diff --git a/dan/manager/metrics.py b/dan/manager/metrics.py
index 565f307211ab31eff18eda16abdeae4d4514bb57..abb8c166dad8afa8b68fa4ff724b1820324f71e1 100644
--- a/dan/manager/metrics.py
+++ b/dan/manager/metrics.py
@@ -149,7 +149,6 @@ class MetricManager:
)
elif metric_name in [
"loss",
- "loss_ctc",
"loss_ce",
]:
value = float(
@@ -219,7 +218,6 @@ class MetricManager:
]
metrics["nb_words_no_punct"] = [len(gt) for gt in split_gt]
elif metric_name in [
- "loss_ctc",
"loss_ce",
"loss",
]:
diff --git a/dan/manager/ocr.py b/dan/manager/ocr.py
index bd4640731aafa905b05b596233e917e1971a7209..45d851c3bb558796c29db980a8d84380d61c6908 100644
--- a/dan/manager/ocr.py
+++ b/dan/manager/ocr.py
@@ -14,7 +14,6 @@ from dan.ocr.utils import LM_str_to_ind
from dan.utils import (
pad_image,
pad_image_width_random,
- pad_image_width_right,
pad_images,
pad_sequences_1D,
rand,
@@ -45,19 +44,12 @@ class OCRDatasetManager(DatasetManager):
self.tokens = {
"pad": params["config"]["padding_token"],
}
- if self.params["config"]["charset_mode"].lower() == "ctc":
- self.tokens["blank"] = len(self.charset)
- self.tokens["pad"] = (
- self.tokens["pad"] if self.tokens["pad"] else len(self.charset) + 1
- )
- self.params["config"]["padding_token"] = self.tokens["pad"]
- elif self.params["config"]["charset_mode"] == "seq2seq":
- self.tokens["end"] = len(self.charset)
- self.tokens["start"] = len(self.charset) + 1
- self.tokens["pad"] = (
- self.tokens["pad"] if self.tokens["pad"] else len(self.charset) + 2
- )
- self.params["config"]["padding_token"] = self.tokens["pad"]
+ self.tokens["end"] = len(self.charset)
+ self.tokens["start"] = len(self.charset) + 1
+ self.tokens["pad"] = (
+ self.tokens["pad"] if self.tokens["pad"] else len(self.charset) + 2
+ )
+ self.params["config"]["padding_token"] = self.tokens["pad"]
def get_merged_charsets(self):
"""
@@ -161,37 +153,7 @@ class OCRDataset(GenericDataset):
sample["img_shape"] / self.reduce_dims_factor
).astype(int)
- # Padding to handle CTC requirements
if self.set_name == "train":
- max_label_len = 0
- height = 1
- ctc_padding = False
- if "CTC_line" in self.params["config"]["constraints"]:
- max_label_len = sample["label_len"]
- ctc_padding = True
- if "CTC_va" in self.params["config"]["constraints"]:
- max_label_len = max(sample["line_label_len"])
- ctc_padding = True
- if "CTC_pg" in self.params["config"]["constraints"]:
- max_label_len = sample["label_len"]
- height = max(sample["img_reduced_shape"][0], 1)
- ctc_padding = True
- if (
- ctc_padding
- and 2 * max_label_len + 1 > sample["img_reduced_shape"][1] * height
- ):
- sample["img"] = pad_image_width_right(
- sample["img"],
- int(
- np.ceil((2 * max_label_len + 1) / height)
- * self.reduce_dims_factor[1]
- ),
- self.padding_value,
- )
- sample["img_shape"] = sample["img"].shape
- sample["img_reduced_shape"] = np.ceil(
- sample["img_shape"] / self.reduce_dims_factor
- ).astype(int)
sample["img_reduced_shape"] = [
max(1, t) for t in sample["img_reduced_shape"]
]
diff --git a/dan/manager/training.py b/dan/manager/training.py
index 876e788c23c3ec785319535674e3af985790b4a7..53012a7c22166f5461bc4f55d840b5b191dc8078 100644
--- a/dan/manager/training.py
+++ b/dan/manager/training.py
@@ -341,12 +341,6 @@ class GenericTrainingManager:
if c in old_charset:
new_weights[i] = weights[old_charset.index(c)]
pretrained_chars.append(c)
- if (
- "transfered_charset_last_is_ctc_blank" in self.params["model_params"]
- and self.params["model_params"]["transfered_charset_last_is_ctc_blank"]
- ):
- new_weights[-1] = weights[-1]
- pretrained_chars.append("<blank>")
checkpoint["{}_state_dict".format(state_dict_name)][key] = new_weights
self.models[model_name].load_state_dict(
{key: checkpoint["{}_state_dict".format(state_dict_name)][key]},
@@ -903,7 +897,6 @@ class GenericTrainingManager:
"nb_samples",
"loss",
"loss_ce",
- "loss_ctc",
"loss_ce_end",
]:
metrics[metric_name] = self.sum_ddp_metric(
diff --git a/dan/ocr/document/train.py b/dan/ocr/document/train.py
index 5a5ebd006e166f1f3c7b872bd2a68dc7b9cbe10a..81660d56e6807f1b84f79fab08f7268ae91441f4 100644
--- a/dan/ocr/document/train.py
+++ b/dan/ocr/document/train.py
@@ -114,7 +114,6 @@ def get_config():
"height_divisor": 32, # Image height will be divided by 32
"padding_value": 0, # Image padding value
"padding_token": None, # Label padding value
- "charset_mode": "seq2seq", # add end-of-transcription and start-of-transcription tokens to charset
"constraints": [
"add_eot",
"add_sot",
diff --git a/docs/usage/train/parameters.md b/docs/usage/train/parameters.md
index dacfd9dd07bc1f97666cf79b5d4df2e5340cc71f..4e3fa785c501273e6963536a5e53854fe1e045e2 100644
--- a/docs/usage/train/parameters.md
+++ b/docs/usage/train/parameters.md
@@ -18,7 +18,6 @@ All hyperparameters are specified and editable in the training scripts (meaning
| `dataset_params.config.width_divisor` | Factor to reduce the height of the feature vector before feeding the decoder. | `int` | `32` |
| `dataset_params.config.padding_value` | Image padding value. | `int` | `0` |
| `dataset_params.config.padding_token` | Transcription padding value. | `int` | `None` |
-| `dataset_params.config.charset_mode` | Whether to add end-of-transcription and start-of-transcription tokens to charset. | `str` | `seq2seq` |
| `dataset_params.config.constraints` | Whether to add end-of-transcription and start-of-transcription tokens in labels. | `list` | `["add_eot", "add_sot"]` |
| `dataset_params.config.normalize` | Normalize with mean and variance of training dataset. | `bool` | `True` |
| `dataset_params.config.preprocessings` | List of pre-processing functions to apply to input images. | `list` | (see [dedicated section](#data-preprocessing)) |
diff --git a/tests/conftest.py b/tests/conftest.py
index 777bdda54579cf43ca1e91c7dd85200781a96d64..9d0860f0db8ef0ce617458be26d867808e3637e7 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -72,7 +72,6 @@ def training_config():
"height_divisor": 32, # Image height will be divided by 32
"padding_value": 0, # Image padding value
"padding_token": None, # Label padding value
- "charset_mode": "seq2seq", # add end-of-transcription and start-of-transcription tokens to charset
"constraints": [
"add_eot",
"add_sot",