m53297601
/
PaddleOCR
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

update infer_rec for srn

Browse Source
release/1.1
tink2123 5 years ago
parent 6832ca029f
commit bf4863c950
8 changed files with 197 additions and 161 deletions
Show Stats Download Patch File Download Diff File

41
ppocr/data/rec/dataset_traversal.py
Unescape View File

@ -40,10 +40,12 @@ class LMDBReader(object):
self.image_shape = params['image_shape']
self.loss_type = params['loss_type']
self.max_text_length = params['max_text_length']
self.num_heads = params['num_heads']
self.mode = params['mode']
self.drop_last = False
self.use_tps = False
self.num_heads = None
if "num_heads" in params:
self.num_heads = params['num_heads']
if "tps" in params:
self.ues_tps = True
self.use_distort = False
@ -134,20 +136,6 @@ class LMDBReader(object):
tps=self.use_tps,
infer_mode=True)
yield norm_img
#elif self.mode == 'eval':
# image_file_list = get_image_file_list(self.infer_img)
# for single_img in image_file_list:
# img = cv2.imread(single_img)
# if img.shape[-1]==1 or len(list(img.shape))==2:
# img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
# norm_img = process_image(
# img=img,
# image_shape=self.image_shape,
# char_ops=self.char_ops,
# tps=self.use_tps,
# infer_mode=True
# )
# yield norm_img
else:
lmdb_sets = self.load_hierarchical_lmdb_dataset()
if process_id == 0:
@ -169,14 +157,22 @@ class LMDBReader(object):
outs = []
if self.loss_type == "srn":
outs = process_image_srn(
img, self.image_shape, self.num_heads,
self.max_text_length, label, self.char_ops,
self.loss_type)
img=img,
image_shape=self.image_shape,
num_heads=self.num_heads,
max_text_length=self.max_text_length,
label=label,
char_ops=self.char_ops,
loss_type=self.loss_type)
else:
outs = process_image(
img, self.image_shape, label, self.char_ops,
self.loss_type, self.max_text_length)
img=img,
image_shape=self.image_shape,
label=label,
char_ops=self.char_ops,
loss_type=self.loss_type,
max_text_length=self.max_text_length)
if outs is None:
continue
yield outs
@ -192,8 +188,9 @@ class LMDBReader(object):
if len(batch_outs) == self.batch_size:
yield batch_outs
batch_outs = []
if len(batch_outs) != 0:
yield batch_outs
if not self.drop_last:
if len(batch_outs) != 0:
yield batch_outs
if self.infer_img is None:
return batch_iter_reader

5
ppocr/modeling/architectures/rec_model.py
Unescape View File

@ -58,7 +58,10 @@ class RecModel(object):
self.loss_type = global_params['loss_type']
self.image_shape = global_params['image_shape']
self.max_text_length = global_params['max_text_length']
self.num_heads = global_params["num_heads"]
if "num_heads" in params:
self.num_heads = global_params["num_heads"]
else:
self.num_heads = None
def create_feed(self, mode):
image_shape = deepcopy(self.image_shape)

2
ppocr/modeling/backbones/rec_resnet_vd.py
Unescape View File

@ -32,7 +32,7 @@ class ResNet():
def __init__(self, params):
self.layers = params['layers']
self.is_3x3 = True
supported_layers = [18, 34, 50, 101, 152]
supported_layers = [18, 34, 50, 101, 152, 200]
assert self.layers in supported_layers, \
"supported layers are {} but input layer is {}".format(supported_layers, self.layers)

194
ppocr/modeling/heads/rec_srn_all_head.py
View File

File diff suppressed because it is too large Load Diff

39
ppocr/modeling/losses/rec_srn_loss.py
Unescape View File

@ -35,24 +35,21 @@ class SRNLoss(object):
lbl_weight = others['lbl_weight']
casted_label = fluid.layers.cast(x=label, dtype='int64')
cost_word = fluid.layers.cross_entropy(input=word_predict, label=casted_label)
cost_gsrm = fluid.layers.cross_entropy(input=gsrm_predict, label=casted_label)
cost_vsfd = fluid.layers.cross_entropy(input=predict, label=casted_label)
#cost_word = cost_word * lbl_weight
#cost_gsrm = cost_gsrm * lbl_weight
#cost_vsfd = cost_vsfd * lbl_weight
cost_word = fluid.layers.reshape(x=fluid.layers.reduce_sum(cost_word), shape=[1])
cost_gsrm = fluid.layers.reshape(x=fluid.layers.reduce_sum(cost_gsrm), shape=[1])
cost_vsfd = fluid.layers.reshape(x=fluid.layers.reduce_sum(cost_vsfd), shape=[1])
sum_cost = fluid.layers.sum([cost_word, cost_vsfd * 2.0, cost_gsrm * 0.15])
#sum_cost = fluid.layers.sum([cost_word * 3.0, cost_vsfd, cost_gsrm * 0.15])
#sum_cost = cost_word
#fluid.layers.Print(cost_word,message="word_cost")
#fluid.layers.Print(cost_vsfd,message="img_cost")
return [sum_cost,cost_vsfd,cost_word]
#return [sum_cost, cost_vsfd, cost_word]
cost_word = fluid.layers.cross_entropy(
input=word_predict, label=casted_label)
cost_gsrm = fluid.layers.cross_entropy(
input=gsrm_predict, label=casted_label)
cost_vsfd = fluid.layers.cross_entropy(
input=predict, label=casted_label)
cost_word = fluid.layers.reshape(
x=fluid.layers.reduce_sum(cost_word), shape=[1])
cost_gsrm = fluid.layers.reshape(
x=fluid.layers.reduce_sum(cost_gsrm), shape=[1])
cost_vsfd = fluid.layers.reshape(
x=fluid.layers.reduce_sum(cost_vsfd), shape=[1])
sum_cost = fluid.layers.sum(
[cost_word, cost_vsfd * 2.0, cost_gsrm * 0.15])
return [sum_cost, cost_vsfd, cost_word]

