def plot_horizontal_bar_from_cm(confusion_matrix = None, classes = []):
# plt.rcdefaults()
# plt.subplots(figsize = (10, 30))
width = 0.30
# Example data
y_lables = 12 * ['a', 'b', 'c', 'd']
y_pos = list(range(len(y_lables)))
print(y_pos)
true_positives = 12 * [8.84036186, 12.94095337, 11.19919226, 10.64395389]
false_negatives = 12 * [1, 1, 1, 1]
false_positives = 12 * [2, 13, 13, 3]
TP = plt.barh(y_pos, true_positives, width, color = 'green', label = 'TP')
FN = plt.barh(y_pos, false_negatives, width, label = 'FN', left = TP)
plt.barh(y_pos, false_positives, width, label = 'FP', left = FN)
# ax.barh([p + width for p in y_pos], false_negatives, width, label = 'FN')
# ax.barh([p + width * 2 for p in y_pos], false_positives, width, label = 'FP')
plt.set_yticks([p + 1.5 * width for p in y_pos])
plt.set_yticklabels(y_lables)
plt.invert_yaxis() # labels read top-to-bottom
plt.set_xlabel('Performance')
plt.set_title('How fast do you want to go today?')
plt.legend(['TP', 'FN', 'FP'], loc = 'upper right')
plt.show()
def fix_hplot(df, statistic, xlabel, ylabel, hue, fontsize):
data = df.query('statistic == "{}"'.format(statistic))
pastel = ["#92C6FF", "#97F0AA", "#FF9F9A",
"#D0BBFF", "#FFFEA3", "#B0E0E6"]
pal = dict(Validation=pastel[0], Test=pastel[2])
plt = sns.barplot(x='data', y='label', data=data, hue=None, errwidth=1.0, capsize=0.15)
[ticklabel.set_fontsize(fontsize) for ticklabel in (plt.get_yticklabels())]
[ticklabel.set_fontsize(fontsize) for ticklabel in (plt.get_xticklabels())]
plt.set_yticklabels([ticklabel._text.capitalize() for ticklabel in (plt.get_yticklabels())], ha='left')
plt.get_yaxis().get_label().set_fontsize(fontsize)
plt.get_xaxis().get_label().set_fontsize(fontsize)
plt.get_yaxis().set_tick_params(pad=fontsize*10-10)
plt.set_ylabel(ylabel)
plt.set_xlabel(xlabel)
plt.set_xlim(0, 1.05)
return plt
color='red',marker='o',linestyle='None',markersize=3)
plt.loglog(slams,sflam_star,\
color='white',marker='o',linestyle='None',markersize=3)
# label='IRAS F16544$-$1604')
#plt.figtext(0.4,0.83,'WISE')
#print plt.gca()
axes=fig1.gca()
for label in axes.get_xticklabels() + axes.get_yticklabels():
label.set_fontsize('x-small')
if (i < nx):
plt.set_xlabel(r'$\lambda$ ($\mu$m)',size='x-small')
if (np.mod(i,nx)==0):
plt.set_ylabel(r'$\lambda F_\lambda$ (erg/s/cm$^{2}$)',size='x-small')
else:
plt.set_yticklabels('',visible='False',size='x-small')
plt.axis([0.4,150,1e-14,1e-6])
# print 'i = ', i
plt.text(0.7,5e-8,'M%d'%(i+1),size='small')
print "%12s &%15s &\$%4.1f\pm%4.1f\$ &\$%4.1f\pm%4.1f\$ &\$%4.1f\pm%4.1f\$ &\$%4.1f\pm%4.1f\$ &\$%4.1f\pm%4.1f\$ &\$%4.2f\pm%4.2f\$ &\$%4.2f\pm%4.2f\$ &\$%4.2f\pm%4.2f\$ &\$%4.2f\pm%4.2f\$(%s%s) &\$%4.2f\pm%4.2f\$(%s%s)\\\\"%\
(scat['2MASS_name'][sstar],scat['object_type'][sstar],\
scat['J_flux_c'][sstar],scat['J_D_flux_c'][sstar],\
scat['H_flux_c'][sstar],scat['H_D_flux_c'][sstar],\
scat['Ks_flux_c'][sstar],scat['Ks_D_flux_c'][sstar],\
w1f[wstar],w1df[wstar],\
w2f[wstar],w2df[wstar],\
w3f[wstar],w3df[wstar],\
w4f[wstar],w4df[wstar],\
scat['MP1_flux_c'][sstar],scat['MP1_D_flux_c'][sstar],\
scat['MP2_flux_c'][sstar],scat['MP2_D_flux_c'][sstar],\
scat['MP2_Q_det_c'][sstar],scat['MP2_imtype'][sstar],\
#plt.bar(np.arange(69),weights[0,:],label = features)
#
order = np.argsort(np.abs(weights[:]))[::-1]
plt.bar(np.arange(20),weights[order[:20]])
plt.xticks(np.arange(20), features[order[:20]],fontsize =12)
