README.md

@@ -4,7 +4,6 @@
 
 For more details about this package, make sure to see the documentation available at https://teklia.gitlab.io/atr/dan/.
 
-
 ## Installation
 
 To use DAN in your own scripts, install it using pip:
@@ -55,7 +54,9 @@ See the [dedicated section](https://teklia.gitlab.io/atr/dan/usage/datasets/form
 See the [dedicated section](https://teklia.gitlab.io/atr/dan/usage/train/) on the official DAN documentation.
 
 ### Synthetic data generation
+
 See the [dedicated section](https://teklia.gitlab.io/atr/dan/usage/generate/) on the official DAN documentation.
 
 ### Model prediction
+
 See the [dedicated section](https://teklia.gitlab.io/atr/dan/usage/predict/) on the official DAN documentation.

dan/manager/dataset.py

@@ -272,13 +272,6 @@ class GenericDataset(Dataset):
                 )
                 if load_in_memory:
                     samples[-1]["img"] = GenericDataset.load_image(filename)
-                if type(gt[filename]) is dict:
-                    if "lines" in gt[filename].keys():
-                        samples[-1]["raw_line_seg_label"] = gt[filename]["lines"]
-                    if "paragraphs" in gt[filename].keys():
-                        samples[-1]["paragraphs_label"] = gt[filename]["paragraphs"]
-                    if "pages" in gt[filename].keys():
-                        samples[-1]["pages_label"] = gt[filename]["pages"]
         return samples
 
     def apply_preprocessing(self, preprocessings):
@@ -421,15 +414,6 @@ def apply_preprocessing(sample, preprocessings):
                 temp_img = np.expand_dims(temp_img, axis=2)
             img = temp_img
             resize_ratio = [ratio, ratio]
-    if resize_ratio != [1, 1] and "raw_line_seg_label" in sample:
-        for li in range(len(sample["raw_line_seg_label"])):
-            for side, ratio in zip(
-                (["bottom", "top"], ["right", "left"]), resize_ratio
-            ):
-                for s in side:
-                    sample["raw_line_seg_label"][li][s] = (
-                        sample["raw_line_seg_label"][li][s] * ratio
-                    )
 
     sample["img"] = img
     sample["resize_ratio"] = resize_ratio

dan/manager/metrics.py

@@ -152,8 +152,6 @@ class MetricManager:
                 "loss",
                 "loss_ctc",
                 "loss_ce",
-                "syn_max_lines",
-                "syn_prob_lines",
             ]:
                 value = float(
                     np.average(
@@ -225,8 +223,6 @@ class MetricManager:
                 "loss_ctc",
                 "loss_ce",
                 "loss",
-                "syn_max_lines",
-                "syn_prob_lines",
             ]:
                 metrics[metric_name] = [
                     values[metric_name],

dan/manager/training.py

@@ -4,8 +4,6 @@ import json
 import os
 import pickle
 import random
-import sys
-from datetime import date
 from time import time
 
 import numpy as np
@@ -523,7 +521,6 @@ class GenericTrainingManager:
             return
         params = copy.deepcopy(self.params)
         params = class_to_str_dict(params)
-        params["date"] = date.today().strftime("%d/%m/%Y")
         total_params = 0
         for model_name in self.models.keys():
             current_params = compute_nb_params(self.models[model_name])
@@ -533,21 +530,6 @@ class GenericTrainingManager:
             ]
             total_params += current_params
         params["model_params"]["total_params"] = "{:,}".format(total_params)
-
-        params["hardware"] = dict()
-        if self.device != "cpu":
-            for i in range(self.params["training_params"]["nb_gpu"]):
-                params["hardware"][str(i)] = "{} {}".format(
-                    torch.cuda.get_device_name(i), torch.cuda.get_device_properties(i)
-                )
-        else:
-            params["hardware"]["0"] = "CPU"
-        params["software"] = {
-            "python_version": sys.version,
-            "pytorch_version": torch.__version__,
-            "cuda_version": torch.version.cuda,
-            "cudnn_version": torch.backends.cudnn.version(),
-        }
         with open(path, "w") as f:
             json.dump(params, f, indent=4)
 
@@ -871,8 +853,9 @@ class GenericTrainingManager:
             with open(path, "w") as f:
                 yaml.dump(metrics, stream=f)
 
-            # Log mlflow artifacts
-            mlflow.log_artifact(path, "predictions")
+            if mlflow_logging:
+                # Log mlflow artifacts
+                mlflow.log_artifact(path, "predictions")
 
     def output_pred(self, name):
         path = os.path.join(
@@ -1104,14 +1087,7 @@ class Manager(OCRManager):
         reduced_size = [s[:2] for s in batch_data["imgs_reduced_shape"]]
         y_len = batch_data["labels_len"]
 
-        # add errors in teacher forcing
-        if (
-            "teacher_forcing_error_rate" in self.params["training_params"]
-            and self.params["training_params"]["teacher_forcing_error_rate"] is not None
-        ):
-            error_rate = self.params["training_params"]["teacher_forcing_error_rate"]
-            simulated_y_pred, y_len = self.add_label_noise(y, y_len, error_rate)
-        elif "label_noise_scheduler" in self.params["training_params"]:
+        if "label_noise_scheduler" in self.params["training_params"]:
             error_rate = (
                 self.params["training_params"]["label_noise_scheduler"][
                     "min_error_rate"
@@ -1188,12 +1164,6 @@ class Manager(OCRManager):
             "str_x": str_x,
             "loss": sum_loss.item(),
             "loss_ce": loss_ce.item(),
-            "syn_max_lines": self.dataset.train_dataset.get_syn_max_lines()
-            if self.params["dataset_params"]["config"]["synthetic_data"]
-            else 0,
-            "syn_prob_lines": self.dataset.train_dataset.get_syn_proba_lines()
-            if self.params["dataset_params"]["config"]["synthetic_data"]
-            else 0,
         }
 
         return values
@@ -1247,10 +1217,6 @@ class Manager(OCRManager):
             else:
                 features = self.models["encoder"](x)
             features_size = features.size()
-            coverage_vector = torch.zeros(
-                (features.size(0), 1, features.size(2), features.size(3)),
-                device=self.device,
-            )
             pos_features = self.models["decoder"].features_updater.get_pos_features(
                 features
             )
@@ -1279,7 +1245,6 @@ class Manager(OCRManager):
                 confidence_scores.append(
                     torch.max(torch.softmax(pred[:, :], dim=1), dim=1).values
                 )
-                coverage_vector = torch.clamp(coverage_vector + weights, 0, 1)
                 predicted_tokens = torch.cat(
                     [
                         predicted_tokens,

dan/ocr/document/train.py

@@ -212,8 +212,6 @@ def get_config():
                 "cer",
                 "wer",
                 "wer_no_punct",
-                "syn_max_lines",
-                "syn_prob_lines",
             ],  # Metrics name for training
             "eval_metrics": [
                 "cer",

dan/predict/prediction.py

@@ -148,10 +148,6 @@ class DAN:
 
             features = self.encoder(input_tensor.float())
             features_size = features.size()
-            coverage_vector = torch.zeros(
-                (features.size(0), 1, features.size(2), features.size(3)),
-                device=self.device,
-            )
             pos_features = self.decoder.features_updater.get_pos_features(features)
             features = torch.flatten(pos_features, start_dim=2, end_dim=3).permute(
                 2, 0, 1
@@ -179,7 +175,6 @@ class DAN:
                 confidence_scores.append(
                     torch.max(torch.softmax(pred[:, :], dim=1), dim=1).values
                 )
-                coverage_vector = torch.clamp(coverage_vector + weights, 0, 1)
                 predicted_tokens = torch.cat(
                     [
                         predicted_tokens,