def eval_cs2satt(path="cs2satt_good.pt"):
net = ConvSeq2SeqAtt(nclass=100).float()
optimizer = optim.Adam(net.parameters())
checkpoint = torch.load(path)
net.load_state_dict(checkpoint["model_state_dict"])
optimizer.load_state_dict(checkpoint["optimizer_state_dict"])
epoch = checkpoint["epoch"]
loss = checkpoint["loss"]
print(f"model current epoch: {epoch} with loss: {loss}")
net.eval()
with torch.no_grad():
while 1:
data = next(dataset)
images = data["the_inputs"]
labels = data["the_labels"]
input_length = data["input_length"]
label_length = data["label_length"]
preds = net(images.float(), labels, 0).detach().permute(1, 0, 2)
for i in range(len(preds)):
print("labels", labels[i])
print("preds", preds[i].argmax(1))
print(labels_to_text(preds[i, :, :].argmax(1), string.printable))
cv2.imshow("im", images[i].permute(1, 2, 0).numpy())
cv2.waitKey(0)
cv2.destroyAllWindows()
def decode_batch(test_func, word_batch):
out = test_func([word_batch])[0]
ret = []
for j in range(out.shape[0]):
out_best = list(np.argmax(out[j, 2:], 1))
out_best = [k for k, g in itertools.groupby(out_best)]
outstr = labels_to_text(out_best, ALPHABET)
ret.append(outstr)
return ret
def show_edit_distance(self, num):
num_left = num
mean_norm_ed = 0.0
mean_ed = 0.0
while num_left > 0:
word_batch = next(self.text_img_gen)[0]
num_proc = min(word_batch["the_input"].shape[0], num_left)
decoded_res = decode_batch(
self.test_func, word_batch["the_input"][0:num_proc]
)
for j in range(num_proc):
pred = decoded_res[j].strip()
truth = labels_to_text(word_batch["the_labels"][j], ALPHABET)
edit_dist = editdistance.eval(pred, truth)
mean_ed += float(edit_dist)
mean_norm_ed += float(edit_dist) / max(len(truth), len(pred))
num_left -= num_proc
mean_norm_ed = mean_norm_ed / num
mean_ed = mean_ed / num
print(
"\nOut of %d samples: Mean edit distance: "
"%.3f / Mean normalized edit distance: %0.3f" % (num, mean_ed, mean_norm_ed)
)
def calculate_map(all_detections,
all_text_preds,
all_annotations,
all_text_annots,
binary,
generator,
retinanet,
iou_threshold=0.5,
score_threshold=0.05,
max_detections=400,
save_path=None):
average_precisions = {}
cers = []
if binary:
n_classes = 1
else:
n_classes = generator.num_classes()
for label in range(n_classes): #generator.num_classes()):
false_positives = np.zeros((0, ))
true_positives = np.zeros((0, ))
scores = np.zeros((0, ))
num_annotations = 0.0
for i in range(len(generator)):
if binary:
detections = np.concatenate(
all_detections[i][:]) #all_detections[i][label]
annotations = np.concatenate(all_annotations[i][:]) #label]
text_dets = np.concatenate([
all_text_preds.get(i, {}).get(lab, {})
for lab in range(generator.num_classes())
])
text_annots = np.concatenate([
all_text_annots.get(i, {}).get(lab, {})
for lab in range(generator.num_classes())
])
else:
detections = all_detections[i][label]
annotations = all_annotations[i][label]
text_dets = all_text_preds.get(i,
{}).get(label, torch.zeros(1))
text_annots = all_text_annots[i][label]
num_annotations += annotations.shape[0]
detected_annotations = []
for j in range(len(detections)):
#for d in detections:
d = detections[j]
scores = np.append(scores, d[4])
if annotations.shape[0] == 0:
false_positives = np.append(false_positives, 1)
true_positives = np.append(true_positives, 0)
continue
overlaps = compute_overlap(np.expand_dims(d, axis=0),
annotations)
assigned_annotation = np.argmax(overlaps, axis=1)
max_overlap = overlaps[0, assigned_annotation]
if max_overlap >= iou_threshold and assigned_annotation not in detected_annotations:
false_positives = np.append(false_positives, 0)
true_positives = np.append(true_positives, 1)
if retinanet.module.binary_classifier:
pred_str = labels_to_text(text_dets[j].astype('int'),
retinanet.module.alphabet)
lab_str = labels_to_text(
text_annots[assigned_annotation[0]].astype('int'),
retinanet.module.alphabet)
else:
pred_str = labels_to_text(text_dets[j].astype('int'),
retinanet.module.alphabet)
lab_str = labels_to_text(
text_annots[assigned_annotation[0]].astype('int'),
retinanet.module.alphabet)
if len(lab_str) > 0:
cer = float(
editdistance.eval(pred_str, lab_str) /
len(lab_str))
else:
cer = 1.
