-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathmain_dmm_causal.py
201 lines (163 loc) · 9.86 KB
/
main_dmm_causal.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
#####################################################
# Creator: Anubhab Ghosh
# Feb 2023
#####################################################
# Import necessary libraries
import sys
import os
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(os.path.dirname(SCRIPT_DIR))
import argparse
from parse import parse
import numpy as np
import json
from utils.utils import NDArrayEncoder
#import matplotlib.pyplot as plt
import torch
import pickle as pkl
from torch import nn
from torch.utils.data import DataLoader, Dataset
from utils.utils import load_saved_dataset, Series_Dataset, obtain_tr_val_test_idx, create_splits_file_name, \
create_file_paths, check_if_dir_or_file_exists, load_splits_file, get_dataloaders, NDArrayEncoder
# Import the parameters
from config.parameters_opt import get_parameters, get_H_DANSE
#from utils.plot_functions import plot_measurement_data, plot_measurement_data_axes, plot_state_trajectory, plot_state_trajectory_axes
# Import estimator model and functions
from src.dmm_causal import DMM, train_dmm
def main():
usage = "Train a causal DMM using trajectories of SSMs \n"\
"python3.8 main_danse.py --mode [train/test] --rnn_model_type [gru/lstm/rnn] --dataset_type [LinearSSM/LorenzSSM] \n"\
"--datafile [fullpath to datafile] --splits [fullpath to splits file]"
parser = argparse.ArgumentParser(description="Input a string indicating the mode of the script \n"\
"train - training and testing is done, test-only evlaution is carried out")
parser.add_argument("--mode", help="Enter the desired mode", type=str)
parser.add_argument("--rnn_model_type", help="Enter the desired model (rnn/lstm/gru)", type=str)
parser.add_argument("--dataset_type", help="Enter the type of dataset (LinearSSM/LorenzSSM)", type=str)
parser.add_argument("--model_file_saved", help="In case of testing mode, Enter the desired model checkpoint with full path (gru/lstm/rnn)", type=str, default=None)
parser.add_argument("--datafile", help="Enter the full path to the dataset", type=str)
parser.add_argument("--splits", help="Enter full path to splits file", type=str)
args = parser.parse_args()
mode = args.mode
model_type = args.rnn_model_type
datafile = args.datafile
dataset_type = args.dataset_type
datafolder = "".join(datafile.split("/")[i]+"/" for i in range(len(datafile.split("/")) - 1))
model_file_saved = args.model_file_saved
splits_file = args.splits
print("datafile name: {}".format(datafile.split('/')[-1]))
# Dataset parameters obtained from the 'datafile' variable
_, n_states, n_obs, _, T, N_samples, sigma_e2_dB, smnr_dB = parse("{}_m_{:d}_n_{:d}_{}_data_T_{:d}_N_{:d}_sigmae2_{:f}dB_smnr_{:f}dB.pkl", datafile.split('/')[-1])
ngpu = 1 # Comment this out if you want to run on cpu and the next line just set device to "cpu"
device = torch.device("cuda:0" if (torch.cuda.is_available() and ngpu>0) else "cpu")
print("Device Used:{}".format(device))
ssm_parameters_dict, est_parameters_dict = get_parameters(
n_states=n_states,
n_obs=n_obs,
device=device
)
batch_size = est_parameters_dict["dmm"]["batch_size"] # Set the batch size
estimator_options = est_parameters_dict["dmm"] # Get the options for the estimator
if not os.path.isfile(datafile):
print("Dataset is not present, run 'generate_data.py / run_generate_data.sh' to create the dataset")
#plot_trajectories(Z_pM, ncols=1, nrows=10)
else:
print("Dataset already present!")
