cGen/eval_pretrain.from_file.py at main · ksenia007/cGen · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
"""Code to get embeddings from the checkpoint, using the provided files with labels as the guide
"""
import itertools
import json
import logging
import os
import uuid
from argparse import ArgumentParser
import torch.nn as nn

import h5py
import numpy as np
import pytorch_lightning as pl
import torch
import yaml

import model as model_module


# Ensure that all operations are deterministic on GPU (if used) for reproducibility
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
device = torch.device(
    "cuda:0") if torch.cuda.is_available() else torch.device("cpu")


def load_all_weights(model, checkpoint, cpu=False):
    """Function to load all the weights (strict load) for evaluation."""
    if cpu:
        checkpoint_weights = torch.load(checkpoint, map_location=torch.device('cpu'))
    else:
        checkpoint_weights = torch.load(checkpoint)
    print('checkpoint weights ', list(checkpoint_weights['state_dict'].keys()))
    for k in list(checkpoint_weights['state_dict'].keys()):
        new_key = k.replace(
            "backbone.backbone.", "backbone.")
        new_key = new_key.replace(
            "backbone.contr_layer.", "contr_layer.")
        new_key = new_key.replace(
                    "backbone.backbone.backbone.", "backbone.")
        checkpoint_weights['state_dict'][new_key] = checkpoint_weights['state_dict'].pop(
            k)
    model.load_state_dict(checkpoint_weights['state_dict'], strict=True)
    return model


def eval(**kwargs):
    """
    Evaluate on randomly sampled validation set for now.
    """
    cp_path = kwargs['checkpoint']
    if 'output_folder' in kwargs:
        output_folder = kwargs['output_folder']
        _, cp_filename = os.path.split(cp_path)
    else:
        output_folder, cp_filename = os.path.split(cp_path)

    logging.basicConfig(
        filename=output_folder + '/eval_log.log', level=logging.DEBUG)
    logging.info("Output folder will be {0}".format(output_folder))

    model_args = {
         'model_type': kwargs['model_type'],
         'model_name': kwargs['model_name'],
         'sequence_length': kwargs['seq_length'],
    }
    if 'final_mlp_size' in kwargs:
        model_args['final_mlp_size'] = kwargs['final_mlp_size']
    model = model_module.ModelContrastive(**model_args)

    if torch.cuda.is_available():
        model = load_all_weights(model, kwargs['checkpoint'])
        print(model)
        model.cuda()
    else:
        model = load_all_weights(model, kwargs['checkpoint'], cpu=True)
        print(model)
        logging.info("Running on CPU")
    model.eval()

    logging.info("Running eval...")

    data_dir = './ds_train_hg38/pretrain_dataset'
    if 'data_dir' in kwargs:
        data_dir = kwargs['data_dir']
    seq_len_file = kwargs['seq_length_file']
    seq_len = kwargs['seq_length']

    if seq_len<seq_len_file:
        # we want a shorter sequence
        tail = (seq_len_file-seq_len)//2

    sampled_h5_file = os.path.join(data_dir, 'sc_sort.sampled_N=200k.seqlen={0}.h5'.format(seq_len_file))
    enh_h5_file = os.path.join(data_dir, 'F5.hg38.enhancers.seqlen={0}.h5'.format(seq_len_file))

    print(sampled_h5_file)
    sampled_predictions = []
    sampled_predictions_preproj = []
    with h5py.File(sampled_h5_file, 'r') as fh:
        seqs = fh['sequences']
        for s in seqs:
            s = torch.Tensor(s.reshape((1, s.shape[0], s.shape[1])))
            if seq_len<seq_len_file:
                s = s[:,tail:tail+seq_len,:]
            pred1, pred2 = model(s.float().permute(0, 2, 1).to(device))
            # specify what we want to save
            sampled_predictions.append(pred1.cpu().detach().numpy())
            sampled_predictions_preproj.append(pred2.cpu().detach().numpy())
    sampled_predictions = np.vstack(sampled_predictions)
    sampled_predictions_preproj = np.vstack(sampled_predictions_preproj)
    print(sampled_predictions.shape)
    np.save(os.path.join(output_folder, '{0}.sc_sort.sampled_N=200k.npy'.format(
        cp_filename)), sampled_predictions)
    np.save(os.path.join(output_folder, '{0}.preproj.sc_sort.sampled_N=200k.npy'.format(
        cp_filename)), sampled_predictions_preproj)
    del sampled_predictions
    del sampled_predictions_preproj

    print(enh_h5_file)
    enh_predictions = []
    enh_predictions_preproj = []
    with h5py.File(enh_h5_file, 'r') as fh:
        seqs = fh['sequences']
        for s in seqs:
            s = torch.Tensor(s.reshape((1, s.shape[0], s.shape[1])))
            if seq_len<seq_len_file:
                s = s[:,tail:tail+seq_len,:]
            pred1, pred2 = model(s.float().permute(0, 2, 1).to(device))
            enh_predictions.append(pred1.cpu().detach().numpy())
            enh_predictions_preproj.append(pred2.cpu().detach().numpy())
    enh_predictions = np.vstack(enh_predictions)
    enh_predictions_preproj = np.vstack(enh_predictions_preproj)
    print(enh_predictions.shape)
    np.save(os.path.join(output_folder, '{0}.F5.hg38.enhancers.npy'.format(
        cp_filename)), enh_predictions)
    np.save(os.path.join(output_folder, '{0}.preproj.F5.hg38.enhancers.npy'.format(
        cp_filename)), enh_predictions_preproj)


if __name__ == "__main__":
    parser = ArgumentParser()
    parser.add_argument(
        "--setup", help='A required .yaml file with the training setup')
    args = parser.parse_args()

    with open(args.setup) as f:
        setup_args = yaml.safe_load(f)

    eval(**setup_args)