-
Notifications
You must be signed in to change notification settings - Fork 0
/
util_mag.py
255 lines (209 loc) · 9.49 KB
/
util_mag.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
import torch
import torch.nn as nn
import torch.optim as optim
import torch_geometric as tg
import e3nn
from e3nn import o3
from data_helpers import DataPeriodicNeighbors
from e3nn.nn.models.gate_points_2101 import Convolution, Network
from e3nn.o3 import Irreps
from pymatgen.core.structure import Structure
from pymatgen.ext.matproj import MPRester
import pymatgen.analysis.magnetism.analyzer as pg
import numpy as np
import pickle
import matplotlib.pyplot as plt
from sklearn.metrics import average_precision_score
from sklearn.metrics import classification_report
from sklearn.metrics import f1_score
from sklearn.metrics import accuracy_score
import io
import random
import math
import sys
import time
import os
import datetime
LETTERS = {'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M', 'N', 'O', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z'}
#################
# Model-related #
#################
class AtomEmbeddingAndSumLastLayer(torch.nn.Module):
def __init__(self, atom_type_in, atom_type_out, model):
super().__init__()
self.linear = torch.nn.Linear(atom_type_in, 128)
self.model = model
self.relu = torch.nn.ReLU()
self.linear2 = torch.nn.Linear(128, 96)
self.linear3 = torch.nn.Linear(96, 64)
self.linear4 = torch.nn.Linear(64, 45)
#self.linear5 = torch.nn.Linear(45, 32)
#self.softmax = torch.nn.LogSoftmax(dim=1)
def forward(self, x, *args, batch=None, **kwargs):
output = self.linear(x)
output = self.relu(output)
print(f"Input: {x}")
output = self.linear2(output)
output = self.relu(output)
output = self.linear3(output)
output = self.relu(output)
output = self.linear4(output)
output = self.relu(output)
output = self.model({'x': output, 'batch': batch, **kwargs})
if batch is None:
N = output.shape[0]
batch = output.new_ones(N)
print(f"Output: {output}")
return output
def create_dataloaders(data, batch_size=1):
indices = np.arange(len(data))
np.random.shuffle(indices)
index_tr, index_va, index_te = np.split(
indices, [int(.8 * len(indices)), int(.9 * len(indices))])
assert set(index_tr).isdisjoint(set(index_te))
assert set(index_tr).isdisjoint(set(index_va))
assert set(index_te).isdisjoint(set(index_va))
pickle.dump((index_tr, index_va, index_te), open("data_splits.p", "wb"))
dataloader = tg.loader.DataLoader([data[i] for i in index_tr], batch_size=batch_size, shuffle=True)
dataloader_valid = tg.loader.DataLoader([data[i] for i in index_va], batch_size=batch_size)
return index_tr, index_va, index_te, dataloader, dataloader_valid
def loglinspace(rate, step, end=None):
t = 0
while end is None or t <= end:
yield t
t = int(t + 1 + step * (1 - math.exp(-t * rate / step)))
def evaluate(model, loss_fn, dataloader, device, cost_multiplier=1.0):
model.eval()
loss_cumulative = 0.
with torch.no_grad():
for _, d in enumerate(dataloader):
d.to(device)
output = model(x=d.x, batch=d.batch, pos=d.pos, z=d.pos.new_ones((d.pos.shape[0], 3)))
if d.y.item() == 2:
loss = cost_multiplier*loss_fn(output, d.y).cpu()
print("Multiplied Loss Index \n")
elif d.y.item() == 0 or d.y.item() == 1:
loss = loss_fn(output, d.y).cpu()
print("Standard Loss Index \n")
else:
print("Lost datapoint \n")
loss_cumulative = loss_cumulative + loss.detach().item()
return loss_cumulative / len(dataloader)
def train(model, optimizer, loss_fn, dataloader, dataloader_valid, scheduler, max_iter=101, device="cpu"):
model.to(device)
checkpoint_generator = loglinspace(3.3, 5)
checkpoint = next(checkpoint_generator)
start_time = time.time()
dynamics = []
for step in range(max_iter):
model.train()
loss_cumulative = 0.
