feat(model, train): improved architecture, overfit prevention, re-trained the model
Browse files- README.md +36 -0
- model.py +23 -18
- performance.json +649 -115
- performance_plot.png +0 -0
- plots.py +3 -3
- predictions.csv +0 -0
- train.py +0 -394
- train_dist.py +40 -7
README.md
CHANGED
|
@@ -1,2 +1,38 @@
|
|
| 1 |
# Places-ResNet
|
|
|
|
| 2 |
My experiment training a ResNet-inspired model for image classification using PyTorch.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# Places-ResNet
|
| 2 |
+
|
| 3 |
My experiment training a ResNet-inspired model for image classification using PyTorch.
|
| 4 |
+
|
| 5 |
+
**Key terms: distributed training, residual layers, convolutional layers, batch normalization, dropout, pooling, SGD, label smoothing, learning rate scheduling, early stopping, data augmentation.**
|
| 6 |
+
|
| 7 |
+
Training time was approximately 10 hours (108 epochs) using **distributed training** across university server GPUs.
|
| 8 |
+
|
| 9 |
+
## Dataset:
|
| 10 |
+
|
| 11 |
+
**MIT MiniPlaces Dataset:** Contains 100,000 training images, 10,000 validation images, and 10,000 testing images across 100 scene categories. Each image is 128x128 pixels.
|
| 12 |
+
|
| 13 |
+
## Model:
|
| 14 |
+
|
| 15 |
+
I implemented a 13-layer ResNet-inspired model for image classification. The architecture consists of:
|
| 16 |
+
- Initial **convolutional layer** with 64 filters, followed by **batch normalization, max pooling, and dropout**
|
| 17 |
+
- 3 stages of **residual blocks**, each with 4 convolutional layers
|
| 18 |
+
- Each **residual block** has two 3x3 **convolutional layers** with **batch normalization and dropout**
|
| 19 |
+
- The number of filters increases from 64 in the first stage, to 128, 256, and 512 in the later stages
|
| 20 |
+
- **Global average pooling and dropout** before a final **fully connected layer**
|
| 21 |
+
|
| 22 |
+
The total number of trainable model parameters is 29,678,180.
|
| 23 |
+
|
| 24 |
+
## Training:
|
| 25 |
+
|
| 26 |
+
The training setup used a **distributed training** approach, with **early stopping** to prevent overfitting. **Data augmentation** techniques were applied to the training and validation sets. An **SGD optimizer** with **label smoothing** was used, along with a **ReduceLROnPlateau learning rate scheduler**.
|
| 27 |
+
|
| 28 |
+
## Performance:
|
| 29 |
+
|
| 30 |
+
Best model checkpoint results (epoch 108):
|
| 31 |
+
|
| 32 |
+
- Training Loss: 2.3231, Training Accuracy: 53.02%
|
| 33 |
+
- Validation Loss: 2.3426, Validation Accuracy: 54.09%
|
| 34 |
+
- Top-5 Validation Accuracy: 81.34%
|
| 35 |
+
|
| 36 |
+
Achieving a **Top-1 accuracy of 54.09% and Top-5 accuracy of 81.34%**
|
| 37 |
+
|
| 38 |
+
|
model.py
CHANGED
|
@@ -3,14 +3,15 @@ import torch.nn as nn
|
|
| 3 |
|
| 4 |
|
| 5 |
class ResidualBlock(nn.Module):
|
| 6 |
-
def __init__(self, in_channels, out_channels):
|
| 7 |
super(ResidualBlock, self).__init__()
|
| 8 |
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
|
| 9 |
self.bn1 = nn.BatchNorm2d(out_channels)
|
|
|
|
| 10 |
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
|
