docling-ibm-models 1.1.6__py3-none-any.whl → 1.2.0__py3-none-any.whl

docling_ibm_models/layoutmodel/layout_predictor.py

@@ -14,29 +14,6 @@ MODEL_CHECKPOINT_FN = "model.pt"
 DEFAULT_NUM_THREADS = 4
 
 
-# Classes:
-CLASSES_MAP = {
-    0: "background",
-    1: "Caption",
-    2: "Footnote",
-    3: "Formula",
-    4: "List-item",
-    5: "Page-footer",
-    6: "Page-header",
-    7: "Picture",
-    8: "Section-header",
-    9: "Table",
-    10: "Text",
-    11: "Title",
-    12: "Document Index",
-    13: "Code",
-    14: "Checkbox-Selected",
-    15: "Checkbox-Unselected",
-    16: "Form",
-    17: "Key-Value Region",
-}
-
-
 class LayoutPredictor:
     r"""
     Document layout prediction using ONNX
@@ -69,6 +46,31 @@ class LayoutPredictor:
         ------
         FileNotFoundError when the model's ONNX file is missing
         """
+        # Initialize classes map:
+        self._classes_map = {
+            0: "background",
+            1: "Caption",
+            2: "Footnote",
+            3: "Formula",
+            4: "List-item",
+            5: "Page-footer",
+            6: "Page-header",
+            7: "Picture",
+            8: "Section-header",
+            9: "Table",
+            10: "Text",
+            11: "Title",
+            12: "Document Index",
+            13: "Code",
+            14: "Checkbox-Selected",
+            15: "Checkbox-Unselected",
+            16: "Form",
+            17: "Key-Value Region",
+        }
+
+        # Blacklisted classes
+        self._black_classes = set(["Form", "Key-Value Region"])
+
         # Set basic params
         self._threshold = 0.6  # Score threshold
         self._image_size = 640
@@ -159,13 +161,19 @@ class LayoutPredictor:
         )
 
         # Yield output
-        for label, box, score in zip(labels[0], boxes[0], scores[0]):
+        for label_idx, box, score in zip(labels[0], boxes[0], scores[0]):
+            # Filter out blacklisted classes
+            label = self._classes_map[label_idx]
+            if label in self._black_classes:
+                continue
+
+            # Check against threshold
             if score > self._threshold:
                 yield {
                     "l": box[0] / self._image_size * w,
                     "t": box[1] / self._image_size * h,
                     "r": box[2] / self._image_size * w,
                     "b": box[3] / self._image_size * h,
-                    "label": CLASSES_MAP[label],
+                    "label": label,
                     "confidence": score,
                 }

docling_ibm_models/tableformer/data_management/matching_post_processor.py

@@ -4,6 +4,7 @@
 #
 import json
 import logging
+import math
 import statistics
 
 import docling_ibm_models.tableformer.settings as s
@@ -403,45 +404,63 @@ class MatchingPostProcessor:
                 # Push horizontally
                 if x1_min < x2_min:
                     # Move box1 to the left and box2 to the right
-                    box1["bbox"][2] -= overlap_x
-                    box2["bbox"][0] += overlap_x
+                    box1["bbox"][2] -= math.ceil(overlap_x / 2) + 2
+                    box2["bbox"][0] += math.floor(overlap_x / 2)
                 else:
                     # Move box2 to the left and box1 to the right
-                    box2["bbox"][2] -= overlap_x
-                    box1["bbox"][0] += overlap_x
+                    box2["bbox"][2] -= math.ceil(overlap_x / 2) + 2
+                    box1["bbox"][0] += math.floor(overlap_x / 2)
             else:
                 # Push vertically
                 if y1_min < y2_min:
                     # Move box1 up and box2 down
-                    box1["bbox"][3] -= overlap_y
-                    box2["bbox"][1] += overlap_y
+                    box1["bbox"][3] -= math.ceil(overlap_y / 2) + 2
+                    box2["bbox"][1] += math.floor(overlap_y / 2)
                 else:
                     # Move box2 up and box1 down
-                    box2["bbox"][3] -= overlap_y
-                    box1["bbox"][1] += overlap_y
+                    box2["bbox"][3] -= math.ceil(overlap_y / 2) + 2
+                    box1["bbox"][1] += math.floor(overlap_y / 2)
+
+            # Will flip coordinates in proper order, if previous operations reversed it
+            box1["bbox"] = [
+                min(box1["bbox"][0], box1["bbox"][2]),
+                min(box1["bbox"][1], box1["bbox"][3]),
+                max(box1["bbox"][0], box1["bbox"][2]),
+                max(box1["bbox"][1], box1["bbox"][3]),
+            ]
+            box2["bbox"] = [
+                min(box2["bbox"][0], box2["bbox"][2]),
+                min(box2["bbox"][1], box2["bbox"][3]),
+                max(box2["bbox"][0], box2["bbox"][2]),
+                max(box2["bbox"][1], box2["bbox"][3]),
+            ]
 
