def forward(self, images, image_scales, transitions=None):
feature_list = self.encoder(images, image_scales)
image_features = feature_list[0]
assert len(
feature_list) == 1, 'current model only support batch size 1'
sample = Sample()
sample.dataset_name = "coco"
sample.dataset_type = "test"
sample.image_feature_0 = image_features
# it seems answers work as a place holder here
# hence, it does not matter what it's size is
sample.answers = torch.zeros((1, 10), dtype=torch.long)
sample_list = SampleList([sample])
sample_list = sample_list.to(device)
# set_trace()
if transitions is not None:
sample_list.transitions = transitions
output = self.decoder(sample_list)
tokens = output['captions']
caption = tokens.tolist()[0]
caption = self.decoder.caption_processor(caption)['caption']
return caption
def predict(self, url, feat_name, get_features=False):
with torch.no_grad():
detectron_features = get_detectron_features([url],
self.detection_model,
False, feat_name,
self.cuda_device)
# returns a single-element list
detectron_features = detectron_features[0]
sample = Sample()
sample.dataset_name = "coco"
sample.dataset_type = "test"
sample.image_feature_0 = detectron_features
sample.answers = torch.zeros((5, 10), dtype=torch.long)
sample_list = SampleList([sample])
sample_list = sample_list.to(self.cuda_device)
tokens = self.caption_model(sample_list)["captions"]
gc.collect()
torch.cuda.empty_cache()
if not get_features:
return tokens
else:
return tokens, detectron_features
def test_nucleus_sampling(self):
vocab = text_utils.VocabFromText(self.VOCAB_EXAMPLE_SENTENCES)
model_config = self.config.model_attributes.butd
model = TestDecoderModel(model_config, vocab)
model.build()
model.to("cuda")
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
expected_tokens = [
1.0000e+00, 2.9140e+03, 5.9210e+03, 2.2040e+03, 5.0550e+03,
9.2240e+03, 4.5120e+03, 1.8200e+02, 3.6490e+03, 6.4090e+03,
2.0000e+00
]
self.assertEqual(tokens[0].tolist(), expected_tokens)
def test_caption_bleu4(self):
path = os.path.join(
os.path.abspath(__file__),
"../../../pythia/common/defaults/configs/datasets/captioning/coco.yml",
)
with open(os.path.abspath(path)) as f:
config = yaml.load(f, Loader=yaml.FullLoader)
config = ConfigNode(config)
captioning_config = config.dataset_attributes.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.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
self.assertAlmostEqual(
caption_bleu4.calculate(expected, predicted).item(), 0.3928, 4)
def get_item(self, idx):
data = self.vqamb_data[idx]
current_sample = Sample()
# store queston and image id
current_sample.img_id = data['id']
current_sample.qa_id = data['qa_id']
# process question
question = data["question"]
tokens = tokenize(question, remove=["?"], keep=["'s"])
processed = self.text_processor({"tokens": tokens})
current_sample.text = processed["text"]
# process answers
processed = self.answer_processor({"answers": [data['answer']]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"][1:] # remove unknown index
# Detectron features ----------------
# TODO: read in detectron image instead if detectron is to be built
detectron_path = self.detectron_folder + str(data['id'])
if self.config.spatial:
point = data['point']
# current_sample.point = point
detectron_path += ',' + str(point['x']) + ',' + str(point['y'])
detectron_path += '.pt'
detectron_feat = torch.load(detectron_path, map_location=torch.device('cpu'))
# Pad features to fixed length
if self.config.pad_detectron:
if detectron_feat.shape[0] > 100:
detectron_feat = detectron_feat[:100]
elif detectron_feat.shape[0] < 100:
pad = torch.zeros(100 - detectron_feat.shape[0], detectron_feat.shape[1])
detectron_feat = torch.cat([detectron_feat, pad], dim=0)
current_sample.image_feature_0 = detectron_feat
# ---------------------------------------------
return current_sample
def predict(self, url):
with torch.no_grad():
detectron_features = self.get_detectron_features(url)
sample = Sample()
sample.dataset_name = "coco"
sample.dataset_type = "test"
sample.image_feature_0 = detectron_features
sample.answers = torch.zeros((5, 10), dtype=torch.long)
sample_list = SampleList([sample])
sample_list = sample_list.to("cuda")
tokens = self.pythia_model(sample_list)["captions"]
gc.collect()
torch.cuda.empty_cache()
return tokens
def get_item(self, idx):
data = self.questions[idx]
# Each call to get_item 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 = Sample()
image_id = self.annotations[idx][0]