| 11 |
self.bn2 = nn.BatchNorm2d(out_channels)
|
|
|
|
| 12 |
|
| 13 |
-
# Skip connection (identity mapping)
|
| 14 |
self.skip_connection = nn.Sequential()
|
| 15 |
if in_channels != out_channels:
|
| 16 |
self.skip_connection = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
|
|
@@ -18,53 +19,57 @@ class ResidualBlock(nn.Module):
|
|
| 18 |
def forward(self, x):
|
| 19 |
residual = self.skip_connection(x)
|
| 20 |
out = nn.functional.relu(self.bn1(self.conv1(x)))
|
|
|
|
| 21 |
out = self.bn2(self.conv2(out))
|
| 22 |
-
out
|
|
|
|
| 23 |
out = nn.functional.relu(out)
|
| 24 |
return out
|
| 25 |
|
| 26 |
|
| 27 |
class MyModel(nn.Module):
|
| 28 |
-
def __init__(self, num_classes=100):
|
| 29 |
super(MyModel, self).__init__()
|
|
|
|
| 30 |
|
| 31 |
-
# Initial convolutional layer
|
| 32 |
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)
|
| 33 |
self.bn1 = nn.BatchNorm2d(64)
|
| 34 |
self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
|
|
|
|
| 35 |
|
| 36 |
-
#
|
| 37 |
-
self.block1 = self._resnet_layers(64, 128, num_blocks=
|
| 38 |
-
self.block2 = self._resnet_layers(128, 256, num_blocks=
|
| 39 |
-
self.block3 = self._resnet_layers(256, 512, num_blocks=
|
| 40 |
|
| 41 |
-
# Global average pooling
|
| 42 |
self.global_avg_pool = nn.AdaptiveAvgPool2d(1)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
|
| 44 |
-
# Combine features
|
| 45 |
self.features = nn.Sequential(
|
| 46 |
self.conv1,
|
| 47 |
self.bn1,
|
| 48 |
nn.ReLU(),
|
| 49 |
self.pool1,
|
|
|
|
| 50 |
self.block1,
|
| 51 |
self.block2,
|
| 52 |
self.block3,
|
| 53 |
-
self.global_avg_pool
|
|
|
|
| 54 |
)
|
| 55 |
|
| 56 |
-
# Fully connected layer
|
| 57 |
-
self.fc = nn.Linear(512, num_classes)
|
| 58 |
-
|
| 59 |
@staticmethod
|
| 60 |
def _resnet_layers(in_channels, out_channels, num_blocks):
|
| 61 |
return nn.Sequential(
|
| 62 |
-
ResidualBlock(in_channels, out_channels),
|
| 63 |
-
*[ResidualBlock(out_channels, out_channels) for _ in range(num_blocks)]
|
| 64 |
)
|
| 65 |
|
| 66 |
def forward(self, x):
|
| 67 |
x = self.features(x)
|
| 68 |
-
x = torch.flatten(x, 1)
|
| 69 |
x = self.fc(x)
|
| 70 |
return x
|
|
|
|
| 3 |
|
| 4 |
|
| 5 |
class ResidualBlock(nn.Module):
|
| 6 |
+
def __init__(self, in_channels, out_channels, dropout_rate=0.2):
|
| 7 |
super(ResidualBlock, self).__init__()
|
| 8 |
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1)
|
| 9 |
self.bn1 = nn.BatchNorm2d(out_channels)
|
| 10 |
+
self.dropout1 = nn.Dropout2d(p=dropout_rate)
|
| 11 |
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1)
|
| 12 |
self.bn2 = nn.BatchNorm2d(out_channels)
|
| 13 |
+
self.dropout2 = nn.Dropout2d(p=dropout_rate)
|
| 14 |
|
|
|
|
| 15 |
self.skip_connection = nn.Sequential()
|
| 16 |
if in_channels != out_channels:
|
| 17 |
self.skip_connection = nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
|
|
|
|
| 19 |
def forward(self, x):
|
| 20 |
residual = self.skip_connection(x)
|
| 21 |
out = nn.functional.relu(self.bn1(self.conv1(x)))
|
| 22 |
+
out = self.dropout1(out)
|
| 23 |
out = self.bn2(self.conv2(out))
|
| 24 |
+
out = self.dropout2(out)
|
| 25 |
+
out += residual
