Comments (6)
Hi,
we do not write any transcription into the test files, as they are not needed. We only use the test data to evaluate and not to train or tune parameters. By not including the transcriptions, we wanted to avoid to accidentally using it for training or tuning.
We train on train.1.h5 and train.2.h while train_valid.h5 is used for validation, for example for early stopping to prevent overfitting.
Once the training is done, we use the trained network and forward the network on the valid.h5 or test.h5 data, which will create h5 files with posterior probabilities. These are then the input to a decoder which does the recognition. In the demo, there is a script which does the recognition in a very simple way (without language model or lexicon and with a greedy search strategy).
The final evaluation is done with the result of the decoder, outside of RETURNN.
from returnn.
Thank you. Before the input to a decoder for recognition, does test transcriptions also go through same regex operations. Or in scoring is same regex operation is applied?
from returnn.
If you need to post-process the recognition output depends on the recognizer. The important thing is that your output matches the format of the ground truth. I'm not sure anymore, if we changed the tokenization of the transcriptions or if it was directly provided like this, but here are a few samples of the ground truth which we used to give you an idea how tokenization was done for IAM (I can try to look up more details if necessary):
d06-030: Amen . MOST people would probably regard tiredness as a purely physical thing . The cure for which is sleep . This is only partly true . Many people wake up tired of a morning and no amount of rest seems to make any difference . Sleep , to be effective , must be of that child-like quality which comes from innocence .
d06-050: Nor is she necessarily being deceitful . She really did feel tired until the mind got the necessary injection of a fresh - and an attractive - interest ! Tiredness has , therefore , as much to do with our mental state as with our physical exhaustion . A disturbed mind can bring the healthiest body to a sense of fatigue . They wonder why they get no rest at night , even if they do sleep .
d06-082: If we ever stopped to consider how much energy - and time - we lose this way in the course of a day we would be staggered . Some of it is inevitable , and we do not want to become too pernickety . Nevertheless , we could all probably be a little more orderly for we so frequently just muddle through .
d04-075: The Messusah , which is nailed on the doorposts contains those two lessons . The third lesson of the Shmah ( Numbers 15 , 37-41 ) is called " the going-out of the land of Egypt " . It starts with the story of the fringes the " tsitsits " . This lesson used to be read only in the morning . And the Talmud tells a story , which is also in the Hagadah that Rabbi Eleasar ben Assarja said , " I am nearly seventy and I had not succeeded that
from returnn.
Thank you for the help. It is not necessary, i was just curious.
from returnn.
Great!
Do you have further problems/questions? Otherwise I'd like to close the issue for now. You can reopen a new one if you get further problems at a later point.
from returnn.
Ok, yeah I do not have any further question.
from returnn.
Related Issues (20)
- Dataset batching like ESPnet support
- torch.nn.functional.conv2d: RuntimeError: GET was unable to find an engine to execute this computation HOT 1
- TensorFlow 2.14 degradation in WER HOT 2
- Updates for recent TensorFlow version
- Hang in dataset iterator HOT 5
- Log GPU device for torch backend HOT 2
- torch.onnx.export requires input_names and output_names to be in order HOT 12
- RF weight dropout HOT 6
- Support for larger scale datasets HOT 33
- RuntimeError: CUDA error: unknown error
- PyTorch debug_add_check_numerics_ops
- Compilation of custom operations failing on TF 2.15/CUDA 12 HOT 5
- `ConcatFilesDataset` combines poorly with `MetaDataset` HOT 6
- RF torch `lstm` fails with torch amp option. HOT 6
- `DistributeFilesDataset` with sharding on file level HOT 6
- ConcatFilesDataset: Reshuffle files per subepoch after every full epoch HOT 2
- `ConcatFilesDataset` needs a better name HOT 10
- RF BatchNorm running var small diff between TF-layers, pure RF and direct PyTorch, biased vs unbiased
- `DistributeFilesDataset`, allow kwargs in `get_sub_epoch_dataset` HOT 10
- Tensor deepcopy does not copy raw_tensor
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from returnn.