-
Notifications
You must be signed in to change notification settings - Fork 0
/
deepclusterer.py
339 lines (282 loc) · 12.6 KB
/
deepclusterer.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
import pickle
from deepcluster.util import AverageMeter
import deepcluster.models.mini_models as dc_models
import faiss
import ipdb
import torch
import torch.nn as nn
from itertools import chain
import numpy as np
import os
import torch.utils.data
import torch.utils.data.sampler
import mixture.models as mix_models
import matplotlib.pyplot as plt
class IdentityLinear(nn.Module):
def __init__(self, in_features, out_features, bias=True):
super(IdentityLinear, self).__init__()
self.linear_positive = nn.Linear(in_features, out_features, bias=bias)
self.linear_negative = nn.Linear(in_features, out_features, bias=bias)
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
positive = self.relu(self.linear_positive(x))
negative = -self.relu(self.linear_negative(-x))
return positive + negative
class MLPNet(nn.Module):
r"""
A variant of AlexNet.
The changes with respect to the original AlexNet are:
- LRN (local response normalization) layers are not included
- The Fully Connected (FC) layers (fc6 and fc7) have smaller dimensions
due to the lower resolution of mini-places images (128x128) compared
with ImageNet images (usually resized to 256x256)
"""
def __init__(self, num_classes, **kwargs):
super(MLPNet, self).__init__()
self.hidden_size = 256
self.feature_size = 256
self.num_classes = num_classes
self.encoder = nn.Sequential(
# linears
nn.Linear(in_features=2, out_features=self.hidden_size, bias=True),
nn.ReLU(inplace=True),
nn.Linear(in_features=self.hidden_size, out_features=self.hidden_size, bias=True),
nn.ReLU(inplace=True),
nn.Linear(in_features=self.hidden_size, out_features=self.feature_size, bias=True),
nn.ReLU(inplace=True),
# for identity init
# IdentityLinear(2, self.hidden_size, True),
# IdentityLinear(self.hidden_size, self.hidden_size, True),
# IdentityLinear(self.hidden_size, self.hidden_size, True),
# IdentityLinear(self.hidden_size, self.hidden_size, True),
# IdentityLinear(self.hidden_size, self.feature_size, True),
# nn.ReLU(inplace=True)
)
self.top_layer = nn.Sequential(
nn.Linear(self.feature_size, num_classes),
)
self.init_model()
def init_model(self):
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
nn.init.kaiming_normal(m.weight.data)
if m.bias is not None:
m.bias.data.zero_()
elif classname == 'Linear':
# init normal
# nn.init.normal_(m.weight.data, std=0.01)
# init orthgonal
# nn.init.orthogonal_(m.weight.data, gain=1)
# init
# init identity + normal
nn.init.eye_(m.weight.data)
normal_init = torch.zeros_like(m.weight.data)
nn.init.normal_(normal_init, std=0.01)
m.weight.data.add_(normal_init)
if m.bias is not None:
m.bias.data.zero_()
self.apply(weights_init)
return self
def forward(self, input):
x = self.encoder(input)
if self.top_layer:
x = self.top_layer(x)
return x
def prepare_for_inference(self):
self.top_layer = None
assert type(list(self.encoder.children())[-1]) is nn.ReLU
self.encoder = nn.Sequential(*list(self.encoder.children())[:-1])
self.eval()
def prepare_for_training(self):
assert self.top_layer is None
encoder = list(self.encoder.children())
encoder.append(nn.ReLU(inplace=True).cuda())
self.encoder = nn.Sequential(*encoder)
self.top_layer = nn.Linear(in_features=self.feature_size, out_features=self.num_classes)
self.top_layer.weight.data.normal_(0, 0.01)
self.top_layer.bias.data.zero_()
self.top_layer.cuda()
self.train()
class DeepClusterer:
def __init__(self, args, clusterer):
self.args = args
self.clusterer = clusterer
self.model = MLPNet(args.num_components)
self.pca = None
# optimization
self.lr = 0.01
self.momentum = 0.9
self.weight_decay = 1e-5
self.num_epochs = 10
self.batch_size_trajectory = 64
self.num_workers = 1
self.model.top_layer = None # this gets its own optimizer later
self.model_optimizer = torch.optim.SGD(
filter(lambda x: x.requires_grad, self.model.parameters()),
lr=self.lr,
momentum=self.momentum,
weight_decay=self.weight_decay
)
# data
self.num_trajectories = -1
self.episode_length = args.episode_length
self.feature_size = self.model.feature_size
self.visualize_features = False
def preprocess_trajectories(self, trajectories):
if isinstance(trajectories, list) and isinstance(trajectories[0], list):
trajectories = list(chain(*trajectories))
trajectories = torch.stack(trajectories, dim=0)
elif isinstance(trajectories, np.ndarray):
trajectories = torch.FloatTensor(trajectories)
else:
raise ValueError
self.num_trajectories = trajectories.shape[0]
assert self.episode_length == trajectories.shape[1]
return trajectories
def fit(self, trajectories):
trajectories = self.preprocess_trajectories(trajectories)
states = trajectories.reshape([-1, trajectories.shape[-1]])
dataset = torch.utils.data.TensorDataset(trajectories)
dataloader = torch.utils.data.DataLoader(dataset,
shuffle=False,
