Add normalization to the positions of the detected objects
Browse files- vqa-lxmert.py +17 -15
vqa-lxmert.py
CHANGED
@@ -77,7 +77,7 @@ class VqaV2Lxmert(datasets.GeneratorBasedBuilder):
|
|
77 |
"question_id": datasets.Value("int32"),
|
78 |
"image_id": datasets.Value("string"),
|
79 |
"features": datasets.Array2D(_SHAPE_FEATURES, dtype="float32"),
|
80 |
-
"
|
81 |
"answer_type": datasets.Value("string"),
|
82 |
"label": datasets.Sequence(
|
83 |
{
|
@@ -117,22 +117,24 @@ class VqaV2Lxmert(datasets.GeneratorBasedBuilder):
|
|
117 |
reader = csv.DictReader(f, FIELDNAMES, delimiter="\t")
|
118 |
for i, item in enumerate(reader):
|
119 |
features = {}
|
120 |
-
|
121 |
-
|
122 |
-
num_boxes =
|
123 |
-
|
124 |
-
(
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
("boxes", (num_boxes, 4), np.float32),
|
129 |
-
("features", (num_boxes, -1), np.float32),
|
130 |
-
]
|
131 |
-
for key, shape, dtype in decode_config:
|
132 |
-
features[key] = np.frombuffer(base64.b64decode(item[key]), dtype=dtype).reshape(shape)
|
133 |
id2features[item["img_id"]] = features
|
134 |
return id2features
|
135 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
136 |
def _generate_examples(self, filepath, imgfeat):
|
137 |
""" Yields examples as (key, example) tuples."""
|
138 |
id2features = self._load_features(imgfeat)
|
@@ -148,7 +150,7 @@ class VqaV2Lxmert(datasets.GeneratorBasedBuilder):
|
|
148 |
"question_id": d["question_id"],
|
149 |
"image_id": d["img_id"],
|
150 |
"features": img_features["features"],
|
151 |
-
"
|
152 |
"answer_type": d["answer_type"],
|
153 |
"label": {
|
154 |
"ids": ids,
|
|
|
77 |
"question_id": datasets.Value("int32"),
|
78 |
"image_id": datasets.Value("string"),
|
79 |
"features": datasets.Array2D(_SHAPE_FEATURES, dtype="float32"),
|
80 |
+
"normalized_boxes": datasets.Array2D(_SHAPE_BOXES, dtype="float32"),
|
81 |
"answer_type": datasets.Value("string"),
|
82 |
"label": datasets.Sequence(
|
83 |
{
|
|
|
117 |
reader = csv.DictReader(f, FIELDNAMES, delimiter="\t")
|
118 |
for i, item in enumerate(reader):
|
119 |
features = {}
|
120 |
+
img_h = int(item["img_h"])
|
121 |
+
img_w = int(item["img_w"])
|
122 |
+
num_boxes = int(item["num_boxes"])
|
123 |
+
features["features"] = np.frombuffer(base64.b64decode(item["features"]), dtype=np.float32).reshape(
|
124 |
+
(num_boxes, -1)
|
125 |
+
)
|
126 |
+
boxes = np.frombuffer(base64.b64decode(item["boxes"]), dtype=np.float32).reshape((num_boxes, 4))
|
127 |
+
features["normalized_boxes"] = self._normalize_boxes(boxes, img_h, img_w)
|
|
|
|
|
|
|
|
|
|
|
128 |
id2features[item["img_id"]] = features
|
129 |
return id2features
|
130 |
|
131 |
+
def _normalize_boxes(self, boxes, img_h, img_w):
|
132 |
+
""" Normalize the input boxes given the original image size."""
|
133 |
+
normalized_boxes = boxes.copy()
|
134 |
+
normalized_boxes[:, (0, 2)] /= img_w
|
135 |
+
normalized_boxes[:, (1, 3)] /= img_h
|
136 |
+
return normalized_boxes
|
137 |
+
|
138 |
def _generate_examples(self, filepath, imgfeat):
|
139 |
""" Yields examples as (key, example) tuples."""
|
140 |
id2features = self._load_features(imgfeat)
|
|
|
150 |
"question_id": d["question_id"],
|
151 |
"image_id": d["img_id"],
|
152 |
"features": img_features["features"],
|
153 |
+
"normalized_boxes": img_features["normalized_boxes"],
|
154 |
"answer_type": d["answer_type"],
|
155 |
"label": {
|
156 |
"ids": ids,
|