def test_beam_search(self):
vocab = text_utils.VocabFromText(self.VOCAB_EXAMPLE_SENTENCES)
model_config = self.config.model_config.butd
model = TestDecoderModel(model_config, vocab)
model.build()
model.eval()
expected_tokens = {
1: [1.0, 23.0, 1.0, 24.0, 29.0, 37.0, 40.0, 17.0, 29.0, 2.0],
2: [1.0, 0.0, 8.0, 1.0, 28.0, 25.0, 2.0],
8: [1.0, 34.0, 1.0, 13.0, 1.0, 2.0],
16: [1.0, 25.0, 18.0, 2.0],
}
for batch_size in [1, 2, 8, 16]:
samples = []
for _ in range(batch_size):
sample = Sample()
sample.dataset_name = "coco"
sample.dataset_type = "test"
sample.image_feature_0 = torch.randn(100, 2048)
sample.answers = torch.zeros((5, 10), dtype=torch.long)
samples.append(sample)
sample_list = SampleList(samples)
tokens = model(sample_list)["captions"]
self.assertEqual(np.trim_zeros(tokens[0].tolist()),
expected_tokens[batch_size])
def test_caption_bleu4(self):
path = os.path.join(
os.path.abspath(__file__),
"../../../mmf/configs/datasets/coco/defaults.yaml",
)
config = load_yaml(os.path.abspath(path))
captioning_config = config.dataset_config.coco
caption_processor_config = captioning_config.processors.caption_processor
vocab_path = os.path.join(os.path.abspath(__file__), "..", "..",
"data", "vocab.txt")
caption_processor_config.params.vocab.type = "random"
caption_processor_config.params.vocab.vocab_file = os.path.abspath(
vocab_path)
caption_processor = CaptionProcessor(caption_processor_config.params)
registry.register("coco_caption_processor", caption_processor)
caption_bleu4 = metrics.CaptionBleu4Metric()
expected = Sample()
predicted = dict()
# Test complete match
expected.answers = torch.empty((5, 5, 10))
expected.answers.fill_(4)
predicted["scores"] = torch.zeros((5, 10, 19))
predicted["scores"][:, :, 4] = 1.0
self.assertEqual(
caption_bleu4.calculate(expected, predicted).item(), 1.0)
# Test partial match
expected.answers = torch.empty((5, 5, 10))
expected.answers.fill_(4)
predicted["scores"] = torch.zeros((5, 10, 19))
predicted["scores"][:, 0:5, 4] = 1.0
predicted["scores"][:, 5:, 18] = 1.0
self.assertAlmostEqual(
caption_bleu4.calculate(expected, predicted).item(), 0.3928, 4)
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
plot = sample_info["plot"]
if isinstance(plot, list):
plot = plot[0]
processed_sentence = self.text_processor({"text": plot})
current_sample.text = processed_sentence["text"]
if "input_ids" in processed_sentence:
current_sample.update(processed_sentence)
if self._use_images is True:
current_sample.image = self.image_db[idx]["images"][0]
processed = self.answer_processor({"answers": sample_info["genres"]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"]
return current_sample
def test_nucleus_sampling(self):
vocab = text_utils.VocabFromText(self.VOCAB_EXAMPLE_SENTENCES)
model_config = self.config.model_config.butd
model = TestDecoderModel(model_config, vocab)
model.build()
model.eval()
sample = Sample()
sample.dataset_name = "coco"
sample.dataset_type = "test"
sample.image_feature_0 = torch.randn(100, 2048)
sample.answers = torch.zeros((5, 10), dtype=torch.long)
sample_list = SampleList([sample])
tokens = model(sample_list)["captions"]
# these are expected tokens for sum_threshold = 0.5
# Because of a bug fix in https://github.com/pytorch/pytorch/pull/47386
# the torch.Tensor.multinomail will generate different random sequence.
# TODO: Remove this hack after OSS uses later version of PyTorch.
if LegacyVersion(torch.__version__) > LegacyVersion("1.7.1"):
expected_tokens = [1.0, 23.0, 38.0, 30.0, 5.0, 11.0, 2.0]
else:
expected_tokens = [
1.0,
29.0,
11.0,
11.0,
39.0,
10.0,
31.0,
4.0,
19.0,
39.0,
2.0,
]
self.assertEqual(tokens[0].tolist(), expected_tokens)
def __getitem__(self, idx):
data = self.questions[idx]
# Each call to __getitem__ from dataloader returns a Sample class object which
# collated by our special batch collator to a SampleList which is basically
# a attribute based batch in layman terms
current_sample = Sample()
question = data["question"]
tokens = tokenize(question, keep=[";", ","], remove=["?", "."])
processed = self.text_processor({"tokens": tokens})
current_sample.text = processed["text"]
processed = self.answer_processor({"answers": [data["answer"]]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"]
image_path = os.path.join(self.image_path, data["image_filename"])
image = np.true_divide(Image.open(image_path).convert("RGB"), 255)
image = image.astype(np.float32)
current_sample.image = torch.from_numpy(image.transpose(2, 0, 1))
return current_sample
def load_item(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
processed_caption = self.text_processor(
{"text": sample_info["captions"][0]})
current_sample.text = processed_caption["text"]
current_sample.caption_len = torch.tensor(len(
processed_caption["text"]),
dtype=torch.int)
if isinstance(sample_info["image_id"], int):
current_sample.image_id = torch.tensor(sample_info["image_id"],
dtype=torch.int)
else:
current_sample.image_id = sample_info["image_id"]
if self._use_features is True:
features = self.features_db[idx]
current_sample.update(features)
current_sample.answers = torch.stack([processed_caption["text"]])
return current_sample
def __getitem__(self, idx):
sample_info = self.annotation_db[idx]
current_sample = Sample()
plot = sample_info["plot"]
if isinstance(plot, list):
plot = plot[0]
processed_sentence = self.text_processor({"text": plot})
current_sample.text = processed_sentence["text"]
if "input_ids" in processed_sentence:
current_sample.update(processed_sentence)
if self._use_features is True:
features = self.features_db[idx]
if hasattr(self, "transformer_bbox_processor"):
features["image_info_0"] = self.transformer_bbox_processor(
features["image_info_0"])
current_sample.update(features)
processed = self.answer_processor({"answers": sample_info["genres"]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"]
return current_sample