|
| 26 |
out = nn.functional.relu(out)
|
| 27 |
return out
|
| 28 |
|
| 29 |
|
| 30 |
class MyModel(nn.Module):
|
| 31 |
+
def __init__(self, num_classes=100, dropout_rate=0.2):
|
| 32 |
super(MyModel, self).__init__()
|
| 33 |
+
self.dropout_rate = dropout_rate
|
| 34 |
|
|
|
|
| 35 |
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)
|
| 36 |
self.bn1 = nn.BatchNorm2d(64)
|
| 37 |
self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
|
| 38 |
+
self.dropout1 = nn.Dropout2d(p=self.dropout_rate)
|
| 39 |
|
| 40 |
+
# Increase the number of residual blocks
|
| 41 |
+
self.block1 = self._resnet_layers(64, 128, num_blocks=4)
|
| 42 |
+
self.block2 = self._resnet_layers(128, 256, num_blocks=4)
|
| 43 |
+
self.block3 = self._resnet_layers(256, 512, num_blocks=4)
|
| 44 |
|
|
|
|
| 45 |
self.global_avg_pool = nn.AdaptiveAvgPool2d(1)
|
| 46 |
+
self.dropout2 = nn.Dropout(p=self.dropout_rate)
|
| 47 |
+
|
| 48 |
+
# Reduce the size of the fully connected layer
|
| 49 |
+
self.fc = nn.Linear(512, num_classes)
|
| 50 |
|
|
|
|
| 51 |
self.features = nn.Sequential(
|
| 52 |
self.conv1,
|
| 53 |
self.bn1,
|
| 54 |
nn.ReLU(),
|
| 55 |
self.pool1,
|
| 56 |
+
self.dropout1,
|
| 57 |
self.block1,
|
| 58 |
self.block2,
|
| 59 |
self.block3,
|
| 60 |
+
self.global_avg_pool,
|
| 61 |
+
self.dropout2
|
| 62 |
)
|
| 63 |
|
|
|
|
|
|
|
|
|
|
| 64 |
@staticmethod
|
| 65 |
def _resnet_layers(in_channels, out_channels, num_blocks):
|
| 66 |
return nn.Sequential(
|
| 67 |
+
ResidualBlock(in_channels, out_channels, dropout_rate=0.2),
|
| 68 |
+
*[ResidualBlock(out_channels, out_channels, dropout_rate=0.2) for _ in range(num_blocks)]
|
| 69 |
)
|
| 70 |
|
| 71 |
def forward(self, x):
|
| 72 |
x = self.features(x)
|
| 73 |
+
x = torch.flatten(x, 1)
|
| 74 |
x = self.fc(x)
|
| 75 |
return x
|
performance.json
CHANGED
|
@@ -1,176 +1,710 @@
|
|
| 1 |
[
|
| 2 |
{
|
| 3 |
-
"avg_train_loss": 4.
|
| 4 |
-
"train_accuracy": 0.
|
| 5 |
-
"avg_val_loss":
|
| 6 |
-
"val_accuracy": 0.
|
| 7 |
},
|
| 8 |
{
|
| 9 |
-
"avg_train_loss":
|
| 10 |
-
"train_accuracy": 0.
|
| 11 |
-
"avg_val_loss": 3.
|
| 12 |
-
"val_accuracy": 0.
|
| 13 |
},
|
| 14 |
{
|
| 15 |
-
"avg_train_loss":
|
| 16 |
-
"train_accuracy": 0.
|
| 17 |
-
"avg_val_loss": 3.
|
| 18 |
-
"val_accuracy": 0.
|
| 19 |
},
|
| 20 |
{
|
| 21 |
-
"avg_train_loss": 3.
|
| 22 |
-
"train_accuracy": 0.
|
| 23 |
-
"avg_val_loss": 3.
|
| 24 |
-
"val_accuracy": 0.
|
| 25 |
},
|
| 26 |
{
|
| 27 |
-
"avg_train_loss": 3.
|
| 28 |
-
"train_accuracy": 0.
|
| 29 |
-
"avg_val_loss": 3.
|
| 30 |
-
"val_accuracy": 0.
|
| 31 |
},
|
| 32 |
{
|
| 33 |
-
"avg_train_loss": 3.
|
| 34 |
-
"train_accuracy": 0.
|
| 35 |
-
"avg_val_loss": 3.
|
| 36 |
-
"val_accuracy": 0.
|
| 37 |
},
|
| 38 |
{
|
| 39 |
-
"avg_train_loss": 3.
|
| 40 |
-
"train_accuracy": 0.
|
| 41 |
-
"avg_val_loss": 3.
|
| 42 |
-
"val_accuracy": 0.
|
| 43 |
},
|
| 44 |
{
|
| 45 |
-
"avg_train_loss": 3.
|
| 46 |
-
"train_accuracy": 0.
|
| 47 |
-
"avg_val_loss": 3.
|
| 48 |
-
"val_accuracy": 0.
|
| 49 |
},
|
| 50 |
{
|
| 51 |
-
"avg_train_loss":
|
| 52 |
-
"train_accuracy": 0.
|
| 53 |
-
"avg_val_loss": 3.
