openocr-python 0.1.0.dev4__tar.gz → 0.1.2__tar.gz

{openocr_python-0.1.0.dev4/openocr_python.egg-info → openocr_python-0.1.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: openocr-python
-Version: 0.1.0.dev4
+Version: 0.1.2
 Summary: Accurate and Efficient General OCR System
 Home-page: https://github.com/Topdu/OpenOCR
 Author: OCR Team, FVL Lab
@@ -75,7 +75,7 @@ Visit: https://github.com/Topdu/OpenOCR
 
 ```bash
 # Install from PyPI (recommended)
-pip install openocr-python==0.1.0.dev4
+pip install openocr-python==0.1.2
 
 # Or install from source
 git clone https://github.com/Topdu/OpenOCR.git

{openocr_python-0.1.0.dev4 → openocr_python-0.1.2}/QUICKSTART.md

@@ -18,7 +18,7 @@ Visit: https://github.com/Topdu/OpenOCR
 
 ```bash
 # Install from PyPI (recommended)
-pip install openocr-python==0.1.0.dev4
+pip install openocr-python==0.1.2
 
 # Or install from source
 git clone https://github.com/Topdu/OpenOCR.git

{openocr_python-0.1.0.dev4 → openocr_python-0.1.2}/openocr/__init__.py

@@ -13,7 +13,7 @@ sys.path.insert(0, os.path.abspath(os.path.join(__dir__, '..')))
 # from .tools.infer_doc_onnx import OpenDocONNX
 from .openocr import OpenOCR, main
 
-__version__ = '0.1.0.dev4'
+__version__ = '0.1.1'
 
 # Lazy import for demo interfaces to avoid initialization on import
 def launch_openocr_demo(*args, **kwargs):

{openocr_python-0.1.0.dev4 → openocr_python-0.1.2}/openocr/demo_gradio.py

@@ -127,13 +127,88 @@ def launch_demo(share=False, server_port=7860, server_name='0.0.0.0'):
     Returns:
         gr.Blocks: Gradio demo instance
     """
-    css = '.image-container img { width: 100%; max-height: 320px;}'
-
-    with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
+    custom_css = """
+body, .gradio-container {
+    font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "Noto Sans", Helvetica, Arial, sans-serif;
+}
+.app-header {
+    text-align: center;
+    max-width: 1200px;
+    margin: 20px auto !important;
+    padding: 20px;
+}
+.app-header h1 {
+    font-size: 2.5em;
+    font-weight: 700;
+    margin-bottom: 10px;
+}
+.app-header p {
+    font-size: 1.1em;
+    opacity: 0.7;
+    line-height: 1.6;
+}
+.quick-links {
+    text-align: center;
+    padding: 12px 0;
+    border: 1px solid var(--border-color-primary);
+    border-radius: 12px;
+    margin: 16px auto;
+    max-width: 1200px;
+    background: var(--background-fill-secondary);
+}
+.quick-links a {
+    margin: 0 16px;
+    font-size: 15px;
+    font-weight: 600;
+    color: var(--link-text-color);
+    text-decoration: none;
+    transition: all 0.3s ease;
+}
+.quick-links a:hover {
+    opacity: 0.8;
+    text-decoration: underline;
+}
+.image-container img {
+    max-width: 100%;
+    max-height: 480px;
+    width: auto;
+    height: auto;
+    object-fit: contain;
+    display: block;
+    margin: 0 auto;
+}
+.upload-section {
+    border: 2px dashed var(--border-color-primary);
+    border-radius: 12px;
+    padding: 20px;
+    background: var(--background-fill-secondary);
+    transition: all 0.3s ease;
+}
+.upload-section:hover {
+    border-color: var(--color-accent);
+    background: var(--background-fill-primary);
+}
+.gradio-button-primary {
+    font-weight: 600 !important;
+    transition: all 0.3s ease !important;
+}
+.gradio-button-primary:hover {
+    transform: translateY(-2px);
+    box-shadow: var(--shadow-drop-lg) !important;
+}
+"""
+
+    with gr.Blocks(css=custom_css, theme=gr.themes.Soft()) as demo:
         gr.HTML("""
-                <h1 style='text-align: center;'><a href="https://github.com/Topdu/OpenOCR">OpenOCR</a></h1>
-                <p style='text-align: center;'>准确高效的通用 OCR 系统 （由<a href="https://fvl.fudan.edu.cn">FVL实验室</a> <a href="https://github.com/Topdu/OpenOCR">OCR Team</a> 创建） <a href="https://github.com/Topdu/OpenOCR/tree/main?tab=readme-ov-file#quick-start">[本地快速部署]</a></p>"""
-                )
+        <div class="app-header">
+            <h1><a href="https://github.com/Topdu/OpenOCR">OpenOCR</a></h1>
+            <p>Accurate and Efficient General OCR System (built by <a href="https://fvl.fudan.edu.cn">FVL Lab</a> <a href="https://github.com/Topdu/OpenOCR">OCR Team</a>)</p>
+        </div>
+        <div class="quick-links">
+            <a href="https://github.com/Topdu/OpenOCR" target="_blank">📖 GitHub</a>
+            <a href="https://github.com/Topdu/OpenOCR/tree/main?tab=readme-ov-file#quick-start" target="_blank">🚀 Quick Start</a>
+        </div>
+        """)
         with gr.Row():
             with gr.Column(scale=1):
                 input_image = gr.Image(label='Input image',
@@ -142,21 +217,21 @@ def launch_demo(share=False, server_port=7860, server_name='0.0.0.0'):
                 examples = gr.Examples(examples=e2e_img_example,
                                        inputs=input_image,
                                        label='Examples')
-                downstream = gr.Button('🚀 运行识别', variant='primary')
+                downstream = gr.Button('🚀 Run Recognition', variant='primary')
 