             return box1, box2
 
         def do_boxes_overlap(box1, box2):
-            # print("{} - {}".format(box1["bbox"], box2["bbox"]))
-            # Extract coordinates from the bounding boxes
-            x1_min, y1_min, x1_max, y1_max = box1["bbox"]
-            x2_min, y2_min, x2_max, y2_max = box2["bbox"]
-            # Check if one box is to the left of the other
-            if x1_max < x2_min or x2_max < x1_min:
+            B1 = box1["bbox"]
+            B2 = box2["bbox"]
+            if (
+                (B1[0] >= B2[2])
+                or (B1[2] <= B2[0])
+                or (B1[3] <= B2[1])
+                or (B1[1] >= B2[3])
+            ):
                 return False
-            # Check if one box is above the other
-            if y1_max < y2_min or y2_max < y1_min:
-                return False
-            return True
+            else:
+                return True
 
         def find_overlapping_pairs_indexes(bboxes):
             overlapping_indexes = []
             # Compare each box with every other box (combinations)
             for i in range(len(bboxes)):
                 for j in range(i + 1, len(bboxes)):
-                    if do_boxes_overlap(bboxes[i], bboxes[j]):
-                        bboxes[i], bboxes[j] = correct_overlap(bboxes[i], bboxes[j])
+                    if i != j:
+                        if bboxes[i] != bboxes[j]:
+                            if do_boxes_overlap(bboxes[i], bboxes[j]):
+                                bboxes[i], bboxes[j] = correct_overlap(
+                                    bboxes[i], bboxes[j]
+                                )
 
             return overlapping_indexes, bboxes
 
@@ -1144,7 +1163,7 @@ class MatchingPostProcessor:
                 new_pdf_cells.append(pdf_cells[i])
         return new_pdf_cells
 
-    def process(self, matching_details):
+    def process(self, matching_details, correct_overlapping_cells=False):
         r"""
         Do post processing, see details in the comments below
 
@@ -1348,9 +1367,10 @@ class MatchingPostProcessor:
         table_cells_wo = po2
         max_cell_id = po3
 
-        # As the last step - correct cell bboxes in a way that they don't overlap:
-        if len(table_cells_wo) <= 300:  # For performance reasons
-            table_cells_wo = self._find_overlapping(table_cells_wo)
+        if correct_overlapping_cells:
+            # As the last step - correct cell bboxes in a way that they don't overlap:
+            if len(table_cells_wo) <= 300:  # For performance reasons
+                table_cells_wo = self._find_overlapping(table_cells_wo)
 
         self._log().debug("*** final_matches_wo")
         self._log().debug(final_matches_wo)

docling_ibm_models/tableformer/data_management/tf_predictor.py

@@ -523,8 +523,9 @@ class TFPredictor:
         # return the resized image
         return resized, sf
 
-    def multi_table_predict(self, iocr_page, table_bboxes, do_matching=True):
-        # def multi_table_predict(self, iocr_page, page_image, table_bboxes):
+    def multi_table_predict(
+        self, iocr_page, table_bboxes, do_matching=True, correct_overlapping_cells=False
+    ):
         multi_tf_output = []
         page_image = iocr_page["image"]
 