|
| 54 |
-
"val_accuracy": 0.
|
| 55 |
},
|
| 56 |
{
|
| 57 |
-
"avg_train_loss":
|
| 58 |
-
"train_accuracy": 0.
|
| 59 |
-
"avg_val_loss": 3.
|
| 60 |
-
"val_accuracy": 0.
|
| 61 |
},
|
| 62 |
{
|
| 63 |
-
"avg_train_loss":
|
| 64 |
-
"train_accuracy": 0.
|
| 65 |
-
"avg_val_loss": 3.
|
| 66 |
-
"val_accuracy": 0.
|
| 67 |
},
|
| 68 |
{
|
| 69 |
-
"avg_train_loss":
|
| 70 |
-
"train_accuracy": 0.
|
| 71 |
-
"avg_val_loss":
|
| 72 |
-
"val_accuracy": 0.
|
| 73 |
},
|
| 74 |
{
|
| 75 |
-
"avg_train_loss":
|
| 76 |
-
"train_accuracy": 0.
|
| 77 |
-
"avg_val_loss":
|
| 78 |
-
"val_accuracy": 0.
|
| 79 |
},
|
| 80 |
{
|
| 81 |
-
"avg_train_loss":
|
| 82 |
-
"train_accuracy": 0.
|
| 83 |
-
"avg_val_loss":
|
| 84 |
-
"val_accuracy": 0.
|
| 85 |
},
|
| 86 |
{
|
| 87 |
-
"avg_train_loss":
|
| 88 |
-
"train_accuracy": 0.
|
| 89 |
-
"avg_val_loss":
|
| 90 |
-
"val_accuracy": 0.
|
| 91 |
},
|
| 92 |
{
|
| 93 |
-
"avg_train_loss":
|
| 94 |
-
"train_accuracy": 0.
|
| 95 |
-
"avg_val_loss":
|
| 96 |
-
"val_accuracy": 0.
|
| 97 |
},
|
| 98 |
{
|
| 99 |
-
"avg_train_loss":
|
| 100 |
-
"train_accuracy": 0.
|
| 101 |
-
"avg_val_loss":
|
| 102 |
-
"val_accuracy": 0.
|
| 103 |
},
|
| 104 |
{
|
| 105 |
-
"avg_train_loss":
|
| 106 |
-
"train_accuracy": 0.
|
| 107 |
-
"avg_val_loss":
|
| 108 |
-
"val_accuracy": 0.
|
| 109 |
},
|
| 110 |
{
|
| 111 |
-
"avg_train_loss":
|
| 112 |
-
"train_accuracy": 0.
|
| 113 |
-
"avg_val_loss":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 114 |
"val_accuracy": 0.4027000069618225
|
| 115 |
},
|
| 116 |
{
|
| 117 |
-
"avg_train_loss": 2.
|
| 118 |
-
"train_accuracy": 0.
|
| 119 |
-
"avg_val_loss": 2.
|
| 120 |
-
"val_accuracy": 0.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
},
|
| 122 |
{
|
| 123 |
-
"avg_train_loss": 2.
|
| 124 |
-
"train_accuracy": 0.
|
| 125 |
-
"avg_val_loss": 2.
|
| 126 |
-
"val_accuracy": 0.
|
| 127 |
},
|
| 128 |
{
|
| 129 |
-
"avg_train_loss": 2.
|
| 130 |
-
"train_accuracy": 0.
|
| 131 |
-
"avg_val_loss": 2.
|
| 132 |
-
"val_accuracy": 0.
|
| 133 |
},
|
| 134 |
{
|
| 135 |
-
"avg_train_loss": 2.
|
| 136 |
-
"train_accuracy": 0.
|
| 137 |
-
"avg_val_loss": 2.
|
| 138 |
-
"val_accuracy": 0.
|
| 139 |
},
|
| 140 |
{
|
| 141 |
-
"avg_train_loss": 2.
|
| 142 |
-
"train_accuracy": 0.
|
| 143 |
-
"avg_val_loss": 2.
|
| 144 |
-
"val_accuracy": 0.
|
| 145 |
},
|
| 146 |
{
|
| 147 |
-
"avg_train_loss": 2.
|
| 148 |
-
"train_accuracy": 0.
|
| 149 |
-
"avg_val_loss": 2.
|
| 150 |
-
"val_accuracy": 0.
|
| 151 |
},
|
| 152 |
{
|
| 153 |
-
"avg_train_loss": 2.
