feat: new space for LWDETR

app.py

@@ -1,109 +1,71 @@
-
 import torch
-from transformers import pipeline
-
-from PIL import Image
-
-import matplotlib.pyplot as plt
-import matplotlib.patches as patches
-
-from random import choice
-import io
-
-detector50 = pipeline(model="facebook/detr-resnet-50")
-
-detector101 = pipeline(model="facebook/detr-resnet-101")
-
-
+from transformers import AutoImageProcessor, LwDetrForObjectDetection
+import supervision as sv
 import gradio as gr
+import spaces
 
-COLORS = ["#ff7f7f", "#ff7fbf", "#ff7fff", "#bf7fff",
-            "#7f7fff", "#7fbfff", "#7fffff", "#7fffbf",
-            "#7fff7f", "#bfff7f", "#ffff7f", "#ffbf7f"]
-
-fdic = {
-    "family" : "Impact",
-    "style" : "italic",
-    "size" : 15,
-    "color" : "yellow",
-    "weight" : "bold"
-}
-
-
-def get_figure(in_pil_img, in_results):
-    plt.figure(figsize=(16, 10))
-    plt.imshow(in_pil_img)
-    #pyplot.gcf()
-    ax = plt.gca()
-
-    for prediction in in_results:
-        selected_color = choice(COLORS)
-
-        x, y = prediction['box']['xmin'], prediction['box']['ymin'],
-        w, h = prediction['box']['xmax'] - prediction['box']['xmin'], prediction['box']['ymax'] - prediction['box']['ymin']
-
-        ax.add_patch(plt.Rectangle((x, y), w, h, fill=False, color=selected_color, linewidth=3))
-        ax.text(x, y, f"{prediction['label']}: {round(prediction['score']*100, 1)}%", fontdict=fdic)
-
-    plt.axis("off")
 
-    return plt.gcf()
+@spaces.GPU
+def infer(model_name, image):
+    label_annotator = sv.LabelAnnotator(text_padding=4, smart_position=True)
+    box_annotator = sv.BoxAnnotator()
 
+    model_name = f"AnnaZhang/{model_name}"
+    processor = AutoImageProcessor.from_pretrained(model_name)
+    model = LwDetrForObjectDetection.from_pretrained(model_name)
+    inputs = processor(images=image, return_tensors="pt")
+    outputs = model(**inputs)
 
-def infer(model, in_pil_img):
+    # convert outputs (bounding boxes and class logits) to COCO API
+    # let's only keep detections with score > 0.7
+    target_sizes = torch.tensor([image.size[::-1]])
+    results = processor.post_process_object_detection(
+        outputs, target_sizes=target_sizes, threshold=0.7
+    )[0]
+    detections = sv.Detections.from_transformers(
+        transformers_results=results, id2label=model.config.id2label
+    )
+    image = box_annotator.annotate(image, detections)
+    image = label_annotator.annotate(image, detections)
+    return image
 
-    results = None
-    if model == "detr-resnet-101":
-        results = detector101(in_pil_img)
-    else:
-        results = detector50(in_pil_img)
 
-    figure = get_figure(in_pil_img, results)
-
-    buf = io.BytesIO()
-    figure.savefig(buf, bbox_inches='tight')
-    buf.seek(0)
-    output_pil_img = Image.open(buf)
-
-    return output_pil_img
-
-
-with gr.Blocks(title="DETR Object Detection - ClassCat",
-                    css=".gradio-container {background:lightyellow;}"
-               ) as demo:
-    #sample_index = gr.State([])
-
-    gr.HTML("""<div style="font-family:'Times New Roman', 'Serif'; font-size:16pt; font-weight:bold; text-align:center; color:royalblue;">DETR Object Detection</div>""")
-
-    gr.HTML("""<h4 style="color:navy;">1. Select a model.</h4>""")
-
-    model = gr.Radio(["detr-resnet-50", "detr-resnet-101"], value="detr-resnet-50", label="Model name")
-
-    gr.HTML("""<br/>""")
-    gr.HTML("""<h4 style="color:navy;">2-a. Select an example by clicking a thumbnail below.</h4>""")
-    gr.HTML("""<h4 style="color:navy;">2-b. Or upload an image by clicking on the canvas.</h4>""")
+with gr.Blocks() as demo:
+    gr.Markdown("# LWDETR Object Detection")
 
     with gr.Row():
-        input_image = gr.Image(label="Input image", type="pil")
-        output_image = gr.Image(label="Output image with predicted instances", type="pil")
-
-    gr.Examples(['samples/cats.jpg', 'samples/detectron2.png', 'samples/cat.jpg', 'samples/hotdog.jpg'], inputs=input_image)
-
-    gr.HTML("""<br/>""")
-    gr.HTML("""<h4 style="color:navy;">3. Then, click "Infer" button to predict object instances. It will take about 10 seconds (on cpu)</h4>""")
-
-    send_btn = gr.Button("Infer")
+        with gr.Column():
+            model = gr.Radio(
+                [
+                    "lwdetr_tiny_30e_objects365",
+                    "lwdetr_small_30e_objects365",
+                    "lwdetr_medium_30e_objects365",
+                    "lwdetr_large_30e_objects365",
+                    "lwdetr_xlarge_30e_objects365",
+                    "lwdetr_tiny_60e_coco",
+                    "lwdetr_small_60e_coco",
+                    "lwdetr_medium_60e_coco",
+                    "lwdetr_large_60e_coco",
+                    "lwdetr_xlarge_60e_coco",
+                ],
+                value="lwdetr_xlarge_60e_coco",
+                label="Model",
+            )
+            input_image = gr.Image(label="Input Image", type="pil")
+            send_btn = gr.Button("Infer", variant="primary")
+        with gr.Column():
+            output_image = gr.Image(label="Output Image", type="pil")
+
+    gr.Examples(
+        examples=[
+            "samples/cats.jpg",
+            "samples/detectron2.png",
+            "samples/cat.jpg",
+            "samples/hotdog.jpg",
+        ],
+        inputs=input_image,
+    )
     send_btn.click(fn=infer, inputs=[model, input_image], outputs=[output_image])
 
-    gr.HTML("""<br/>""")
-    gr.HTML("""<h4 style="color:navy;">Reference</h4>""")
-    gr.HTML("""<ul>""")
-    gr.HTML("""<li><a href="https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/detr_attention.ipynb" target="_blank">Hands-on tutorial for DETR</a>""")
-    gr.HTML("""</ul>""")
 
-
-#demo.queue()
 demo.launch(debug=True)
-
-
-### EOF ###

requirements.txt

@@ -1,2 +1,4 @@
 torch
-transformers[timm]
+transformers[timm] @ git+https://github.com/huggingface/transformers.git
+supervision
+spaces