fix: for CUDA version >= 12.6

.gitignore

@@ -143,4 +143,7 @@ grounding/config/configs
 grounding/version.py
 
 vis/
-tmp/
+tmp/
+data/
+
+*.pth

gdino_checkpoints/download_ckpts.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+# Define the URLs for the checkpoints
+BASE_URL="https://github.com/IDEA-Research/GroundingDINO/releases/download/"
+swint_ogc_url="${BASE_URL}v0.1.0-alpha/groundingdino_swint_ogc.pth"
+swinb_cogcoor_url="${BASE_URL}v0.1.0-alpha2/groundingdino_swinb_cogcoor.pth"
+
+
+
+# Download each of the four checkpoints using wget
+echo "Downloading groundingdino_swint_ogc.pth checkpoint..."
+wget $swint_ogc_url || { echo "Failed to download checkpoint from $swint_ogc_url"; exit 1; }
+
+echo "Downloading groundingdino_swinb_cogcoor.pth checkpoint..."
+wget $swinb_cogcoor_url || { echo "Failed to download checkpoint from $swinb_cogcoor_url"; exit 1; }
+
+echo "All checkpoints are downloaded successfully."

groundingdino/models/GroundingDINO/backbone/swin_transformer.py

@@ -16,7 +16,7 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 import torch.utils.checkpoint as checkpoint
-from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+from timm.layers import DropPath, to_2tuple, trunc_normal_
 
 from groundingdino.util.misc import NestedTensor
 
@@ -445,7 +445,7 @@ class BasicLayer(nn.Module):
         for blk in self.blocks:
             blk.H, blk.W = H, W
             if self.use_checkpoint:
-                x = checkpoint.checkpoint(blk, x, attn_mask)
+                x = checkpoint.checkpoint(blk, x, attn_mask, use_reentrant=True)
             else:
                 x = blk(x, attn_mask)
         if self.downsample is not None:

groundingdino/models/GroundingDINO/csrc/MsDeformAttn/ms_deform_attn_cuda.cu

@@ -15,11 +15,24 @@
 #include <ATen/cuda/CUDAContext.h>
 #include <cuda.h>
 #include <cuda_runtime.h>
+#include <torch/extension.h>
+#include <torch/version.h>
+
+// Check PyTorch version and define appropriate macros
+#if TORCH_VERSION_MAJOR >= 2 && TORCH_VERSION_MINOR >= 6
+  // PyTorch 2.x and above
+  #define GET_TENSOR_TYPE(x) x.scalar_type()
+  #define IS_CUDA_TENSOR(x) x.device().is_cuda()
+#else
+  // PyTorch 1.x
+  #define GET_TENSOR_TYPE(x) x.type()
+  #define IS_CUDA_TENSOR(x) x.type().is_cuda()
+#endif
 
 namespace groundingdino {
 
 at::Tensor ms_deform_attn_cuda_forward(
-    const at::Tensor &value, 
+    const at::Tensor &value,
     const at::Tensor &spatial_shapes,
     const at::Tensor &level_start_index,
     const at::Tensor &sampling_loc,
@@ -32,11 +45,11 @@ at::Tensor ms_deform_attn_cuda_forward(
     AT_ASSERTM(sampling_loc.is_contiguous(), "sampling_loc tensor has to be contiguous");
     AT_ASSERTM(attn_weight.is_contiguous(), "attn_weight tensor has to be contiguous");
 
-    AT_ASSERTM(value.type().is_cuda(), "value must be a CUDA tensor");
-    AT_ASSERTM(spatial_shapes.type().is_cuda(), "spatial_shapes must be a CUDA tensor");
-    AT_ASSERTM(level_start_index.type().is_cuda(), "level_start_index must be a CUDA tensor");
-    AT_ASSERTM(sampling_loc.type().is_cuda(), "sampling_loc must be a CUDA tensor");
-    AT_ASSERTM(attn_weight.type().is_cuda(), "attn_weight must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(value), "value must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(spatial_shapes), "spatial_shapes must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(level_start_index), "level_start_index must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(sampling_loc), "sampling_loc must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(attn_weight), "attn_weight must be a CUDA tensor");
 
     const int batch = value.size(0);
     const int spatial_size = value.size(1);
@@ -51,7 +64,7 @@ at::Tensor ms_deform_attn_cuda_forward(
     const int im2col_step_ = std::min(batch, im2col_step);
 
