├── utils.py
├── .gitignore
├── README.md
├── layers.py
├── mnist.py
├── example.svg
└── example.ipynb


/utils.py:
--------------------------------------------------------------------------------
 1 | import cmath
 2 | 
 3 | 
 4 | def solve_quadratic(a, b, c):
 5 |     delta = (b**2) - (4*a*c)
 6 |     solution1 = (-b-cmath.sqrt(delta))/(2*a)
 7 |     solution2 = (-b+cmath.sqrt(delta))/(2*a)
 8 | 
 9 |     return max(solution1.real, solution2.real)
10 | 
11 | 
12 | def get_equivalent_compression(input_dim, output_dim, nhu, nhLayers, compression):
13 |     '''
14 |     Attempts to find a suitable hidden layer dimension
15 |     to match the number of parameters in a model compressed
16 |     with HashedNets.
17 |     '''
18 |     if nhLayers == 1 or compression == 1.0:
19 |         return compression
20 | 
21 |     # Number of compressed parameters for the HashedNet
22 |     # Assumes we hash all biases apart from the output
23 |     N = input_dim * nhu + (nhLayers - 1) * nhu**2 + nhu * output_dim
24 |     biases = nhu + nhu * (nhLayers - 1)
25 |     compressed_N = N * compression + biases * compression + output_dim
26 | 
27 |     # Solve for compression rate (nhu * compress)
28 |     # inp*nhu*x + nhu*x + layers*nhu*x*nhu*x + layers*nhu*x + oup*nhu*x + oup*x
29 |     # (inp*nhu + nhu + layers*nhu + oup*nhu + oup)*x + layers*nhu*x*nhu*x
30 |     a = (nhLayers - 1) * nhu**2
31 |     b = nhu * (input_dim + 1 + (nhLayers - 1) + output_dim) + output_dim
32 |     c = -compressed_N
33 | 
34 |     equiv_compression = solve_quadratic(a, b, c)
35 | 
36 |     c_nhu = nhu * equiv_compression
37 |     equiv_N = (input_dim * c_nhu + c_nhu
38 |                + (nhLayers -1 ) * c_nhu**2
39 |                + (nhLayers - 1) * c_nhu
40 |                + c_nhu * output_dim + output_dim)
41 |     assert abs(equiv_N - compressed_N) < 10, 'Equiv: {} vs. compressed {}'.format(equiv_N, compressed_N)
42 | 
43 |     return equiv_compression
44 | 
45 | 


--------------------------------------------------------------------------------
/.gitignore:
--------------------------------------------------------------------------------
  1 | # Byte-compiled / optimized / DLL files
  2 | __pycache__/
  3 | *.py[cod]
  4 | *$py.class
  5 | 
  6 | # C extensions
  7 | *.so
  8 | 
  9 | # Distribution / packaging
 10 | .Python
 11 | build/
 12 | develop-eggs/
 13 | dist/
 14 | downloads/
 15 | eggs/
 16 | .eggs/
 17 | lib/
 18 | lib64/
 19 | parts/
 20 | sdist/
 21 | var/
 22 | wheels/
 23 | *.egg-info/
 24 | .installed.cfg
 25 | *.egg
 26 | MANIFEST
 27 | 
 28 | # PyInstaller
 29 | #  Usually these files are written by a python script from a template
 30 | #  before PyInstaller builds the exe, so as to inject date/other infos into it.
 31 | *.manifest
 32 | *.spec
 33 | 
 34 | # Installer logs
 35 | pip-log.txt
 36 | pip-delete-this-directory.txt
 37 | 
 38 | # Unit test / coverage reports
 39 | htmlcov/
 40 | .tox/
 41 | .coverage
 42 | .coverage.*
 43 | .cache
 44 | nosetests.xml
 45 | coverage.xml
 46 | *.cover
 47 | .hypothesis/
 48 | .pytest_cache/
 49 | 
 50 | # Translations
 51 | *.mo
 52 | *.pot
 53 | 
 54 | # Django stuff:
 55 | *.log
 56 | local_settings.py
 57 | db.sqlite3
 58 | 
 59 | # Flask stuff:
 60 | instance/
 61 | .webassets-cache
 62 | 
 63 | # Scrapy stuff:
 64 | .scrapy
 65 | 
 66 | # Sphinx documentation
 67 | docs/_build/
 68 | 
 69 | # PyBuilder
 70 | target/
 71 | 
 72 | # Jupyter Notebook
 73 | .ipynb_checkpoints
 74 | 
 75 | # pyenv
 76 | .python-version
 77 | 
 78 | # celery beat schedule file
 79 | celerybeat-schedule
 80 | 
 81 | # SageMath parsed files
 82 | *.sage.py
 83 | 
 84 | # Environments
 85 | .env
 86 | .venv
 87 | env/
 88 | venv/
 89 | ENV/
 90 | env.bak/
 91 | venv.bak/
 92 | 
 93 | # Spyder project settings
 94 | .spyderproject
 95 | .spyproject
 96 | 
 97 | # Rope project settings
 98 | .ropeproject
 99 | 
100 | # mkdocs documentation
101 | /site
102 | 
103 | # mypy
104 | .mypy_cache/
105 | 


--------------------------------------------------------------------------------
/README.md:
--------------------------------------------------------------------------------
 1 | # HashedNets
 2 | This is a PyTorch implementation of [HashedNets](http://www.jmlr.org/proceedings/papers/v37/chenc15.pdf) by [Chen et al. (2015)](#References). The original authors  have published a (Lua)Torch implementation [here](https://www.cse.wustl.edu/~ychen/HashedNets/).
 3 | 
 4 | HashedNets implements parameter sharing by tying weights that collide in the same buckets given a hash function. The output dimensions of each layer remains the same, while the number of unique values in the weight matrices drop with the compression factor. The aim is to maintain performance as the model is compressed by making use of redundancy in the parameters. The benefit of using a hash function is that we only require the hash seed in order to record which parameters are tied, saving space.
 5 | 
 6 | ## Example results on MNIST
 7 | This is an approximate reproduction of a figure from the original paper (produced in `example.ipynb`). Note that the results will be slightly different from the original paper given the different implementations, and the fact that I have not run Bayesian hyperparameter optimisation which they do in the paper or averaged over multiple runs.
 8 | 
 9 | ![Example results](./example.svg)
10 | 
11 | ## Dependencies
12 | - [PyTorch](https://pytorch.org)
13 | - [xxhash](https://pypi.org/project/xxhash/)
14 | - [TorchVision](https://pytorch.org) (to run the MNIST example)
15 | 
16 | Tested with Python 3.6.8, PyTorch 1.0.1.post2, xxhash 1.3.0.
17 | 
18 | ## Usage
19 | The main component of interest is probably the `HashLinear` layer in `layers.py`. See `mnist.py` for an example model using the layer.
20 | 
21 | To see possible arguments to the script, run:
22 | ```sh
23 | python3 mnist.py --help
24 | ```
25 | 
26 | To run the MNIST example with default hyperparameters, with or without hashing:
27 | ```sh
28 | python3 mnist.py --compress 0.015625  # No hashing -> Test accuracy: 94.09%
29 | python3 mnist.py --compress 0.015625 --hashed  # With hashing -> Test accuracy: 96.83%
30 | ```
31 | 
32 | ## References
33 | - Chen, W., Wilson, J., Tyree, S., Weinberger, K. and Chen, Y., 2015, June. [Compressing neural networks with the hashing trick](http://www.jmlr.org/proceedings/papers/v37/chenc15.pdf). In International Conference on Machine Learning (pp. 2285-2294).
34 | 


--------------------------------------------------------------------------------
/layers.py:
--------------------------------------------------------------------------------
  1 | import torch
  2 | import torch.nn as nn
  3 | from torch.nn import Parameter
  4 | from torch.nn import init
  5 | import torch.nn.functional as F
  6 | import xxhash
  7 | import math
  8 | 
  9 | 
 10 | class HashLinear(nn.Module):
 11 |     '''
 12 |     This layer implements a linear hashed network as in
 13 |      - Chen, W., Wilson, J., Tyree, S., Weinberger, K. and Chen, Y., 2015,
 14 |        Compressing neural networks with the hashing trick.
 15 |        In International Conference on Machine Learning (pp. 2285-2294).
 16 | 
 17 |     It is largely based on the above authors (Lua)Torch implementation:
 18 |     https://www.cse.wustl.edu/~ychen/HashedNets
 19 | 
 20 |     Note that some static hashing parameters are wrapped with
 21 |     `Parameter(..., requires_grad=False)` so that they get sent
 22 |     to device along with the layer. I.e. check for requires_grad
 23 |     when computing total number of parameters.
 24 |     '''
 25 |     def __init__(self, in_features, out_features, compression,
 26 |                  xi=True, hash_bias=True, bias=True, hash_seed=2):
 27 |         super(HashLinear, self).__init__()
 28 | 
 29 |         self.hash_seed = hash_seed
 30 |         self.use_bias = bias
 31 |         self.hash_bias = hash_bias
 32 |         self.xi = xi
 33 | 
 34 |         self.in_features = in_features
 35 |         self.out_features = out_features
 36 | 
 37 |         #  Virtual sizes
 38 |         self.size_w = in_features * out_features
 39 |         self.size_b = out_features
 40 | 
 41 |         #  Compressed sizes
 42 |         self.hsize_w = math.ceil(self.size_w * compression)
 43 | 
 44 |         if self.hash_bias:
 45 |             self.hsize_b = math.ceil(self.size_b * compression)
 46 |         else:
 47 |             self.hsize_b = self.size_b
 48 | 
 49 |         self.h_weight = Parameter(torch.Tensor(self.hsize_w))
 50 |         if bias:
 51 |             self.h_bias = Parameter(torch.Tensor(self.hsize_b))
 52 | 
 53 |         self.xxhash = xxhash
 54 | 
 55 |         self.hash_config('W')
 56 |         if bias and self.hash_bias:
 57 |             self.hash_config('B')
 58 | 
 59 |         self.reset_parameters()
 60 | 
 61 |     def hash_config(self, WorB):
 62 |         '''
 63 |         Returns virtual matrices with indices into the compressed
 64 |         size given by the hashing function.
 65 |         '''
 66 |         assert WorB == 'W' or WorB == 'B'
 67 | 
 68 |         if WorB == 'W':
 69 |             h_size = self.hsize_w
 70 |             dim1 = self.out_features
 71 |             dim2 = self.in_features
 72 |             self.idxW = self.hash_func(h_size, dim1, dim2, 'idxW')
 73 |         elif WorB == 'B':
 74 |             h_size = self.hsize_b
 75 |             dim1 = self.out_features
 76 |             dim2 = 1
 77 |             self.idxB = self.hash_func(h_size, dim1, dim2, 'idxB').squeeze()
 78 | 
 79 |         if self.xi:
 80 |             # Returns 1 and -1
 81 |             if WorB == 'W':
 82 |                 self.xiW = Parameter(self.hash_func(2, dim1, dim2,
 83 |                                           'xiW').add(1).mul(2).add(-3).float(),
 84 |                                      requires_grad=False)
 85 |             elif WorB == 'B':
 86 |                 self.xiB = Parameter(self.hash_func(2, dim1, dim2,
 87 |                                           'xiB').add(1).mul(2).add(-3).float().squeeze(),
 88 |                                      requires_grad=False)
 89 | 
 90 |     def hash_func(self, hN, size_out, size_in, extra_str=''):
 91 |         '''
 92 |         Hash matrix indices to an index in the compressed vector
 93 |         representation.
 94 | 
 95 |         Returns a matrix of indices with size size_out x size_in,
 96 |         where the indices are in the range [0,hN).
 97 |         '''
 98 |         idx = torch.LongTensor(size_out, size_in)
 99 |         for i in range(size_out):
100 |             for j in range(size_in):
101 |                 key = '{}_{}{}'.format(i, j, extra_str)
102 | 
103 |                 # Wrap hashed values to the compressed range
104 |                 idx[i, j] = self.xxhash.xxh32(key, self.hash_seed).intdigest() % hN
105 | 
106 |         return idx
107 | 
108 |     def reset_parameters(self):
109 |         stdv = 1. / math.sqrt(self.in_features)
110 |         init.uniform_(self.h_weight, -stdv, stdv)
111 |         if self.use_bias:
112 |             init.uniform_(self.h_bias, -stdv, stdv)
113 | 
114 |     def forward(self, x):
115 |         # self.idxW is a matrix of the full size of type LongTensor,
116 |         # which contains indices that selects from the elements in h_weight
117 |         if self.use_bias:
118 |             if self.hash_bias:
119 |                 if self.xi:
120 |                     return F.linear(x, self.h_weight[self.idxW] * self.xiW, self.h_bias[self.idxB] * self.xiB)
121 |                 else:
122 |                     return F.linear(x, self.h_weight[self.idxW], self.h_bias[self.idxB])
123 |             else:
124 |                 if self.xi:
125 |                     return F.linear(x, self.h_weight[self.idxW] * self.xiW, self.h_bias)
126 |                 else:
127 |                     return F.linear(x, self.h_weight[self.idxW], self.h_bias)
128 |         else:
129 |             if self.xi:
130 |                 return F.linear(x, self.h_weight[self.idxW] * self.xiW, None)
131 |             else:
132 |                 return F.linear(x, self.h_weight[self.idxW], None)
133 | 


--------------------------------------------------------------------------------
/mnist.py:
--------------------------------------------------------------------------------
  1 | import argparse
  2 | import random
  3 | import math
  4 | 
  5 | import torch
  6 | import torch.nn as nn
  7 | import torch.nn.functional as F
  8 | import torch.optim as optim
  9 | from torchvision import datasets, transforms
 10 | from torch.utils.data.sampler import SubsetRandomSampler
 11 | 
 12 | from layers import HashLinear
 13 | from utils import get_equivalent_compression
 14 | 
 15 | 
 16 | def parse_arguments():
 17 |     parser = argparse.ArgumentParser(description='PyTorch HashedNets',
 18 |                                      formatter_class=argparse.ArgumentDefaultsHelpFormatter)
 19 |     parser.add_argument('--seed', type=int, default=1,
 20 |                         help='random seed')
 21 |     parser.add_argument('--nhLayers', type=int, default=1,
 22 |                         help='# hidden layers, excluding input/output layers')
 23 |     parser.add_argument('--nhu', type=int, default=1000,
 24 |                         help='Number of hidden units')
 25 |     parser.add_argument('--hashed', default=False, action='store_true',
 26 |                         help='Enable hashing')
 27 |     parser.add_argument('--compress', type=float, default=0.03125,
 28 |                         help='Compression rate')
 29 |     parser.add_argument('--hash-bias', default=False, action='store_true',
 30 |                         help='Hash bias terms')
 31 |     parser.add_argument('--lr', type=float, default=0.01,
 32 |                         help='Learning rate at t=0')
 33 |     parser.add_argument('--decay-factor', type=float, default=0.1,
 34 |                         help='Learning rate decay factor')
 35 |     parser.add_argument('--batch-size', type=int, default=50,
 36 |                         help='Mini-batch size (1 = pure stochastic')
 37 |     parser.add_argument('--validation-percent', type=float, default=0.1,
 38 |                         help='Percent of training data used for validation')
 39 |     parser.add_argument('--momentum', type=float, default=0.9,
 40 |                         help='Momentum (SGD only)')
 41 |     parser.add_argument('--dropout', type=float, default=0.25,
 42 |                         help='Dropout rate')
 43 |     parser.add_argument('--l2reg', type=float, default=0.0,
 44 |                         help='l2 regularisation')
 45 |     parser.add_argument('--epochs', type=int, default=50,
 46 |                         help='Maximum # of epochs')
 47 |     parser.add_argument('--patience', type=int, default=2,
 48 |                         help='Number of epochs to wait before scaling lr.')
 49 |     parser.add_argument('--no-xi', default=True, action='store_false',
 50 |                         help='Do not use auxiliary hash (sign factor)')
 51 |     parser.add_argument('--hash-seed', type=int, default=2,
 52 |                         help='Seed for hash functions')
 53 |     parser.add_argument('--no-cuda', action='store_true', default=False,
 54 |                         help='disables CUDA training')
 55 |     parser.add_argument('--save-model', action='store_true', default=False,
 56 |                         help='For Saving the current Model')
 57 |     args = parser.parse_args()
 58 | 
 59 |     print(args)
 60 |     return args
 61 | 
 62 | 
 63 | def load_data(batch_size, kwargs):
 64 |     '''
 65 |     Load MNIST data. Largely from PyTorch MNIST example.
 66 |     '''
 67 |     train_dataset = datasets.MNIST('../data',
 68 |                                    train=True, download=True,
 69 |                                    transform=transforms.Compose([
 70 |                                        transforms.ToTensor(),
 71 |                                        transforms.Normalize((0.1307,), (0.3081,))
 72 |                                    ]))
 73 | 
 74 |     num_train = len(train_dataset)
 75 |     indices = list(range(num_train))
 76 |     random.shuffle(indices)
 77 |     validation_percent = 0.1
 78 |     split = int(math.floor(validation_percent * num_train))
 79 |     train_idx, valid_idx = indices[split:], indices[:split]
 80 | 
 81 |     train_sampler = SubsetRandomSampler(train_idx)
 82 |     valid_sampler = SubsetRandomSampler(valid_idx)
 83 | 
 84 |     train_loader = torch.utils.data.DataLoader(
 85 |         train_dataset,
 86 |         batch_size=batch_size, sampler=train_sampler, **kwargs)
 87 | 
 88 |     valid_loader = torch.utils.data.DataLoader(
 89 |         train_dataset,
 90 |         batch_size=batch_size, sampler=valid_sampler, **kwargs)
 91 | 
 92 |     test_loader = torch.utils.data.DataLoader(
 93 |         datasets.MNIST('../data', train=False,
 94 |                        transform=transforms.Compose([
 95 |                            transforms.ToTensor(),
 96 |                            transforms.Normalize((0.1307,), (0.3081,))
 97 |                        ])),
 98 |         batch_size=batch_size, shuffle=True, **kwargs)
 99 | 
100 |     return train_loader, valid_loader, test_loader
101 | 
102 | 
103 | def train(model, device, train_loader, optimizer, epoch, log_interval=5):
104 |     '''
105 |     One epoch of training.
106 |     '''
107 |     model.train()
108 |     train_loss = 0.0
109 |     for batch_idx, (data, target) in enumerate(train_loader):
110 |         data, target = data.to(device), target.to(device)
111 |         optimizer.zero_grad()
112 |         output = model(data)
113 |         loss = F.nll_loss(output, target)
114 |         loss.backward()
115 |         optimizer.step()
116 |         if batch_idx % log_interval == 0:
117 |             print('Train Epoch: {} [{}/{} ({:.2f}%)]\tLoss: {:.6f}'.format(
118 |                 epoch, batch_idx * len(data), len(train_loader.sampler),
119 |                 100. * batch_idx / len(train_loader), loss.item()), end='\r')
120 |         train_loss += loss.item()
121 | 
122 |     return train_loss / len(train_loader)
123 | 
124 | 
125 | def evaluate(model, device, loader):
126 |     model.eval()
127 |     loss = 0
128 |     correct = 0
129 |     with torch.no_grad():
130 |         for data, target in loader:
131 |             data, target = data.to(device), target.to(device)
132 |             output = model(data)
133 |             loss += F.nll_loss(output, target, reduction='sum').item()
134 |             pred = output.argmax(dim=1, keepdim=True)
135 |             correct += pred.eq(target.view_as(pred)).sum().item()
136 | 
137 |     loss /= len(loader.sampler)
138 |     accuracy = 100. * correct / len(loader.sampler)
139 | 
140 |     return loss, accuracy
141 | 
142 | 
143 | class Net(nn.Module):
144 |     '''
145 |     Standard feedforward network with ReLU activations
146 |     and optional interleaving dropout layers.
