new load fn to initialize HD LUT and exemplars

app.py

@@ -8,12 +8,16 @@ import pandas as pd
 from glob import glob
 import random
 from datetime import datetime
+import numpy as np
+from numpy.random import MT19937
+from numpy.random import RandomState, SeedSequence
 
 clip_model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32") 
 clip_processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
 
-# pipe = pipeline("zero-shot-image-classification", model="openai/clip-vit-base-patch32")
-# images="dog.jpg"
+HYPERDIMS = 1024
+VALUE_BITS = 8
+POS_BITS = 9 # CLIP features are 512 dims
 
 def extract_features(image):
     PIL_image = Image.fromarray(np.uint8(image)).convert('RGB')
@@ -45,6 +49,29 @@ def update_table_up(img, img_name, df, state):
 def update_table_down(img, img_name, df, state):
     return update_table(img, img_name, df, state, 0)
 
+def make_LUT(nvalues, dims, rs):
+    lut = np.zeros(shape=(nvalues+1, dims))
+    lut[0, :] = rs.binomial(n=1, p=0.5, size=(dims))
+    for row in range(1, nvalues):
+        lut[row, :] = lut[row-1, :]
+        # flip few randomly
+        rand_idx = rs.choice(dims, size=dims//nvalues, replace=False)
+        lut[row, rand_idx] = 1 - lut[row, rand_idx]
+        assert np.abs(lut[row, :] - lut[row-1, :]).sum() ==dims//nvalues 
+    unique_rows = np.unique(lut, axis=0)
+    assert len(unique_rows) == len(lut)
+    return lut
+
+def load_fn(images, rng_state, exemplars_state, lut_state):
+    rs = RandomState(MT19937(SeedSequence(123456789)))
+    rng_state[0] = rs
+    exemplars_state[0] = rs.binomial(n=1, p=0.5, size=HYPERDIMS)
+    exemplars_state[1] = rs.binomial(n=1, p=0.5, size=HYPERDIMS)
+    lut_state[0] = make_LUT(2**VALUE_BITS, HYPERDIMS, rs)
+    lut_state[1] = make_LUT(2**POS_BITS, HYPERDIMS, rs)
+    return images[-1], images[-1], rng_state, exemplars_state, lut_state
+
+
     
 with gr.Blocks(title="End-User Personalization") as demo:
     img_list = glob('images/**/*.jpg')
@@ -63,5 +90,11 @@ with gr.Blocks(title="End-User Personalization") as demo:
     upvote.click(update_table_up, inputs=[image_display, image_fname, annotated_samples, images], outputs=[image_display, image_fname, annotated_samples, images])
     downvote.click(update_table_down, inputs=[image_display, image_fname, annotated_samples, images], outputs=[image_display, image_fname, annotated_samples, images])
 
-    demo.load(lambda: (images.value[-1], images.value[-1]), inputs=[], outputs=[image_display, image_fname])
+    rng = gr.State([None])
+    exemplars_state = gr.State([None, None])
+    lut_state = gr.State([None, None])
+
+    demo.load(load_fn, inputs=[images, rng, exemplars_state, lut_state], outputs=[image_display, image_fname, rng, exemplars_state, lut_state])
+    
+    personalize = gr.Button("Personalize")
 demo.launch(show_error=True, debug=True)