def test_metric_tracker_best():
"""
测试 metric tracker
:return:
"""
metric_tracker = MetricTracker(patient=None)
for metric in METRICS:
metric_tracker.add_metric(**metric)
expect = {"epoch": 3,
"train_metric": {"acc": 0.85},
"train_model_target_metric": ModelTargetMetric(metric_name="acc", metric_value=0.85),
"validation_metric": {"acc": 0.60},
"validation_model_target_metric": ModelTargetMetric(metric_name="acc", metric_value=0.60)}
best = metric_tracker.best()
ASSERT.assertEqual(expect["epoch"], best.epoch)
ASSERT.assertDictEqual(expect["train_metric"], best.train_metric)
ASSERT.assertDictEqual(expect["validation_metric"], best.validation_metric)
ASSERT.assertEqual(expect["train_model_target_metric"].name, best.train_model_target_metric.name)
ASSERT.assertEqual(expect["train_model_target_metric"].value, best.train_model_target_metric.value)
ASSERT.assertEqual(expect["validation_model_target_metric"].name, best.validation_model_target_metric.name)
ASSERT.assertEqual(expect["validation_model_target_metric"].value, best.validation_model_target_metric.value)
def test_dynamic_rnn(sequence_embedding):
sequence, mask = sequence_embedding
hidden_size = 4
batch_size = 3
sequence_len = 3
rnn = RNN(input_size=2,
hidden_size=4,
num_layers=2,
batch_first=True,
bidirectional=True)
dynamic_rnn = DynamicRnn(rnn=rnn)
rnn_output: DynamicRnnOutput = dynamic_rnn(sequence=sequence, mask=mask)
logging.info(json2str(rnn_output))
last_layer_h_n: torch.Tensor = rnn_output.last_layer_h_n
last_layer_h_n_expect_size = (batch_size, hidden_size * 2)
ASSERT.assertEqual(last_layer_h_n_expect_size, last_layer_h_n.size())
ASSERT.assertTrue(rnn_output.last_layer_c_n is None)
sequence_encoding_expect_size = (batch_size, sequence_len, hidden_size * 2)
senquence_encoding = rnn_output.output
ASSERT.assertEqual(sequence_encoding_expect_size,
senquence_encoding.size())
def test_metric_tracker_save_and_load():
metric_tracker = MetricTracker(patient=1)
for metric in METRICS:
metric_tracker.add_metric(**metric)
if metric["epoch"] > 4:
ASSERT.assertTrue(metric_tracker.early_stopping(metric["epoch"]))
else:
ASSERT.assertFalse(metric_tracker.early_stopping(metric["epoch"]))
if metric_tracker.early_stopping(metric["epoch"]):
break
saved_file_path = os.path.join(ROOT_PATH, "data/easytext/tests/trainer/metric_tracker.json")
metric_tracker.save(saved_file_path)
loaded_metric_tracker = MetricTracker.from_file(saved_file_path)
best = metric_tracker.best()
loaded_best = loaded_metric_tracker.best()
ASSERT.assertEqual(best.epoch, loaded_best.epoch)
ASSERT.assertDictEqual(best.train_metric, loaded_best.train_metric)
ASSERT.assertDictEqual(best.validation_metric, loaded_best.validation_metric)
ASSERT.assertEqual(best.train_model_target_metric.name, loaded_best.train_model_target_metric.name)
ASSERT.assertEqual(best.train_model_target_metric.value, loaded_best.train_model_target_metric.value)
ASSERT.assertEqual(best.validation_model_target_metric.name, loaded_best.validation_model_target_metric.name)
ASSERT.assertEqual(best.validation_model_target_metric.value, loaded_best.validation_model_target_metric.value)
