-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_uci_dataset_tests.py
352 lines (298 loc) · 15.9 KB
/
run_uci_dataset_tests.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
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
import argparse
import numpy as np
import pandas as pd
import os
from pathlib import Path
import seaborn as sns
import matplotlib.pyplot as plt
import tensorflow as tf
import time
from scipy import stats
import re
import edl
import data_loader
import trainers
import models
from models.toy.h_params import h_params
parser = argparse.ArgumentParser()
parser.add_argument("--num-trials", default=20, type=int,
help="Number of trials to repreat training for \
statistically significant results.")
parser.add_argument("--num-epochs", default=40, type=int)
parser.add_argument('--datasets', nargs='+', default=["yacht"],
choices=['boston', 'concrete', 'energy-efficiency',
'kin8nm', 'naval', 'power-plant', 'protein',
'wine', 'yacht'])
parser.add_argument("--load-pkl", action='store_true',
help="Load predictions for a cached pickle file or \
recompute from scratch by feeding the data through \
trained models")
args = parser.parse_args()
output_dir = "figs/uci"
"""" ================================================"""
#training_schemes = [trainers.Likelihood, trainers.Likelihood, trainers.Evidential, trainers.Ensemble]
#method_names = ["Gaussian", "Laplace", "Evidential", "Ensemble"]
training_schemes = [trainers.Likelihood]
method_names = ["Gaussian"]
datasets = args.datasets
num_trials = args.num_trials
num_epochs = args.num_epochs
dev = "/cpu:0" # for small datasets/models cpu is faster than gpu
"""" ================================================"""
def predict(method_name, model, x):
if method_name == "Dropout":
preds = tf.stack([model(x, training=True) for _ in range(n_samples)], axis=0) #forward pass
mu, var = tf.nn.moments(preds, axes=0)
return mu, tf.sqrt(var)
elif method_name == "Evidential":
outputs = model(x, training=False)
mu, v, alpha, beta = tf.split(outputs, 4, axis=-1)
sigma = tf.sqrt(beta/(v*(alpha + 1e-6 - 1)))
print ("Evidential predict shape : ", mu.shape, sigma.shape)
return mu, sigma
elif method_name == "Ensemble":
preds = tf.stack([f(x) for f in model], axis=0)
y, sigmas = tf.split(preds, 2, axis=-1)
mu = tf.reduce_mean(y, axis=0)
sigma = tf.math.reduce_std(sigmas, axis=0)
var = tf.reduce_mean(sigmas**2 + tf.square(y), axis=0) - tf.square(mu)
#preds = tf.stack([f(x) for f in model], axis=0)
print ("Ensemble preds shape ", preds.shape)
#mu, var = tf.nn.moments(preds, 0)
print ("Ensemble predict shape : ", mu.shape, var.shape)
return mu, tf.sqrt(var)
elif method_name == "Gaussian":
outputs = model(x, training=False)
mu, var = tf.split(outputs, 2, axis=-1)
return mu, tf.sqrt(var)
elif method_name == "Laplace":
outputs = model(x, training=False)
mu, b = tf.split(outputs, 2, axis=-1)
print ("Laplace predict shape : ", mu.shape, b.shape)
return mu, b
else:
raise ValueError("Unknown model")
def get_prediction_summary(dataset, method_name, model, x_batch, y_batch):
#First collect predictions
mu_batch, sigma_batch = predict(method_name, model, x_batch)
#mu_batch = np.clip(mu_batch, 0, 1)
mu_batch = np.squeeze(mu_batch.numpy())
sigma_batch = np.squeeze(sigma_batch.numpy())
y_batch = np.squeeze(y_batch)
print (" Prediction summary : ", method_name, mu_batch.shape, sigma_batch.shape, y_batch.shape)
### Save the predictions to some dataframes for later analysis
summary = [{"Dataset": dataset, "Method": method_name,"Target": y, "Mu": mu, "Sigma": sigma}
