-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtest_graph_encode_norm_ecfp.py
165 lines (142 loc) · 6.4 KB
/
test_graph_encode_norm_ecfp.py
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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
import cPickle as pickle
import sys
import numpy as np
import torch
import torch.cuda
from rdkit import Chem
from sklearn import metrics
from sklearn.model_selection import train_test_split
from torch import nn
from torch import optim
from torch.utils.data import DataLoader
from models.normed_encoded_basic_model_ecfp import BasicModel
from models.graph_model_wrapper import GraphWrapper
from mol_graph import *
from mol_graph import GraphEncoder
from pre_process.data_loader import GraphDataSet, collate_2d_graphs, collate_2d_tensors, collate_2d_ecfp_graphs
from pre_process.load_dataset import load_classification_dataset, load_ecfp_dataset
from mpnn_functions.encoders.bond_autoencoder import BondAutoEncoder
from mpnn_functions.encoders.atom_autoencoder import AtomAutoEncoder
import tqdm
def filter_dataset(data, labels, size_cutoff):
uniq, count = np.unique(labels, return_counts=True)
mask = np.isin(labels, uniq[count > size_cutoff])
new_label_dict = dict(zip(uniq[count > size_cutoff], range(len(uniq[count > size_cutoff]))))
filtered_dataset = []
new_labels = []
for graph, cond, label in zip(data, mask, labels):
if cond:
new_label = new_label_dict[label]
new_labels.append(new_label)
graph.label = new_label
filtered_dataset.append(graph)
return filtered_dataset, new_labels, sum(count > size_cutoff)
def count_model_params(model):
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
return sum([np.prod(p.size()) for p in model_parameters])
def save_model(model, model_name, model_att, model_metrics):
# type: (nn.Module, dict) -> None
torch.save(model.state_dict(), 'basic_model' + str(model_name) + '.state_dict')
with open('basic_model_attributes.pickle', 'wb') as out_file:
pickle.dump(model_att, out_file)
with open('basic_model_' + str(model_name) + '_stats.pickle', 'wb') as out_file:
pickle.dump(model_metrics, out_file)
def test_model(model, dataset):
model.eval()
labels = []
true_labels = []
with torch.no_grad():
for batch in tqdm.tqdm(dataset):
labels = labels + model(batch).max(dim=-1)[1].cpu().data.numpy().tolist()
true_labels = true_labels + batch['labels'].cpu().data.numpy().tolist()
return (
metrics.accuracy_score(true_labels, labels),
metrics.precision_score(true_labels, labels, average='micro'),
metrics.recall_score(true_labels, labels, average='micro')
)
seed = 317
torch.manual_seed(seed)
data_file = sys.argv[1]
mgf = MolGraphFactory(Mol2DGraph.TYPE, AtomFeatures(), BondFeatures())
data = load_ecfp_dataset(data_file+'.csv', 'InChI', Chem.MolFromInchi, mgf, 'target')
# graph_encoder = GraphEncoder()
# with open('basic_model_graph_encoder.pickle', 'wb') as out:
# pickle.dump(graph_encoder, out)
# np.savez_compressed(data_file, data=data, no_labels=no_labels, all_labels=all_labels)
# data, all_labels, no_labels = filter_dataset(data, all_labels, 99)
model_attributes = {
'afm': 8,
'bfm': 2,
'mfm': 2*8,
'adj': data[0].adj.shape[-1],
'out': 4*8,
'classification_output': 16384
}
ae = AtomAutoEncoder()
# ae.load_state_dict(torch.load('./atom_autoencoder.state_dict', map_location=lambda storage, loc: storage))
be = BondAutoEncoder()
# be.load_state_dict(torch.load('./bond_autoencoder.state_dict', map_location=lambda storage, loc: storage))
model = nn.Sequential(
GraphWrapper(BasicModel(model_attributes['afm'], model_attributes['bfm'], model_attributes['mfm'],
model_attributes['adj'], model_attributes['out'], atom_encoder=ae.encoder,
bond_encoder=be.encoder)),
nn.Linear(model_attributes['out'], model_attributes['classification_output'])
)
model.float() # convert to half precision
# for layer in model.modules():
# if isinstance(layer, nn.BatchNorm1d):
# layer.float()
model.apply(BasicModel.init_weights)
ae.load_state_dict(torch.load('./atom_autoencoder.state_dict', map_location=lambda storage, loc: storage))
be.load_state_dict(torch.load('./bond_autoencoder.state_dict', map_location=lambda storage, loc: storage))
print "Model has: {} parameters".format(count_model_params(model))
if torch.cuda.is_available():
model.cuda()
criterion = nn.MSELoss()
optimizer = optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-5)
model.train()
train, test = train_test_split(data, test_size=0.1, random_state=seed)
del data
train, val = train_test_split(train, test_size=0.1, random_state=seed)
train = GraphDataSet(train)
val = GraphDataSet(val)
test = GraphDataSet(test)
train = DataLoader(train, 128, shuffle=True, collate_fn=collate_2d_ecfp_graphs)
val = DataLoader(val, 128, shuffle=False, collate_fn=collate_2d_ecfp_graphs)
test = DataLoader(test, 128, shuffle=False, collate_fn=collate_2d_ecfp_graphs)
epoch_losses = []
break_con = False
for epoch in tqdm.trange(1000):
model.train()
epoch_loss = 0
for batch in tqdm.tqdm(train):
model.zero_grad()
loss = criterion(torch.sigmoid(model(batch) * batch['mask']), batch['labels'])
epoch_loss += loss.item()
loss.backward()
optimizer.step()
epoch_losses.append(epoch_loss)
val_loss = 0
with torch.no_grad():
for batch in tqdm.tqdm(val):
model.zero_grad()
loss = criterion(torch.sigmoid(model(batch) * batch['mask']), batch['labels'])
val_loss += loss.item()
tqdm.tqdm.write("Train loss: {}, Val loss: {}".format(epoch_loss, val_loss))
# acc, pre, rec = test_model(model, train)
# f1 = 2 * (pre * rec) / (pre + rec)
# tqdm.tqdm.write(
# "epoch {} training loss: {}, acc: {}, pre: {}, rec: {}, F1: {}".format(epoch, epoch_loss, acc,
# pre, rec, f1))
# acc, pre, rec = test_model(model, val)
# f1 = 2 * (pre * rec) / (pre + rec)
# tqdm.tqdm.write(
# "epoch {} validation acc: {}, pre: {}, rec: {}, F1: {}".format(epoch, acc,
# pre, rec, f1))
# if not np.isnan(f1) and f1 > 0.8:
# save_model(model, 'epoch_'+str(epoch), model_attributes, {'acc': acc, 'pre': pre, 'rec': rec, 'f1': f1})
#
# acc, pre, rec = test_model(model, test)
# f1 = 2 * (pre * rec) / (pre + rec)
# tqdm.tqdm.write(
# "Testing acc: {}, pre: {}, rec: {}, F1: {}".format(acc, pre, rec, f1))