def log_pdf(doc, score, scores, predict_text, answer_text):
fname = get_pdf_path(doc.slug)
try:
pdf = pdfplumber.open(fname)
except Exception:
# If the file's not there, that's fine -- we use available PDFs to
# define what to see
print(f"Cannot open pdf {fname}")
return
print(f"Rendering output for {fname}")
# Get the correct answers: find the indices of the token(s) labelled 1
target_idx = [idx for (idx, val) in enumerate(doc.labels) if val == 1]
# Draw the machine output: get a score for each token
page_images = []
for pagenum, page in enumerate(pdf.pages):
im = page.to_image(resolution=300)
# training data has 0..1 for page range (see create-training-data.py)
num_pages = len(pdf.pages)
if num_pages > 1:
current_page = pagenum / float(num_pages - 1)
else:
current_page = 0.0
# Draw guesses
rel_score = scores / score
page_match = np.isclose(doc.tokens["page"], current_page)
for token in doc.tokens[page_match & (rel_score > 0.5)].itertuples():
if rel_score[token.Index] == 1:
w = 5
s = "magenta"
elif rel_score[token.Index] >= 0.75:
w = 3
s = "red"
else:
w = 1
s = "red"
im.draw_rect(docrow_to_bbox(token), stroke=s, stroke_width=w, fill=None)
# Draw target tokens
target_toks = [
doc.tokens.iloc[i]
for i in target_idx
if np.isclose(doc.tokens.iloc[i]["page"], current_page)
]
rects = [docrow_to_bbox(t) for t in target_toks]
im.draw_rects(rects, stroke="blue", stroke_width=3, fill=None)
page_images.append(wandb.Image(im.annotated, caption="page " + str(pagenum)))
# get best matching score of any token in the training data
match = doc.tokens[SINGLE_CLASS_PREDICTION].max()
caption = (
f"{doc.slug} guessed:{predict_text} answer:{answer_text} match:{match:.2f}"
)
verdict = dollar_match(predict_text, answer_text)
return verdict, caption, page_images
def test_docrow_to_bbox(x0, y0, x1, y1, mh):
t = BoundingBox(x0=x0, x1=x1, y0=y0, y1=y1)
bbox0 = docrow_to_bbox(t, min_height=None)
bbox1 = docrow_to_bbox(t)
bbox2 = docrow_to_bbox(t, min_height=mh)
for box in (bbox0, bbox1, bbox2):
assert box.x0 == Decimal(x0)
assert box.x1 == Decimal(x1)
assert box.y1 == Decimal(y1)
assert bbox0.y0 == Decimal(y0)
# Floating point arithmetic, yo.
assert bbox1.y1 - bbox1.y0 >= 10 or isclose(bbox1.y1 - bbox1.y0, 10)
assert bbox2.y1 - bbox2.y0 >= mh or isclose(bbox2.y1 - bbox2.y0, mh)
def render_tokenized_pdf(doc):
fname = get_pdf_path(doc.slug)
try:
pdf = pdfplumber.open(fname)
except Exception:
# If the file's not there, that's fine -- we use available PDFs to
# define what to see
print(f"Cannot open pdf {fname}")
return
page_images = [{
"image": page.to_image(resolution=300),
"rects": [],
"lines": []
} for page in pdf.pages]
for token in doc.tokens.itertuples():
page_num = int(token.page)
if page_num < len(page_images):
page_images[page_num]["rects"].append(docrow_to_bbox(token))
for indices in np.argwhere(doc.adjacency_matrix):
first_index, second_index = indices
if first_index != second_index:
first_token = doc.tokens.iloc[first_index]
second_token = doc.tokens.iloc[second_index]
page = int(first_token.page)
line = (
(Decimal(float(first_token.x0)),
Decimal(float(first_token.y1))),
(Decimal(float(second_token.x0)),
Decimal(float(second_token.y1))),
)
page_images[page_num]["lines"].append(line)
for page in page_images:
image, rects, lines = page["image"], page["rects"], page["lines"]
image.draw_rects(rects, stroke="blue", stroke_width=3, fill=None)
print(f"first lines = {lines[:5]}")
image.draw_lines(lines, stroke="green", stroke_width=3)
return [page["image"] for page in page_images]