forked from sj-li/MS-TCN2
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_predict_eval.py
188 lines (148 loc) · 6.86 KB
/
train_predict_eval.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
import torch
from model import Trainer
from batch_gen import BatchGenerator
import numpy as np
import os
import argparse
import random
from clearml import Task
from eval import eval
task = Task.init(project_name='ProjectCV', task_name='TrainPredictEval', reuse_last_task_id=False)
task.set_user_properties(
{"name": "backbone", "description": "network type", "value": "mstcn++"}
)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
seed = 1538574472
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', default="valid")
parser.add_argument('--fold', default='1')
parser.add_argument('--features_dim', default='1280', type=int)
parser.add_argument('--bz', default='1', type=int)
parser.add_argument('--lr', default='0.0005', type=float)
parser.add_argument('--num_f_maps', default='64', type=int)
# Need input
parser.add_argument('--num_epochs', type=int)
parser.add_argument('--num_layers_PG', type=int)
parser.add_argument('--num_layers_R', type=int)
parser.add_argument('--num_R', type=int)
parser.add_argument('--weight_type', type=str)
parser.add_argument('--experimental', type=int)
args = parser.parse_args()
dataset = args.dataset
folds = args.fold.split(",")
# print(folds)
num_epochs = args.num_epochs
features_dim = args.features_dim
bz = args.bz
lr = args.lr
num_layers_PG = args.num_layers_PG
num_layers_R = args.num_layers_R
num_R = args.num_R
num_f_maps = args.num_f_maps
weight_type = args.weight_type
experimental = args.experimental
learn_from_domain = False
w = None
if weight_type == "framewise": # give more weight to lower resolution dilations
w = torch.tensor(np.array(range(num_layers_R, 0, -1)) / sum(range(num_layers_R, 0, -1)), dtype=torch.float32, device=device)
elif weight_type == "smooth": # give more weight to higher resolution dilations
w = torch.tensor(np.array(range(1, num_layers_R + 1)) / sum(range(1, num_layers_R + 1)), dtype=torch.float32, device=device)
elif weight_type == "none":
w = None
elif weight_type == "learned":
w = num_layers_R # here you can control the initialization of the learned parameters if you modify the Weighting class properly
elif weight_type == "uniform": # weight equally all dilations
w = torch.tensor(np.ones(num_layers_R) / num_layers_R, dtype=torch.float32, device=device)
elif weight_type == "learned_smooth": # weight equally all dilations
w = torch.tensor(np.array(range(1, num_layers_R + 1)) / sum(range(1, num_layers_R + 1)), dtype=torch.float32, device=device)
learn_from_domain = True
elif weight_type == "learned_framewise": # weight equally all dilations
w = torch.tensor(np.array(range(num_layers_R, 0, -1)) / sum(range(num_layers_R, 0, -1)), dtype=torch.float32, device=device)
learn_from_domain = True
elif weight_type == "learned_uniform": # weight equally all dilations
w = torch.tensor(np.ones(num_layers_R) / num_layers_R, dtype=torch.float32, device=device)
learn_from_domain = True
elif weight_type == "learned_framewise_exp": # weight equally all dilations
arr = np.array(range(num_layers_R, 0, -1))
arr_exps = np.exp(arr)
arr_exps = arr_exps/sum(arr_exps)
w = torch.tensor(arr_exps, dtype=torch.float32, device=device)
learn_from_domain = True
elif weight_type == "learned_smooth_exp": # weight equally all dilations
arr = np.array(range(1, num_layers_R + 1))
arr_exps = np.exp(arr)
arr_exps = arr_exps / sum(arr_exps)
w = torch.tensor(arr_exps, dtype=torch.float32, device=device)
learn_from_domain = True
elif weight_type == "learned_poly": # weight polynomial at center
arr = np.arange(0, 1, step=0.1)
arr_exps = (arr - 0.45) ** 2 + 0.5
arr_exps = arr_exps / sum(arr_exps)
w = torch.tensor(arr_exps, dtype=torch.float32, device=device)
learn_from_domain = True
print("weight:{}".format(w))
sample_rate = 1
mapping_file = "/datashare/APAS/mapping_gestures.txt"
gt_path = '/datashare/APAS/transcriptions_gestures/'
file_ptr = open(mapping_file, 'r')
actions = file_ptr.read().split('\n')[:-1]
file_ptr.close()
actions_dict = dict()
for a in actions:
actions_dict[a.split()[1]] = int(a.split()[0])
num_classes = len(actions_dict)
fold_files = [(f"/datashare/APAS/folds/valid {i}.txt",
f"/datashare/APAS/folds/test {i}.txt",
f"/datashare/APAS/features/fold{i}/") for i in folds]
def fold_split(features_path, val_path, test_path, foldi):
with open(val_path, 'r') as f:
val_files = [vid.split('.')[0] + '.npy' for vid in f.readlines()]
with open(test_path, 'r') as f:
test_files = [vid.split('.')[0] + '.npy' for vid in f.readlines()]
other_test_files = []
for i in range(5):
if i!= foldi:
curr_path = f"/datashare/APAS/folds/test {i}.txt"
with open(curr_path, 'r') as f:
curr_files = [vid.split('.')[0] + '.npy' for vid in f.readlines()]
other_test_files.extend(curr_files)
train_files = list(set(other_test_files) - set(test_files + val_files))
print('train files:')
print(train_files)
print('val files:')
print(val_files)
print('test files:')
print(test_files)
return train_files, val_files, test_files
test_files = []
i = 0
for val_path_fold, test_path_fold, features_path_fold in fold_files:
print(f"{dataset} : {i}")
vid_list_file, vid_list_file_val, vid_list_file_test = fold_split(features_path_fold, val_path_fold,
test_path_fold, foldi=i)
fold = features_path_fold.split("/")[-2]
i+=1
model_dir = f"./models/test/{dataset}{fold}"
results_dir = f"./results/test/{dataset}{fold}"
if not os.path.exists(model_dir):
os.makedirs(model_dir)
if not os.path.exists(results_dir):
os.makedirs(results_dir)
trainer = Trainer(num_layers_PG, num_layers_R, num_R, num_f_maps, features_dim, num_classes, f"fold{fold}", f"fold{fold}",
refinement_weighting=w, experimental=experimental, fold=fold, learn_from_domain=learn_from_domain)
batch_gen_train = BatchGenerator(num_classes, actions_dict, gt_path, features_path_fold,
sample_rate)
batch_gen_train.read_data(vid_list_file)
batch_gen_val = BatchGenerator(num_classes, actions_dict, gt_path, features_path_fold, sample_rate)
batch_gen_val.read_data(vid_list_file_val)
trainer.train(model_dir, batch_gen_train, batch_gen_val, num_epochs=num_epochs, batch_size=bz, learning_rate=lr,
device=device, weighting_method=weight_type)
trainer.predict(model_dir, results_dir, features_path_fold, vid_list_file_test, num_epochs, actions_dict, device,
sample_rate, weighting_method=weight_type)
test_files.append(vid_list_file_test)
eval(dataset, folds, test_files, weight_type)