-                # 添加参数调节组件
+                # Parameter adjustment components
                 with gr.Column():
                     with gr.Row():
                         det_input_size_textbox = gr.Number(
                             label='Detection Input Size',
                             value=960,
-                            info='检测网络输入尺寸的最长边，默认为960。')
+                        info='Max side length of detection input, default 960.')
                         det_score_mode_dropdown = gr.Dropdown(
                             ['slow', 'fast'],
                             value='slow',
                             label='Detection Score Mode',
-                            info=
-                            '文本框的置信度计算模式，默认为 slow。slow 模式计算速度较慢，但准确度较高。fast 模式计算速度较快，但准确度较低。'
+                        info=
+                            'Confidence score mode for text boxes, default slow. Slow mode is more accurate but slower. Fast mode is faster but less accurate.'
                         )
                     with gr.Row():
                         rec_drop_score_slider = gr.Slider(
@@ -165,14 +240,14 @@ def launch_demo(share=False, server_port=7860, server_name='0.0.0.0'):
                             value=0.4,
                             step=0.01,
                             label='Recognition Drop Score',
-                            info='识别置信度阈值，默认值为0.4。低于该阈值的识别结果和对应的文本框被丢弃。')
+                        info='Recognition confidence threshold, default 0.4. Results below this threshold will be discarded.')
                         mask_thresh_slider = gr.Slider(
                             0.0,
                             1.0,
                             value=0.3,
                             step=0.01,
                             label='Mask Threshold',
-                            info='Mask 阈值，用于二值化 mask，默认值为0.3。如果存在文本截断时，请调低该值。')
+                        info='Mask threshold for binarization, default 0.3. Lower this value if text is truncated.')
                     with gr.Row():
                         box_thresh_slider = gr.Slider(
                             0.0,
@@ -180,21 +255,20 @@ def launch_demo(share=False, server_port=7860, server_name='0.0.0.0'):
                             value=0.6,
                             step=0.01,
                             label='Box Threshold',
-                            info='文本框置信度阈值，默认值为0.6。如果存在文本被漏检时，请调低该值。')
+                        info='Text box confidence threshold, default 0.6. Lower this value if text boxes are missed.')
                         unclip_ratio_slider = gr.Slider(
                             1.5,
                             2.0,
                             value=1.5,
                             step=0.05,
                             label='Unclip Ratio',
-                            info='文本框解析时的膨胀系数，默认值为1.5。值越大文本框越大。')
+                        info='Expansion ratio for text box parsing, default 1.5. Larger values produce larger text boxes.')
 
