Hi, I have just run the new version of your code (TF and Keras) at my dataset. The perplexity is ~ 1.3, which is very nice. But when I want to test the model I get ~1000. After looking into visualization it look like the model always predict the same sequence of letters, no matter what input is (both on train test and validation set which the model did not see).

This is my command which I use for testing:

python src/launcher.py --phase=test --data-path=/home/ubuntu/Data/anno_v2/labels_vals.txt --data-base-dir=/home/ubuntu/Data/anno_v2 --log-path=logtest.txt --load-model --visualize

In the other hand, when I continue training, everything is ok. The only think which vary between this two runs is 'decoder' input: while testing it is predicted words from previous iterations while in 'train' it is ground truth. Do you know what goes wrong? Or I overfitt to data and now first letter does create a all sequence, no matter what input is (my data have 15k examples)?

Hi guys,

My training model can't reach the accuracy of your release model.tar. My training parameter is here:

python src/launcher.py \
    --phase=train \
    --data-path=${data_path} \
    --data-base-dir=${data_base_dir} \
    --log-path=log_train.txt \
    --load-model \
    --model-dir=$model_dir \
    --num-epoch=800 \
    --target-embedding-size=10 

When the steps reach translate.ckpt-48000, I use it to evaluate svt. But the accuracy is every poor of 6.1%(39.365199 out of 647 correct) comparing with your result of 68%(437.192016 out of 647 correct).

I don't know why. What's your training parameters?

waiting for your answers

Hi guys,

I've download the evaluation_data and model you supported, and verify the test result by the command python src/launcher.py --phase=test --visualize --data-path=evaluation_data/svt/test.txt --data-base-dir=evaluation_data/svt --log-path=log.txt --load-model --model-dir=model --output-dir=results --target-embedding-size=10
(BTW. should add --target-embedding-size=10 option or else it will fails by "Assign requires shapes of both tensors to match. lhs shape= [39,20] rhs shape= [39,10]")

But the result log.txt shows "188.780794 out of 647 correct" which means only 29% accuracy rate you get. Why is it so bad? Or Did I something wrong?

python -V Python 2.7.10

python src/launcher.py --phase=train --data-path=sample/sample.txt --data-base-dir=sample --log-path=log.txt --no-load-model Using TensorFlow backend. 2017-01-09 11:52:13,224 root INFO loading data 2017-01-09 11:52:13,225 root INFO data_path: sample/sample.txt 2017-01-09 11:52:13,225 root INFO phase: train 2017-01-09 11:52:13,225 root INFO batch_size: 64 2017-01-09 11:52:13,225 root INFO num_epoch: 1000 2017-01-09 11:52:13,225 root INFO steps_per_checkpoint 500 2017-01-09 11:52:13,225 root INFO target_vocab_size: 39 2017-01-09 11:52:13,225 root INFO model_dir: train 2017-01-09 11:52:13,225 root INFO target_embedding_size: 10 2017-01-09 11:52:13,226 root INFO attn_num_hidden: 128 2017-01-09 11:52:13,226 root INFO attn_num_layers: 2 2017-01-09 11:52:13,226 root INFO buckets 2017-01-09 11:52:13,226 root INFO [(16, 11), (27, 17), (35, 19), (64, 22), (80, 32)] Traceback (most recent call last): File "src/launcher.py", line 142, in main(sys.argv[1:], exp_config.ExpConfig) File "src/launcher.py", line 138, in main old_model_version = parameters.old_model_version) File "/Users/logiz/www/test/Attention-OCR/src/model/model.py", line 119, in init cnn_model = CNN(self.img_data) File "/Users/logiz/www/test/Attention-OCR/src/model/cnn.py", line 30, in init self._build_network(input_tensor) File "/Users/logiz/www/test/Attention-OCR/src/model/cnn.py", line 43, in _build_network input_layer.set_input(input_tensor=input_tensor) AttributeError: 'InputLayer' object has no attribute 'set_input'

how to fix,thank you.

