tokenspace / compare-allids-embeds.py
ppbrown's picture
tweaked graph title
537a599
raw
history blame
2.1 kB
#!/bin/env python
"""
Load in two pre-calculated embeddings files.
(eg: *.allid.*)
Typically, I generate files that cover the full range of tokenids,
0-49405
(using generate-allid-embeddings.py(XL).py )
This then goes through the full range and calculate distances
between each.
Display a graph of the distances.
Also print out things like the mean distance
"""
import sys
import torch
from safetensors import safe_open
file1=sys.argv[1]
file2=sys.argv[2]
device=torch.device("cuda")
print(f"reading {file1} embeddings now",file=sys.stderr)
model = safe_open(file1,framework="pt",device="cuda")
embs1=model.get_tensor("embeddings")
embs1.to(device)
print("Shape of loaded embeds =",embs1.shape)
print(f"reading {file2} embeddings now",file=sys.stderr)
model = safe_open(file2,framework="pt",device="cuda")
embs2=model.get_tensor("embeddings")
embs2.to(device)
print("Shape of loaded embeds =",embs2.shape)
if torch.equal(embs1 , embs2):
print("HEY! Both files are identical!")
exit(0)
print(f"calculating distances...")
# This calculates a full cross matrix of ALL distances to ALL other points
# in other tensor
##targetdistances = torch.cdist( embs1,embs2, p=2)
targetdistances = torch.norm(embs2 - embs1, dim=1)
#print(targetdistances.shape)
#tl=targetdistances.tolist()
#print(tl[:10])
print("sum of all distances=",torch.sum(targetdistances))
embs1_avg=torch.mean(embs1,dim=0)
embs2_avg=torch.mean(embs2,dim=0)
avg_dist= torch.cdist( embs1_avg.unsqueeze(0),embs2_avg.unsqueeze(0), p=2)
print("However, the distance between the avg-point of each is:",avg_dist)
print("Mean of all the distances:" + str(torch.mean(targetdistances,dim=0)))
######################################
import PyQt5
import matplotlib
matplotlib.use('QT5Agg') # Set the backend to QT5Agg
import matplotlib.pyplot as plt
junk, ax = plt.subplots()
graph1=targetdistances.tolist()
ax.set_title(f"{file1} vs \n{file2}")
ax.plot(graph1, label="Distance between same tokenID")
ax.set_ylabel("Distance")
ax.set_xlabel("CLIP TokenID")
ax.legend()
plt.show()