Hugging Face's logo

HyperbeeAI
/
vision-bnn-benchmarks-hf

Image Classification
vision
Model card Files Community
vision-bnn-benchmarks-hf
File size: 4,801 Bytes 
5096607
 
4e45d68
5096607
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
###########################################################################
# Computer vision - Binary neural networks demo software by HyperbeeAI.   #
# Copyrights © 2023 Hyperbee.AI Inc. All rights reserved. hello@hyperbee.ai #
###########################################################################
import torch, sys, time
import torch.nn as nn
import torch.optim as optim

# bizden
import layers, models, dataloader
from library.utils import compute_batch_accuracy, compute_set_accuracy

bs = 100;
train_loader, test_loader = dataloader.load_cifar100(batch_size=bs, num_workers=6, shuffle=False, act_8b_mode=False);

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model  = models.maxim_nas()
model  = model.to(device) 

# first, ftp2qat but it's Ap
sq = 0.985

layer_attribute = getattr(model, 'conv1_1')
layer_attribute.configure_layer_base(weight_bits=8, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat');
setattr(model, 'conv1_1', layer_attribute)

layer_attribute = getattr(model, 'conv1_2')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat');
setattr(model, 'conv1_2', layer_attribute)

layer_attribute = getattr(model, 'conv1_3')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat');
setattr(model, 'conv1_3', layer_attribute)

layer_attribute = getattr(model, 'conv2_1')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat_ap');
setattr(model, 'conv2_1', layer_attribute)

layer_attribute = getattr(model, 'conv2_2')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat_ap');
setattr(model, 'conv2_2', layer_attribute)

layer_attribute = getattr(model, 'conv3_1')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat_ap');
setattr(model, 'conv3_1', layer_attribute)

layer_attribute = getattr(model, 'conv3_2')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat_ap');
setattr(model, 'conv3_2', layer_attribute)

layer_attribute = getattr(model, 'conv4_1')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat_ap');
setattr(model, 'conv4_1', layer_attribute)

layer_attribute = getattr(model, 'conv4_2')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat');
setattr(model, 'conv4_2', layer_attribute)

layer_attribute = getattr(model, 'conv5_1')
layer_attribute.configure_layer_base(weight_bits=2, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat');
setattr(model, 'conv5_1', layer_attribute)

layer_attribute = getattr(model, 'fc')
layer_attribute.configure_layer_base(weight_bits=8, bias_bits=8, shift_quantile=sq)
layer_attribute.mode_fpt2qat('qat');
setattr(model, 'fc', layer_attribute)

model.to(device) # somehow new parameters are left out, so they need a reload to the GPU
#########################################################################################

# then, load trained checkpoint
checkpoint = torch.load('training_checkpoint.pth.tar');
model.load_state_dict(checkpoint['state_dict'])

print('')
print('Computing test set accuracy, training checkpoint')
test_acc = compute_set_accuracy(model, test_loader)

print('')
print('Test accuracy:', test_acc*100.0)
print('')

train_loader, test_loader = dataloader.load_cifar100(batch_size=bs, num_workers=6, shuffle=False, act_8b_mode=True);

# then, qat2hw
model  = model.to(device) 
for layer_string in dir(model):
    layer_attribute = getattr(model, layer_string)
    if isinstance(layer_attribute, layers.shallow_base_layer):
        print('Generating HW parameters for:', layer_string)
        if(layer_attribute.mode == 'qat'):
            layer_attribute.mode_qat2hw('eval');
        elif(layer_attribute.mode == 'qat_ap'):
            layer_attribute.mode_qat_ap2hw('eval');
        setattr(model, layer_string, layer_attribute)
model.to(device) # somehow new parameters are left out, so they need a reload

print('')
print('Computing test set accuracy, hardware checkpoint')
test_acc = compute_set_accuracy(model, test_loader)

torch.save({
            'epoch': 123456789,
            'extras': {'best epoch':123456789, 'best_top1':100*test_acc.cpu().numpy(), 'clipping_method':'MAX_BIT_SHIFT', 'current_top1':100*test_acc.cpu().numpy()},
            'state_dict': model.state_dict(),
            'arch': 'ai85nascifarnet'
            }, 'hardware_checkpoint.pth.tar')

print('')
print('Test accuracy:', test_acc*100.0)