pay20y/sar_tf

This is an implementation of Show, Attend and Read with tensorflow

sar_tf's Issues

train with multi gpus

How can i train with multi gpus?
I set parser.add_argument('-g', '--gpus', type=str, default='0,1')
but nvidia-smi shows only gpu0 is computing

Scalars on TensorBoard

Is the library saving scalars? I'm not able to see them on TensorBoard.

Am I missing something?

Two identical experiments have completely different results

I used the same data and the same configuration to train my handwritten Chinese recognition. The first time I iterated 4000 times, I got 50% acc, and it can be increased to about 80% later, but in the second experiment，the loss didn't decrease, the accuracy was not increase at all, it was always 5%

pretrained model

hi~can you provide the pretrained model ?Thanks

the role of mask in attention operation

I am reading torch implementation, your implementation and the pytorch implementation. I found that there are mask in your implementation and torch implementation, but there is no mask in pytorch implementation. Is the role of mask is to get the valid ones? If there is no mask, what will the performance and the result be like?

I am training the pytorch implementation on handwritten dataset, I found that there is a lot of repeat in the decoded result, as below shown. is is the reason that I didn't use mask in the procedure of attention operation?

groundtruth:  the^fragile^nature
prediction:  the^fragile^fragile^fragile^fragile^fragile^fragile^fragile^fragile^fragile^fragi

The result of training test and offline test is different in same data

The same data, when used as the validation during training, the accuracy of the test is 91.5%. After finshed training and I used the same data for inference test (the checkpoint model is same), but the accuracy was only 84.5%. Why is there a difference of 7 points. Filtering was used in both times.

AttributeError: Can't pickle local object 'GeneratorEnqueuer.start.<locals>.data_generator_task'

Seems like the library has problems with multiprocessing on Windows.

Even using the parameter -j "1" that should elimitate the multiprocessing, the script returns this error

Traceback (most recent call last):
  File "train.py", line 232, in <module>
    main_train(args)
  File "train.py", line 186, in main_train
    train_data = get_batch_data(train_data_list, args.train_batch_size)
  File "train.py", line 51, in get_batch_data
    _data = next(data)
  File "C:\Users\mikyl\Desktop\SAR\data_provider\data_generator.py", line 124, in get_batch
    enqueuer.start(max_queue_size=4, workers=num_workers)
  File "C:\Users\mikyl\Desktop\SAR\data_provider\generator_enqueuer.py", line 78, in start
    thread.start()
  File "C:\Users\mikyl\AppData\Local\Programs\Python\Python36\lib\multiprocessing\process.py", line 105, in start
    self._popen = self._Popen(self)
  File "C:\Users\mikyl\AppData\Local\Programs\Python\Python36\lib\multiprocessing\context.py", line 223, in _Popen
    return _default_context.get_context().Process._Popen(process_obj)
  File "C:\Users\mikyl\AppData\Local\Programs\Python\Python36\lib\multiprocessing\context.py", line 322, in _Popen
    return Popen(process_obj)
  File "C:\Users\mikyl\AppData\Local\Programs\Python\Python36\lib\multiprocessing\popen_spawn_win32.py", line 65, in __init__
    reduction.dump(process_obj, to_child)
  File "C:\Users\mikyl\AppData\Local\Programs\Python\Python36\lib\multiprocessing\reduction.py", line 60, in dump
    ForkingPickler(file, protocol).dump(obj)
AttributeError: Can't pickle local object 'GeneratorEnqueuer.start.<locals>.data_generator_task'

I'm working on a pull request

Can you show me code to visualize the attention map?

ArTLoader, Syn90KLoader and SynTextLoader missing

Hi,
executing test.py the script stops because the following modules are missing.

from data_provider.art_loader import ArTLoader
from data_provider.syn90k_loader import Syn90KLoader
from data_provider.syntext_loader import SynTextLoader

Where I can find them?

Name of the output node with the inferenced word

Hi, what's the name of the node containing the predicted word?

训练数据集

请问你可以提供一下训练数据集的百度云链接吗？我下载不到这些数据。。。

Experiments or performance on Chinese

Hi, thanks for sharing this great work. I wonder if you have tried this model on Chinese dataset? If so, did it perform well?

libgcc_s.so.1 must be installed for pthread_cancel to work

Did you ever get this bug when running pytorch code. When I ungrade pytorch from 1.2.0 to 1.1.0, this error happened.

download pretrained model without Baidu

Hi, I can't understand Chinese. Is there another way to download the pretrained model?

Mismatch with paper

Greetings,
I noticed in your back bone implementation you used 2x1 maxpool instead of 1x2, which is mentioned in the paper. Could you explain why? Thanks!

中文训练

请问是否支持中文的识别

Same character repeated in test prediction

Hi all,
I'm seeing a strange behaviour in the training. When I do the evaluation test, the result is very different from the training test.

