allow gpu

CLIP/clip/__init__.py

@@ -148,7 +148,7 @@ class ClipWrapper:
         text_labels,
         horizontal_flipping=False,
         positive_attn_only: bool = False,
-        tile_batch_size=16,
+        tile_batch_size=32,
         prompt_batch_size=32,
         tile_interpolate_batch_size=16,
         **kwargs

app.py

@@ -43,12 +43,14 @@ def generate_relevancy(
     img = np.asarray(Image.fromarray(img).resize((244 * 4, 244 * 4)))
     assert img.dtype == np.uint8
     h, w, c = img.shape
+    start = time()
     grads = ClipWrapper.get_clip_saliency(
         img=img,
         text_labels=np.array(labels),
         prompts=prompts,
         **saliency_configs[saliency_config](h),
     )[0]
+    print("inference took", float(time() - start))
     if subtract_mean:
         grads -= grads.mean(axis=0)
     grads = grads.cpu().numpy()
@@ -78,7 +80,7 @@ def generate_relevancy(
 
 iface = gr.Interface(
     title="Semantic Abstraction Multi-scale Relevancy Extractor",
-    description="""A CPU-only demo of [Semantic Abstraction](https://semantic-abstraction.cs.columbia.edu/)'s Multi-Scale Relevancy Extractor. To run GPU inference locally, use the [official codebase release](https://github.com/columbia-ai-robotics/semantic-abstraction).
+    description="""A demo of [Semantic Abstraction](https://semantic-abstraction.cs.columbia.edu/)'s Multi-Scale Relevancy Extractor. To run GPU inference locally, use the [official codebase release](https://github.com/columbia-ai-robotics/semantic-abstraction).
 
 This relevancy extractor builds heavily on [Chefer et al.'s codebase](https://github.com/hila-chefer/Transformer-MM-Explainability) and [CLIP on Wheels' codebase](https://cow.cs.columbia.edu/).""",
     fn=generate_relevancy,

requirements.txt

@@ -1,8 +1,9 @@
 ftfy
 matplotlib
+--extra-index-url https://download.pytorch.org/whl/cu113
 torch
-tqdm 
 torchvision
+tqdm 
 regex
 numpy 
 Pillow