-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathplot_memorizing_associations_task_relu.py
164 lines (135 loc) · 6.94 KB
/
plot_memorizing_associations_task_relu.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
"""Plot layer outputs of the model for the memorizing associations tasks with ReLU activation"""
import argparse
import random
import sys
from collections import OrderedDict
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.backends.cudnn as cudnn
from data.memorizing_associations_dataset import MemorizingAssociationsDataset
from functions.plasticity_functions import InvertedOjaWithSoftUpperBound
from models.network_models import MemorizingAssociationsReLU
from utils.checkpoint import load_checkpoint
def main():
parser = argparse.ArgumentParser(description='Memorizing associations task plotting')
parser.add_argument('--checkpoint_path', default='', type=str, metavar='PATH',
help='Path to checkpoint (default: none)')
parser.add_argument('--check_params', default=1, type=int, choices=[0, 1], metavar='CHECK_PARAMS',
help='When loading from a checkpoint check if the model was trained with the same parameters '
'as requested now (default: 1)')
parser.add_argument('--sequence_length', default=10, type=int, metavar='SEQUENCE_LENGTH',
help='The number of vector-label pairs (default: 10)')
parser.add_argument('--num_classes', default=10, type=int, metavar='NUM_CLASSES',
help='The number of classes (default: 10)')
parser.add_argument('--feature_size', default=10, type=int, metavar='FEATURE_SIZE',
help='Size of the input features (default: 10)')
parser.add_argument('--embedding_size', default=80, type=int, metavar='N',
help='Embedding size (default: 80)')
parser.add_argument('--memory_size', default=100, type=int, metavar='N',
help='Size of the memory matrix (default: 100)')
parser.add_argument('--w_max', default=1.0, type=float, metavar='N',
help='Soft maximum of Hebbian weights (default: 1.0)')
parser.add_argument('--gamma_pos', default=0.1, type=float, metavar='N',
help='Write factor of Hebbian rule (default: 0.1)')
parser.add_argument('--gamma_neg', default=0.1, type=float, metavar='N',
help='Forget factor of Hebbian rule (default: 0.1)')
parser.add_argument('--dataset_seed', default=42, type=int, metavar='N',
help='Seed for creating the dataset (default: 42)')
parser.add_argument('--seed', default=None, type=int, metavar='N',
help='Seed for initializing (default: none)')
args = parser.parse_args()
if args.seed is not None:
random.seed(args.seed)
torch.manual_seed(args.seed)
cudnn.deterministic = True
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Data loading code
dataset = MemorizingAssociationsDataset(sequence_length=args.sequence_length, num_classes=args.num_classes,
feature_size=args.feature_size, inf_data=False, dataset_size=1,
seed=args.dataset_seed)
# Create the model
model = MemorizingAssociationsReLU(
input_size=args.feature_size,
output_size=args.num_classes,
num_embeddings=args.num_classes,
embedding_size=args.embedding_size,
memory_size=args.memory_size,
plasticity_rule=InvertedOjaWithSoftUpperBound(w_max=args.w_max,
gamma_pos=args.gamma_pos,
gamma_neg=args.gamma_neg))
# Load checkpoint
if args.checkpoint_path:
print("=> loading checkpoint '{}'".format(args.checkpoint_path))
checkpoint = load_checkpoint(args.checkpoint_path, device)
best_acc = checkpoint['best_acc']
print("Best accuracy {}".format(best_acc))
if args.check_params:
for key, val in vars(args).items():
if key not in ['check_params', 'seed', 'dataset_seed', 'checkpoint_path', 'example']:
if vars(checkpoint['params'])[key] != val:
print("=> You tried to load a model that was trained on different parameters as you requested "
"now. You may disable this check by setting `check_params` to 0. Aborting...")
sys.exit()
new_state_dict = OrderedDict()
for k, v in checkpoint['state_dict'].items():
if k.startswith('module.'):
k = k[len('module.'):] # remove `module.`
new_state_dict[k] = v
model.load_state_dict(new_state_dict)
# Switch to evaluate mode
model.eval()
# Get dataset example and run the model
sample, sequence_length = dataset[0]
features = sample['features'].unsqueeze(0)
labels = sample['labels'].unsqueeze(0)
query = sample['query'].unsqueeze(0)
outputs, encoding_outputs, writing_outputs, reading_outputs = model(features, labels, query)
# Get the outputs
features_encoded = encoding_outputs[0].detach().numpy()
labels_encoded = encoding_outputs[1].detach().numpy()
query_encoded = encoding_outputs[2].detach().numpy()
mem = writing_outputs[0].detach().numpy()
write_key = writing_outputs[1].detach().numpy()
write_val = writing_outputs[2].detach().numpy()
read_key = reading_outputs[0].detach().numpy()
read_val = reading_outputs[1].detach().numpy()
outputs = outputs.detach().numpy()
# Make some plots
fig, ax = plt.subplots(nrows=3, ncols=1, sharex='all')
ax[0].pcolormesh(features[0].T, cmap='binary')
ax[0].set_ylabel('features')
ax[1].pcolormesh(labels[0, np.newaxis], cmap='binary')
ax[1].set_ylabel('labels')
ax[2].pcolormesh(query.T, cmap='binary')
ax[2].set_ylabel('query')
plt.tight_layout()
fig, ax = plt.subplots(nrows=3, ncols=1, sharex='col')
ax[0].pcolormesh(features_encoded[0].T, cmap='binary')
ax[0].set_ylabel('features')
ax[1].pcolormesh(labels_encoded[0].T, cmap='binary')
ax[1].set_ylabel('labels')
ax[2].pcolormesh(query_encoded[0].T, cmap='binary')
ax[2].set_ylabel('query')
plt.tight_layout()
fig, ax = plt.subplots(nrows=2, ncols=1, sharex='col')
ax[0].pcolormesh(write_key[0].T, cmap='binary')
ax[0].set_ylabel('write keys')
ax[1].pcolormesh(write_val[0].T, cmap='binary')
ax[1].set_ylabel('write values')
plt.tight_layout()
fig, ax = plt.subplots(nrows=1, ncols=1, sharex='all')
ax.matshow(mem[0], cmap='RdBu')
plt.tight_layout()
fig, ax = plt.subplots(nrows=2, ncols=1, sharex='col')
ax[0].pcolormesh(read_key[0].T, cmap='binary')
ax[0].set_ylabel('read keys')
ax[1].pcolormesh(read_val[0].T, cmap='binary')
ax[1].set_ylabel('read values')
plt.tight_layout()
fig, ax = plt.subplots(nrows=1, ncols=1, sharex='all')
ax.pcolormesh(outputs[:, None], cmap='binary')
plt.tight_layout()
plt.show()
if __name__ == '__main__':
main()