|
| 154 |
-
"train_accuracy": 0.
|
| 155 |
-
"avg_val_loss": 2.
|
| 156 |
-
"val_accuracy": 0.
|
| 157 |
},
|
| 158 |
{
|
| 159 |
-
"avg_train_loss": 2.
|
| 160 |
-
"train_accuracy": 0.
|
| 161 |
-
"avg_val_loss": 2.
|
| 162 |
-
"val_accuracy": 0.
|
| 163 |
},
|
| 164 |
{
|
| 165 |
-
"avg_train_loss": 2.
|
| 166 |
-
"train_accuracy": 0.
|
| 167 |
-
"avg_val_loss": 2.
|
| 168 |
-
"val_accuracy": 0.
|
| 169 |
},
|
| 170 |
{
|
| 171 |
-
"avg_train_loss": 2.
|
| 172 |
-
"train_accuracy": 0.
|
| 173 |
-
"avg_val_loss": 2.
|
| 174 |
-
"val_accuracy": 0.
|
| 175 |
}
|
| 176 |
]
|
|
|
|
| 1 |
[
|
| 2 |
{
|
| 3 |
+
"avg_train_loss": 4.452685312236971,
|
| 4 |
+
"train_accuracy": 0.03584,
|
| 5 |
+
"avg_val_loss": 4.1456576963014244,
|
| 6 |
+
"val_accuracy": 0.07199999690055847
|
| 7 |
},
|
| 8 |
{
|
| 9 |
+
"avg_train_loss": 4.22897495028308,
|
| 10 |
+
"train_accuracy": 0.06454,
|
| 11 |
+
"avg_val_loss": 3.984914272646361,
|
| 12 |
+
"val_accuracy": 0.10199999809265137
|
| 13 |
},
|
| 14 |
{
|
| 15 |
+
"avg_train_loss": 4.097483895318892,
|
| 16 |
+
"train_accuracy": 0.08686,
|
| 17 |
+
"avg_val_loss": 3.8411196938043908,
|
| 18 |
+
"val_accuracy": 0.13650000095367432
|
| 19 |
},
|
| 20 |
{
|
| 21 |
+
"avg_train_loss": 3.993970947192453,
|
| 22 |
+
"train_accuracy": 0.1051,
|
| 23 |
+
"avg_val_loss": 3.7907390353045884,
|
| 24 |
+
"val_accuracy": 0.14059999585151672
|
| 25 |
},
|
| 26 |
{
|
| 27 |
+
"avg_train_loss": 3.9059129904603105,
|
| 28 |
+
"train_accuracy": 0.1232,
|
| 29 |
+
"avg_val_loss": 3.656763390649723,
|
| 30 |
+
"val_accuracy": 0.164900004863739
|
| 31 |
},
|
| 32 |
{
|
| 33 |
+
"avg_train_loss": 3.8222918010428737,
|
| 34 |
+
"train_accuracy": 0.13996,
|
| 35 |
+
"avg_val_loss": 3.5380990716475473,
|
| 36 |
+
"val_accuracy": 0.19670000672340393
|
| 37 |
},
|
| 38 |
{
|
| 39 |
+
"avg_train_loss": 3.756975746520645,
|
| 40 |
+
"train_accuracy": 0.1505,
|
| 41 |
+
"avg_val_loss": 3.4915640142899527,
|
| 42 |
+
"val_accuracy": 0.20239999890327454
|
| 43 |
},
|
| 44 |
{
|
| 45 |
+
"avg_train_loss": 3.6912049436203356,
|
| 46 |
+
"train_accuracy": 0.16672,
|
| 47 |
+
"avg_val_loss": 3.4289465795589398,
|
| 48 |
+
"val_accuracy": 0.2176000028848648
|
| 49 |
},
|
| 50 |
{
|
| 51 |
+
"avg_train_loss": 3.6327530968829493,
|
| 52 |
+
"train_accuracy": 0.18014,
|
| 53 |
+
"avg_val_loss": 3.3895603614517404,
|
| 54 |
+
"val_accuracy": 0.2402999997138977
|
| 55 |
},
|
| 56 |
{
|
| 57 |
+
"avg_train_loss": 3.578799500489784,
|
| 58 |
+
"train_accuracy": 0.19102,
|
| 59 |
+
"avg_val_loss": 3.3035911849782438,
|
| 60 |
+
"val_accuracy": 0.2533000111579895
|
| 61 |
},
|
| 62 |
{
|
| 63 |
+
"avg_train_loss": 3.5295029982276587,
|
| 64 |
+
"train_accuracy": 0.20518,
|
| 65 |
+
"avg_val_loss": 3.2572307345233384,
|
| 66 |
+
