-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathreranking_and_explaining.py
265 lines (215 loc) · 11.8 KB
/
reranking_and_explaining.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
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
# !/usr/bin/python3
# @File: train.py
# --coding:utf-8--
# @Author:yuwang
# @Email:as1003208735@foxmail.com
# @Time: 2022.03.27.19
import os
import argparse
import torch
from model.RetroAGT import RetroAGT
from data.datasets import RerankingDataset
from torch.utils.data import DataLoader
from tqdm import tqdm
from torch.nn.functional import one_hot
from model.util import bond_fea2type
import nums_from_string
os.environ['CUDA_LAUNCH_BLOCKING'] = '1'
BOND_TYPE_INT_TO_BOND_ORDER = {0: 0, 1: 1, 2: 1.5, 3: 2, 4: 3}
@torch.no_grad()
def main():
parser = argparse.ArgumentParser()
# dataset configuration
parser.add_argument('--dataset', type=str, default="GLN_200topk_200maxk_noGT_19260817_test")
parser.add_argument('--model_path', type=str,
default='your_dir/checkpoints/epoch=374-step=29250.ckpt')
parser.add_argument('--not_fast_read', default=False, action='store_true')
parser.add_argument('--lg_path', type=str,
default='your_dir/processed/leaving_group.pt')
parser.add_argument('--num_workers', type=int, default=16)
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--cuda', type=int, default=1)
parser.add_argument('--need_action', default=False, action='store_true')
args = parser.parse_args()
dataset = RerankingDataset(root="./data/reranking", dataset=args.dataset, lg_path=args.lg_path)
dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers)
print("Building Model...")
model = build_model(args).to(args.cuda)
print("Finished Model...")
print("Predicting...")
model.eval()
model.state = 'reranking'
origin_top_k = TopKAccRecorder()
model_top_k = TopKAccRecorder()
model_correct_action = []
model_best_action = []
for batch_idx, batch in enumerate(tqdm(dataloader, desc='Epoch')):
batch = data_to_device(batch, args.cuda)
bsz = batch['origin_idx'].shape[0]
pred_state = []
# because the re-ranking task is aware of center cnt, we add rxn_cnt condition as prior knowledge
for rxn_cnt in range(10):
batch['center_cnt'] = torch.tensor([rxn_cnt] * bsz).to(args.cuda)
rc_adj_prob, h_prob, lg_prob, rc_fea = model(batch)
rc_adj_prob = rc_adj_prob.sigmoid()
h_prob = h_prob.softmax(dim=-1)
lg_prob = lg_prob.softmax(dim=-1)
pred_state.append({
'rc_adj_prob': rc_adj_prob,
'h_prob': h_prob,
'lg_prob': lg_prob,
'rc_fea': rc_fea,
})
product_bond_order_adj = bond_fea2type(batch['product']['bond_adj'] - 1)
with tqdm(total=bsz, leave=False) as pbar:
pbar.set_description('Step')
for idx_data in range(bsz):
cur_product = []
predictions = dataset.get_predictions(batch['origin_idx'][idx_data],
end_idx=50)
origin_rank = batch['origin_rank'][idx_data]
if predictions:
origin_top_k.count(origin_rank, n_pred=len(predictions))
solutions = []
n_pro = int(batch['product']['n_atom'][idx_data])
for pred, rank in tqdm(predictions, desc='Top_K_Predictions', leave=False):
pred = data_to_device(pred, args.cuda)
center_cnt = int(pred['center_cnt'])
cur_state_dict = pred_state[center_cnt]
cur_rc_adj = cur_state_dict['rc_adj_prob'][idx_data, :n_pro, :n_pro]
cur_h = cur_state_dict['h_prob'][idx_data, :n_pro]
lg_dic = pred['lg']
n_lg = int(lg_dic['n_atom'])
shared_atom_fea_lg, masked_adj_lg = model.emb(lg_dic['atom_fea'].unsqueeze(0),
lg_dic['bond_adj'].unsqueeze(0),
lg_dic['dist_adj'].unsqueeze(0),
torch.tensor([center_cnt], dtype=torch.long,
device=lg_dic['atom_fea'].device),
None,
dist3d_adj=None,
contrast=True)
shared_atom_fea_lg = model.shared_encoder(shared_atom_fea_lg, masked_adj_lg)
shared_atom_fea_lg[:, 1:] += model.gate_embedding(pred['gate_token'])
ct_fea_lg = model.ct_encoder(shared_atom_fea_lg, masked_adj_lg)[:, 1:]
# rc_fea = self.ct_encoder(shared_atom_fea, masked_adj)[:, 1:]
ct_prob_fea = model.ct_adj_fn(cur_state_dict['rc_fea'][[idx_data], 1:], ct_fea_lg, None)[:, :,
:model.max_ct_atom]
ct_prob = model.ct_out_fn(ct_prob_fea).squeeze(-1) + \
torch.where(batch['product']['atom_fea'][[idx_data], 0] > 0, 0., -1e3)[:, :, None]
# Find LG
cur_lg_idx = int(pred['lg_id'])
if cur_lg_idx < 0:
lg_fea2 = model.lg_encoder(shared_atom_fea_lg, masked_adj_lg)[:, 0]
lg_prob2 = model.lg_out_fn(lg_fea2)
cur_lg_idx = lg_prob2.argmax(dim=-1).item()
cur_lg_score = cur_state_dict['lg_prob'][idx_data, cur_lg_idx]
cur_lg_score = -cur_lg_score.log()
ct_prob = ct_prob.squeeze(0).sigmoid()[:n_pro, :n_lg]
