SAM2 AMG example mIoU, perf numbers and more SAM2 model annotations (#1196)

Author: cpuhrsch

examples/sam2_amg_server/README.md

@@ -2,3 +2,101 @@
 ```
 curl -X POST http://127.0.0.1:5000/upload -F 'image=@/path/to/file.jpg' --output path/to/output.png
 ```
+
+## Example script to collect rles
+
+Start the server
+
+```
+python server.py ~/checkpoints/sam2 --port <your_port> --host <your_hostname> --fast
+```
+
+Collect the rles
+
+```
+xargs -I {} curl -s -w "\n" -X POST http://<your_hostname>:<your_port>/upload_rle -F 'image=@{}' < image_paths > rle_masks
+```
+
+## mIoU scores on random subset of sav validation dataset
+
+Experiments run on H100 and with batch size 1
+
+| mode | mIoU | mask count mismatch | avg. ms per request |
+| ---  |---   | ------------------ | ----------------- |
+| baseline | 1.0 | 0 | 786 |
+| ao | 1.0 | 0 | 738 |
+| fast | 0.95 | 190 | 563 |
+| furious | 0 | 1000 | 204 |
+
+mask count mismatch counts the number of requests where the number of masks differ from the baseline.
+For example, the baseline may have chosen to segment an image into 18 masks, but the fast variant produces 17 or 19.
+We exclude these examples from the mIoU calculation.
+
+### 1. Create a random subset of 1000 images
+```
+find sav_val -type f > sav_val_image_paths
+shuf -n 1000 sav_val_image_paths > sav_val_image_paths_shuf_1000
+```
+
+### 2. Use the baseline (https://github.com/facebookresearch/sam2) to generate rles
+
+Make sure you've installed https://github.com/facebookresearch/sam2
+
+Start server
+```
+python server.py ~/checkpoints/sam2 --port <your_port> --host <your_hostname> --baseline
+```
+
+Generate and save rles (one line per json via `-w "\n"`)
+```
+$ time xargs -I {} curl -s -w "\n" -X POST http://<your_hostname>:<your_port>/upload_rle -F 'image=@{}' < sav_val_image_paths_shuf_1000 > results/sav_val_masks_baseline_shuf_1000
+
+real    13m6.374s
+user    0m3.349s
+sys     0m4.137s
+```
+
+### 3. Start server with torchao variant of SAM2
+Start server
+```
+python server.py ~/checkpoints/sam2 --port <your_port> --host <your_hostname>
+```
+
+Generate and save rles (one line per json via `-w "\n"`)
+```
+$ time xargs -I {} curl -s -w "\n" -X POST http://<your_hostname>:<your_port>/upload_rle -F 'image=@{}' < sav_val_image_paths_shuf_1000 > results/sav_val_masks_shuf_1000
+
+real    12m18.916s
+user    0m3.506s
+sys     0m4.350s
+```
+
+### 4. Start server with torchao variant of SAM2 and `--fast` optimizations
+Start server
+```
+python server.py ~/checkpoints/sam2 --port <your_port> --host <your_hostname> --fast
+```
+
+Generate and save rles (one line per json via `-w "\n"`)
+```
+$ time xargs -I {} curl -s -w "\n" -X POST http://<your_hostname>:<your_port>/upload_rle -F 'image=@{}' < sav_val_image_paths_shuf_1000 > results/sav_val_masks_fast_shuf_1000
+
+real    9m23.912s
+user    0m3.271s
+sys     0m4.138s
+```
+
+### 5. Start server with torchao variant of SAM2 and `--fast` and `--furious` optimizations
+Start server
+```
+python server.py ~/checkpoints/sam2 --port <your_port> --host <your_hostname> --fast --furious
+```
+
+Generate and save rles (one line per json via `-w "\n"`)
+```
+$ time xargs -I {} curl -s -w "\n" -X POST http://<your_hostname>:<your_port>/upload_rle -F 'image=@{}' < sav_val_image_paths_shuf_1000 > results/sav_val_masks_fast_furious_shuf_1000
+
+real    3m24.383s
+user    0m3.583s
+sys     0m4.519s
+```