     AT_ASSERTM(batch % im2col_step_ == 0, "batch(%d) must divide im2col_step(%d)", batch, im2col_step_);
-    
+
     auto output = at::zeros({batch, num_query, num_heads, channels}, value.options());
 
     const int batch_n = im2col_step_;
@@ -62,7 +75,7 @@ at::Tensor ms_deform_attn_cuda_forward(
     for (int n = 0; n < batch/im2col_step_; ++n)
     {
         auto columns = output_n.select(0, n);
-        AT_DISPATCH_FLOATING_TYPES(value.type(), "ms_deform_attn_forward_cuda", ([&] {
+        AT_DISPATCH_FLOATING_TYPES(GET_TENSOR_TYPE(value), "ms_deform_attn_forward_cuda", ([&] {
             ms_deformable_im2col_cuda(at::cuda::getCurrentCUDAStream(),
                 value.data<scalar_t>() + n * im2col_step_ * per_value_size,
                 spatial_shapes.data<int64_t>(),
@@ -82,7 +95,7 @@ at::Tensor ms_deform_attn_cuda_forward(
 
 
 std::vector<at::Tensor> ms_deform_attn_cuda_backward(
-    const at::Tensor &value, 
+    const at::Tensor &value,
     const at::Tensor &spatial_shapes,
     const at::Tensor &level_start_index,
     const at::Tensor &sampling_loc,
@@ -98,12 +111,12 @@ std::vector<at::Tensor> ms_deform_attn_cuda_backward(
     AT_ASSERTM(attn_weight.is_contiguous(), "attn_weight tensor has to be contiguous");
     AT_ASSERTM(grad_output.is_contiguous(), "grad_output tensor has to be contiguous");
 
-    AT_ASSERTM(value.type().is_cuda(), "value must be a CUDA tensor");
-    AT_ASSERTM(spatial_shapes.type().is_cuda(), "spatial_shapes must be a CUDA tensor");
-    AT_ASSERTM(level_start_index.type().is_cuda(), "level_start_index must be a CUDA tensor");
-    AT_ASSERTM(sampling_loc.type().is_cuda(), "sampling_loc must be a CUDA tensor");
-    AT_ASSERTM(attn_weight.type().is_cuda(), "attn_weight must be a CUDA tensor");
-    AT_ASSERTM(grad_output.type().is_cuda(), "grad_output must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(value), "value must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(spatial_shapes), "spatial_shapes must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(level_start_index), "level_start_index must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(sampling_loc), "sampling_loc must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(attn_weight), "attn_weight must be a CUDA tensor");
+    AT_ASSERTM(IS_CUDA_TENSOR(grad_output), "grad_output must be a CUDA tensor");
 
     const int batch = value.size(0);
     const int spatial_size = value.size(1);
@@ -128,11 +141,11 @@ std::vector<at::Tensor> ms_deform_attn_cuda_backward(
     auto per_sample_loc_size = num_query * num_heads * num_levels * num_point * 2;
     auto per_attn_weight_size = num_query * num_heads * num_levels * num_point;
     auto grad_output_n = grad_output.view({batch/im2col_step_, batch_n, num_query, num_heads, channels});
-    
+
     for (int n = 0; n < batch/im2col_step_; ++n)
     {
         auto grad_output_g = grad_output_n.select(0, n);
-        AT_DISPATCH_FLOATING_TYPES(value.type(), "ms_deform_attn_backward_cuda", ([&] {
+        AT_DISPATCH_FLOATING_TYPES(GET_TENSOR_TYPE(value), "ms_deform_attn_backward_cuda", ([&] {
             ms_deformable_col2im_cuda(at::cuda::getCurrentCUDAStream(),
                                     grad_output_g.data<scalar_t>(),
                                     value.data<scalar_t>() + n * im2col_step_ * per_value_size,
@@ -153,4 +166,4 @@ std::vector<at::Tensor> ms_deform_attn_cuda_backward(
     };
 }
 