I was checking the last version that should support tensorflow 11 and keras 1.1.1 that I have in my system

I run on full synth90 corpus the sample training (without GRU option)

python src/launcher.py \
	--phase=train \
	--data-path=90kDICT32px/new_annotation_train.txt \
	--data-base-dir=90kDICT32px \
	--log-path=log_sy.txt \
	--attn-num-hidden 256 \
	--batch-size 32 \
	--model-dir=model_x \
	--initial-learning-rate=1.0 \
	--num-epoch=20000 \
	--gpu-id=0 \
--target-embedding-size=10

It seem to converge better than the old version. After 70k iterations I get pretty nice perplexity

2016-12-08 13:12:07,393 root  INFO     current_step: 78998
2016-12-08 13:12:07,589 root  INFO     step_time: 0.196019, step_loss: 0.080441, step perplexity: 1.083765
2016-12-08 13:12:08,620 root  INFO     current_step: 78999
2016-12-08 13:12:08,884 root  INFO     step_time: 0.264444, step_loss: 0.071078, step perplexity: 1.073665
2016-12-08 13:12:08,885 root  INFO     global step 79000 step-time 0.26 loss 0.094580  perplexity 1.10
2016-12-08 13:12:08,885 root  INFO     Saving model, current_step: 79000

However, when testing on SVT, I get a very bad performance

2016-12-08 13:35:25,947 root  INFO     step_time: 0.049570, loss: 3.750340, step perplexity: 42.535551
2016-12-08 13:35:25,950 root  INFO     62.324373 out of 647 correct

Interestingly, the model provided by the authors works great with --old-model option. So it is not the decoder bug but rather some problem in training that does not show up in training log but does affect the final performance.

I wonder if someone tried to traing/test the new version using tensorflow 11 and keras 1.1.1 ? Thanks

hello, I tried to train your data with your code, I use the tensorflow as the backend and I meet a problem. how can i solve it? I’d like to receive your answer, thanks!

File "~/anaconda2/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 496, in __iter__
 raise TypeError("'Tensor' object is not iterable.")
TypeError: 'Tensor' object is not iterable.

I was trying to evaluate the performance of your tool on SVT test data

As suggested by README, I downloaded the model from http://www.cs.cmu.edu/~yuntiand/model.tgz and the data from http://www.cs.cmu.edu/~yuntiand/evaluation_data.tgz

After running

python src/launcher.py --phase=test --visualize --data-path=evaluation_data/svt/test.txt --data-base-dir=evaluation_data/svt --log-path=log.txt --load-model --model-dir=model --output-dir=results

I get

INFO     3.166667 out of 6 correct

...

INFO     437.192016 out of 647 correct

I was expecting to get about 80% accuracy, but the results seem to be worse.

Do you know what is the expected accuracy of this tool with the provided model and how it can possibly improved to achieve state-of-the art result?

Thanks

Hi, I've trained a model following the toy example. And using the model to test the iiit5k data. While error happens: "../Attention-OCR/src/model/model.py" line 182, in init self.saver_all = tf.train.Saver(tf.all_variables()) an exception is throwed: tensorflow.python.framework.errors.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shap=[1024] rhs shape=[2048]

When using the author's model trained on Synth 90K to test iiit5k , everythin is ok and finally get "2034.98 out of 3000 correct".

Could anybody explain this? Thanks a lot.