for y,mu,sigma in zip(y_batch, mu_batch, sigma_batch)]
return summary
"""" ================================================"""
def compute_predictions():
RMSE = np.zeros((len(datasets), len(training_schemes), num_trials))
NLL = np.zeros((len(datasets), len(training_schemes), num_trials))
df_pred_uci = pd.DataFrame(columns=["Dataset", "Method", "Target", "Mu", "Sigma"] )
for di, dataset in enumerate(datasets):
for ti, trainer_obj in enumerate(training_schemes):
for n in range(num_trials):
(x_train, y_train), (x_test, y_test), y_scale = data_loader.load_dataset(dataset, return_as_tensor=False)
batch_size = h_params[dataset]["batch_size"]
num_iterations = num_epochs * x_train.shape[0]//batch_size
print ("Num of iterations :", num_iterations)
done = False
while not done:
with tf.device(dev):
model_generator = models.get_correct_model(dataset="toy", trainer=trainer_obj)
model, opts = model_generator.create(input_shape=x_train.shape[1:])
if method_names[ti] == "Laplace": #training scheme is likelihood; as its 2nd in list
print ("Trainienr lalpace likelihood")
trainer = trainer_obj(model, opts, "laplace", dataset, learning_rate=h_params[dataset]["learning_rate"])
elif method_names[ti] == "Gaussian":
print ("Trainienr Gaussian likelihood")
trainer = trainer_obj(model, opts, "gaussian", dataset, learning_rate=h_params[dataset]["learning_rate"])
else:
trainer = trainer_obj(model, opts, dataset, learning_rate=h_params[dataset]["learning_rate"])
model, rmse, nll = trainer.train(x_train, y_train, x_test, y_test, y_scale, iters=num_iterations, batch_size=batch_size, verbose=True)
#Compute on validation data and save predictions
summary_to_add = get_prediction_summary(
dataset, method_names[ti], model, x_test, y_test)
df_pred_uci = df_pred_uci.append(summary_to_add, ignore_index=True)
del model
tf.keras.backend.clear_session()
done = False if np.isinf(nll) or np.isnan(nll) else True
print("saving {} {}".format(rmse, nll))
RMSE[di, ti, n] = rmse
NLL[di, ti, n] = nll
RESULTS = np.hstack((RMSE, NLL))
mu = RESULTS.mean(axis=-1)
error = np.std(RESULTS, axis=-1)
print("==========================")
print("[{}]: {} pm {}".format(dataset, mu, error))
print("==========================")
print("TRAINERS: {}\nDATASETS: {}".format([trainer.__name__ for trainer in training_schemes], datasets))
print("MEAN: \n{}".format(mu))
print("ERROR: \n{}".format(error))
return df_pred_uci
def gen_paper_plots(df_pred_uci):
cm = 1/2.54 # centimeters in inches
sns.set()
sns.set_style("white")
sns.set_style("ticks")
sns.despine()
sns.set_context("paper")
sns.color_palette("tab10")
print ("Generating point plots")
print (df_pred_uci.head())
dataset_names = df_pred_uci["Dataset"].unique()
new_dataset_names = [label.replace('-', '-\n') for label in dataset_names]
#================================================
print (f"Generating Interval Score")
df_pred_uci["lower"] = df_pred_uci["Mu"] - 2*df_pred_uci["Sigma"]
df_pred_uci["lower"].mask(df_pred_uci["Method"]=="Laplace", df_pred_uci["Mu"] - 3*df_pred_uci["Sigma"], inplace=True)
df_pred_uci["upper"] = df_pred_uci["Mu"] + 2*df_pred_uci["Sigma"]
df_pred_uci["upper"].mask(df_pred_uci["Method"]=="Laplace", df_pred_uci["Mu"] + 3*df_pred_uci["Sigma"], inplace=True)
df_pred_uci["Interval Score"] = df_pred_uci["upper"] - df_pred_uci["lower"] \
+ (2/0.95)*(df_pred_uci["lower"]-df_pred_uci["Target"])*(df_pred_uci["Target"]<df_pred_uci["lower"]) \
+ (2/0.95)*(df_pred_uci["Target"] - df_pred_uci["upper"])*(df_pred_uci["Target"]>df_pred_uci["upper"])