@@ -546,7 +547,12 @@ class TFPredictor:
             # Predict
             if do_matching:
                 tf_responses, predict_details = self.predict(
-                    iocr_page, table_bbox, table_image, scale_factor, None
+                    iocr_page,
+                    table_bbox,
+                    table_image,
+                    scale_factor,
+                    None,
+                    correct_overlapping_cells,
                 )
             else:
                 tf_responses, predict_details = self.predict_dummy(
@@ -733,7 +739,13 @@ class TFPredictor:
         return tf_output, matching_details
 
     def predict(
-        self, iocr_page, table_bbox, table_image, scale_factor, eval_res_preds=None
+        self,
+        iocr_page,
+        table_bbox,
+        table_image,
+        scale_factor,
+        eval_res_preds=None,
+        correct_overlapping_cells=False,
     ):
         r"""
         Predict the table out of an image in memory
@@ -744,6 +756,8 @@ class TFPredictor:
             Docling provided table data
         eval_res_preds : dict
             Ready predictions provided by the evaluation results
+        correct_overlapping_cells : boolean
+            Enables or disables last post-processing step, that fixes cell bboxes to remove overlap
 
         Returns
         -------
@@ -834,7 +848,9 @@ class TFPredictor:
             ):  # There are at least some pdf cells to match with
                 if self.enable_post_process:
                     AggProfiler().begin("post_process", self._prof)
-                    matching_details = self._post_processor.process(matching_details)
+                    matching_details = self._post_processor.process(
+                        matching_details, correct_overlapping_cells
+                    )
                     AggProfiler().end("post_process", self._prof)
 
         # Generate the expected Docling responses

docling_ibm_models/tableformer/models/table04_rs/bbox_decoder_rs.py

@@ -157,7 +157,12 @@ class BBoxDecoder(nn.Module):
             predictions_classes.append(self._class_embed(h))
         if len(predictions_bboxes) > 0:
             predictions_bboxes = torch.stack([x[0] for x in predictions_bboxes])
+        else:
+            predictions_bboxes = torch.empty(0)
+
         if len(predictions_classes) > 0:
             predictions_classes = torch.stack([x[0] for x in predictions_classes])
+        else:
+            predictions_classes = torch.empty(0)
 
         return predictions_classes, predictions_bboxes

docling_ibm_models/tableformer/models/table04_rs/transformer_rs.py

@@ -149,11 +149,11 @@ class Tag_Transformer(nn.Module):
         self._positional_encoding = PositionalEncoding(embed_dim)
         self._td_encode = td_encode
 
+        encoder_layer = nn.TransformerEncoderLayer(
+            d_model=embed_dim, nhead=n_heads, dim_feedforward=dim_ff
+        )
         self._encoder = nn.TransformerEncoder(
-            nn.TransformerEncoderLayer(
-                d_model=embed_dim, nhead=n_heads, dim_feedforward=dim_ff
-            ),
-            num_layers=encoder_layers,
+            encoder_layer, num_layers=encoder_layers, enable_nested_tensor=False
         )
 
         self._decoder = TMTransformerDecoder(

docling_ibm_models/tableformer/otsl.py

@@ -123,6 +123,9 @@ def otsl_check_right(rs_split, x, y):
 
 
 def otsl_to_html(rs_list, logdebug):
+    if len(rs_list) == 0:
+        return []
+
     if rs_list[0] not in ["fcel", "ched", "rhed", "srow", "ecel"]:
         # Most likely already HTML...
         return rs_list

docling_ibm_models/tableformer/utils/app_profiler.py

@@ -6,6 +6,8 @@ import time
 from collections import deque
 from statistics import mean, median
 
+from docling_ibm_models.tableformer.utils.mem_monitor import MemMonitor
+
 
 class SingletonClass(type):
     r"""
@@ -37,11 +39,13 @@ class Profiler:
     def __init__(self):
         self._section_dts = {}  # section name -> sum(section intervals)
         self._section_calls = {}  # section name -> number of invocations
-        self._section_kB = {}  # section name -> max kB of used heap
+        self._section_kB = {}  # section name -> max kB of used heap (resident set size)
 
