sample covariances (Tensors) indexed by target class
"""
device = process_device_arg(device)
M = {} # Means
S = {} # Covariances
## We need to get all targets in advance, in order to filter.
## Here we assume targets is the full dataset targets (ignoring subsets, etc)
## so we need to find effective targets.
if targets is None:
targets, classnames, indices = extract_data_targets(data)
else:
assert (indices
is not None), "If targets are provided, so must be indices"
if classnames is None:
classnames = sorted([a.item() for a in torch.unique(targets)])
effective_targets = targets[indices]
if nworkers > 1:
import torch.multiprocessing as mp # Ugly, sure. But useful.
mp.set_start_method('spawn', force=True)
M = mp.Manager().dict() # Alternatively, M = {}; M.share_memory
S = mp.Manager().dict()
processes = []
for i, c in enumerate(classnames): # No. of processes
label_indices = indices[effective_targets == i]
p = mp.Process(target=_single_label_stats,
args=(data, i, c, label_indices, M, S),
kwargs={
'device': device,
'online': online
})
p.start()
processes.append(p)
for p in processes:
p.join()
else:
for i, c in enumerate(classnames):
label_indices = indices[effective_targets == i]
μ, Σ, n = _single_label_stats(data,
i,
c,
label_indices,
device=device,
dtype=dtype,
embedding=embedding,
online=online,
diagonal_cov=diagonal_cov)
M[i], S[i] = μ, Σ
if to_tensor:
## Warning: this assumes classes are *exactly* {0,...,n}, might break things
## downstream if data is missing some classes
M = torch.stack(
[μ.to(device) for i, μ in sorted(M.items()) if μ is not None],
dim=0)
S = torch.stack(
[Σ.to(device) for i, Σ in sorted(S.items()) if Σ is not None],
dim=0)
### Shift the Covariance matrix's diagonal to ensure PSD'ness
if eigen_correction:
logger.warning('Applying eigenvalue correction to Covariance Matrix')
λ = eigen_correction_scale
for i in range(S.shape[0]):
if eigen_correction == 'constant':
S[i] += torch.diag(λ * torch.ones(S.shape[1], device=device))
elif eigen_correction == 'jitter':
S[i] += torch.diag(
λ *
torch.ones(S.shape[1], device=device).uniform_(0.99, 1.01))
elif eigen_correction == 'exact':
s, v = torch.symeig(S[i])
print(s.min())
s, v = torch.lobpcg(S[i], largest=False)
print(s.min())
s = torch.eig(S[i], eigenvectors=False).eigenvalues
print(s.min())
pdb.set_trace()
s_min = s.min()
if s_min <= 1e-10:
S[i] += torch.diag(λ * torch.abs(s_min) *
torch.ones(S.shape[1], device=device))
raise NotImplemented()
return M, S
def r_kernel(samples1,samples2,score_p_func,threshold=30,num_selection=None,**kwargs):
# for kernel_smooth active slice
samples1 = samples1.clone().detach()
samples1.requires_grad_()
if 'lobpcg' in kwargs:
flag_lobpcg=kwargs['lobpcg']
else:
flag_lobpcg=False
if 'fix_sample' in kwargs:
flag_fix_sample = kwargs['fix_sample']
else:
flag_fix_sample = False
if 'kernel' in kwargs:
kernel = kwargs['kernel']
else:
kernel = SE_kernel_multi
if flag_fix_sample:
# sample1 =sample 2
samples2=samples1.clone().detach()
samples2.requires_grad_()
else:
samples2=samples2.clone().detach()
samples2.requires_grad_()
score_p=score_p_func(samples1) # N
score_p=torch.autograd.grad(score_p.sum(),samples1)[0] # N x D
with torch.no_grad():
score_p = score_p.reshape((score_p.shape[0], 1, score_p.shape[1])) # N x 1 x D
# median distance
median_dist = median_heruistic(samples1[0:100, :], samples1[0:100, :].clone())
bandwidth = 2 * torch.pow(0.5 * median_dist, 0.5)
kernel_hyper_KSD = {
'bandwidth': 2*bandwidth
}
K = kernel(torch.unsqueeze(samples1, dim=1), torch.unsqueeze(samples2, dim=0)
