-
Notifications
You must be signed in to change notification settings - Fork 0
/
Trainer.py
204 lines (175 loc) · 10.7 KB
/
Trainer.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
# Created by Adam Goldbraikh - Scalpel Lab Technion
# parts of the code were adapted from: https://github.com/sj-li/MS-TCN2?utm_source=catalyzex.com
import torch
from torch import optim
import torch.nn as nn
import math
import wandb
from datetime import datetime
import tqdm
from model import RecognitionModel
from metrics import *
class Trainer:
def __init__(self, num_classes_list, device="cuda", debugging=False, **kwargs):
"""
Initiates a Trainer object. CHANGE - the arguments passed to this method were changed according to our needs.
:param num_classes_list: list of classes
:param debugging: True for debugging mode, else False
:param device: cpu or cuda
:param kwargs: key-word arguments needed for the model.
"""
self.model = RecognitionModel(device=device, **kwargs)
self.debugging = debugging
self.device = device
self.ce = nn.CrossEntropyLoss(ignore_index=-100)
self.num_classes_list = num_classes_list
def train(self, save_dir, batch_gen, num_epochs, batch_size, learning_rate, optimizer, eval_dict, split, args):
number_of_seqs = len(batch_gen.list_of_train_examples)
number_of_batches = math.ceil(number_of_seqs / batch_size)
eval_results_list = []
train_results_list = []
print(args.dataset + " " + args.group + " " + args.dataset + " dataset " + "split: " + split)
self.model.train()
self.model.to(self.device)
eval_rate = eval_dict["eval_rate"]
# CHANGE - enabled a choice of optimizers (through yaml parameters or command line arguments)
if optimizer == 1:
optimizer = optim.Adam(self.model.parameters(), lr=learning_rate)
elif optimizer == 2:
optimizer = optim.SGD(self.model.parameters(), lr=learning_rate)
else:
raise NotImplementedError()
# CHANGE - recording the best epoch so far to avoid saving models in each epoch
max_epoch = 0
max_metric = 0
for epoch in range(num_epochs):
pbar = tqdm.tqdm(total=number_of_batches)
epoch_loss = 0
correct1 = 0
total1 = 0
while batch_gen.has_next():
# CHANGE - different outputs for different usage of videos (single video input or 2 video inputs)
if args.include_video in [0, 1]:
batch_input, batch_videos, batch_target_gestures, mask = batch_gen.next_batch(batch_size)
batch_input, batch_videos, batch_target_gestures, mask = batch_input.to(self.device), batch_videos.to(self.device), \
batch_target_gestures.to(self.device), mask.to(self.device)
elif args.include_video == 2:
batch_input, batch_videos, batch_aux_videos, batch_target_gestures, mask = batch_gen.next_batch(batch_size)
batch_input, batch_videos, batch_aux_videos, batch_target_gestures, mask = batch_input.to(self.device),\
batch_videos.to(self.device),\
batch_aux_videos.to(self.device),\
batch_target_gestures.to(self.device),\
mask.to(self.device)
else:
raise NotImplementedError()
optimizer.zero_grad()
lengths = torch.sum(mask[:, 0, :], dim=1).to(dtype=torch.int64).to(device='cpu')