"val_accuracy": 0.2711000144481659
|
| 67 |
},
|
| 68 |
{
|
| 69 |
+
"avg_train_loss": 3.49012098257499,
|
| 70 |
+
"train_accuracy": 0.21266,
|
| 71 |
+
"avg_val_loss": 3.194729382478738,
|
| 72 |
+
"val_accuracy": 0.29030001163482666
|
| 73 |
},
|
| 74 |
{
|
| 75 |
+
"avg_train_loss": 3.444543719901453,
|
| 76 |
+
"train_accuracy": 0.22732,
|
| 77 |
+
"avg_val_loss": 3.1562333891663372,
|
| 78 |
+
"val_accuracy": 0.2962000072002411
|
| 79 |
},
|
| 80 |
{
|
| 81 |
+
"avg_train_loss": 3.409660898206179,
|
| 82 |
+
"train_accuracy": 0.23328,
|
| 83 |
+
"avg_val_loss": 3.1249258306962027,
|
| 84 |
+
"val_accuracy": 0.303600013256073
|
| 85 |
},
|
| 86 |
{
|
| 87 |
+
"avg_train_loss": 3.365924084278019,
|
| 88 |
+
"train_accuracy": 0.2484,
|
| 89 |
+
"avg_val_loss": 3.1003157217291335,
|
| 90 |
+
"val_accuracy": 0.3172000050544739
|
| 91 |
},
|
| 92 |
{
|
| 93 |
+
"avg_train_loss": 3.3438522516918914,
|
| 94 |
+
"train_accuracy": 0.25078,
|
| 95 |
+
"avg_val_loss": 3.0775443934187106,
|
| 96 |
+
"val_accuracy": 0.3222000002861023
|
| 97 |
},
|
| 98 |
{
|
| 99 |
+
"avg_train_loss": 3.310065208188713,
|
| 100 |
+
"train_accuracy": 0.26132,
|
| 101 |
+
"avg_val_loss": 3.0591898085195806,
|
| 102 |
+
"val_accuracy": 0.3260999917984009
|
| 103 |
},
|
| 104 |
{
|
| 105 |
+
"avg_train_loss": 3.2758816282462586,
|
| 106 |
+
"train_accuracy": 0.2669,
|
| 107 |
+
"avg_val_loss": 3.0423757818680777,
|
| 108 |
+
"val_accuracy": 0.3319999873638153
|
| 109 |
},
|
| 110 |
{
|
| 111 |
+
"avg_train_loss": 3.2471999869017343,
|
| 112 |
+
"train_accuracy": 0.276,
|
| 113 |
+
"avg_val_loss": 3.0076527655879155,
|
| 114 |
+
"val_accuracy": 0.3409999907016754
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"avg_train_loss": 3.2196996800429987,
|
| 118 |
+
"train_accuracy": 0.28116,
|
| 119 |
+
"avg_val_loss": 2.9659501087816458,
|
| 120 |
+
"val_accuracy": 0.3515999913215637
|
| 121 |
+
},
|
| 122 |
+
{
|
| 123 |
+
"avg_train_loss": 3.195254132875701,
|
| 124 |
+
"train_accuracy": 0.28704,
|
| 125 |
+
"avg_val_loss": 2.9751936514166335,
|
| 126 |
+
"val_accuracy": 0.35580000281333923
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"avg_train_loss": 3.171722950532918,
|
| 130 |
+
"train_accuracy": 0.29404,
|
| 131 |
+
"avg_val_loss": 2.9623639070535006,
|
| 132 |
+
"val_accuracy": 0.3531999886035919
|
| 133 |
+
},
|
| 134 |
+
{
|
| 135 |
+
"avg_train_loss": 3.1580248549771124,
|
| 136 |
+
"train_accuracy": 0.2984,
|
| 137 |
+
"avg_val_loss": 2.909536240976068,
|
| 138 |
+
"val_accuracy": 0.3702000081539154
|
| 139 |
+
},
|
| 140 |
+
{
|
| 141 |
+
"avg_train_loss": 3.125413342814921,
|
| 142 |
+
"train_accuracy": 0.30764,
|
| 143 |
+
"avg_val_loss": 2.8885996371884888,
|
| 144 |
+
"val_accuracy": 0.3765999972820282
|
| 145 |
+
},
|
