-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathplot.py
163 lines (123 loc) · 7.95 KB
/
plot.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
#!/usr/bin/python
from __future__ import print_function
import sys
import os
import glob
import pickle
import collections
import matplotlib.pyplot as plt
import numpy as np
if sys.version_info < (3,):
range = xrange
EPSILON = 0.01
NUM_BUCKETS = 500
def make_bucket(vals, epsilon, x):
return np.abs(vals - x) < epsilon
def bucket_mean(vals, bucket):
return np.mean(vals[bucket])
def bucket_max(vals, bucket):
bucket_vals = vals[bucket]
if not len(bucket_vals):
return np.nan
return np.max(bucket_vals)
def sample_delta(samples, optimized_samples, bucket):
return np.mean(optimized_samples[bucket] - samples[bucket])
def bucket_count(bucket):
count = np.sum(bucket)
return count if count != 0 else np.nan
# def expand_bucket(bucket):
# first_true_i = np.argmax(bucket)
# if not first_true_i:
# return bucket
# new_bucket = np.copy(bucket)
# new_bucket[first_true_i - 1] = True
# return new_bucket
# def sign(x):
# return x / abs(x)
def norm(arr):
return arr/np.nanmax(arr)
def produce_trace(run):
linear = np.linspace(-1, 1, NUM_BUCKETS)
trace = collections.defaultdict(list)
trace["linear"] = linear
for x in linear:
random_proxy_bucket = make_bucket(run["random_proxy_vals"], EPSILON, x)
trace["random_samples"].append(bucket_count(random_proxy_bucket))
trace["random_real_vs_proxy"].append(bucket_mean(run["random_real_vals"], random_proxy_bucket))
trace["random_proxy"].append(bucket_mean(run["random_proxy_vals"], random_proxy_bucket))
optimized_bucket = make_bucket(run["optimized_proxy_vals"], EPSILON, x)
trace["optimized_samples"].append(bucket_count(optimized_bucket))
trace["optimized_real"].append(bucket_mean(run["optimized_real_vals"], optimized_bucket))
trace["optimized_proxy"].append(bucket_mean(run["optimized_proxy_vals"], optimized_bucket))
delta_bucket = make_bucket(run["optimized_proxy_vals"] - run["random_proxy_vals"], EPSILON, x)
trace["deltas"].append(bucket_mean(run["optimized_real_vals"] - run["random_real_vals"], delta_bucket))
trace["delta_samples"].append(bucket_count(delta_bucket))
if run["hyper_parameters"]["INPUT_SIZE"] == 1:
samples_bucket = make_bucket(run["samples"], EPSILON, x)
trace["real_values"].append(bucket_mean(run["random_real_vals"], samples_bucket))
trace["proxy_values"].append(bucket_mean(run["random_proxy_vals"], samples_bucket))
trace["optimized_count"].append(bucket_count(make_bucket(run["optimized_samples"], EPSILON, x)))
# randoptimized_bucket = make_bucket(run["randoptimized_proxy_vals"], EPSILON, x)
# trace["randoptimized_proxy"].append(bucket_mean(run["randoptimized_proxy_vals"], randoptimized_bucket))
# trace["randoptimized_real"].append(bucket_mean(run["randoptimized_real_vals"], randoptimized_bucket))
# randoptimized_samples.append(bucket_count(randoptimized_bucket))
# random_sampled_proxy_bucket = make_bucket(run["random_sampled_proxy_vals"], EPSILON, x)
# trace["random_real_vs_sampled_proxy"].append(bucket_mean(run["random_real_vals"], random_sampled_proxy_bucket))
# trace["random_sampled_proxy"].append(bucket_mean(run["random_sampled_proxy_vals"], random_sampled_proxy_bucket))
# sampled_proxy_optimized_bucket = make_bucket(run["sampled_proxy_optimized_proxy_vals"], EPSILON, x)
# trace["sampled_proxy_optimized_samples"].append(bucket_count(sampled_proxy_optimized_bucket))
# trace["sampled_proxy_optimized_real"].append(bucket_mean(run["sampled_proxy_optimized_real_vals"], sampled_proxy_optimized_bucket))
# trace["sampled_proxy_optimized_proxy"].append(bucket_mean(run["sampled_proxy_optimized_proxy_vals"], sampled_proxy_optimized_bucket))
random_grad_opt_proxy_bucket = make_bucket(run["random_grad_opt_proxy_vals"], EPSILON, x)
trace["random_real_vs_grad_opt_proxy"].append(bucket_mean(run["random_real_vals"], random_grad_opt_proxy_bucket))
trace["random_grad_opt_proxy"].append(bucket_mean(run["random_grad_opt_proxy_vals"], random_grad_opt_proxy_bucket))
grad_opt_proxy_optimized_bucket = make_bucket(run["grad_opt_proxy_optimized_proxy_vals"], EPSILON, x)
trace["grad_opt_proxy_optimized_samples"].append(bucket_count(grad_opt_proxy_optimized_bucket))
trace["grad_opt_proxy_optimized_real"].append(bucket_mean(run["grad_opt_proxy_optimized_real_vals"], grad_opt_proxy_optimized_bucket))
trace["grad_opt_proxy_optimized_proxy"].append(bucket_mean(run["grad_opt_proxy_optimized_proxy_vals"], grad_opt_proxy_optimized_bucket))
return trace
def get_trace_for_file(run_path):
print("Processing:", run_path)
with open(run_path, "rb") as f:
return produce_trace(pickle.load(f))
if __name__ == "__main__":
run_paths = glob.glob("runs/*.pickle")
traces = collections.defaultdict(list)
for run_path in run_paths:
trace = get_trace_for_file(run_path)
for trace_name, trace_list in trace.items():
trace_array = np.asarray(trace_list)
traces[trace_name].append(trace_array)
