docling-ibm-models 1.1.7__tar.gz → 1.2.0__tar.gz

{docling_ibm_models-1.1.7 → docling_ibm_models-1.2.0}/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-1.1.7 → docling_ibm_models-1.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: docling-ibm-models
-Version: 1.1.7
+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-1.1.7 → docling_ibm_models-1.2.0}/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-1.1.7 → docling_ibm_models-1.2.0}/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-1.1.7 → docling_ibm_models-1.2.0}/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-1.2.0/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-1.1.7 → docling_ibm_models-1.2.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "docling-ibm-models"
-version = "1.1.7"  # DO NOT EDIT, updated automatically
+version = "1.2.0"  # DO NOT EDIT, updated automatically
 description = "This package contains the AI models used by the Docling PDF conversion package"
 authors = ["Nikos Livathinos <nli@zurich.ibm.com>", "Maxim Lysak <mly@zurich.ibm.com>", "Ahmed Nassar <ahn@zurich.ibm.com>", "Christoph Auer <cau@zurich.ibm.com>", "Michele Dolfi <dol@zurich.ibm.com>", "Peter Staar <taa@zurich.ibm.com>"]
 license = "MIT"

docling_ibm_models-1.1.7/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"