for j, d in enumerate(dataloader):
d.to(device)
output = model(x=d.x, batch=d.batch, pos=d.pos, z=d.pos.new_ones((d.pos.shape[0], 3)))
loss = loss_fn(output, d.y).cpu()
print(f"Iteration {step+1:4d} batch {j+1:5d} / {len(dataloader):5d} " + f"batch loss = {loss.data}", end="\r", flush=True)
loss_cumulative = loss_cumulative + loss.detach().item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
end_time = time.time()
wall = end_time - start_time
if step == checkpoint:
checkpoint = next(checkpoint_generator)
assert checkpoint > step
valid_avg_loss = evaluate(model, loss_fn, dataloader_valid, device)
train_avg_loss = evaluate(model, loss_fn, dataloader, device)
dynamics.append({
'step': step,
'wall': wall,
'batch': {'loss': loss.item(),},
'valid': {'loss': valid_avg_loss,},
'train': {'loss': train_avg_loss,},
})
yield {
'dynamics': dynamics,
'state': model.state_dict()
}
print(f"Iteration {step+1:4d} batch {j+1:5d} / {len(dataloader):5d} " +
f"train loss = {train_avg_loss:8.3f} " +
f"valid loss = {valid_avg_loss:8.3f} " +
f"elapsed time = {time.strftime('%H:%M:%S', time.gmtime(wall))}")
with open('loss.txt', 'a') as f:
f.write(f"train average loss: {str(train_avg_loss)} \n")
f.write(f" validation average loss: {str(valid_avg_loss)} \n")
scheduler.step()
def plots(run_name):
saved = torch.load(run_name+'_trial_run_full_data.torch')
steps = [d['step'] + 1 for d in saved['dynamics']]
valid = [d['valid']['loss'] for d in saved['dynamics']]
train = [d['train']['loss'] for d in saved['dynamics']]
plt.plot(steps, train, 'o-', label="train")
plt.plot(steps, valid, 'o-', label="valid")
plt.legend()
plt.savefig(run_name+'_hist.png', dpi=300)
def run_write_data(stage, indices, data, model, device, formula_list_mp, id_list):
composition_dict = {}
sites_dict = {}
y_output = []
y_pred = []
y_actual = []
# only used if stage=='testing'
y_score = []
for _, index in enumerate(indices):
d = tg.data.Batch.from_data_list([data[index]])
d.to(device)
# run the model on the current batch
# pos: position of the nodes (atoms)
# z: attributes of nodes, initialized as blank
output_vec = model(x=d.x, batch=d.batch, pos=d.pos, z=d.pos.new_ones((d.pos.shape[0], 3)))
y_output.append(output_vec)
# if this is the test set, we should also prepare y_actual and y_score to return
if stage == 'testing':
y_actual.append(d.y.item())
y_score.append(output_vec)
with open(f'{stage}_results.txt', 'a') as f:
f.write(f"Output for below sample: {torch.exp(output_vec)} \n")
# find the output encoding
if max(output_vec[0][0], output_vec[0][1], output_vec[0][2]) == output_vec[0][0]:
output = 0
elif max(output_vec[0][0], output_vec[0][1], output_vec[0][2]) == output_vec[0][1]:
output = 1
else:
output = 2
y_pred.append(output)
with open(f'{stage}_results.txt', 'a') as f:
f.write(f"{id_list[index]} {formula_list_mp[index]} Prediction: {output} Actual: {d.y} \n")
correct_flag = d.y.item() == output
# Accuracy per element calculation
current_element = ""
for char_index in range(len(formula_list_mp[index])):
print("Entered Loop")
formula = formula_list_mp[index]
if formula[char_index] in LETTERS:
current_element += formula[char_index]
print(f"Using char: {formula[char_index]}")
if char_index + 1 == len(formula) or formula[char_index + 1].isupper() or formula[char_index + 1] not in LETTERS:
print(f"printing to dict {current_element}")
if correct_flag:
current_entry = composition_dict.get(current_element, [0, 0])
current_entry = [current_entry[0] + 1, current_entry[1] + 1]
else:
current_entry = composition_dict.get(current_element, [0, 0])
current_entry = [current_entry[0], current_entry[1] + 1]
composition_dict[current_element] = current_entry
current_element = ""
# Raw output (i.e. triples)
pickle.dump(y_output, open(f'{stage}_output_raw.p', 'wb'))
# Accuracy per element depiction
with open(f'{stage}_composition_info.txt', 'a') as f:
f.write(f"{stage.capitalize()} Composition Ratios: \n")
for key, value in composition_dict.items():
f.write(f"Element: {key} Ratio: {value[0]}/{value[1]} Fraction: {value[0]/value[1]}\n")
# Accuracy per nsites depiction
with open(f'{stage}_nsites_info.txt', 'a') as f:
f.write(f"{stage.capitalize()} Nsites Info: \n")
for key, value in sites_dict.items():
f.write(f"nsites: {key} Ratio: {value[0]}/{value[1]} Fraction: {value[0]/value[1]}\n")
if stage == 'testing':
return y_actual, y_pred