-
Notifications
You must be signed in to change notification settings - Fork 0
/
run_policy_learning_gcn.py
278 lines (209 loc) · 9.83 KB
/
run_policy_learning_gcn.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
import torch
import os
import numpy as np
import pandas as pd
from datasets.utils_gcn import create_trace_dataloader
from torch_geometric.data import Batch
NUM_TOTAL_ACTIONS = 4
WIN_HIST = 3
def train(lfd_params, model, verbose=False, input_dtype="video", ablation=False):
# Create DataLoaders
assert input_dtype in ["video", "iad", "gcn"], "ERROR: run_videos.py: input_dtype must be 'video' or 'itr'"
if input_dtype == "video":
from datasets.dataset_video_trace import DatasetVideoTrace as CustomDataset
elif input_dtype == "iad":
from datasets.dataset_iad_trace import DatasetIADTrace as CustomDataset
else:
from datasets.dataset_gcn_trace import DatasetGCNTrace as CustomDataset
dataset = CustomDataset(lfd_params, lfd_params.application.file_directory, "train", eval=False, trace_path=lfd_params.application.trace_file, verbose=True,
backbone=lfd_params.model.model_id, num_segments=lfd_params.input_frames, ablation=ablation)
data_loader = create_trace_dataloader(dataset, lfd_params, "train", shuffle=True)
# put model on GPU
params = list(model.parameters())
net = torch.nn.DataParallel(model, device_ids=lfd_params.gpus).cuda()
net.train()
# define loss function
criterion = torch.nn.CrossEntropyLoss().cuda()
# define optimizer
optimizer = torch.optim.Adam(params, lr=lfd_params.lr)
# Train Network
loss_record = []
with torch.autograd.detect_anomaly():
epoch = lfd_params.epochs
for e in range(epoch):
cumulative_loss = 0
for i, data_packet in enumerate(data_loader):
obs, act, obs_filename, act_filename = data_packet
obs = list(obs)
# constrain size to a history of 5 timesteps
obs = obs[-WIN_HIST:]
act = act[:, -WIN_HIST:]
obs = Batch.from_data_list(obs)
# obtain label
label = act[:, -1]
label = torch.as_tensor(label).cuda()
label = torch.argmax(label, dim=1)
# hide label
act[:, -1] = 0
# compute output
logits = net(obs, act.float())
# get loss
loss = criterion(logits, label.long().cuda())
loss.backward()
# optimize SGD
optimizer.step()
optimizer.zero_grad()
if verbose:
print("epoch: {:3d}/{:3d}".format(e, epoch))
print("loss:", loss.cpu().detach().numpy())
print("expected:", label.cpu().detach().numpy())
print("pred:", np.argmax(logits.cpu().detach().numpy(), axis=1))
print("logits:")
print(logits.cpu().detach().numpy())
cumulative_loss += loss.cpu().detach().numpy()
print("e:", e, "loss:", cumulative_loss)
loss_record.append(cumulative_loss)
# show loss over time, output placed in Log Directory
import matplotlib.pyplot as plt
plt.plot(loss_record)
# add bells and whistles to plt
plt.title(model.filename)
plt.ylabel("loss")
plt.tight_layout()
# make sure log_dir exists
log_dir = model.filename
if not os.path.exists(log_dir):
os.makedirs(log_dir)
# save plt to file
fig_filename = os.path.join(log_dir, "train_loss.png")
plt.savefig(fig_filename)
