-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathproto_net.py
211 lines (166 loc) · 7.02 KB
/
proto_net.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
""" Prototype Network """
import os
import sys
import time
import math
import librosa
import numpy as np
import matplotlib.pyplot as plt
import random
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as cp
from torch.nn.parameter import Parameter
import datetime
import _pickle as cPickle
from prototype_memory import *
import copy
# seed = 0
# torch.backends.cudnn.deterministic = True
# random.seed(seed)
# torch.manual_seed(seed)
# torch.cuda.manual_seed(seed)
# np.random.seed(seed)
class ProtoNet(nn.Module):
def __init__(self, extractor,n_dim,n_classes,memory=None, margin = None):
super(ProtoNet,self).__init__()
self.n_dim = n_dim
self.n_classes = n_classes
self.extractor = extractor.cuda()
if not memory:
self.memory = PrototypeMemory()
#self.memory.zero_initialization(self.n_dim, self.n_base_classes)
self.margin = margin
def forward_offline(self, support, y_support, query,hidden, hidden_query=None, DSoftmax=False):
if self.training:
_,_, z_support = self.extractor.forward(support,hidden,support.size(0))
self.memory.initialize_prototypes(z_support,np.argmax(y_support.data.cpu(),axis=1))
if hidden_query:
_,h, z_query = self.extractor.forward(query,hidden_query,query.size(0))
else:
_,h, z_query = self.extractor.forward(query,hidden,query.size(0))
#print(type(np.array(list(self.memory.prototypes.values()))),np.shape(np.array(list(self.memory.prototypes.values()))))
z_proto = torch.from_numpy(np.array(list(self.memory.prototypes.values()))).float().cuda()
#print(np.shape(z_proto),np.shape(z_query))
dists = self.compute_euclidean(z_query,z_proto)
#print(np.shape(dists))
#dists = self.compute_euclidean1(z_query,z_proto)
#assert dists == dists1
if self.training and DSoftmax:
return dists, h
else:
#print('not training')
p_y = F.softmax(-dists,dim=1)
return p_y,h
def forward_inference(self, support, y_support, query, hidden,hidden_query=None):
## training model on base data
## training model on base data
self.extractor.eval()
if hidden_query:
_,h, z_query = self.extractor.forward(query,hidden_query,query.size(0))
else:
_,h, z_query = self.extractor.forward(query,hidden,query.size(0))
_,_, z_support = self.extractor.forward(support,hidden,support.size(0))
self.inference_prototypes(z_support,np.argmax(y_support.data.cpu(),axis=1))
#print(type(np.array(list(self.memory.prototypes.values()))),np.shape(np.array(list(self.memory.prototypes.values()))))
z_proto = torch.squeeze(torch.from_numpy(copy.deepcopy(np.array(list(self.prototypes.values())))).float()).cuda()
dists = self.compute_euclidean(z_query,z_proto)
#dists = self.compute_euclideanFIX(z_query)
#print(np.shape(dists))
#dists = self.compute_euclidean1(z_query,z_proto)
#assert dists == dists1
p_y = F.softmax(-dists,dim=1)
return p_y,h
def inference_prototypes(self, X, y):
classes = np.sort(np.unique(y))
self.prototypes = dict()
self.counters = dict()
#print(classes)
for c in classes:
p_mean = X[y==c].mean(0)
#print(np.shape(X[y==c]))
self.prototypes[c] = copy.deepcopy(list(p_mean.data.cpu().numpy().flatten()))
self.counters[c] = len(X[y==c])
def prototype_update_momentum(self, support, y_support, momentum,hidden):
self.extractor.eval()
_,_, z_support = self.extractor.forward(support,hidden,support.size(0))
self.inference_prototypes(z_support, np.argmax(y_support.data.cpu(),axis=1))
self.memory.update_prototypes_momentum(self.prototypes,momentum)
def forward_online(self, support, y_support, query, hidden_support, hidden_query, lwf=False):
if self.training:
#print("Training ... ")
_,_, z_support = self.extractor.forward(support,hidden_support,support.size(0))
self.memory.update_prototypes(z_support.data.cpu().numpy(),np.argmax(y_support.data.cpu(),axis=1))
_,h, z_query = self.extractor.forward(query,hidden_query,query.size(0))
#z_proto = torch.squeeze(torch.from_numpy(np.array(list(self.memory.prototypes.values()))).float()).cuda()
#dists = self.compute_euclidean(z_query,z_proto)
#print(dists)
#print(log_p_y)
dists = self.compute_euclideanFIX(z_query)
#print(dists)
#print(np.shape(dists))
#dists = self.compute_euclidean1(z_query,z_proto)
#assert dists == dists1
p_y = F.softmax(-dists,dim=1)
#print(log_p_y)
#sys.exit()
if self.training:
if lwf:
return dists, h, z_support
return p_y,h,dists
else:
if lwf:
return dists, h
return p_y, h
def forward_online_QUERY(self, query, hidden_query, lwf=False):
_,h, z_query = self.extractor.forward(query,hidden_query,query.size(0))
z_proto = torch.squeeze(torch.from_numpy(np.array(list(self.memory.prototypes.values()))).float()).cuda()
#dists = self.compute_euclidean(z_query,z_proto)
dists = self.compute_euclideanFIX(z_query)
#print(np.shape(dists))
#dists = self.compute_euclidean1(z_query,z_proto)
#assert dists == dists1
if lwf:
return dists, h
p_y = F.softmax(-dists,dim=1)
return p_y, h
def online_update_prototypes(self, support, y_support, hidden_support):
_,_,z_support = self.extractor.forward(support,hidden_support,support.size(0))
self.memory.update_prototypes(z_support.data.cpu().numpy(), np.argmax(y_support.data.cpu(),axis=1))
return z_support
#z_proto = torch.squeeze(torch.from_numpy(np.array(list(self.memory.prototypes.values()))).float()).cuda()
#dists = self.compute_euclidean(z_query, z_proto)
#log_p_y = F.softmax(-dists,dim=1)
#return log_p_y, h, z_support
def compute_euclideanFIX(self, z_query):
dists = torch.ones((len(z_query),self.n_classes))*float('inf')
dists = dists.float().cuda()
#print("CURRENT CLASSES IN PROTOTYPE MEMORY: ", list(self.memory.prototypes.keys()))
for c in self.memory.prototypes.keys():
z_proto = torch.from_numpy(self.memory.prototypes[c][None,:]).float().cuda()
dist = self.compute_euclidean(z_query,z_proto)
#print(np.shape(dist))
dists[:,c] = torch.squeeze(dist)
#print(np.shape(dists))
return dists
def update_protoMemory(self, z_support, y_support):
#_,_, z_support = self.extractor.forward(support,hidden,support.size(0))
self.memory.initialize_prototypes(z_support,np.argmax(y_support.data.cpu(),axis=1))
def compute_euclidean(self,query, proto):
#print(np.shape(query), np.shape(proto))
# query_n = torch.linalg.norm(query, dim=1, keepdims=True)
# proto_n = torch.linalg.norm(proto, dim=1, keepdims=True)
#import pdb; pdb.set_trace()
x = query.unsqueeze(1).expand(query.size(0),proto.size(0),query.size(1))
y = proto.unsqueeze(0).expand(query.size(0),proto.size(0),query.size(1))
#print(np.shape(x),np.shape(y))
return torch.pow(x-y,2).sum(2)
def set_memory(self, memory):
self.memory = memory
def compute_euclidean1(self,a,b):
a2 = tf.cast(tf.reduce_sum(tf.square(a),[-1],keepdims=True),dtype=tf.float32)
ab = tf.cast(tf.matmul(a,b, transpose_b=True), dtype=tf.float32)
b2 = tf.cast(tf.repeat(tf.reduce_sum(tf.square(b),[-1],keepdims=True), len(a),axis=0), dtype=tf.float32)
#print(np.shape(a),np.shape(b),np.shape(a2),np.shape(ab),np.shape(b2))
return a2 - 2*ab + b2