Add DeepSeek OCR notebook

[morteza89]

No description provided.

[Copilot]

🟡 Changes recommended

One or more issues must be addressed before approval.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Adds an isolated DeepSeek OCR notebook workflow for OpenVINO conversion, inference, quantization, and an interactive Gradio demo.

Changes:

• Adds model selection, download, conversion, and inference notebook cells.
• Implements OpenVINO model conversion and runtime helpers.
• Adds encoder implementations and isolated Python environment configuration.

File summaries

File	Description
openvino_notebooks/deepseek-ocr/README.md	Updated as part of this pull request.
openvino_notebooks/deepseek-ocr/pyproject.toml	Updated as part of this pull request.
openvino_notebooks/deepseek-ocr/ov_deepseek_ocr_helper.py	Updated as part of this pull request.
openvino_notebooks/deepseek-ocr/gradio_helper.py	Updated as part of this pull request.
openvino_notebooks/deepseek-ocr/deepseek-ocr.ipynb	Updated as part of this pull request.
openvino_notebooks/deepseek-ocr/deepencoderv2.py	Updated as part of this pull request.
openvino_notebooks/deepseek-ocr/deepencoder.py	Updated as part of this pull request.
openvino_notebooks/deepseek-ocr/.python-version	Updated as part of this pull request.
openvino_notebooks/deepseek-ocr/.gitignore	Updated as part of this pull request.

Review details

Suppressed comments (11)

openvino_notebooks/deepseek-ocr/deepseek-ocr.ipynb:431

• The notebook text misspells the OpenVINO API as core.complie_model; readers following this description cannot find that method. Correct it to core.compile_model.

    "DeepSeek-OCR is PyTorch model. OpenVINO supports PyTorch models via conversion to OpenVINO Intermediate Representation (IR). [OpenVINO model conversion API](https://docs.openvino.ai/2024/openvino-workflow/model-preparation.html#convert-a-model-with-python-convert-model) should be used for these purposes. `ov.convert_model` function accepts original PyTorch model instance and example input for tracing and returns `ov.Model` representing this model in OpenVINO framework. Converted model can be used for saving on disk using `ov.save_model` function or directly loading on device using `core.complie_model`. \n",

openvino_notebooks/deepseek-ocr/gradio_helper.py:240

• For an empty PDF, len(doc) - 1 is -1, so this clamp still selects page 0 and load_page(0) raises. Handle zero-page PDFs before clamping/loading and avoid exposing a 1-based selector for a document with no pages.

            page_idx = max(0, min(int(page_num) - 1, len(doc) - 1))
            page = doc.load_page(page_idx)

openvino_notebooks/deepseek-ocr/gradio_helper.py:78

• Model-generated <|det|> contents are parsed without validation; a malformed coordinate payload makes ast.literal_eval raise and aborts the whole Gradio request. Parse each reference defensively and skip invalid boxes, as the notebook helper does.

            coords = ast.literal_eval(ref[2])

openvino_notebooks/deepseek-ocr/gradio_helper.py:24

• These presets are not all valid for OVDeepseekOCR2ForCausalLM: its extract_image_features accepts only 144 or 256 queries, while 512/640/1280 inputs produce 64/100/400 queries and raise ValueError. Since this demo is used with either model variant, make the dropdown model-specific or restrict it to dimensions supported by the loaded model.

    "Gundam": {"base_size": 1024, "image_size": 640, "crop_mode": True},
    "Tiny": {"base_size": 512, "image_size": 512, "crop_mode": False},
    "Small": {"base_size": 640, "image_size": 640, "crop_mode": False},
    "Base": {"base_size": 1024, "image_size": 1024, "crop_mode": False},
    "Large": {"base_size": 1280, "image_size": 1280, "crop_mode": False},

openvino_notebooks/deepseek-ocr/ov_deepseek_ocr2_helper.py:1460

• dynamic_preprocess is called without image_size, so it always creates 768×768 local tiles, while the token sequence below uses the caller's image_size (the default is 640). For a high-resolution image this produces 144 local feature vectors but only 100 image-token slots, causing masked_scatter_ to fail. Keep preprocessing and token counting tied to the same supported tile size and validate that size.

                        images_crop_raw, crop_ratio = dynamic_preprocess(image)

openvino_notebooks/deepseek-ocr/ov_deepseek_ocr2_helper.py:335

• The no-normalization branch calls nn.Identity(), but this module never imports nn; constructing BasicImageTransform(..., normalize=False) therefore raises NameError. Use None for the no-op branch (the following condition already omits it) or import torch.nn as nn.

        normalize = normalize_transform(mean, std) if normalize else nn.Identity()

openvino_notebooks/deepseek-ocr/ov_deepseek_ocr2_helper.py:1377

• infer defaults output_path to an empty string and then calls os.makedirs(output_path, exist_ok=True), which raises FileNotFoundError when callers rely on the default. Use output_path="." as the default or guard the directory creation.

        output_path="",

openvino_notebooks/deepseek-ocr/ov_deepseek_ocr2_helper.py:1430

• The elif prompt branch creates a conversation without images, so load_pil_images returns an empty list; the next images[0] access then always raises IndexError for the text-only call path. Reject missing image_file before this branch or implement a real text-only path.

        image_draw = images[0].copy()

openvino_notebooks/deepseek-ocr/ov_deepseek_ocr_helper.py:735

• nncf.compress_weights returns the compressed ov.Model, but its return value is discarded here before saving. As a result, the requested vision INT8 compression is not included in the saved IR; the language-model path below correctly assigns the return value.

            nncf.compress_weights(ov_model, **quantization_config["vision"])

openvino_notebooks/deepseek-ocr/ov_deepseek_ocr_helper.py:1269

• infer defaults output_path to an empty string and then calls os.makedirs(output_path, exist_ok=True), which raises FileNotFoundError when callers rely on the default. Use output_path="." as the default or guard the directory creation.

        output_path="",

openvino_notebooks/deepseek-ocr/ov_deepseek_ocr_helper.py:1322

• The elif prompt branch creates a conversation without images, so load_pil_images returns an empty list; the next images[0] access then always raises IndexError for the text-only call path. Reject missing image_file before this branch or implement a real text-only path.

        image_draw = images[0].copy()

• Files reviewed: 10/11 changed files
• Comments generated: 7
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.