-                    # 模型选择组件
                     model_type_dropdown = gr.Dropdown(
                         ['mobile', 'server'],
                         value='mobile',
                         label='Model Type',
-                        info='选择 OCR 模型类型：高效率模型mobile，高精度模型server。')
+                        info='Select OCR model type: mobile for efficiency, server for accuracy.')
 
             with gr.Column(scale=1):
                 img_mask = gr.Image(label='mask',

{openocr_python-0.1.0.dev4 → openocr_python-0.1.2}/openocr/demo_opendoc.py

@@ -8,11 +8,9 @@ import gradio as gr
 from PIL import Image
 from pathlib import Path
 
-import sys
-sys.path.append(os.path.dirname(os.path.abspath(__file__)))
-from .tools.infer_doc_onnx import OpenDocONNX
-from .tools.utils.logging import get_logger
-from .tools.download_example_images import get_example_images_path
+from tools.infer_doc_onnx import OpenDocONNX
+from tools.utils.logging import get_logger
+from tools.download_example_images import get_example_images_path
 
 logger = get_logger(name='opendoc_gradio')
 
@@ -28,18 +26,18 @@ def get_pipeline(
     use_gpu=None,
     auto_download=True
 ) -> OpenDocONNX:
-    """获取或初始化OpenDocONNX流水线
+    """Get or initialize OpenDocONNX pipeline.
 
     Args:
-        layout_model_path: 版面检测ONNX模型路径
-        unirec_encoder_path: UniRec编码器ONNX模型路径
-        unirec_decoder_path: UniRec解码器ONNX模型路径
-        tokenizer_mapping_path: Tokenizer映射文件路径
+        layout_model_path: Path to layout detection ONNX model.
+        unirec_encoder_path: Path to UniRec encoder ONNX model.
+        unirec_decoder_path: Path to UniRec decoder ONNX model.
+        tokenizer_mapping_path: Path to tokenizer mapping file.
         use_gpu: Whether to use GPU. If None, auto-detect.
-        auto_download: If True, automatically download missing models
+        auto_download: If True, automatically download missing models.
 
     Returns:
-        OpenDocONNX: 初始化好的OpenDocONNX实例
+        OpenDocONNX: Initialized OpenDocONNX instance.
     """
     global pipeline
     if pipeline is None:
@@ -63,13 +61,13 @@ current_pipeline = None
 def process_image(
     image_path: str | None
 ) -> tuple[Image.Image | None, str, str, str | None, str, str]:
-    """处理图片并进行OCR识别
+    """Process image and perform OCR recognition.
 
     Args:
-        image_path: 图片文件路径，None表示无图片
+        image_path: Image file path, None if no image.
 
     Returns:
-        tuple: (可视化图片, Markdown内容(base64图片), JSON内容, ZIP文件路径, 原始Markdown, Markdown内容(base64图片))
+        tuple: (visualization image, Markdown content with base64 images, JSON content, ZIP file path, raw Markdown)
     """
     global current_pipeline
 
@@ -271,10 +269,18 @@ body, .gradio-container {
     background: var(--background-fill-primary);
 }
 #vis_output {
-    min-height: 400px;
     border-radius: 12px;
     overflow: hidden;
 }
+.image-container img, #vis_output img {
+    max-width: 100%;
+    max-height: 480px;
+    width: auto;
+    height: auto;
+    object-fit: contain;
+    display: block;
+    margin: 0 auto;
+}
 #md_preview {
     max-height: 600px;
     min-height: 200px;
@@ -349,10 +355,10 @@ LATEX_DELIMS = [
 
 # Define the Gradio Interface
 def create_demo() -> gr.Blocks:
-    """创建Gradio演示界面
+    """Create Gradio demo interface.
 
     Returns:
-        gr.Blocks: Gradio Blocks应用实例
+        gr.Blocks: Gradio Blocks application instance.
     """
     # Get example images path and download if necessary
     example_img_dir = get_example_images_path(demo_type='doc')
@@ -388,7 +394,7 @@ def create_demo() -> gr.Blocks:
             with gr.Column(scale=4, elem_classes=['upload-section']):
                 input_img = gr.Image(type='filepath',
                                      label='📤 Upload Document Image',
-                                     height=400)
+                                     elem_classes=['image-container'])
 
                 # Add examples if available
                 if example_images:

{openocr_python-0.1.0.dev4 → openocr_python-0.1.2}/openocr/demo_unirec.py

@@ -12,7 +12,7 @@ from tools.infer_unirec_onnx import UniRecONNX, clean_special_tokens
 from tools.download_example_images import get_example_images_path
 from tools.to_markdown import MarkdownConverter
 
