Skip to content
Snippets Groups Projects
Commit 82d7946e authored by Manon Blanco's avatar Manon Blanco Committed by Yoann Schneider
Browse files

Add a new configuration to tests the training command

parent 1591b8ef
No related branches found
No related tags found
1 merge request!301Add a new configuration to tests the training command
Hide whitespace changes
Inline Side-by-side
configs/tests.json 0 → 100644
+ 105
− 0
View file @ 82d7946e
{
"dataset": {
"datasets": {
"training": "tests/data/training/training_dataset"
},
"train": {
"name": "training-train",
"datasets": [
["training", "train"]
]
},
"val": {
"training-val": [
["training", "val"]
]
},
"test": {
"training-test": [
["training", "test"]
]
},
"max_char_prediction": 30,
"tokens": null
},
"model": {
"transfered_charset": true,
"additional_tokens": 1,
"encoder": {
"dropout": 0.5,
"nb_layers": 5
},
"h_max": 500,
"w_max": 1000,
"decoder": {
"l_max": 15000,
"dec_num_layers": 8,
"dec_num_heads": 4,
"dec_res_dropout": 0.1,
"dec_pred_dropout": 0.1,
"dec_att_dropout": 0.1,
"dec_dim_feedforward": 256,
"attention_win": 100,
"enc_dim": 256
}
},
"training": {
"data": {
"batch_size": 2,
"load_in_memory": true,
"worker_per_gpu": 4,
"preprocessings": [
{
"type": "max_resize",
"max_width": 2000,
"max_height": 2000
}
],
"augmentation": true
},
"device": {
"use_ddp": false,
"ddp_port": "20027",
"use_amp": true,
"nb_gpu": 0,
"force": "cpu"
},
"metrics": {
"train": [
"loss_ce",
"cer",
"wer",
"wer_no_punct"
],
"eval": [
"cer",
"wer",
"wer_no_punct"
]
},
"validation": {
"eval_on_valid": true,
"eval_on_valid_interval": 2,
"set_name_focus_metric": "training-val"
},
"output_folder": "dan_trained_model",
"gradient_clipping": {},
"max_nb_epochs": 4,
"load_epoch": "last",
"optimizers": {
"all": {
"args": {
"lr": 0.0001,
"amsgrad": false
}
}
},
"lr_schedulers": null,
"label_noise_scheduler": {
"min_error_rate": 0.2,
"max_error_rate": 0.2,
"total_num_steps": 5e4
},
"transfer_learning": null
}
}
docs/get_started/development.md
+ 13
− 1
View file @ 82d7946e
......@@ -13,7 +13,9 @@ pip install pre-commit
pre-commit run -a
```
## Run tests
## Tests
### Unit tests
Tests are executed with [tox](https://tox.wiki) using [pytest](https://pytest.org).
......@@ -24,6 +26,16 @@ tox
To recreate tox virtual environment (e.g. a dependencies update), you may run `tox -r`.
### Commands
As unit tests do not test *everything*, it is sometimes necessary to use DAN commands directly to test developments.
The library already has all the documents needed to run the [training command](../usage/train/index.md) on a minimalist dataset. You can use the configuration available at `configs/tests.json`. It is already populated with the parameters used in the unit tests.
```shell
teklia-dan train --config configs/tests.json
```
## Documentation
This documentation uses [Sphinx](http://www.sphinx-doc.org/) and was generated using [MkDocs](https://mkdocs.org/) and [mkdocstrings](https://mkdocstrings.github.io/).
......
tests/conftest.py
+ 1
− 106
View file @ 82d7946e
......@@ -184,112 +184,7 @@ def mock_database(tmp_path_factory):
@pytest.fixture
def training_config():
config = {
"dataset": {
"datasets": {
"training": str(FIXTURES / "training" / "training_dataset"),
},
"train": {
"name": "training-train",
"datasets": [
("training", "train"),
],
},
"val": {
"training-val": [
("training", "val"),
],
},
"test": {
"training-test": [
("training", "test"),
],
},
"max_char_prediction": 30, # max number of token prediction
"tokens": None,
},
"model": {
"transfered_charset": True, # Transfer learning of the decision layer based on charset of the line HTR model
"additional_tokens": 1, # for decision layer = [<eot>, ], only for transferred charset
"encoder": {
"dropout": 0.5, # dropout rate for encoder
"nb_layers": 5, # encoder
},
"h_max": 500, # maximum height for encoder output (for 2D positional embedding)
"w_max": 1000, # maximum width for encoder output (for 2D positional embedding)
"decoder": {
"l_max": 15000, # max predicted sequence (for 1D positional embedding)
"dec_num_layers": 8, # number of transformer decoder layers
"dec_num_heads": 4, # number of heads in transformer decoder layers
"dec_res_dropout": 0.1, # dropout in transformer decoder layers
"dec_pred_dropout": 0.1, # dropout rate before decision layer
"dec_att_dropout": 0.1, # dropout rate in multi head attention
"dec_dim_feedforward": 256, # number of dimension for feedforward layer in transformer decoder layers
"attention_win": 100, # length of attention window
"enc_dim": 256, # dimension of extracted features
},
},
"training": {
"data": {
"batch_size": 2, # mini-batch size for training
"load_in_memory": True, # Load all images in CPU memory
"worker_per_gpu": 4, # Num of parallel processes per gpu for data loading
"preprocessings": [
{
"type": "max_resize",
"max_width": 2000,
"max_height": 2000,
}
],
"augmentation": True,
},
"device": {
"use_ddp": False, # Use DistributedDataParallel
"ddp_port": "20027",
"use_amp": True, # Enable automatic mix-precision
"nb_gpu": 0,
"force": "cpu", # `cpu` for debug purposes
},
"metrics": {
"train": [
"loss_ce",
"cer",
"wer",
"wer_no_punct",
], # Metrics name for training
"eval": [
"cer",
"wer",
"wer_no_punct",
], # Metrics name for evaluation on validation set during training
},
"validation": {
"eval_on_valid": True, # Whether to eval and logs metrics on validation set during training or not
"eval_on_valid_interval": 2, # Interval (in epochs) to evaluate during training
"set_name_focus_metric": "training-val",
},
"output_folder": "dan_trained_model", # folder name for checkpoint and results
"gradient_clipping": {},
"max_nb_epochs": 4, # maximum number of epochs before to stop
"load_epoch": "last", # ["best", "last"]: last to continue training, best to evaluate
"optimizers": {
"all": {
"args": {
"lr": 0.0001,
"amsgrad": False,
},
},
},
"lr_schedulers": None, # Learning rate schedulers
# Keep teacher forcing rate to 20% during whole training
"label_noise_scheduler": {
"min_error_rate": 0.2,
"max_error_rate": 0.2,
"total_num_steps": 5e4,
},
"transfer_learning": None,
},
}
config = json.loads((FIXTURES.parent.parent / "configs" / "tests.json").read_text())
update_config(config)
return config
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment