wrap layout models for prediction (image resize or tiling) all in TF

(to avoid back and forth between CPU and GPU memory when looping
 over image patches)

- `patch_encoder`: define `Model` subclasses which take an existing
  (layout segmentation) model in the constructor, and define a new
  `call()` using the existing model in a GPU-only `tf.function`:
  * `wrap_layout_model_resized`: just `tf.image.resize()` from
    input image to model size, then predict, then resize back
  * `wrap_layout_model_patched`: ditto if smaller than model size;
    otherwise use `tf.image.extract_patches` for patching in a
    sliding-window approach, then predict patches one by one, then
    `tf.scatter_nd` to reconstruct to image size
- when compiling `tf.function` graph, make sure to use input signature
  with variable image size, but avoid retracing each new size sample
- in `EynollahModelZoo.load_model` for relevant model types,
  also wrap the loaded model
  * by `wrap_layout_model_resized` under model name + `_resized`
  * by `wrap_layout_model_patched` under model name + `_patched`
- introduce `do_prediction_new_concept_autosize`,
  replacing `do_prediction/_new_concept`,
  but using passed model's `predict` directly without
  resizing or tiling to model size
- instead of `do_prediction/_new_concept(True, ...)`,
  now call `do_prediction_new_concept_autosize`,
  but with `_patched` appended to model name
- instead of `do_prediction/_new_concept(False, ...)`,
  now call `do_prediction_new_concept_autosize`,
  but with `_resized` appended to model name

src/eynollah/eynollah.py

@@ -914,8 +914,34 @@ class Eynollah:
         gc.collect()
         return prediction, confidence
 
+    def do_prediction_new_concept_autosize(
+            self, img, model,
+            thresholding_for_heading=False,
+            thresholding_for_artificial_class=False,
+            threshold_art_class=0.1,
+            artificial_class=4,
+    ):
 
+        self.logger.debug("enter do_prediction_new_concept (%s)", model.name)
+        img = img / 255.0
+
+        prediction = model.predict(img[np.newaxis]).numpy()[0]
+        confidence = prediction[:, :, 1]
+        segmentation = np.argmax(prediction, axis=2).astype(np.uint8)
+
+        if thresholding_for_artificial_class:
+            seg_mask_label(segmentation,
+                           prediction[:, :, artificial_class] >= threshold_art_class,
+                           label=artificial_class,
+                           only=True,
+                           skeletonize=True,
+                           dilate=3)
+        if thresholding_for_heading:
+            seg_mask_label(segmentation,
+                           prediction[:, :, 2] >= 0.2,
+                           label=2)
         gc.collect()
+        return segmentation, confidence
 
     def extract_page(self):
         self.logger.debug("enter extract_page")
@@ -990,7 +1016,9 @@ class Eynollah:
         self.logger.debug("enter extract_text_regions")
         img_height_h = img.shape[0]
         img_width_h = img.shape[1]
-        model_region = self.model_zoo.get("region_fl") if patches else self.model_zoo.get("region_fl_np")
+        #model_name = "region_fl" if patches else "region_fl_np"
+        model_name = "region_fl_patched" if patches else "region_fl_np_resized"
+        model_region = self.model_zoo.get(model_name)
 
         thresholding_for_heading = True
         img = otsu_copy_binary(img).astype(np.uint8)
@@ -1010,9 +1038,8 @@ class Eynollah:
             else:
                 img = resize_image(img, int(img_height_h * 2500 / float(img_width_h)), 2500).astype(np.uint8)
 
-        prediction_regions = self.do_prediction(patches, img, model_region,
+        prediction_regions, _ = self.do_prediction_new_concept_autosize(
-                                                marginal_of_patch_percent=0.1,
+            img, model_region,
-                                                n_batch_inference=3,
             thresholding_for_heading=thresholding_for_heading)
         prediction_regions = resize_image(prediction_regions, img_height_h, img_width_h)
         self.logger.debug("exit extract_text_regions")
@@ -1162,9 +1189,10 @@ class Eynollah:
     def textline_contours(self, img, use_patches, num_col_classifier=None):
         self.logger.debug('enter textline_contours')
 
-        prediction_textline = self.do_prediction(use_patches, img, self.model_zoo.get("textline"),
+        prediction_textline, _ = self.do_prediction_new_concept_autosize(
-                                                 marginal_of_patch_percent=0.15,
+            img, self.model_zoo.get("textline_patched" if use_patches else "textline_resized"),
-                                                 n_batch_inference=3,
+            artificial_class=2,
+            thresholding_for_artificial_class=True,
             threshold_art_class=self.threshold_art_class_textline)
 
