-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathlearning_cte.py
141 lines (115 loc) · 5.22 KB
/
learning_cte.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
from haiku.nets import ResNet18
import haiku as hk
import os
import pandas as pd
import optax
import argparse
import jax.numpy as jnp
import jax.random as jrand
import jax
import pickle
import os
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
from tqdm import tqdm
from core.data.images import loader
from core.utils.meters import AverageMeter
from core.data.load import get_bdy_states_rknn
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
DIRNAME = 'perception-ckpts-50-epochs'
os.makedirs(DIRNAME, exist_ok=True)
def main():
rng = jrand.PRNGKey(23)
train_loader, test_loader = loader(train_bs=64, test_bs=128)
imgs, _ = next(iter(train_loader))
opt_init, opt_update = optax.chain(
optax.scale_by_adam(b1=0.9, b2=0.999, eps=1e-8),
optax.scale(-0.001)
)
forward = hk.transform_with_state(_forward)
params, nn_state = forward.init(rng, imgs, is_training=True)
opt_state = opt_init(params)
def loss_fn(params, batch, nn_state, is_training=True):
imgs, labels = batch
preds, nn_state = forward.apply(params, nn_state, None, imgs, is_training=is_training)
loss = (1. / preds.shape[0]) * jnp.sum(jnp.square(preds - labels.reshape(preds.shape)))
return loss, (loss, nn_state)
@jax.jit
def update(params, batch, nn_state, opt_state):
grad, (loss, nn_state) = jax.grad(loss_fn, has_aux=True)(params, batch, nn_state)
updates, opt_state = opt_update(grad, opt_state, params)
params = optax.apply_updates(params, updates)
return loss, params, opt_state, nn_state
# We varied it between 50 - 1500 epoch
for epoch in range(50):
train_loss_meter = AverageMeter('train_loss')
test_loss_meter = AverageMeter('test_loss')
print(f'Beginning epoch {epoch}\n')
for batch_idx, batch in enumerate(train_loader):
loss, params, opt_state, nn_state = update(params, batch, nn_state, opt_state)
train_loss_meter.update(loss, n=batch[0].shape[0])
if batch_idx % 100 == 0:
print(f'Batch {batch_idx}/{len(train_loader)}\tLoss: {train_loss_meter.avg:.3f}')
all_preds, all_labels = [], []
for imgs, labels in tqdm(test_loader):
preds, _ = forward.apply(params, nn_state, None, imgs, is_training=False)
test_loss = (1. / preds.shape[0]) * jnp.sum(jnp.square(preds - labels.reshape(preds.shape)))
# test_loss, _ = loss_fn(params, batch, nn_state, is_training=False)
test_loss_meter.update(test_loss, n=batch[0].shape[0])
all_preds.append(jnp.asarray(preds))
all_labels.append(jnp.asarray(labels.reshape(preds.shape)))
plot_preds(all_preds, all_labels, test_loader.dataset.states)
# save checkpoint at the end of the epoch
save_checkpoint(params, nn_state)
print(f'\nFinished epoch {epoch}')
print(f'{"-"*50}')
print(f'Avg training loss: {train_loss_meter.avg:.4f}\t', end='')
print(f'Avg test loss: {test_loss_meter.avg:.4f}')
print(f'{"-"*50}\n')
# For saving output data into a text file
with open(f'{DIRNAME}/output_no_crop.txt', 'a') as f:
output_str = f'\nFinished epoch {epoch}\n'
output_str += f'{"-" * 50}\n'
output_str += f'Avg training loss: {train_loss_meter.avg:.4f}\t'
output_str += f'Avg test loss: {test_loss_meter.avg:.4f}\n'
output_str += f'{"-"*50}\n'
# Write output to file
f.write(output_str)
def _forward(images, is_training):
return ResNet18(1, resnet_v2=True)(images, is_training)
def save_checkpoint(params, nn_state):
os.makedirs(DIRNAME, exist_ok=True)
with open(os.path.join(DIRNAME, 'params.npy'), 'wb') as f:
pickle.dump(params, f, protocol=pickle.HIGHEST_PROTOCOL)
with open(os.path.join(DIRNAME, 'nn_state.npy'), 'wb') as f:
pickle.dump(nn_state, f, protocol=pickle.HIGHEST_PROTOCOL)
def load_checkpoint():
with open(os.path.join(DIRNAME, 'params.npy'), 'rb') as f:
params = pickle.load(f)
with open(os.path.join(DIRNAME, 'nn_state.npy'), 'rb') as f:
nn_state = pickle.load(f)
return params, nn_state
def plot_preds(all_preds, all_labels, state_df):
preds, labels = np.vstack(all_preds), np.vstack(all_labels)
df = pd.DataFrame(
list(zip(np.squeeze(preds), np.squeeze(labels))),
columns=['Predictions', 'Labels'])
full_df = pd.concat([df, state_df], axis=1)
get_bdy_states_rknn(full_df, ['Predictions', 'speed(m/s)', 'theta_e', 'd'], 200)
fig, ax = plt.subplots()
sns.set(style='darkgrid', font_scale=1.5)
sns.scatterplot(data=full_df, x='Labels', y='Predictions', hue='Safe', ax=ax)
line = np.linspace(-2, 2, 1000)
ax.plot(line, line, 'k--')
ax.set_xlabel(r'True $c_e$')
ax.set_ylabel(r'Predicted $c_e$')
handles, labels = ax.get_legend_handles_labels()
label_dict = {'True': 'Safe', True: 'Safe', 'False': 'Unsafe', False: 'Unsafe'}
new_labels = [label_dict[label] for label in labels]
ax.legend(handles=handles[0:], labels=new_labels)
plt.subplots_adjust(bottom=0.15)
plt.savefig(f'{DIRNAME}/perception-ckpts.png')
plt.close()
if __name__ == '__main__':
main()