plt.xticks(rotation="vertical")
plt.ylabel("Feature weights")
plt.savefig("/neurospin/brainomics/2016_schizConnect/analysis/all_studies+VIP/Freesurfer/all_subjects/results/ROIs_analysis/weights/svm_weights_top20.png")
plt.barh(np.arange(20),weights[order[:20]])
plt.set_yticks(np.arange(20))
plt.set_yticklabels(features[order[:20]])
plt.xticks(np.arange(20), features[order[:20]],fontsize =12)
plt.xticks(rotation="vertical")
plt.ylabel("Feature weights")
plt.rc('font', family='serif')
plt.figure
plt.grid()
fig, ax = plt.subplots()
# Example data
features_names = features[order[:20]]
y_pos = np.arange(len(features_names))
performance = weights[order[:20]]
import pickle
import matplotlib.pyplot as plt
import numpy as np
# Example data
people = ('TP', 'FP', 'FN', 'TN')
y_pos = np.arange(len(people))
performance = (64354, 2701, 17007, 2417826)
plt.barh(y_pos, performance, align='center',
color='green', ecolor='black')
plt.set_yticks(y_pos)
plt.set_yticklabels(people)
plt.invert_yaxis() # labels read top-to-bottom
plt.set_xlabel('Performance')
plt.set_title('Precission - 0.959 Recal - 0.790')
plt.show()
fig1.clf()
i = 0
j = 0
for i in range(len(files)):
plt = fig1.add_subplot(ny,nx,j+1)
histogram = hist(files[i])
bin_edges=histogram[0]
x = histogram[1]
C = colour
plt.scatter(bin_edges[:-1],x,c=u'r')
axes=fig1.gca()
plt.set_xticklabels('',visible='False',size='small')
plt.set_yticklabels('',visible='True',size='small')
if (j==4):
plt.set_xlabel(r'Temperature (K)',size='small')
plt.set_xticklabels('',visible='True',size='small')
plt.axis([5,60,0,1.1E-3])
plt.set_ylabel('Normalised Number of pixels',size='small')
j = j + 1
fig1.show()
def main(TRAIN=False,
TUNING=False,
ANCHOR=False,
LIME=True,
STATISTICS=False,
PROTODASH=False):
# read poems using simplereader
poems_english = readPoems('tsv/english.tsv')
poems_german = readPoems('tsv/emotion.german.tsv')
poems_chinese = readPoems('tsv/chinese.tsv')
print(len(poems_english))
print(len(poems_german))
print(len(poems_chinese))
# set up label dictionary
label_dict = {
'Sadness': 0,
'Humor': 1,
'Suspense': 2,
'Nostalgia': 3,
'Uneasiness': 4,
'Annoyance': 5,
'Awe / Sublime': 6,
'Awe/Sublime': 6,
'Vitality': 7,
'Beauty / Joy': 8,
'Beauty/Joy': 8
}
# array of stanzas
stanzas = []
# array of most prominent label for each stanza
labels = []
# list of languages
lang = []
# extract sentences with one label
for poem in itertools.chain(poems_english, poems_german, poems_chinese):
for stanza in poem[1:]:
if poem in poems_english:
lang.append(0)
elif poem in poems_german:
lang.append(1)
else:
lang.append(2)
labelsPerStanza = []
currentStanzaIndex = len(stanzas)
newStanza = 1
for line in stanza:
if newStanza:
stanzas.append(line[0])
newStanza = 0
else:
stanzas[currentStanzaIndex] += " " + line[0]
labelsPerStanza.extend(line[1].split(" --- "))
if len(line) > 2:
labelsPerStanza.extend(line[2].split(" --- "))
counter = [0, 0, 0, 0, 0, 0, 0, 0, 0]
for label in labelsPerStanza:
counter[label_dict[label]] += 1
labels.append(np.argmax(counter))
# plot dataset statistics
if STATISTICS is True:
df = pd.DataFrame({
"stanzas": stanzas,
"labels": labels,
"languages": lang
})
bar_labels = [lab.replace(" ", "") for lab in label_dict.keys()]
ger_values = df.loc[df["languages"] == 1, "labels"].value_counts()
en_values = df.loc[df["languages"] == 0, "labels"].value_counts()