# CHANGE - mask is passed to model forward method (for usage with the Transformers). The lengths are
# also calculated inside the model when LSTM is used.
videos_input = [batch_videos, batch_aux_videos] if args.include_video == 2 else [batch_videos]
predictions1 = self.model(batch_input, videos_input, mask[:, 0, :])
predictions1 = (predictions1[0] * mask).unsqueeze_(0)
loss = 0
for p in predictions1:
loss += self.ce(p.transpose(2, 1).contiguous().view(-1, self.num_classes_list[0]), batch_target_gestures.view(-1))
epoch_loss += loss.item()
loss.backward()
optimizer.step()
_, predicted1 = torch.max(predictions1[-1].data, 1)
for i in range(len(lengths)):
correct1 += (predicted1[i][:lengths[i]] == batch_target_gestures[i][:lengths[i]]).float().sum().item()
total1 += lengths[i]
pbar.update(1)
batch_gen.reset()
pbar.close()
dt_string = datetime.now().strftime("%d/%m/%Y %H:%M:%S")
print(colored(dt_string, 'green',
attrs=['bold']) + " " + "[epoch %d]: train loss = %f, train acc = %f" % (epoch + 1,
epoch_loss / len(
batch_gen.list_of_train_examples),
float(correct1) / total1))
train_results = {"epoch": epoch, "train loss": epoch_loss / len(batch_gen.list_of_train_examples),
"train acc": float(correct1) / total1}
if args.upload:
wandb.log(train_results)
train_results_list.append(train_results)
if (epoch) % eval_rate == 0:
print(colored("epoch: " + str(epoch + 1) + " model evaluation", 'red', attrs=['bold']))
results = {"epoch": epoch}
results.update(self.evaluate(eval_dict, batch_gen))
eval_results_list.append(results)
if args.upload:
wandb.log(results)
# CHANGE - moved the timing of model saving to allow saving the best performing model so far
if not self.debugging:
# save the best model so far
if results["mean_metric"] > max_metric:
torch.save(self.model.state_dict(), save_dir + "/epoch-" + str(epoch + 1) + ".model")
torch.save(optimizer.state_dict(), save_dir + "/epoch-" + str(epoch + 1) + ".opt")
# remove previously saved model
if os.path.exists(save_dir + "/epoch-" + str(max_epoch + 1) + ".model") and epoch > 0:
os.remove(save_dir + "/epoch-" + str(max_epoch + 1) + ".model")
os.remove(save_dir + "/epoch-" + str(max_epoch + 1) + ".opt")
# update the the max values
max_epoch = epoch
max_metric = results["mean_metric"]
return eval_results_list, train_results_list
def evaluate(self, eval_dict, batch_gen):
results = {}
device = eval_dict["device"]
features_path = eval_dict["features_path"]
videos_path = eval_dict["videos_path"] # CHANGE - added video path to eval_dict for video features integration
include_video = eval_dict["include_video"] # CHANGE - added include_video to eval_dict for video features integration
sample_rate = eval_dict["sample_rate"]
actions_dict_gesures = eval_dict["actions_dict_gestures"]
ground_truth_path_gestures = eval_dict["gt_path_gestures"]
self.model.eval()
with torch.no_grad():
self.model.to(device)
list_of_vids = batch_gen.list_of_valid_examples
recognition1_list = []
for seq in list_of_vids:
features = np.load(features_path + seq.split('.')[0] + '.npy')
features = features[:, ::sample_rate]
input_x = torch.tensor(features, dtype=torch.float)
input_x.unsqueeze_(0)
input_x = input_x.to(device)
if include_video in [1, 2]:
side_video_path = f"{videos_path}{seq.split('.')[0]}_side.pt"
side_video_tensor = torch.load(side_video_path)
side_video_tensor.unsqueeze_(0)
input_y = side_video_tensor.to(device)
if include_video in [0, 2]:
top_video_path = f"{videos_path}{seq.split('.')[0]}_top.pt"
top_video_tensor = torch.load(top_video_path)
top_video_tensor.unsqueeze_(0)
input_z = top_video_tensor.to(device)
# CHANGE - added number of frames to deal with inconsistencies between kinematic and video features
num_frames = min(input_x.shape[2], input_y.shape[1] if include_video in [1, 2] else np.inf,
input_z.shape[1] if include_video in [0, 2] else np.inf)
# CHANGE - according to our change in the training loop, a mask should be passed to the model, instead
# of the length. Also passing the video input.
if include_video == 0:
videos_tensors = [input_z[:, :num_frames, :]]
elif include_video == 1:
videos_tensors = [input_y[:, :num_frames, :]]
elif include_video == 2:
videos_tensors = [input_y[:, :num_frames, :], input_z[:, :num_frames, :]]
else:
raise NotImplementedError()
predictions1 = self.model(input_x[:, :, :num_frames], videos_tensors, torch.ones((1, num_frames), device=device))
predictions1 = predictions1[0].unsqueeze_(0)
predictions1 = torch.nn.Softmax(dim=2)(predictions1)
_, predicted1 = torch.max(predictions1[-1].data, 1)
predicted1 = predicted1.squeeze()
recognition1 = []
for i in range(len(predicted1)):
recognition1 = np.concatenate((recognition1, [list(actions_dict_gesures.keys())[
list(actions_dict_gesures.values()).index(predicted1[i].item())]] * sample_rate))
recognition1_list.append(recognition1)
print("gestures results")
results1, _ = metric_calculation(ground_truth_path=ground_truth_path_gestures, recognition_list=recognition1_list,
list_of_videos=list_of_vids, suffix="gesture")
results.update(results1)
self.model.train()
return results