# clear plt so I don't draw on top of my multiple images.
plt.clf()
return model
def evaluate_single_action(lfd_params, model, mode="evaluation", verbose=False, input_dtype="video"):
# Create DataLoaders
assert input_dtype in ["video", "iad", "gcn"], "ERROR: run_videos.py: input_dtype must be 'video' or 'itr'"
if input_dtype == "video":
from datasets.dataset_video_trace import DatasetVideoTrace as CustomDataset
elif input_dtype == "iad":
from datasets.dataset_iad_trace import DatasetIADTrace as CustomDataset
else:
from datasets.dataset_gcn_trace import DatasetGCNTrace as CustomDataset
dataset = CustomDataset(lfd_params, lfd_params.application.file_directory, mode, trace_path=lfd_params.application.trace_file, verbose=True,
backbone=model.backbone_id, num_segments=lfd_params.args.num_segments)
data_loader = create_trace_dataloader(dataset, lfd_params, mode, shuffle=False)
# put model on GPU
net = torch.nn.DataParallel(model, device_ids=lfd_params.gpus).cuda()
net.eval()
# Train Network
expected_label_list = []
predicted_label_list = []
obs_filename_list = []
trace_id = []
with torch.no_grad():
for i, data_packet in enumerate(data_loader):
obs, act, obs_filenames, _ = data_packet
obs = list(obs)
for j in range(1, act.shape[1]):
o = obs[:j]
a = act[:, :j]
# constrain size to a history of 5 timesteps
o = o[-WIN_HIST:]
a = a[:, -WIN_HIST:]
o = Batch.from_data_list(o)
# obtain label
label = a[:, -1]
label = torch.as_tensor(label).cuda()
label = torch.argmax(label, dim=1)
# hide label
a[:, -1] = 0
# compute output
logits = net(o, a.float())
# get label information
expected_label = label.cpu().detach().numpy()
predicted_label = np.argmax(logits.cpu().detach().numpy(), axis=1)
# add data to lists to be returned
expected_label_list.append(expected_label)
predicted_label_list.append(predicted_label)
obs_filename_list.append(obs_filenames[j-1])
trace_id.append(i)
if verbose:
print("file: {:3d}/{:3d}".format(i, len(data_loader)))
print("expected_label:", expected_label)
print("predicted_label:", predicted_label)
print("logits:")
print(logits.cpu().detach().numpy())
# return Pandas dataframe
return pd.DataFrame({
"expected_label": expected_label_list,
"predicted_label": predicted_label_list,
"obs_filename_list": obs_filename_list,
"trace_id": trace_id,
})
def evaluate_action_trace(lfd_params, model, mode="evaluation", verbose=False, input_dtype="video", ablation=False):
# Create DataLoaders
assert input_dtype in ["video", "iad", "gcn"], "ERROR: run_videos.py: input_dtype must be 'video' or 'itr'"
if input_dtype == "video":
from datasets.dataset_video_trace import DatasetVideoTrace as CustomDataset
elif input_dtype == "iad":
from datasets.dataset_iad_trace import DatasetIADTrace as CustomDataset
else:
from datasets.dataset_gcn_trace import DatasetGCNTrace as CustomDataset
dataset = CustomDataset(lfd_params, lfd_params.application.file_directory, mode, eval=True, trace_path=lfd_params.application.trace_file,
verbose=True, backbone=lfd_params.model.model_id, num_segments=lfd_params.input_frames,
ablation=ablation)
data_loader = create_trace_dataloader(dataset, lfd_params, mode, shuffle=False)
# put model on GPU
net = torch.nn.DataParallel(model, device_ids=lfd_params.gpus).cuda()
net.eval()
# Train Network
expected_label_list = []
predicted_label_list = []
obs_filename_list = []
with torch.no_grad():
for i, data_packet in enumerate(data_loader):
obs, act, obs_filenames, _ = data_packet
predicted_action_history = []
for j in range(1, act.shape[1]+1):
o = obs[:j]
a = act[:, :j]
o = list(o)
o = Batch.from_data_list(o)
# obtain label
label = a[:, -1]
label = torch.argmax(label, dim=1)
label = torch.as_tensor(label).cuda()
# prepare a_history
a_history = np.zeros((1, (len(predicted_action_history)+1), NUM_TOTAL_ACTIONS))
for k in range(len(predicted_action_history)):
a_history[0, k, predicted_action_history[k]] = 1
a_history = torch.from_numpy(a_history)
# compute output
logits = net(o, a_history.float())
# get label information
expected_label = label.cpu().detach().numpy()
predicted_label = np.argmax(logits.cpu().detach().numpy(), axis=1)
predicted_action_history.append(predicted_label)
# add data to lists to be returned
act = act.cpu().detach().numpy()
for j in range(act.shape[1]):
if len(expected_label_list) <= j:
expected_label_list.append([])
predicted_label_list.append([])
obs_filename_list.append([])
expected_label_list[j].append(np.argmax(act[0, j]))
predicted_label_list[j].append(predicted_action_history[j][0])
obs_filename_list[j].append(obs_filenames[j])
if verbose:
print("file: {:3d}/{:3d}".format(i, len(data_loader)))
print("expected_label:", expected_label)
print("predicted_label:", predicted_label)
print("logits:")
print(logits.cpu().detach().numpy())
df_dict = {}
for i in range(len(expected_label_list)):
df_dict["expected_label_" + str(i)] = expected_label_list[i]
df_dict["predicted_label_" + str(i)] = predicted_label_list[i]
df_dict["obs_filename_" + str(i)] = obs_filename_list[i]
# return Pandas dataframe
return pd.DataFrame(df_dict)