def test_pretrained_vocabulary(pretrained_vocabulary):
"""
测试预训练词汇表
"""
ASSERT.assertEqual(4, pretrained_vocabulary.size)
ASSERT.assertEqual(4, len(pretrained_vocabulary))
ASSERT.assertEqual(2, pretrained_vocabulary.index("我"))
ASSERT.assertEqual(3, pretrained_vocabulary.index("美丽"))
ASSERT.assertEqual((pretrained_vocabulary.size, 3),
pretrained_vocabulary.embedding_matrix.size())
expect_embedding_dict = {
"a": [1.0, 2.0, 3.0],
"b": [4.0, 5.0, 6.0],
"美丽": [7.0, 8.0, 9.0]
}
ASSERT.assertListEqual(
expect_embedding_dict["美丽"], pretrained_vocabulary.embedding_matrix[
pretrained_vocabulary.index("美丽")].tolist())
zero_vec = [0.] * 3
for index in [
pretrained_vocabulary.index("我"),
pretrained_vocabulary.padding_index,
pretrained_vocabulary.index(pretrained_vocabulary.unk)
]:
ASSERT.assertListEqual(
zero_vec, pretrained_vocabulary.embedding_matrix[index].tolist())
def test_metric_tracker_patient():
metric_tracker = MetricTracker(patient=1)
for metric in METRICS:
metric_tracker.add_metric(**metric)
if metric["epoch"] > 4:
ASSERT.assertTrue(metric_tracker.early_stopping(metric["epoch"]))
else:
ASSERT.assertFalse(metric_tracker.early_stopping(metric["epoch"]))
if metric_tracker.early_stopping(metric["epoch"]):
break
expect = {"epoch": 3,
"train_metric": {"acc": 0.85},
"train_model_target_metric": ModelTargetMetric(metric_name="acc", metric_value=0.85),
"validation_metric": {"acc": 0.60},
"validation_model_target_metric": ModelTargetMetric(metric_name="acc", metric_value=0.60)}
best = metric_tracker.best()
ASSERT.assertEqual(expect["epoch"], best.epoch)
ASSERT.assertDictEqual(expect["train_metric"], best.train_metric)
ASSERT.assertDictEqual(expect["validation_metric"], best.validation_metric)
ASSERT.assertEqual(expect["train_model_target_metric"].name, best.train_model_target_metric.name)
ASSERT.assertEqual(expect["train_model_target_metric"].value, best.train_model_target_metric.value)
ASSERT.assertEqual(expect["validation_model_target_metric"].name, best.validation_model_target_metric.name)
ASSERT.assertEqual(expect["validation_model_target_metric"].value, best.validation_model_target_metric.value)
def test_masked_softmax():
"""
测试 masked softmax
:return:
"""
vector = torch.FloatTensor([[1., 2., 3.], [4., 5., 6.]])
mask = torch.ByteTensor([[1, 1, 0], [1, 1, 1]])
result = masked_softmax(vector=vector, mask=mask)
expect1 = np.exp(np.array([1., 2.]))
expect1 = expect1 / np.sum(expect1)
expect1 = np.concatenate([expect1, np.array([0.])], axis=-1).tolist()
result1 = result[0].tolist()
ASSERT.assertEqual(len(expect1), len(result1))
for expect_data, result_data in zip(expect1, result1):
ASSERT.assertAlmostEqual(expect_data, result_data)
expect2 = np.exp(np.array([4., 5., 6.]))
expect2 = expect2 / np.sum(expect2)
expect2 = expect2.tolist()
result2 = result[1].tolist()
ASSERT.assertEqual(len(expect2), len(result2))
for expect_data, result_data in zip(expect2, result2):
ASSERT.assertAlmostEqual(expect_data, result_data)
def test_multi_input_lstm_cell():
"""
测试 MultiInputLSTMCell