Below an example:

Iter 17 train loss= 10.839604
################################################################################
train prediction         train labels
1AAAAA   HAIRANE
13       93
13       26
13A      JAM
15       84
13       61
13       1970
13       293
13       127
13       70
13       DERIBA
13       854
13       2033
13       3
13A      SZWARNOG
13       69
13AAA    ESP
13       32
13       8
13       MOUKRIME
13       4
13       13
13AAAA   FRA
11       184
11       72
13       237
15       241
13AAA    YATSEVICH
13       68
13       53
15       SAW
13       20
15       2357
13       250
################################################################################
test prediction          test labels
111111111111111111111111111111   354
111111111111111111111111111111   3
111111111111111111111111111111   22
111111111111111111111111111111   1103
111111111111111111111111111111   208
111111111111111111111111111111   102
111111111111111111111111111111   1213
111111111111111111111111111111   37
111111111111111111111111111111   12
111111111111111111111111111111   72
111111111111111111111111111111   F131
111111111111111111111111111111   18
111111111111111111111111111111   227
111111111111111111111111111111   2478
111111111111111111111111111111   752
111111111111111111111111111111   3
111111111111111111111111111111   14
111111111111111111111111111111   106
111111111111111111111111111111   12
111111111111111111111111111111   441
111111111111111111111111111111   11
111111111111111111111111111111   34
111111111111111111111111111111   9
111111111111111111111111111111   48
111111111111111111111111111111   54
111111111111111111111111111111   405
111111111111111111111111111111   GER
111111111111111111111111111111   21
111111111111111111111111111111   211
111111111111111111111111111111   64
111111111111111111111111111111   240
111111111111111111111111111111   7
################################################################################
Evaluation Iter 17 test loss: 252054992.000000 train accuracy: 0.010000 test accuracy 0.000000
Iter 17 save to checkpoint

Do you know why in the test prediction it is repeated the same number for all the rows?

Performance issues in the programs

Hello, I found a performance issue in the definition of __call__, module/Decoder.py, tf.fill and will created repeatedly during program execution, resulting in reduced efficiency. I think it should be created before the loop in __call__.

The same issue exists here

Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.

dataset of pre-trained model

Hi hear,
thanks for your implementation,
I have a question, which datasets you use for pretrain model?
if i want to train with my own dataset, whether i resume with pretrain model or train from scratch with Syn90k,SynthText,SynthAdd and then fine tune with my own?
best regards

Model tensors are linked to batch size

When freezing the model, the frozen graph contains the information of the training batch size.

We should somehow refactor the model to avoid this behaviour when freezing

freeze.py error

Hello,
When I tried to freeze the graph and save as pb, I got the following error:

  File "/mnt/DATA/project/xxx/SAR_TF-master/sar_model.py", line 55, in inference
    input_widths_list = tf.unstack(input_widths, axis=0)
  File "/data/project/anaconda3/lib/python3.6/site-packages/tensorflow/python/ops/array_ops.py", line 1015, in unstack
    raise ValueError("Cannot infer num from shape %s" % value_shape)
ValueError: Cannot infer num from shape (?,)

I fixed the batch size=1 and the error solved:

    input_train_images = tf.placeholder(dtype=tf.float32, shape=[1, args.height, args.width, 3], name="input_train_images")
    input_train_images_width = tf.placeholder(dtype=tf.float32, shape=[1], name="input_train_width")
    input_train_labels = tf.placeholder(dtype=tf.int32, shape=[1, args.max_len], name="input_train_labels")
    input_train_labels_mask = tf.placeholder(dtype=tf.int32, shape=[1, args.max_len], name="input_train_labels_mask")

    input_val_images = tf.placeholder(dtype=tf.float32, shape=[1, args.height, args.width, 3],name="input_val_images")
    input_val_images_width = tf.placeholder(dtype=tf.float32, shape=[1], name="input_val_width")
    input_val_labels = tf.placeholder(dtype=tf.int32, shape=[1, args.max_len], name="input_val_labels")
    input_val_labels_mask = tf.placeholder(dtype=tf.int32, shape=[1, args.max_len], name="input_val_labels_mask")

Is there any way to keep the batch size as None and save the grapb as pb? Thanks

Reduce model size

Is there a way to reduce the size of the output model?

AttributeError: module 'tensorflow.math' has no attribute 'floor'

Hi,
I get this error when I try to start the training. Do you know how to fix it?

Traceback (most recent call last):
  File "train.py", line 232, in <module>
    main_train(args)
  File "train.py", line 103, in main_train
    batch_size=args.train_batch_size, reuse=False)
  File "C:\Users\mikyl\Desktop\SAR\sar_model.py", line 40, in __call__
    encoder_state,  feature_map, mask_map = self.inference(input_images, input_widths, batch_size)
  File "C:\Users\mikyl\Desktop\SAR\sar_model.py", line 50, in inference
    input_widths = tf.cast(tf.math.floor(input_widths * (fea_W / img_W)), tf.int32)
AttributeError: module 'tensorflow.math' has no attribute 'floor'

Why the test accuracy of CUTE80 is 80.49%, not 83.6%? thanks.

How can I reproduce ICDAR13 acc?

When I train with ICDAR13 dataset with default config file, the train acc reach 1 quickly, but val acc is very low , which means overfitting .
So how can i reproduce the result on ICDAR2013?
thx