ch_values = df.loc[df["languages"] == 2, "labels"].value_counts()
print(type(df.loc[df["languages"] == 1, "labels"].value_counts()))
ger_values[3] = 0
ger_values.sort_index(inplace=True)
en_values.sort_index(inplace=True)
ch_values.sort_index(inplace=True)
width = 0.5
fig, ax = plt.subplots()
plt.grid(zorder=0, alpha=0.7)
ax.bar(bar_labels, ger_values, width, label='German')
ax.bar(bar_labels,
en_values,
width,
bottom=ger_values,
label='English')
ax.bar(bar_labels,
ch_values,
width,
bottom=en_values + ger_values,
label='Chinese')
ax.set_ylabel('Number of stanzas', fontsize=18)
ax.legend(prop={'size': 18})
ax.tick_params(axis='both', which='major', labelsize=18)
plt.xticks(rotation=16)
plt.show()
# transform labels into one hot encodings
one_hot_labels = to_categorical(labels)
# analyze distribution of labels in dataset
df = pd.DataFrame({"labels": labels})
print(df['labels'].value_counts())
# use pretrained multilingual model to encode sentences
model = SentenceTransformer('distiluse-base-multilingual-cased-v1')
embeddings = model.encode(stanzas)
# shuffle data and split into train and test set
all_data = [(embeddings[i], one_hot_labels[i], i)
for i in range(len(embeddings))]
unshuffled_data = all_data
random.shuffle(all_data)
embeddings = [emb for emb, _, _ in all_data]
labels = [lab for _, lab, _ in all_data]
indices = [idx for _, _, idx in all_data]
train_data = np.array(embeddings[:int(0.75 * len(embeddings))])
train_labels = np.array(labels[:int(0.75 * len(embeddings))])
dev_data = np.array(
embeddings[int(0.75 * len(embeddings)):int(0.875 * len(embeddings))])
dev_labels = np.array(
labels[int(0.75 * len(embeddings)):int(0.875 * len(embeddings))])
test_data = np.array(embeddings[int(0.875 * len(embeddings)):])
test_labels = np.array(labels[int(0.875 * len(embeddings)):])
# Hyperparameter Tuning
if TUNING is True:
learning_rates = [0.001, 0.01, 0.1]
epochs = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
middle_nodes = [20, 50, 100, 150, 200]
losses = []
accuracies = []
max_loss = 100000
min_acc = 0
max_config = None
for lr in learning_rates:
for epoch in epochs:
for middle_node in middle_nodes:
print("Training with following hyperparameters:", lr,
epoch, middle_node)
adam = Adam(learning_rate=lr)
mdl = Sequential()
mdl.add(
Dense(middle_node,
input_dim=512,
kernel_initializer="uniform",
activation="relu"))
mdl.add(
Dense(9,
activation="softmax",
kernel_initializer="uniform"))
mdl.compile(loss="categorical_crossentropy",
optimizer=adam,
metrics=["categorical_accuracy"])
mdl.fit(train_data, train_labels, epochs=epoch, verbose=1)
print("evaluating on dev set...")
(loss, accuracy) = mdl.evaluate(dev_data,
dev_labels,
verbose=1)
print("loss: {:.4f}, accuracy: {:.4f}%".format(
loss, accuracy * 100))
losses.append(loss)
accuracies.append(accuracy)
if accuracy > min_acc:
min_acc = accuracy
max_config = (lr, epoch, middle_node)
print(max_config)
max_config = (0.01, 7, 150)
mdl = Sequential()
if TRAIN is True:
# use final model
adam = Adam(learning_rate=max_config[0])
mdl = Sequential()
mdl.add(
Dense(max_config[2],
input_dim=512,
kernel_initializer="uniform",
activation="relu"))
mdl.add(Dense(9, activation="softmax", kernel_initializer="uniform"))
mdl.compile(loss="categorical_crossentropy",
optimizer=adam,
metrics=["categorical_accuracy"])
mdl.fit(train_data, train_labels, epochs=max_config[1], verbose=1)
print("evaluating on test set...")
