-
Notifications
You must be signed in to change notification settings - Fork 0
/
rosetta.py
348 lines (290 loc) · 11.4 KB
/
rosetta.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
# Standard
import argparse
from collections import namedtuple
import datetime
# ML
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
# DL
import torch
import torch.nn.functional as F
import torch.utils.data as data_utils
from torch import optim
from torch.utils.tensorboard import SummaryWriter
import datetime
# In house
from models import RecursiveNN, RecursiveNN_Linear, ModelBlock, weights_init
from feature_extraction import FeatureExtractor
from utils import load_features, create_loader
logdir = "./logs/"
def train_model(
model, train_loader, optimizer, epoch, log_interval=100, scheduler=None, writer=None
):
"""Manage the training process of the model for one epoch."""
tloss = 0
"""SMOOTH L1 LOSS: Creates a criterion that uses a squared term if the absolute
element-wise error falls below 1 and an L1 term otherwise.
It is less sensitive to outliers than the `MSELoss` and in some cases
prevents exploding gradients """
for batch_idx, (lsr, feats, targets) in enumerate(train_loader):
optimizer.zero_grad()
outputs = model(lsr, feats)
loss = F.smooth_l1_loss(outputs, targets.view(-1))
# loss = F.mse_loss(outputs, targets.view(-1))
tloss += loss.item()
loss.backward()
optimizer.step()
# print(
# "Train Epoch: {:02d} -- Batch: {:03d} -- Loss: {:.4f}".format(
# epoch, batch_idx, tloss
# )
# )
# Write loss to tensorboard
if writer != None:
writer.add_scalar("Train/Loss", tloss, epoch)
if scheduler is not None:
scheduler.step()
def test_model(model, test_loader, epoch, writer=None, score=False):
"""Output test loss and/or score for one epoch."""
test_loss = 0.0
with torch.no_grad():
for lsr, feats, targets in test_loader:
outputs = model(lsr, feats)
test_loss += F.mse_loss(outputs, targets.view(-1)).item()
# print("\nTest set: Average loss: {:.4f}\n".format(test_loss))
# Write loss to tensorboard
if writer != None:
writer.add_scalar("Test/Loss", test_loss, epoch)
if score: # For evolutionary algorithms
df = pd.DataFrame({"real": targets.view(-1), "preds": outputs}).fillna(0)
df = df.corr().fillna(0)
score = abs(df["preds"]["real"])
return test_loss, score
else:
return test_loss
class Rosetta:
"""Rosetta stone regressor.
Main class orchestrating whole regression pipeline."""
def __init__(self, mode="extract", bSave="T", bUseConv=False):
self.mode = mode
if bSave is "T":
self.bSave = False
else:
self.bSave = False
self.bUseConv = bUseConv
# Saved model
self.model = None
self.full_data = True
# Define all hyperparameters
if self.bUseConv:
self.params = {
"step_size": 5,
"gamma": 0.8,
"batch_size_train": 64,
"batch_size_test": 128,
"lr": 4e-04,
"epochs": 40,
"NBaseline": 10,
"upsampling_factor": 3000,
"upsample": False,
"conv_dict": {
"InChannels": [2],
"OutChannels": [2],
"Ksze": [1],
"Stride": [1],
"Padding": [0],
"MaxPoolDim": 1,
"MaxPoolBool": False,
},
"conv_ffnn_dict": {
"laser_hidden_layers": [64, 16],
"mixture_hidden_layers": [32, 32, 1],
},
}
else:
self.params = {
"N1": 40,
"N2": 20,
"batch_size_test": 100,
"batch_size_train": 500,
"dropout": 0,
"epochs": 30,
"leaky_relu": True,
"lr": 0.0003,
"out_features": 27,
"score": 0.09012854763115952,
"upsample": False,
"upsampling_factor": 5000,
"step_size": 40,
"gamma": 0.5,
}
def upsample(self, data):
"""Upsample data to make score distribution more uniform."""