# Average the traces.
for k in traces:
traces_k_array = np.asarray(traces[k])
traces[k] = np.nanmean(traces_k_array, axis=0)
plt.figure(figsize=(16, 12))
plt.plot(traces["linear"], traces["random_real_vs_proxy"], label="Real Utility vs. Proxy Utility (Random Data)")
plt.plot(traces["linear"], traces["optimized_real"], label="Real Utility vs. Proxy Utility (Optimized Data)")
plt.plot(traces["linear"], traces["optimized_real"] - traces["random_real_vs_proxy"], label="Real Utility vs. Proxy Utility (Optimized Data - Random Data) [Goodhart Error]")
# plt.plot(traces["linear"], traces["optimized_proxy"] - traces["random_proxy"], label="Proxy Utility (Optimized - Random) [Bucket Error]")
# plt.plot(traces["linear"], traces["deltas"], label="Opt Real - Rand Real vs. Opt Proxy - Rand Proxy [Optimization Delta]")
# plt.plot(traces["linear"], norm(traces["random_samples"]), label="Random Samples")
# plt.plot(traces["linear"], norm(traces["optimized_samples"]), label="Optimized Samples")
# plt.plot(traces["linear"], norm(traces["randoptimized_samples"]), label="Randoptimized Samples")
# plt.plot(traces["linear"], traces["randoptimized_real"], label="Real Utility (Randoptimized)")
# plt.plot(traces["linear"], norm(traces["delta_samples"]), label="Delta Samples")
# plt.plot(traces["linear"], traces["real_values"], label="Real Utility")
# plt.plot(traces["linear"], traces["proxy_values"], label="Proxy Utility")
# plt.plot(traces["linear"], norm(traces["optimized_count"]), label="Optimized Points")
# plt.plot(traces["linear"], traces["random_real_vs_sampled_proxy"], label="Real Utility vs. Sampled Proxy Utility (Random Data)")
# plt.plot(traces["linear"], traces["sampled_proxy_optimized_real"], label="Real Utility vs. Sampled Proxy Utility (Sampled Proxy Optimized Data)")
# plt.plot(traces["linear"], traces["sampled_proxy_optimized_real"] - traces["random_real_vs_sampled_proxy"], label="Real Utility vs. Sampled Proxy Utility (Sampled Proxy Optimized Data - Random Data) [Goodhart Error]")
# plt.plot(traces["linear"], traces["random_real_vs_grad_opt_proxy"], label="Real Utility vs. Grad Opt Proxy Utility (Random Data)")
# plt.plot(traces["linear"], traces["grad_opt_proxy_optimized_real"], label="Real Utility vs. Grad Opt Proxy Utility (Grad Opt Proxy Optimized Data)")
# plt.plot(traces["linear"], traces["grad_opt_proxy_optimized_real"] - traces["random_real_vs_grad_opt_proxy"], label="Real Utility vs. Grad Opt Proxy Utility (Grad Opt Proxy Optimized Data - Random Data) [Goodhart Error]")
plt.legend()
# plt.ylabel("Proxy Utility")
plt.ylabel("Real Utility")
plt.xlabel("Proxy Utility")
# plt.xlabel("Input Value")
plt.grid()
plot_file = os.path.join(os.path.dirname(__file__), "output.png")
plt.savefig(plot_file)
plt.show()