log train set stats; log quant error mean; increase val bits and dims;

app.py

@@ -15,8 +15,8 @@ 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")
 
-HYPERDIMS = 1024
-VALUE_BITS = 8
+HYPERDIMS = 4096
+VALUE_BITS = 10
 POS_BITS = 9 # CLIP features are 512 dims
 val_bins = np.linspace(start=-1., stop=1., num=2**VALUE_BITS)
 print(val_bins.shape, val_bins.min(), val_bins.max(), 'val bins')
@@ -89,7 +89,7 @@ def quantize_embeds(embeds):
     quantized_embeds = np.reshape(quantized_embeds_flat, embeds.shape)
     closest_bin = np.reshape(closest_bin, embeds.shape)
     print(closest_bin.shape, 'values are in bins', closest_bin.min(), 'to', closest_bin.max())
-    print('abs quant error', np.abs(embeds - quantized_embeds).sum())
+    print('abs quant error avg', np.abs(embeds - quantized_embeds).mean())
     return quantized_embeds, closest_bin
 
 def update_exemplars(df, rng, exemplars, lut):
@@ -106,6 +106,7 @@ def update_exemplars(df, rng, exemplars, lut):
     embeds_train = np.concatenate([embeds[labels_zero_train_idx], embeds[labels_one_train_idx]], axis=0)
     labels_train = np.concatenate([labels[labels_zero_train_idx], labels[labels_one_train_idx]], axis=0)
     print('Training set ', embeds_train.shape, labels_train.shape)
+    print(np.sum(labels_train == 0), " zeros and ", np.sum(labels_train == 1).sum(), " ones")
     labels_zero_test_idx = np.setdiff1d(labels_zero_idx, labels_zero_train_idx)
     labels_one_test_idx = np.setdiff1d(labels_one_idx, labels_one_train_idx)
     embeds_test = np.concatenate([embeds[labels_zero_test_idx], embeds[labels_one_test_idx]], axis=0)