-# 创建全局 markdown_converter 实例
+# Create global markdown_converter instance
 markdown_converter = MarkdownConverter()
 
 # LaTeX delimiters for formula rendering
@@ -81,7 +81,7 @@ def stream_recognize_image(input_image, max_length=2048):
         max_length: Maximum generation length
     """
     if input_image is None:
-        yield '请先上传一张图片。', '**请先上传一张图片。**'
+        yield 'Please upload an image first.', '**Please upload an image first.**'
         return
 
     # Convert to PIL Image if needed
@@ -138,7 +138,7 @@ def stream_recognize_image(input_image, max_length=2048):
         current_text = model.tokenizer.decode(generated_ids[-1:], skip_special_tokens=False)
         cleaned_text += clean_special_tokens(current_text)
 
-        yield cleaned_text, '_正在识别中，请稍候..._'
+        yield cleaned_text, '_Recognizing, please wait..._'
 
         # Check for EOS
         if next_token_id == eos_token_id:
@@ -149,7 +149,7 @@ def stream_recognize_image(input_image, max_length=2048):
         formatted_result = format_markdown_output(cleaned_text)
         yield formatted_result, formatted_result
     else:
-        yield '识别失败，请重试。', '**识别失败，请重试。**'
+        yield 'Recognition failed, please try again.', '**Recognition failed, please try again.**'
 
 
 def format_markdown_output(markdown_text):
@@ -161,7 +161,7 @@ def format_markdown_output(markdown_text):
     - Basic markdown formatting
     """
     if not markdown_text:
-        return '_等待识别结果..._'
+        return '_Waiting for recognition results..._'
     if '<table>' in markdown_text:
         markdown_text = markdown_converter._handle_table(markdown_text)
     if '\\(' in markdown_text or '\\[' in markdown_text:
@@ -192,28 +192,106 @@ def create_demo() -> gr.Blocks:
                 example_images.append(os.path.join(example_img_dir, file))
         example_images = sorted(example_images)
 
-    with gr.Blocks(theme=gr.themes.Soft()) as demo:
+    custom_css = """
+body, .gradio-container {
+    font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", "Noto Sans", Helvetica, Arial, sans-serif;
+}
+.app-header {
+    text-align: center;
+    max-width: 1200px;
+    margin: 20px auto !important;
+    padding: 20px;
+}
+.app-header h1 {
+    font-size: 2.5em;
+    font-weight: 700;
+    margin-bottom: 10px;
+}
+.app-header p {
+    font-size: 1.1em;
+    opacity: 0.7;
+    line-height: 1.6;
+}
+.quick-links {
+    text-align: center;
+    padding: 12px 0;
+    border: 1px solid var(--border-color-primary);
+    border-radius: 12px;
+    margin: 16px auto;
+    max-width: 1200px;
+    background: var(--background-fill-secondary);
+}
+.quick-links a {
+    margin: 0 16px;
+    font-size: 15px;
+    font-weight: 600;
+    color: var(--link-text-color);
+    text-decoration: none;
+    transition: all 0.3s ease;
+}
+.quick-links a:hover {
+    opacity: 0.8;
+    text-decoration: underline;
+}
+.upload-section {
+    border: 2px dashed var(--border-color-primary);
+    border-radius: 12px;
+    padding: 20px;
+    background: var(--background-fill-secondary);
+    transition: all 0.3s ease;
+}
+.upload-section:hover {
+    border-color: var(--color-accent);
+    background: var(--background-fill-primary);
+}
+.image-container img {
+    max-width: 100%;
+    max-height: 480px;
+    width: auto;
+    height: auto;
+    object-fit: contain;
+    display: block;
+    margin: 0 auto;
+}
+.gradio-button-primary {
+    font-weight: 600 !important;
+    transition: all 0.3s ease !important;
+}
+.gradio-button-primary:hover {
+    transform: translateY(-2px);
+    box-shadow: var(--shadow-drop-lg) !important;
+}
+"""
+
+    with gr.Blocks(css=custom_css, theme=gr.themes.Soft()) as demo:
         gr.HTML("""
-                <h1 style='text-align: center;'><a href="https://github.com/Topdu/OpenOCR">UniRec-0.1B: Unified Text and Formula Recognition with 0.1B Parameters</a></h1>
-                <p style='text-align: center;'>0.1B超轻量模型统一文本与公式识别（由<a href="https://fvl.fudan.edu.cn">FVL实验室</a> <a href="https://github.com/Topdu/OpenOCR">OCR Team</a> 创建）</p>
-                <p style='text-align: center;'><a href="https://github.com/Topdu/OpenOCR/blob/main/docs/unirec.md">[本地GPU部署]</a>获取快速识别体验</p>"""
-                )
-        gr.Markdown('上传一张图片，点击"运行识别"按钮进行文本和公式识别。')
+        <div class="app-header">
+            <h1><a href="https://github.com/Topdu/OpenOCR">UniRec-0.1B</a></h1>
+            <p>Unified Text and Formula Recognition with 0.1B Parameters (built by <a href="https://fvl.fudan.edu.cn">FVL Lab</a> <a href="https://github.com/Topdu/OpenOCR">OCR Team</a>)</p>
+        </div>
+        <div class="quick-links">
+            <a href="https://github.com/Topdu/OpenOCR" target="_blank">📖 GitHub</a>
+            <a href="https://arxiv.org/pdf/2512.21095" target="_blank">📄 Paper</a>
+            <a href="https://huggingface.co/topdu/unirec-0.1b" target="_blank">🤗 Model</a>
+            <a href="https://github.com/Topdu/OpenOCR/blob/main/docs/unirec.md" target="_blank">🚀 Local GPU Deployment</a>
+        </div>
+        """)
+        gr.Markdown('Upload an image and click the "Run Recognition" button to recognize text and formulas.')
         with gr.Row():
-            with gr.Column(scale=4):  # 左侧竖排：图片 + 按钮
-                image_input = gr.Image(label='上传图片 or 粘贴截图', type='pil')
+            with gr.Column(scale=4, elem_classes=['upload-section']):  # Left column: image + buttons
+                image_input = gr.Image(label='Upload Image or Paste Screenshot', type='pil', elem_classes=['image-container'])
 