cers.append(cer)
detected_annotations.append(assigned_annotation)
else:
false_positives = np.append(false_positives, 1)
true_positives = np.append(true_positives, 0)
# no annotations -> AP for this class is 0 (is this correct?)
if num_annotations == 0:
average_precisions[label] = 0, 0
continue
# sort by score
indices = np.argsort(-scores)
false_positives = false_positives[indices]
true_positives = true_positives[indices]
# compute false positives and true positives
false_positives = np.cumsum(false_positives)
true_positives = np.cumsum(true_positives)
# compute recall and precision
recall = true_positives / num_annotations
precision = true_positives / np.maximum(
true_positives + false_positives,
np.finfo(np.float64).eps)
# compute average precision
average_precision = _compute_ap(recall, precision)
average_precisions[label] = average_precision, num_annotations
if binary: #retinanet.module.binary_classifier:
for label in range(1): #generator.num_classes()):
label_name = 'Text' #generator.label_to_name(label)
print('{}: {}'.format(label_name, average_precisions[label][0]))
else:
for label in range(generator.num_classes()):
label_name = generator.label_to_name(label)
print('{}: {}'.format(label_name, average_precisions[label][0]))
mAPs = []
for k, v in average_precisions.items():
mAPs.append(v[0])
mAP = np.mean(mAPs)
print('mAP', mAP)
print("Per box CER", np.mean(cers))
cer = np.mean(cers)
return mAP, np.mean(cers)
def generate_pagexml(image_id, data, retinanet, score_threshold, nms_threshold,
dataset_val):
image_name = image_id + '.jpg'
im_file_out = 'pagexmls/' + image_name
alphabet = retinanet.alphabet
#retinanet.score_threshold = torch.tensor(score_threshold).cuda().float()
colors = get_n_random_colors(len(dataset_val.labels))
gtxml_name = os.path.join(image_name.split('/')[-1].split('.')[-2])
pxml = pagexml.PageXML()
unnormalize = UnNormalizer()
with torch.no_grad():
st = time.time()
im = data['img']
im = im.cuda().float()
print(retinanet.htr_gt_box)
if retinanet.htr_gt_box:
scores, classification, transformed_anchors, transcriptions = retinanet(
[im, data['annot']])
score_threshold = 0
else:
scores, classification, transformed_anchors, transcriptions = retinanet(
im)
n_boxes_predicted = transformed_anchors.shape[0]
print(n_boxes_predicted, "BOXES PREDICTED")
img = np.array(255 * unnormalize(data['img'][0, :, :, :])).copy()
img[img < 0] = 0
img[img > 255] = 255
img = np.transpose(img, (1, 2, 0))
img = cv2.cvtColor(img.astype(np.uint8), cv2.COLOR_BGR2RGB)
width = img.shape[1]
height = img.shape[0]
cv2.imwrite(im_file_out, img)
conf = pagexml.ptr_double()
pxml.newXml('retinanet_dets', image_name, width, height)
page = pxml.selectNth("//_:Page", 0)
reg = pxml.addTextRegion(page)
pxml.setCoordsBBox(reg, 0, 0, width, height, conf)
line = pxml.addTextLine(reg)
pxml.setCoordsBBox(line, 0, 0, width, height, conf)
words = []
for k in range(len(dataset_val.labels)):
cv2.putText(img, dataset_val.labels[k], (25, 25 + k * 15),
cv2.FONT_HERSHEY_PLAIN, 1, colors[k], 2)
transcriptions = np.argmax(transcriptions.cpu(), axis=-1)
for box_id in range(n_boxes_predicted):
# Initialize object for setting confidence values
box = {}
bbox = transformed_anchors[box_id, :]
transcription = transcriptions[box_id, :]
x1 = int(bbox[0])
y1 = int(bbox[1])
x2 = int(bbox[2])
y2 = int(bbox[3])
label_name = dataset_val.labels[int(classification[box_id])]
cv2.rectangle(img, (x1, y1), (x2, y2),
color=colors[int(classification[box_id])],
thickness=2)
# Add a text region to the Page
word = pxml.addWord(line, "ID" + str(box_id))
# Set text region bounding box with a confidence
pxml.setCoordsBBox(word, x1, y1, x2 - x1, y2 - y1, conf)
#pxml.setCoordsBBox( reg,x1, y1, x2-x1, y2-y1, conf )
#transcription = transcripts[j]
transcription = labels_to_text(transcription, alphabet)
draw_caption(img, (x1, y1, x2, y2), transcription)
# Set the text for the text region
conf.assign(0.9)
pxml.setTextEquiv(word, transcription, conf)
# Add property to text region
pxml.setProperty(word, "category", label_name)
# Add a second page with a text region and specific id
#page = pxml.addPage("example_image_2.jpg", 300, 300)