fig = plt.figure(figsize=(14.2*cm, 14.2*cm/2.0))
gs = fig.add_gridspec(1, 9)
for i,dataset_name in enumerate(dataset_names):
ax = fig.add_subplot(gs[0,i])
g = sns.pointplot(x="Method", y="Interval Score", hue="Method",
markers=["o", "x", "*", "D"],
linestyles=["-","--","-.",":"],
data=df_pred_uci[df_pred_uci.Dataset == dataset_name], legend=False)
#g.set(yscale="log")
g.get_legend().remove()
g.set_xlabel(new_dataset_names[i], fontsize='xx-small')
g.axes.get_xaxis().set_ticks([])
g.axes.get_yaxis().set_ticks([])
if i != 0:
g.axes.get_yaxis().get_label().set_visible(False)
g.axes.get_yaxis().set_ticks([])
handles, labels = g.get_legend_handles_labels()
fig.legend(handles, labels, bbox_to_anchor=(0.55, 0.92), loc='lower center', ncol=5, fancybox=True, shadow=True)
plt.savefig(os.path.join(output_dir, f"Interval_score_point_uci.pdf"), bbox_inches='tight')
plt.show()
plt.clf()
#RMSE
df_pred_uci["RMSE"] = (df_pred_uci["Mu"] - df_pred_uci["Target"])**2
fig = plt.figure(figsize=(14.2*cm, 14.2*cm/2.0))
gs = fig.add_gridspec(1, 9)
for i,dataset_name in enumerate(dataset_names):
ax = fig.add_subplot(gs[0,i])
g = sns.pointplot(x="Method", y="RMSE", hue="Method",
markers=["o", "x", "*", "D"],
linestyles=["-","--","-.",":"],
data=df_pred_uci[df_pred_uci.Dataset == dataset_name], legend=False)
g.get_legend().remove()
g.set_xlabel(new_dataset_names[i], fontsize='xx-small')
g.axes.get_xaxis().set_ticks([])
g.axes.get_yaxis().set_ticks([])
if i != 0:
g.axes.get_yaxis().get_label().set_visible(False)
handles, labels = g.get_legend_handles_labels()
fig.legend(handles, labels, bbox_to_anchor=(0.55, 0.92), loc='lower center', ncol=5, fancybox=True, shadow=True)
plt.savefig(os.path.join(output_dir, "RMSE_point_uci.pdf"), bbox_inches='tight')
plt.show()
def simplified_paper_plot(df_pred_uci):
cm = 1/2.54 # centimeters in inches
sns.set()
sns.set_style("white")
sns.set_style("ticks")
sns.despine()
sns.set_context("paper")
sns.color_palette("tab10")
print ("Generating point plots")
print (df_pred_uci.head())
dataset_names = df_pred_uci["Dataset"].unique()
new_dataset_names = [label.replace('-', '-\n') for label in dataset_names]
print (new_dataset_names, dataset_names)
#================================================
print (f"Generating Interval Score")
df_pred_uci["lower"] = df_pred_uci["Mu"] - 2*df_pred_uci["Sigma"]
df_pred_uci["lower"].mask(df_pred_uci["Method"]=="Laplace", df_pred_uci["Mu"] - 3*df_pred_uci["Sigma"], inplace=True)
df_pred_uci["upper"] = df_pred_uci["Mu"] + 2*df_pred_uci["Sigma"]
df_pred_uci["upper"].mask(df_pred_uci["Method"]=="Laplace", df_pred_uci["Mu"] + 3*df_pred_uci["Sigma"], inplace=True)
df_pred_uci["Interval Score"] = df_pred_uci["upper"] - df_pred_uci["lower"] \
+ (2/0.95)*(df_pred_uci["lower"]-df_pred_uci["Target"])*(df_pred_uci["Target"]<df_pred_uci["lower"]) \
+ (2/0.95)*(df_pred_uci["Target"] - df_pred_uci["upper"])*(df_pred_uci["Target"]>df_pred_uci["upper"])
fig = plt.figure(figsize=(14.2*cm, 14.2*cm/2.0))
gs = fig.add_gridspec(2, 1)
ax = fig.add_subplot(gs[0,0])
#RMSE
df_pred_uci["RMSE"] = (df_pred_uci["Mu"] - df_pred_uci["Target"])**2
g = sns.pointplot(x="Dataset", y="RMSE", hue="Method",
markers=["*","D","o", "x" ],
linestyles=["-","--","-.",":"],
dodge=0.45,
join=False, ci='sd',palette=["C2", "C3", "C0", "C1"],
data=df_pred_uci, legend=False)
g.set_xticklabels([])
g.get_legend().remove()
# Improve the legend
handles, labels = g.get_legend_handles_labels()
print (labels)
fig.legend(handles, labels, bbox_to_anchor=(0.5, 0.85), loc='lower center', ncol=4, fancybox=True, shadow=True)
