-
Notifications
You must be signed in to change notification settings - Fork 0
/
trainer.py
165 lines (148 loc) · 6.44 KB
/
trainer.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
from collections import defaultdict
import torch
import torch.nn as nn
from mighty.monitor.accuracy import AccuracyEmbedding
from mighty.trainer import TrainerEmbedding, TrainerGrad
from mighty.utils.common import batch_to_cuda, clone_cpu
from mighty.utils.data import DataLoader
from mighty.utils.stub import OptimizerStub
from mighty.utils.var_online import MeanOnlineLabels
from nn.kwta import WTAInterface, IterativeWTASoft
from nn.monitor import MonitorIWTA
from nn.utils import compute_loss, l0_sparsity
from mighty.monitor.accuracy import calc_accuracy, AccuracyEmbedding
class TrainerIWTA(TrainerEmbedding):
watch_modules = TrainerEmbedding.watch_modules + (WTAInterface,)
N_CHOOSE = 100
LEARNING_RATE = 0.001
def __init__(self,
model: nn.Module,
criterion: nn.Module,
data_loader: DataLoader,
optimizer=OptimizerStub(),
**kwargs):
super().__init__(model=model,
criterion=criterion,
data_loader=data_loader,
optimizer=optimizer,
accuracy_measure=AccuracyEmbedding(cache=True),
**kwargs)
self.mutual_info.save_activations = self.mi_save_activations_y
self.cached_labels = []
self.cached_output = defaultdict(list)
self.cached_output_prev = {}
self.loss_x = None
self.accuracy_x = None
def log_trainer(self):
super().log_trainer()
self.monitor.log(f"LEARNING_RATE={self.LEARNING_RATE}")
self.monitor.log(f"N_CHOOSE={self.N_CHOOSE}")
def mi_save_activations_y(self, module, tin, tout):
"""
A hook to save the activates at a forward pass.
"""
if not self.mutual_info.is_updating:
return
h, y = tout
layer_name = self.mutual_info.layer_to_name[module]
tout_clone = clone_cpu(y.detach().float())
tout_clone = tout_clone.flatten(start_dim=1)
self.mutual_info.activations[layer_name].append(tout_clone)
def _init_monitor(self, mutual_info):
monitor = MonitorIWTA(
mutual_info=mutual_info,
normalize_inverse=self.data_loader.normalize_inverse
)
return monitor
def full_forward_pass(self, train=True):
if not train:
return None
return super().full_forward_pass(train=train)
def update_contribution(self, h, y):
freq = dict(y=y.mean(dim=0), h=h.mean(dim=0))
for name, param in self.model.named_parameters():
param.update_contribution(freq[name[-1]])
def train_batch(self, batch):
x, labels = batch
h, y = self.model(x)
self.update_contribution(h, y)
loss = self._get_loss(batch, (h, y))
if isinstance(self.model, IterativeWTASoft):
loss.backward()
self.optimizer.step(closure=None)
else:
self.model.update_weights(x, h, y, n_choose=self.N_CHOOSE,
lr=self.LEARNING_RATE)
return loss
def _update_cached(self):
labels = torch.cat(self.cached_labels)
convergence = {}
sparsity = {}
for name, output in self.cached_output.items():
output = torch.cat(output)
mean = [output[labels == label].mean(dim=0)
for label in labels.unique()]
mean = torch.stack(mean)
self.monitor.clusters_heatmap(mean, title=f"Embeddings '{name}'")
# self.monitor.plot_assemblies(output, labels, name=name)
loss = compute_loss(output, labels)
self.monitor.update_loss(loss, mode=f'pairwise {name}')
sparsity[name] = l0_sparsity(output)
output = output.int()
if name in self.cached_output_prev:
xor = (self.cached_output_prev[name] ^ output).sum(dim=1)
convergence[name] = xor.float().mean().item() / output.size(1)
self.cached_output_prev[name] = output
self.monitor.update_output_convergence(convergence)
self.monitor.update_sparsity(sparsity)
self.monitor.update_loss(loss=self.loss_x, mode='pairwise x')
self.cached_output.clear()
self.cached_labels.clear()
if self.timer.epoch == self.timer.n_epochs:
print(f"convergence={convergence}")
print(f"sparsity={sparsity}")
def training_started(self):
# self.monitor.weights_heatmap(self.model)
self.monitor.update_weight_sparsity(self.model.weight_sparsity())
self.monitor.update_s_w(self.model.s_w())
x_centroids = AccuracyEmbedding()
x = []
for x_batch, labels in self.data_loader.eval():
x_batch, labels = batch_to_cuda((x_batch, labels))
x.append(x_batch)
x_centroids.partial_fit(x_batch, labels)
h, y = self.model(x_batch)
self.cached_labels.append(labels)
self.cached_output['h'].append(h)
self.cached_output['y'].append(y)
x = torch.cat(x)
self.monitor.log(f"sparsity x: {l0_sparsity(x):.3f}")
labels = torch.cat(self.cached_labels)
self.accuracy_x = calc_accuracy(labels, x_centroids.predict(x))
print(f"accuracy x: {self.accuracy_x:.3f}")
self.loss_x = compute_loss(x, labels)
self._update_cached()
def _epoch_finished(self, loss):
# self.monitor.weights_heatmap(self.model)
# self.monitor.update_permanences_removed(self.model.permanences_removed())
self.monitor.update_contribution(self.model.weight_contribution())
self.monitor.update_kwta_thresholds(self.model.kwta_thresholds())
self.monitor.update_weight_sparsity(self.model.weight_sparsity())
self.monitor.update_s_w(self.model.s_w())
self._update_cached()
self.model.epoch_finished()
TrainerGrad._epoch_finished(self, loss)
def _on_forward_pass_batch(self, batch, output, train):
h, y = output
if train:
x, labels = batch
self.cached_labels.append(labels)
self.cached_output['h'].append(h)
self.cached_output['y'].append(y)
TrainerGrad._on_forward_pass_batch(self, batch, y, train)
def _get_loss(self, batch, output):
# In case of unsupervised learning, '_get_loss' is overridden
# accordingly.
input, labels = batch
h, y = output
return self.criterion(y, labels)