batch_size=self.batch_size_trajectory,
num_workers=self.num_workers,
pin_memory=False)
for i_epoch in range(self.num_epochs):
print('epoch {}'.format(i_epoch))
# infer features, preprocess features
# if i_epoch != 0:
features = self.compute_features(dataloader)
assert features.ndim == 3
features = np.reshape(features, [-1, features.shape[-1]])
# else:
# features = states
# features = self.preprocess_features(features)
# visualize features
if self.visualize_features:
self.visualize(features, i_epoch)
# cluster features
labels = self.clusterer.fit_predict(features, group=self.episode_length)
# print('EM lower bound: {}'.format(self.clusterer.lower_bound_))
print('k-means inertia: {}'.format(self.clusterer.inertia_))
if i_epoch != self.num_epochs - 1:
# assign pseudo-labels
labels = torch.LongTensor(labels)
train_dataset = torch.utils.data.TensorDataset(states, labels)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=self.batch_size_trajectory*self.episode_length,
num_workers=self.num_workers,
shuffle=True
)
# weigh loss according to inverse-frequency of cluster
weight = torch.zeros(self.args.num_components, dtype=torch.float32)
unique_labels, counts = np.unique(labels.numpy(), return_counts=True)
inverse_counts = 1 / counts
p_label = inverse_counts / np.sum(inverse_counts)
for i, label in enumerate(unique_labels):
weight[label] = p_label[i]
loss_function = nn.CrossEntropyLoss(weight=weight).cuda()
# train model
self.train(train_dataloader, loss_function)
def compute_features(self, dataloader):
self.model.cuda()
self.model.prepare_for_inference()
features = np.zeros((self.num_trajectories, self.episode_length, self.feature_size), dtype=np.float32)
with torch.no_grad():
for i, (input_tensor,) in enumerate(dataloader):
feature_batch = self.model(input_tensor.cuda()).cpu().numpy()
if i < len(dataloader) - 1:
features[i * self.batch_size_trajectory: (i + 1) * self.batch_size_trajectory] = feature_batch
else:
features[i * self.batch_size_trajectory:] = feature_batch
self.model.cpu()
return features
def train(self, loader, loss_function):
losses = AverageMeter()
self.model.prepare_for_training()
# create optimizer for top layer
optimizer_tl = torch.optim.SGD(
self.model.top_layer.parameters(),
lr=self.lr,
weight_decay=self.weight_decay
)
self.model.cuda()
for _ in range(10):
for i, (input_tensor, label) in enumerate(loader):
output = self.model(input_tensor.cuda())
loss = loss_function(output, label.cuda())
losses.update(loss.item(), label.shape[0])
self.model_optimizer.zero_grad()
optimizer_tl.zero_grad()
loss.backward()
self.model_optimizer.step()
optimizer_tl.step()
self.model.cpu()
print('classification loss: {}'.format(losses.avg))
def visualize(self, features, iteration, trajectories=None):
# trajectories = self.preprocess_trajectories(trajectories)
# states = trajectories.reshape([-1, trajectories.shape[-1]])
#
# dataset = torch.utils.data.TensorDataset(trajectories)
# dataloader = torch.utils.data.DataLoader(dataset,
# shuffle=False,
# batch_size=self.batch_size,
# num_workers=self.num_workers,
# pin_memory=False)
#
# features = self.compute_features(dataloader)
features = features.reshape([-1, features.shape[-1]])
#
#
#
# fig, axes = plt.subplots(1, 1, sharex='all', sharey='all', figsize=[5, 5])
#
# xs = np.arange(start=-10, stop=10, step=1, dtype=np.float32)
# ys = np.arange(start=-10, stop=10, step=1, dtype=np.float32)
# x, y = np.meshgrid(xs, ys)
#
# c = np.linspace(0, 1, num=len(xs)*len(ys)).reshape([len(xs), len(ys)])
# ipdb.set_trace()
# plt.scatter(states[:, 0], states[:, 1], s=2**2)
# plt.savefig('./vis/deepcluster_states.png')
#
# plt.clf()
os.makedirs(self.args.log_dir, exist_ok=True)
plt.scatter(features[:, 0], features[:, 1], s=2**2)
plt.savefig(os.path.join(self.args.log_dir, 'features_{}.png'.format(iteration)))
plt.close('all')
if __name__ == '__main__':
# filename = '/home/kylehsu/experiments/umrl/output/point2d/20190108/context-all_mog_K50_T50_lambda0.5_ent0.1_N1000/history.pkl'
# history = pickle.load(open(filename, 'rb'))
# trajectories = history['trajectories']
import argparse
parser = argparse.ArgumentParser()
args = parser.parse_args()
args.num_components = 20
args.episode_length = 100
args.seed = 1
# args.log_dir = './output/deepcluster/cluster-kmeans_init-normal_layers5_h4_f2'
args.log_dir = './output/deepcluster/debug'
# set seeds
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
np.random.seed(args.seed)
from mixture.models import BayesianGaussianMixture, GaussianMixture, KMeans
# clusterer = GaussianMixture(n_components=args.num_components,
# covariance_type='full',
# verbose=1,
# verbose_interval=100,
# max_iter=1000,
# n_init=1)
clusterer = KMeans(n_clusters=args.num_components,
n_init=1,
max_iter=300,
verbose=0,
algorithm='full')
dc = DeepClusterer(args, clusterer)
trajectories = np.load('./mixture/data_itr20.pkl')
trajectories = trajectories / 10
dc.fit(trajectories)
# dc.visualize(trajectories)