         # section name -> beginning of the last interval
         self._last_begin = {}
 
+        self._mem_monitor = MemMonitor()
+
     def begin(self, section_name, enable=True):
         r"""
         Mark the beginning of an interval
@@ -83,13 +87,20 @@ class Profiler:
         if section_name not in self._last_begin:
             return False
 
+        # Get memory
+        kB = self._mem_monitor.get_memory()
+        if isinstance(kB, dict):
+            kB = kB["resident"]
+
         dt = time.time() - self._last_begin[section_name]
         if section_name not in self._section_dts:
             self._section_dts[section_name] = dt
             self._section_calls[section_name] = 1
+            self._section_kB[section_name] = kB
         else:
             self._section_dts[section_name] += dt
             self._section_calls[section_name] += 1
+            self._section_kB[section_name] = max(kB, self._section_kB[section_name])
 
         return True
 

docling_ibm_models/tableformer/utils/mem_monitor.py

@@ -0,0 +1,175 @@
+#
+# Copyright IBM Corp. 2024 - 2024
+# SPDX-License-Identifier: MIT
+#
+import os
+import platform
+import re
+
+
+class MemMonitor:
+    r"""
+    Memory monitor for Linux
+
+    It supports 2 approaches for extracting memory information:
+    - linux-native: It parse the `/proc` pseudo-files. It is available only for Linux
+    - psutil: Use the `psutil` library
+
+    ## Linux-Native approach
+
+    The linux-native approach implements 2 methods to extract the memory fields:
+
+    1. The `get_memory()` method:
+
+    - It is very fast
+    - It parses the `/proc/<pid>/statm` pseudo-file
+    - It Contains the following fields:
+        size       (1) total program size
+                   (same as VmSize in /proc/[pid]/status)
+        resident   (2) resident set size
+                   (same as VmRSS in /proc/[pid]/status)
+        shared     (3) number of resident shared pages (i.e., backed by a file)
+                   (same as RssFile+RssShmem in /proc/[pid]/status)
+        text       (4) text (code)
+        lib        (5) library (unused since Linux 2.6; always 0)
+        data       (6) data + stack
+        dt         (7) dirty pages (unused since Linux 2.6; always 0)
+
+
+    2. The `get_memory_full()` method:
+
+    - It is slower to parse but contains more detailed information
+    - It uses regex to parse the `/proc/<pid>/status` pseudo-file
+    - It contains the following fields:
+        VmPeak: Peak virtual memory size.
+        VmSize: Virtual memory size.
+        VmLck: Locked memory size (see mlock(2)).
+        VmPin: Pinned memory size (since Linux 3.2). These are pages that can't be moved because
+               something needs to directly access physical memory.
+        VmHWM: Peak resident set size ("high water mark").
+        VmRSS: Resident set size.  Note that the value here is the sum of RssAnon, RssFile, and
+               RssShmem.
+        RssAnon: Size of resident anonymous memory.  (since Linux 4.5).
+        RssFile: Size of resident file mappings.  (since Linux 4.5).
+        RssShmem: Size of resident shared memory (includes System V shared memory, mappings from
+                  tmpfs(5), and shared anonymous mappings).  (since Linux 4.5).
+        VmData, VmStk, VmExe: Size of data, stack, and text segments.
+        VmLib: Shared library code size.
+        VmPTE: Page table entries size (since Linux 2.6.10).
+        VmPMD: Size of second-level page tables (added in Linux 4.0; removed in Linux 4.15).
+        VmSwap: Swapped-out virtual memory size by anonymous private pages; shmem swap usage is
+                not included (since Linux 2.6.34).
+
+
+    ## The psutil library
+
+    - Apparently the psutil library parses the `/proc/<pid>/statm`
+    - The memory_info() function returns the fields: rss, vms, shared, text, lib, data, dirty
+
+
+    ## Field mappings
+
+    These are the fields returned by psutil memory_info() and their mapping in the /proc files:
+    (I put ? when I am not 100% about the mapping)
+
