def get_y_pred(model_path, test_data_path, is_decaf6=False, top_k=1, bagging=False, preprocessing=None):
model = init(model_path, is_decaf6=is_decaf6)
target_size = (224, 224)
if is_decaf6:
target_size = (227, 227)
dico = get_dico()
y_true = []
y_pred = []
s = 0
for t in list(os.walk(test_data_path)):
s += len(t[2])
style_names = os.listdir(test_data_path)
print('Calculating predictions...')
bar = ProgressBar(max_value=s)
i = 0
for style_name in style_names:
style_path = join(test_data_path, style_name)
img_names = os.listdir(style_path)
label = dico.get(style_name)
for img_name in img_names:
img = load_img(join(style_path, img_name), target_size=target_size)
x = img_to_array(img)
if bagging:
pred = _bagging_predict(x, model, preprocessing)
else:
x = _preprocess_input(x, preprocessing)
pred = model.predict(x[np.newaxis, ...])
args_sorted = np.argsort(pred)[0][::-1]
y_true.append(label)
y_pred.append([a for a in args_sorted[:top_k]])
i += 1
bar.update(i)
return np.asarray(y_pred), y_true
def get_scores_labels(model, test_data_path, is_decaf=False):
target_size = (224, 224)
if is_decaf:
target_size = (227, 227)
dico = get_dico()
y = []
scores = []
s = 0
for t in list(os.walk(test_data_path)):
s += len(t[2])
style_names = os.listdir(test_data_path)
print('Calculating predictions...')
bar = ProgressBar(max_value=s)
i = 0
for style_name in style_names:
style_path = join(test_data_path, style_name)
img_names = os.listdir(style_path)
label = dico.get(style_name)
for img_name in img_names:
img = load_img(join(style_path, img_name), target_size=target_size)
x = img_to_array(img)
pred = model.predict(x[np.newaxis, ...])
y.append(label)
scores.append(pred[0])
i += 1
bar.update(i)
index_shuf = range(len(y))
y = [y[i] for i in index_shuf]
scores = [scores[i] for i in index_shuf]
return np.asarray(scores), y
def plot_confusion_matrix(labels, preds):
conf_arr = confusion_matrix(labels, preds)
dico = get_dico()
new_conf_arr = []
for row in conf_arr:
new_conf_arr.append(row / sum(row))
plt.matshow(new_conf_arr)
plt.yticks(range(25), dico.keys())
plt.xticks(range(25), dico.keys(), rotation=90)
plt.colorbar()
plt.show()
def plot_confusion_matrix(labels,preds):
conf_arr = confusion_matrix(labels, preds)
dico = get_dico()
new_conf_arr = []
for row in conf_arr:
new_conf_arr.append(row / sum(row))
plt.matshow(new_conf_arr)
plt.yticks(range(25), dico.keys())
plt.xticks(range(25), dico.keys(), rotation=90)
plt.colorbar()
plt.savefig('./conf_mat.eps',format='eps', dpi=1000,bbox_inches='tight')
def get_per_class_accuracy(labels, preds):
names = []
accs = []
dico = get_dico()
inv_dico = invert_dico(dico)
for value in set(labels):
s = 0
n = 0
for i in range(len(labels)):
if (labels[i] == value):
n = n + 1
if (preds[i] == value):
s = s + 1
names.append(inv_dico.get(value))
accs.append(s / n * 100, )
return accs, names
def plot_confusion_matrix(labels,preds):
print("build matrix")
conf_arr = confusion_matrix(labels, preds)
dico = get_dico()
new_conf_arr = []
for row in conf_arr:
new_conf_arr.append(row / sum(row))
plt.matshow(new_conf_arr)
plt.yticks(range(25), dico.keys())
plt.xticks(range(25), dico.keys(), rotation=90)
plt.colorbar()
plt.show()
plt.savefig('../../assets/confusion_matrix.png', dpi = 300)
def get_pred(model, image_path, is_decaf6=False, top_k=1,bagging=DEFAULT_BAGGING,preprocessing=DEFAULT_PREPROCESSING):
target_size = (224, 224)
if is_decaf6:
target_size = (227, 227)
img = load_img(image_path, target_size=target_size)
x = img_to_array(img)
if bagging:
pred = _bagging_predict(x, model,preprocessing=preprocessing)
else:
x = _preprocess_input(x,preprocessing=preprocessing)
pred = model.predict(x[np.newaxis, ...])
dico = get_dico()
inv_dico = invert_dico(dico)
args_sorted = np.argsort(pred)[0][::-1]
preds = [inv_dico.get(a) for a in args_sorted[:top_k]]
pcts = [pred[0][a] for a in args_sorted[:top_k]]
return preds,pcts
def get_y_pred(model_path, test_data_path,top_k=1):
model = load_model(model_path)
target_size =(224,224)
dico = get_dico()
y_true = []
y_pred = []
s = 0
for t in list(os.walk(test_data_path)):
s += len(t[2])
style_names = os.listdir(test_data_path)
if '.DS_Store' in style_names:
style_names.remove('.DS_Store')
print('Calculating predictions...')
bar = ProgressBar(max_value=s)
i = 0
for style_name in style_names:
style_path = join(test_data_path, style_name)
img_names = os.listdir(style_path)
if '.DS_Store' in img_names:
img_names.remove('.DS_Store')
label = dico.get(style_name)
for img_name in img_names:
img = load_img(join(style_path, img_name),target_size=target_size)
x = img_to_array(img)
#With bagging
pred = _bagging_predict(style_name, img_name, x,model)
args_sorted = np.argsort(pred)[0][::-1]
y_true.append(label)
y_pred.append([a for a in args_sorted[:top_k]])
i += 1
bar.update(i)
return np.asarray(y_pred), y_true
args.data_path,
is_decaf=args.isdecaf)
n_per_classes = 87
new_scores = []
new_labels = []
for classe in set(labels):
n_class = 0
for i, label in enumerate(labels):
if label == classe and n_class < n_per_classes:
n_class += 1
new_scores.append(scores[i, :])
new_labels.append(label)
print(len(new_scores), len(new_labels))
dico = get_dico()
inv_dico = {v: k for k, v in dico.items()}
APs = []
new_scores = np.asarray(new_scores)
for classe in set(labels):
temp_labels = []
for i, label in enumerate(new_labels):
if label == classe:
temp_labels.append(1)
else:
temp_labels.append(0)
temp_scores = new_scores[:, classe]
score = average_precision_score(np.asarray(temp_labels), temp_scores)
APs.append(score)
print(np.sum(np.asarray(temp_labels)), len(temp_labels),