147 |     '''
148 |     def __init__(self, input_dim, output_dim, nhLayers=1, nhu=1000,
149 |                  compress=1.0, dropout=0.25):
150 |         super(Net, self).__init__()
151 |         self.nhLayers = nhLayers
152 |         self.input_dim = input_dim
153 | 
154 |         c_nhu = round(nhu * compress)
155 | 
156 |         self.dropout0 = nn.Dropout(dropout)
157 |         self.linear1 = nn.Linear(input_dim, c_nhu)
158 |         self.dropout1 = nn.Dropout(dropout)
159 | 
160 |         for layer in range(2, nhLayers + 1):
161 |             setattr(self, 'linear' + str(layer), nn.Linear(c_nhu, c_nhu))
162 |             setattr(self, 'dropout' + str(layer), nn.Dropout(dropout))
163 | 
164 |         self.linear_out = nn.Linear(c_nhu, output_dim)
165 | 
166 |     def forward(self, x):
167 |         x = x.reshape(-1, self.input_dim)
168 |         x = self.dropout0(x)
169 |         x = F.relu(self.linear1(x))
170 |         x = self.dropout1(x)
171 | 
172 |         for layer in range(2, self.nhLayers + 1):
173 |             x = F.relu(getattr(self, 'linear' + str(layer))(x))
174 |             x = getattr(self, 'dropout' + str(layer))(x)
175 | 
176 |         x = self.linear_out(x)
177 |         return F.log_softmax(x, dim=1)
178 | 
179 | 
180 | class HashedNet(nn.Module):
181 |     '''
182 |     Feedforward network with hashed linear layers,
183 |     ReLU activations and optional interleaving dropout layers.
184 |     '''
185 |     def __init__(self, input_dim, output_dim, nhLayers=1, nhu=1000,
186 |                  compress=1.0, dropout=0.25, hash_seed=2):
187 |         super(HashedNet, self).__init__()
188 |         self.nhLayers = nhLayers
189 |         self.input_dim = input_dim
190 | 
191 |         self.dropout0 = nn.Dropout(dropout)
192 |         self.linear1 = HashLinear(input_dim, nhu, compress)
193 |         self.dropout1 = nn.Dropout(dropout)
194 | 
195 |         for layer in range(2, nhLayers + 1):
196 |             setattr(self, 'linear' + str(layer), HashLinear(nhu, nhu, compress,
197 |                                                             hash_seed))
198 |             setattr(self, 'dropout' + str(layer), nn.Dropout(dropout))
199 | 
200 |         self.linear_out = HashLinear(nhu, output_dim, compress,
201 |                                      hash_bias=False, hash_seed=hash_seed)
202 | 
203 |     def forward(self, x):
204 |         x = x.reshape(-1, self.input_dim)
205 |         x = self.dropout0(x)
206 |         x = F.relu(self.linear1(x))
207 |         x = self.dropout1(x)
208 | 
209 |         for layer in range(2, self.nhLayers + 1):
210 |             x = F.relu(getattr(self, 'linear' + str(layer))(x))
211 |             x = getattr(self, 'dropout' + str(layer))(x)
212 | 
213 |         x = self.linear_out(x)
214 |         return F.log_softmax(x, dim=1)
215 | 
216 | 
217 | def main():
218 |     args = parse_arguments()
219 | 
220 |     use_cuda = torch.cuda.is_available()
221 |     torch.manual_seed(1)
222 |     random.seed(1)
223 | 
224 |     device = torch.device("cuda" if use_cuda else "cpu")
225 |     kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
226 | 
227 |     tr_loader, val_loader, test_loader = load_data(args.batch_size, kwargs)
228 |     input_dim = 784
229 |     output_dim = 10
230 | 
231 |     if args.hashed:
232 |         model = HashedNet(input_dim, output_dim, args.nhLayers, args.nhu,
233 |                           args.compress, args.dropout, args.hash_seed).to(device)
234 |     else:
235 |         eq_compress = get_equivalent_compression(input_dim, output_dim,
236 |                                                  args.nhu, args.nhLayers, args.compress)
237 |         model = Net(input_dim, output_dim, args.nhLayers, args.nhu,
238 |                     eq_compress, args.dropout).to(device)
239 | 
240 |     optimizer = optim.SGD(model.parameters(), lr=args.lr,
241 |                           momentum=args.momentum,
242 |                           weight_decay=args.l2reg)
243 | 
244 |     scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,
245 |                                                      factor=args.decay_factor,
246 |                                                      patience=args.patience,
247 |                                                      verbose=True)
248 | 
249 |     print('The number of parameters is: {}'.format(
250 |         sum(p.numel() for p in model.parameters() if p.requires_grad)))
251 | 
252 |     for epoch in range(1, args.epochs + 1):
253 |         tr_loss = train(model, device, tr_loader, optimizer, epoch)
254 |         val_loss, val_acc = evaluate(model, device, val_loader)
255 |         scheduler.step(val_loss)
256 |         print('Epoch {} Train loss: {:.3f} Val loss: {:.3f} Val acc: {:.2f}%'.format(
257 |               epoch, tr_loss, val_loss, val_acc))
258 | 
259 |     test_loss, test_acc = evaluate(model, device, test_loader)
260 |     print('Test loss: {:.3f} Test acc: {:.2f}%'.format(test_loss, test_acc))
261 | 
262 |     if (args.save_model):
263 |         torch.save(model.state_dict(), "mnist.pt")
264 | 
265 | 
266 | if __name__ == '__main__':
267 |     main()
268 | 


--------------------------------------------------------------------------------
/example.svg:
--------------------------------------------------------------------------------
   1 | <?xml version="1.0" encoding="utf-8" standalone="no"?>
   2 | <!DOCTYPE svg PUBLIC "-//W3C//DTD SVG 1.1//EN"
   3 |   "http://www.w3.org/Graphics/SVG/1.1/DTD/svg11.dtd">
   4 | <!-- Created with matplotlib (https://matplotlib.org/) -->
   5 | <svg height="288pt" version="1.1" viewBox="0 0 432 288" width="432pt" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink">
   6 |  <defs>
   7 |   <style type="text/css">
   8 | *{stroke-linecap:butt;stroke-linejoin:round;}
   9 |   </style>
  10 |  </defs>
  11 |  <g id="figure_1">
  12 |   <g id="patch_1">
  13 |    <path d="M 0 288 
  14 | L 432 288 
  15 | L 432 0 
  16 | L 0 0 
  17 | z
  18 | " style="fill:#ffffff;"/>
  19 |   </g>
  20 |   <g id="axes_1">
  21 |    <g id="patch_2">
  22 |     <path d="M 54 252 
  23 | L 388.8 252 
  24 | L 388.8 34.56 
  25 | L 54 34.56 
  26 | z
  27 | " style="fill:#ffffff;"/>
  28 |    </g>
  29 |    <g id="matplotlib.axis_1">
  30 |     <g id="xtick_1">
  31 |      <g id="line2d_1">
  32 |       <path clip-path="url(#pe986d3e2bd)" d="M 69.218182 252 
  33 | L 69.218182 34.56 
  34 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
  35 |      </g>
  36 |      <g id="line2d_2">
  37 |       <defs>
  38 |        <path d="M 0 0 
  39 | L 0 3.5 
  40 | " id="m1bf0a3ebac" style="stroke:#000000;stroke-width:0.8;"/>
  41 |       </defs>
  42 |       <g>
  43 |        <use style="stroke:#000000;stroke-width:0.8;" x="69.218182" xlink:href="#m1bf0a3ebac" y="252"/>
  44 |       </g>
  45 |      </g>
  46 |      <g id="text_1">
  47 |       <!-- 1/64 -->
  48 |       <defs>
  49 |        <path d="M 12.40625 8.296875 
  50 | L 28.515625 8.296875 
  51 | L 28.515625 63.921875 
  52 | L 10.984375 60.40625 
  53 | L 10.984375 69.390625 
  54 | L 28.421875 72.90625 
  55 | L 38.28125 72.90625 
  56 | L 38.28125 8.296875 
  57 | L 54.390625 8.296875 
  58 | L 54.390625 0 
  59 | L 12.40625 0 
  60 | z
  61 | " id="DejaVuSans-49"/>
  62 |        <path d="M 25.390625 72.90625 
  63 | L 33.6875 72.90625 
  64 | L 8.296875 -9.28125 
  65 | L 0 -9.28125 
  66 | z
  67 | " id="DejaVuSans-47"/>
  68 |        <path d="M 33.015625 40.375 
  69 | Q 26.375 40.375 22.484375 35.828125 
  70 | Q 18.609375 31.296875 18.609375 23.390625 
  71 | Q 18.609375 15.53125 22.484375 10.953125 
  72 | Q 26.375 6.390625 33.015625 6.390625 
  73 | Q 39.65625 6.390625 43.53125 10.953125 
  74 | Q 47.40625 15.53125 47.40625 23.390625 
  75 | Q 47.40625 31.296875 43.53125 35.828125 
  76 | Q 39.65625 40.375 33.015625 40.375 
  77 | z
  78 | M 52.59375 71.296875 
  79 | L 52.59375 62.3125 
  80 | Q 48.875 64.0625 45.09375 64.984375 
  81 | Q 41.3125 65.921875 37.59375 65.921875 
  82 | Q 27.828125 65.921875 22.671875 59.328125 
  83 | Q 17.53125 52.734375 16.796875 39.40625 
  84 | Q 19.671875 43.65625 24.015625 45.921875 
  85 | Q 28.375 48.1875 33.59375 48.1875 
  86 | Q 44.578125 48.1875 50.953125 41.515625 
  87 | Q 57.328125 34.859375 57.328125 23.390625 
  88 | Q 57.328125 12.15625 50.6875 5.359375 
  89 | Q 44.046875 -1.421875 33.015625 -1.421875 
  90 | Q 20.359375 -1.421875 13.671875 8.265625 
  91 | Q 6.984375 17.96875 6.984375 36.375 
  92 | Q 6.984375 53.65625 15.1875 63.9375 
  93 | Q 23.390625 74.21875 37.203125 74.21875 
  94 | Q 40.921875 74.21875 44.703125 73.484375 
  95 | Q 48.484375 72.75 52.59375 71.296875 
  96 | z
  97 | " id="DejaVuSans-54"/>
  98 |        <path d="M 37.796875 64.3125 
  99 | L 12.890625 25.390625 
 100 | L 37.796875 25.390625 
 101 | z
 102 | M 35.203125 72.90625 
 103 | L 47.609375 72.90625 
 104 | L 47.609375 25.390625 
 105 | L 58.015625 25.390625 
 106 | L 58.015625 17.1875 
 107 | L 47.609375 17.1875 
 108 | L 47.609375 0 
 109 | L 37.796875 0 
 110 | L 37.796875 17.1875 
 111 | L 4.890625 17.1875 
 112 | L 4.890625 26.703125 
 113 | z
 114 | " id="DejaVuSans-52"/>
 115 |       </defs>
 116 |       <g transform="translate(57.990057 266.598437)scale(0.1 -0.1)">
 117 |        <use xlink:href="#DejaVuSans-49"/>
 118 |        <use x="63.623047" xlink:href="#DejaVuSans-47"/>
 119 |        <use x="97.314453" xlink:href="#DejaVuSans-54"/>
 120 |        <use x="160.9375" xlink:href="#DejaVuSans-52"/>
 121 |       </g>
 122 |      </g>
 123 |     </g>
 124 |     <g id="xtick_2">
 125 |      <g id="line2d_3">
 126 |       <path clip-path="url(#pe986d3e2bd)" d="M 145.309091 252 
 127 | L 145.309091 34.56 
 128 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 129 |      </g>
 130 |      <g id="line2d_4">
 131 |       <g>
 132 |        <use style="stroke:#000000;stroke-width:0.8;" x="145.309091" xlink:href="#m1bf0a3ebac" y="252"/>
 133 |       </g>
 134 |      </g>
 135 |      <g id="text_2">
 136 |       <!-- 1/32 -->
 137 |       <defs>
 138 |        <path d="M 40.578125 39.3125 
 139 | Q 47.65625 37.796875 51.625 33 
 140 | Q 55.609375 28.21875 55.609375 21.1875 
 141 | Q 55.609375 10.40625 48.1875 4.484375 
 142 | Q 40.765625 -1.421875 27.09375 -1.421875 
 143 | Q 22.515625 -1.421875 17.65625 -0.515625 
 144 | Q 12.796875 0.390625 7.625 2.203125 
 145 | L 7.625 11.71875 
 146 | Q 11.71875 9.328125 16.59375 8.109375 
 147 | Q 21.484375 6.890625 26.8125 6.890625 
 148 | Q 36.078125 6.890625 40.9375 10.546875 
 149 | Q 45.796875 14.203125 45.796875 21.1875 
 150 | Q 45.796875 27.640625 41.28125 31.265625 
 151 | Q 36.765625 34.90625 28.71875 34.90625 
 152 | L 20.21875 34.90625 
 153 | L 20.21875 43.015625 
 154 | L 29.109375 43.015625 
 155 | Q 36.375 43.015625 40.234375 45.921875 
 156 | Q 44.09375 48.828125 44.09375 54.296875 
 157 | Q 44.09375 59.90625 40.109375 62.90625 
 158 | Q 36.140625 65.921875 28.71875 65.921875 
 159 | Q 24.65625 65.921875 20.015625 65.03125 
 160 | Q 15.375 64.15625 9.8125 62.3125 
 161 | L 9.8125 71.09375 
 162 | Q 15.4375 72.65625 20.34375 73.4375 
 163 | Q 25.25 74.21875 29.59375 74.21875 
 164 | Q 40.828125 74.21875 47.359375 69.109375 
 165 | Q 53.90625 64.015625 53.90625 55.328125 
 166 | Q 53.90625 49.265625 50.4375 45.09375 
 167 | Q 46.96875 40.921875 40.578125 39.3125 
 168 | z
 169 | " id="DejaVuSans-51"/>
 170 |        <path d="M 19.1875 8.296875 
 171 | L 53.609375 8.296875 
 172 | L 53.609375 0 
 173 | L 7.328125 0 
 174 | L 7.328125 8.296875 
 175 | Q 12.9375 14.109375 22.625 23.890625 
 176 | Q 32.328125 33.6875 34.8125 36.53125 
 177 | Q 39.546875 41.84375 41.421875 45.53125 
 178 | Q 43.3125 49.21875 43.3125 52.78125 
 179 | Q 43.3125 58.59375 39.234375 62.25 
 180 | Q 35.15625 65.921875 28.609375 65.921875 
 181 | Q 23.96875 65.921875 18.8125 64.3125 
 182 | Q 13.671875 62.703125 7.8125 59.421875 
 183 | L 7.8125 69.390625 
 184 | Q 13.765625 71.78125 18.9375 73 
 185 | Q 24.125 74.21875 28.421875 74.21875 
 186 | Q 39.75 74.21875 46.484375 68.546875 
 187 | Q 53.21875 62.890625 53.21875 53.421875 
 188 | Q 53.21875 48.921875 51.53125 44.890625 
 189 | Q 49.859375 40.875 45.40625 35.40625 
 190 | Q 44.1875 33.984375 37.640625 27.21875 
 191 | Q 31.109375 20.453125 19.1875 8.296875 
 192 | z
 193 | " id="DejaVuSans-50"/>
 194 |       </defs>
 195 |       <g transform="translate(134.080966 266.598437)scale(0.1 -0.1)">
 196 |        <use xlink:href="#DejaVuSans-49"/>
 197 |        <use x="63.623047" xlink:href="#DejaVuSans-47"/>
 198 |        <use x="97.314453" xlink:href="#DejaVuSans-51"/>
 199 |        <use x="160.9375" xlink:href="#DejaVuSans-50"/>
 200 |       </g>
 201 |      </g>
 202 |     </g>
 203 |     <g id="xtick_3">
 204 |      <g id="line2d_5">
 205 |       <path clip-path="url(#pe986d3e2bd)" d="M 221.4 252 
 206 | L 221.4 34.56 
 207 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 208 |      </g>
 209 |      <g id="line2d_6">
 210 |       <g>
 211 |        <use style="stroke:#000000;stroke-width:0.8;" x="221.4" xlink:href="#m1bf0a3ebac" y="252"/>
 212 |       </g>
 213 |      </g>
 214 |      <g id="text_3">
 215 |       <!-- 1/16 -->
 216 |       <g transform="translate(210.171875 266.598437)scale(0.1 -0.1)">
 217 |        <use xlink:href="#DejaVuSans-49"/>
 218 |        <use x="63.623047" xlink:href="#DejaVuSans-47"/>
 219 |        <use x="97.314453" xlink:href="#DejaVuSans-49"/>
 220 |        <use x="160.9375" xlink:href="#DejaVuSans-54"/>
 221 |       </g>
 222 |      </g>
 223 |     </g>
 224 |     <g id="xtick_4">
 225 |      <g id="line2d_7">
 226 |       <path clip-path="url(#pe986d3e2bd)" d="M 297.490909 252 
 227 | L 297.490909 34.56 
 228 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 229 |      </g>
 230 |      <g id="line2d_8">
 231 |       <g>
 232 |        <use style="stroke:#000000;stroke-width:0.8;" x="297.490909" xlink:href="#m1bf0a3ebac" y="252"/>
 233 |       </g>
 234 |      </g>
 235 |      <g id="text_4">
 236 |       <!-- 1/8 -->
 237 |       <defs>
 238 |        <path d="M 31.78125 34.625 
 239 | Q 24.75 34.625 20.71875 30.859375 
 240 | Q 16.703125 27.09375 16.703125 20.515625 
 241 | Q 16.703125 13.921875 20.71875 10.15625 
 242 | Q 24.75 6.390625 31.78125 6.390625 
 243 | Q 38.8125 6.390625 42.859375 10.171875 
 244 | Q 46.921875 13.96875 46.921875 20.515625 
 245 | Q 46.921875 27.09375 42.890625 30.859375 
 246 | Q 38.875 34.625 31.78125 34.625 
 247 | z
 248 | M 21.921875 38.8125 
 249 | Q 15.578125 40.375 12.03125 44.71875 
 250 | Q 8.5 49.078125 8.5 55.328125 
 251 | Q 8.5 64.0625 14.71875 69.140625 
 252 | Q 20.953125 74.21875 31.78125 74.21875 
 253 | Q 42.671875 74.21875 48.875 69.140625 
 254 | Q 55.078125 64.0625 55.078125 55.328125 
 255 | Q 55.078125 49.078125 51.53125 44.71875 
 256 | Q 48 40.375 41.703125 38.8125 
 257 | Q 48.828125 37.15625 52.796875 32.3125 
 258 | Q 56.78125 27.484375 56.78125 20.515625 
 259 | Q 56.78125 9.90625 50.3125 4.234375 
 260 | Q 43.84375 -1.421875 31.78125 -1.421875 
 261 | Q 19.734375 -1.421875 13.25 4.234375 
 262 | Q 6.78125 9.90625 6.78125 20.515625 
 263 | Q 6.78125 27.484375 10.78125 32.3125 
 264 | Q 14.796875 37.15625 21.921875 38.8125 
 265 | z
 266 | M 18.3125 54.390625 
 267 | Q 18.3125 48.734375 21.84375 45.5625 
 268 | Q 25.390625 42.390625 31.78125 42.390625 
 269 | Q 38.140625 42.390625 41.71875 45.5625 
 270 | Q 45.3125 48.734375 45.3125 54.390625 
 271 | Q 45.3125 60.0625 41.71875 63.234375 
 272 | Q 38.140625 66.40625 31.78125 66.40625 
 273 | Q 25.390625 66.40625 21.84375 63.234375 
 274 | Q 18.3125 60.0625 18.3125 54.390625 
 275 | z
 276 | " id="DejaVuSans-56"/>
 277 |       </defs>
 278 |       <g transform="translate(289.444034 266.598437)scale(0.1 -0.1)">
 279 |        <use xlink:href="#DejaVuSans-49"/>
 280 |        <use x="63.623047" xlink:href="#DejaVuSans-47"/>
 281 |        <use x="97.314453" xlink:href="#DejaVuSans-56"/>
 282 |       </g>
 283 |      </g>
 284 |     </g>
 285 |     <g id="xtick_5">
 286 |      <g id="line2d_9">
 287 |       <path clip-path="url(#pe986d3e2bd)" d="M 373.581818 252 
 288 | L 373.581818 34.56 
 289 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 290 |      </g>
 291 |      <g id="line2d_10">
 292 |       <g>
 293 |        <use style="stroke:#000000;stroke-width:0.8;" x="373.581818" xlink:href="#m1bf0a3ebac" y="252"/>
 294 |       </g>
 295 |      </g>
 296 |      <g id="text_5">
 297 |       <!-- 1 -->
 298 |       <g transform="translate(370.400568 266.598437)scale(0.1 -0.1)">
 299 |        <use xlink:href="#DejaVuSans-49"/>
 300 |       </g>
 301 |      </g>
 302 |     </g>
 303 |     <g id="text_6">
 304 |      <!-- Compression ratio -->
 305 |      <defs>
 306 |       <path d="M 64.40625 67.28125 
 307 | L 64.40625 56.890625 
 308 | Q 59.421875 61.53125 53.78125 63.8125 
 309 | Q 48.140625 66.109375 41.796875 66.109375 
 310 | Q 29.296875 66.109375 22.65625 58.46875 
 311 | Q 16.015625 50.828125 16.015625 36.375 
 312 | Q 16.015625 21.96875 22.65625 14.328125 
 313 | Q 29.296875 6.6875 41.796875 6.6875 
 314 | Q 48.140625 6.6875 53.78125 8.984375 
 315 | Q 59.421875 11.28125 64.40625 15.921875 
 316 | L 64.40625 5.609375 
 317 | Q 59.234375 2.09375 53.4375 0.328125 
 318 | Q 47.65625 -1.421875 41.21875 -1.421875 
 319 | Q 24.65625 -1.421875 15.125 8.703125 
 320 | Q 5.609375 18.84375 5.609375 36.375 
 321 | Q 5.609375 53.953125 15.125 64.078125 
 322 | Q 24.65625 74.21875 41.21875 74.21875 
 323 | Q 47.75 74.21875 53.53125 72.484375 
 324 | Q 59.328125 70.75 64.40625 67.28125 
 325 | z
 326 | " id="DejaVuSans-67"/>
 327 |       <path d="M 30.609375 48.390625 
 328 | Q 23.390625 48.390625 19.1875 42.75 
 329 | Q 14.984375 37.109375 14.984375 27.296875 
 330 | Q 14.984375 17.484375 19.15625 11.84375 
 331 | Q 23.34375 6.203125 30.609375 6.203125 
 332 | Q 37.796875 6.203125 41.984375 11.859375 
 333 | Q 46.1875 17.53125 46.1875 27.296875 
 334 | Q 46.1875 37.015625 41.984375 42.703125 
 335 | Q 37.796875 48.390625 30.609375 48.390625 
 336 | z
 337 | M 30.609375 56 
 338 | Q 42.328125 56 49.015625 48.375 
 339 | Q 55.71875 40.765625 55.71875 27.296875 
 340 | Q 55.71875 13.875 49.015625 6.21875 
 341 | Q 42.328125 -1.421875 30.609375 -1.421875 
 342 | Q 18.84375 -1.421875 12.171875 6.21875 
 343 | Q 5.515625 13.875 5.515625 27.296875 
 344 | Q 5.515625 40.765625 12.171875 48.375 
 345 | Q 18.84375 56 30.609375 56 
 346 | z
 347 | " id="DejaVuSans-111"/>
 348 |       <path d="M 52 44.1875 
 349 | Q 55.375 50.25 60.0625 53.125 
 350 | Q 64.75 56 71.09375 56 
 351 | Q 79.640625 56 84.28125 50.015625 
 352 | Q 88.921875 44.046875 88.921875 33.015625 
 353 | L 88.921875 0 
 354 | L 79.890625 0 
 355 | L 79.890625 32.71875 
 356 | Q 79.890625 40.578125 77.09375 44.375 
 357 | Q 74.3125 48.1875 68.609375 48.1875 
 358 | Q 61.625 48.1875 57.5625 43.546875 
 359 | Q 53.515625 38.921875 53.515625 30.90625 
 360 | L 53.515625 0 
 361 | L 44.484375 0 
 362 | L 44.484375 32.71875 
 363 | Q 44.484375 40.625 41.703125 44.40625 
 364 | Q 38.921875 48.1875 33.109375 48.1875 
 365 | Q 26.21875 48.1875 22.15625 43.53125 
 366 | Q 18.109375 38.875 18.109375 30.90625 
 367 | L 18.109375 0 
 368 | L 9.078125 0 
 369 | L 9.078125 54.6875 
 370 | L 18.109375 54.6875 
 371 | L 18.109375 46.1875 
 372 | Q 21.1875 51.21875 25.484375 53.609375 
 373 | Q 29.78125 56 35.6875 56 
 374 | Q 41.65625 56 45.828125 52.96875 
 375 | Q 50 49.953125 52 44.1875 
 376 | z
 377 | " id="DejaVuSans-109"/>
 378 |       <path d="M 18.109375 8.203125 
 379 | L 18.109375 -20.796875 
 380 | L 9.078125 -20.796875 
 381 | L 9.078125 54.6875 
 382 | L 18.109375 54.6875 
 383 | L 18.109375 46.390625 
 384 | Q 20.953125 51.265625 25.265625 53.625 
 385 | Q 29.59375 56 35.59375 56 
 386 | Q 45.5625 56 51.78125 48.09375 
 387 | Q 58.015625 40.1875 58.015625 27.296875 
 388 | Q 58.015625 14.40625 51.78125 6.484375 
 389 | Q 45.5625 -1.421875 35.59375 -1.421875 
 390 | Q 29.59375 -1.421875 25.265625 0.953125 
 391 | Q 20.953125 3.328125 18.109375 8.203125 
 392 | z
 393 | M 48.6875 27.296875 
 394 | Q 48.6875 37.203125 44.609375 42.84375 
 395 | Q 40.53125 48.484375 33.40625 48.484375 
 396 | Q 26.265625 48.484375 22.1875 42.84375 
 397 | Q 18.109375 37.203125 18.109375 27.296875 
 398 | Q 18.109375 17.390625 22.1875 11.75 
 399 | Q 26.265625 6.109375 33.40625 6.109375 
 400 | Q 40.53125 6.109375 44.609375 11.75 
 401 | Q 48.6875 17.390625 48.6875 27.296875 
 402 | z
 403 | " id="DejaVuSans-112"/>
 404 |       <path d="M 41.109375 46.296875 
 405 | Q 39.59375 47.171875 37.8125 47.578125 
 406 | Q 36.03125 48 33.890625 48 
 407 | Q 26.265625 48 22.1875 43.046875 
 408 | Q 18.109375 38.09375 18.109375 28.8125 
 409 | L 18.109375 0 
 410 | L 9.078125 0 
 411 | L 9.078125 54.6875 
 412 | L 18.109375 54.6875 
 413 | L 18.109375 46.1875 
 414 | Q 20.953125 51.171875 25.484375 53.578125 
 415 | Q 30.03125 56 36.53125 56 
 416 | Q 37.453125 56 38.578125 55.875 
 417 | Q 39.703125 55.765625 41.0625 55.515625 
 418 | z
 419 | " id="DejaVuSans-114"/>
 420 |       <path d="M 56.203125 29.59375 
 421 | L 56.203125 25.203125 
 422 | L 14.890625 25.203125 
 423 | Q 15.484375 15.921875 20.484375 11.0625 
 424 | Q 25.484375 6.203125 34.421875 6.203125 
 425 | Q 39.59375 6.203125 44.453125 7.46875 
 426 | Q 49.3125 8.734375 54.109375 11.28125 
 427 | L 54.109375 2.78125 
 428 | Q 49.265625 0.734375 44.1875 -0.34375 
 429 | Q 39.109375 -1.421875 33.890625 -1.421875 
 430 | Q 20.796875 -1.421875 13.15625 6.1875 
 431 | Q 5.515625 13.8125 5.515625 26.8125 
 432 | Q 5.515625 40.234375 12.765625 48.109375 
 433 | Q 20.015625 56 32.328125 56 
 434 | Q 43.359375 56 49.78125 48.890625 
 435 | Q 56.203125 41.796875 56.203125 29.59375 
 436 | z
 437 | M 47.21875 32.234375 
 438 | Q 47.125 39.59375 43.09375 43.984375 
 439 | Q 39.0625 48.390625 32.421875 48.390625 
 440 | Q 24.90625 48.390625 20.390625 44.140625 
 441 | Q 15.875 39.890625 15.1875 32.171875 
 442 | z
 443 | " id="DejaVuSans-101"/>
 444 |       <path d="M 44.28125 53.078125 
 445 | L 44.28125 44.578125 
 446 | Q 40.484375 46.53125 36.375 47.5 
 447 | Q 32.28125 48.484375 27.875 48.484375 
 448 | Q 21.1875 48.484375 17.84375 46.4375 
 449 | Q 14.5 44.390625 14.5 40.28125 
 450 | Q 14.5 37.15625 16.890625 35.375 
 451 | Q 19.28125 33.59375 26.515625 31.984375 
 452 | L 29.59375 31.296875 
 453 | Q 39.15625 29.25 43.1875 25.515625 
 454 | Q 47.21875 21.78125 47.21875 15.09375 
 455 | Q 47.21875 7.46875 41.1875 3.015625 
 456 | Q 35.15625 -1.421875 24.609375 -1.421875 
 457 | Q 20.21875 -1.421875 15.453125 -0.5625 
 458 | Q 10.6875 0.296875 5.421875 2 
 459 | L 5.421875 11.28125 
 460 | Q 10.40625 8.6875 15.234375 7.390625 
 461 | Q 20.0625 6.109375 24.8125 6.109375 
 462 | Q 31.15625 6.109375 34.5625 8.28125 
 463 | Q 37.984375 10.453125 37.984375 14.40625 
 464 | Q 37.984375 18.0625 35.515625 20.015625 
 465 | Q 33.0625 21.96875 24.703125 23.78125 
 466 | L 21.578125 24.515625 
 467 | Q 13.234375 26.265625 9.515625 29.90625 
 468 | Q 5.8125 33.546875 5.8125 39.890625 
 469 | Q 5.8125 47.609375 11.28125 51.796875 
 470 | Q 16.75 56 26.8125 56 
 471 | Q 31.78125 56 36.171875 55.265625 
 472 | Q 40.578125 54.546875 44.28125 53.078125 
 473 | z
 474 | " id="DejaVuSans-115"/>
 475 |       <path d="M 9.421875 54.6875 
 476 | L 18.40625 54.6875 
 477 | L 18.40625 0 
 478 | L 9.421875 0 
 479 | z
 480 | M 9.421875 75.984375 
 481 | L 18.40625 75.984375 
 482 | L 18.40625 64.59375 
 483 | L 9.421875 64.59375 
 484 | z
 485 | " id="DejaVuSans-105"/>
 486 |       <path d="M 54.890625 33.015625 
 487 | L 54.890625 0 
 488 | L 45.90625 0 
 489 | L 45.90625 32.71875 
 490 | Q 45.90625 40.484375 42.875 44.328125 
 491 | Q 39.84375 48.1875 33.796875 48.1875 
 492 | Q 26.515625 48.1875 22.3125 43.546875 
 493 | Q 18.109375 38.921875 18.109375 30.90625 
 494 | L 18.109375 0 
 495 | L 9.078125 0 
 496 | L 9.078125 54.6875 
 497 | L 18.109375 54.6875 
 498 | L 18.109375 46.1875 
 499 | Q 21.34375 51.125 25.703125 53.5625 
 500 | Q 30.078125 56 35.796875 56 
 501 | Q 45.21875 56 50.046875 50.171875 
 502 | Q 54.890625 44.34375 54.890625 33.015625 
 503 | z
 504 | " id="DejaVuSans-110"/>
 505 |       <path id="DejaVuSans-32"/>
 506 |       <path d="M 34.28125 27.484375 
 507 | Q 23.390625 27.484375 19.1875 25 
 508 | Q 14.984375 22.515625 14.984375 16.5 
 509 | Q 14.984375 11.71875 18.140625 8.90625 
 510 | Q 21.296875 6.109375 26.703125 6.109375 
 511 | Q 34.1875 6.109375 38.703125 11.40625 
 512 | Q 43.21875 16.703125 43.21875 25.484375 
 513 | L 43.21875 27.484375 
 514 | z
 515 | M 52.203125 31.203125 
 516 | L 52.203125 0 
 517 | L 43.21875 0 
 518 | L 43.21875 8.296875 
 519 | Q 40.140625 3.328125 35.546875 0.953125 
 520 | Q 30.953125 -1.421875 24.3125 -1.421875 
 521 | Q 15.921875 -1.421875 10.953125 3.296875 
 522 | Q 6 8.015625 6 15.921875 
 523 | Q 6 25.140625 12.171875 29.828125 
 524 | Q 18.359375 34.515625 30.609375 34.515625 
 525 | L 43.21875 34.515625 
 526 | L 43.21875 35.40625 
 527 | Q 43.21875 41.609375 39.140625 45 
 528 | Q 35.0625 48.390625 27.6875 48.390625 
 529 | Q 23 48.390625 18.546875 47.265625 
 530 | Q 14.109375 46.140625 10.015625 43.890625 
 531 | L 10.015625 52.203125 
 532 | Q 14.9375 54.109375 19.578125 55.046875 
 533 | Q 24.21875 56 28.609375 56 
 534 | Q 40.484375 56 46.34375 49.84375 
 535 | Q 52.203125 43.703125 52.203125 31.203125 
 536 | z
 537 | " id="DejaVuSans-97"/>
 538 |       <path d="M 18.3125 70.21875 
 539 | L 18.3125 54.6875 
 540 | L 36.8125 54.6875 
 541 | L 36.8125 47.703125 
 542 | L 18.3125 47.703125 
 543 | L 18.3125 18.015625 
 544 | Q 18.3125 11.328125 20.140625 9.421875 
 545 | Q 21.96875 7.515625 27.59375 7.515625 
 546 | L 36.8125 7.515625 
 547 | L 36.8125 0 
 548 | L 27.59375 0 
 549 | Q 17.1875 0 13.234375 3.875 
 550 | Q 9.28125 7.765625 9.28125 18.015625 
 551 | L 9.28125 47.703125 
 552 | L 2.6875 47.703125 
 553 | L 2.6875 54.6875 
 554 | L 9.28125 54.6875 
 555 | L 9.28125 70.21875 
 556 | z
 557 | " id="DejaVuSans-116"/>
 558 |      </defs>
 559 |      <g transform="translate(175.730469 280.276563)scale(0.1 -0.1)">
 560 |       <use xlink:href="#DejaVuSans-67"/>
 561 |       <use x="69.824219" xlink:href="#DejaVuSans-111"/>
 562 |       <use x="131.005859" xlink:href="#DejaVuSans-109"/>
 563 |       <use x="228.417969" xlink:href="#DejaVuSans-112"/>
 564 |       <use x="291.894531" xlink:href="#DejaVuSans-114"/>
 565 |       <use x="332.976562" xlink:href="#DejaVuSans-101"/>
 566 |       <use x="394.5" xlink:href="#DejaVuSans-115"/>
 567 |       <use x="446.599609" xlink:href="#DejaVuSans-115"/>
 568 |       <use x="498.699219" xlink:href="#DejaVuSans-105"/>
 569 |       <use x="526.482422" xlink:href="#DejaVuSans-111"/>
 570 |       <use x="587.664062" xlink:href="#DejaVuSans-110"/>
 571 |       <use x="651.042969" xlink:href="#DejaVuSans-32"/>
 572 |       <use x="682.830078" xlink:href="#DejaVuSans-114"/>
 573 |       <use x="723.943359" xlink:href="#DejaVuSans-97"/>
 574 |       <use x="785.222656" xlink:href="#DejaVuSans-116"/>
 575 |       <use x="824.431641" xlink:href="#DejaVuSans-105"/>
 576 |       <use x="852.214844" xlink:href="#DejaVuSans-111"/>
 577 |      </g>
 578 |     </g>
 579 |    </g>
 580 |    <g id="matplotlib.axis_2">
 581 |     <g id="ytick_1">
 582 |      <g id="line2d_11">
 583 |       <path clip-path="url(#pe986d3e2bd)" d="M 54 211.016991 
 584 | L 388.8 211.016991 
 585 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 586 |      </g>
 587 |      <g id="line2d_12">
 588 |       <defs>
 589 |        <path d="M 0 0 
 590 | L -3.5 0 
 591 | " id="mff23e23d23" style="stroke:#000000;stroke-width:0.8;"/>
 592 |       </defs>
 593 |       <g>
 594 |        <use style="stroke:#000000;stroke-width:0.8;" x="54" xlink:href="#mff23e23d23" y="211.016991"/>
 595 |       </g>
 596 |      </g>
 597 |      <g id="text_7">
 598 |       <!-- 2 -->
 599 |       <g transform="translate(40.6375 214.816209)scale(0.1 -0.1)">
 600 |        <use xlink:href="#DejaVuSans-50"/>
 601 |       </g>
 602 |      </g>
 603 |     </g>
 604 |     <g id="ytick_2">
 605 |      <g id="line2d_13">
 606 |       <path clip-path="url(#pe986d3e2bd)" d="M 54 168.41511 
 607 | L 388.8 168.41511 
 608 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 609 |      </g>
 610 |      <g id="line2d_14">
 611 |       <g>
 612 |        <use style="stroke:#000000;stroke-width:0.8;" x="54" xlink:href="#mff23e23d23" y="168.41511"/>
 613 |       </g>
 614 |      </g>
 615 |      <g id="text_8">
 616 |       <!-- 3 -->
 617 |       <g transform="translate(40.6375 172.214328)scale(0.1 -0.1)">
 618 |        <use xlink:href="#DejaVuSans-51"/>
 619 |       </g>
 620 |      </g>
 621 |     </g>
 622 |     <g id="ytick_3">
 623 |      <g id="line2d_15">
 624 |       <path clip-path="url(#pe986d3e2bd)" d="M 54 125.813229 
 625 | L 388.8 125.813229 
 626 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 627 |      </g>
 628 |      <g id="line2d_16">
 629 |       <g>
 630 |        <use style="stroke:#000000;stroke-width:0.8;" x="54" xlink:href="#mff23e23d23" y="125.813229"/>
 631 |       </g>
 632 |      </g>
 633 |      <g id="text_9">
 634 |       <!-- 4 -->
 635 |       <g transform="translate(40.6375 129.612448)scale(0.1 -0.1)">
 636 |        <use xlink:href="#DejaVuSans-52"/>
 637 |       </g>
 638 |      </g>
 639 |     </g>
 640 |     <g id="ytick_4">
 641 |      <g id="line2d_17">
 642 |       <path clip-path="url(#pe986d3e2bd)" d="M 54 83.211348 
 643 | L 388.8 83.211348 
 644 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 645 |      </g>
 646 |      <g id="line2d_18">
 647 |       <g>
 648 |        <use style="stroke:#000000;stroke-width:0.8;" x="54" xlink:href="#mff23e23d23" y="83.211348"/>
 649 |       </g>
 650 |      </g>
 651 |      <g id="text_10">
 652 |       <!-- 5 -->
 653 |       <defs>
 654 |        <path d="M 10.796875 72.90625 
 655 | L 49.515625 72.90625 
 656 | L 49.515625 64.59375 
 657 | L 19.828125 64.59375 
 658 | L 19.828125 46.734375 
 659 | Q 21.96875 47.46875 24.109375 47.828125 
 660 | Q 26.265625 48.1875 28.421875 48.1875 
 661 | Q 40.625 48.1875 47.75 41.5 
 662 | Q 54.890625 34.8125 54.890625 23.390625 
 663 | Q 54.890625 11.625 47.5625 5.09375 
 664 | Q 40.234375 -1.421875 26.90625 -1.421875 
 665 | Q 22.3125 -1.421875 17.546875 -0.640625 
 666 | Q 12.796875 0.140625 7.71875 1.703125 
 667 | L 7.71875 11.625 
 668 | Q 12.109375 9.234375 16.796875 8.0625 
 669 | Q 21.484375 6.890625 26.703125 6.890625 
 670 | Q 35.15625 6.890625 40.078125 11.328125 
 671 | Q 45.015625 15.765625 45.015625 23.390625 
 672 | Q 45.015625 31 40.078125 35.4375 
 673 | Q 35.15625 39.890625 26.703125 39.890625 
 674 | Q 22.75 39.890625 18.8125 39.015625 
 675 | Q 14.890625 38.140625 10.796875 36.28125 
 676 | z
 677 | " id="DejaVuSans-53"/>
 678 |       </defs>
 679 |       <g transform="translate(40.6375 87.010567)scale(0.1 -0.1)">
 680 |        <use xlink:href="#DejaVuSans-53"/>
 681 |       </g>
 682 |      </g>
 683 |     </g>
 684 |     <g id="ytick_5">
 685 |      <g id="line2d_19">
 686 |       <path clip-path="url(#pe986d3e2bd)" d="M 54 40.609467 
 687 | L 388.8 40.609467 
 688 | " style="fill:none;stroke:#b0b0b0;stroke-linecap:square;stroke-width:0.8;"/>
 689 |      </g>
 690 |      <g id="line2d_20">
 691 |       <g>
 692 |        <use style="stroke:#000000;stroke-width:0.8;" x="54" xlink:href="#mff23e23d23" y="40.609467"/>
 693 |       </g>
 694 |      </g>
 695 |      <g id="text_11">
 696 |       <!-- 6 -->
 697 |       <g transform="translate(40.6375 44.408686)scale(0.1 -0.1)">
 698 |        <use xlink:href="#DejaVuSans-54"/>
 699 |       </g>
 700 |      </g>
 701 |     </g>
 702 |     <g id="text_12">
 703 |      <!-- Error (%) -->
 704 |      <defs>
 705 |       <path d="M 9.8125 72.90625 
 706 | L 55.90625 72.90625 
 707 | L 55.90625 64.59375 
 708 | L 19.671875 64.59375 
 709 | L 19.671875 43.015625 
 710 | L 54.390625 43.015625 
 711 | L 54.390625 34.71875 
 712 | L 19.671875 34.71875 
 713 | L 19.671875 8.296875 
 714 | L 56.78125 8.296875 
 715 | L 56.78125 0 
 716 | L 9.8125 0 
 717 | z
 718 | " id="DejaVuSans-69"/>
 719 |       <path d="M 31 75.875 
 720 | Q 24.46875 64.65625 21.28125 53.65625 
 721 | Q 18.109375 42.671875 18.109375 31.390625 
 722 | Q 18.109375 20.125 21.3125 9.0625 
 723 | Q 24.515625 -2 31 -13.1875 
 724 | L 23.1875 -13.1875 
 725 | Q 15.875 -1.703125 12.234375 9.375 
 726 | Q 8.59375 20.453125 8.59375 31.390625 
 727 | Q 8.59375 42.28125 12.203125 53.3125 
 728 | Q 15.828125 64.359375 23.1875 75.875 
 729 | z
 730 | " id="DejaVuSans-40"/>
 731 |       <path d="M 72.703125 32.078125 
 732 | Q 68.453125 32.078125 66.03125 28.46875 
 733 | Q 63.625 24.859375 63.625 18.40625 
 734 | Q 63.625 12.0625 66.03125 8.421875 
 735 | Q 68.453125 4.78125 72.703125 4.78125 
 736 | Q 76.859375 4.78125 79.265625 8.421875 
 737 | Q 81.6875 12.0625 81.6875 18.40625 
 738 | Q 81.6875 24.8125 79.265625 28.4375 
 739 | Q 76.859375 32.078125 72.703125 32.078125 
 740 | z
 741 | M 72.703125 38.28125 
 742 | Q 80.421875 38.28125 84.953125 32.90625 
 743 | Q 89.5 27.546875 89.5 18.40625 
 744 | Q 89.5 9.28125 84.9375 3.921875 
 745 | Q 80.375 -1.421875 72.703125 -1.421875 
 746 | Q 64.890625 -1.421875 60.34375 3.921875 
 747 | Q 55.8125 9.28125 55.8125 18.40625 
 748 | Q 55.8125 27.59375 60.375 32.9375 
 749 | Q 64.9375 38.28125 72.703125 38.28125 
 750 | z
 751 | M 22.3125 68.015625 
 752 | Q 18.109375 68.015625 15.6875 64.375 
 753 | Q 13.28125 60.75 13.28125 54.390625 
 754 | Q 13.28125 47.953125 15.671875 44.328125 
 755 | Q 18.0625 40.71875 22.3125 40.71875 
 756 | Q 26.5625 40.71875 28.96875 44.328125 
 757 | Q 31.390625 47.953125 31.390625 54.390625 
 758 | Q 31.390625 60.6875 28.953125 64.34375 
 759 | Q 26.515625 68.015625 22.3125 68.015625 
 760 | z
 761 | M 66.40625 74.21875 
 762 | L 74.21875 74.21875 
 763 | L 28.609375 -1.421875 
 764 | L 20.796875 -1.421875 
 765 | z
 766 | M 22.3125 74.21875 
 767 | Q 30.03125 74.21875 34.609375 68.875 
 768 | Q 39.203125 63.53125 39.203125 54.390625 
 769 | Q 39.203125 45.171875 34.640625 39.84375 
 770 | Q 30.078125 34.515625 22.3125 34.515625 
 771 | Q 14.546875 34.515625 10.03125 39.859375 
 772 | Q 5.515625 45.21875 5.515625 54.390625 
 773 | Q 5.515625 63.484375 10.046875 68.84375 
 774 | Q 14.59375 74.21875 22.3125 74.21875 
 775 | z
 776 | " id="DejaVuSans-37"/>
 777 |       <path d="M 8.015625 75.875 
 778 | L 15.828125 75.875 
 779 | Q 23.140625 64.359375 26.78125 53.3125 
 780 | Q 30.421875 42.28125 30.421875 31.390625 
 781 | Q 30.421875 20.453125 26.78125 9.375 
 782 | Q 23.140625 -1.703125 15.828125 -13.1875 
 783 | L 8.015625 -13.1875 
 784 | Q 14.5 -2 17.703125 9.0625 
 785 | Q 20.90625 20.125 20.90625 31.390625 
 786 | Q 20.90625 42.671875 17.703125 53.65625 
 787 | Q 14.5 64.65625 8.015625 75.875 
 788 | z
 789 | " id="DejaVuSans-41"/>
 790 |      </defs>
 791 |      <g transform="translate(34.557812 165.904219)rotate(-90)scale(0.1 -0.1)">
 792 |       <use xlink:href="#DejaVuSans-69"/>
 793 |       <use x="63.183594" xlink:href="#DejaVuSans-114"/>
 794 |       <use x="104.28125" xlink:href="#DejaVuSans-114"/>
 795 |       <use x="145.363281" xlink:href="#DejaVuSans-111"/>
 796 |       <use x="206.544922" xlink:href="#DejaVuSans-114"/>
 797 |       <use x="247.658203" xlink:href="#DejaVuSans-32"/>
 798 |       <use x="279.445312" xlink:href="#DejaVuSans-40"/>
 799 |       <use x="318.458984" xlink:href="#DejaVuSans-37"/>
 800 |       <use x="413.478516" xlink:href="#DejaVuSans-41"/>
 801 |      </g>
 802 |     </g>
 803 |    </g>
 804 |    <g id="line2d_21">
 805 |     <path clip-path="url(#pe986d3e2bd)" d="M 69.218182 162.024828 
 806 | L 145.309091 200.36652 
 807 | L 221.4 217.833292 
 808 | L 297.490909 231.465893 
 809 | L 373.581818 234.022006 
 810 | " style="fill:none;stroke:#1f77b4;stroke-linecap:square;stroke-width:2;"/>
 811 |     <defs>
 812 |      <path d="M -3 3 
 813 | L 3 3 
 814 | L 3 -3 
 815 | L -3 -3 
 816 | z
 817 | " id="m6c88cf564b" style="stroke:#1f77b4;stroke-linejoin:miter;"/>
 818 |     </defs>
 819 |     <g clip-path="url(#pe986d3e2bd)">
 820 |      <use style="fill:#1f77b4;stroke:#1f77b4;stroke-linejoin:miter;" x="69.218182" xlink:href="#m6c88cf564b" y="162.024828"/>
 821 |      <use style="fill:#1f77b4;stroke:#1f77b4;stroke-linejoin:miter;" x="145.309091" xlink:href="#m6c88cf564b" y="200.36652"/>
 822 |      <use style="fill:#1f77b4;stroke:#1f77b4;stroke-linejoin:miter;" x="221.4" xlink:href="#m6c88cf564b" y="217.833292"/>
 823 |      <use style="fill:#1f77b4;stroke:#1f77b4;stroke-linejoin:miter;" x="297.490909" xlink:href="#m6c88cf564b" y="231.465893"/>
 824 |      <use style="fill:#1f77b4;stroke:#1f77b4;stroke-linejoin:miter;" x="373.581818" xlink:href="#m6c88cf564b" y="234.022006"/>
 825 |     </g>
 826 |    </g>
 827 |    <g id="line2d_22">
 828 |     <path clip-path="url(#pe986d3e2bd)" d="M 69.218182 44.443636 
 829 | L 145.309091 145.836113 
 830 | L 221.4 199.088464 
 831 | L 297.490909 221.667461 
 832 | L 373.581818 242.116364 
 833 | " style="fill:none;stroke:#ff7f0e;stroke-dasharray:7.4,3.2;stroke-dashoffset:0;stroke-width:2;"/>
 834 |     <defs>
 835 |      <path d="M -0 4.242641 
 836 | L 2.545584 0 
 837 | L 0 -4.242641 
 838 | L -2.545584 -0 
 839 | z
 840 | " id="m9efc44ff44" style="stroke:#ff7f0e;stroke-linejoin:miter;"/>
 841 |     </defs>
 842 |     <g clip-path="url(#pe986d3e2bd)">
 843 |      <use style="fill:#ff7f0e;stroke:#ff7f0e;stroke-linejoin:miter;" x="69.218182" xlink:href="#m9efc44ff44" y="44.443636"/>
 844 |      <use style="fill:#ff7f0e;stroke:#ff7f0e;stroke-linejoin:miter;" x="145.309091" xlink:href="#m9efc44ff44" y="145.836113"/>
 845 |      <use style="fill:#ff7f0e;stroke:#ff7f0e;stroke-linejoin:miter;" x="221.4" xlink:href="#m9efc44ff44" y="199.088464"/>
 846 |      <use style="fill:#ff7f0e;stroke:#ff7f0e;stroke-linejoin:miter;" x="297.490909" xlink:href="#m9efc44ff44" y="221.667461"/>
 847 |      <use style="fill:#ff7f0e;stroke:#ff7f0e;stroke-linejoin:miter;" x="373.581818" xlink:href="#m9efc44ff44" y="242.116364"/>
 848 |     </g>
 849 |    </g>
 850 |    <g id="patch_3">
 851 |     <path d="M 54 252 
 852 | L 54 34.56 
 853 | " style="fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;"/>
 854 |    </g>
 855 |    <g id="patch_4">
 856 |     <path d="M 388.8 252 
 857 | L 388.8 34.56 
 858 | " style="fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;"/>
 859 |    </g>
 860 |    <g id="patch_5">
 861 |     <path d="M 54 252 
 862 | L 388.8 252 
 863 | " style="fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;"/>
 864 |    </g>
 865 |    <g id="patch_6">
 866 |     <path d="M 54 34.56 
 867 | L 388.8 34.56 
 868 | " style="fill:none;stroke:#000000;stroke-linecap:square;stroke-linejoin:miter;stroke-width:0.8;"/>
 869 |    </g>
 870 |    <g id="text_13">
 871 |     <!-- 1 hidden layer, 1000 units -->
 872 |     <defs>
 873 |      <path d="M 54.890625 33.015625 
 874 | L 54.890625 0 
 875 | L 45.90625 0 
 876 | L 45.90625 32.71875 
 877 | Q 45.90625 40.484375 42.875 44.328125 
 878 | Q 39.84375 48.1875 33.796875 48.1875 
 879 | Q 26.515625 48.1875 22.3125 43.546875 
 880 | Q 18.109375 38.921875 18.109375 30.90625 
 881 | L 18.109375 0 
 882 | L 9.078125 0 
 883 | L 9.078125 75.984375 
 884 | L 18.109375 75.984375 
 885 | L 18.109375 46.1875 
 886 | Q 21.34375 51.125 25.703125 53.5625 
 887 | Q 30.078125 56 35.796875 56 
 888 | Q 45.21875 56 50.046875 50.171875 
 889 | Q 54.890625 44.34375 54.890625 33.015625 
 890 | z
 891 | " id="DejaVuSans-104"/>
 892 |      <path d="M 45.40625 46.390625 
 893 | L 45.40625 75.984375 
 894 | L 54.390625 75.984375 
 895 | L 54.390625 0 
 896 | L 45.40625 0 
 897 | L 45.40625 8.203125 
 898 | Q 42.578125 3.328125 38.25 0.953125 
 899 | Q 33.9375 -1.421875 27.875 -1.421875 
 900 | Q 17.96875 -1.421875 11.734375 6.484375 
 901 | Q 5.515625 14.40625 5.515625 27.296875 
 902 | Q 5.515625 40.1875 11.734375 48.09375 
 903 | Q 17.96875 56 27.875 56 
 904 | Q 33.9375 56 38.25 53.625 
 905 | Q 42.578125 51.265625 45.40625 46.390625 
 906 | z
 907 | M 14.796875 27.296875 
 908 | Q 14.796875 17.390625 18.875 11.75 
 909 | Q 22.953125 6.109375 30.078125 6.109375 
 910 | Q 37.203125 6.109375 41.296875 11.75 
 911 | Q 45.40625 17.390625 45.40625 27.296875 
 912 | Q 45.40625 37.203125 41.296875 42.84375 
 913 | Q 37.203125 48.484375 30.078125 48.484375 
 914 | Q 22.953125 48.484375 18.875 42.84375 
 915 | Q 14.796875 37.203125 14.796875 27.296875 
 916 | z
 917 | " id="DejaVuSans-100"/>
 918 |      <path d="M 9.421875 75.984375 
 919 | L 18.40625 75.984375 
 920 | L 18.40625 0 
 921 | L 9.421875 0 
 922 | z
 923 | " id="DejaVuSans-108"/>
 924 |      <path d="M 32.171875 -5.078125 
 925 | Q 28.375 -14.84375 24.75 -17.8125 
 926 | Q 21.140625 -20.796875 15.09375 -20.796875 
 927 | L 7.90625 -20.796875 
 928 | L 7.90625 -13.28125 
 929 | L 13.1875 -13.28125 
 930 | Q 16.890625 -13.28125 18.9375 -11.515625 
 931 | Q 21 -9.765625 23.484375 -3.21875 
 932 | L 25.09375 0.875 
 933 | L 2.984375 54.6875 
 934 | L 12.5 54.6875 
 935 | L 29.59375 11.921875 
 936 | L 46.6875 54.6875 
 937 | L 56.203125 54.6875 
 938 | z
 939 | " id="DejaVuSans-121"/>
 940 |      <path d="M 11.71875 12.40625 
 941 | L 22.015625 12.40625 
 942 | L 22.015625 4 
 943 | L 14.015625 -11.625 
 944 | L 7.71875 -11.625 
 945 | L 11.71875 4 
 946 | z
 947 | " id="DejaVuSans-44"/>
 948 |      <path d="M 31.78125 66.40625 
 949 | Q 24.171875 66.40625 20.328125 58.90625 
 950 | Q 16.5 51.421875 16.5 36.375 
 951 | Q 16.5 21.390625 20.328125 13.890625 
 952 | Q 24.171875 6.390625 31.78125 6.390625 
 953 | Q 39.453125 6.390625 43.28125 13.890625 
 954 | Q 47.125 21.390625 47.125 36.375 
 955 | Q 47.125 51.421875 43.28125 58.90625 
 956 | Q 39.453125 66.40625 31.78125 66.40625 
 957 | z
 958 | M 31.78125 74.21875 
 959 | Q 44.046875 74.21875 50.515625 64.515625 
 960 | Q 56.984375 54.828125 56.984375 36.375 
 961 | Q 56.984375 17.96875 50.515625 8.265625 
 962 | Q 44.046875 -1.421875 31.78125 -1.421875 
 963 | Q 19.53125 -1.421875 13.0625 8.265625 
 964 | Q 6.59375 17.96875 6.59375 36.375 
 965 | Q 6.59375 54.828125 13.0625 64.515625 
 966 | Q 19.53125 74.21875 31.78125 74.21875 
 967 | z
 968 | " id="DejaVuSans-48"/>
 969 |      <path d="M 8.5 21.578125 
 970 | L 8.5 54.6875 
 971 | L 17.484375 54.6875 
 972 | L 17.484375 21.921875 
 973 | Q 17.484375 14.15625 20.5 10.265625 
 974 | Q 23.53125 6.390625 29.59375 6.390625 
 975 | Q 36.859375 6.390625 41.078125 11.03125 
 976 | Q 45.3125 15.671875 45.3125 23.6875 
 977 | L 45.3125 54.6875 
 978 | L 54.296875 54.6875 
 979 | L 54.296875 0 
 980 | L 45.3125 0 
 981 | L 45.3125 8.40625 
 982 | Q 42.046875 3.421875 37.71875 1 
 983 | Q 33.40625 -1.421875 27.6875 -1.421875 
 984 | Q 18.265625 -1.421875 13.375 4.4375 
 985 | Q 8.5 10.296875 8.5 21.578125 
 986 | z
 987 | M 31.109375 56 
 988 | z
 989 | " id="DejaVuSans-117"/>
 990 |     </defs>
 991 |     <g transform="translate(142.393125 28.56)scale(0.12 -0.12)">
 992 |      <use xlink:href="#DejaVuSans-49"/>
 993 |      <use x="63.623047" xlink:href="#DejaVuSans-32"/>
 994 |      <use x="95.410156" xlink:href="#DejaVuSans-104"/>
 995 |      <use x="158.789062" xlink:href="#DejaVuSans-105"/>
 996 |      <use x="186.572266" xlink:href="#DejaVuSans-100"/>
 997 |      <use x="250.048828" xlink:href="#DejaVuSans-100"/>
 998 |      <use x="313.525391" xlink:href="#DejaVuSans-101"/>
 999 |      <use x="375.048828" xlink:href="#DejaVuSans-110"/>
1000 |      <use x="438.427734" xlink:href="#DejaVuSans-32"/>
1001 |      <use x="470.214844" xlink:href="#DejaVuSans-108"/>
1002 |      <use x="497.998047" xlink:href="#DejaVuSans-97"/>
1003 |      <use x="559.277344" xlink:href="#DejaVuSans-121"/>
1004 |      <use x="618.457031" xlink:href="#DejaVuSans-101"/>
1005 |      <use x="679.980469" xlink:href="#DejaVuSans-114"/>
1006 |      <use x="721.09375" xlink:href="#DejaVuSans-44"/>
1007 |      <use x="752.880859" xlink:href="#DejaVuSans-32"/>
1008 |      <use x="784.667969" xlink:href="#DejaVuSans-49"/>
1009 |      <use x="848.291016" xlink:href="#DejaVuSans-48"/>
1010 |      <use x="911.914062" xlink:href="#DejaVuSans-48"/>
1011 |      <use x="975.537109" xlink:href="#DejaVuSans-48"/>
1012 |      <use x="1039.160156" xlink:href="#DejaVuSans-32"/>
1013 |      <use x="1070.947266" xlink:href="#DejaVuSans-117"/>
1014 |      <use x="1134.326172" xlink:href="#DejaVuSans-110"/>
1015 |      <use x="1197.705078" xlink:href="#DejaVuSans-105"/>
1016 |      <use x="1225.488281" xlink:href="#DejaVuSans-116"/>
1017 |      <use x="1264.697266" xlink:href="#DejaVuSans-115"/>
1018 |     </g>
1019 |    </g>
1020 |    <g id="legend_1">
1021 |     <g id="patch_7">
1022 |      <path d="M 283.9625 71.91625 
1023 | L 381.8 71.91625 
1024 | Q 383.8 71.91625 383.8 69.91625 
1025 | L 383.8 41.56 
1026 | Q 383.8 39.56 381.8 39.56 
1027 | L 283.9625 39.56 
1028 | Q 281.9625 39.56 281.9625 41.56 
1029 | L 281.9625 69.91625 
1030 | Q 281.9625 71.91625 283.9625 71.91625 
1031 | z
1032 | " style="fill:#ffffff;opacity:0.8;stroke:#cccccc;stroke-linejoin:miter;"/>
1033 |     </g>
1034 |     <g id="line2d_23">
1035 |      <path d="M 285.9625 47.658437 
1036 | L 315.9625 47.658437 
1037 | " style="fill:none;stroke:#1f77b4;stroke-linecap:square;stroke-width:2;"/>
1038 |     </g>
1039 |     <g id="line2d_24">
1040 |      <g>
1041 |       <use style="fill:#1f77b4;stroke:#1f77b4;stroke-linejoin:miter;" x="300.9625" xlink:href="#m6c88cf564b" y="47.658437"/>
1042 |      </g>
1043 |     </g>
1044 |     <g id="text_14">
1045 |      <!-- Hashed NN -->
1046 |      <defs>
1047 |       <path d="M 9.8125 72.90625 
1048 | L 19.671875 72.90625 
1049 | L 19.671875 43.015625 
1050 | L 55.515625 43.015625 
1051 | L 55.515625 72.90625 
1052 | L 65.375 72.90625 
1053 | L 65.375 0 
1054 | L 55.515625 0 
1055 | L 55.515625 34.71875 
1056 | L 19.671875 34.71875 
1057 | L 19.671875 0 
1058 | L 9.8125 0 
1059 | z
1060 | " id="DejaVuSans-72"/>
1061 |       <path d="M 9.8125 72.90625 
1062 | L 23.09375 72.90625 
1063 | L 55.421875 11.921875 
1064 | L 55.421875 72.90625 
1065 | L 64.984375 72.90625 
1066 | L 64.984375 0 
1067 | L 51.703125 0 
1068 | L 19.390625 60.984375 
1069 | L 19.390625 0 
1070 | L 9.8125 0 
1071 | z
1072 | " id="DejaVuSans-78"/>
1073 |      </defs>
1074 |      <g transform="translate(323.9625 51.158437)scale(0.1 -0.1)">
1075 |       <use xlink:href="#DejaVuSans-72"/>
1076 |       <use x="75.195312" xlink:href="#DejaVuSans-97"/>
1077 |       <use x="136.474609" xlink:href="#DejaVuSans-115"/>
1078 |       <use x="188.574219" xlink:href="#DejaVuSans-104"/>
1079 |       <use x="251.953125" xlink:href="#DejaVuSans-101"/>
1080 |       <use x="313.476562" xlink:href="#DejaVuSans-100"/>
1081 |       <use x="376.953125" xlink:href="#DejaVuSans-32"/>
1082 |       <use x="408.740234" xlink:href="#DejaVuSans-78"/>
1083 |       <use x="483.544922" xlink:href="#DejaVuSans-78"/>
1084 |      </g>
1085 |     </g>
1086 |     <g id="line2d_25">
1087 |      <path d="M 285.9625 62.336562 
1088 | L 315.9625 62.336562 
1089 | " style="fill:none;stroke:#ff7f0e;stroke-dasharray:7.4,3.2;stroke-dashoffset:0;stroke-width:2;"/>
1090 |     </g>
1091 |     <g id="line2d_26">
1092 |      <g>
1093 |       <use style="fill:#ff7f0e;stroke:#ff7f0e;stroke-linejoin:miter;" x="300.9625" xlink:href="#m9efc44ff44" y="62.336562"/>
1094 |      </g>
1095 |     </g>
1096 |     <g id="text_15">
1097 |      <!-- Equiv. NN -->
1098 |      <defs>
1099 |       <path d="M 14.796875 27.296875 
1100 | Q 14.796875 17.390625 18.875 11.75 
1101 | Q 22.953125 6.109375 30.078125 6.109375 
1102 | Q 37.203125 6.109375 41.296875 11.75 
1103 | Q 45.40625 17.390625 45.40625 27.296875 
1104 | Q 45.40625 37.203125 41.296875 42.84375 
1105 | Q 37.203125 48.484375 30.078125 48.484375 
1106 | Q 22.953125 48.484375 18.875 42.84375 
1107 | Q 14.796875 37.203125 14.796875 27.296875 
1108 | z
1109 | M 45.40625 8.203125 
1110 | Q 42.578125 3.328125 38.25 0.953125 
1111 | Q 33.9375 -1.421875 27.875 -1.421875 
1112 | Q 17.96875 -1.421875 11.734375 6.484375 
1113 | Q 5.515625 14.40625 5.515625 27.296875 
1114 | Q 5.515625 40.1875 11.734375 48.09375 
1115 | Q 17.96875 56 27.875 56 
1116 | Q 33.9375 56 38.25 53.625 
1117 | Q 42.578125 51.265625 45.40625 46.390625 
1118 | L 45.40625 54.6875 
1119 | L 54.390625 54.6875 
1120 | L 54.390625 -20.796875 
1121 | L 45.40625 -20.796875 
1122 | z
1123 | " id="DejaVuSans-113"/>
1124 |       <path d="M 2.984375 54.6875 
1125 | L 12.5 54.6875 
1126 | L 29.59375 8.796875 
1127 | L 46.6875 54.6875 
1128 | L 56.203125 54.6875 
1129 | L 35.6875 0 
1130 | L 23.484375 0 
1131 | z
1132 | " id="DejaVuSans-118"/>
1133 |       <path d="M 10.6875 12.40625 
1134 | L 21 12.40625 
1135 | L 21 0 
1136 | L 10.6875 0 
1137 | z
1138 | " id="DejaVuSans-46"/>
1139 |      </defs>
1140 |      <g transform="translate(323.9625 65.836562)scale(0.1 -0.1)">
1141 |       <use xlink:href="#DejaVuSans-69"/>
1142 |       <use x="63.183594" xlink:href="#DejaVuSans-113"/>
1143 |       <use x="126.660156" xlink:href="#DejaVuSans-117"/>
1144 |       <use x="190.039062" xlink:href="#DejaVuSans-105"/>
1145 |       <use x="217.822266" xlink:href="#DejaVuSans-118"/>
1146 |       <use x="276.892578" xlink:href="#DejaVuSans-46"/>
1147 |       <use x="308.679688" xlink:href="#DejaVuSans-32"/>
1148 |       <use x="340.466797" xlink:href="#DejaVuSans-78"/>
1149 |       <use x="415.271484" xlink:href="#DejaVuSans-78"/>
1150 |      </g>
1151 |     </g>
1152 |    </g>
1153 |   </g>
1154 |  </g>
1155 |  <defs>
1156 |   <clipPath id="pe986d3e2bd">
1157 |    <rect height="217.44" width="334.8" x="54" y="34.56"/>
1158 |   </clipPath>
1159 |  </defs>
1160 | </svg>
1161 | 


--------------------------------------------------------------------------------
/example.ipynb:
--------------------------------------------------------------------------------
  1 | {
  2 |  "cells": [
  3 |   {
  4 |    "cell_type": "code",
  5 |    "execution_count": 1,
  6 |    "metadata": {},
  7 |    "outputs": [],
  8 |    "source": [
  9 |     "import torch\n",
 10 |     "import torch.optim as optim\n",
 11 |     "import random\n",
 12 |     "from mnist import load_data, train, evaluate, Net, HashedNet\n",
 13 |     "from utils import get_equivalent_compression\n",
 14 |     "\n",
 15 |     "use_cuda = torch.cuda.is_available()\n",
 16 |     "torch.manual_seed(1)\n",
 17 |     "random.seed(1)"
 18 |    ]
 19 |   },
 20 |   {
 21 |    "cell_type": "code",
 22 |    "execution_count": 2,
 23 |    "metadata": {
 24 |     "scrolled": false
 25 |    },
 26 |    "outputs": [
 27 |     {
 28 |      "name": "stdout",
 29 |      "output_type": "stream",
 30 |      "text": [
 31 |       "Compression rate: 0.015625\n",
 32 |       "The number of parameters is: 12730\n",
 33 |       "Epoch 1 Train loss: 0.698 Val loss: 0.298 Val acc: 91.05%\n",
 34 |       "Epoch 2 Train loss: 0.540 Val loss: 0.262 Val acc: 92.23%\n",
 35 |       "Epoch 3 Train loss: 0.526 Val loss: 0.263 Val acc: 91.97%\n",
 36 |       "Epoch 4 Train loss: 0.515 Val loss: 0.255 Val acc: 91.85%\n",
 37 |       "Epoch 5 Train loss: 0.505 Val loss: 0.243 Val acc: 92.58%\n",
 38 |       "Epoch 6 Train loss: 0.498 Val loss: 0.243 Val acc: 92.82%\n",
 39 |       "Epoch 7 Train loss: 0.501 Val loss: 0.243 Val acc: 92.80%\n",
 40 |       "Epoch 8 Train loss: 0.497 Val loss: 0.239 Val acc: 92.87%\n",
 41 |       "Epoch 9 Train loss: 0.493 Val loss: 0.239 Val acc: 92.70%\n",
 42 |       "Epoch 10 Train loss: 0.490 Val loss: 0.235 Val acc: 93.00%\n",
 43 |       "Epoch 11 Train loss: 0.493 Val loss: 0.237 Val acc: 93.20%\n",
 44 |       "Epoch 12 Train loss: 0.496 Val loss: 0.242 Val acc: 92.73%\n",
 45 |       "Epoch    12: reducing learning rate of group 0 to 1.0000e-03.\n",
 46 |       "Epoch 13 Train loss: 0.479 Val loss: 0.238 Val acc: 92.75%\n",
 47 |       "Epoch 14 Train loss: 0.446 Val loss: 0.221 Val acc: 93.62%\n",
 48 |       "Epoch 15 Train loss: 0.435 Val loss: 0.220 Val acc: 93.50%\n",
 49 |       "Epoch 16 Train loss: 0.438 Val loss: 0.216 Val acc: 93.70%\n",
 50 |       "Epoch 17 Train loss: 0.438 Val loss: 0.216 Val acc: 93.62%\n",
 51 |       "Epoch 18 Train loss: 0.431 Val loss: 0.215 Val acc: 93.78%\n",
 52 |       "Epoch 19 Train loss: 0.428 Val loss: 0.216 Val acc: 93.73%\n",
 53 |       "Epoch 20 Train loss: 0.433 Val loss: 0.216 Val acc: 93.62%\n",
 54 |       "Epoch    20: reducing learning rate of group 0 to 1.0000e-04.\n",
 55 |       "Epoch 21 Train loss: 0.430 Val loss: 0.216 Val acc: 93.78%\n",
 56 |       "Epoch 22 Train loss: 0.431 Val loss: 0.215 Val acc: 93.80%\n",
 57 |       "Epoch 23 Train loss: 0.422 Val loss: 0.214 Val acc: 93.70%\n",
 58 |       "Epoch 24 Train loss: 0.426 Val loss: 0.213 Val acc: 93.83%\n",
 59 |       "Epoch 25 Train loss: 0.428 Val loss: 0.213 Val acc: 93.78%\n",
 60 |       "Epoch 26 Train loss: 0.424 Val loss: 0.214 Val acc: 93.72%\n",
 61 |       "Epoch 27 Train loss: 0.427 Val loss: 0.214 Val acc: 93.77%\n",
 62 |       "Epoch    27: reducing learning rate of group 0 to 1.0000e-05.\n",
 63 |       "Epoch 28 Train loss: 0.427 Val loss: 0.214 Val acc: 93.75%\n",
 64 |       "Epoch 29 Train loss: 0.424 Val loss: 0.213 Val acc: 93.75%\n",
 65 |       "Epoch 30 Train loss: 0.427 Val loss: 0.213 Val acc: 93.77%\n",
 66 |       "Epoch 31 Train loss: 0.427 Val loss: 0.213 Val acc: 93.72%\n",
 67 |       "Epoch 32 Train loss: 0.428 Val loss: 0.213 Val acc: 93.73%\n",
 68 |       "Epoch 33 Train loss: 0.427 Val loss: 0.213 Val acc: 93.73%\n",
 69 |       "Epoch 34 Train loss: 0.430 Val loss: 0.213 Val acc: 93.73%\n",
 70 |       "Epoch 35 Train loss: 0.427 Val loss: 0.213 Val acc: 93.73%\n",
 71 |       "Epoch 36 Train loss: 0.423 Val loss: 0.213 Val acc: 93.75%\n",
 72 |       "Epoch    36: reducing learning rate of group 0 to 1.0000e-06.\n",
 73 |       "Epoch 37 Train loss: 0.426 Val loss: 0.213 Val acc: 93.72%\n",
 74 |       "Epoch 38 Train loss: 0.428 Val loss: 0.213 Val acc: 93.72%\n",
 75 |       "Epoch 39 Train loss: 0.422 Val loss: 0.213 Val acc: 93.72%\n",
 76 |       "Epoch 40 Train loss: 0.426 Val loss: 0.213 Val acc: 93.72%\n",
 77 |       "Epoch 41 Train loss: 0.423 Val loss: 0.213 Val acc: 93.72%\n",
 78 |       "Epoch    41: reducing learning rate of group 0 to 1.0000e-07.\n",
 79 |       "Epoch 42 Train loss: 0.423 Val loss: 0.213 Val acc: 93.72%\n",
 80 |       "Epoch 43 Train loss: 0.426 Val loss: 0.213 Val acc: 93.72%\n",
 81 |       "Epoch 44 Train loss: 0.426 Val loss: 0.213 Val acc: 93.72%\n",
 82 |       "Epoch    44: reducing learning rate of group 0 to 1.0000e-08.\n",
 83 |       "Epoch 45 Train loss: 0.425 Val loss: 0.213 Val acc: 93.72%\n",
 84 |       "Epoch 46 Train loss: 0.430 Val loss: 0.213 Val acc: 93.72%\n",
 85 |       "Epoch 47 Train loss: 0.425 Val loss: 0.213 Val acc: 93.72%\n",
 86 |       "Epoch 48 Train loss: 0.422 Val loss: 0.213 Val acc: 93.72%\n",
 87 |       "Epoch 49 Train loss: 0.424 Val loss: 0.213 Val acc: 93.72%\n",
 88 |       "Epoch 50 Train loss: 0.426 Val loss: 0.213 Val acc: 93.72%\n",
 89 |       "Test loss: 0.208 Test acc: 94.09%\n",
 90 |       "The number of parameters is: 12433\n",
 91 |       "Epoch 1 Train loss: 0.566 Val loss: 0.233 Val acc: 93.00%\n",
 92 |       "Epoch 2 Train loss: 0.496 Val loss: 0.223 Val acc: 93.68%\n",
 93 |       "Epoch 3 Train loss: 0.491 Val loss: 0.237 Val acc: 93.37%\n",
 94 |       "Epoch 4 Train loss: 0.505 Val loss: 0.226 Val acc: 93.83%\n",
 95 |       "Epoch     4: reducing learning rate of group 0 to 1.0000e-03.\n",
 96 |       "Epoch 5 Train loss: 0.532 Val loss: 0.229 Val acc: 93.95%\n",
 97 |       "Epoch 6 Train loss: 0.302 Val loss: 0.153 Val acc: 95.38%\n",
 98 |       "Epoch 7 Train loss: 0.269 Val loss: 0.147 Val acc: 95.78%\n",
 99 |       "Epoch 8 Train loss: 0.260 Val loss: 0.143 Val acc: 95.70%\n",
100 |       "Epoch 9 Train loss: 0.248 Val loss: 0.140 Val acc: 95.82%\n",
101 |       "Epoch 10 Train loss: 0.244 Val loss: 0.134 Val acc: 95.98%\n",
102 |       "Epoch 11 Train loss: 0.244 Val loss: 0.134 Val acc: 96.18%\n",
103 |       "Epoch 12 Train loss: 0.235 Val loss: 0.132 Val acc: 96.00%\n",
104 |       "Epoch 13 Train loss: 0.237 Val loss: 0.129 Val acc: 96.00%\n",
105 |       "Epoch 14 Train loss: 0.229 Val loss: 0.127 Val acc: 96.18%\n",
106 |       "Epoch 15 Train loss: 0.227 Val loss: 0.126 Val acc: 96.12%\n",
107 |       "Epoch 16 Train loss: 0.227 Val loss: 0.126 Val acc: 96.20%\n",
108 |       "Epoch 17 Train loss: 0.226 Val loss: 0.125 Val acc: 96.27%\n",
109 |       "Epoch 18 Train loss: 0.225 Val loss: 0.121 Val acc: 96.35%\n",
110 |       "Epoch 19 Train loss: 0.220 Val loss: 0.120 Val acc: 96.47%\n",
111 |       "Epoch 20 Train loss: 0.219 Val loss: 0.120 Val acc: 96.48%\n",
112 |       "Epoch 21 Train loss: 0.217 Val loss: 0.119 Val acc: 96.52%\n",
113 |       "Epoch 22 Train loss: 0.214 Val loss: 0.118 Val acc: 96.40%\n",
114 |       "Epoch 23 Train loss: 0.219 Val loss: 0.119 Val acc: 96.55%\n",
115 |       "Epoch 24 Train loss: 0.213 Val loss: 0.117 Val acc: 96.43%\n",
116 |       "Epoch 25 Train loss: 0.211 Val loss: 0.119 Val acc: 96.45%\n",
117 |       "Epoch 26 Train loss: 0.211 Val loss: 0.115 Val acc: 96.48%\n",
118 |       "Epoch 27 Train loss: 0.215 Val loss: 0.116 Val acc: 96.57%\n",
119 |       "Epoch 28 Train loss: 0.209 Val loss: 0.114 Val acc: 96.63%\n",
120 |       "Epoch 29 Train loss: 0.207 Val loss: 0.115 Val acc: 96.48%\n",
121 |       "Epoch 30 Train loss: 0.204 Val loss: 0.116 Val acc: 96.47%\n",
122 |       "Epoch    30: reducing learning rate of group 0 to 1.0000e-04.\n",
123 |       "Epoch 31 Train loss: 0.206 Val loss: 0.115 Val acc: 96.50%\n",
124 |       "Epoch 32 Train loss: 0.201 Val loss: 0.112 Val acc: 96.65%\n",
125 |       "Epoch 33 Train loss: 0.192 Val loss: 0.112 Val acc: 96.57%\n",
126 |       "Epoch 34 Train loss: 0.194 Val loss: 0.111 Val acc: 96.58%\n",
127 |       "Epoch 35 Train loss: 0.194 Val loss: 0.112 Val acc: 96.67%\n",
128 |       "Epoch 36 Train loss: 0.195 Val loss: 0.112 Val acc: 96.77%\n",
129 |       "Epoch    36: reducing learning rate of group 0 to 1.0000e-05.\n",
130 |       "Epoch 37 Train loss: 0.191 Val loss: 0.111 Val acc: 96.67%\n",
131 |       "Epoch 38 Train loss: 0.197 Val loss: 0.111 Val acc: 96.60%\n",
132 |       "Epoch 39 Train loss: 0.193 Val loss: 0.111 Val acc: 96.68%\n",
133 |       "Epoch    39: reducing learning rate of group 0 to 1.0000e-06.\n",
134 |       "Epoch 40 Train loss: 0.192 Val loss: 0.111 Val acc: 96.65%\n",
135 |       "Epoch 41 Train loss: 0.190 Val loss: 0.111 Val acc: 96.67%\n",
136 |       "Epoch 42 Train loss: 0.195 Val loss: 0.111 Val acc: 96.67%\n",
137 |       "Epoch    42: reducing learning rate of group 0 to 1.0000e-07.\n",
138 |       "Epoch 43 Train loss: 0.192 Val loss: 0.111 Val acc: 96.67%\n",
139 |       "Epoch 44 Train loss: 0.195 Val loss: 0.111 Val acc: 96.67%\n",
140 |       "Epoch 45 Train loss: 0.193 Val loss: 0.111 Val acc: 96.67%\n",
141 |       "Epoch    45: reducing learning rate of group 0 to 1.0000e-08.\n",
142 |       "Epoch 46 Train loss: 0.193 Val loss: 0.111 Val acc: 96.67%\n",
143 |       "Epoch 47 Train loss: 0.194 Val loss: 0.111 Val acc: 96.67%\n",
144 |       "Epoch 48 Train loss: 0.192 Val loss: 0.111 Val acc: 96.67%\n",
145 |       "Epoch 49 Train loss: 0.194 Val loss: 0.111 Val acc: 96.67%\n",
146 |       "Epoch 50 Train loss: 0.196 Val loss: 0.111 Val acc: 96.67%\n",
147 |       "Test loss: 0.108 Test acc: 96.85%\n",
148 |       "Compression rate: 0.03125\n",
149 |       "The number of parameters is: 24655\n",
150 |       "Epoch 1 Train loss: 0.503 Val loss: 0.213 Val acc: 93.43%\n",
151 |       "Epoch 2 Train loss: 0.365 Val loss: 0.185 Val acc: 94.63%\n",
152 |       "Epoch 3 Train loss: 0.337 Val loss: 0.182 Val acc: 94.88%\n",
153 |       "Epoch 4 Train loss: 0.319 Val loss: 0.162 Val acc: 95.18%\n",
154 |       "Epoch 5 Train loss: 0.316 Val loss: 0.161 Val acc: 95.37%\n",
155 |       "Epoch 6 Train loss: 0.307 Val loss: 0.152 Val acc: 95.43%\n",
156 |       "Epoch 7 Train loss: 0.301 Val loss: 0.147 Val acc: 95.65%\n",
157 |       "Epoch 8 Train loss: 0.295 Val loss: 0.154 Val acc: 95.45%\n",
158 |       "Epoch 9 Train loss: 0.293 Val loss: 0.149 Val acc: 95.38%\n",
159 |       "Epoch 10 Train loss: 0.291 Val loss: 0.141 Val acc: 95.88%\n",
160 |       "Epoch 11 Train loss: 0.289 Val loss: 0.151 Val acc: 95.50%\n",
161 |       "Epoch 12 Train loss: 0.279 Val loss: 0.142 Val acc: 95.85%\n",
162 |       "Epoch    12: reducing learning rate of group 0 to 1.0000e-03.\n",
163 |       "Epoch 13 Train loss: 0.282 Val loss: 0.150 Val acc: 95.35%\n",
164 |       "Epoch 14 Train loss: 0.255 Val loss: 0.128 Val acc: 96.40%\n",
165 |       "Epoch 15 Train loss: 0.246 Val loss: 0.127 Val acc: 96.42%\n",
166 |       "Epoch 16 Train loss: 0.244 Val loss: 0.124 Val acc: 96.50%\n",
167 |       "Epoch 17 Train loss: 0.244 Val loss: 0.124 Val acc: 96.40%\n",
168 |       "Epoch 18 Train loss: 0.241 Val loss: 0.122 Val acc: 96.40%\n",
169 |       "Epoch 19 Train loss: 0.243 Val loss: 0.124 Val acc: 96.42%\n",
170 |       "Epoch 20 Train loss: 0.238 Val loss: 0.124 Val acc: 96.58%\n",
171 |       "Epoch 21 Train loss: 0.239 Val loss: 0.122 Val acc: 96.57%\n"
172 |      ]
173 |     },
174 |     {
175 |      "name": "stdout",
176 |      "output_type": "stream",
177 |      "text": [
178 |       "Epoch 22 Train loss: 0.239 Val loss: 0.122 Val acc: 96.45%\n",
179 |       "Epoch 23 Train loss: 0.239 Val loss: 0.123 Val acc: 96.47%\n",
180 |       "Epoch    23: reducing learning rate of group 0 to 1.0000e-04.\n",
181 |       "Epoch 24 Train loss: 0.236 Val loss: 0.122 Val acc: 96.40%\n",
182 |       "Epoch 25 Train loss: 0.235 Val loss: 0.122 Val acc: 96.43%\n",
183 |       "Epoch 26 Train loss: 0.237 Val loss: 0.122 Val acc: 96.50%\n",
184 |       "Epoch    26: reducing learning rate of group 0 to 1.0000e-05.\n",
185 |       "Epoch 27 Train loss: 0.233 Val loss: 0.122 Val acc: 96.45%\n",
186 |       "Epoch 28 Train loss: 0.234 Val loss: 0.122 Val acc: 96.45%\n",
187 |       "Epoch 29 Train loss: 0.233 Val loss: 0.122 Val acc: 96.45%\n",
188 |       "Epoch    29: reducing learning rate of group 0 to 1.0000e-06.\n",
189 |       "Epoch 30 Train loss: 0.233 Val loss: 0.122 Val acc: 96.47%\n",
190 |       "Epoch 31 Train loss: 0.233 Val loss: 0.122 Val acc: 96.47%\n",
191 |       "Epoch 32 Train loss: 0.238 Val loss: 0.122 Val acc: 96.47%\n",
192 |       "Epoch    32: reducing learning rate of group 0 to 1.0000e-07.\n",
193 |       "Epoch 33 Train loss: 0.237 Val loss: 0.122 Val acc: 96.47%\n",
194 |       "Epoch 34 Train loss: 0.232 Val loss: 0.122 Val acc: 96.47%\n",
195 |       "Epoch 35 Train loss: 0.238 Val loss: 0.122 Val acc: 96.47%\n",
196 |       "Epoch    35: reducing learning rate of group 0 to 1.0000e-08.\n",
197 |       "Epoch 36 Train loss: 0.231 Val loss: 0.122 Val acc: 96.47%\n",
198 |       "Epoch 37 Train loss: 0.232 Val loss: 0.122 Val acc: 96.47%\n",
199 |       "Epoch 38 Train loss: 0.238 Val loss: 0.122 Val acc: 96.47%\n",
200 |       "Epoch 39 Train loss: 0.236 Val loss: 0.122 Val acc: 96.47%\n",
201 |       "Epoch 40 Train loss: 0.236 Val loss: 0.122 Val acc: 96.47%\n",
202 |       "Epoch 41 Train loss: 0.232 Val loss: 0.122 Val acc: 96.47%\n",
203 |       "Epoch 42 Train loss: 0.238 Val loss: 0.122 Val acc: 96.47%\n",
204 |       "Epoch 43 Train loss: 0.232 Val loss: 0.122 Val acc: 96.47%\n",
205 |       "Epoch 44 Train loss: 0.237 Val loss: 0.122 Val acc: 96.47%\n",
206 |       "Epoch 45 Train loss: 0.232 Val loss: 0.122 Val acc: 96.47%\n",
207 |       "Epoch 46 Train loss: 0.235 Val loss: 0.122 Val acc: 96.47%\n",
208 |       "Epoch 47 Train loss: 0.236 Val loss: 0.122 Val acc: 96.47%\n",
209 |       "Epoch 48 Train loss: 0.233 Val loss: 0.122 Val acc: 96.47%\n",
210 |       "Epoch 49 Train loss: 0.237 Val loss: 0.122 Val acc: 96.47%\n",
211 |       "Epoch 50 Train loss: 0.233 Val loss: 0.122 Val acc: 96.47%\n",
212 |       "Test loss: 0.122 Test acc: 96.47%\n",
213 |       "The number of parameters is: 24855\n",
214 |       "Epoch 1 Train loss: 0.397 Val loss: 0.158 Val acc: 95.30%\n",
215 |       "Epoch 2 Train loss: 0.284 Val loss: 0.135 Val acc: 96.00%\n",
216 |       "Epoch 3 Train loss: 0.261 Val loss: 0.125 Val acc: 96.25%\n",
217 |       "Epoch 4 Train loss: 0.261 Val loss: 0.126 Val acc: 96.20%\n",
218 |       "Epoch 5 Train loss: 0.245 Val loss: 0.124 Val acc: 96.02%\n",
219 |       "Epoch 6 Train loss: 0.242 Val loss: 0.116 Val acc: 96.37%\n",
220 |       "Epoch 7 Train loss: 0.238 Val loss: 0.122 Val acc: 96.37%\n",
221 |       "Epoch 8 Train loss: 0.232 Val loss: 0.118 Val acc: 96.62%\n",
222 |       "Epoch     8: reducing learning rate of group 0 to 1.0000e-03.\n",
223 |       "Epoch 9 Train loss: 0.233 Val loss: 0.129 Val acc: 96.33%\n",
224 |       "Epoch 10 Train loss: 0.173 Val loss: 0.100 Val acc: 97.12%\n",
225 |       "Epoch 11 Train loss: 0.150 Val loss: 0.098 Val acc: 97.28%\n",
226 |       "Epoch 12 Train loss: 0.148 Val loss: 0.096 Val acc: 97.12%\n",
227 |       "Epoch 13 Train loss: 0.145 Val loss: 0.095 Val acc: 97.22%\n",
228 |       "Epoch 14 Train loss: 0.142 Val loss: 0.092 Val acc: 97.27%\n",
229 |       "Epoch 15 Train loss: 0.140 Val loss: 0.092 Val acc: 97.17%\n",
230 |       "Epoch 16 Train loss: 0.137 Val loss: 0.092 Val acc: 97.25%\n",
231 |       "Epoch 17 Train loss: 0.139 Val loss: 0.090 Val acc: 97.25%\n",
232 |       "Epoch 18 Train loss: 0.137 Val loss: 0.089 Val acc: 97.30%\n",
233 |       "Epoch 19 Train loss: 0.133 Val loss: 0.091 Val acc: 97.27%\n",
234 |       "Epoch 20 Train loss: 0.130 Val loss: 0.087 Val acc: 97.28%\n",
235 |       "Epoch 21 Train loss: 0.130 Val loss: 0.087 Val acc: 97.27%\n",
236 |       "Epoch 22 Train loss: 0.130 Val loss: 0.089 Val acc: 97.20%\n",
237 |       "Epoch 23 Train loss: 0.129 Val loss: 0.086 Val acc: 97.30%\n",
238 |       "Epoch 24 Train loss: 0.131 Val loss: 0.088 Val acc: 97.27%\n",
239 |       "Epoch 25 Train loss: 0.126 Val loss: 0.087 Val acc: 97.33%\n",
240 |       "Epoch 26 Train loss: 0.123 Val loss: 0.086 Val acc: 97.38%\n",
241 |       "Epoch 27 Train loss: 0.126 Val loss: 0.085 Val acc: 97.37%\n",
242 |       "Epoch 28 Train loss: 0.127 Val loss: 0.085 Val acc: 97.35%\n",
243 |       "Epoch 29 Train loss: 0.125 Val loss: 0.085 Val acc: 97.37%\n",
244 |       "Epoch 30 Train loss: 0.125 Val loss: 0.084 Val acc: 97.37%\n",
245 |       "Epoch 31 Train loss: 0.125 Val loss: 0.083 Val acc: 97.47%\n",
246 |       "Epoch 32 Train loss: 0.121 Val loss: 0.082 Val acc: 97.55%\n",
247 |       "Epoch 33 Train loss: 0.124 Val loss: 0.085 Val acc: 97.33%\n",
248 |       "Epoch 34 Train loss: 0.122 Val loss: 0.084 Val acc: 97.35%\n",
249 |       "Epoch    34: reducing learning rate of group 0 to 1.0000e-04.\n",
250 |       "Epoch 35 Train loss: 0.118 Val loss: 0.084 Val acc: 97.45%\n",
251 |       "Epoch 36 Train loss: 0.118 Val loss: 0.083 Val acc: 97.40%\n",
252 |       "Epoch 37 Train loss: 0.116 Val loss: 0.082 Val acc: 97.38%\n",
253 |       "Epoch 38 Train loss: 0.116 Val loss: 0.081 Val acc: 97.45%\n",
254 |       "Epoch 39 Train loss: 0.116 Val loss: 0.081 Val acc: 97.42%\n",
255 |       "Epoch 40 Train loss: 0.117 Val loss: 0.082 Val acc: 97.40%\n",
256 |       "Epoch 41 Train loss: 0.117 Val loss: 0.082 Val acc: 97.42%\n",
257 |       "Epoch    41: reducing learning rate of group 0 to 1.0000e-05.\n",
258 |       "Epoch 42 Train loss: 0.119 Val loss: 0.081 Val acc: 97.47%\n",
259 |       "Epoch 43 Train loss: 0.117 Val loss: 0.081 Val acc: 97.47%\n",
260 |       "Epoch 44 Train loss: 0.118 Val loss: 0.081 Val acc: 97.47%\n",
261 |       "Epoch    44: reducing learning rate of group 0 to 1.0000e-06.\n",
262 |       "Epoch 45 Train loss: 0.116 Val loss: 0.081 Val acc: 97.47%\n",
263 |       "Epoch 46 Train loss: 0.117 Val loss: 0.081 Val acc: 97.47%\n",
264 |       "Epoch 47 Train loss: 0.117 Val loss: 0.081 Val acc: 97.47%\n",
265 |       "Epoch    47: reducing learning rate of group 0 to 1.0000e-07.\n",
266 |       "Epoch 48 Train loss: 0.115 Val loss: 0.081 Val acc: 97.47%\n",
267 |       "Epoch 49 Train loss: 0.115 Val loss: 0.081 Val acc: 97.47%\n",
268 |       "Epoch 50 Train loss: 0.116 Val loss: 0.081 Val acc: 97.47%\n",
269 |       "Test loss: 0.077 Test acc: 97.75%\n",
270 |       "Compression rate: 0.0625\n",
271 |       "The number of parameters is: 49300\n",
272 |       "Epoch 1 Train loss: 0.404 Val loss: 0.168 Val acc: 94.87%\n",
273 |       "Epoch 2 Train loss: 0.256 Val loss: 0.134 Val acc: 95.82%\n",
274 |       "Epoch 3 Train loss: 0.225 Val loss: 0.121 Val acc: 96.32%\n",
275 |       "Epoch 4 Train loss: 0.210 Val loss: 0.115 Val acc: 96.35%\n",
276 |       "Epoch 5 Train loss: 0.197 Val loss: 0.104 Val acc: 96.83%\n",
277 |       "Epoch 6 Train loss: 0.192 Val loss: 0.104 Val acc: 96.87%\n",
278 |       "Epoch 7 Train loss: 0.187 Val loss: 0.098 Val acc: 96.88%\n",
279 |       "Epoch 8 Train loss: 0.179 Val loss: 0.100 Val acc: 96.77%\n",
280 |       "Epoch 9 Train loss: 0.178 Val loss: 0.087 Val acc: 97.15%\n",
281 |       "Epoch 10 Train loss: 0.173 Val loss: 0.095 Val acc: 97.00%\n",
282 |       "Epoch 11 Train loss: 0.170 Val loss: 0.091 Val acc: 97.08%\n",
283 |       "Epoch    11: reducing learning rate of group 0 to 1.0000e-03.\n",
284 |       "Epoch 12 Train loss: 0.167 Val loss: 0.090 Val acc: 97.20%\n",
285 |       "Epoch 13 Train loss: 0.141 Val loss: 0.083 Val acc: 97.35%\n",
286 |       "Epoch 14 Train loss: 0.140 Val loss: 0.081 Val acc: 97.47%\n",
287 |       "Epoch 15 Train loss: 0.137 Val loss: 0.080 Val acc: 97.55%\n",
288 |       "Epoch 16 Train loss: 0.133 Val loss: 0.080 Val acc: 97.58%\n",
289 |       "Epoch 17 Train loss: 0.133 Val loss: 0.079 Val acc: 97.50%\n",
290 |       "Epoch 18 Train loss: 0.135 Val loss: 0.078 Val acc: 97.57%\n",
291 |       "Epoch 19 Train loss: 0.133 Val loss: 0.079 Val acc: 97.52%\n",
292 |       "Epoch 20 Train loss: 0.132 Val loss: 0.078 Val acc: 97.67%\n",
293 |       "Epoch 21 Train loss: 0.129 Val loss: 0.079 Val acc: 97.63%\n",
294 |       "Epoch 22 Train loss: 0.129 Val loss: 0.078 Val acc: 97.53%\n",
295 |       "Epoch 23 Train loss: 0.129 Val loss: 0.078 Val acc: 97.60%\n",
296 |       "Epoch 24 Train loss: 0.130 Val loss: 0.078 Val acc: 97.62%\n",
297 |       "Epoch 25 Train loss: 0.127 Val loss: 0.078 Val acc: 97.73%\n",
298 |       "Epoch 26 Train loss: 0.129 Val loss: 0.077 Val acc: 97.63%\n",
299 |       "Epoch 27 Train loss: 0.129 Val loss: 0.076 Val acc: 97.72%\n",
300 |       "Epoch 28 Train loss: 0.127 Val loss: 0.077 Val acc: 97.77%\n",
301 |       "Epoch 29 Train loss: 0.128 Val loss: 0.076 Val acc: 97.72%\n",
302 |       "Epoch    29: reducing learning rate of group 0 to 1.0000e-04.\n",
303 |       "Epoch 30 Train loss: 0.127 Val loss: 0.076 Val acc: 97.77%\n",
304 |       "Epoch 31 Train loss: 0.128 Val loss: 0.076 Val acc: 97.80%\n",
305 |       "Epoch 32 Train loss: 0.126 Val loss: 0.076 Val acc: 97.77%\n",
306 |       "Epoch 33 Train loss: 0.124 Val loss: 0.076 Val acc: 97.77%\n",
307 |       "Epoch 34 Train loss: 0.121 Val loss: 0.075 Val acc: 97.77%\n",
308 |       "Epoch 35 Train loss: 0.125 Val loss: 0.075 Val acc: 97.75%\n",
309 |       "Epoch 36 Train loss: 0.122 Val loss: 0.075 Val acc: 97.75%\n",
310 |       "Epoch 37 Train loss: 0.127 Val loss: 0.075 Val acc: 97.77%\n",
311 |       "Epoch 38 Train loss: 0.123 Val loss: 0.075 Val acc: 97.75%\n",
312 |       "Epoch 39 Train loss: 0.123 Val loss: 0.075 Val acc: 97.75%\n",
313 |       "Epoch 40 Train loss: 0.121 Val loss: 0.075 Val acc: 97.77%\n",
314 |       "Epoch    40: reducing learning rate of group 0 to 1.0000e-05.\n",
315 |       "Epoch 41 Train loss: 0.123 Val loss: 0.075 Val acc: 97.77%\n",
316 |       "Epoch 42 Train loss: 0.126 Val loss: 0.075 Val acc: 97.75%\n",
317 |       "Epoch 43 Train loss: 0.123 Val loss: 0.075 Val acc: 97.75%\n"
318 |      ]
319 |     },
320 |     {
321 |      "name": "stdout",
322 |      "output_type": "stream",
323 |      "text": [
324 |       "Epoch    43: reducing learning rate of group 0 to 1.0000e-06.\n",
325 |       "Epoch 44 Train loss: 0.125 Val loss: 0.075 Val acc: 97.73%\n",
326 |       "Epoch 45 Train loss: 0.123 Val loss: 0.075 Val acc: 97.73%\n",
327 |       "Epoch 46 Train loss: 0.125 Val loss: 0.075 Val acc: 97.73%\n",
328 |       "Epoch    46: reducing learning rate of group 0 to 1.0000e-07.\n",
329 |       "Epoch 47 Train loss: 0.124 Val loss: 0.075 Val acc: 97.73%\n",
330 |       "Epoch 48 Train loss: 0.123 Val loss: 0.075 Val acc: 97.73%\n",
331 |       "Epoch 49 Train loss: 0.123 Val loss: 0.075 Val acc: 97.73%\n",
332 |       "Epoch    49: reducing learning rate of group 0 to 1.0000e-08.\n",
333 |       "Epoch 50 Train loss: 0.123 Val loss: 0.075 Val acc: 97.73%\n",
334 |       "Test loss: 0.078 Test acc: 97.72%\n",
335 |       "The number of parameters is: 49698\n",
336 |       "Epoch 1 Train loss: 0.330 Val loss: 0.134 Val acc: 95.75%\n",
337 |       "Epoch 2 Train loss: 0.205 Val loss: 0.118 Val acc: 96.47%\n",
338 |       "Epoch 3 Train loss: 0.179 Val loss: 0.101 Val acc: 96.68%\n",
339 |       "Epoch 4 Train loss: 0.166 Val loss: 0.097 Val acc: 97.25%\n",
340 |       "Epoch 5 Train loss: 0.157 Val loss: 0.092 Val acc: 97.15%\n",
341 |       "Epoch 6 Train loss: 0.151 Val loss: 0.083 Val acc: 97.60%\n",
342 |       "Epoch 7 Train loss: 0.143 Val loss: 0.091 Val acc: 97.33%\n",
343 |       "Epoch 8 Train loss: 0.138 Val loss: 0.089 Val acc: 97.40%\n",
344 |       "Epoch 9 Train loss: 0.137 Val loss: 0.082 Val acc: 97.60%\n",
345 |       "Epoch 10 Train loss: 0.135 Val loss: 0.087 Val acc: 97.33%\n",
346 |       "Epoch 11 Train loss: 0.132 Val loss: 0.079 Val acc: 97.73%\n",
347 |       "Epoch 12 Train loss: 0.129 Val loss: 0.082 Val acc: 97.65%\n",
348 |       "Epoch 13 Train loss: 0.124 Val loss: 0.083 Val acc: 97.50%\n",
349 |       "Epoch    13: reducing learning rate of group 0 to 1.0000e-03.\n",
350 |       "Epoch 14 Train loss: 0.127 Val loss: 0.079 Val acc: 97.68%\n",
351 |       "Epoch 15 Train loss: 0.098 Val loss: 0.072 Val acc: 97.87%\n",
352 |       "Epoch 16 Train loss: 0.091 Val loss: 0.068 Val acc: 97.93%\n",
353 |       "Epoch 17 Train loss: 0.084 Val loss: 0.068 Val acc: 97.98%\n",
354 |       "Epoch 18 Train loss: 0.080 Val loss: 0.067 Val acc: 97.97%\n",
355 |       "Epoch 19 Train loss: 0.081 Val loss: 0.068 Val acc: 97.97%\n",
356 |       "Epoch 20 Train loss: 0.080 Val loss: 0.067 Val acc: 97.87%\n",
357 |       "Epoch 21 Train loss: 0.078 Val loss: 0.066 Val acc: 98.10%\n",
358 |       "Epoch 22 Train loss: 0.077 Val loss: 0.066 Val acc: 97.97%\n",
359 |       "Epoch 23 Train loss: 0.075 Val loss: 0.065 Val acc: 98.15%\n",
360 |       "Epoch 24 Train loss: 0.076 Val loss: 0.065 Val acc: 98.08%\n",
361 |       "Epoch 25 Train loss: 0.073 Val loss: 0.064 Val acc: 98.10%\n",
362 |       "Epoch 26 Train loss: 0.074 Val loss: 0.064 Val acc: 98.10%\n",
363 |       "Epoch 27 Train loss: 0.072 Val loss: 0.063 Val acc: 98.02%\n",
364 |       "Epoch 28 Train loss: 0.073 Val loss: 0.062 Val acc: 98.07%\n",
365 |       "Epoch 29 Train loss: 0.071 Val loss: 0.063 Val acc: 98.08%\n",
366 |       "Epoch 30 Train loss: 0.071 Val loss: 0.062 Val acc: 98.13%\n",
367 |       "Epoch    30: reducing learning rate of group 0 to 1.0000e-04.\n",
368 |       "Epoch 31 Train loss: 0.071 Val loss: 0.063 Val acc: 97.93%\n",
369 |       "Epoch 32 Train loss: 0.069 Val loss: 0.062 Val acc: 98.02%\n",
370 |       "Epoch 33 Train loss: 0.067 Val loss: 0.062 Val acc: 98.05%\n",
371 |       "Epoch 34 Train loss: 0.070 Val loss: 0.062 Val acc: 98.03%\n",
372 |       "Epoch 35 Train loss: 0.069 Val loss: 0.062 Val acc: 98.07%\n",
373 |       "Epoch 36 Train loss: 0.069 Val loss: 0.062 Val acc: 98.10%\n",
374 |       "Epoch 37 Train loss: 0.070 Val loss: 0.062 Val acc: 98.10%\n",
375 |       "Epoch 38 Train loss: 0.068 Val loss: 0.062 Val acc: 98.12%\n",
376 |       "Epoch 39 Train loss: 0.070 Val loss: 0.062 Val acc: 98.12%\n",
377 |       "Epoch 40 Train loss: 0.066 Val loss: 0.062 Val acc: 98.10%\n",
378 |       "Epoch 41 Train loss: 0.069 Val loss: 0.062 Val acc: 98.08%\n",
379 |       "Epoch 42 Train loss: 0.070 Val loss: 0.062 Val acc: 98.10%\n",
380 |       "Epoch    42: reducing learning rate of group 0 to 1.0000e-05.\n",
381 |       "Epoch 43 Train loss: 0.068 Val loss: 0.062 Val acc: 98.13%\n",
382 |       "Epoch 44 Train loss: 0.067 Val loss: 0.062 Val acc: 98.13%\n",
383 |       "Epoch 45 Train loss: 0.068 Val loss: 0.062 Val acc: 98.13%\n",
384 |       "Epoch    45: reducing learning rate of group 0 to 1.0000e-06.\n",
385 |       "Epoch 46 Train loss: 0.067 Val loss: 0.062 Val acc: 98.13%\n",
386 |       "Epoch 47 Train loss: 0.068 Val loss: 0.062 Val acc: 98.13%\n",
387 |       "Epoch 48 Train loss: 0.067 Val loss: 0.062 Val acc: 98.13%\n",
388 |       "Epoch    48: reducing learning rate of group 0 to 1.0000e-07.\n",
389 |       "Epoch 49 Train loss: 0.070 Val loss: 0.062 Val acc: 98.13%\n",
390 |       "Epoch 50 Train loss: 0.066 Val loss: 0.062 Val acc: 98.13%\n",
391 |       "Test loss: 0.065 Test acc: 98.16%\n",
392 |       "Compression rate: 0.125\n",
393 |       "The number of parameters is: 99385\n",
394 |       "Epoch 1 Train loss: 0.355 Val loss: 0.145 Val acc: 95.60%\n",
395 |       "Epoch 2 Train loss: 0.197 Val loss: 0.108 Val acc: 96.67%\n",
396 |       "Epoch 3 Train loss: 0.165 Val loss: 0.098 Val acc: 96.72%\n",
397 |       "Epoch 4 Train loss: 0.147 Val loss: 0.090 Val acc: 96.92%\n",
398 |       "Epoch 5 Train loss: 0.139 Val loss: 0.084 Val acc: 97.53%\n",
399 |       "Epoch 6 Train loss: 0.128 Val loss: 0.075 Val acc: 97.67%\n",
400 |       "Epoch 7 Train loss: 0.122 Val loss: 0.076 Val acc: 97.67%\n",
401 |       "Epoch 8 Train loss: 0.116 Val loss: 0.078 Val acc: 97.43%\n",
402 |       "Epoch 9 Train loss: 0.111 Val loss: 0.074 Val acc: 97.68%\n",
403 |       "Epoch 10 Train loss: 0.107 Val loss: 0.070 Val acc: 98.00%\n",
404 |       "Epoch 11 Train loss: 0.104 Val loss: 0.070 Val acc: 97.77%\n",
405 |       "Epoch 12 Train loss: 0.100 Val loss: 0.065 Val acc: 98.05%\n",
406 |       "Epoch 13 Train loss: 0.098 Val loss: 0.071 Val acc: 97.80%\n",
407 |       "Epoch 14 Train loss: 0.096 Val loss: 0.065 Val acc: 97.83%\n",
408 |       "Epoch 15 Train loss: 0.094 Val loss: 0.067 Val acc: 97.88%\n",
409 |       "Epoch 16 Train loss: 0.092 Val loss: 0.064 Val acc: 97.97%\n",
410 |       "Epoch 17 Train loss: 0.091 Val loss: 0.067 Val acc: 97.83%\n",
411 |       "Epoch 18 Train loss: 0.089 Val loss: 0.068 Val acc: 97.80%\n",
412 |       "Epoch    18: reducing learning rate of group 0 to 1.0000e-03.\n",
413 |       "Epoch 19 Train loss: 0.090 Val loss: 0.070 Val acc: 97.77%\n",
414 |       "Epoch 20 Train loss: 0.077 Val loss: 0.065 Val acc: 97.97%\n",
415 |       "Epoch 21 Train loss: 0.071 Val loss: 0.063 Val acc: 98.08%\n",
416 |       "Epoch 22 Train loss: 0.069 Val loss: 0.062 Val acc: 98.12%\n",
417 |       "Epoch 23 Train loss: 0.066 Val loss: 0.062 Val acc: 98.13%\n",
418 |       "Epoch 24 Train loss: 0.065 Val loss: 0.061 Val acc: 98.20%\n",
419 |       "Epoch 25 Train loss: 0.065 Val loss: 0.060 Val acc: 98.20%\n",
420 |       "Epoch 26 Train loss: 0.064 Val loss: 0.060 Val acc: 98.23%\n",
421 |       "Epoch 27 Train loss: 0.064 Val loss: 0.060 Val acc: 98.23%\n",
422 |       "Epoch    27: reducing learning rate of group 0 to 1.0000e-04.\n",
423 |       "Epoch 28 Train loss: 0.064 Val loss: 0.061 Val acc: 98.15%\n",
424 |       "Epoch 29 Train loss: 0.064 Val loss: 0.060 Val acc: 98.12%\n",
425 |       "Epoch 30 Train loss: 0.063 Val loss: 0.060 Val acc: 98.13%\n",
426 |       "Epoch    30: reducing learning rate of group 0 to 1.0000e-05.\n",
427 |       "Epoch 31 Train loss: 0.061 Val loss: 0.060 Val acc: 98.10%\n",
428 |       "Epoch 32 Train loss: 0.061 Val loss: 0.060 Val acc: 98.10%\n",
429 |       "Epoch 33 Train loss: 0.063 Val loss: 0.060 Val acc: 98.10%\n",
430 |       "Epoch    33: reducing learning rate of group 0 to 1.0000e-06.\n",
431 |       "Epoch 34 Train loss: 0.061 Val loss: 0.060 Val acc: 98.12%\n",
432 |       "Epoch 35 Train loss: 0.062 Val loss: 0.060 Val acc: 98.12%\n",
433 |       "Epoch 36 Train loss: 0.062 Val loss: 0.060 Val acc: 98.12%\n",
434 |       "Epoch    36: reducing learning rate of group 0 to 1.0000e-07.\n",
435 |       "Epoch 37 Train loss: 0.063 Val loss: 0.060 Val acc: 98.12%\n",
436 |       "Epoch 38 Train loss: 0.063 Val loss: 0.060 Val acc: 98.12%\n",
437 |       "Epoch 39 Train loss: 0.063 Val loss: 0.060 Val acc: 98.12%\n",
438 |       "Epoch    39: reducing learning rate of group 0 to 1.0000e-08.\n",
439 |       "Epoch 40 Train loss: 0.062 Val loss: 0.060 Val acc: 98.12%\n",
440 |       "Epoch 41 Train loss: 0.061 Val loss: 0.060 Val acc: 98.12%\n",
441 |       "Epoch 42 Train loss: 0.063 Val loss: 0.060 Val acc: 98.12%\n",
442 |       "Epoch 43 Train loss: 0.064 Val loss: 0.060 Val acc: 98.12%\n",
443 |       "Epoch 44 Train loss: 0.061 Val loss: 0.060 Val acc: 98.12%\n",
444 |       "Epoch 45 Train loss: 0.062 Val loss: 0.060 Val acc: 98.12%\n",
445 |       "Epoch 46 Train loss: 0.060 Val loss: 0.060 Val acc: 98.12%\n",
446 |       "Epoch 47 Train loss: 0.062 Val loss: 0.060 Val acc: 98.12%\n",
447 |       "Epoch 48 Train loss: 0.062 Val loss: 0.060 Val acc: 98.12%\n",
448 |       "Epoch 49 Train loss: 0.062 Val loss: 0.060 Val acc: 98.12%\n",
449 |       "Epoch 50 Train loss: 0.062 Val loss: 0.060 Val acc: 98.12%\n",
450 |       "Test loss: 0.057 Test acc: 98.25%\n",
451 |       "The number of parameters is: 99385\n",
452 |       "Epoch 1 Train loss: 0.303 Val loss: 0.125 Val acc: 96.20%\n",
453 |       "Epoch 2 Train loss: 0.169 Val loss: 0.097 Val acc: 97.15%\n",
454 |       "Epoch 3 Train loss: 0.139 Val loss: 0.086 Val acc: 97.38%\n",
455 |       "Epoch 4 Train loss: 0.123 Val loss: 0.092 Val acc: 97.22%\n",
456 |       "Epoch 5 Train loss: 0.113 Val loss: 0.084 Val acc: 97.17%\n",
457 |       "Epoch 6 Train loss: 0.105 Val loss: 0.083 Val acc: 97.67%\n",
458 |       "Epoch 7 Train loss: 0.102 Val loss: 0.077 Val acc: 97.75%\n",
459 |       "Epoch 8 Train loss: 0.095 Val loss: 0.078 Val acc: 97.63%\n",
460 |       "Epoch 9 Train loss: 0.087 Val loss: 0.070 Val acc: 98.02%\n",
461 |       "Epoch 10 Train loss: 0.087 Val loss: 0.068 Val acc: 97.88%\n",
462 |       "Epoch 11 Train loss: 0.087 Val loss: 0.074 Val acc: 98.03%\n",
463 |       "Epoch 12 Train loss: 0.080 Val loss: 0.069 Val acc: 98.02%\n",
464 |       "Epoch    12: reducing learning rate of group 0 to 1.0000e-03.\n",
465 |       "Epoch 13 Train loss: 0.077 Val loss: 0.070 Val acc: 97.90%\n"
466 |      ]
467 |     },
468 |     {
469 |      "name": "stdout",
470 |      "output_type": "stream",
471 |      "text": [
472 |       "Epoch 14 Train loss: 0.065 Val loss: 0.065 Val acc: 98.20%\n",
473 |       "Epoch 15 Train loss: 0.055 Val loss: 0.064 Val acc: 98.27%\n",
474 |       "Epoch 16 Train loss: 0.052 Val loss: 0.064 Val acc: 98.22%\n",
475 |       "Epoch 17 Train loss: 0.051 Val loss: 0.063 Val acc: 98.27%\n",
476 |       "Epoch 18 Train loss: 0.047 Val loss: 0.062 Val acc: 98.25%\n",
477 |       "Epoch 19 Train loss: 0.047 Val loss: 0.063 Val acc: 98.25%\n",
478 |       "Epoch 20 Train loss: 0.048 Val loss: 0.062 Val acc: 98.25%\n",
479 |       "Epoch 21 Train loss: 0.045 Val loss: 0.062 Val acc: 98.32%\n",
480 |       "Epoch 22 Train loss: 0.046 Val loss: 0.061 Val acc: 98.22%\n",
481 |       "Epoch 23 Train loss: 0.045 Val loss: 0.061 Val acc: 98.32%\n",
482 |       "Epoch 24 Train loss: 0.045 Val loss: 0.061 Val acc: 98.27%\n",
483 |       "Epoch 25 Train loss: 0.045 Val loss: 0.060 Val acc: 98.35%\n",
484 |       "Epoch 26 Train loss: 0.044 Val loss: 0.059 Val acc: 98.28%\n",
485 |       "Epoch 27 Train loss: 0.042 Val loss: 0.060 Val acc: 98.22%\n",
486 |       "Epoch 28 Train loss: 0.041 Val loss: 0.060 Val acc: 98.30%\n",
487 |       "Epoch    28: reducing learning rate of group 0 to 1.0000e-04.\n",
488 |       "Epoch 29 Train loss: 0.042 Val loss: 0.061 Val acc: 98.32%\n",
489 |       "Epoch 30 Train loss: 0.041 Val loss: 0.060 Val acc: 98.35%\n",
490 |       "Epoch 31 Train loss: 0.041 Val loss: 0.060 Val acc: 98.33%\n",
491 |       "Epoch    31: reducing learning rate of group 0 to 1.0000e-05.\n",
492 |       "Epoch 32 Train loss: 0.042 Val loss: 0.060 Val acc: 98.33%\n",
493 |       "Epoch 33 Train loss: 0.042 Val loss: 0.060 Val acc: 98.33%\n",
494 |       "Epoch 34 Train loss: 0.040 Val loss: 0.060 Val acc: 98.33%\n",
495 |       "Epoch    34: reducing learning rate of group 0 to 1.0000e-06.\n",
496 |       "Epoch 35 Train loss: 0.041 Val loss: 0.060 Val acc: 98.33%\n",
497 |       "Epoch 36 Train loss: 0.038 Val loss: 0.060 Val acc: 98.33%\n",
498 |       "Epoch 37 Train loss: 0.041 Val loss: 0.060 Val acc: 98.33%\n",
499 |       "Epoch    37: reducing learning rate of group 0 to 1.0000e-07.\n",
500 |       "Epoch 38 Train loss: 0.039 Val loss: 0.060 Val acc: 98.33%\n",
501 |       "Epoch 39 Train loss: 0.041 Val loss: 0.060 Val acc: 98.33%\n",
502 |       "Epoch 40 Train loss: 0.040 Val loss: 0.060 Val acc: 98.33%\n",
503 |       "Epoch    40: reducing learning rate of group 0 to 1.0000e-08.\n",
504 |       "Epoch 41 Train loss: 0.041 Val loss: 0.060 Val acc: 98.33%\n",
505 |       "Epoch 42 Train loss: 0.042 Val loss: 0.060 Val acc: 98.33%\n",
506 |       "Epoch 43 Train loss: 0.042 Val loss: 0.060 Val acc: 98.33%\n",
507 |       "Epoch 44 Train loss: 0.039 Val loss: 0.060 Val acc: 98.33%\n",
508 |       "Epoch 45 Train loss: 0.041 Val loss: 0.060 Val acc: 98.33%\n",
509 |       "Epoch 46 Train loss: 0.041 Val loss: 0.060 Val acc: 98.33%\n",
510 |       "Epoch 47 Train loss: 0.041 Val loss: 0.060 Val acc: 98.33%\n",
511 |       "Epoch 48 Train loss: 0.042 Val loss: 0.060 Val acc: 98.33%\n",
512 |       "Epoch 49 Train loss: 0.040 Val loss: 0.060 Val acc: 98.33%\n",
513 |       "Epoch 50 Train loss: 0.039 Val loss: 0.060 Val acc: 98.33%\n",
514 |       "Test loss: 0.053 Test acc: 98.48%\n",
515 |       "Compression rate: 1\n",
516 |       "The number of parameters is: 795010\n",
517 |       "Epoch 1 Train loss: 0.294 Val loss: 0.132 Val acc: 95.90%\n",
518 |       "Epoch 2 Train loss: 0.140 Val loss: 0.094 Val acc: 97.07%\n",
519 |       "Epoch 3 Train loss: 0.106 Val loss: 0.075 Val acc: 97.63%\n",
520 |       "Epoch 4 Train loss: 0.087 Val loss: 0.072 Val acc: 97.95%\n",
521 |       "Epoch 5 Train loss: 0.077 Val loss: 0.065 Val acc: 98.02%\n",
522 |       "Epoch 6 Train loss: 0.065 Val loss: 0.067 Val acc: 98.00%\n",
523 |       "Epoch 7 Train loss: 0.059 Val loss: 0.061 Val acc: 98.12%\n",
524 |       "Epoch 8 Train loss: 0.055 Val loss: 0.057 Val acc: 98.12%\n",
525 |       "Epoch 9 Train loss: 0.048 Val loss: 0.057 Val acc: 98.25%\n",
526 |       "Epoch 10 Train loss: 0.045 Val loss: 0.059 Val acc: 98.18%\n",
527 |       "Epoch 11 Train loss: 0.041 Val loss: 0.054 Val acc: 98.32%\n",
528 |       "Epoch 12 Train loss: 0.037 Val loss: 0.053 Val acc: 98.40%\n",
529 |       "Epoch 13 Train loss: 0.035 Val loss: 0.057 Val acc: 98.35%\n",
530 |       "Epoch 14 Train loss: 0.033 Val loss: 0.057 Val acc: 98.20%\n",
531 |       "Epoch 15 Train loss: 0.034 Val loss: 0.053 Val acc: 98.53%\n",
532 |       "Epoch 16 Train loss: 0.030 Val loss: 0.051 Val acc: 98.38%\n",
533 |       "Epoch 17 Train loss: 0.029 Val loss: 0.048 Val acc: 98.48%\n",
534 |       "Epoch 18 Train loss: 0.027 Val loss: 0.052 Val acc: 98.50%\n",
535 |       "Epoch 19 Train loss: 0.026 Val loss: 0.055 Val acc: 98.33%\n",
536 |       "Epoch    19: reducing learning rate of group 0 to 1.0000e-03.\n",
537 |       "Epoch 20 Train loss: 0.025 Val loss: 0.057 Val acc: 98.45%\n",
538 |       "Epoch 21 Train loss: 0.020 Val loss: 0.052 Val acc: 98.57%\n",
539 |       "Epoch 22 Train loss: 0.018 Val loss: 0.050 Val acc: 98.70%\n",
540 |       "Epoch    22: reducing learning rate of group 0 to 1.0000e-04.\n",
541 |       "Epoch 23 Train loss: 0.018 Val loss: 0.049 Val acc: 98.73%\n",
542 |       "Epoch 24 Train loss: 0.016 Val loss: 0.049 Val acc: 98.75%\n",
543 |       "Epoch 25 Train loss: 0.016 Val loss: 0.048 Val acc: 98.75%\n",
544 |       "Epoch 26 Train loss: 0.016 Val loss: 0.048 Val acc: 98.73%\n",
545 |       "Epoch 27 Train loss: 0.016 Val loss: 0.048 Val acc: 98.73%\n",
546 |       "Epoch 28 Train loss: 0.016 Val loss: 0.048 Val acc: 98.73%\n",
547 |       "Epoch 29 Train loss: 0.016 Val loss: 0.048 Val acc: 98.75%\n",
548 |       "Epoch 30 Train loss: 0.015 Val loss: 0.048 Val acc: 98.73%\n",
549 |       "Epoch 31 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
550 |       "Epoch    31: reducing learning rate of group 0 to 1.0000e-05.\n",
551 |       "Epoch 32 Train loss: 0.014 Val loss: 0.048 Val acc: 98.73%\n",
552 |       "Epoch 33 Train loss: 0.015 Val loss: 0.048 Val acc: 98.73%\n",
553 |       "Epoch 34 Train loss: 0.015 Val loss: 0.048 Val acc: 98.73%\n",
554 |       "Epoch    34: reducing learning rate of group 0 to 1.0000e-06.\n",
555 |       "Epoch 35 Train loss: 0.015 Val loss: 0.048 Val acc: 98.72%\n",
556 |       "Epoch 36 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
557 |       "Epoch 37 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
558 |       "Epoch    37: reducing learning rate of group 0 to 1.0000e-07.\n",
559 |       "Epoch 38 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
560 |       "Epoch 39 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
561 |       "Epoch 40 Train loss: 0.015 Val loss: 0.048 Val acc: 98.72%\n",
562 |       "Epoch    40: reducing learning rate of group 0 to 1.0000e-08.\n",
563 |       "Epoch 41 Train loss: 0.017 Val loss: 0.048 Val acc: 98.72%\n",
564 |       "Epoch 42 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
565 |       "Epoch 43 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
566 |       "Epoch 44 Train loss: 0.015 Val loss: 0.048 Val acc: 98.72%\n",
567 |       "Epoch 45 Train loss: 0.017 Val loss: 0.048 Val acc: 98.72%\n",
568 |       "Epoch 46 Train loss: 0.015 Val loss: 0.048 Val acc: 98.72%\n",
569 |       "Epoch 47 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
570 |       "Epoch 48 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
571 |       "Epoch 49 Train loss: 0.016 Val loss: 0.048 Val acc: 98.72%\n",
572 |       "Epoch 50 Train loss: 0.015 Val loss: 0.048 Val acc: 98.72%\n",
573 |       "Test loss: 0.045 Test acc: 98.73%\n",
574 |       "The number of parameters is: 795010\n",
575 |       "Epoch 1 Train loss: 0.289 Val loss: 0.122 Val acc: 96.62%\n",
576 |       "Epoch 2 Train loss: 0.142 Val loss: 0.088 Val acc: 97.23%\n",
577 |       "Epoch 3 Train loss: 0.110 Val loss: 0.080 Val acc: 97.37%\n",
578 |       "Epoch 4 Train loss: 0.090 Val loss: 0.075 Val acc: 97.57%\n",
579 |       "Epoch 5 Train loss: 0.078 Val loss: 0.066 Val acc: 97.95%\n",
580 |       "Epoch 6 Train loss: 0.069 Val loss: 0.065 Val acc: 97.87%\n",
581 |       "Epoch 7 Train loss: 0.062 Val loss: 0.062 Val acc: 98.07%\n",
582 |       "Epoch 8 Train loss: 0.055 Val loss: 0.060 Val acc: 98.13%\n",
583 |       "Epoch 9 Train loss: 0.055 Val loss: 0.058 Val acc: 98.25%\n",
584 |       "Epoch 10 Train loss: 0.048 Val loss: 0.058 Val acc: 98.45%\n",
585 |       "Epoch 11 Train loss: 0.045 Val loss: 0.058 Val acc: 98.22%\n",
586 |       "Epoch    11: reducing learning rate of group 0 to 1.0000e-03.\n",
587 |       "Epoch 12 Train loss: 0.041 Val loss: 0.059 Val acc: 98.23%\n",
588 |       "Epoch 13 Train loss: 0.032 Val loss: 0.053 Val acc: 98.50%\n",
589 |       "Epoch 14 Train loss: 0.031 Val loss: 0.052 Val acc: 98.47%\n",
590 |       "Epoch 15 Train loss: 0.027 Val loss: 0.051 Val acc: 98.55%\n",
591 |       "Epoch 16 Train loss: 0.025 Val loss: 0.051 Val acc: 98.52%\n",
592 |       "Epoch 17 Train loss: 0.024 Val loss: 0.050 Val acc: 98.45%\n",
593 |       "Epoch 18 Train loss: 0.024 Val loss: 0.049 Val acc: 98.53%\n",
594 |       "Epoch 19 Train loss: 0.023 Val loss: 0.049 Val acc: 98.58%\n",
595 |       "Epoch 20 Train loss: 0.023 Val loss: 0.049 Val acc: 98.50%\n",
596 |       "Epoch 21 Train loss: 0.022 Val loss: 0.050 Val acc: 98.52%\n",
597 |       "Epoch    21: reducing learning rate of group 0 to 1.0000e-04.\n",
598 |       "Epoch 22 Train loss: 0.023 Val loss: 0.049 Val acc: 98.60%\n",
599 |       "Epoch 23 Train loss: 0.021 Val loss: 0.049 Val acc: 98.60%\n",
600 |       "Epoch 24 Train loss: 0.022 Val loss: 0.049 Val acc: 98.60%\n",
601 |       "Epoch    24: reducing learning rate of group 0 to 1.0000e-05.\n",
602 |       "Epoch 25 Train loss: 0.021 Val loss: 0.049 Val acc: 98.60%\n",
603 |       "Epoch 26 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
604 |       "Epoch 27 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
605 |       "Epoch    27: reducing learning rate of group 0 to 1.0000e-06.\n",
606 |       "Epoch 28 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
607 |       "Epoch 29 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
608 |       "Epoch 30 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
609 |       "Epoch    30: reducing learning rate of group 0 to 1.0000e-07.\n",
610 |       "Epoch 31 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
611 |       "Epoch 32 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n"
612 |      ]
613 |     },
614 |     {
615 |      "name": "stdout",
616 |      "output_type": "stream",
617 |      "text": [
618 |       "Epoch 33 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
619 |       "Epoch    33: reducing learning rate of group 0 to 1.0000e-08.\n",
620 |       "Epoch 34 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
621 |       "Epoch 35 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
622 |       "Epoch 36 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
623 |       "Epoch 37 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
624 |       "Epoch 38 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
625 |       "Epoch 39 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
626 |       "Epoch 40 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
627 |       "Epoch 41 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
628 |       "Epoch 42 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
629 |       "Epoch 43 Train loss: 0.023 Val loss: 0.049 Val acc: 98.62%\n",
630 |       "Epoch 44 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
631 |       "Epoch 45 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
632 |       "Epoch 46 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
633 |       "Epoch 47 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
634 |       "Epoch 48 Train loss: 0.022 Val loss: 0.049 Val acc: 98.62%\n",
635 |       "Epoch 49 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
636 |       "Epoch 50 Train loss: 0.021 Val loss: 0.049 Val acc: 98.62%\n",
637 |       "Test loss: 0.048 Test acc: 98.54%\n"
638 |      ]
639 |     }
640 |    ],
641 |    "source": [
642 |     "device = torch.device(\"cuda\" if use_cuda else \"cpu\")\n",
643 |     "kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}\n",
644 |     "train_loader, valid_loader, test_loader = load_data(batch_size=50, kwargs=kwargs)\n",
645 |     "\n",
646 |     "def train_nn(compress, hashed):\n",
647 |     "    input_dim = 784\n",
648 |     "    output_dim = 10\n",
649 |     "    \n",
650 |     "    if hashed:\n",
651 |     "        model = HashedNet(input_dim, output_dim, 1, 1000,\n",
652 |     "                          compress, dropout=0.25).to(device)\n",
653 |     "    else:\n",
654 |     "        eq_compress = get_equivalent_compression(input_dim, output_dim,\n",
655 |     "                                                 1000, 1, compress)\n",
656 |     "        model = Net(input_dim, output_dim, 1, 1000,\n",
657 |     "                    eq_compress, 0.25).to(device)\n",
658 |     "\n",
659 |     "    optimizer = optim.SGD(model.parameters(), lr=0.01,\n",
660 |     "                          momentum=0.9,\n",
661 |     "                          weight_decay=0.0)\n",
662 |     "\n",
663 |     "    scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,\n",
664 |     "                                                     factor=0.1,\n",
665 |     "                                                     patience=2,\n",
666 |     "                                                     verbose=True)\n",
667 |     "\n",
668 |     "    print('The number of parameters is: {}'.format(\n",
669 |     "        sum(p.numel() for p in model.parameters() if p.requires_grad)))\n",
670 |     "\n",
671 |     "    for epoch in range(1, 50 + 1):\n",
672 |     "        tr_loss = train(model, device, train_loader, optimizer, epoch, log_interval=50)\n",
673 |     "        val_loss, val_acc = evaluate(model, device, valid_loader)\n",
674 |     "        scheduler.step(val_loss)\n",
675 |     "        print('Epoch {} Train loss: {:.3f} Val loss: {:.3f} Val acc: {:.2f}%'.format(\n",
676 |     "              epoch, tr_loss, val_loss, val_acc))\n",
677 |     "\n",
678 |     "    test_loss, test_acc = evaluate(model, device, test_loader)\n",
679 |     "    print('Test loss: {:.3f} Test acc: {:.2f}%'.format(test_loss, test_acc))\n",
680 |     "    \n",
681 |     "    return test_loss, test_acc\n",
682 |     "\n",
683 |     "compression_rates = [1/64, 1/32, 1/16, 1/8, 1]\n",
684 |     "nn_records = []\n",
685 |     "hashednn_records = []\n",
686 |     "for compression in compression_rates:\n",
687 |     "    print(\"Compression rate: {}\".format(compression))\n",
688 |     "    nn_records.append(train_nn(compression, hashed=False))\n",
689 |     "    hashednn_records.append(train_nn(compression, hashed=True))"
690 |    ]
691 |   },
692 |   {
693 |    "cell_type": "code",
694 |    "execution_count": 43,
695 |    "metadata": {},
696 |    "outputs": [
697 |     {
698 |      "data": {
699 |       "image/png": "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\n",
700 |       "text/plain": [
701 |        "<Figure size 432x288 with 1 Axes>"
702 |       ]
703 |      },
704 |      "metadata": {
705 |       "needs_background": "light"
706 |      },
707 |      "output_type": "display_data"
708 |     }
709 |    ],
710 |    "source": [
711 |     "import matplotlib.pyplot as plt\n",
712 |     "\n",
713 |     "fig, ax = plt.subplots(figsize=(6,4))\n",
714 |     "ax.plot([100.0-x[1] for x in hashednn_records], 's-', linewidth=2, label='Hashed NN')\n",
715 |     "ax.plot([100.0-x[1] for x in nn_records], 'd--', linewidth=2, label='Equiv. NN')\n",
716 |     "ax.set_xticklabels(['1/64', '1/32', '1/16', '1/8', '1'])\n",
717 |     "ax.set_xticks(range(len(compression_rates)))\n",
718 |     "ax.set_xlabel('Compression ratio')\n",
719 |     "ax.set_ylabel('Error (%)')\n",
720 |     "plt.title(\"1 hidden layer, 1000 units\")\n",
721 |     "plt.legend(handlelength=3)\n",
722 |     "plt.grid()\n",
723 |     "plt.savefig('example.svg')\n",
724 |     "plt.show()"
725 |    ]
726 |   }
727 |  ],
728 |  "metadata": {
729 |   "kernelspec": {
730 |    "display_name": "Python [conda env:pytorch3.7] *",
731 |    "language": "python",
732 |    "name": "conda-env-pytorch3.7-py"
733 |   },
734 |   "language_info": {
735 |    "codemirror_mode": {
736 |     "name": "ipython",
737 |     "version": 3
738 |    },
739 |    "file_extension": ".py",
740 |    "mimetype": "text/x-python",
741 |    "name": "python",
742 |    "nbconvert_exporter": "python",
743 |    "pygments_lexer": "ipython3",
744 |    "version": "3.6.8"
745 |   }
746 |  },
747 |  "nbformat": 4,
748 |  "nbformat_minor": 2
749 | }
750 | 


--------------------------------------------------------------------------------