-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_dsp_h5file.py
168 lines (138 loc) · 6.52 KB
/
train_dsp_h5file.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
import matplotlib.pyplot as plt
from utils_dataset import set_random_seed, Dataset_features, MNISTDataset_current
from torch import torch
from torchvision.datasets import MNIST
from torch.utils.data import DataLoader
import numpy as np
from pathlib import Path
import tqdm
import os
import torch
from tqdm import trange
import argparse
import h5py
import torch.nn as nn
from utils_dataset import extract_interval
import json
def train_classifier(train_dl, test_dl, n_epochs, device):
examples = enumerate(train_dl)
batch_idx, (example_data, example_targets) = next(examples)
classifier = nn.Sequential(nn.Linear(example_data.shape[1], train_dl.dataset.n_classes, )).to(device)
optimizer = torch.optim.Adam(classifier.parameters(), lr=0.001)
epochs = tqdm.trange(n_epochs, desc=f"Classifier", leave=False, position=1)
batches = tqdm.tqdm(train_dl, desc="Epoch", disable=False)
loss = nn.CrossEntropyLoss()
loss_coll = []
acc_coll = []
for epoch in range(n_epochs):
loss_list = []
acc_list = []
batches.reset()
batches.set_description('Training')
for batch_idx, (data, target) in enumerate(train_dl):
target = target.squeeze().long()
out = classifier(data.float())
loss_val = loss(out, target)
loss_val.backward()
optimizer.step()
optimizer.zero_grad()
loss_list.append(loss_val.item())
batches.update()
batches.reset(total=len(test_dl))
batches.set_description('Testing')
with torch.no_grad():
for batch_idx, (data, target) in enumerate(test_dl):
data = data.float()
target = target.long()
out = classifier(data)
acc = torch.mean((torch.argmax(out, dim=1) == target).to(torch.int16), dtype=torch.float)
acc_list.append(acc.item())
# if epoch == (n_epochs - 1):
# list_acc_last_epoch.append(acc.tolist())
batches.update()
loss_coll.append(np.mean(loss_list))
acc_coll.append(np.mean(acc_list))
epochs.set_postfix_str(f"Loss: {np.mean(loss_list):.3f}, Acc: {np.mean(acc_list):.3f}")
epochs.update()
return acc_list # return test accuracy of last training epoch, last batch
def main(args):
print(f'Running with data type on {args.data_type}')
# Device: ----------------------------------------------------------------------------------------------------------
device = torch.device(args.device)
if device == torch.device("cuda") and not (torch.cuda.is_available()):
device == torch.device('cpu')
print('Could not find CUDA. Running on CPU')
print(f'Running on {device}')
if device == torch.device("cuda"):
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.empty_cache()
else:
torch.set_default_tensor_type('torch.FloatTensor')
# Folders: ---------------------------------------------------------------------------------------------------------
folder = Path('dataset_analysis')
folder_run = Path(os.path.join(folder, args.dataset))
folder_fig = folder_run.joinpath('fig')
folder_run.mkdir(parents=True, exist_ok=True)
folder_fig.mkdir(parents=True, exist_ok=True)
folder_data = folder_run.joinpath('data')
folder_data.mkdir(parents=True, exist_ok=True)
# Dataseet: --------------------------------------------------------------------------------------------------------
generator = set_random_seed(args.seed, add_generator=True, device=device)
path_to_dataset = Path(args.path_to_dataset)
path_to_dataset.mkdir(parents=True, exist_ok=True)
dict_dataset = {}
df = {'center': [], 'span': [], 'accuracy': []}
list_datasets = os.listdir(path_to_dataset.joinpath(args.dataset, args.data_type))
sweep = tqdm.tqdm(total=len(list_datasets), desc=f"{args.data_type} sweeping", position=0, leave=True)
for dataset_name in list_datasets:
filename_dataset = path_to_dataset.joinpath(args.dataset,
args.data_type,
dataset_name)
train_dataset = Dataset_features(h5py.File(filename_dataset.joinpath('train.h5'), 'r'), device=device)
test_dataset = Dataset_features(h5py.File(filename_dataset.joinpath('test.h5'), 'r'), device=device)
dict_dataset['train_loader'] = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
generator=generator)
dict_dataset['test_loader'] = DataLoader(test_dataset,
batch_size=args.batch_size,
shuffle=True,
generator=generator)
acc = train_classifier(dict_dataset['train_loader'], dict_dataset['test_loader'], args.n_epochs, device)
df['center'].extend([train_dataset.center] * len(acc))
df['span'].extend([train_dataset.span] * len(acc))
df['accuracy'].extend(acc)
sweep.update()
with open(path_to_dataset.joinpath(args.dataset, f'{args.data_type}.json'), 'w') as f:
# write the dictionary to the file in JSON format
json.dump(df, f)
print('Done')
if __name__ == "__main__":
parser = argparse.ArgumentParser('dataloader')
parser.add_argument('--seed',
type=int,
default=10)
parser.add_argument('--dataset',
type=str,
default='Braille',
choices=['MNIST', 'Braille'])
parser.add_argument('--data_type',
type=str,
default='frequency',
choices=['frequency', 'amplitude', 'slope'])
parser.add_argument('--batch_size',
type=int,
default=128)
parser.add_argument('--n_epochs',
type=int,
default=100)
parser.add_argument('--device',
type=str,
choices=['cpu', 'cuda'],
default='cpu')
parser.add_argument('--path_to_dataset',
type=str,
default='/media/p308783/bics/Nicoletta/tactile_encoding/')
args = parser.parse_args()
# Run:
main(args)