def test_padding(small_dataset, model):
# mAP should be similar after padding to a square
preds_normal = []
preds_padded = []
for (fpath, img) in small_dataset:
response = get_prediction(fpath, model)
result = [box for box in response.json()[0]]
preds_normal.append(result)
tmp_img = Image.open(fpath)
w, h = tmp_img.size
target = max((w, h))
dw = (target - w) / 2
dh = (target - h) / 2
tmp_img = pad(tmp_img, (target, target))
with tempfile.NamedTemporaryFile(suffix=".jpg") as tmp:
padded_path = tmp.name
tmp_img.save(padded_path)
response = get_prediction(padded_path, model)
result = [box for box in response.json()[0]]
# half every box:
for box in result:
box["x1"] -= dw
box["y1"] -= dh
box["x2"] -= dw
box["y2"] -= dh
preds_padded.append(result)
mAP_normal = small_dataset.eval(preds_normal)
mAP_padded = small_dataset.eval(preds_padded)
print(f"Padding diff: {abs(mAP_normal - mAP_padded)}")
assert abs(mAP_normal - mAP_padded) < 0.05 # 5% diff seems acceptable
def test_padding(small_dataset, model):
# mAP should be similar after padding to a square
preds_normal = []
preds_padded = []
for (fpath, img) in small_dataset:
response = get_prediction(fpath, model)
result = [ann for ann in response.json()[0]]
preds_normal.append(result)
tmp_img = Image.open(fpath)
w, h = tmp_img.size
target = max((w, h))
dw = (target - w) / 2
dh = (target - h) / 2
tmp_img = pad(tmp_img, (target, target))
with tempfile.NamedTemporaryFile(suffix=".jpg") as tmp:
padded_path = tmp.name
tmp_img.save(padded_path)
response = get_prediction(padded_path, model)
result = [ann for ann in response.json()[0]]
# unpad every coord:
for ann in result:
ann["segmentation"]["segmentation"] = [
p - dw if i % 2 == 0 else p - dh
for i, p in enumerate(ann["segmentation"]["segmentation"])
]
preds_padded.append(result)
mAP_normal = small_dataset.eval(preds_normal)
mAP_padded = small_dataset.eval(preds_padded)
print(f"Padding diff: {abs(mAP_normal - mAP_padded)}")
assert abs(mAP_normal - mAP_padded) < 0.07 # 5% diff seems acceptable
def image_to_unicode(im: Image, invert: bool = False) -> str:
char_for = dict_unicode_inverted if invert else dict_unicode
width = im.width
height = im.height + (im.height % 2)
padded_im = pad(image=im, size=(width, height))
a = padded_im.load()
output_rows = []
for r in range(0, height, 2):
char_list = [char_for[(a[c, r], a[c, r + 1])] for c in range(width)]
row = "".join(char_list)
output_rows.append(row)
output_str = "\n".join(output_rows)
return output_str
def test_padding(small_dataset, model):
# mAP should be similar after padding to a square
preds_normal = []
preds_padded = []
for fpath, img in small_dataset:
response = get_prediction(fpath, model)
preds_normal.append(response.json()[0])
tmp_img = Image.open(fpath)
w, h = tmp_img.size
target = max((w, h))
dw = (target - w) / 2
dh = (target - h) / 2
tmp_img = pad(tmp_img, (target, target))
with tempfile.NamedTemporaryFile(suffix=".jpg") as tmp:
padded_path = tmp.name
tmp_img.save(padded_path)
response = get_prediction(padded_path, model)
preds_padded.append(response.json()[0])
score_normal = small_dataset.eval(preds_normal)
score_padded = small_dataset.eval(preds_padded)
print(f"Padding diff: {abs(score_normal - score_padded)}")
assert abs(score_normal - score_padded) < 0.1 # 10% diff seems acceptable