ax = fig.add_subplot(gs[1,0])
g = sns.pointplot(x="Dataset", y="Interval Score", hue="Method",
markers=["*","D","o", "x" ],
linestyles=["-","--","-.",":"],
dodge=0.45,
join=False, ci='sd',palette=["C2", "C3", "C0", "C1"],
data=df_pred_uci, legend=False)
g.set(yscale="log")
g.get_legend().remove()
g.set_xticklabels(new_dataset_names, fontsize='x-small')
plt.savefig(os.path.join(output_dir, "RMSE_IS_Comb_point_uci.pdf"))
plt.show()
def gen_plots(df_pred_uci):
print (f"Generating Entropy plot")
print (df_pred_uci.head())
print (df_pred_uci[df_pred_uci['Method']=="Ensemble"].head())
df_pred_uci["Sigma"] = pd.to_numeric(df_pred_uci.Sigma, errors='coerce')
df_pred_uci["Mu"] = pd.to_numeric(df_pred_uci.Mu, errors='coerce')
df_pred_uci["Target"] = pd.to_numeric(df_pred_uci.Target, errors='coerce')
# No use of platting Entropy so disabling
#df_pred_uci["Entropy"] = 0.5*np.log( 2 * np.pi * np.exp(1.) * (df_pred_uci["Sigma"])**2 )
#df_pred_uci["Entropy"].mask(df_pred_uci["Method"]=="Laplace", np.log(2*df_pred_uci["Sigma"]*np.exp(1.)), inplace=True) # entropy for laplace distirbution
### Plot PDF for each Dataset
#g = sns.FacetGrid(df_pred_uci, col="Dataset", hue="Method")
#g.map(sns.distplot, "Entropy").add_legend()
#plt.savefig(os.path.join(".", f"entropy_pdf_per_method.pdf"))
#plt.show()
#sns.catplot(x="Dataset", y="Entropy", hue="Method", data=df_pred_uci, kind="box", whis=0.5, showfliers=False)
#plt.savefig(os.path.join(output_dir, f"entropy_box_uci.pdf"))
#plt.show()
print ("Plot error distribution")
df_pred_uci["RMSE"] = (df_pred_uci["Mu"] - df_pred_uci["Target"])**2
g = sns.catplot(x="Dataset", y="RMSE", hue="Method", data=df_pred_uci, kind="box", whis=0.5, showfliers=False, aspect=4.0)
plt.savefig(os.path.join(output_dir, f"RMSE_box_uci.pdf"))
plt.show()
g = sns.catplot(x="Method", y="RMSE", col="Dataset", data=df_pred_uci, kind="box", whis=0.5, showfliers=False, aspect=0.3)
plt.savefig(os.path.join(output_dir, f"RMSE_box_uci_panel.pdf"))
plt.show()
print (f"Generating Interval Score")
df_pred_uci["lower"] = df_pred_uci["Mu"] - 2*df_pred_uci["Sigma"]
df_pred_uci["lower"].mask(df_pred_uci["Method"]=="Laplace", df_pred_uci["Mu"] - 3*df_pred_uci["Sigma"], inplace=True)
df_pred_uci["upper"] = df_pred_uci["Mu"] + 2*df_pred_uci["Sigma"]
df_pred_uci["upper"].mask(df_pred_uci["Method"]=="Laplace", df_pred_uci["Mu"] + 3*df_pred_uci["Sigma"], inplace=True)
df_pred_uci["Interval Score"] = df_pred_uci["upper"] - df_pred_uci["lower"] \
+ (2/0.95)*(df_pred_uci["lower"]-df_pred_uci["Target"])*(df_pred_uci["Target"]<df_pred_uci["lower"]) \
+ (2/0.95)*(df_pred_uci["Target"] - df_pred_uci["upper"])*(df_pred_uci["Target"]>df_pred_uci["upper"])
g = sns.catplot(x="Dataset", y="Interval Score", hue="Method", data=df_pred_uci, kind="box", whis=0.5, showfliers=False, aspect=4.0)
g.set(yscale="log")
plt.savefig(os.path.join(output_dir, f"Interval_score_box_uci.pdf"))
plt.show()
g = sns.catplot(x="Method", y="Interval Score", col="Dataset", data=df_pred_uci, kind="box", whis=0.5, showfliers=False, aspect=0.7)
g.set(yscale="log")
plt.savefig(os.path.join(output_dir, f"Interval_score_box_uci_panel.pdf"))
plt.show()
if args.load_pkl:
print("Loading!")
df_pred_uci = pd.read_pickle("cached_uci_results.pkl")
else:
df_pred_uci = compute_predictions()
df_pred_uci.to_pickle("cached_uci_results.pkl")
"""=============================="""
Path(output_dir).mkdir(parents=True, exist_ok=True)
#gen_plots(df_pred_uci)
#gen_paper_plots(df_pred_uci)
simplified_paper_plot(df_pred_uci)