"""
input_size = 2
hidden_size = 3
cell = MultiInputLSTMCell(input_size=input_size,
hidden_size=hidden_size,
bias=True)
with torch.no_grad():
weight_ih_value = list()
for i in range(input_size):
weight_ih_value.append([
j * 0.37
for j in range(i * hidden_size * 3, (i + 1) * hidden_size * 3)
])
cell.weight_ih.copy_(torch.tensor(weight_ih_value, dtype=torch.float))
alpha_weight_ih_value = list()
for i in range(input_size):
alpha_weight_ih_value.append([
j * 0.23 for j in range(i * hidden_size, (i + 1) * hidden_size)
])
cell.alpha_weight_ih.copy_(
torch.tensor(alpha_weight_ih_value, dtype=torch.float))
torch.nn.init.constant(cell.bias, val=1.0)
torch.nn.init.constant(cell.alpha_bias, val=0.5)
char_input = torch.tensor([[0.2, 0.4]], dtype=torch.float)
h = torch.tensor([[0.2, 0.11, 0.15]], dtype=torch.float)
c = torch.tensor([[0.5, 0.6, 0.7]], dtype=torch.float)
word_c_input = [
torch.tensor([[0.7, 0.5, 0.2]], dtype=torch.float),
torch.tensor([[0.3, 0.4, 1.5]], dtype=torch.float)
]
output_hc = cell(input_=char_input, c_input=word_c_input, hx=(h, c))
expect_size = (1, hidden_size)
ASSERT.assertEqual(expect_size, output_hc[0].size())
ASSERT.assertEqual(expect_size, output_hc[1].size())
expects = [[0.5728, 0.5523, 0.7130], [0.6873, 0.6506, 0.9345]]
for expect, hc in zip(expects, output_hc):
for e_i, hc_i in zip(expect, hc[0].tolist()):
ASSERT.assertAlmostEqual(e_i, hc_i, places=4)
def test_label_f1_metric():
"""
测试 label f1 metric
"""
predictions = torch.tensor([0, 0, 1, 1, 2, 2, 3, 3, 4, 4])
gold_labels = torch.tensor([0, 1, 1, 2, 2, 3, 3, 4, 4, 1])
labels = [0, 1, 2, 3, 4]
f1_metric = LabelF1Metric(labels=labels, label_vocabulary=None)
metrics = f1_metric(prediction_labels=predictions,
gold_labels=gold_labels,
mask=None)
logging.debug(json2str(metrics))
ASSERT.assertEqual((len(labels) + 1) * 3, len(metrics))
precision_0 = metrics[f"{LabelF1Metric.PRECISION}-0"]
recall_0 = metrics[f"{LabelF1Metric.RECALL}-0"]
f1_0 = metrics[f"{LabelF1Metric.F1}-0"]
expect_precision_0 = 1. / 2.
ASSERT.assertAlmostEqual(expect_precision_0, precision_0)
expect_recall_0 = 1. / 1.
ASSERT.assertAlmostEqual(expect_recall_0, recall_0)
expect_f1_0 = 2. * expect_precision_0 * expect_recall_0 / (
expect_precision_0 + expect_recall_0)
ASSERT.assertAlmostEqual(expect_f1_0, f1_0)
expect_precision_overall = 5. / 10.
expect_recall_overall = 5. / 10
precision_overall = metrics[LabelF1Metric.PRECISION_OVERALL]
recall_overall = metrics[LabelF1Metric.RECALL_OVERALL]
ASSERT.assertAlmostEqual(expect_precision_overall, precision_overall)
ASSERT.assertAlmostEqual(expect_recall_overall, recall_overall)
predictions = torch.tensor([0, 2])
gold_labels = torch.tensor([0, 1])
f1_metric(prediction_labels=predictions,
gold_labels=gold_labels,
mask=None)
precision_0 = f1_metric.metric[f"{LabelF1Metric.PRECISION}-0"]
recall_0 = f1_metric.metric[f"{LabelF1Metric.RECALL}-0"]
f1_0 = f1_metric.metric[f"{LabelF1Metric.F1}-0"]
expect_precision_0 = (1. + 1.) / (2. + 1.)
ASSERT.assertAlmostEqual(expect_precision_0, precision_0)
expect_recall_0 = (1. + 1.) / (1. + 1.)