+    | psutil  | /proc/$$/status    | /proc/$$/statm |
+    |---------|--------------------|----------------|
+    | rss     | VmRSS              | resident       |
+    | vms     | VmSize             | size           |
+    | shared  | RssFile + RssShmem | shared         |
+    | text    | VmExe ?            | text           |
+    | lib     | RssShmem ?         | lib            |
+    | data    | VmData + VmStk     | data           |
+    | dirty   | VmSwap ?           | dt             |
+
+    """
+
+    def __init__(self, enable=True):
+        self._enable = enable
+        self._pid = os.getpid()
+
+        # Create regex for each memory field of the /proc/status pseudo-file
+        self._status_fields = [
+            "VmPeak",
+            "VmSize",
+            "VmLck",
+            "VmPin",
+            "VmHWM",
+            "VmRSS",
+            "RssAnon",
+            "RssFile",
+            "RssShmem",
+            "VmData",
+            "VmStk",
+            "VmExe",
+            "VmLib",
+            "VmPTE",
+            "VmPMD",
+            "VmSwap",
+        ]
+        self._status_regex = {}
+        for mem_field in self._status_fields:
+            regex_str = r"({}:)(\s+)(\d*)(.*)".format(mem_field)
+            self._status_regex[mem_field] = re.compile(regex_str)
+
+    def get_memory_full(self) -> dict:
+        r"""
+        - Parse /proc/<pid>status to get all memory info.
+        - The method returns a dict with the fields self._status_fields
+        - This method is SLOW. Unless you need the full memory info, better to use `get_memory`
+
+        The returned values are in kB
+        """
+        if not self._enable:
+            return -2
+        if platform.system() != "Linux":
+            return -1
+        pid_fn = "/proc/{}/status".format(self._pid)
+
+        # Dict to collect all memory fields
+        memory = {}
+        with open(pid_fn, "r") as fn:
+            for ll in fn:
+                for mem_field in self._status_fields:
+                    regex = self._status_regex[mem_field]
+                    m = regex.match(ll)
+                    if m is not None:
+                        memory[mem_field] = int(m.group(3))
+                if len(memory) == len(self._status_fields):
+                    break
+
+        return memory
+
+    def get_memory(self) -> dict:
+        r"""
+        - Parse /proc/<pid>statm to get the most important memory fields
+        - This is a fast implementation.
+        - The method returns a dict with the fields:
+            "size", "resident", "shared", "text", "lib", "data", "dt"
+        - Check the documentation at the top for a mapping across the various fields
+
+        The returned values are in kB
+        """
+        if not self._enable:
+            return -2
+        if platform.system() != "Linux":
+            return -1
+        pid_fn = "/proc/{}/statm".format(self._pid)
+
+        # Dict to collect all memory fields
+        memory = {}
+        with open(pid_fn, "r") as fn:
+            ll = fn.read()
+            # The values are in pages
+            # Each page is 4096 bytes (4kB)
+            data = [int(x) << 2 for x in ll.split(" ")]
+            memory = {
+                "size": data[0],
+                "resident": data[1],
+                "shared": data[2],
+                "text": data[3],
+                "lib": data[4],
+                "data": data[5],
+                "dt": data[6],
+            }
+        return memory

{docling_ibm_models-1.1.6.dist-info → docling_ibm_models-1.2.0.dist-info}/LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) [year] [fullname]
+Copyright (c) 2024 International Business Machines
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

{docling_ibm_models-1.1.6.dist-info → docling_ibm_models-1.2.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: docling-ibm-models
-Version: 1.1.6
+Version: 1.2.0
 Summary: This package contains the AI models used by the Docling PDF conversion package
 License: MIT
 Keywords: docling,convert,document,pdf,layout model,segmentation,table structure,table former