         #prediction_textline_longshot = self.do_prediction(False, img, self.model_zoo.get("textline"))
@@ -1242,15 +1270,17 @@ class Eynollah:
             if self.image_org.shape[0]/self.image_org.shape[1] > 2.5:
                 self.logger.debug("resized to %dx%d for %d cols",
                                   img_resized.shape[1], img_resized.shape[0], num_col_classifier)
-                prediction_regions_org, confidence_matrix = self.do_prediction_new_concept(
+                prediction_regions_org, confidence_matrix = \
-                    True, img_resized, self.model_zoo.get("region_1_2"), n_batch_inference=1,
+                    self.do_prediction_new_concept_autosize(
+                        img_resized, self.model_zoo.get("region_1_2_patched"),
                         thresholding_for_artificial_class=True,
                         threshold_art_class=self.threshold_art_class_layout)
             else:
                 prediction_regions_org = np.zeros((self.image_org.shape[0], self.image_org.shape[1]))
                 confidence_matrix = np.zeros((self.image_org.shape[0], self.image_org.shape[1]))
-                prediction_regions_page, confidence_matrix_page = self.do_prediction_new_concept(
+                prediction_regions_page, confidence_matrix_page = \
-                    False, self.image_page_org_size, self.model_zoo.get("region_1_2"), n_batch_inference=1,
+                    self.do_prediction_new_concept_autosize(
+                        self.image_page_org_size, self.model_zoo.get("region_1_2_resized"),
                         thresholding_for_artificial_class=True,
                         threshold_art_class=self.threshold_art_class_layout)
                 ys = slice(*self.page_coord[0:2])
@@ -1263,8 +1293,9 @@ class Eynollah:
             img_resized = resize_image(img_bin, int(new_h * img_bin.shape[0] /img_bin.shape[1]), new_h)
             self.logger.debug("resized to %dx%d (new_h=%d) for %d cols",
                               img_resized.shape[1], img_resized.shape[0], new_h, num_col_classifier)
-            prediction_regions_org, confidence_matrix = self.do_prediction_new_concept(
+            prediction_regions_org, confidence_matrix = \
-                True, img_resized, self.model_zoo.get("region_1_2"), n_batch_inference=2,
+                self.do_prediction_new_concept_autosize(
+                    img_resized, self.model_zoo.get("region_1_2_patched"),
                     thresholding_for_artificial_class=True,
                     threshold_art_class=self.threshold_art_class_layout)
         ###prediction_regions_org = self.do_prediction(True, img_bin, self.model_zoo.get_model("region"),
@@ -1664,8 +1695,7 @@ class Eynollah:
         img_org = np.copy(img)
         img_height_h = img_org.shape[0]
         img_width_h = img_org.shape[1]
-        patches = False
+        prediction_table, _ = self.do_prediction_new_concept_autosize(img, self.model_zoo.get("table_resized"))
-        prediction_table, _ = self.do_prediction_new_concept(patches, img, self.model_zoo.get("table"))
         prediction_table = prediction_table.astype(np.int16)
         return prediction_table
 

src/eynollah/model_zoo/model_zoo.py

@@ -14,7 +14,12 @@ from tensorflow.keras.models import Model as KerasModel
 from tensorflow.keras.models import load_model
 from tabulate import tabulate
 
-from ..patch_encoder import PatchEncoder, Patches
+from ..patch_encoder import (
+    PatchEncoder,
+    Patches,
+    wrap_layout_model_patched,
+    wrap_layout_model_resized,
+)
 from .specs import EynollahModelSpecSet
 from .default_specs import DEFAULT_MODEL_SPECS
 from .types import AnyModel, T
@@ -125,7 +130,12 @@ class EynollahModelZoo:
                 model = load_model(
                     model_path, compile=False, custom_objects={"PatchEncoder": PatchEncoder, "Patches": Patches}
                 )
+            model._name = model_category
         self._loaded[model_category] = model
+        if model_category in ['region_1_2', 'table', 'region_fl_np']:
+            self._loaded[model_category + '_resized'] = wrap_layout_model_resized(model)
+        if model_category in ['region_1_2', 'textline']:
+            self._loaded[model_category + '_patched'] = wrap_layout_model_patched(model)
         return model  # type: ignore
 
     def get(self, model_category: str, model_type: Optional[Type[T]] = None) -> T:

src/eynollah/patch_encoder.py

@@ -1,7 +1,7 @@
 import os
 os.environ['TF_USE_LEGACY_KERAS'] = '1' # avoid Keras 3 after TF 2.15
 import tensorflow as tf
-from tensorflow.keras import layers
+from tensorflow.keras import layers, models
 
 class PatchEncoder(layers.Layer):
 