ASSERT.assertAlmostEqual(expect_recall_0, recall_0)
expect_f1_0 = 2. * expect_precision_0 * expect_recall_0 / (
expect_precision_0 + expect_recall_0)
ASSERT.assertAlmostEqual(expect_f1_0, f1_0)
def test_sequence_max_label_index_decoder():
label_vocabulary = LabelVocabulary(
[["B-T", "B-T", "B-T", "I-T", "I-T", "O"]],
padding=LabelVocabulary.PADDING)
b_index = label_vocabulary.index("B-T")
ASSERT.assertEqual(0, b_index)
i_index = label_vocabulary.index("I-T")
ASSERT.assertEqual(1, i_index)
o_index = label_vocabulary.index("O")
ASSERT.assertEqual(2, o_index)
# [[O, B, I], [B, B, I], [B, I, I], [B, I, O]]
batch_sequence_logits = torch.tensor(
[[[0.2, 0.3, 0.4], [0.7, 0.2, 0.3], [0.2, 0.3, 0.1]],
[[0.8, 0.3, 0.4], [0.7, 0.2, 0.3], [0.2, 0.3, 0.1]],
[[0.8, 0.3, 0.4], [0.1, 0.7, 0.3], [0.2, 0.3, 0.1]],
[[0.8, 0.3, 0.4], [0.1, 0.7, 0.3], [0.2, 0.3, 0.5]]],
dtype=torch.float)
expect_sequence_labels = [["O", "B-T", "I-T"], ["B-T", "B-T", "I-T"],
["B-T", "I-T", "I-T"], ["B-T", "I-T", "O"]]
expect = list()
for expect_sequence_label in expect_sequence_labels:
expect.append(
[label_vocabulary.index(label) for label in expect_sequence_label])
decoder = SequenceMaxLabelIndexDecoder(label_vocabulary=label_vocabulary)
label_indices = decoder(logits=batch_sequence_logits, mask=None)
ASSERT.assertEqual(expect, label_indices.tolist())
def test_cnn_seq2vec_output_dim():
"""
测试 cnn 输出维度
:return:
"""
kernel_size = (1, 2, 3, 4, 5)
encoder = CnnSeq2Vec(embedding_dim=7,
num_filters=13,
kernel_sizes=kernel_size)
tokens = torch.rand(4, 8, 7)
vector = encoder(sequence=tokens, mask=None)
expect = (4, 13 * len(kernel_size))
ASSERT.assertEqual(expect, vector.size())
def event_type_vocabulary():
event_types = [["A", "B", "C"], ["A", "B"], ["A"]]
vocabulary = Vocabulary(tokens=event_types,
padding="",
unk="Negative",
special_first=True)
ASSERT.assertEqual(4, vocabulary.size)
ASSERT.assertEqual(0, vocabulary.index(vocabulary.unk))
ASSERT.assertEqual(1, vocabulary.index("A"))
ASSERT.assertEqual(2, vocabulary.index("B"))
ASSERT.assertEqual(3, vocabulary.index("C"))
return vocabulary
def test_component_evaluate_factory():
Registry().clear_objects()
config_json_file_path = "data/easytext/tests/component/training.json"
config_json_file_path = os.path.join(ROOT_PATH, config_json_file_path)
with open(config_json_file_path, encoding="utf-8") as f:
param_dict = json.load(f, object_pairs_hook=OrderedDict)
factory = ComponentFactory(is_training=False)
parsed_dict = factory.create(config=param_dict)
my_component = parsed_dict["my_component"]
ASSERT.assertEqual("evaluate_3", my_component.value)
def test_glove_loader():
pretrained_file_path = "data/easytext/tests/pretrained/word_embedding_sample.3d.txt"
pretrained_file_path = os.path.join(ROOT_PATH, pretrained_file_path)
glove_loader = GloveLoader(embedding_dim=3,
pretrained_file_path=pretrained_file_path)
embedding_dict = glove_loader.load()
expect_embedding_dict = {
"a": [1.0, 2.0, 3.0],
"b": [4.0, 5.0, 6.0],
"美丽": [7.0, 8.0, 9.0]
}
ASSERT.assertDictEqual(expect_embedding_dict, embedding_dict)
ASSERT.assertEqual(glove_loader.embedding_dim, 3)
def test_label_vocabulary():
"""
测试 label vocabulary
:return:
"""
vocabulary = LabelVocabulary([["A", "B", "C"], ["D", "E"]], padding="")
ASSERT.assertEqual(vocabulary.size, 5)
vocabulary = LabelVocabulary([["A", "B", "C"], ["D", "E"]],
padding=LabelVocabulary.PADDING)
ASSERT.assertEqual(vocabulary.size, 6)
ASSERT.assertEqual(vocabulary.label_size, 5)
ASSERT.assertEqual(vocabulary.index(vocabulary.padding), 5)
for index, w in enumerate(["A", "B", "C", "D", "E"]):
ASSERT.assertEqual(vocabulary.index(w), index)
def test_component_factory():
Registry().clear_objects()
model_json_file_path = "data/easytext/tests/component/model.json"
model_json_file_path = os.path.join(ROOT_PATH, model_json_file_path)
with open(model_json_file_path, encoding="utf-8") as f:
