-
Notifications
You must be signed in to change notification settings - Fork 0
/
XAI.py
426 lines (303 loc) · 13.6 KB
/
XAI.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
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
import numpy as np
import torch
from torch.utils.data import TensorDataset, random_split, DataLoader
from chytorch.utils.data import MoleculeDataset, collate_molecules, chained_collate
from Model import GT
import chython
import os
import copy
from tqdm import tqdm
from scipy.optimize import minimize
from chytorch.utils.data import MoleculeDataset
from chytorch.utils.data import collate_molecules, chained_collate, SMILESDataset
from IPython.display import clear_output
from torch.autograd.functional import jacobian
import random
import pytorch_lightning as pl
import torch.nn as nn
def set_global_seed(seed):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
pl.seed_everything(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
return seed
def l1_inf_norm(X):
l1_norm = np.linalg.norm(X, ord=1)
l_inf_norm = np.linalg.norm(X, ord=np.inf)
l1_inf_norm = np.sqrt(l1_norm * l_inf_norm)
return l1_inf_norm
def objective_function(v, x):
w = np.outer(np.ones(len(x)), v)
return np.linalg.norm(x-w)
def get_minimal(x):
x = torch.tensor(x).detach()
initial_guess = x.mean(dim = 0).numpy()
result = minimize(objective_function, initial_guess, args = x.numpy(), tol = 1e-5, method = 'BFGS')
minimized_value = result.x
return l1_inf_norm(x.numpy()-np.array(minimized_value))
def get_iterator(molecules, is_prepared_as_packed_chython = False, return_unpacked_molecules = False):
if not is_prepared_as_packed_chython:
molecules_prepared = []
molecules_unpacked = []
for mol in molecules:
try:
s = chython.smiles(mol).pack()
molecules_prepared.append(s)
molecules_unpacked.append(chython.MoleculeContainer.unpack(s))
except:
print(mol, "IS NOT VALID")
else:
molecules_prepared = molecules
data = TensorDataset(MoleculeDataset(molecules_prepared, unpack = True))
dlts = DataLoader(data, collate_fn=chained_collate(collate_molecules), shuffle=False, batch_size=1)
if return_unpacked_molecules:
return dlts, molecules_unpacked
else:
return dlts
def get_dtl_from_smiles(smiles):
mol = MoleculeDataset(SMILESDataset(smiles), unpack=False)
data = TensorDataset(mol)
dlts = DataLoader(data, collate_fn=chained_collate(collate_molecules), shuffle=False, batch_size=1)
return dlts
def get_Jacobian_ij(model, batch, i, j, n = None, device = 'cuda'):
batch[0] = batch[0].to(device)
model.to(device)
for name, param in model.named_parameters:
param.requires_grad_(True)
model.eval()
atoms, neighbors, distances = batch[0]
encoder = model.encoder
xo = encoder.atoms_encoder(atoms) + encoder.neighbors_encoder(neighbors)
xo.requires_grad(True)
d_mask = encoder.distance_encoders[0](distances).permute(0, 3, 1, 2).flatten(end_dim=1)
d_mask.requires_grad(True)
if n is None:
n = len(encoder.layers)
layer = encoder.layers[0]
x, a = layer(xo, d_mask)
for j in range(1,n):
layer = encoder.layers[j]
x, a = layer(x, d_mask)
gradients = [[torch.autograd.grad(outputs=x[0,i,k], inputs=xo[0,j,h], grad_outputs = 1.0, retain_graph=True, create_graph=True).item() for k in range(x.size(-1))] for h in range(xo.size(-1))]
return gradients
def get_rank_residuals(model, batch, n = None, device = 'cuda'):
model.eval()
if n is None:
n = len( model.encoder.layers)
ranks = []
for i in range(1, n):
result, a_, x, xo, L = get_nth_layer(model, batch, x_ = True, n = i, device = device, double_cast = False)
ranks.append(get_minimal(x[0].detach().cpu().numpy())/l1_inf_norm(x[0].detach().cpu().numpy()))
return ranks
def get_metrics(eigvecs_r, eigvecs_l, eigvals_r, t = 0.9):
# Laplacian eigenvectors are N-dimensional with N nodes in the graph, but Rollout comes from an Attention matrix that considers
# the CLS token as well, so the first degree of freedom of each eigenvector is cut out and the vector is renormalized
# also for the same reason we have N laplacian eigenvectors and eigenvalues,
# so we discard the last eigenvector and eigenvalue from Rollout (it is the least important one for the considered ordering)
a = np.real(eigvecs_l[1:,:])/np.linalg.norm(np.real(eigvecs_l[1:,:]),axis = 0)
b = np.real(eigvecs_r[1:,:])/np.linalg.norm(np.real(eigvecs_r[1:,:]), axis = 0)
# we compute the outer product, or overlap matrix C_ij = |<a_i|l_j>| and we do not consider the row and column relative to the trivial eigenvectors
outer = np.abs(np.einsum('ij,jk->ik', np.conj(b.T), (a)))[1:,1:]