(loss, accuracy) = mdl.evaluate(test_data, test_labels, verbose=1)
print("loss={:.4f}, accuracy: {:.4f}%".format(loss, accuracy * 100))
#print("precision={:.4f}%".format(precision * 100))
#print("recall={:.4f}%".format(recall * 100))
# mdl.save('emotion_classifier')
#mdl = keras.models.load_model('emotion_classifier')
(loss, accuracy) = mdl.evaluate(test_data, test_labels, verbose=1)
y_pred = mdl.predict(test_data, batch_size=test_data.shape[0])
wrong_classified_idx = []
for j, idx in enumerate(indices[int(0.875 * len(embeddings)):]):
if np.argmax(y_pred[j]) != np.where(test_labels[j] == 1.0)[0]:
wrong_classified_idx.append(idx)
print("These stanzas were wronlgy classified:")
print(wrong_classified_idx)
wrong_classified_en = [idx for idx in wrong_classified_idx if idx < 167]
wrong_classified_ger = [
idx for idx in wrong_classified_idx if (idx >= 167 and idx < 688)
]
wrong_classified_ch = [idx for idx in wrong_classified_idx if idx >= 688]
total_en = [
idx for idx in indices[int(0.875 * len(embeddings)):] if idx < 167
]
total_ger = [
idx for idx in indices[int(0.875 * len(embeddings)):]
if (idx >= 167 and idx < 688)
]
total_ch = [
idx for idx in indices[int(0.875 * len(embeddings)):] if idx >= 688
]
print("Number of wrongly classified stanzas - English: ",
len(wrong_classified_en))
print("Number of wrongly classified stanzas - German: ",
len(wrong_classified_ger))
print("Number of wrongly classified stanzas - Chinese: ",
len(wrong_classified_ch))
print("Total - English: ", len(total_en))
print("Total - German: ", len(total_ger))
print("Total - Chinese: ", len(total_ch))
class_names = [
'Sadness', 'Humor', 'Suspense', 'Nostalgia', 'Uneasiness', 'Annoyance',
'Awe / Sublime', 'Vitality', 'Beauty / Joy'
]
examples = [592, 9, 5]
# ------------------------------------------------------------LIME--------------------------------------------------------------------------------------------
# apply LIME to obtain explanations for a specific instance
def pipeline(stanza, mdl=mdl, model=model):
embedded = model.encode(stanza)
return mdl.predict(embedded, batch_size=embedded.shape[0])
if LIME is True:
# apply LIME to 10 uncorreclty classified stanzas
for idx in examples:
print("True Label: ", one_hot_labels[idx])
emb = np.array(model.encode(stanzas[idx]))
emb = emb.reshape((512, 1))
emb = emb.T
print("Predicted Probabilities: ", mdl.predict(emb, batch_size=1))
explainer = LimeTextExplainer(class_names=class_names)
exp = explainer.explain_instance(stanzas[idx],
pipeline,
num_features=6,
top_labels=2)
top_labs = exp.available_labels()
print("Explanation for class {}".format(top_labs[0]))
print('\n'.join(map(str, exp.as_list(label=top_labs[0]))))
print("Explanation for class {}".format(top_labs[1]))
print('\n'.join(map(str, exp.as_list(label=top_labs[1]))))
fig = exp.as_pyplot_figure(top_labs[0])
plt.show()
fig_2 = exp.as_pyplot_figure(top_labs[1])
plt.show()
# apply LIME to different correctly classified stanzas
idx = 5
print("True Label: ", one_hot_labels[idx])
emb = np.array(model.encode(stanzas[idx]))
emb = emb.reshape((512, 1))
emb = emb.T
print("Predicted Probabilities: ", mdl.predict(emb, batch_size=1))
print(mdl.predict(emb, batch_size=1).sum())
explainer = LimeTextExplainer(class_names=class_names)
exp = explainer.explain_instance(stanzas[idx],
pipeline,
num_features=6,
top_labels=2)
pickle.dump(exp, open("explanation.pkl", "wb"))
top_labs = exp.available_labels()
print("Explanation for class {}".format(top_labs[0]))
print('\n'.join(map(str, exp.as_list(label=top_labs[0]))))
print("Explanation for class {}".format(top_labs[1]))
print('\n'.join(map(str, exp.as_list(label=top_labs[1]))))
fig = exp.as_pyplot_figure(top_labs[0])
plt.legend(prop={'size': 600})
plt.tick_params(axis='both', which='major', labelsize=600)
plt.set_yticklabels(x, fontsize=600)
plt.show()
fig_2 = exp.as_pyplot_figure(top_labs[1])
plt.legend(prop={'size': 20})
plt.tick_params(axis='both', which='major', labelsize=20)
plt.show()
# ----------------------------------------------------------ANCHOR---------------------------------------------------------------------------------------------
def predict_label(stanza):
