Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Clean up data loading and randomization; add scheduler #29

Closed
wants to merge 1 commit into from
Closed

Clean up data loading and randomization; add scheduler #29

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 32 additions & 36 deletions lib/audio_net.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,18 +1,22 @@
import csv
import os
import random
import time

import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
import torch.optim as optim
from numpy.random import default_rng
from torch.utils.data import DataLoader, Dataset
from torch.utils.data.sampler import SubsetRandomSampler
import os
from lib.machinelearning import *
import numpy as np
import csv

from config.config import *
import torch.optim as optim
import time
from lib.combine_models import connect_model
from lib.key_poller import KeyPoller
import random
from lib.machinelearning import *


class TinyAudioNet(nn.Module):

Expand Down Expand Up @@ -69,18 +73,20 @@ class AudioNetTrainer:
dataset_size = 0

optimizers = []
validation_loaders = []
train_loaders = []
schedulers = []
validation_loader: torch.utils.data.DataLoader = None
train_loader: torch.utils.data.DataLoader = None
criterion = nn.NLLLoss()
batch_size = 256
validation_split = .2
max_epochs = 300
random_seeds = []
random_seed = 42
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
dataset = False
train_indices = []
input_size = 120
data_seed = 42

def __init__(self, dataset, net_count = 1, audio_settings = None):
self.net_count = net_count
Expand All @@ -94,23 +100,20 @@ def __init__(self, dataset, net_count = 1, audio_settings = None):

split = int(np.floor(self.validation_split * self.dataset_size))

# Split the dataset into validation and training data loaders
dataset_rng = torch.Generator().manual_seed(self.data_seed)
train_dataset, val_dataset = torch.utils.data.random_split(dataset, [self.dataset_size - split, split], generator=dataset_rng)
Comment on lines +104 to +105
Copy link
Contributor Author

@pokey pokey Dec 23, 2022 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note that the dataset randomizer now has its own seed, and we do the split once, rather than doing it once per ensemble member


np.random.seed(self.random_seed)

for i in range(self.net_count):
self.nets.append(TinyAudioNet(self.input_size, len(self.dataset_labels), True))
self.optimizers.append(optim.SGD(self.nets[i].parameters(), lr=0.003, momentum=0.9, nesterov=True))
self.random_seeds.append(random.randint(0, 100000))


# Split the dataset into validation and training data loaders
indices = list(range(self.dataset_size))
np.random.seed(self.random_seeds[i])
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
self.train_indices.append( train_indices)

train_sampler = SubsetRandomSampler(self.train_indices[i])
valid_sampler = SubsetRandomSampler(val_indices)
self.train_loaders.append(torch.utils.data.DataLoader(dataset, batch_size=self.batch_size, sampler=train_sampler))
self.validation_loaders.append(torch.utils.data.DataLoader(dataset, batch_size=self.batch_size, sampler=valid_sampler))
optimizer = optim.SGD(self.nets[i].parameters(), lr=0.003, momentum=0.9, nesterov=True)
self.optimizers.append(optimizer)
self.schedulers.append( torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min'))
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added a scheduler


self.train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=self.batch_size, shuffle=True)
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We create a shuffling data loader, so we don't need to do shuffling during training; that happens for us automatically

self.validation_loader = torch.utils.data.DataLoader(val_dataset, batch_size=self.batch_size)

def train(self, filename):
best_accuracy = []
Expand All @@ -122,8 +125,6 @@ def train(self, filename):
starttime = int(time.time())
combined_model = TinyAudioNetEnsemble(self.nets).to(self.device)

input_size = 120

with open(REPLAYS_FOLDER + "/model_training_" + filename + str(starttime) + ".csv", 'a', newline='') as csvfile:
headers = ['epoch', 'loss', 'avg_validation_accuracy']
headers.extend(self.dataset_labels)
Expand All @@ -138,14 +139,9 @@ def train(self, filename):
running_loss.append(0.0)
self.nets[j].train(True)

# Reshuffle the indexes in the training batch to ensure the net does not memories the order of data being fed in
np.random.shuffle(self.train_indices[j])
train_sampler = SubsetRandomSampler(self.train_indices[j])
self.train_loaders[j] = torch.utils.data.DataLoader(self.dataset, batch_size=self.batch_size, sampler=train_sampler)

i = 0
with torch.set_grad_enabled(True):
for local_batch, local_labels in self.train_loaders[j]:
for local_batch, local_labels in self.train_loader:
# Transfer to GPU
local_batch, local_labels = local_batch.to(self.device), local_labels.to(self.device)

Expand Down Expand Up @@ -196,7 +192,7 @@ def train(self, filename):
accuracy_batch['correct'][dataset_label] = 0
accuracy_batch['percent'][dataset_label] = 0

for local_batch, local_labels in self.validation_loaders[j]:
for local_batch, local_labels in self.validation_loader:
# Transfer to GPU
local_batch, local_labels = local_batch.to(self.device), local_labels.to(self.device)

Expand Down Expand Up @@ -228,6 +224,7 @@ def train(self, filename):
for j in range(self.net_count):
epoch_loss.append(epoch_validation_loss[j] / ( self.dataset_size * self.validation_split ) )
accuracy.append( correct[j] / ( self.dataset_size * self.validation_split ) )
self.schedulers[j].step(epoch_loss[j])
print('[Net: %d] Validation loss: %.4f accuracy %.3f' % (j + 1, epoch_loss[j], accuracy[j]))

print('[Combined] Sum validation loss: %.4f average accuracy %.3f' % (np.sum(epoch_loss), combined_correct / ( self.dataset_size * self.validation_split )))
Expand All @@ -253,7 +250,6 @@ def train(self, filename):
'last_row': csv_row,
'loss': epoch_loss[j],
'epoch': epoch,
'random_seed': self.random_seeds[j],
}, os.path.join(CLASSIFIER_FOLDER, current_filename) + '-weights.pth.tar')

# Persist a new combined model with the best weights if new best weights are given
Expand Down