# get the max in each column, namely max_i{C_ij}.
diag = np.max(outer, axis = 0)
# turn this into a mask of 0 and 1 based on the threshold value
which_diag = copy.deepcopy(diag)
# get which laplacian modes overlap well
which_laplacian = np.argwhere(which_diag>t).reshape(-1)
# turn the threshold into a (0,1) mask for the eigenvalues
which_diag[which_diag<t] = 0.0
which_diag[which_diag!=0.0] = 1.
# get the eigenvalues and compute \zeta = \eta*number_of_laplacians
eigvals = np.absolute(eigvals_r[1:])
fraction = (which_diag*eigvals).sum()/(eigvals.sum())
quantity = fraction*(which_diag).sum()
number_of_laplacians = np.sum(which_diag)
# in case something goes wrong
if np.isnan(quantity):
print('something went wrong')
quantity = 0
return quantity, fraction, number_of_laplacians, which_laplacian
def transfer_weights(source_path, target_model, freeze=False, device='cuda'):
source_state_dict = torch.load(source_path, map_location=device)['state_dict']
target_state_dict = target_model.state_dict()
transferred_keys = []
for k, v in source_state_dict.items():
if k in target_state_dict:
target_state_dict[k] = v
transferred_keys.append(k)
if k.replace('encoder.', 'encoder.embedding.', 1) in target_state_dict:
target_state_dict[k.replace('encoder.', 'encoder.embedding', 1)] = v
transferred_keys.append(k.replace('encoder.', 'encoder.embedding', 1))
else:
if k.startswith('net.0.') and k.replace('net.0.', 'encoder.', 1) in target_state_dict:
new_key = k.replace('net.0.', 'encoder.', 1)
if v.size() == target_state_dict[new_key].size():
target_state_dict[new_key] = v
transferred_keys.append(new_key)
if k.startswith('net.0.') and k.replace('net.0.', 'encoder.embedding.', 1) in target_state_dict:
new_key = k.replace('net.0.', 'encoder.embedding.', 1)
if v.size() == target_state_dict[new_key].size():
target_state_dict[new_key] = v
transferred_keys.append(new_key)
target_state_dict_2 = target_model.state_dict()
filtered_state_dict = {k: v for k, v in target_state_dict.items() if k in target_state_dict_2 and v.size() == target_state_dict_2[k].size()}
target_model.load_state_dict(filtered_state_dict, strict=False)
if freeze:
for name, param in target_model.named_parameters():
if name in transferred_keys:
param.requires_grad = False
target_model.to(device)
return target_model, transferred_keys
def get_nth_layer(model, batch, rollout = True, x_ = False, laplacian = True, n = -1, latent = False, device = 'cpu', double_cast = True):
batch[0] = batch[0].to(device)
model.to(device)
if double_cast:
model = model.double()
model.eval()
model.freeze()
atoms, neighbors, distances = batch[0]
L = None
if laplacian:
A = copy.deepcopy(distances[0].float().detach().cpu().numpy())
#A = A[1:, 1:]
A[A==2] = 0
A[A>3] = 0
A[A==3] = 1
A[0,0] = 0
D = np.diag(np.sum(A, axis = 1))
L = D - A
x = model.encoder.atoms_encoder(atoms) + model.encoder.neighbors_encoder(neighbors)
xo = x
d_mask = model.encoder.distance_encoder(distances).permute(0, 3, 1, 2).flatten(end_dim=1)
a_ = None
m = model.hparams['nhead']
if n == -1:
n = model.hparams['num_layers']
for j in range(0,n):
lr = model.encoder.layers[j]
if double_cast:
x = x.double()
d_mask = d_mask.double()
x, a = lr(x, d_mask, need_weights=rollout) # noqa
a = 0.5*(a + torch.eye(a.size(2), device = device).view(1,a.size(1), -1))
if a_ is None:
a_ = a
else:
a_ = torch.bmm(a, a_)
a_= a_.detach().tolist()
zero_mask = atoms != 0
zero_mask = zero_mask.to(device)
#x = x[:,1:,:]
result = x[zero_mask[:, :]].detach().cpu().numpy()
#result = result.reshape(result.shape[0]*result.shape[1],result.shape[2]).detach().cpu().numpy()