image_folder = image_id.split('_')[0]
caption = self.annotations[idx][1]
tokens = tokenize(caption)
tokens = ['<s>'] + tokens + ['</s>']
# use text_processor to process caption
# pad sequence, convert token to indices and add SOS, EOS token
# text_processor already contains a pre-processor to tokenize caption
caption_p = self.text_processor({'tokens': tokens})
sample.text = caption_p['text']
sample.caption_len = torch.tensor(len(tokens), dtype=torch.int)
# sample.target = caption_p['text']
sample.answers = torch.stack([caption_p['text']])
# generate image features
image_path = os.path.join(self.image_dir, image_folder, image_id)
image, image_scale = self._image_transform(image_path)
with torch.no_grad():
image_features = self.feature_extractor([image], [image_scale])
image_features = image_features[0]
sample.image_feature_0 = image_features.cpu()
return sample
def get_item(self, idx):
data = self.vqamb_data[idx]
current_sample = Sample()
# store queston and image id
current_sample.img_id = data['id']
# current_sample.qa_id = data['qa_id']
# store points
current_sample.point = data['point'] # data['points']
bbox = data['bbox']
current_sample.gt_bbox = torch.Tensor([bbox['x'], bbox['y'], bbox['x'] + bbox['w'], bbox['y'] + bbox['h']])
# process question
question = data["pt_question"]
tokens = tokenize(question, remove=["?"], keep=["'s"])
processed = self.text_processor({"tokens": tokens})
current_sample.text = processed["text"]
# process answers
processed = self.answer_processor({"answers": [data['ans']]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"][1:] # remove unknown index
# Detectron features ----------------
# TODO: read in detectron image instead if detectron is to be built
detectron_path = self.detectron_folder + str(data['id'])
point = data['point'] # point = data['points'][0]
if 'pt' in self.detectron_folder:
detectron_path += ',' + str(point['x']) + ',' + str(point['y'])
detectron_path += '.pt'
detectron_feat = torch.load(detectron_path, map_location=torch.device('cpu'))
# Pad features to fixed length
if self.config.pad_detectron:
if detectron_feat.shape[0] > 100:
detectron_feat = detectron_feat[:100]
elif detectron_feat.shape[0] < 100:
pad = torch.zeros(100 - detectron_feat.shape[0], detectron_feat.shape[1])
detectron_feat = torch.cat([detectron_feat, pad], dim=0)
current_sample.image_feature_0 = detectron_feat
# ---------------------------------------------
# read in bounding boxes (hardcoded for now)
bbox_path = ''
bbox_path += str(data['id']) + ',' + str(point['x']) + ',' + str(point['y']) + '.pt'
bboxes = torch.load(bbox_path, map_location=torch.device('cpu'))
if bboxes.shape[0] > 100:
bboxes = bboxes[:100]
elif bboxes.shape[0] < 100:
pad = torch.zeros(100 - bboxes.shape[0], bboxes.shape[1])
bboxes = torch.cat([bboxes, pad], dim=0)
current_sample.pt_bbox = bboxes
# read in image bounding boxes
bbox_path = ''
bbox_path += str(data['id']) + '.pt' # + ',' + str(point['x']) + ',' + str(point['y']) + '.pt'
bboxes = torch.load(bbox_path, map_location=torch.device('cpu'))
if bboxes.shape[0] > 100:
bboxes = bboxes[:100]
elif bboxes.shape[0] < 100:
pad = torch.zeros(100 - bboxes.shape[0], bboxes.shape[1])
bboxes = torch.cat([bboxes, pad], dim=0)
current_sample.img_bbox = bboxes
# Context features --------------------
if self.config.use_context:
context_path = self.context_folder + str(data['id'])
context_path += ',' + str(point['x']) + ',' + str(point['y'])
context_path += '.pt'
context_feat = torch.load(context_path, map_location=torch.device('cpu'))
context_feat = context_feat.squeeze()
orig_dim = context_feat.shape[0]
if self.config.pad_context:
if context_feat.shape[0] > 100:
context_feat = context_feat[:100]
elif context_feat.shape[0] < 100:
pad = torch.zeros(100 - context_feat.shape[0], context_feat.shape[1])
context_feat = torch.cat([context_feat, pad], dim=0)
current_sample.context_feature_0 = context_feat
# ---------------------------------------------
return current_sample
def get_item(self, idx):
data = self.vqamb_data[idx]
current_sample = Sample()
# store queston and image id
current_sample.img_id = data['id']
current_sample.qa_id = data['qa_index']
# store points
current_sample.points = data['points']
obj = data['all_objs'][0]
xmin, ymin, xmax, ymax = obj['x'], obj[
'y'], obj['x'] + obj['w'], obj['y'] + obj['h']
current_sample.gt_bbox = torch.Tensor([xmin, ymin, xmax, ymax])
# process question
question = data["question"]
tokens = tokenize(question, remove=["?"])