-} // namespace groundingdino
+} // namespace groundingdino

groundingdino/models/GroundingDINO/fuse_modules.py

@@ -8,7 +8,7 @@
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
-from timm.models.layers import DropPath
+from timm.layers import DropPath
 
 
 class FeatureResizer(nn.Module):

groundingdino/models/GroundingDINO/transformer.py

@@ -554,6 +554,7 @@ class TransformerEncoder(nn.Module):
                         memory_text,
                         key_padding_mask,
                         text_attention_mask,
+                        use_reentrant=True,
                     )
                 else:
                     output, memory_text = self.fusion_layers[layer_id](
@@ -581,6 +582,7 @@ class TransformerEncoder(nn.Module):
                     spatial_shapes,
                     level_start_index,
                     key_padding_mask,
+                    use_reentrant=True,
                 )
             else:
                 output = layer(
@@ -859,7 +861,7 @@ class DeformableTransformerDecoderLayer(nn.Module):
         return tensor if pos is None else tensor + pos
 
     def forward_ffn(self, tgt):
-        with torch.cuda.amp.autocast(enabled=False):
+        with torch.amp.autocast("cuda", enabled=False):
             tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
         tgt = tgt + self.dropout4(tgt2)
         tgt = self.norm3(tgt)

groundingdino/version.py

@@ -0,0 +1 @@
+__version__ = '0.1.0'