if args.need_action:
cur_action = [f"LGM|Select Leaving Group with Index {cur_lg_idx} and Energy %.2f" % cur_lg_score]
it_score = cur_lg_score - (1 - cur_rc_adj).log().sum() \
- cur_h[:, 3].log().sum() - (1 - ct_prob).log().sum()
if args.need_action:
cur_action += [f"IT|Initial Energy:%.2f" % it_score]
# Find LGC
real_ct = pred['ct_target'][:n_pro, :n_lg].bool()
ct_score = -ct_prob[real_ct].log().sum() + (1 - ct_prob[real_ct]).log().sum()
if args.need_action:
for pro_idx, lg_idx in real_ct.nonzero().tolist():
cur_cost = -ct_prob[pro_idx, lg_idx].log() + (1 - ct_prob[pro_idx, lg_idx]).log()
cur_bond_order = pred["rea_bond_adj"][pro_idx, lg_idx]
cur_bond_order = BOND_TYPE_INT_TO_BOND_ORDER[int(cur_bond_order)]
cur_action += [f"LGC|Add Bonds: between {pro_idx} and {lg_idx + n_pro} with Bond Type "
f"{cur_bond_order} and Cost %.2f" % cur_cost]
# Find RC
real_rc = pred['rc_target'][:n_pro, :n_pro].bool()
rc_score = -cur_rc_adj[real_rc].log().sum() + (1 - cur_rc_adj[real_rc]).log().sum()
if args.need_action:
for pro_idx_1, pro_idx_2 in real_rc.nonzero().tolist():
if pro_idx_1 >= pro_idx_2:
continue
cur_cost = -cur_rc_adj[pro_idx_1, pro_idx_2].log() + (1 - cur_rc_adj[pro_idx_1, pro_idx_2]).log()
cur_cost += -cur_rc_adj[pro_idx_2, pro_idx_1].log() + (1 - cur_rc_adj[pro_idx_2, pro_idx_1]).log()
cur_bond_order = pred["rea_bond_adj"][pro_idx_1, pro_idx_2]
cur_bond_order = BOND_TYPE_INT_TO_BOND_ORDER[int(cur_bond_order)]
origin_bond_order = product_bond_order_adj[idx_data, pro_idx_1, pro_idx_2].item()
cur_action += [f"Replace Bonds: between {pro_idx_1} and {pro_idx_2}, "
f"from Bond Type {origin_bond_order} to Bond Type"
f"{cur_bond_order} with Cost %.2f" % cur_cost]
real_hc = pred['rc_h'][:n_pro]
mask_select = one_hot(real_hc, 7).bool()
hc_score = -cur_h[mask_select].log().sum() + (1 - cur_h[:, 3]).log().sum()
if args.need_action:
for pro_idx, n_hc in mask_select.nonzero().tolist():
cur_cost = -cur_h[pro_idx, n_hc].log() + (1 - cur_h[pro_idx, 3]).log()
cur_action += [f"H number change {n_hc - 3} cost of atom {pro_idx}: %.2f" % cur_cost]
total_score = it_score + ct_score + rc_score + hc_score
if not args.need_action:
cur_action = []
solutions.append((float(total_score), rank, rank == origin_rank, cur_action))
solutions.sort(key=lambda x: x[0])
model_rank = 9999
if solutions:
model_best_action.append(solutions[0])
for rank_idx in range(len(solutions)):
if solutions[rank_idx][2]:
model_rank = rank_idx
model_correct_action.append(solutions[rank_idx])
break
if model_rank == 9999:
model_correct_action.append(([], 10000, False))
model_top_k.count(model_rank, n_pred=len(predictions))
pbar.set_postfix({"origin_1_acc": origin_top_k.top_k_acc_dict[f'top_{1}_acc'],
"model_1_acc": model_top_k.top_k_acc_dict[f'top_{1}_acc'],
"origin_per": f"{origin_top_k.get_percentage_rank():.2f}",
"model_per": f"{model_top_k.get_percentage_rank():.2f}"})
pbar.update(1)
origin_top_k.normalize(len(dataset))
model_top_k.normalize(len(dataset))
print(args)
print("Origin Top K Acc:", origin_top_k.top_k_acc_dict)
print("Model Top K Acc:", model_top_k.top_k_acc_dict)
print("Origin_percentage", f"{origin_top_k.get_percentage_rank():.2f}")
print("Model_percentage", f"{model_top_k.get_percentage_rank():.2f}")
def build_model(args):
model = RetroAGT.load_from_checkpoint(
args.model_path,
lg_path=args.lg_path,
strict=False,
)
return model
class TopKAccRecorder:
def __init__(self):
self.top_k_acc_dict = {f"top_{k}_acc": 0 for k in [1, 3, 5, 10, 50]}
self.cur_k = []
def count(self, rank, n_pred=None):
if isinstance(rank, torch.Tensor):
rank = rank.item()
elif isinstance(rank, str):
rank = int(rank)
if n_pred is not None and rank < 50:
self.cur_k.append((rank+1)/n_pred)
else:
self.cur_k.append(1)
for k in self.top_k_acc_dict.keys():
if rank < nums_from_string.get_nums(k)[0]:
self.top_k_acc_dict[k] += 1
def normalize(self, total):
for k in self.top_k_acc_dict.keys():
self.top_k_acc_dict[k] /= total
def get_percentage_rank(self):
return sum(self.cur_k) / len(self.cur_k)
def prob2energy(prob):
return -prob.log()
def data_to_device(batch, device):
if isinstance(batch, torch.Tensor):
batch = batch.to(device=device)
elif isinstance(batch, tuple):
batch = tuple(data_to_device(ele, device) for ele in batch)
elif isinstance(batch, dict):
batch = {k: data_to_device(v, device) for k, v in batch.items()}
elif isinstance(batch, list):
batch = [data_to_device(ele, device) for ele in batch]
return batch
if __name__ == '__main__':
main()