{docling_ibm_models-1.1.6.dist-info → docling_ibm_models-1.2.0.dist-info}/RECORD

@@ -1,32 +1,32 @@
-docling_ibm_models/layoutmodel/layout_predictor.py,sha256=y1Z51hUKju8SCvejCWWr04xbV_CmJrJ32FgA56tkgNU,5565
+docling_ibm_models/layoutmodel/layout_predictor.py,sha256=JHZbh6HyA2fLqaN0p9Lv3Y9P9dgkeHUqQI-JyyetocE,6042
 docling_ibm_models/tableformer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 docling_ibm_models/tableformer/common.py,sha256=RV2ptqgkfz1OIoN-WqiSeln0pkZ_7zTO9DhOcbvPS5k,6023
 docling_ibm_models/tableformer/data_management/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 docling_ibm_models/tableformer/data_management/data_transformer.py,sha256=lNKkAk0VALbixapCuDDSIQKtA0QPCGQF8AGO3D64new,18263
 docling_ibm_models/tableformer/data_management/functional.py,sha256=UrXsEm4DSc1QXdUPb0tZ7nvbg7mGVjpQhX3pGL6C5bA,20633
-docling_ibm_models/tableformer/data_management/matching_post_processor.py,sha256=-82B4xUJ9uxMDcsX2DJINTy3J0OB9rKXzHKtf-J3GHI,57205
+docling_ibm_models/tableformer/data_management/matching_post_processor.py,sha256=41GLMlkMAY1pkc-elP3ktFgZLCHjscghaHfgIVn2168,57998
 docling_ibm_models/tableformer/data_management/tf_cell_matcher.py,sha256=kzOjSmXkYrxc0de8wHbDJMvwKXelxYf4OccHTRqnpco,21081
 docling_ibm_models/tableformer/data_management/tf_dataset.py,sha256=6_qSsYt6qoE2JBzUNrJfCDX3Kgg7tyrv3kimGLdEQ5o,49890
-docling_ibm_models/tableformer/data_management/tf_predictor.py,sha256=O0cDcOWJU39aKoQi1oDKAdLmilhmNBoTZzt1TWdJ8m4,39334
+docling_ibm_models/tableformer/data_management/tf_predictor.py,sha256=Ha--59Rfs3V78p3q__q5cuEoewrTld18qhX8VqAQrYc,39730
 docling_ibm_models/tableformer/data_management/transforms.py,sha256=_i1HXkX8LAuHbeGRrg8kF9yFNJRQZOKmWzxKt559ABQ,13268
 docling_ibm_models/tableformer/models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 docling_ibm_models/tableformer/models/common/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 docling_ibm_models/tableformer/models/common/base_model.py,sha256=SbCjeEvDmGnyoKYhB5pYeg2LFVQdArglfrhqkuW1nUw,10030
 docling_ibm_models/tableformer/models/table04_rs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-docling_ibm_models/tableformer/models/table04_rs/bbox_decoder_rs.py,sha256=mMZSGk0PjQ4_fhuP44_WJVzfxyLky6S3zlVxCrNKRgc,5991
+docling_ibm_models/tableformer/models/table04_rs/bbox_decoder_rs.py,sha256=JV9rFh9caT3qnwWlZ0CZpw5aiiNzyTbfVp6H6JMxS0Q,6117
