float64->float16 converter, pkl committed

app.ini

@@ -0,0 +1,3 @@
+[DEFAULT]
+pkl = sample_fbi_s1e1.f16.pkl
+url = https://static.renyi.hu/ai-shared/daniel/sameenergy/

app.py

@@ -1,17 +1,54 @@
 import sys
+import argparse
+import configparser
+
+import pickle
 import gradio as gr
 import numpy as np
-import pickle
-import clip
 import torch
+import clip
+
+
+CONFIG_PATH = "app.ini"
+
 
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--pkl', type=str, help='input pickle produced by create_embedding.py')
+    parser.add_argument('--url', type=str, help='the base URL for the images')
+    args = parser.parse_args()
+    return args
 
-pickle_filename, base_url = sys.argv[1:]
-# base_url = "https://static.renyi.hu/ai-shared/daniel/sameenergy/"
+
+def parse_config_file():
+    config = configparser.ConfigParser()
+    config.read(CONFIG_PATH)
+    config_args = argparse.Namespace(**config['DEFAULT'])
+    return config_args
+
+
+if len(sys.argv) == 1:
+    print(f"no command line arguments, using {CONFIG_PATH}")
+    args = parse_config_file()
+else:
+    print("using command line arguments, ignoring ini file")
+    args = parse_args()
+
+assert "pkl" in args and args.pkl is not None
+assert "url" in args and args.url is not None
+assert args.url.endswith("/")
+
+print("arguments:", args)
+
+
+pickle_filename, base_url = args.pkl, args.url
 
 
 data = pickle.load(open(pickle_filename, "rb"))
-embeddings = data["embeddings"]
+# the data might be float16 so that the pkl is small,
+# but we use float32 in-memory to avoid numerical issues.
+# tbh i'm not sure there are any such issues.
+embeddings = data["embeddings"].astype(np.float32)
 image_features = torch.Tensor(embeddings)
 image_features /= image_features.norm(dim=-1, keepdim=True)
 
@@ -70,10 +107,8 @@ with gr.Blocks(css="footer {visibility: hidden}") as demo:
     gallery = gr.Gallery(label="Images", show_label=False, elem_id="gallery"
         ).style(columns=4, container=False)
 
-    # demo.load(empty_gallery, None, [gallery, state])
-
     with gr.Row():
-        selected = gr.Number(placeholder=0, show_label=False)
+        selected = gr.Number(0, show_label=False)
         image_query_button = gr.Button("Show similar")
 
     text_query_button.click(image_retrieval_from_text, [text], [gallery, state])
@@ -86,4 +121,4 @@ with gr.Blocks(css="footer {visibility: hidden}") as demo:
 
 
 if __name__ == "__main__":
-    demo.launch(height=3000)
+    demo.launch()

convert.py

@@ -0,0 +1,14 @@
+import sys
+import pickle
+import numpy as np
+
+filename, = sys.argv[1:]
+data = pickle.load(open(filename, "rb"))
+embeddings = data["embeddings"]
+filenames = data["filenames"]
+embeddings = embeddings.astype(np.float16)
+data["embeddings"] = embeddings
+
+output_filename = filename[:-4] + ".f16.pkl"
+with open(output_filename, "wb") as f:
+    pickle.dump(data, f)

readme.sh

@@ -10,8 +10,15 @@ scp -q -r -P 2820 sample_fbi_s1e1 hexagon.renyi.hu:./ai-shared/daniel/sameenergy
 # https://static.renyi.hu/ai-shared/daniel/sameenergy/sample_fbi_s1e1/x_BRIDGE_ADRIATIC/Dobogoko_Esztergom/Videk_ut_Dobogoko_Esztergom_014.jpg
 
 # run CLIP:
-cat sample_fbi_s1e1.txt | python create_embeddings.py sample_fbi_s1e1.pkl no-thumbs
+time cat sample_fbi_s1e1.txt | python create_embeddings.py sample_fbi_s1e1.pkl no-thumbs
 # -> sample_fbi_s1e1.pkl contains embeddings and filenames.
+# some 12 images/sec on CPU.
 
 # gradio app:
-python app.py sample_fbi_s1e1.pkl https://static.renyi.hu/ai-shared/daniel/sameenergy/
+python app.py --pkl sample_fbi_s1e1.pkl --url https://static.renyi.hu/ai-shared/daniel/sameenergy/
+# or
+python app.py
+# ...and then it takes these from app.ini
+
+python convert.py sample_fbi_s1e1.pkl
+# -> creates sample_fbi_s1e1.f16.pkl dumbed from float64 to float16.

sample_fbi_s1e1.f16.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ea091f8e382bc09ede33174285d7e1906046549b7c5291860463e001650175f
+size 6570022