Z_XY = load_saved_dataset(filename=datafile)
Z_XY_dataset = Series_Dataset(Z_XY_dict=Z_XY)
ssm_model = Z_XY["ssm_model"]
estimator_options['C_w'] = ssm_model.Cw # Get the covariance matrix of the measurement noise from the model information
estimator_options['H'] = get_H_DANSE(type_=dataset_type, n_states=n_states, n_obs=n_obs) # Get the sensing matrix from the model info
print(estimator_options['H'], estimator_options['C_w'])
if not os.path.isfile(splits_file):
tr_indices, val_indices, test_indices = obtain_tr_val_test_idx(dataset=Z_XY_dataset,
tr_to_test_split=0.9,
tr_to_val_split=0.833)
print(len(tr_indices), len(val_indices), len(test_indices))
splits = {}
splits["train"] = tr_indices
splits["val"] = val_indices
splits["test"] = test_indices
splits_file_name = create_splits_file_name(dataset_filename=datafile,
splits_filename=splits_file
)
print("Creating split file at:{}".format(splits_file_name))
with open(splits_file_name, 'wb') as handle:
pkl.dump(splits, handle, protocol=pkl.HIGHEST_PROTOCOL)
else:
print("Loading the splits file from {}".format(splits_file))
splits = load_splits_file(splits_filename=splits_file)
tr_indices, val_indices, test_indices = splits["train"], splits["val"], splits["test"]
train_loader, val_loader, test_loader = get_dataloaders(dataset=Z_XY_dataset,
batch_size=batch_size,
tr_indices=tr_indices,
val_indices=val_indices,
test_indices=test_indices)
print("No. of training, validation and testing batches: {}, {}, {}".format(len(train_loader),
len(val_loader),
len(test_loader)))
#ngpu = 1 # Comment this out if you want to run on cpu and the next line just set device to "cpu"
#device = torch.device("cuda:0" if (torch.cuda.is_available() and ngpu>0) else "cpu")
#print("Device Used:{}".format(device))
logfile_path = "./log/"
modelfile_path = "./models/"
#NOTE: Currently this is hardcoded into the system
main_exp_name = "{}_dmm_{}_opt_{}_m_{}_n_{}_T_{}_N_{}_sigmae2_{}dB_smnr_{}dB_lmindel".format(
dataset_type,
estimator_options["inference_mode"],
model_type,
n_states,
n_obs,
T,
N_samples,
sigma_e2_dB,
smnr_dB
)
#print(params)
tr_log_file_name = "training.log"
te_log_file_name = "testing.log"
flag_log_dir, flag_log_file = check_if_dir_or_file_exists(os.path.join(logfile_path, main_exp_name),
file_name=tr_log_file_name)
print("Is log-directory present:? - {}".format(flag_log_dir))
print("Is log-file present:? - {}".format(flag_log_file))
flag_models_dir, _ = check_if_dir_or_file_exists(os.path.join(modelfile_path, main_exp_name),
file_name=None)
print("Is model-directory present:? - {}".format(flag_models_dir))
#print("Is file present:? - {}".format(flag_file))
tr_logfile_name_with_path = os.path.join(os.path.join(logfile_path, main_exp_name), tr_log_file_name)
te_logfile_name_with_path = os.path.join(os.path.join(logfile_path, main_exp_name), te_log_file_name)
if flag_log_dir == False:
print("Creating {}".format(os.path.join(logfile_path, main_exp_name)))
os.makedirs(os.path.join(logfile_path, main_exp_name), exist_ok=True)
if flag_models_dir == False:
print("Creating {}".format(os.path.join(modelfile_path, main_exp_name)))
os.makedirs(os.path.join(modelfile_path, main_exp_name), exist_ok=True)
modelfile_path = os.path.join(modelfile_path, main_exp_name) # Modify the modelfile path to add full model file
if mode.lower() == "train":
model_dmm = DMM(**estimator_options)
tr_verbose = True
# Starting model training
tr_losses, val_losses, _, _, _ = train_dmm(
model=model_dmm,
train_loader=train_loader,
val_loader=val_loader,
options=estimator_options,
optimizer_params=estimator_options['optimizer_params'],
logfile_path=tr_logfile_name_with_path,
modelfile_path=modelfile_path,
save_chkpoints="some",
device=device,
tr_verbose=tr_verbose
)
#if tr_verbose == True:
# plot_losses(tr_losses=tr_losses, val_losses=val_losses, logscale=False)
losses_model = {}
losses_model["tr_losses"] = tr_losses
losses_model["val_losses"] = val_losses
with open(os.path.join(os.path.join(logfile_path, main_exp_name),
'dmm_{}_{}_losses_eps{}.json'.format(estimator_options['inference_mode'],
estimator_options['rnn_model_type'],
estimator_options['optimizer_params']['num_epochs'],)), 'w') as f:
f.write(json.dumps(losses_model, cls=NDArrayEncoder, indent=2))
return None
if __name__ == "__main__":
main()