 docling_ibm_models/tableformer/models/table04_rs/encoder04_rs.py,sha256=iExmqJ0Pn0lJU3nWb_x8abTn42GctMqE55_YA2ppgvc,1975
 docling_ibm_models/tableformer/models/table04_rs/tablemodel04_rs.py,sha256=7iGkrTNLzjC1yn1zuA3N6DvBvbrcO_BR5tmHG3RKmXs,12159
-docling_ibm_models/tableformer/models/table04_rs/transformer_rs.py,sha256=4106qxxH0w92CVOFzFuCb87tRMvqAUP3X3F1WT5Z47A,6371
-docling_ibm_models/tableformer/otsl.py,sha256=k8l1hYWvcCkcnWbLxuBUYEcigYBFTRqiM2GBAHcUDok,21024
+docling_ibm_models/tableformer/models/table04_rs/transformer_rs.py,sha256=nhnYFlXT5KyJMdB4qMo5r8GimWXVy0lcqcmoHPEl-KE,6416
+docling_ibm_models/tableformer/otsl.py,sha256=oE_s2QHTE74jXD0vsXCuya_woReabUOBg6npprEqt58,21069
 docling_ibm_models/tableformer/settings.py,sha256=UlpsP0cpJZR2Uk48lgysYy0om3fr8Xt3z1xzvlTw5j4,3067
 docling_ibm_models/tableformer/test_dataset_cache.py,sha256=zvVJvUnYz4GxAQfPUmLTHUbqj0Yhi2vwgOBnsRgt1rI,818
 docling_ibm_models/tableformer/test_prepare_image.py,sha256=oPmU93-yWIkCeUYulGQ1p676Vq-zcjw2EX24WA5lspA,3155
 docling_ibm_models/tableformer/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-docling_ibm_models/tableformer/utils/app_profiler.py,sha256=13dvwo5byzfP2ejqGBFwtOWNnS-0EZVUrjTpeevbke8,7993
+docling_ibm_models/tableformer/utils/app_profiler.py,sha256=Pb7o1zcikKXh7ninaNt4_nVa1xuUrogZxbTr6U6jkEE,8392
+docling_ibm_models/tableformer/utils/mem_monitor.py,sha256=ycZ07fUBVVKKLTVGF54jGPDM2aTkKuZWk1kMbOS0wwQ,6353
 docling_ibm_models/tableformer/utils/torch_utils.py,sha256=uN0rK9mSXy1ewBnBnILrWebJhhVU4N-XJZBqNiLJwlQ,8893
 docling_ibm_models/tableformer/utils/utils.py,sha256=8Bxf1rEn977lFbY9NX0r5xh9PvxIRipQZX_EZW92XfA,10980
-docling_ibm_models/tableformer/utils/variance.py,sha256=USjRwaMsCmzvc6PeWskaAJnUjbliRVd_MqNKLjMDQw8,4675
-docling_ibm_models-1.1.6.dist-info/LICENSE,sha256=ACwmltkrXIz5VsEQcrqljq-fat6ZXAMepjXGoe40KtE,1069
-docling_ibm_models-1.1.6.dist-info/METADATA,sha256=3TIbhffZPFPwIH8oaKnEviMergfpg4Wppoas7bC36yU,7172
-docling_ibm_models-1.1.6.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
-docling_ibm_models-1.1.6.dist-info/RECORD,,
+docling_ibm_models-1.2.0.dist-info/LICENSE,sha256=mBb7ErEcM8VS9OhiGHnQ2kk75HwPhr54W1Oiz3965MY,1088
+docling_ibm_models-1.2.0.dist-info/METADATA,sha256=j_ccZliZ-e99bOg1MVoshV2f_ZxmfqKsIE-JRW2N2tI,7172
+docling_ibm_models-1.2.0.dist-info/WHEEL,sha256=sP946D7jFCHeNz5Iq4fL4Lu-PrWrFsgfLXbbkciIZwg,88
+docling_ibm_models-1.2.0.dist-info/RECORD,,