25
ppocr/utils/character.py
Unescape View File

@ -149,38 +149,29 @@ def cal_predicts_accuracy(char_ops,
acc = acc_num * 1.0 / img_num
return acc, acc_num, img_num
def cal_predicts_accuracy_srn(char_ops,
preds,
labels,
max_text_len,
is_debug=False):
preds,
labels,
max_text_len,
is_debug=False):
acc_num = 0
img_num = 0
total_len = preds.shape[0]
img_num = int(total_len / max_text_len)
#print (img_num)
for i in range(img_num):
cur_label = []
cur_pred = []
for j in range(max_text_len):
if labels[j + i * max_text_len] != 37: #0
if labels[j + i * max_text_len] != 37: #0
cur_label.append(labels[j + i * max_text_len][0])
else:
break
if is_debug:
for j in range(max_text_len):
if preds[j + i * max_text_len] != 37: #0
cur_pred.append(preds[j + i * max_text_len][0])
else:
break
print ("cur_label: ", cur_label)
print ("cur_pred: ", cur_pred)
for j in range(max_text_len + 1):
if j < len(cur_label) and preds[j + i * max_text_len][0] != cur_label[j]:
if j < len(cur_label) and preds[j + i * max_text_len][
0] != cur_label[j]:
break
elif j == len(cur_label) and j == max_text_len:
acc_num += 1

4
tools/eval_utils/eval_rec_utils.py
Unescape View File

@ -123,8 +123,8 @@ def eval_rec_run(exe, config, eval_info_dict, mode):
def test_rec_benchmark(exe, config, eval_info_dict):
" Evaluate lmdb dataset "
eval_data_list = ['IIIT5k_3000', 'SVT', 'IC03_860', \
'IC13_857', 'IC15_1811', 'IC15_2077','SVTP', 'CUTE80']
eval_data_list = ['IIIT5k_3000', 'SVT', 'IC03_860','IC03_867', \
'IC13_857', 'IC13_1015', 'IC15_1811', 'IC15_2077','SVTP', 'CUTE80']
eval_data_dir = config['TestReader']['lmdb_sets_dir']
total_evaluation_data_number = 0
total_correct_number = 0

48
tools/infer_rec.py
Unescape View File

@ -64,7 +64,6 @@ def main():
exe = fluid.Executor(place)
rec_model = create_module(config['Architecture']['function'])(params=config)
startup_prog = fluid.Program()
eval_prog = fluid.Program()
with fluid.program_guard(eval_prog, startup_prog):
@ -86,10 +85,36 @@ def main():
for i in range(max_img_num):
logger.info("infer_img:%s" % infer_list[i])
img = next(blobs)
predict = exe.run(program=eval_prog,
feed={"image": img},
fetch_list=fetch_varname_list,
return_numpy=False)
if loss_type != "srn":
predict = exe.run(program=eval_prog,
feed={"image": img},
fetch_list=fetch_varname_list,
return_numpy=False)
else:
encoder_word_pos_list = []
gsrm_word_pos_list = []
gsrm_slf_attn_bias1_list = []
gsrm_slf_attn_bias2_list = []
encoder_word_pos_list.append(img[1])
gsrm_word_pos_list.append(img[2])
gsrm_slf_attn_bias1_list.append(img[3])
gsrm_slf_attn_bias2_list.append(img[4])
encoder_word_pos_list = np.concatenate(
encoder_word_pos_list, axis=0).astype(np.int64)
gsrm_word_pos_list = np.concatenate(
gsrm_word_pos_list, axis=0).astype(np.int64)
gsrm_slf_attn_bias1_list = np.concatenate(
gsrm_slf_attn_bias1_list, axis=0).astype(np.float32)
gsrm_slf_attn_bias2_list = np.concatenate(
gsrm_slf_attn_bias2_list, axis=0).astype(np.float32)
predict = exe.run(program=eval_prog, \
feed={'image': img[0], 'encoder_word_pos': encoder_word_pos_list,
'gsrm_word_pos': gsrm_word_pos_list, 'gsrm_slf_attn_bias1': gsrm_slf_attn_bias1_list,
'gsrm_slf_attn_bias2': gsrm_slf_attn_bias2_list}, \
fetch_list=fetch_varname_list, \
return_numpy=False)
if loss_type == "ctc":
preds = np.array(predict[0])
preds = preds.reshape(-1)
@ -114,7 +139,18 @@ def main():
score = np.mean(probs[0, 1:end_pos[1]])
preds = preds.reshape(-1)
preds_text = char_ops.decode(preds)
elif loss_type == "srn":
cur_pred = []
preds = np.array(predict[0])
preds = preds.reshape(-1)
probs = np.array(predict[1])
ind = np.argmax(probs, axis=1)
valid_ind = np.where(preds != 37)[0]
if len(valid_ind) == 0:
continue
score = np.mean(probs[valid_ind, ind[valid_ind]])
preds = preds[:valid_ind[-1] + 1]
preds_text = char_ops.decode(preds)
logger.info("\t index: {}".format(preds))
logger.info("\t word : {}".format(preds_text))
logger.info("\t score: {}".format(score))

Write Preview
Loading…
Cancel
Save