config = json.load(f, object_pairs_hook=OrderedDict)
factory = ComponentFactory(is_training=True)
parserd_dict = factory.create(config=config)
model = parserd_dict["model"]
ASSERT.assertTrue(model.linear is not None)
ASSERT.assertEqual((2, 4),
(model.linear.in_features, model.linear.out_features))
def test_default_typename():
"""
测试，当 component 构建的时候，某个参数是 object
:return:
"""
Registry().clear_objects()
config_json_file_path = "data/easytext/tests/component/default_typename.json"
config_json_file_path = os.path.join(ROOT_PATH, config_json_file_path)
with open(config_json_file_path, encoding="utf-8") as f:
param_dict = json.load(f, object_pairs_hook=OrderedDict)
factory = ComponentFactory(is_training=False)
parsed_dict = factory.create(config=param_dict)
default_typename = parsed_dict["default_typename"]
ASSERT.assertEqual(10, default_typename.value)
def test_vocabulary_speical_first():
"""
测试 vocabulary speical first
:return:
"""
batch_tokens = [["我", "和", "你"], ["在", "我"]]
vocabulary = Vocabulary(batch_tokens,
padding=Vocabulary.PADDING,
unk=Vocabulary.UNK,
special_first=True,
min_frequency=1,
max_size=None)
ASSERT.assertEqual(vocabulary.size, 6)
ASSERT.assertEqual(vocabulary.padding, vocabulary.PADDING)
ASSERT.assertEqual(vocabulary.unk, vocabulary.UNK)
ASSERT.assertEqual(vocabulary.index(vocabulary.padding), 0)
ASSERT.assertEqual(vocabulary.index(vocabulary.unk), 1)
def test_fill():
"""
测试 bio
:return:
"""
pairs = [(1, 2), (2, 4)]
for begin, end in pairs:
sl = ["O"] * 10
tag = "Test"
BIO.fill(sequence_label=sl, begin_index=begin, end_index=end, tag=tag)
for i in range(begin, end):
if i == begin:
ASSERT.assertEqual(sl[i], f"B-{tag}")
else:
ASSERT.assertEqual(sl[i], f"I-{tag}")
def test_gat_without_hidden():
"""
测试 gat
:return:
"""
torch.manual_seed(7)
torch.cuda.manual_seed_all(7)
in_features = 2
out_features = 4
gat = GAT(in_features=in_features,
out_features=out_features,
dropout=0.,
alpha=0.1,
num_heads=3,
hidden_size=None)
nodes = torch.tensor([[[0.1, 0.2], [0.3, 0.4], [0.5, 0.6]],
[[0.7, 0.8], [0.9, 0.10], [0.11, 0.12]]],
dtype=torch.float)
adj = torch.tensor(
[[[0, 1, 0], [1, 0, 0], [0, 0, 0]], [[0, 1, 1], [1, 0, 1], [1, 1, 0]]],
dtype=torch.long)
output_nodes: torch.Tensor = gat(nodes=nodes, adj=adj)
expect_size = (nodes.size(0), nodes.size(1), out_features)
ASSERT.assertEqual(expect_size, output_nodes.size())
expect = torch.tensor([[[-1.6478, -0.3935, -2.6613, -2.7653],
[-1.3204, -0.8394, -1.8519, -1.9375],
[-1.6478, -0.3935, -2.6613, -2.7653]],
[[-1.9897, -0.4203, -2.4447, -2.1232],
[-2.1944, -0.1897, -3.4053, -3.5697],
[-2.9364, -0.0878, -4.1695, -4.1617]]],
dtype=torch.float)
ASSERT.assertTrue(
tensor_util.is_tensor_equal(expect, output_nodes, epsilon=1e-4))
def test_gat_with_hidden():
"""
测试 gat
:return:
"""
torch.manual_seed(7)
torch.cuda.manual_seed_all(7)
in_features = 2
out_features = 4
gat = GAT(in_features=in_features,
out_features=out_features,
dropout=0.,
alpha=0.1,
num_heads=3,
hidden_size=3)
nodes = torch.tensor([[[0.1, 0.2], [0.3, 0.4], [0.5, 0.6]],
[[0.7, 0.8], [0.9, 0.10], [0.11, 0.12]]],
dtype=torch.float)
adj = torch.tensor(
[[[0, 1, 0], [1, 0, 0], [0, 0, 0]], [[0, 1, 1], [1, 0, 1], [1, 1, 0]]],
dtype=torch.long)
output_nodes: torch.Tensor = gat(nodes=nodes, adj=adj)
expect_size = (nodes.size(0), nodes.size(1), out_features)
ASSERT.assertEqual(expect_size, output_nodes.size())
expect = torch.tensor([[[-1.3835, -1.4764, -1.2033, -1.5113],
[-1.3316, -1.5785, -1.1564, -1.5368],
[-1.3475, -1.5467, -1.1706, -1.5279]],
[[-1.3388, -1.6693, -1.4427, -1.1610],
[-1.4288, -1.6525, -1.6607, -0.9707],
[-1.4320, -1.4422, -1.6465, -1.1025]]])
ASSERT.assertTrue(
tensor_util.is_tensor_equal(expect, output_nodes, epsilon=1e-4))
def test_cnn_seq2vec():
"""
测试 cnn seq2vec
:return:
"""
encoder = CnnSeq2Vec(embedding_dim=2, num_filters=1, kernel_sizes=(1, 2))
for name, parameter in encoder.named_parameters():
parameter.data.fill_(1.)