                 # Add examples if available
                 if example_images:
                     gr.Examples(
                         examples=example_images,
                         inputs=image_input,
-                        label='📚 示例图片'
+                    label='📚 Example Images'
                     )
 
                 with gr.Row():
-                    run_button = gr.Button('🚀 运行识别', variant='primary')
-                    clear_button = gr.Button('🗑️ 清空', variant='secondary')
+                    run_button = gr.Button('🚀 Run Recognition', variant='primary')
+                    clear_button = gr.Button('🗑️ Clear', variant='secondary')
 
             with gr.Column(scale=6):
                 with gr.Tabs():
@@ -223,20 +301,20 @@ def create_demo() -> gr.Blocks:
                                                 lines=20)
                     with gr.Tab('📝 Markdown Preview'):
                         markdown_render = gr.Markdown(
-                                value='_渲染后的表格/公式将显示在这里..._',
+                                value='_Rendered tables/formulas will be displayed here..._',
                                 latex_delimiters=LATEX_DELIMS,
                                 elem_id='md_preview')
 
-        # 点击运行按钮后触发
+        # Trigger recognition on button click
         run_button.click(
             stream_recognize_image,
             inputs=[image_input],
             outputs=[markdown_output, markdown_render]
         )
 
-        # 清空按钮功能：清空图片和输出结果
+        # Clear button: reset image and output
         def clear_all():
-            return None, '', '_渲染后的表格/公式将显示在这里..._'
+            return None, '', '_Rendered tables/formulas will be displayed here..._'
 
         clear_button.click(
             clear_all,

{openocr_python-0.1.0.dev4 → openocr_python-0.1.2}/openocr/tools/infer_det.py

@@ -35,11 +35,12 @@ DOWNLOAD_URL_DET_ONNX = 'https://github.com/Topdu/OpenOCR/releases/download/deve
 
 def check_and_download_model(model_name: str, url: str):
     """
-    检查预训练模型是否存在，若不存在则从指定 URL 下载到固定缓存目录。
+    检查预训练模型是否存在，若不存在则从 ModelScope 或 HuggingFace 下载到固定缓存目录。
+    优先从 ModelScope 下载，失败后回退到 HuggingFace。
 
