def eval_embed_space(run,
net_type,
metric,
rating_metric,
epochs,
dset,
rating_norm='none',
cross_validation=False,
n_groups=5):
# init
Embed = FileManager.Embed(net_type)
embed_source = [
Embed(run + 'c{}'.format(c), dset) for c in range(n_groups)
]
idx_hubness, idx_symmetry, idx_concentration, idx_contrast, idx_kummar, idx_featCorr, idx_sampCorr \
= [[] for i in range(n_groups)], [[] for i in range(n_groups)], [[] for i in range(n_groups)], \
[[] for i in range(n_groups)], [[] for i in range(n_groups)], [[] for i in range(n_groups)], \
[[] for i in range(n_groups)]
valid_epochs = [[] for i in range(n_groups)]
# calculate
Ret = Retriever(title='{}'.format(run), dset=dset)
for i, source in enumerate(embed_source):
embd, epoch_mask = Ret.load_embedding(source, multi_epcch=True)
for e in epochs:
try:
epoch_idx = np.argwhere(e == epoch_mask)[0][0]
Ret.fit(metric=metric, epoch=e)
indices, distances = Ret.ret_nbrs()
# hubness
idx_hubness[i].append(calc_hubness(indices))
# symmetry
idx_symmetry[i].append(calc_symmetry(indices))
# kumar index
tau, l_e = kumar(distances, res=0.01)
idx_kummar[i].append(tau)
# concentration & contrast
idx_concentration[i].append(concentration(distances))
idx_contrast[i].append(relative_contrast_imp(distances))
valid_epochs[i].append(e)
# correlation
idx_featCorr[i].append(features_correlation(embd[epoch_idx]))
idx_sampCorr[i].append(samples_correlation(embd[epoch_idx]))
except:
print("Epoch {} - no calculated embedding".format(e))
valid_epochs[i] = np.array(valid_epochs[i])
idx_hubness[i] = np.array(list(zip(*idx_hubness[i])))
idx_symmetry[i] = np.array(list(zip(*idx_symmetry[i])))
idx_concentration[i] = np.array(list(zip(*idx_concentration[i])))
idx_contrast[i] = np.array(list(zip(*idx_contrast[i])))
idx_kummar[i] = np.array([idx_kummar[i]])
idx_featCorr[i] = np.array([idx_featCorr[i]])
idx_sampCorr[i] = np.array([idx_sampCorr[i]])
combined_epochs = [
i for i, c in enumerate(np.bincount(np.concatenate(valid_epochs)))
if c > 3
]
idx_hubness = mean_cross_validated_index(idx_hubness, valid_epochs,
combined_epochs)
idx_symmetry = mean_cross_validated_index(idx_symmetry, valid_epochs,
combined_epochs)
idx_concentration = np.zeros_like(
idx_hubness
) #mean_cross_validated_index(idx_concentration, valid_epochs, combined_epochs)
idx_contrast = np.zeros_like(
idx_hubness
) # mean_cross_validated_index(idx_contrast, valid_epochs, combined_epochs)
idx_kummar = np.zeros_like(
idx_hubness
) # mean_cross_validated_index(idx_kummar, valid_epochs, combined_epochs)
idx_featCorr = np.zeros_like(
idx_hubness
) # mean_cross_validated_index(idx_featCorr, valid_epochs, combined_epochs)
idx_sampCorr = np.zeros_like(
idx_hubness
) # mean_cross_validated_index(idx_sampCorr, valid_epochs, combined_epochs)
return combined_epochs, idx_hubness, idx_symmetry, idx_concentration, idx_contrast, idx_kummar, idx_featCorr, idx_sampCorr
plt.figure("Distances - {}".format(name))
p = [None] * 9
for i in range(9):
p[i] = plt.subplot(3, 3, i + 1)
# init
Embed = FileManager.Embed(net_type)
embed_source = [
Embed(run + 'c{}'.format(c), dset) for c in range(n_groups)
]
idx_hubness, idx_symmetry, idx_concentration, idx_contrast, idx_kummar, valid_epochs = [], [], [], [], [], []
# calculate
Ret = Retriever(title='{}'.format(run), dset=dset)
embd, epoch_mask = Ret.load_embedding(embed_source, multi_epcch=True)
for e in [60]: # epochs:
# full
Ret.fit(metric=metric, epoch=e)
_, distances = Ret.ret_nbrs()
plot_row(0, distances)
# benign
Ret.fit(metric=metric, epoch=e, label=0)
_, distances = Ret.ret_nbrs()
plot_row(3, distances, label=0)
# malignant
Ret.fit(metric=metric, epoch=e, label=1)
_, distances = Ret.ret_nbrs()
plot_row(6, distances, label=1)
#p[-1].legend(run_names)
print('Done distance analysis')
#plt.ylabel('Normalized')
#plt.show()
for metric, m, in zip(metrics, range(len(metrics))):
norm = 'None'
if len(metric) > 5 and metric[-4:] == 'Norm':
metric = metric[:-5]
norm = 'Norm'
elif len(metric) > 6 and metric[-5:] == 'Scale':
metric = metric[:-6]
norm = 'Scale'
#distance_matrix = calc_distance_matrix(rating_normalize(rating, norm), method=metric)
Ret.fit(len(data) - 1, metric=metric, normalization=norm)
indices, distances = Ret.ret_nbrs()
plt.figure()
plt.hist(distances.flatten(), bins=500)
plt.title('distance distribution')
plt.figure()
# Hubness
K = [3, 5, 7, 11, 17]
#K = 1+np.array(range(20))
h = np.zeros(len(K))
plt.figure('k_occ')
plt.title('k_occ distribution')