tokens = torch.FloatTensor([[[0.7, 0.8], [0.1, 1.5]]])
vector = encoder(sequence=tokens, mask=None)
vector = vector.view(-1).tolist()
expect = torch.tensor([[0.1 + 1.5 + 1.,
0.7 + 0.8 + 0.1 + 1.5 + 1.]]).view(-1).tolist()
ASSERT.assertEqual(len(expect), len(vector))
for i in range(len(vector)):
ASSERT.assertAlmostEqual(expect[i], vector[i])
def __call__(self, instances: Iterable[Instance]) -> ModelInputs:
x = list()
labels = list()
for instance in instances:
x_data = instance["x"]
x.append(torch.tensor([x_data], dtype=torch.float))
if x_data - 50 > 0:
labels.append(1)
else:
labels.append(0)
x = torch.stack(x)
batch_size = x.size(0)
ASSERT.assertEqual(x.dim(), 2)
ASSERT.assertListEqual([batch_size, 1], [x.size(0), x.size(1)])
labels = torch.tensor(labels)
ASSERT.assertEqual(labels.dim(), 1)
ASSERT.assertEqual(batch_size, labels.size(0))
model_inputs = ModelInputs(batch_size=batch_size,
model_inputs={"x": x},
labels=labels)
return model_inputs
def test_decode():
"""
测试 模型输出的 batch logits 解码
:return:
"""
# [[O, B, I], [B, B, I], [B, I, I], [B, I, O]]
batch_sequence_logits = torch.tensor([[[0.2, 0.3, 0.4], [0.7, 0.2, 0.3], [0.2, 0.3, 0.1]],
[[0.8, 0.3, 0.4], [0.7, 0.2, 0.3], [0.2, 0.3, 0.1]],
[[0.8, 0.3, 0.4], [0.1, 0.7, 0.3], [0.2, 0.3, 0.1]],
[[0.8, 0.3, 0.4], [0.1, 0.7, 0.3], [0.2, 0.3, 0.5]]],
dtype=torch.float)
expect = [[{"label": "T", "begin": 1, "end": 3}],
[{"label": "T", "begin": 0, "end": 1}, {"label": "T", "begin": 1, "end": 3}],
[{"label": "T", "begin": 0, "end": 3}],
[{"label": "T", "begin": 0, "end": 2}]]
vocabulary = LabelVocabulary([["B-T", "B-T", "B-T", "I-T", "I-T", "O"]],
padding=LabelVocabulary.PADDING)
b_index = vocabulary.index("B-T")
ASSERT.assertEqual(0, b_index)
i_index = vocabulary.index("I-T")
ASSERT.assertEqual(1, i_index)
o_index = vocabulary.index("O")
ASSERT.assertEqual(2, o_index)
spans = BIO.decode(batch_sequence_logits=batch_sequence_logits,
mask=None,
vocabulary=vocabulary)
ASSERT.assertListEqual(expect, spans)
def test_decode_decode_label_index_to_span():
"""
测试解码 golden label index
:return:
"""
vocabulary = LabelVocabulary([["B-T", "B-T", "B-T", "I-T", "I-T", "O"]],
padding=LabelVocabulary.PADDING)
b_index = vocabulary.index("B-T")
ASSERT.assertEqual(0, b_index)
i_index = vocabulary.index("I-T")
ASSERT.assertEqual(1, i_index)
o_index = vocabulary.index("O")
ASSERT.assertEqual(2, o_index)
golden_labels = torch.tensor([[0, 1, 2, 0],
[2, 0, 1, 1]])
expect = [[{"label": "T", "begin": 0, "end": 2}, {"label": "T", "begin": 3, "end": 4}],
[{"label": "T", "begin": 1, "end": 4}]]
spans = BIO.decode_label_index_to_span(batch_sequence_label_index=golden_labels,
mask=None,
vocabulary=vocabulary)
ASSERT.assertListEqual(expect, spans)
def test_label_f1_metric_with_mask():
"""
测试 label f1 metric
"""
predictions = torch.tensor([0, 1, 2, 3])