I downloaded the sample data in the Attention-OCR folder and depressed there. I executed such command line and encountered such error. TypeError: unsupported operand type(s) for *: 'NoneType' and 'int' below is the trackback information: File "src/launcher.py", line 129, in main(sys.argv[1:], exp_config.ExpConfig) File "src/launcher.py", line 125, in main session = sess) File "/data/clzhai/Attention-OCR-master/src/model/model.py", line 118, in init cnn_model = CNN(self.img_data) File "/data/clzhai/Attention-OCR-master/src/model/cnn.py", line 31, in init self._build_network(input_tensor) File "/data/clzhai/Attention-OCR-master/src/model/cnn.py", line 51, in _build_network border_mode='same')) File "/data/clzhai/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/models.py", line 308, in add output_tensor = layer(self.outputs[0]) File "/data/clzhai/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/engine/topology.py", line 487, in call self.build(input_shapes[0]) File "/data/clzhai/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/layers/convolutional.py", line 410, in build self.W = self.init(self.W_shape, name='{}_W'.format(self.name)) File "/data/clzhai/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/initializations.py", line 57, in glorot_uniform fan_in, fan_out = get_fans(shape, dim_ordering=dim_ordering) File "/data/clzhai/anaconda2/envs/tensorflow/lib/python2.7/site-packages/keras/initializations.py", line 15, in get_fans receptive_field_size = np.prod(shape[2:]) File "/data/clzhai/anaconda2/envs/tensorflow/lib/python2.7/site-packages/numpy/core/fromnumeric.py", line 2492, in prod out=out, keepdims=keepdims) File "/data/clzhai/anaconda2/envs/tensorflow/lib/python2.7/site-packages/numpy/core/_methods.py", line 35, in _prod return umr_prod(a, axis, dtype, out, keepdims) TypeError: unsupported operand type(s) for *: 'NoneType' and 'int'

hi guys,

When I run "python src/launcher.py --phase=train --data-path=sample/sample.txt --data-base-dir=sample --log-path=log.txt --no-load-model", the following errors came out:

Traceback (most recent call last): File "src/launcher.py", line 129, in main(sys.argv[1:], exp_config.ExpConfig) File "src/launcher.py", line 125, in main session = sess) File "/home/aaron/projects/Attention-OCR/src/model/model.py", line 164, in init for old_value, new_value in cnn_model.model.updates: File "/home/aaron/anaconda2/envs/tensorflow/lib/python2.7/site-packages/tensorflow/python/framework/ops.py", line 495, in iter raise TypeError("'Tensor' object is not iterable.") TypeError: 'Tensor' object is not iterable.

I train the model on my own dataset, which contains 10k plate images. When I set the batch_size as 256, the step perplexity reaches to 1.001, however, the step perplexity of the trained model increases to 10 if the batch_size is set to 2. If I fix the batch_size to 256, does the model really converge? I wonder that a large batch_size is not appropriate.

2021-07-29 10:45:49,016 root INFO step_time: 0.138798, loss: 3.648251, step perplexity: 38.407426 2021-07-29 10:45:49,174 root INFO 0.000000 out of 614 correct 2021-07-29 10:45:49,315 root INFO step_time: 0.133353, loss: 3.665064, step perplexity: 39.058630 2021-07-29 10:45:49,397 root INFO 0.000000 out of 615 correct 2021-07-29 10:45:49,551 root INFO step_time: 0.148111, loss: 3.655409, step perplexity: 38.683334 2021-07-29 10:45:49,598 root INFO 0.000000 out of 616 correct

Hi,

I try to visualise attention masks based on the output of SavedModel. What is the meaning of real_len variable in visualise_attention function and why is the attention mask convolved with [0.199547,0.200226,0.200454,0.200226,0.199547]?

Hi, can this model used in recognition of printed mathematical expression recognition? like

im2latex-100k? https://paperswithcode.com/dataset/im2latex-100k

I try to train my own dataset. I get loss, perflexity and precision might be good. But when I run model to test on same dataset I trained. It shows bad. Any help? Thanks Beside, I test author's pre-train model, it happens: 2019-12-11 05:06:25,839 root INFO 0.000000 out of 647 correct

After training, many files were created in 'train' folder as shown in the picture. How do I select and test the model I want?

Which parameters represent the path to the model that I want to test?

python src/launcher.py --phase=test --visualize --data-path=evaluation_data/svt/test.txt --data-base-dir=evaluation_data/svt --log-path=log.txt --load-model --model-dir=model --output-dir=results