examples/sam2_amg_server/compare_rle_lists.py

@@ -0,0 +1,41 @@
+import fire
+import torch
+import json
+from sam2.utils.amg import rle_to_mask
+
+"""
+Script to calculate mIoU given two lists of rles from upload_rle endpoint
+of server.
+"""
+
+
+def iou(mask1, mask2):
+    assert mask1.dim() == 2
+    assert mask2.dim() == 2
+    intersection = torch.logical_and(mask1, mask2)
+    union = torch.logical_or(mask1, mask2)
+    return (intersection.sum(dim=(-1, -2)) / union.sum(dim=(-1, -2)))
+
+
+def main(path0, path1):
+    fail_count = 0
+    miou_sum = 0.0
+    miou_count = 0
+    with open(path0, 'r') as f0, open(path1, 'r') as f1:
+        for line0, line1 in zip(f0, f1):
+            masks0 = json.loads(line0)
+            masks1 = json.loads(line1)
+            if masks0.keys() != masks1.keys():
+                fail_count += 1
+                continue
+            for mask0, mask1 in zip(masks0.values(), masks1.values()):
+                mask0 = torch.from_numpy(rle_to_mask(mask0))
+                mask1 = torch.from_numpy(rle_to_mask(mask1))
+                miou_sum += iou(mask0, mask1).item()
+                miou_count += 1
+
+    print(f"fail_count: {fail_count} mIoU: {miou_sum / miou_count}")
+
+
+if __name__ == "__main__":
+    fire.Fire(main)

examples/sam2_amg_server/dog_rle.json

@@ -30,6 +30,15 @@
 
 # torch.set_float32_matmul_precision('high')
 
+
+def iou(mask1, mask2):
+    assert mask1.dim() == 2
+    assert mask2.dim() == 2
+    intersection = torch.logical_and(mask1, mask2)
+    union = torch.logical_or(mask1, mask2)
+    return (intersection.sum(dim=(-1, -2)) / union.sum(dim=(-1, -2)))
+
+
 def show_anns(anns):
     if len(anns) == 0:
         return
@@ -49,17 +58,44 @@ def show_anns(anns):
     return torch.stack(ms)
 
 
-def main(checkpoint_path, fast=False, furious=False, benchmark=False, verbose=False, points_per_batch=64):
+def profiler_runner(path, fn, *args, **kwargs):
+    with torch.profiler.profile(
+            activities=[torch.profiler.ProfilerActivity.CPU,
+                        torch.profiler.ProfilerActivity.CUDA],
+            record_shapes=True) as prof:
+        result = fn(*args, **kwargs)
+    prof.export_chrome_trace(path)
+    return result
+
+
+def main(checkpoint_path,
+         baseline=False,
+         fast=False,
+         furious=False,
+         unittest=False,
+         benchmark=False,
+         profile=None,
+         verbose=False,
+         points_per_batch=64,
+         port=5000,
+         host="127.0.0.1",
+         dry=False):
     if verbose:
-        logging.basicConfig(level=logging.INFO)
+        logging.basicConfig(level=logging.INFO,
+                            format='%(asctime)s - %(levelname)s - %(message)s',
+                            datefmt='%Y-%m-%d %H:%M:%S')
     logging.info(f"Running with fast set to {fast} and furious set to {furious}")
+    logging.info(f"Running with port {port} and host {host}")
 
-    if fast:
-        from torchao._models.sam2.build_sam import build_sam2
-        from torchao._models.sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
-    else:
+    if baseline:
+        logging.info(f"Importing sam2 from outside of torchao. If this errors, install https://github.com/facebookresearch/sam2")
         from sam2.build_sam import build_sam2
         from sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
+        from sam2.utils.amg import rle_to_mask
+    else:
+        from torchao._models.sam2.build_sam import build_sam2
+        from torchao._models.sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
+        from torchao._models.sam2.utils.amg import rle_to_mask
 
     device = "cuda"
     from pathlib import Path
@@ -70,7 +106,7 @@ def main(checkpoint_path, fast=False, furious=False, benchmark=False, verbose=Fa
     sam2 = build_sam2(model_cfg, sam2_checkpoint, device=device, apply_postprocessing=False)
 
     logging.info(f"Using {points_per_batch} points_per_batch")
-    mask_generator = SAM2AutomaticMaskGenerator(sam2, points_per_batch=points_per_batch)
+    mask_generator = SAM2AutomaticMaskGenerator(sam2, points_per_batch=points_per_batch, output_mode="uncompressed_rle")
 