, kernel_hyper=kernel_hyper_KSD) # N x N
Term1 = (torch.unsqueeze(K, dim=-1) * score_p) # N x N x D
Term2 = repulsive_SE_kernel_multi(samples1, samples2, kernel_hyper=kernel_hyper_KSD,
K=torch.unsqueeze(K, dim=-1)) # N x N xD
force = torch.mean(Term1 + 1* Term2, dim=0) # N x D
score_diff_1 = force.unsqueeze(-1) # 1000 x D x 1
score_diff_2 = force.unsqueeze(-2) # 1000 x 1 x D
score_mat_kernel = torch.matmul(score_diff_1, score_diff_2)
mean_score_mat = score_mat_kernel.mean(0) # D x D
if flag_lobpcg==True and num_selection is not None and (num_selection*3)<mean_score_mat.shape[0]:
# approximate eigenvectors
eig_value_appro, eig_vec_appro = torch.lobpcg(mean_score_mat, k=num_selection, niter=10)
r=eig_vec_appro.t()
else:
eig_value, eig_vec = torch.symeig(mean_score_mat, eigenvectors=True)
# Selection top large eigenvectors
r_th = threshold
r_candidate = eig_vec.t()
r_candidate = r_candidate.flip(0)
eig_value_cand = eig_value.flip(0)
max_eig_value = eig_value_cand[0]
if num_selection is None:
for d in range(samples1.shape[-1]):
eig_cand = eig_value_cand[d]
if torch.abs(max_eig_value / (eig_cand + 1e-10)) < r_th:
r_comp = r_candidate[d, :].unsqueeze(0) # 1x D
if d == 0:
r = r_comp
else:
r = torch.cat((r, r_comp), dim=0) # n x D
else:
for d in range(num_selection):
r_comp = r_candidate[d, :].unsqueeze(0) # 1x D
if d == 0:
r = r_comp
else:
r = torch.cat((r, r_comp), dim=0) # n x D
return r
def Poincare_g_kernel_SVD(samples1,samples2,score_p_func,**kwargs):
# active slice with kernel smooth estimated score_q
samples1 = samples1.clone().detach()
samples1.requires_grad_()
if 'lobpcg' in kwargs:
# enable fast approximation for eigenvalue decomposition
flag_lobpcg=kwargs['lobpcg']
else:
flag_lobpcg=False
if 'fix_sample' in kwargs:
flag_fix_sample=kwargs['fix_sample']
else:
flag_fix_sample=False
if 'kernel' in kwargs:
kernel=kwargs['kernel']
else:
kernel=SE_kernel_multi
if 'r' in kwargs:
r=kwargs['r']
else:
r=torch.eye(samples1.shape[-1]) # default value for r (eye initialization)
if flag_fix_sample: # samples1=samples2
samples2=samples1.clone().detach()
samples2.requires_grad_()
else:
samples2=samples2.clone().detach()
samples2.requires_grad_()
num_r=r.shape[0]
median_dist = median_heruistic(samples1[0:100, :], samples1[0:100, :].clone())
bandwidth = 2 * torch.pow(0.5 * median_dist, 0.5)
kernel_hyper_KSD = {
'bandwidth': 1*bandwidth
}
score_p=score_p_func(samples1)# N
score_p=torch.autograd.grad(score_p.sum(),samples1,create_graph=True,retain_graph=True)[0]# N x D
# compute kernel matrix
K = kernel(torch.unsqueeze(samples1, dim=1), torch.unsqueeze(samples2, dim=0),kernel_hyper=kernel_hyper_KSD) # N x N
K_exp=K.unsqueeze(-1)
score_p_exp = score_p.unsqueeze(-2) # N x 1 x D
samples1_exp = samples1.unsqueeze(-2) # N x 1 x D
samples2_exp = samples2.unsqueeze(-3) # 1 x N2 x D
for d in range(num_r):
r_candidate = r[d, :].unsqueeze(0).unsqueeze(0) # 1 x 1 x D
Term1=torch.einsum('ijk,pqr,ilr->q',K_exp,r_candidate,score_p_exp)
# only for RBF kernel. We explicit derive its derivative of this kernel,
Term2_c1=-2./(bandwidth**2+1e-9)*torch.einsum('ijk,imr,pqr->m',K_exp,samples1_exp,r_candidate)
Term2_c2=-2./(bandwidth**2+1e-9)*torch.einsum('ijk,mjr,pqr->m',K_exp,samples2_exp,r_candidate)
Term2=Term2_c1-Term2_c2
force = (Term1 + 1*Term2) # 1
grad_force = torch.autograd.grad(force.sum(), samples2, retain_graph=True)[0] # sam2 x D
H = grad_force.unsqueeze(-1) * grad_force.unsqueeze(-2) # sam2 x D x D
H = H.mean(0) # D x D
if flag_lobpcg:
_, eig_vec_appro = torch.lobpcg(H.clone().detach(), k=1, niter=30) # approximated eigen decomposition
if d == 0:
g_comp = eig_vec_appro.t() # 1 x D
g = g_comp
else:
g_comp = eig_vec_appro.t() # 1 x D
g = torch.cat((g, g_comp), dim=0) # r x D
else:
eig_value, eig_vec = torch.symeig(H.clone().detach(), eigenvectors=True) # eigenvalue decomposition
vec_candidate = eig_vec.t().flip(0)
if d == 0:
g_comp = vec_candidate[0, :].unsqueeze(0) # 1 x D
g = g_comp
else:
g_comp = vec_candidate[0, :].unsqueeze(0) # 1 x D
g = torch.cat((g, g_comp), dim=0) # r x D
return g
def wrapper_for_recNystrom(similarity_matrix,
K,
num_imp_samples,
runs=1,
mode="normal",
normalize="rows",
expand=False):
eps = 1e-3
mult = 1.5
error_list = []
abs_error_list = []
n = len(similarity_matrix)
list_of_available_indices = list(range(len(similarity_matrix)))
avg_min_eig = 0
for r in range(runs):
if mode == "eigI":
if expand:
new_num = int(np.sqrt(num_imp_samples * n))
sample_indices_bar = np.sort(
random.sample(list_of_available_indices, new_num))
min_eig_A = torch.lobpcg(similarity_matrix[np.ix_(sample_indices_bar, sample_indices_bar)].expand(1, -1, -1), \
k=1, largest=False, method="ortho")[0]
min_eig_A = min(0, min_eig_A) - eps
else:
pass
pass
if mode == "eigI":
pass
C, W, error, minEig = recursiveNystrom(similarity_matrix, num_imp_samples, \
correction=True, minEig=mult*min_eig_A, expand_eigs=True, eps=eps)
# rank_l_K = (C @ W) @ C.T
rank_l_K = torch.matmul(torch.matmul(C, W), torch.transpose(C, 0, 1))
# if np.iscomplexobj(rank_l_K):
# rank_l_K = np.absolute(rank_l_K)
# if normalize == "rows":
# rank_l_K = utils.row_norm_matrix(rank_l_K)
# similarity_matrix = utils.row_norm_matrix(similarity_matrix)
# pass
# if normalize == "laplacian":
# rank_l_K = utils.laplacian_norm_matrix(similarity_matrix, rank_l_K)
# pass
if normalize == "original":
pass
abs_error = torch.norm(K - rank_l_K) / torch.norm(K)
error_list.append(error.detach().cpu().numpy())
abs_error_list.append(abs_error.detach().cpu().numpy())
avg_min_eig += min_eig_A
if r < runs - 1:
del rank_l_K
pass
avg_min_eig = avg_min_eig / len(error_list)
avg_error = np.sum(np.array(error_list)) / len(error_list)
avg_abs_error = np.sum(np.array(abs_error_list)) / len(abs_error_list)
return avg_error, avg_abs_error, avg_min_eig, rank_l_K
# K = np.random.random((1000,600))
# K = torch.from_numpy(K).to(0)
# gamma = 40
# K = torch.cdist(K, K)
# K = torch.exp(-gamma*K)
# # print(K)
# s = 100
# #"""
# avg_error, avg_abs_error, avg_min_eig, rank_l_K = \
# wrapper_for_recNystrom(K, K, s, runs=10, mode="eigI", normalize="original", expand=True)
# print(avg_error, avg_abs_error, avg_min_eig)
#"""
def compute_spectral_emb(adj, K):
A = to_scipy(adj.to("cpu"))
L = from_scipy(sp.csgraph.laplacian(A, normed=True))
_, spectral_emb = torch.lobpcg(L, K)
return spectral_emb.to(adj.device)
def main(argv):
start_time = time.time()
num_cores = multiprocessing.cpu_count()
step = 50
norm_type = "original"
expand_eigs = True
mode = "eigI"
runs_ = 3
code_mode = "CPU"
from utils import read_file, read_mat_file
"""
20ng2_new_K_set1.mat oshumed_K_set1.mat recipe_K_set1.mat recipe_trainData.mat twitter_K_set1.mat twitter_set1.mat
"""
# approximation_type = "leverage_"
approximation_type = "uniform_"
if approximation_type == "leverage_":
from recursiveNystromGPU import wrapper_for_recNystrom as simple_nystrom
pass
if approximation_type == "uniform_":
from Nystrom import simple_nystrom
pass
filename = "stsb"
#similarity_matrix = read_file(pred_id_count=id_count, file_=filename+".npy")
print("Reading file ...")
# similarity_matrix = read_mat_file(file_="WordMoversEmbeddings/mat_files/20ng2_new_K_set1.mat", version="v7.3")
similarity_matrix = read_file("../GYPSUM/"+filename+"_predicts_0.npy")
print("File read. Beginning preprocessing ...")
#number_of_runs = id_count / step
error_list = []
abs_error_list = []
avg_min_eig_vec = []
# check for similar rows or columns
unique_rows, indices = np.unique(similarity_matrix, axis=0, return_index=True)
similarity_matrix_O = similarity_matrix[indices][:, indices]
similarity_matrix = (similarity_matrix_O + similarity_matrix_O.T) / 2.0
if filename == "rte":
similarity_matrix = 1-similarity_matrix
if code_mode == "GPU":
import torch
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
similarity_matrix = torch.from_numpy(similarity_matrix).to(device)
print("Preprocessing done.")
# print()
similarity_matrix_O = deepcopy(similarity_matrix)
# similarity_matrix_O = similarity_matrix
# """
def top_level_function(id_count):
error, abs_error, avg_min_eig, rank_l_k = simple_nystrom(\
similarity_matrix, similarity_matrix_O, id_count, runs=runs_, \
mode='eigI', normalize=norm_type, expand=expand_eigs)
# processed_list = [error, abs_error, avg_min_eig, rank_l_k]
print(error, abs_error)
return error, abs_error, avg_min_eig
id_count = 500 #len(similarity_matrix) #1000
inputs = tqdm(range(10, id_count, 10))
# for k in tqdm(range(10, id_count, 10)):
# error, abs_error, avg_min_eig, _ = simple_nystrom(similarity_matrix, similarity_matrix_O, k, runs=runs_, mode='eigI', normalize=norm_type, expand=expand_eigs)
# error_list.append(error)
# abs_error_list.append(abs_error)
# avg_min_eig_vec.append(avg_min_eig)
# del _
# pass
print("Beginning approximation parallely ...")
e = Parallel(n_jobs=num_cores, backend="threading")(map(delayed(top_level_function), inputs))
print("Approximation done. Beginning write out to files.")
for i in range(len(e)):
tuple_out = e[i]
error_list.append(tuple_out[0])
abs_error_list.append(tuple_out[1])
avg_min_eig_vec.append(tuple_out[2])
# print(len(error_list), len(abs_error_list), len(avg_min_eig_vec))
error_list = np.array(error_list)
abs_error_list = np.array(abs_error_list)
print("check for difference: ", np.linalg.norm(error_list-abs_error_list))
# min_eig = np.real(np.min(np.linalg.eigvals(similarity_matrix)))
if mode == "GPU":
min_eig = torch.lobpcg(similarity_matrix.expand(1, -1, -1), k=1, largest=False, method="ortho")[0].cpu().numpy()
else:
min_eig = np.real(np.min(np.linalg.eigvals(similarity_matrix)))
min_eig = round(min_eig, 2)
if mode == "GPU":
avg_min_eig_vec = [x.cpu().numpy() for x in avg_min_eig_vec]
else:
pass
# display
# sns.set()
# flatui = ["#9b59b6", "#3498db", "#95a5a6", "#e74c3c", "#34495e", "#2ecc71"]
# cmap = ListedColormap(sns.color_palette(flatui).as_hex())
x_axis = list(range(10, id_count, 10))
# fig, ax = plt.subplots(figsize=(15, 8))
# for label in (ax.get_xticklabels() + ax.get_yticklabels()):
# label.set_fontsize(16)
plt.rc('axes', titlesize=13)
plt.rc('axes', labelsize=13)
plt.rc('xtick', labelsize=13)
plt.rc('ytick', labelsize=13)
plt.rc('legend', fontsize=11)
STYLE_MAP = {"symmetrized error": {"color": "#4d9221", "marker": "s", "markersize": 7, 'label': 'avg error wrt similarity matrix', 'linewidth': 1},
"true error": {"color": "#7B3294", "marker": ".", "markersize": 7, 'label': 'avg error wrt true matrix', 'linewidth': 1},
"min eig": {"color": "#f1a340", "marker": ".", "markersize": 7, 'label': "True minimum eigenvalue = "+str(min_eig), 'linewidth': 1},
}
def plot_me():
plt.gcf().clear()
scale_ = 0.55
new_size = (scale_ * 10, scale_ * 8.5)
plt.gcf().set_size_inches(new_size)
sim_error_pairs = [(x, y) for x, y in zip(x_axis, error_list)]
true_error_pairs = [(x, y) for x, y in zip(x_axis, abs_error_list)]
arr1 = np.array(sim_error_pairs)
arr2 = np.array(true_error_pairs)
print(arr1.shape, arr2.shape)
plt.plot(arr1[:, 0], arr1[:, 1], **STYLE_MAP['symmetrized error'])
plt.plot(arr2[:, 0], arr2[:, 1], **STYLE_MAP['true error'])
plt.locator_params(axis='x', nbins=6)
plt.xlabel("Number of landmark samples")
plt.ylabel("Approximation error")
plt.title("Plot of average errors using Nystrom on "+filename+" BERT", fontsize=13)
plt.tight_layout()
plt.legend(loc='upper right')
plt.savefig("./test1.pdf")
# plt.savefig("figures/final_"+approximation_type+"_nystrom_errors_"+filename+".pdf")
# plt.close()
# plt.gcf().clear()
# scale_ = 0.55
# new_size = (scale_ * 10, scale_ * 8.5)
# plt.gcf().set_size_inches(new_size)
# eigval_estimate_pairs = [(x, y) for x, y in zip(x_axis, list(np.squeeze(avg_min_eig_vec)))]
# arr1 = np.array(eigval_estimate_pairs)
# plt.plot(arr1[:, 0], arr1[:, 1], **STYLE_MAP['min eig'])
# plt.locator_params(axis='x', nbins=6)
# plt.xlabel("Number of landmark samples")
# plt.ylabel("Minimum eigenvalue estimate")
# plt.tight_layout(rect=[0, 0.03, 1, 0.95])
# plt.legend(loc='upper right')
# plt.title("Plot of minimum eigenvalue estimate", fontsize=13)
# plt.savefig("./test2.pdf")
# # plt.savefig("figures/final_"+approximation_type+filename+"_min_eigenvalue_estimate.pdf")
# plt.close()
# plt.plot(x, error_list, label="average errors")
# plt.plot(x, abs_error_list, label="average errors wrt original similarity_matrix")
# plt.xlabel("number of reduced samples", fontsize=20)
# plt.ylabel("error score", fontsize=20)
# plt.legend(loc="upper right", fontsize=20)
# plt.title("plot of average errors using Nystrom on "+filename+" BERT", fontsize=20)
# if mode == "eigI":
# plt.savefig("figures/"+approximation_type+"nystrom_errors_new_"+mode+"_"+norm_type+"_"+str(int(expand_eigs))+"_"+filename+".pdf")
# else:
# plt.savefig("figures/"+approximation_type+"nystrom_errors_new_"+mode+"_"+norm_type+"_"+filename+".pdf")
# plt.clf()
# plt.plot(x, np.squeeze(avg_min_eig_vec), label="average min eigenvalues")
# plt.xlabel("number of reduced samples", fontsize=20)
# plt.ylabel("minimum eigenvalues", fontsize=20)
# plt.legend(loc="upper right", fontsize=20)
# plt.title("plot of average eigenvalues for original values: "+str(min_eig), fontsize=20)
# plt.savefig("figures/"+approximation_type+filename+"_min_eigenvalue_estimate.pdf")
# #"""
plot_me()
end_time = time.time()
print("total time for execution:", end_time-start_time)
