def evaluate_confusion(bs: int, file: str, fixed: int, model_path: str,
test_bin) -> None:
"""
Evaluates the confusion matrix for a given number of features
:param bs: batch size
:param file: file where the confusion matrix will be written
:param fixed: number of features to be considered
:param model_path: string pointing to the .h5 keras model of the network.
If empty will default to data_dir/model.h5
:param test_bin: path to the test dataset that will be used
"""
test = BinaryDs(test_bin, read_only=True).open()
binary = test.get_categories() <= 2
model = load_model(model_path)
generator = DataGenerator(test, bs, fake_pad=True, pad_len=fixed,
predict=True)
expected = get_expected(bs, test)
predicted = model.predict(generator, verbose=1)
if binary:
predicted = np.round(predicted).flatten().astype(np.int8)
else:
predicted = np.argmax(predicted, axis=1)
matrix = np.array(tf.math.confusion_matrix(expected, predicted))
with open(file, "w") as f:
np.savetxt(f, X=matrix, fmt="%d")
test.close()
def count_categories(dataset: BinaryDs) -> List[int]:
examples = dataset.get_examples_no()
amount = 1000
read_total = int(examples / amount)
remainder = examples % amount
categories = []
for i in range(read_total):
buffer = dataset.read(i * amount, amount)
for val in buffer:
category = val[0]
while len(categories) <= category:
categories.append(0)
categories[category] += 1
if remainder > 0:
buffer = dataset.read(read_total * amount, remainder)
for val in buffer:
category = val[0]
while len(categories) <= category:
categories.append(0)
categories[category] += 1
assert len(categories) == dataset.get_categories()
return categories