main.py

@@ -0,0 +1,137 @@
+from pathlib import Path
+
+import cv2
+import numpy as np
+import supervision as sv
+import torch
+import yaml
+from pdf2image import convert_from_bytes, convert_from_path
+from PIL import Image
+from tqdm import tqdm
+
+from groundingdino.util.inference import Model, preprocess_caption
+
+GROUNDING_DINO_CONFIG = "groundingdino/config/GroundingDINO_SwinT_OGC.py"
+GROUNDING_DINO_CHECKPOINT = "gdino_checkpoints/groundingdino_swint_ogc.pth"
+BOX_THRESHOLD = 0.4
+TEXT_THRESHOLD = 0.25
+
+
+class PdfConverterService:
+    """
+    A service to convert PDF files to images and resize images,
+    using pypdfium2 for PDFs and Pillow for images.
+    """
+
+    def __init__(self, dpi: int):
+        self._pdfium_initialized = False  # Track if PDFium needs explicit init/deinit
+        self._dpi = dpi
+
+    def convert_pdf_to_jpg(self, file_path: str) -> list[Image.Image]:
+        """
+        Converts a PDF file to JPG images at different scales.
+        """
+        pil_images = convert_from_path(file_path, dpi=self._dpi)
+        return pil_images
+
+    def resize_image(self, img: Image.Image, size: tuple[int, int]) -> Image.Image:
+        """
+        Resizes a PIL Image to the specified size.
+        """
+        return img.resize(size, Image.LANCZOS)
+
+    @staticmethod
+    def save_image_as_png(img: Image.Image, file_path: str):
+        """
+        Saves a PIL Image as a PNG file.
+        """
+        img.save(file_path, format="PNG")
+
+    @staticmethod
+    def to_cv2_image(img: Image.Image):
+        open_cv_image = np.array(img.convert("RGB"))
+        return open_cv_image[:, :, ::-1].copy()
+
+    def convert_pdf_bytes_to_jpg(self, pdf_bytes: bytes) -> list[Image.Image]:
+        """
+        Converts PDF bytes to JPG images at different scales.
+        """
+        pil_images = convert_from_bytes(pdf_bytes, dpi=self._dpi)
+        return pil_images
+
+
+def main(
+    data_dir: str | Path,
+    text: str = "ID card. Carte Vitale. Bank details. Human face.",
+    concept_list_yaml: str | None = None,
+    device: str = "cuda:0" if torch.cuda.is_available() else "cpu",
+):
+    output_dir = Path("outputs") / "extract"
+    output_dir.mkdir(parents=True, exist_ok=True)
+    if concept_list_yaml:
+        print(f"Overriding concepts !")
+        with open(concept_list_yaml, "r") as f:
+            concepts = yaml.load(f)
+        text = "".join([f" {x}." for x in concepts])
+
+    print(f"List of concepts to detect: {text}")
+
+    if isinstance(data_dir, str):
+        data_dir = Path(data_dir)
+
+    for img_path in tqdm(
+        data_dir.glob("*.pdf"), total=len(list(data_dir.glob("*.pdf")))
+    ):
+        pdf_convertor = PdfConverterService(120)
+        if img_path.suffix == ".pdf":
+            imgs = pdf_convertor.convert_pdf_to_jpg(str(img_path))
+            img = imgs[0]
+            pdf_convertor.save_image_as_png(img, img_path.parent / "test.png")
+            img = pdf_convertor.to_cv2_image(img)
+        else:
+            img = cv2.imread(str(img_path))
+
+        # image_source, image = load_image(str(img_path.parent / "test.png"))
+
+        grounding_model = Model(
+            model_config_path=GROUNDING_DINO_CONFIG,
+            model_checkpoint_path=GROUNDING_DINO_CHECKPOINT,
+            device=device,
+        )
+        caption = preprocess_caption(text)
+        detections, labels = grounding_model.predict_with_caption(
+            image=img,
+            caption=caption,
+            box_threshold=BOX_THRESHOLD,
+            text_threshold=TEXT_THRESHOLD,
+        )
+        confidences = detections.confidence.tolist()
+        class_names = labels
+
+        labels = [
+            f"{class_name} {confidence:.2f}"
+            for class_name, confidence in zip(class_names, confidences)
+        ]
+
+        for i, bbox in enumerate(detections.xyxy):
+            x_min, y_min, x_max, y_max = tuple(bbox)
+            patch = img[int(y_min) : int(y_max), int(x_min) : int(x_max)]
+            patch_img_path = str(
+                Path("outputs") / "extract" / f"{img_path.stem}_{i:d}.png"
+            )
+            cv2.imwrite(patch_img_path, patch)
+
+        box_annotator = sv.BoxAnnotator(color_lookup=sv.ColorLookup.INDEX)
+        annotated_frame = box_annotator.annotate(
+            scene=img.copy(), detections=detections
+        )
+        label_annotator = sv.LabelAnnotator(color_lookup=sv.ColorLookup.INDEX)
+        annotated_frame = label_annotator.annotate(
+            scene=annotated_frame, detections=detections, labels=labels
+        )
+
+        cv2.imwrite(str(Path("outputs") / f"{img_path.stem}.png"), annotated_frame)
+
+
+if __name__ == "__main__":
+    main("data")

pyproject.toml

@@ -0,0 +1,25 @@
+[project]
+name = "groundingdino"
+version = "0.1.0"
+description = "Add your description here"
+readme = "README.md"
+requires-python = ">=3.13"
+dependencies = [
+    "torch>=2.3.1",
+    "torchvision>=0.18.1",
+    "numpy>=1.24.4",
+    "tqdm>=4.66.1",
+    "hydra-core>=1.3.2",
+    "iopath>=0.1.10",
+    "pillow>=9.4.0",
+    "opencv-python-headless>=4.8.0",
+    "supervision>=0.26.1",
+    "pycocotools>=2.0.10",
+    "transformers>=4.55.1",
+    "addict>=2.4.0",
+    "yapf>=0.43.0",
+    "timm>=1.0.19",
+    "pdf2image>=1.17.0",
+    "pip>=25.2",
+    "setuptools>=80.9.0",
+]