gold_labels = torch.tensor([0, 0, 0, 2])
mask = torch.tensor([1, 1, 1, 0], dtype=torch.long)
labels = [0, 1, 2, 3]
f1_metric = LabelF1Metric(labels=labels, label_vocabulary=None)
metrics = f1_metric(prediction_labels=predictions,
gold_labels=gold_labels,
mask=mask)
logging.debug(json2str(metrics))
ASSERT.assertEqual((len(labels) + 1) * 3, len(metrics))
precision_0 = metrics[f"{LabelF1Metric.PRECISION}-0"]
recall_0 = metrics[f"{LabelF1Metric.RECALL}-0"]
f1_0 = metrics[f"{LabelF1Metric.F1}-0"]
expect_precision_0 = 1. / 1.
ASSERT.assertAlmostEqual(expect_precision_0, precision_0)
expect_recall_0 = 1. / 3.
ASSERT.assertAlmostEqual(expect_recall_0, recall_0)
expect_f1_0 = 2. * expect_precision_0 * expect_recall_0 / (
expect_precision_0 + expect_recall_0)
ASSERT.assertAlmostEqual(expect_f1_0, f1_0)
expect_precision_overall = 1. / 3.
expect_recall_overall = 1. / 3.
precision_overall = metrics[LabelF1Metric.PRECISION_OVERALL]
recall_overall = metrics[LabelF1Metric.RECALL_OVERALL]
ASSERT.assertAlmostEqual(expect_precision_overall, precision_overall)
ASSERT.assertAlmostEqual(expect_recall_overall, recall_overall)
def test_attention_seq2vec_no_mask(inputs):
"""
测试 attention seq2vec
:return:
"""
sequence, mask = inputs
encoder = AttentionSeq2Vec(input_size=2,
query_hidden_size=3,
value_hidden_size=None)
encoder.wk.weight = Parameter(FloatTensor([
[0.1, 0.2],
[0.3, 0.4],
[0.5, 0.6]
]))
encoder.wk.bias = Parameter(FloatTensor([0.2, 0.4, 0.6]))
encoder.attention.weight = Parameter(FloatTensor(
[
[0.6, 0.2, 7]
]
))
vec = encoder(sequence=sequence, mask=None)
print(vec)
ASSERT.assertEqual((2, 2), vec.size())
expect = torch.tensor([[4.8455, 5.2867],
[5.7232, 3.6037]])
vec1d = vec.view(-1).tolist()
expect1d = expect.view(-1).tolist()
for expect_data, vec_data in zip(expect1d, vec1d):
ASSERT.assertAlmostEqual(expect_data, vec_data, delta=1e-4)
def test_graph_attention_layer():
torch.manual_seed(7)
torch.cuda.manual_seed_all(7)
in_features = 2
out_features = 4
gat_layer = GraphAttentionLayer(in_features=in_features,
out_features=out_features,
dropout=0.0,
alpha=0.1)
nodes = torch.tensor([[[0.1, 0.2], [0.3, 0.4], [0.5, 0.6]],
[[0.7, 0.8], [0.9, 0.10], [0.11, 0.12]]],
dtype=torch.float)
adj = torch.tensor(
[[[0, 1, 0], [1, 0, 0], [0, 0, 0]], [[0, 1, 1], [1, 0, 1], [1, 1, 0]]],
dtype=torch.long)
outputs: torch.Tensor = gat_layer(input=nodes, adj=adj)
expect_size = (nodes.size(0), nodes.size(1), out_features)
ASSERT.assertEqual(expect_size, outputs.size())
# 下面的 expect 是从原论文中测试得到的结果，直接拿来用
expect = torch.tensor([[[0.2831, 0.3588, -0.5131, -0.2058],
[0.1606, 0.1292, -0.2264, -0.0951],
[0.2831, 0.3588, -0.5131, -0.2058]],
[[-0.0748, 0.5025, -0.3840, -0.1192],
[0.2959, 0.4624, -0.6123, -0.2405],
[0.1505, 0.8668, -0.8609, -0.3059]]],
dtype=torch.float)
ASSERT.assertTrue(