embedded = model.encode(stanza)
probs = mdl.predict(embedded, batch_size=embedded.shape[0])
return [np.argmax(probs[0])]
def predict_second_label(stanza, predicted_label):
embedded = model.encode(stanza)
probs = mdl.predict(embedded, batch_size=embedded.shape[0])
probs[0][np.argmax(probs[0])] = 0
return [np.argmax(probs)]
if ANCHOR is True:
ids = np.zeros(3)
print()
# for i in examples:
# lowest = 500
# lowest_id = 500
# for j in range(len(stanzas)):
# if len(stanzas[j]) < lowest:
# if j not in ids and len(stanzas[j]) > 85 and j < 174:
# lowest = len(stanzas[j])
# lowest_id = j
# ids[i] = lowest_id
# print("Ausgewähltes Stanza: ", stanzas[lowest_id])
# print("Länge: ", len(stanzas[lowest_id]), " id: ", lowest_id)
# print()
nlp = spacy.load('en_core_web_lg')
explainer = anchor_text.AnchorText(nlp,
class_names,
use_unk_distribution=True)
print("GPU's: ", get_available_gpus())
for idx in examples:
print()
print("------------STANZA-", idx, "------------")
print()
text = stanzas[idx]
print(predict_label([text]))
pred = explainer.class_names[predict_label([text])[0]]
alternative = explainer.class_names[predict_second_label(
[text],
predict_label([text])[0])[0]]
print('Prediction: %s' % pred)
print("Stanza: ", stanzas[idx], " True Label: ", labels[idx])
exp = explainer.explain_instance(text,
predict_label,
threshold=0.95)
print('Anchor: %s' % (' AND '.join(exp.names())))
print('Precision: %.2f' % exp.precision())
print()
print('Examples where anchor applies and model predicts %s:' %
pred)
print()
print('\n'.join(
[x[0] for x in exp.examples(only_same_prediction=True)]))
print()
print('Examples where anchor applies and model predicts %s:' %
alternative)
print()
print('\n'.join([
x[0] for x in exp.examples(partial_index=0,
only_different_prediction=True)
]))
# ----------------------------------------------------------PROTODASH------------------------------------------------------------------------------------------
if PROTODASH is True:
for idx in examples:
from aix360.algorithms.protodash import ProtodashExplainer
def predict_label(stanza):
embedded = model.encode(stanza)
embedded = embedded.reshape((512, 1))
embedded = embedded.T
probs = mdl.predict(embedded, batch_size=1)
return [np.argmax(probs)]
def index_to_vector(index):
for k, data in enumerate(all_data):
if data[2] == index:
return embeddings[k]
return None
explainer = ProtodashExplainer()
num_prototypes = 5
print(train_data.shape)
vector = index_to_vector(idx)
vector = vector.reshape((1, 512))
(weights, proto_ind, _) = explainer.explain(vector,
train_data,
m=num_prototypes)
weights = np.around(weights / np.sum(weights), 2)
print()
print("example: ", stanzas[idx])
print("prototypes with weights:")
print()
print()
for i in range(num_prototypes):
j = proto_ind[i]
print(weights[i], stanzas[indices[j]])
all_indices = [idx]
for i in range(num_prototypes):
j = proto_ind[i]
stanza_ind = indices[j]
all_indices.append(stanza_ind)
for l in all_indices:
print()
print(stanzas[l])
print("Predicted Label: ", predict_label(stanzas[l]))
print("True Label: ", np.argmax(one_hot_labels[l]))
# building graphix
plt.figure()
plt.subplots()
# plot figure 1
#ax1=plt.subplot(221)
plt.xlabel(r'Stress, $S$ (GPa)', fontsize=labelFontSizeX)
plt.ylabel('Probability density', fontsize=labelFontSizeY)
#plt.title('Histogram of Stress')
plt.grid(True)
n, bins, patches = plt.hist(data.x, num.bins, facecolor='green', normed=True, alpha=0.5)
data.linspace = np.linspace(bins[0], bins[num.bins], num.weib)
plt.plot(data.x_hist, weib.pdf(data.x_hist, coeff.shape, coeff.scale), 'r--')
ticks = plt.get_yticks()/sum(n)
newTicks = ['%.2f' % a for a in ticks]
plt.set_yticklabels(newTicks)
plt.savefig('xxx6_v2_%s_1.png' % (data.numfile), dpi=300, transparent=True)
# plot figure 2
plt.figure()
#ax2=plt.subplot(222)
plt.xlabel(r'Stress, $S$ (GPa)', fontsize=labelFontSizeX)
plt.ylabel('Cumulative distribution', fontsize=labelFontSizeY)
#plt.title('Histogram of Stress')
plt.grid(True)
plt.ylim(0., 1.)
plt.plot(data.x_hist, data.y_hist, 'go', alpha=0.5)
plt.plot(data.linspace, weib.cdf(data.linspace, coeff.shape, coeff.scale), 'r--')