From e01751a4df836018bbf2eb05481433d2e36d488e Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?M=C3=A9lodie=20Boillet?= <boillet@teklia.com>
Date: Wed, 24 May 2023 13:48:27 +0200
Subject: [PATCH] Remove ctc mode
---
dan/manager/metrics.py | 2 --
dan/manager/ocr.py | 19 ++++++-------------
dan/manager/training.py | 7 -------
dan/ocr/document/train.py | 1 -
docs/usage/train/parameters.md | 1 -
tests/conftest.py | 1 -
6 files changed, 6 insertions(+), 25 deletions(-)
diff --git a/dan/manager/metrics.py b/dan/manager/metrics.py
index ecb94a86..1f78f505 100644
--- a/dan/manager/metrics.py
+++ b/dan/manager/metrics.py
@@ -150,7 +150,6 @@ class MetricManager:
)
elif metric_name in [
"loss",
- "loss_ctc",
"loss_ce",
]:
value = float(
@@ -220,7 +219,6 @@ 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 bd464073..ec98755a 100644
--- a/dan/manager/ocr.py
+++ b/dan/manager/ocr.py
@@ -45,19 +45,12 @@ 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):
"""
diff --git a/dan/manager/training.py b/dan/manager/training.py
index 5821d2e8..1f18aace 100644
--- a/dan/manager/training.py
+++ b/dan/manager/training.py
@@ -343,12 +343,6 @@ 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]},
@@ -920,7 +914,6 @@ 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 5a5ebd00..81660d56 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 dacfd9dd..4e3fa785 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 777bdda5..9d0860f0 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",
--
GitLab