docling_ibm_models/tableformer/utils/variance.py

@@ -1,175 +0,0 @@
-#
-# Copyright IBM Corp. 2024 - 2024
-# SPDX-License-Identifier: MIT
-#
-import logging
-
-import numpy as np
-
-import docling_ibm_models.tableformer.settings as s
-
-LOG_LEVEL = logging.INFO
-
-
-class MyWelford:
-    r"""
-    Running computation of the sample mean and sample variance using Welford's algorithm
-    """
-
-    def __init__(self):
-        self._i = 0  # Running index
-        self._m = 0  # Running mean
-        self._s = 0  # (n - 1) * variance
-
-    def reset(self):
-        r"""
-        Reset the object
-        """
-        self._i = 0
-        self._m = 0
-        self._s = 0
-
-    def add(self, xi):
-        r"""
-        Invoke add each time a new sample arrives
-
-        Inputs:
-           xi: The next sample of data
-        """
-        self._i += 1
-        old_m = self._m
-        self._m = self._m + (xi - self._m) / self._i
-        self._s = self._s + (xi - self._m) * (xi - old_m)
-
-    def results(self):
-        r"""
-        Get the computed mean, variance and standard deviation up to now
-
-        Outputs:
-            m: Sample mean
-            v: Sample variance
-            std: Sample standard deviation
-        """
-        if self._i <= 1:
-            return None, None, None
-
-        # v = self._s / (self._i - 1)  # Sample variance
-        v = self._s / (self._i)  # Population variance
-        std = np.sqrt(v)
-        return self._m, v, std
-
-
-class MyWelfordImg(MyWelford):
-    r"""
-    Welford algorithm to calculate running mean and sample variance for images
-    """
-
-    def __init__(self):
-        super(MyWelfordImg, self).__init__()
-
-    def add(self, img):
-        r"""
-        Input:
-            img: An image numpy array (channel, width, height). The only requirement is to have the
-                 channels as the first dimension and have 3 dimensions in total
-        """
-        channels = img.shape[0]
-        flat_dim = img.shape[1] * img.shape[2]
-        img_r = img.reshape(channels, flat_dim)
-
-        for i in range(flat_dim):
-            super(MyWelfordImg, self).add(img_r[:, i])
-
-
-class ChanVarianceImg:
-    r"""
-    Chan's algorithm to compute a running variance with support of sub-samples
-    In this implementation each sub-sample is an images
-
-    Math for the original paper:
-    https://github.ibm.com/nli/variance_formulae
-    """
-
-    def __init__(self):
-        r""" """
-        self._first = True
-        # Size of the calculated dataset
-        self._n = 0
-        # Sum of the samples for the 3 image channels
-        self._t = 0
-        # Sum of the square differences of the deviations of the samples from the mean
-        self._s = 0
-
-    def add(self, img):
-        r"""
-        Add the provided image to the computation of the dataset statistics
-
-        Input:
-            img: An image numpy array (channel, width, height). The only requirement is to have the
-                 channels as the first dimension and have 3 dimensions in total
-        """
-        ch = img.shape[0]
-        n = img.shape[1] * img.shape[2]
-        img = img.reshape(ch, n)
-        img_t = img.sum(axis=1)
-        img_t_v = img_t.reshape(ch, 1)
-        diff = (img - (img_t_v / n)) ** 2
-        img_s = diff.sum(axis=1)
-
-        if not self._first:
-            c = (self._n / (n * (self._n + n))) * (
-                ((n / self._n) * self._t - img_t) ** 2
-            )
-            self._s += img_s + c
-            self._t += img_t
-        else:
-            self._s = img_s
-            self._t = img_t
-            self._first = False
-        self._n += n
-
-    def results(self):
-        r"""
-        Get the computed statistics
-
-        Output:
-            mean: Mean for the complete dataset
-            var: Population variance for the complete dataset
-            std: Population standard deviation for the complete dataset
-        """
-        mean = list(self._t / self._n)
-        var = list(self._s / self._n)  # Population variance
-        std = list(np.sqrt(var))
-
-        return mean, var, std
-
-    def reset(self):
-        r"""
-        Reset the object to start over again
-        """
-        self._n = 0
-        self._t = 0
-        self._s = 0
-        self._first = True
-
-
-if __name__ == "__main__":
-    logger = s.get_custom_logger("variance", LOG_LEVEL)
-
-    n = 50000
-    channels = 3
-    width = 448
-    height = 448
-
-    my = ChanVarianceImg()
-    # Generate random images
-    for i in range(n):
-        logger.info(i)
-        img = 255 * np.random.rand(channels, width, height)
-        my.add(img)
-
-    # Calculate the statistics
-    m, v, std = my.results()
-    assert m.shape == (3,), "Wrong mean dimension"
-    assert v.shape == (3,), "Wrong variance dimension"
-    assert std.shape == (3,), "Wrong std dimension"