calc_fid.py

import numpy as np
from bio_embeddings.project import tsne_reduce
import matplotlib.pyplot as plt
from bio_embeddings.embed import ProtTransBertBFDEmbedder, ESM1bEmbedder
import pandas as pd
import csv
import os
import numpy
from numpy import cov
from numpy import trace
from numpy import iscomplexobj
from scipy.linalg import sqrtm
from evodiff.utils import parse_txt
from evodiff.plot import plot_embedding
import umap

# Need to run PGP first on generated seqs , this performs downstream analysis
# https://github.com/hefeda/PGP

def calculate_fid(act1, act2):
    """calculate frechet inception distance"""
    # calculate mean and covariance statistics
    mu1, sigma1 = act1.mean(axis=0), cov(act1, rowvar=False)
    mu2, sigma2 = act2.mean(axis=0), cov(act2, rowvar=False)
    # calculate sum squared difference between means
    ssdiff = numpy.sum((mu1 - mu2)**2.0)
    # calculate sqrt of product between cov
    covmean = sqrtm(sigma1.dot(sigma2))
    # check and correct imaginary numbers from sqrt
    if iscomplexobj(covmean):
        covmean = covmean.real
    # calculate score
    fid = ssdiff + trace(sigma1 + sigma2 - 2.0 * covmean)
    return fid

project_run='large' #'large' or 'small'

# Calculate FID between test dataset sample and generated seqs
if project_run=='large':
    project_dir = '../PGP/PGP_OUT_LARGE/'
    runs = ['blosum-new', 'uniform-new', 'oaardm-backup', 'soardm', 'carp', 'ref', 'valid', 'esm-1b', 'esm2',
            'rfdiff', 'foldingdiff-new']
    c = ['#D0D0D0', "#b0e16d", '#63C2B5', '#46A7CB', '#1B479D', 'plum', 'firebrick', 'grey', 'mediumpurple',
         '#89194B', '#F8961D', 'darkgoldenrod', 'darkslateblue', 'darkgoldenrod', 'firebrick']
elif project_run=='small':
    project_dir = '../PGP/PGP_OUT/'
    runs = ['blosum-new', 'uniform-new', 'oaardm', 'soardm', 'carp', 'ref', 'valid']
    c = ['#D0D0D0', "#b0e16d", '#63C2B5', '#46A7CB', '#1B479D', 'plum', 'firebrick', 'grey']

test_fasta = project_dir + 'test3/seqs.txt'
test_sequences = parse_txt(test_fasta)
len_test = len(test_sequences)
sequences = test_sequences
colors = ['#D0D0D0'] * len(sequences)
for i, run in enumerate(runs):
    gen_file = project_dir + run + '/seqs.txt'
    gen_df = pd.read_csv(gen_file, names=['sequences'])
    gen_sequences = list(gen_df.sequences)
    [sequences.append(s) for s in gen_sequences]
    [colors.append(c[i]) for s in gen_sequences]
runs = ['test'] + runs
runs

# Fit UMAP to train embeddings, then fit to each model
embedder = ProtTransBertBFDEmbedder()
embeddings = embedder.embed_many([s for s in sequences])
embeddings = list(embeddings)
reduced_embeddings = [ProtTransBertBFDEmbedder.reduce_per_protein(e) for e in embeddings]
projection = umap.UMAP(n_components=2, n_neighbors=25, random_state=42).fit(reduced_embeddings[:len_test])
train_proj_emb = projection.transform(reduced_embeddings[:len_test])
# Plot and save to file
for i in range(len(runs)-1):
    begin = len_test + (1000 * (i))
    end = len_test + (1000 * (i + 1))
    print(begin, end)
    run_proj_emb = projection.transform(reduced_embeddings[begin:end])
    print(len(run_proj_emb))
    print(run_proj_emb.shape)
    plot_embedding(train_proj_emb, run_proj_emb, c, i, runs, project_run)

# Calculate FID
reduced_embeddings = np.array(reduced_embeddings)
test_embeddings = reduced_embeddings[:len_test]
reduced_embeddings_by_model = reduced_embeddings[len_test:].reshape(len(runs)-1,-1,1024)  # 7 runs x 300 sample x 1024 params
print("test shape", test_embeddings.shape)
print("rest shape", reduced_embeddings_by_model.shape)
fids = []
for i in range(len(reduced_embeddings_by_model)): # compare all to test
    curr_fid = calculate_fid(test_embeddings, reduced_embeddings_by_model[i])
    fids.append(curr_fid)
    print(f'{runs[i+1]} to test, {curr_fid : 0.2f}')