From f3718eb103e09486826f57ddab828996b1aac326 Mon Sep 17 00:00:00 2001
From: manonBlanco <blanco@teklia.com>
Date: Mon, 12 Jun 2023 17:25:48 +0200
Subject: [PATCH] Remove padding value and padding token parameters from
training configuration
---
dan/manager/dataset.py | 8 --------
dan/manager/ocr.py | 13 ++-----------
dan/ocr/document/train.py | 2 --
dan/utils.py | 9 +++------
docs/usage/train/parameters.md | 2 --
tests/conftest.py | 2 --
6 files changed, 5 insertions(+), 31 deletions(-)
diff --git a/dan/manager/dataset.py b/dan/manager/dataset.py
index 9d987c68..525e637c 100644
--- a/dan/manager/dataset.py
+++ b/dan/manager/dataset.py
@@ -18,7 +18,6 @@ class DatasetManager:
def __init__(self, params, device: str):
self.params = params
self.dataset_class = None
- self.img_padding_value = params["config"]["padding_value"]
self.my_collate_function = None
# Whether data should be copied on GPU via https://pytorch.org/docs/stable/generated/torch.Tensor.pin_memory.html
@@ -224,13 +223,6 @@ class GenericDataset(Dataset):
if self.load_in_memory:
self.apply_preprocessing(params["config"]["preprocessings"])
- self.padding_value = params["config"]["padding_value"]
- if self.padding_value == "mean":
- if self.mean is None:
- _, _ = self.compute_std_mean()
- self.padding_value = self.mean
- self.params["config"]["padding_value"] = self.padding_value
-
self.curriculum_config = None
def __len__(self):
diff --git a/dan/manager/ocr.py b/dan/manager/ocr.py
index 533d074c..6bd79268 100644
--- a/dan/manager/ocr.py
+++ b/dan/manager/ocr.py
@@ -25,14 +25,9 @@ class OCRDatasetManager(DatasetManager):
params["charset"] if "charset" in params else self.get_merged_charsets()
)
- self.tokens = {
- "pad": params["config"]["padding_token"],
- }
+ self.tokens = {"pad": len(self.charset) + 2}
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):
@@ -143,7 +138,6 @@ class OCRDataset(GenericDataset):
sample["img"], sample["img_position"] = pad_image(
sample["img"],
- padding_value=self.padding_value,
new_width=self.params["config"]["padding"]["min_width"],
new_height=self.params["config"]["padding"]["min_height"],
pad_width=pad_width,
@@ -177,7 +171,6 @@ class OCRCollateFunction:
"""
def __init__(self, config):
- self.img_padding_value = float(config["padding_value"])
self.label_padding_value = config["padding_token"]
self.config = config
@@ -190,9 +183,7 @@ class OCRCollateFunction:
self.config["padding_mode"] if "padding_mode" in self.config else "br"
)
imgs = [batch_data[i]["img"] for i in range(len(batch_data))]
- imgs = pad_images(
- imgs, padding_value=self.img_padding_value, padding_mode=padding_mode
- )
+ imgs = pad_images(imgs, padding_mode=padding_mode)
imgs = torch.tensor(imgs).float().permute(0, 3, 1, 2)
formatted_batch_data = {
diff --git a/dan/ocr/document/train.py b/dan/ocr/document/train.py
index d0a5fc47..ebfa34de 100644
--- a/dan/ocr/document/train.py
+++ b/dan/ocr/document/train.py
@@ -107,8 +107,6 @@ def get_config():
"worker_per_gpu": 4, # Num of parallel processes per gpu for data loading
"width_divisor": 8, # Image width will be divided by 8
"height_divisor": 32, # Image height will be divided by 32
- "padding_value": 0, # Image padding value
- "padding_token": None, # Label padding value
"preprocessings": [
{
"type": "to_RGB",
diff --git a/dan/utils.py b/dan/utils.py
index 8bffd18a..147bcee0 100644
--- a/dan/utils.py
+++ b/dan/utils.py
@@ -25,7 +25,7 @@ def pad_sequences_1D(data, padding_value):
return padded_data
-def pad_images(data, padding_value, padding_mode="br"):
+def pad_images(data, padding_mode="br"):
"""
data: list of numpy array
mode: "br"/"tl"/"random" (bottom-right, top-left, random)
@@ -34,9 +34,7 @@ def pad_images(data, padding_value, padding_mode="br"):
y_lengths = [x.shape[1] for x in data]
longest_x = max(x_lengths)
longest_y = max(y_lengths)
- padded_data = (
- np.ones((len(data), longest_x, longest_y, data[0].shape[2])) * padding_value
- )
+ padded_data = np.ones((len(data), longest_x, longest_y, data[0].shape[2])) * 0
for i, xy_len in enumerate(zip(x_lengths, y_lengths)):
x_len, y_len = xy_len
if padding_mode == "br":
@@ -56,7 +54,6 @@ def pad_images(data, padding_value, padding_mode="br"):
def pad_image(
image,
- padding_value,
new_height=None,
new_width=None,
pad_width=None,
@@ -90,7 +87,7 @@ def pad_image(
)
if not (pad_width == 0 and pad_height == 0):
- padded_image = np.ones((h + pad_height, w + pad_width, c)) * padding_value
+ padded_image = np.ones((h + pad_height, w + pad_width, c)) * 0
if padding_mode == "br":
hi, wi = 0, 0
elif padding_mode == "tl":
diff --git a/docs/usage/train/parameters.md b/docs/usage/train/parameters.md
index 8d97ae63..12cba43d 100644
--- a/docs/usage/train/parameters.md
+++ b/docs/usage/train/parameters.md
@@ -16,8 +16,6 @@ All hyperparameters are specified and editable in the training scripts (meaning
| `dataset_params.config.worker_per_gpu` | Number of parallel processes per gpu for data loading. | `int` | `4` |
| `dataset_params.config.height_divisor` | Factor to reduce the width of the feature vector before feeding the decoder. | `int` | `8` |
| `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.preprocessings` | List of pre-processing functions to apply to input images. | `list` | (see [dedicated section](#data-preprocessing)) |
| `dataset_params.config.augmentation` | Configuration for data augmentation. | `dict` | (see [dedicated section](#data-augmentation)) |
diff --git a/tests/conftest.py b/tests/conftest.py
index e660869c..8fa1e79c 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -70,8 +70,6 @@ def training_config():
"load_in_memory": True, # Load all images in CPU memory
"width_divisor": 8, # Image width will be divided by 8
"height_divisor": 32, # Image height will be divided by 32
- "padding_value": 0, # Image padding value
- "padding_token": None, # Label padding value
"preprocessings": [
{
"type": "to_RGB",
--
GitLab