/
utils.py
318 lines (258 loc) · 16.4 KB
/
utils.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
# -*- coding: utf-8 -*-
#@title imports & settings { form-width: "10%" }
import torch
import torchvision
import matplotlib.pyplot as plt
from matplotlib import lines as plt_lines
from matplotlib import markers as plt_markers
#@title Data load { form-width: "10%" }
from model import Network
def data_load(batch_size_train = 1000 , batch_size_test = 1000):
trainset = torchvision.datasets.MNIST( 'files/', train=True, download=True,
transform=torchvision.transforms.Compose([
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(
(0.1307,), (0.3081,))
]))
if batch_size_train <= 1001:
trainset = torch.utils.data.Subset(trainset, list(range(0, batch_size_train*50,50)))
else:
trainset = torch.utils.data.Subset(trainset, list(range(0, batch_size_train)))
train_loader = torch.utils.data.DataLoader(trainset, batch_size=batch_size_train, shuffle=False, num_workers = 0)
test_loader = torch.utils.data.DataLoader(
torchvision.datasets.MNIST( 'files/', train=False, download=True,
transform=torchvision.transforms.Compose([
torchvision.transforms.ToTensor(),
torchvision.transforms.Normalize(
(0.1307,), (0.3081,))
])),
batch_size=batch_size_test, shuffle=True)
return train_loader,test_loader
#@title neural network { form-width: "10%" }
#@title graphs { form-width: "10%" }
def plot_networks_loss_graph(base_network , networks_to_compare = [], download = False, networks_type_str = 'Regular', scale = 'regular'):
#figure settings
#plt.rcParams.update({'font.size': 15})
plt.rcParams.update({'legend.fontsize': 9})
plt.figure(figsize=(3.2, 3))
plt.ylabel('Training Loss')
plt.xlabel('$\eta_{0}$ Iterations')
#plot
plt.plot(base_network.train_time_stamps, base_network.train_losses, label='$\eta_{0}$')
for i,network in enumerate(networks_to_compare):
plt.plot(base_network.train_time_stamps, network.train_losses, label='$\eta_{0}/$'+str(network.lr_ratio_from_base))
plt.legend()
if scale == 'log':
plt.yscale('log')
if networks_type_str == 'Linear':
plt.title('Fully Connected, Linear Activation', pad=10)
if networks_type_str == 'ReLU':
plt.title('Fully Connected, ReLU Activation', pad=10)
if networks_type_str == 'conv_subsample':
plt.title('Convolutional, ReLU Activation, No Pooling', pad=10)
if networks_type_str == 'conv_max_pool':
plt.title('Convolutional, ReLU Activation, Max Pooling', pad=10)
plt.xlim(left=0,right=base_network.train_time_stamps[-1])
#plt.tight_layout(pad=0.05)
plt.tight_layout()
if download == True:
if scale == 'log':
plt.savefig('LossGraph'+networks_type_str+'_log_scale.pdf',bbox_inches = "tight")
else:
plt.savefig('LossGraph'+networks_type_str+'.pdf',bbox_inches = "tight")
#files.download('LossGraph'+networks_type_str+'.pdf')
plt.show()
def plot_networks_distance_graph(base_network, networks_to_compare = [], download = False, networks_type_str = 'Regular', scale = 'regular'):
#figure settings
#plt.rcParams.update({'font.size': 15})
plt.rcParams.update({'legend.fontsize': 9})
plt.figure(figsize=(3.2, 3))
plt.ylabel('Distance')
plt.xlabel('$\eta_{0}$ Iterations')
#initialize lists for plot
base_network_norms = []
networks_distance = []
for compare_network in networks_to_compare:
networks_distance.append([])
#insert values into lists
for i,train_weight in enumerate(base_network.train_weights):
base_network_norms.append(torch.norm(train_weight - base_network.train_weights[0]).cpu().detach().numpy())
for j,compare_network in enumerate(networks_to_compare):
networks_distance[j].append(torch.norm(base_network.train_weights[i]-compare_network.train_weights[i]).cpu().detach().numpy())
#plot
plt.plot(base_network.train_time_stamps[0:],base_network_norms[0:],label='$\eta_{0}$ from init')
for j,compare_network in enumerate(networks_to_compare):
plt.plot(base_network.train_time_stamps[0:],networks_distance[j][0:],label='$\eta_{0}$ from $\eta_{0}/$'+str(compare_network.lr_ratio_from_base))
plt.legend()
if scale == 'log':
plt.yscale('log')
if networks_type_str == 'Linear':
plt.title('Fully Connected, Linear Activation', pad=10)
if networks_type_str == 'ReLU':
plt.title('Fully Connected, ReLU Activation', pad=10)
if networks_type_str == 'conv_subsample':
plt.title('Convolutional, ReLU Activation, No Pooling', pad=10)
if networks_type_str == 'conv_max_pool':
plt.title('Convolutional, ReLU Activation, Max Pooling', pad=10)
plt.xlim(left=0,right=base_network.train_time_stamps[-1])
#plt.tight_layout(pad=0.05)
plt.tight_layout()
if download == True:
if scale == 'log':
plt.savefig('DistGraph'+networks_type_str+'_log_scale.pdf',bbox_inches = "tight")
else:
plt.savefig('DistGraph'+networks_type_str+'.pdf',bbox_inches = "tight")
#files.download('DistGraph'+networks_type_str+'.pdf')
plt.show()
def set_networks_loss_subgraph(base_network , subgraph, networks_to_compare = [], download = False, networks_type_str = 'Regular', scale = 'regular',axis_size=14.5):
subgraph.set_ylabel('Training Loss',fontsize=axis_size)
subgraph.set_xlabel('$\eta_{0}$ Iterations',fontsize=axis_size)
lines = list(plt_lines.lineStyles.keys())
markers = list(plt_markers.MarkerStyle.markers.keys())
#lines = [lines[i] for i in [0,1,3]]
lines = [lines[0],lines[0],lines[1],lines[1],lines[3]]
markers = [markers[i] for i in [2, 3, 4, 12, 14, 21, 24]]
markers = [markers[3],markers[0],markers[1],markers[2],markers[4],markers[5],markers[6]]
#plot
subgraph.plot(base_network.train_time_stamps, base_network.train_losses, label='$\eta_{0}$',linestyle = (0,(1,1)), marker=markers[0],markevery=10,markersize=7)#linewidth = 3
for i,network in enumerate(networks_to_compare):
subgraph.plot(base_network.train_time_stamps, network.train_losses, label='$\eta_{0}/$'+str(network.lr_ratio_from_base),linestyle = (0.25*(i+1),(1,1)), marker=markers[(1+i)%len(markers)],markevery=10)
subgraph.legend()
if scale == 'log':
subgraph.set_yscale('log')
subgraph.set_xlim(left=0,right=base_network.train_time_stamps[-1])
subgraph.set_ylim(bottom = 0)
return subgraph
def set_networks_distance_subgraph(base_network, subgraph, networks_to_compare = [], download = False, networks_type_str = 'Regular', scale = 'regular',axis_size=14.5):
subgraph.set_ylabel('Distance',fontsize=axis_size)
subgraph.set_xlabel('$\eta_{0}$ Iterations',fontsize=axis_size)
lines = list(plt_lines.lineStyles.keys())
lines = [lines[i] for i in [0,1,3]]
markers = list(plt_markers.MarkerStyle.markers.keys())
markers = [markers[i] for i in [2, 3, 4, 12, 14, 21, 24]]
markers = [markers[3],markers[0],markers[1],markers[2],markers[4],markers[5],markers[6]]
#initialize lists for plot
base_network_norms = []
networks_distance = []
for compare_network in networks_to_compare:
networks_distance.append([])
#insert values into lists
for i,train_weight in enumerate(base_network.train_weights):
base_network_norms.append(torch.norm(train_weight - base_network.train_weights[0]).cpu().detach().numpy())
for j,compare_network in enumerate(networks_to_compare):
networks_distance[j].append(torch.norm(base_network.train_weights[i]-compare_network.train_weights[i]).cpu().detach().numpy())
#plot
#subgraph.plot(base_network.train_time_stamps[0:],base_network_norms[0:],label='$\eta_{0}$ from init', linestyle = lines[0], marker=markers[0],markevery=10,markersize=7)
subgraph.plot(base_network.train_time_stamps[0:],base_network_norms[0:],label='$\eta_{0}$ from init', linestyle = (0,(1,1)), marker=markers[0],markevery=10,markersize=7)
for j,compare_network in enumerate(networks_to_compare):
#subgraph.plot(base_network.train_time_stamps[0:],networks_distance[j][0:],label='$\eta_{0}$ from $\eta_{0}/$'+str(compare_network.lr_ratio_from_base),linestyle = lines[(1+j)%len(lines)], marker=markers[(1+j)%len(markers)],markevery=10)
subgraph.plot(base_network.train_time_stamps[0:],networks_distance[j][0:],label='$\eta_{0}$ from $\eta_{0}/$'+str(compare_network.lr_ratio_from_base),linestyle = (0.25*(j+1),(1,1)), marker=markers[(1+j)%len(markers)],markevery=10)
subgraph.legend()
if scale == 'log':
subgraph.set_yscale('log')
subgraph.set_xlim(left=0,right=base_network.train_time_stamps[-1])
subgraph.set_ylim(bottom = 0,top = 3.0)
return subgraph
def plot_all_graphs(base_network, networks_to_compare = [], download = False, networks_type_str = 'Regular', scale = 'regular', figsize = (6,3.65),title_size=15,general_size=10.5,axis_size=14.5,title_height=0.96,fontstyle='normal',fontbold='normal'):
if networks_type_str == 'Linear':
title = 'Fully Connected, Linear Activation'
figsize = (6,3.05)
general_size = 9
if networks_type_str == 'ReLU':
title = 'Fully Connected, Rectified Linear Activation'
figsize = (6, 3.05)
general_size = 9
if networks_type_str == 'conv_subsample':
title = 'Convolutional, Adapted'
if networks_type_str == 'conv_max_pool':
title = 'Convolutional, Off-the-Shelf'
figure, subgraphs = plt.subplots(1, 2, figsize=figsize) # For fully connected used figsize = (6,3.05)
plt.rcParams.update({'legend.fontsize': general_size})
figure.suptitle(title,fontstyle=fontstyle,fontweight=fontbold,fontsize=title_size,y=title_height)
set_networks_loss_subgraph(base_network,subgraphs[0],networks_to_compare=networks_to_compare, download=download, networks_type_str=networks_type_str, scale=scale,axis_size=axis_size)
set_networks_distance_subgraph(base_network,subgraphs[1],networks_to_compare=networks_to_compare , download=download, networks_type_str=networks_type_str, scale=scale,axis_size=axis_size)
plt.tight_layout()
if download == True:
plt.savefig('plots/'+'Graph_'+networks_type_str.lower()+'.png',bbox_inches = "tight")
plt.show()
def experiment(compare_ratios = [],batch_size_train = 1000,n_epochs = 10000,learning_rate = 0.01,deviation = 0.01,layers_size=[28*28,50,50,10],activation='Linear', download = False, scale = 'log'):
#data
train_loader , test_loader = data_load(batch_size_train)
#initialize networks
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
network = Network(train_loader, test_loader, label=activation + '_h=' + str(learning_rate), deviation=deviation, layers_size=layers_size, activation=activation, batch_size_train=batch_size_train, init_mode='xavier')
network = network.to(device)
network.save_state()
networks_compare = []
for ratio in compare_ratios:
cur_network = torch.load(activation+'_h='+str(learning_rate) + '_model.pth')
cur_network = cur_network.to(device)
cur_network.label = activation+'_h='+str(learning_rate/ratio)
networks_compare.append(cur_network)
#train networks
network.init_train_params(n_epochs,learning_rate)
network.train()
for ratio,network_to_compare in zip(compare_ratios,networks_compare):
network_to_compare.init_train_params(n_epochs * ratio, learning_rate/ratio,lr_ratio_from_base=ratio)
network_to_compare.train()
#plot
#plot_all_graphs(network,networks_to_compare=networks_compare,networks_type_str=activation,download=download, scale=scale)
#old plots
#plot_networks_loss_graph(network,networks_to_compare=networks_compare,networks_type_str=activation,download=download)
#plot_networks_loss_graph(network,networks_to_compare=networks_compare,networks_type_str=activation,download=download, scale='log')
#plot_networks_distance_graph(network , networks_compare,networks_type_str=activation,download=download)
#plot_networks_distance_graph(network , networks_compare,networks_type_str=activation,download=download, scale='log')
def continue_training(network_file_name = 'Linear_h=0.005_model.pth', compare_ratios = [],batch_size_train = 1000,n_epochs = 10000,learning_rate = 0.01,deviation = 0.01,layers_size=[28*28,50,50,10],activation='Linear', download = False):
#data
train_loader , test_loader = data_load(batch_size_train)
network = torch.load(network_file_name)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
network = network.to(device)
network.train()
def load_networks_file_paths(activation = 'Linear',stepsize = 0.001, ratios = [2,5,10,20]):
networks_file_paths = []
networks_file_paths.append(activation+'_h='+str(stepsize)+'_model.pth')
for ratio in ratios:
networks_file_paths.append(activation+'_h='+str(stepsize/ratio)+'_model.pth')
return networks_file_paths
def plot_trained_networks_of_a_single_type(activation = 'Linear' , stepsize = 0.001, ratios = [2,5,10,20] , download = True, scale = 'regular'):
network_file_paths = load_networks_file_paths(activation=activation,stepsize=stepsize,ratios=ratios)
network = torch.load(network_file_paths[0])
networks_compare = [torch.load(network_file_path) for network_file_path in network_file_paths[1:]]
plot_all_graphs(network,networks_to_compare=networks_compare,networks_type_str=activation,download=download, scale=scale)
def plot_trained_networks_of_all_types(stepsize = 0.001, ratios = [2,5,10,20],scale = 'regular'):
plot_trained_networks_of_a_single_type(activation='Linear',stepsize=stepsize,ratios=ratios,scale=scale)
plot_trained_networks_of_a_single_type(activation='ReLU',stepsize=stepsize,ratios=ratios,scale=scale)
plot_trained_networks_of_a_single_type(activation='conv_subsample',stepsize=stepsize,ratios=ratios,scale=scale)
plot_trained_networks_of_a_single_type(activation='conv_max_pool',stepsize=stepsize,ratios=ratios,scale=scale)
def general_all_experiments(compare_ratios=[2,5,10,20],n_epochs=10000,learning_rate=0.001,batch_size_train=1000):
experiment(activation='Linear',compare_ratios=compare_ratios,n_epochs=n_epochs,learning_rate=learning_rate,download=True,batch_size_train=batch_size_train)
experiment(activation='ReLU',compare_ratios=compare_ratios,n_epochs=n_epochs,learning_rate=learning_rate,download=True,batch_size_train=batch_size_train)
experiment(activation='conv_subsample',compare_ratios=compare_ratios,n_epochs=n_epochs,learning_rate=learning_rate,download=True,batch_size_train=batch_size_train)
experiment(activation='conv_max_pool',compare_ratios=compare_ratios,n_epochs=n_epochs,learning_rate=learning_rate,download=True,batch_size_train=batch_size_train)
def run_specific_experiment(experiment_type=None,compare_ratios=[2,5,10,20],n_epochs=10000,learning_rate=0.001,batch_size_train=1000):
if experiment_type == 'fully_connected_linear':
activation = 'Linear'
elif experiment_type == 'fully_connected_relu':
activation = 'ReLU'
elif experiment_type == 'conv_subsample':
activation = 'conv_subsample'
elif experiment_type == 'conv_maxpool':
activation = 'conv_max_pool'
else:
print("no such experiment :\'",experiment_type,"\'")
return
experiment(activation=activation, compare_ratios=compare_ratios, n_epochs=n_epochs, learning_rate=learning_rate,
download=True, batch_size_train=batch_size_train)
plot_trained_networks_of_a_single_type(activation=activation,stepsize=learning_rate,ratios=compare_ratios)
def super_fast_experiments_to_check_graph_style():
general_all_experiments(compare_ratios=[2,3],n_epochs=1000,learning_rate=0.01,batch_size_train=20)
def semi_short_experiments_1hour_run():
general_all_experiments(compare_ratios=[2,4],n_epochs=10000,learning_rate=0.001,batch_size_train=100)
def full_scale_experiments():
random_seed = 1
torch.backends.cudnn.enabled = False
torch.manual_seed(random_seed)
if torch.cuda.is_available():
torch.set_default_tensor_type('torch.cuda.FloatTensor')
general_all_experiments(compare_ratios=[2,5,10,20],n_epochs=10000,learning_rate=0.001,batch_size_train=1000)