@@ -44,3 +44,117 @@ class Patches(layers.Layer):
         return dict(patch_size_x=self.patch_size_x,
                     patch_size_y=self.patch_size_y,
                     **super().get_config())
+
+class wrap_layout_model_resized(models.Model):
+    """
+    replacement for layout model using resizing to model width/height and back
+
+    (accepts arbitrary width/height input [B, H, W, 3], returns same size segmentation [B, H, W, C])
+    """
+    def __init__(self, model):
+        super().__init__(name=model.name + '_resized')
+        self.model = model
+        self.height = model.layers[-1].output_shape[1]
+        self.width = model.layers[-1].output_shape[2]
+
+    def call(self, img, training=False):
+        height = tf.shape(img)[1]
+        width = tf.shape(img)[2]
+        img_resized = tf.image.resize(img,
+                                      (self.height, self.width),
+                                      antialias=True)
+        pred_resized = self.model(img_resized)
+        pred = tf.image.resize(pred_resized,
+                               (height, width))
+        return pred
+
+    @tf.function(reduce_retracing=True,
+                 #jit_compile=True, (ScaleAndTranslate is not supported by XLA)
+                 input_signature=[tf.TensorSpec([1, None, None, 3],
+                                                dtype=tf.float32)])
+    def predict(self, x):
+        return self(x)
+
+class wrap_layout_model_patched(models.Model):
+    """
+    replacement for layout model using sliding window for patches
+
+    (accepts arbitrary width/height input [B, H, W, 3], returns same size segmentation [B, H, W, C])
+    """
+    def __init__(self, model):
+        super().__init__(name=model.name + '_patched')
+        self.model = model
+        self.height = model.layers[-1].output_shape[1]
+        self.width = model.layers[-1].output_shape[2]
+        self.classes = model.layers[-1].output_shape[3]
+        # equivalent of marginal_of_patch_percent=0.1 ...
+        self.stride_x = int(self.width * (1 - 0.1))
+        self.stride_y = int(self.height * (1 - 0.1))
+        offset_height = (self.height - self.stride_y) // 2
+        offset_width = (self.width - self.stride_x) // 2
+        window = tf.image.pad_to_bounding_box(
+            tf.ones((self.stride_y, self.stride_x, 1), dtype=tf.int32),
+            offset_height, offset_width,
+            self.height, self.width)
+        self.window = tf.expand_dims(window, axis=0)
+
+    def call(self, img, training=False):
+        height = tf.shape(img)[1]
+        width = tf.shape(img)[2]
+        if (height < self.height or
+            width < self.width):
+            img_resized = tf.image.resize(img,
+                                          (self.height, self.width),
+                                          antialias=True)
+            pred_resized = self.model(img_resized)
+            pred = tf.image.resize(pred_resized,
+                                   (height, width))
+            return pred
+
+        img_patches = tf.image.extract_patches(
+            images=img,
+            sizes=[1, self.height, self.width, 1],
+            strides=[1, self.stride_y, self.stride_x, 1],
+            rates=[1, 1, 1, 1],
+            padding='SAME')
+        img_patches = tf.squeeze(img_patches)
+        new_shape = (-1, self.height, self.width, 3)
+        img_patches = tf.reshape(img_patches, shape=new_shape)
+        # may be too large:
+        #pred_patches = self.model(img_patches)
+        # so rebatch to fit in memory:
+        img_patches = tf.expand_dims(img_patches, 1)
+        pred_patches = tf.map_fn(self.model, img_patches,
+                                 parallel_iterations=1,
+                                 infer_shape=False)
+        pred_patches = tf.squeeze(pred_patches, 1)
+        # calculate corresponding indexes for reconstruction
+        x = tf.range(width)
+        y = tf.range(height)
+        x, y = tf.meshgrid(x, y)
+        indices = tf.stack([y, x], axis=-1)
+        indices_patches = tf.image.extract_patches(
+            images=tf.expand_dims(indices, axis=0),
+            sizes=[1, self.height, self.width, 1],
+            strides=[1, self.stride_y, self.stride_x, 1],
+            rates=[1, 1, 1, 1],
+            padding='SAME')
+        indices_patches = tf.squeeze(indices_patches)
+        indices_patches = tf.reshape(indices_patches, shape=new_shape[:-1] + (2,))
+
+        # use margins for sliding window approach
+        indices_patches = indices_patches * self.window
+
+        pred = tf.scatter_nd(
+            indices_patches,
+            pred_patches,
+            (height, width, self.classes))
+        pred = tf.expand_dims(pred, axis=0)
+        return pred
+
+    @tf.function(reduce_retracing=True,
+                 #jit_compile=True, (ScaleAndTranslate and ExtractImagePatches not supported by XLA)
+                 input_signature=[tf.TensorSpec([1, None, None, 3],
+                                                dtype=tf.float32)])
+    def predict(self, x):
+        return self(x)