     if furious:
         torch.set_float32_matmul_precision('high')
@@ -107,6 +143,37 @@ def main(checkpoint_path, fast=False, furious=False, benchmark=False, verbose=Fa
         logging.info(f"Running one iteration to compile.")
         masks = mask_generator.generate(example_image)
         logging.info(f"First iteration took {time.time() - t}s.")
+        if unittest:
+            logging.info(f"Running strict comparison to reference mask")
+            import json
+            ref_masks = json.loads(open("dog_rle.json").read())
+            ret_data = {}
+            for mask_id in range(len(masks)):
+                ret_data[f"mask_{mask_id}"] = masks[mask_id]["segmentation"]
+            v0_areas = []
+            v1_areas = []
+            miou_sum = 0.0
+            miou_count = 0
+            for k0 in ref_masks:
+                assert k0 in ret_data, f"Expected {k0} to be in return data"
+                from torchao._models.sam2.utils.amg import area_from_rle
+                v0_area = area_from_rle(ref_masks[k0])
+                v1_area = area_from_rle(ret_data[k0])
+                v0_areas.append(v0_area)
+                v1_areas.append(v1_area)
+                if v0_area != v1_area:
+                    print(f"v0 area {v0_area} doesn't match v1 area {v1_area}")
+                v0_mask = torch.from_numpy(rle_to_mask(ref_masks[k0]))
+                v1_mask = torch.from_numpy(rle_to_mask(ret_data[k0]))
+                if not torch.allclose(v0_mask, v1_mask):
+                    miou_sum += iou(v0_mask, v1_mask)
+                    miou_count += 1
+                    print(f"Masks don't match for key {k0}. IoU is {iou(v0_mask, v1_mask)}")
+            if miou_count == 0:
+                print("Masks exactly match reference.")
+            else:
+                print(f"mIoU is {miou_sum / miou_count}")
+
         if benchmark:
             logging.info(f"Running 3 warumup iterations.")
             for _ in range(3):
@@ -121,7 +188,13 @@ def main(checkpoint_path, fast=False, furious=False, benchmark=False, verbose=Fa
             max_memory_allocated_percentage = int(100 * (max_memory_allocated_bytes / total_memory))
             max_memory_allocated_bytes = max_memory_allocated_bytes >> 20
             print(f"max_memory_allocated_bytes: {max_memory_allocated_bytes}MiB or {max_memory_allocated_percentage}%")
-            return
+
+        if profile is not None:
+            print(f"Saving profile under {profile}")
+            profiler_runner(profile, mask_generator.generate, example_image)
+
+    if dry:
+        return
 
     app = FastAPI()
 
@@ -133,6 +206,25 @@ def main(checkpoint_path, fast=False, furious=False, benchmark=False, verbose=Fa
         allow_methods=["*"],
         allow_headers=["*"],
     )
+
+    @app.post("/upload_rle")
+    async def upload_rle(image: UploadFile = File(...)):
+        # Save the uploaded image to a temporary location
+        temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=f"_{image.filename}")
+        with open(temp_file.name, "wb") as b:
+            shutil.copyfileobj(image.file, b)
+
+        # Read the image back into memory to send as response
+        example_image = cv2.imread(temp_file.name)
+        example_image = cv2.cvtColor(example_image, cv2.COLOR_BGR2RGB)
+        t = time.time()
+        with torch.backends.cuda.sdp_kernel(enable_cudnn=True):
+            masks = mask_generator.generate(example_image)
+        print(f"Took {time.time() - t} to generate a mask for input image.")
+        ret_data = {}
+        for mask_id in range(len(masks)):
+            ret_data[f"mask_{mask_id}"] = masks[mask_id]["segmentation"]
+        return ret_data
 
     @app.post("/upload")
     async def upload_image(image: UploadFile = File(...)):
@@ -143,13 +235,16 @@ async def upload_image(image: UploadFile = File(...)):
 
         # Read the image back into memory to send as response
         example_image = cv2.imread(temp_file.name)
+        example_image = cv2.cvtColor(example_image, cv2.COLOR_BGR2RGB)
         t = time.time()
         with torch.backends.cuda.sdp_kernel(enable_cudnn=True):
             masks = mask_generator.generate(example_image)
         print(f"Took {time.time() - t} to generate a mask for input image.")
         # Save an example
         plt.figure(figsize=(example_image.shape[1]/100., example_image.shape[0]/100.), dpi=100)
         plt.imshow(example_image)
+        for i in range(len(masks)):
+            masks[i]["segmentation"] = rle_to_mask(masks[i]["segmentation"])
         show_anns(masks)
         plt.axis('off')
         plt.tight_layout()
@@ -163,7 +258,7 @@ async def upload_image(image: UploadFile = File(...)):
         return StreamingResponse(BytesIO(image_data), media_type="image/png")
 
 
-    uvicorn.run(app, host="127.0.0.1", port=5000, log_level="info")
+    uvicorn.run(app, host=host, port=port, log_level="info")
 
 if __name__ == "__main__":
     fire.Fire(main)