tensor_util.is_tensor_equal(expect, outputs, epsilon=1e-4))
def test_word_lstm_cell_with_bias():
"""
测试 WordLSTMCell
:return:
"""
input_size = 2
hidden_size = 3
word_lstm_cell = WordLSTMCell(input_size=input_size,
hidden_size=hidden_size,
bias=True)
value = list()
for i in range(input_size):
value.append([
j * 0.37
for j in range(i * hidden_size * 3, (i + 1) * hidden_size * 3)
])
with torch.no_grad():
word_lstm_cell.weight_ih.copy_(torch.tensor(value, dtype=torch.float))
torch.nn.init.constant(word_lstm_cell.bias, val=1.0)
word_input = torch.tensor([[0.2, 0.4]], dtype=torch.float)
h = torch.tensor([[0.2, 0.11, 0.15]], dtype=torch.float)
c = torch.tensor([[0.5, 0.6, 0.7]], dtype=torch.float)
output_c = word_lstm_cell(input_=word_input, hx=(h, c))
expect_size = (1, hidden_size)
ASSERT.assertEqual(expect_size, output_c.size())
expect_output_c = [1.4231, 1.5257, 1.6372]
for e_i, i in zip(expect_output_c, output_c[0].tolist()):
ASSERT.assertAlmostEqual(e_i, i, places=3)
def test_decode_one_sequence_logits_to_label():
"""
测试 decode sequence label
:return:
"""
sequence_logits_list = list()
expect_list = list()
sequence_logits = torch.tensor([[0.2, 0.3, 0.4], [0.7, 0.2, 0.3], [0.2, 0.3, 0.1]],
dtype=torch.float) # O B I 正常
expect = ["O", "B-T", "I-T"]
sequence_logits_list.append(sequence_logits)
expect_list.append(expect)
sequence_logits = torch.tensor([[0.9, 0.3, 0.4], [0.2, 0.8, 0.3], [0.2, 0.3, 0.1]],
dtype=torch.float)
expect = ["B-T", "I-T", "I-T"]
sequence_logits_list.append(sequence_logits)
expect_list.append(expect)
sequence_logits = torch.tensor([[0.9, 0.3, 0.4], [0.2, 0.8, 0.3], [0.2, 0.3, 0.9]],
dtype=torch.float)
expect = ["B-T", "I-T", "O"]
sequence_logits_list.append(sequence_logits)
expect_list.append(expect)
vocabulary = LabelVocabulary([["B-T", "B-T", "B-T", "I-T", "I-T", "O"]],
padding=LabelVocabulary.PADDING)
b_index = vocabulary.index("B-T")
ASSERT.assertEqual(0, b_index)
i_index = vocabulary.index("I-T")
ASSERT.assertEqual(1, i_index)
o_index = vocabulary.index("O")
ASSERT.assertEqual(2, o_index)
for sequence_logits, expect in zip(sequence_logits_list, expect_list):
sequence_label, sequence_label_indices = BIO.decode_one_sequence_logits_to_label(
sequence_logits=sequence_logits,
vocabulary=vocabulary)
ASSERT.assertListEqual(sequence_label, expect)
expect_indices = [vocabulary.index(label) for label in expect]
ASSERT.assertListEqual(sequence_label_indices, expect_indices)
def test_vocabulary():
"""
:return:
"""
batch_tokens = [["我", "和", "你"], ["在", "我"]]
vocabulary = Vocabulary(batch_tokens,
padding="",
unk="",
special_first=True,
min_frequency=1,
max_size=None)
ASSERT.assertEqual(vocabulary.size, 4)
ASSERT.assertTrue(not vocabulary.padding)
ASSERT.assertTrue(not vocabulary.unk)
ASSERT.assertEqual(vocabulary.index("我"), 0)
ASSERT.assertEqual(vocabulary.index("和"), 1)