if x_:
return result, a_, x, xo, L
else:
return result, a_, L
def Laplacian_Rollout_analysis(checkpoint_path, molecules, is_prepared_as_packed_chython = False, return_weights = False, device = 'cpu'):
loaded_path_hyper_dict = torch.load(checkpoint_path)['hyper_parameters']
try:
model = GT(
checkpoint_path = checkpoint_path,
**loaded_path_hyper_dict
)
except:
print('doing the transfer weight thing')
model = GT(
checkpoint_path = None,
**loaded_path_hyper_dict
)
model, w = transfer_weights(checkpoint_path, model, device = device)
model.eval()
model.freeze()
print(len(w))
dataloader = get_dtl_from_smiles(molecules)#, is_prepared_as_packed_chython)
if not is_prepared_as_packed_chython:
molecules_prepared = []
for mol in molecules:
try:
s = chython.smiles(mol).pack()
molecules_prepared.append(s)
except:
print(mol, "IS NOT VALID")
molecules = molecules_prepared
zetas = []
etas = []
number_of_laplacians_ = []
which_laplacians = []
for i, batch in enumerate(dataloader):
model.eval()
_, rollout, L = get_nth_layer(model, batch, latent = False, laplacian = True, device = device)
rollout = np.array(rollout[0])
eigvals_r, eigvecs_r = np.linalg.eig(rollout)
idx_r = np.flip(np.argsort(np.abs(eigvals_r)))
eigvals_r = eigvals_r[idx_r]
eigvecs_r = eigvecs_r[:,idx_r]
eigvals_l, eigvecs_l = np.linalg.eig(L)
idx_l = np.argsort(eigvals_l)
eigvals_l = eigvals_l[idx_l]
eigvecs_l = eigvecs_l[:,idx_l]
zeta, eta, number_of_laplacians, which_laplacian = get_metrics(eigvecs_r, eigvecs_l, eigvals_r, t = 0.9)
zetas.append(zeta)
etas.append(eta)
number_of_laplacians_.append(number_of_laplacians)
which_laplacians.append(which_laplacian)
if return_weights:
return zetas, etas, number_of_laplacians_, which_laplacians, w
else:
return zetas, etas, number_of_laplacians_, which_laplacians
def get_Jacobian_ij(model, batch, n=None, device='cuda'):
batch[0] = batch[0].to(device)
model.to(device)
atoms, neighbors, distances = batch[0]
encoder = model.encoder
for param in encoder.parameters():
param.requires_grad_(False)
encoder.eval()
xo = encoder.atoms_encoder(atoms) + encoder.neighbors_encoder(neighbors)
xo.requires_grad_(True)
d_mask = encoder.distance_encoders[0](distances).permute(0, 3, 1, 2).flatten(end_dim=1)
d_mask.requires_grad_(True)
def get_output(xo, n = n):
if n is None:
n = len(encoder.layers)-1
x = xo
for k in range(n):
layer = encoder.layers[k]
x, _ = layer(x, d_mask)
return x
jac = jacobian(get_output, xo, vectorize = True)
return jac
def get_sensitivities_per_topdistance(ckpt_path, batch, depth = None):
ckpts = os.listdir(ckpt_path)
ckpt = [c for c in ckpts if c.startswith('epoch')][0]
checkpoint_path = f'{ckpt_path}/{ckpt}'
loaded_path_hyper_dict = torch.load(checkpoint_path)['hyper_parameters']
model = GT(
checkpoint_path = checkpoint_path,
**loaded_path_hyper_dict
)
top_dist = batch[0].distances[0].cpu().numpy()
top_dist = top_dist[1:,1:]
sensitivities = []
if len(batch[0].atoms[0])>=4 and len(batch[0].atoms[0])<=50:
jacc = get_Jacobian_ij(model, batch, n = depth).squeeze(0).squeeze(2).transpose(1,2)
jacc = jacc[1:,1:,:,:]
torch.cuda.empty_cache()
for k in tqdm(range(2, np.max(top_dist))):
jac_avgs = []
for atom_idx in range(0, len(top_dist)):
atoms_to_check = np.argwhere(top_dist[atom_idx] == k).reshape(-1)
jac_avg = []
for j in atoms_to_check:
jac = jacc[atom_idx, j].norm().item()
jac_avg.append(jac)
jac_avgs.append(np.mean(jac_avg))
sensitivities.append(np.mean(jac_avgs))
return sensitivities
else:
print('too short or too long')
print(len(batch[0].atoms[0]))