     Args:
-        model_name (str): 模型文件的名称，例如 "model.pt"
-        url (str): 模型文件的下载地址
+        model_name (str): 模型文件的名称，例如 "openocr_det_repvit_ch.pth"
+        url (str): 模型文件的下载地址（用于兼容，实际使用 ModelScope 和 HuggingFace）
 
     Returns:
         str: 模型文件的完整路径
@@ -56,31 +57,60 @@ def check_and_download_model(model_name: str, url: str):
         logger.info(f'Model already exists at: {model_path}')
         return str(model_path)
 
-    # 如果文件不存在，下载模型
-    logger.info(f'Model not found. Downloading from {url}...')
-
     # 创建缓存目录（如果不存在）
     cache_dir.mkdir(parents=True, exist_ok=True)
 
+    # 定义下载源
+    modelscope_repo = 'topdktu/OpenOCR'
+    huggingface_repo = 'topdu/OpenOCR'
+
+    # 尝试从 ModelScope 下载
+    logger.info(f'Model not found. Attempting to download from ModelScope...')
     try:
-        # 下载文件
-        import urllib.request
-        with urllib.request.urlopen(url) as response, open(model_path,
-                                                           'wb') as out_file:
-            out_file.write(response.read())
-        logger.info(f'Model downloaded and saved at: {model_path}')
+        from modelscope.hub.file_download import model_file_download
+        import shutil
+        downloaded_path = model_file_download(
+            model_id=modelscope_repo,
+            file_path=model_name,
+            cache_dir=str(cache_dir.parent)
+        )
+        # 复制文件到目标位置（处理符号链接）
+        if downloaded_path != str(model_path):
+            # 使用 copy2 而不是 move，确保复制实际文件内容
+            shutil.copy2(downloaded_path, str(model_path))
+        logger.info(f'Model downloaded from ModelScope and saved at: {model_path}')
         return str(model_path)
+    except Exception as e:
+        logger.warning(f'Failed to download from ModelScope: {e}')
+        logger.info('Falling back to HuggingFace...')
 
+    # 回退到 HuggingFace
+    try:
+        from huggingface_hub import hf_hub_download
+        import shutil
+        downloaded_path = hf_hub_download(
+            repo_id=huggingface_repo,
+            filename=model_name,
+            cache_dir=str(cache_dir.parent)
+        )
+        # 复制文件到目标位置（HuggingFace 返回的可能是符号链接）
+        if downloaded_path != str(model_path):
+            # 使用 copy2 而不是 move，确保复制实际文件内容而不是符号链接
+            shutil.copy2(downloaded_path, str(model_path))
+        logger.info(f'Model downloaded from HuggingFace and saved at: {model_path}')
+        return str(model_path)
     except Exception as e:
-        logger.error(f'Error downloading the model: {e}')
+        logger.error(f'Failed to download from HuggingFace: {e}')
         # 提示用户手动下载
         logger.error(
             f'Unable to download the model automatically. '
-            f'Please download the model manually from the following URL:\n{url}\n'
+            f'Please download the model manually from:\n'
+            f'  ModelScope: https://modelscope.cn/models/{modelscope_repo}\n'
+            f'  HuggingFace: https://huggingface.co/{huggingface_repo}\n'
             f'and save it to: {model_name} or {model_path}')
         raise RuntimeError(
-            f'Failed to download the model. Please download it manually from {url} '
-            f'and save it to {model_path}') from e
+            f'Failed to download the model from both ModelScope and HuggingFace. '
+            f'Please download it manually and save it to {model_path}') from e
 
 
 def replace_batchnorm(net):

{openocr_python-0.1.0.dev4 → openocr_python-0.1.2}/openocr/tools/infer_e2e.py

@@ -98,6 +98,15 @@ class OpenOCRE2E(object):
             无返回值。
 
         """
+        # Auto-switch backend for server mode
+        if mode == 'server' and backend != 'torch':
+            logger.warning(
+                f"Server mode only supports 'torch' backend, got '{backend}'. "
+                f"Automatically switching to 'torch' backend. "
+                f"Please make sure 'torch' and 'torchvision' are installed: "
+                f"pip install torch torchvision")
+            backend = 'torch'
+
         # Parse use_gpu parameter
         if use_gpu == 'auto':
             try: