-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathS4.2-plot-CNN.py
230 lines (194 loc) · 8.36 KB
/
S4.2-plot-CNN.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
import os
import numpy as np
from tensorboard.backend.event_processing.event_accumulator import EventAccumulator
import matplotlib.pyplot as plt
from mpl_toolkits.axes_grid1.inset_locator import mark_inset
from mpl_toolkits.axes_grid1.inset_locator import inset_axes
from matplotlib.lines import Line2D
MAX_EPOCH = 80
def read_tensorboard_data(logdir, tag, max_step=-1):
# Create EventAccumulator object and reload event files
event_acc = EventAccumulator(logdir)
event_acc.Reload()
# Initialize data dictionary
dict_step_value = {}
# Traverse all scalar events
for scalar_event in event_acc.Scalars(tag):
step = scalar_event.step
value = scalar_event.value
if max_step > 0 and step > max_step:
continue
# Categorize data into different runs based on steps
if step not in dict_step_value:
dict_step_value[step] = []
dict_step_value[step].append(value)
data = {'steps': [], 'values_avg': [], 'values_var': [], 'values_std': [], 'repeat_time': None}
# Check if the number of data points in each run is consistent
unique_counts = set([len(values) for values in dict_step_value.values()])
if len(unique_counts) == 1:
data['repeat_time'] = list(unique_counts)[0]
else:
inconsistent_steps = [step for step, values in dict_step_value.items() if len(values) != max(unique_counts)]
print(f"[Warning] {logdir}: Missing some data (Steps: {min(inconsistent_steps)}--{max(inconsistent_steps)})")
data['repeat_time'] = max(unique_counts)
# Calculate mean, variance, and standard deviation for each step
for step, values in dict_step_value.items():
data['steps'].append(step)
data['values_avg'].append(np.mean(values))
data['values_var'].append(np.var(values))
data['values_std'].append(np.std(values))
return data
def plot_tensorboard_data(data_list, tag, save_path_png, save_path_pdf,
legend_lines, legend_text):
SCALE = 0.6
fig = plt.figure(figsize=(SCALE*6.4, SCALE*4.8))
ax = plt.gca()
axins = inset_axes(ax, width="40%", height="30%",loc='lower left',
bbox_to_anchor=(0.4, 0.2, 1, 1),
bbox_transform=ax.transAxes)
for i, data in enumerate(data_list):
linestyle = data['linestyle']
marker = data['marker']
color = data['color']
running_data = data['data']
steps = running_data['steps']
smooth_loss = []
for loss in running_data['values_avg']:
if len(smooth_loss) == 0:
smooth_loss.append(loss)
else:
gamma = 0.8
smooth_loss.append(gamma*smooth_loss[-1]+(1-gamma)*loss)
smooth_loss = [100*v for v in smooth_loss]
ax.plot(steps, smooth_loss,
linestyle=linestyle, marker=marker, color=color,
markevery=5)
axins.plot(steps, smooth_loss,
linestyle=linestyle, marker=marker, color=color,
markevery=2)
# Adjust display range of the sub-graph
axins.set_xlim(MAX_EPOCH-3, MAX_EPOCH)
axins.set_ylim(98.7, 99.3)
axins.grid(True)
# axins.set_xticks([])
# axins.set_yticks([])
# Establish connection lines between the main and inset axes
# loc1 loc2: four corners of the axes
# 1 (upper right) 2 (upper left) 3 (lower left) 4 (lower right)
mark_inset(ax, axins, loc1=2, loc2=1, fc="none", ec='k', lw=1)
ax.set_ylim(bottom=20)
ax.set_xlabel('Epoch')
ax.set_ylabel('Test Accuracy')
# ax.grid(True)
ax.legend(legend_lines, legend_text,
bbox_to_anchor=(1,0), loc="lower left",
)
plt.savefig(save_path_png, format='png', bbox_inches='tight')
plt.savefig(save_path_pdf, format='pdf', bbox_inches='tight')
print(f'finish: {save_path_pdf}')
plt.clf()
if __name__ == "__main__":
log_directory = 'runs/MNIST-CNN'
tensorboard_tag = 'Accuracy/test'
subdirectories = []
data_list = []
taus = [0.6, 0.7, 0.8]
legend_lines = [
Line2D([0], [0], color=f'C{tau_idx}', lw=4)
for tau_idx in range(len(taus))
]
legend_text = [
fr'$\tau={tau}$' for tau in taus
]
for tau_idx, tau in enumerate(taus):
subdir_path = os.path.join(log_directory, f'Analog SGD-tau={tau}')
tensorboard_data = read_tensorboard_data(subdir_path, tensorboard_tag, MAX_EPOCH)
data_list.append({
'data': tensorboard_data,
'linestyle': '--',
'marker': 'o',
'color': f'C{tau_idx}'
})
mean_final = sum(tensorboard_data['values_avg'][-5:])*100/5
stdv_final = sum(tensorboard_data['values_std'][-5:])*100/5
print(fr'Analog SGD-tau={tau}: {mean_final:.2f} \stdv{{$\pm$ {stdv_final:.2f}}} steps: {len(tensorboard_data["steps"])} repeat: {tensorboard_data["repeat_time"]}')
legend_lines.append(
Line2D([0], [0], linestyle=data_list[-1]['linestyle'],
marker=data_list[-1]['marker'],
markersize=4,
color=f'k')
)
legend_text.append('Analog SGD')
for tau_idx, tau in enumerate(taus):
subdir_path = os.path.join(log_directory, f'TT-v1-tau={tau}')
tensorboard_data = read_tensorboard_data(subdir_path, tensorboard_tag, MAX_EPOCH)
data_list.append({
'data': tensorboard_data,
'linestyle': '-',
'marker': '^',
'color': f'C{tau_idx}'
})
mean_final = sum(tensorboard_data['values_avg'][-5:])*100/5
stdv_final = sum(tensorboard_data['values_std'][-5:])*100/5
print(fr'TT-tau={tau}: {mean_final:.2f} \stdv{{$\pm$ {stdv_final:.2f}}} steps: {len(tensorboard_data["steps"])} repeat: {tensorboard_data["repeat_time"]}')
legend_lines.append(
Line2D([0], [0], linestyle=data_list[-1]['linestyle'],
marker=data_list[-1]['marker'],
markersize=4,
color=f'k')
)
legend_text.append('Tiki-Taka')
# subdir_path = os.path.join(log_directory, f'mp')
# tensorboard_data = read_tensorboard_data(subdir_path, tensorboard_tag, MAX_EPOCH)
# data_list.append({
# 'data': tensorboard_data,
# 'linestyle': '-',
# 'marker': 'D',
# 'color': f'C{len(taus)+1}'
# })
# print(f'MP: {sum(tensorboard_data["values_means"][-5:])*100/5:.2f} steps: {len(tensorboard_data["steps"])}')
# legend_lines.append(
# Line2D([0], [0], linestyle=data_list[-1]['linestyle'],
# marker=data_list[-1]['marker'],
# color=data_list[-1]['color'],
# )
# )
# legend_text.append(r'MP ($\tau=0.6$)')
subdir_path = os.path.join(log_directory, f'FP SGD')
tensorboard_data = read_tensorboard_data(subdir_path, tensorboard_tag, MAX_EPOCH)
data_list.append({
'data': tensorboard_data,
'linestyle': '-.',
'marker': 's',
'color': f'C{len(taus)}'
})
legend_lines.append(
Line2D([0], [0], linestyle=data_list[-1]['linestyle'],
marker=data_list[-1]['marker'],
color=data_list[-1]['color'],
)
)
mean_final = sum(tensorboard_data['values_avg'][-5:])*100/5
stdv_final = sum(tensorboard_data['values_std'][-5:])*100/5
print(fr'FP: {mean_final:.2f} \stdv{{$\pm$ {stdv_final:.2f}}} steps: {len(tensorboard_data["steps"])} repeat: {tensorboard_data["repeat_time"]}')
legend_text.append('Digital SGD')
# linestyle_markers = [
# ('-', 'o', 'blue'),
# ('--', '^', 'orange'),
# ('-.', 's', 'green'),
# (':', 'D', 'red')
# ]
# Generate a separate plot file for each folder
file_dir = ''
dir_png_path = os.path.join(file_dir, 'fig', 'png')
dir_pdf_path = os.path.join(file_dir, 'fig', 'pdf')
if not os.path.isdir(dir_png_path):
os.makedirs(dir_png_path)
if not os.path.isdir(dir_pdf_path):
os.makedirs(dir_pdf_path)
pic_png_path = os.path.join(dir_png_path, 'A03-MNIST-CNN.png')
pic_pdf_path = os.path.join(dir_pdf_path, 'A03-MNIST-CNN.pdf')
# Plot and save the curve
plot_tensorboard_data(data_list, tensorboard_tag,
pic_png_path, pic_pdf_path,
legend_lines, legend_text)