processed = self.text_processor({"tokens": tokens})
current_sample.text = processed["text"]
# process answers
processed = self.answer_processor({"answers": data['all_ans']})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"][
1:] # remove unknown index
# Detectron features ----------------
# TODO: read in detectron image instead if detectron is to be built
detectron_path = self.detectron_folder + str(data['id'])
bbox_path = self.bbox_folder + str(data['id'])
if 'pt' in self.detectron_folder:
point = data['points'][0]
detectron_path += ',' + str(point['x']) + ',' + str(point['y'])
bbox_path += ',' + str(point['x']) + ',' + str(point['y'])
detectron_path += '.pt'
bbox_path += '.pt'
detectron_feat = torch.load(
detectron_path, map_location=torch.device('cpu')).squeeze()
# bbox_feat = torch.load(bbox_path, map_location=torch.device('cpu')).squeeze()
'''if detectron_feat.shape[0] == 2048:
detectron_feat = detectron_feat.unsqueeze(0)
bbox_feat = bbox_feat.unsqueeze(0)
'''
'''
if self.config.grid:
detectron_feat = detectron_feat.view(detectron_feat.shape[0], -1).T
'''
# x_down = max(int(round(pt['x']/600)), 18)
# y_down = int(round(pt['y']/800), 25)
# preproessing for grid features only
# detectron_feat = detectron_feat.view(detectron_feat.shape[0], -1).T
# Pad features to fixed length
if self.config.grid:
MAX_FEAT = 608
else:
MAX_FEAT = 100
if self.config.pad_detectron:
if detectron_feat.shape[0] > MAX_FEAT:
detectron_feat = detectron_feat[:MAX_FEAT]
# bbox_feat = bbox_feat[:MAX_FEAT]
elif detectron_feat.shape[0] < MAX_FEAT:
pad = torch.zeros(MAX_FEAT - detectron_feat.shape[0],
detectron_feat.shape[1])
detectron_feat = torch.cat([detectron_feat, pad], dim=0)
pad = torch.zeros(MAX_FEAT - bbox_feat.shape[0],
bbox_feat.shape[1])
bbox_feat = torch.cat([bbox_feat, pad], dim=0)
'''
else:
if detectron_feat.dim() > 1:
detectron_feat = torch.zeros(2048)
'''
# current_sample.bbox = bbox_feat
current_sample.image_feature_0 = detectron_feat
# ---------------------------------------------
return current_sample
def get_item(self, idx):
data = self.objpart_data[idx]
current_sample = Sample()
# store queston and image id
current_sample.img_id = data['id']
# current_sample.qa_id = data['qa_id']
if data['ans'] == 'part':
current_sample.part = 1
else:
current_sample.part = 0
# store points
current_sample.point = data['point']
# process question
question = data["question"]
tokens = tokenize(question, remove=["?"])
processed = self.text_processor({"tokens": tokens})
current_sample.text = processed["text"]
# process answers
processed = self.answer_processor({"answers": [data['ans']]})
current_sample.answers = processed["answers"]
current_sample.targets = processed["answers_scores"][
1:] # remove unknown index
# Detectron features ----------------
# TODO: read in detectron image instead if detectron is to be built
detectron_path = self.detectron_folder + str(data['id'])
if 'pt' in self.detectron_folder: # hacky way of assessing point supervision
point = data['point']
detectron_path += ',' + str(point['x']) + ',' + str(point['y'])
detectron_path += '.pt'
detectron_feat = torch.load(
detectron_path, map_location=torch.device('cpu')).squeeze()
# hardcode bounding box and read it
# x_down = max(int(round(pt['x']/600)), 18)
# y_down = int(round(pt['y']/800), 25)
# preproessing for grid features only
# detectron_feat = detectron_feat.view(detectron_feat.shape[0], -1).T
# Pad features to fixed length
MAX_FEAT = 100
if self.config.pad_detectron:
if detectron_feat.shape[0] > MAX_FEAT:
detectron_feat = detectron_feat[:MAX_FEAT]
elif detectron_feat.shape[0] < MAX_FEAT:
pad = torch.zeros(MAX_FEAT - detectron_feat.shape[0],
detectron_feat.shape[1])
detectron_feat = torch.cat([detectron_feat, pad], dim=0)
'''
else:
if detectron_feat.dim() > 1:
detectron_feat = torch.zeros(2048)
'''
current_sample.image_feature_0 = detectron_feat
# ---------------------------------------------
return current_sample