nlp = data.feats
scores = data.scores
lsr = data.lsr
if self.params["upsample"]:
# Define parameters
alpha = 0.45
beta = 15
gamma = 0.05
if self.bUseConv:
lsr = lsr.reshape(-1, 2048)
# Retrieve score distribution in 15 bins
n, bins, _ = plt.hist(
scores, 15, density=True, range=(-1, 1), facecolor="g", alpha=0.75
)
# Create upsampling distribution
prob_dist = np.ones(len(n)) - n * alpha
prob_dist = prob_dist ** beta / sum(prob_dist)
# Assign upsampling distribution to each score
probs = np.ones(len(scores))
scores = scores.ravel()
for idx in range(len(bins) - 1):
probs[(scores > bins[idx]) & (scores < bins[idx + 1])] = (
1 * prob_dist[idx]
)
scaled_probs = probs / sum(probs)
# Select indices to upsample
idxs = np.random.choice(
list(range(len(scores))),
p=scaled_probs,
size=self.params["upsampling_factor"],
)
# Create upsampling data subset with random noise
augmented_lsr = np.zeros((len(idxs), lsr.shape[1]))
augemented_nlp = np.zeros((len(idxs), nlp.shape[1]))
augmented_scores = np.zeros((len(idxs), scores.shape[1]))
lsr_std = lsr.std(axis=0)
nlp_std = nlp.std(axis=0)
scores_std = scores.std(axis=0)
for i, value in enumerate(idxs):
augmented_lsr[i, :] = lsr[value, :] + np.random.normal(
0, lsr_std * gamma, lsr.shape[1]
)
augemented_nlp[i, :] = nlp[value, :] + np.random.normal(
0, nlp_std * gamma, nlp.shape[1]
)
augmented_scores[i, :] = scores[value, :] + np.random.normal(
0, scores_std * gamma, scores.shape[1]
)
# Concatenate initial data with upsampled data
final_lsr = np.concatenate([lsr, augmented_lsr], axis=0)
final_nlp = np.concatenate([nlp, augemented_nlp], axis=0)
final_scores = np.concatenate([scores, augmented_scores], axis=0)
if self.bUseConv:
final_lsr = final_lsr.reshape(-1, 2, 1024)
else:
final_lsr = lsr
final_nlp = nlp
final_scores = scores
res = namedtuple("res", ["lsr", "feats", "scores"])(
lsr=final_lsr, feats=final_nlp, scores=final_scores
)
return res
def write_predictions(self):
"""Output the predictions to a text file."""
res = FeatureExtractor("test").run()
model = torch.load("model.pt")
test = namedtuple("res", ["lsr", "feats", "scores"])(
lsr=res.lsr.reshape(-1, 2048), feats=res.feats, scores=res.scores
)
dev_ = data_utils.TensorDataset(
*[
torch.tensor(getattr(test, i)).float()
for i in ["lsr", "feats", "scores"]
]
)
with torch.no_grad():
preds = model.forward(*dev_.tensors[:2]).cpu().numpy()
np.set_printoptions(suppress=True)
np.savetxt("predictions.txt", preds.astype(float), delimiter="\n", fmt="%f")
print("Predictions saved to predictions.txt")
def run(self):
"""Run whole data loading, feature extraction, model training and regressing pipeline."""
if self.mode == "extract":
print("Extracting features")
train = FeatureExtractor("train").run()
dev = FeatureExtractor("dev").run()
print("Saving features")
np.save("saved_features/train_lsr", train.lsr)
np.save("saved_features/train_nlp", train.feats)
np.save("saved_features/train_scores", train.scores)
np.save("saved_features/dev_lsr", dev.lsr)
np.save("saved_features/dev_nlp", dev.feats)
np.save("saved_features/dev_scores", dev.scores)
else: # Load saved extracted features
print("Loading saved features")
split = False if self.full_data else True
train, dev = load_features(split=split, nt=True)
if self.params["upsample"]:
train = self.upsample(train)
train_loader = create_loader(train, self.params["batch_size_train"])
dev_loader = create_loader(dev, validate=True)
# We set a random seed to ensure that results are reproducible.
# Also set a cuda GPU if available
if torch.cuda.is_available():
torch.backends.cudnn.deterministic = True
GPU = True
else:
GPU = False
device_idx = 0
if GPU:
device = torch.device(
"cuda:" + str(device_idx) if torch.cuda.is_available() else "cpu"
)
else:
device = torch.device("cpu")
print(f"Running on {device}")
if self.bUseConv:
model = RecursiveNN(
ModelBlock,
self.params["conv_dict"],
self.params["conv_ffnn_dict"],
BASELINE_dim=self.params["NBaseline"],
)
else:
model = RecursiveNN_Linear(
in_features=2048,
N1=self.params["N1"],
N2=self.params["N2"],
out_features=self.params["out_features"],
dropout=self.params["dropout"],
leaky_relu=self.params["leaky_relu"],
)
model = model.to(device)
weights_initialiser = True
if weights_initialiser:
model.apply(weights_init)
params_net = sum(p.numel() for p in model.parameters() if p.requires_grad)
print("Total number of parameters in Model is: {}".format(params_net))
print(model)
optimizer = optim.Adam(model.parameters(), lr=self.params["lr"])
scheduler = optim.lr_scheduler.StepLR(
optimizer, step_size=self.params["step_size"], gamma=self.params["gamma"]
)
date_string = (
str(datetime.datetime.now())[:16].replace(":", "-").replace(" ", "-")
)
writer = SummaryWriter(logdir + date_string)
print("Running model")
for epoch in range(self.params["epochs"]):
train_model(
model,
train_loader,
optimizer,
epoch,
log_interval=1000,
scheduler=scheduler,
writer=writer,
)
test_loss = test_model(model, dev_loader, epoch, writer=writer)
torch.save(model, "model.pt")
self.model = model
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Process input args")
parser.add_argument("mode", type=str, nargs="+", help="extract or no-extract")
parser.add_argument("save", type=str, nargs="+", help="T / F for save or not save")
args = parser.parse_args().__dict__
Rosetta(args["mode"][0], args["save"][0]).run()