#reg = pxml.addTextRegion( page, "regA" )
#pxml.setCoordsBBox( reg, 15, 12, 76, 128 )
words.append(word)
words = pxml.select('//_:Word')
order, groups = pxml.getLeftRightTopBottomReadingOrder(
words, fake_baseline=True, max_horiz_iou=1, prolong_alpha=0.0)
line = pxml.selectNth('//_:TextLine')
group_idx = 0
idx_in_group = 0
for n in order:
word_idx = order.index(n)
if idx_in_group >= groups[group_idx]:
group_idx += 1
idx_in_group = 0
pxml.setProperty(words[n], 'word_idx', str(word_idx))
pxml.setProperty(words[n], "line", str(group_idx))
pxml.moveElem(words[n], line)
idx_in_group += 1
# Write XML to file
pxml.write('pagexmls/' + gtxml_name + ".xml")
cv2.imwrite(os.path.join('pred_sample_ims',
str(image_id) + '.jpg'), img)
def evaluate(generator,
retinanet,
iou_threshold=0.5,
score_threshold=0.05,
max_detections=400,
save_path=None):
""" Evaluate a given dataset using a given retinanet.
# Arguments
generator : The generator that represents the dataset to evaluate.
retinanet : The retinanet to evaluate.
iou_threshold : The threshold used to consider when a detection is positive or negative.
score_threshold : The score confidence threshold to use for detections.
max_detections : The maximum number of detections to use per image.
save_path : The path to save images with visualized detections to.
# Returns
A dict mapping class names to mAP scores.
"""
# gather all detections and annotations
all_detections, all_text_preds = _get_detections(
generator,
retinanet,
score_threshold=score_threshold,
max_detections=max_detections,
save_path=save_path)
all_annotations, all_text_annots = _get_annotations(generator)
average_precisions = {}
cers = []
for label in range(generator.num_classes()):
false_positives = np.zeros((0, ))
true_positives = np.zeros((0, ))
scores = np.zeros((0, ))
num_annotations = 0.0
for i in range(len(generator)):
detections = all_detections[i][label]
annotations = all_annotations[i][label]
text_dets = all_text_preds[i][label]
text_annots = all_text_annots[i][label]
num_annotations += annotations.shape[0]
detected_annotations = []
for j in range(len(detections)):
#for d in detections:
d = detections[j]
scores = np.append(scores, d[4])
if annotations.shape[0] == 0:
false_positives = np.append(false_positives, 1)
true_positives = np.append(true_positives, 0)
continue
overlaps = compute_overlap(np.expand_dims(d, axis=0),
annotations)
assigned_annotation = np.argmax(overlaps, axis=1)
max_overlap = overlaps[0, assigned_annotation]
if max_overlap >= iou_threshold and assigned_annotation not in detected_annotations:
false_positives = np.append(false_positives, 0)
true_positives = np.append(true_positives, 1)
pred_str = labels_to_text(text_dets[j].astype('int'),
retinanet.module.alphabet)
lab_str = labels_to_text(
text_annots[assigned_annotation[0]].astype('int'),
retinanet.module.alphabet)
if len(lab_str) > 0:
cer = float(
editdistance.eval(pred_str, lab_str) /
len(lab_str))
cers.append(cer)
print("CER", cer)
detected_annotations.append(assigned_annotation)
else:
false_positives = np.append(false_positives, 1)
true_positives = np.append(true_positives, 0)
# no annotations -> AP for this class is 0 (is this correct?)
if num_annotations == 0:
average_precisions[label] = 0, 0
continue
# sort by score
indices = np.argsort(-scores)
false_positives = false_positives[indices]
true_positives = true_positives[indices]
# compute false positives and true positives
false_positives = np.cumsum(false_positives)
true_positives = np.cumsum(true_positives)
# compute recall and precision
recall = true_positives / num_annotations
precision = true_positives / np.maximum(
true_positives + false_positives,
np.finfo(np.float64).eps)
# compute average precision
average_precision = _compute_ap(recall, precision)
average_precisions[label] = average_precision, num_annotations
print('\nmAP:')
for label in range(generator.num_classes()):
label_name = generator.label_to_name(label)
print('{}: {}'.format(label_name, average_precisions[label][0]))
print("Per box CER", np.mean(cers))
return average_precisions