| 146 |
+
{
|
| 147 |
+
"avg_train_loss": 3.11991933300672,
|
| 148 |
+
"train_accuracy": 0.30932,
|
| 149 |
+
"avg_val_loss": 2.8885996371884888,
|
| 150 |
+
"val_accuracy": 0.3788999915122986
|
| 151 |
+
},
|
| 152 |
+
{
|
| 153 |
+
"avg_train_loss": 3.0930506728799143,
|
| 154 |
+
"train_accuracy": 0.3174,
|
| 155 |
+
"avg_val_loss": 2.852832456178303,
|
| 156 |
+
"val_accuracy": 0.3903000056743622
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"avg_train_loss": 3.0655275760099405,
|
| 160 |
+
"train_accuracy": 0.32236,
|
| 161 |
+
"avg_val_loss": 2.868115388894383,
|
| 162 |
+
"val_accuracy": 0.38449999690055847
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"avg_train_loss": 3.0548000393621146,
|
| 166 |
+
"train_accuracy": 0.32792,
|
| 167 |
+
"avg_val_loss": 2.8296788855444026,
|
| 168 |
+
"val_accuracy": 0.3880999982357025
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"avg_train_loss": 3.032537115809253,
|
| 172 |
+
"train_accuracy": 0.3308,
|
| 173 |
+
"avg_val_loss": 2.7847178012509888,
|
| 174 |
+
"val_accuracy": 0.4097999930381775
|
| 175 |
+
},
|
| 176 |
+
{
|
| 177 |
+
"avg_train_loss": 3.009990167434868,
|
| 178 |
+
"train_accuracy": 0.3352,
|
| 179 |
+
"avg_val_loss": 2.7986656864987145,
|
| 180 |
+
"val_accuracy": 0.4027999937534332
|
| 181 |
+
},
|
| 182 |
+
{
|
| 183 |
+
"avg_train_loss": 2.9918381585489455,
|
| 184 |
+
"train_accuracy": 0.34312,
|
| 185 |
+
"avg_val_loss": 2.7371235135235366,
|
| 186 |
+
"val_accuracy": 0.42089998722076416
|
| 187 |
+
},
|
| 188 |
+
{
|
| 189 |
+
"avg_train_loss": 2.9780573729054094,
|
| 190 |
+
"train_accuracy": 0.34844,
|
| 191 |
+
"avg_val_loss": 2.7818868130068237,
|
| 192 |
+
"val_accuracy": 0.4077000021934509
|
| 193 |
+
},
|
| 194 |
+
{
|
| 195 |
+
"avg_train_loss": 2.965448642020945,
|
| 196 |
+
"train_accuracy": 0.34998,
|
| 197 |
+
"avg_val_loss": 2.7794544847705698,
|
| 198 |
"val_accuracy": 0.4027000069618225
|
| 199 |
},
|
| 200 |
{
|
| 201 |
+
"avg_train_loss": 2.949932183451055,
|
| 202 |
+
"train_accuracy": 0.354,
|
| 203 |
+
"avg_val_loss": 2.7443399550039556,
|
| 204 |
+
"val_accuracy": 0.4169999957084656
|
| 205 |
+
},
|
| 206 |
+
{
|
| 207 |
+
"avg_train_loss": 2.934478478968296,
|
| 208 |
+
"train_accuracy": 0.35772,
|
| 209 |
+
"avg_val_loss": 2.7268808099287973,
|
| 210 |
+
"val_accuracy": 0.4259999990463257
|
| 211 |
+
},
|
| 212 |
+
{
|
| 213 |
+
"avg_train_loss": 2.925810183710454,
|
| 214 |
+
"train_accuracy": 0.36066,
|
| 215 |
+
"avg_val_loss": 2.7194407016416138,
|
| 216 |
+
"val_accuracy": 0.4253999888896942
|
| 217 |
+
},
|
| 218 |
+
{
|
| 219 |
+
"avg_train_loss": 2.9101742749933694,
|
| 220 |
+
"train_accuracy": 0.36768,
|
| 221 |
+
"avg_val_loss": 2.682816324354727,
|
| 222 |
+
"val_accuracy": 0.4357999861240387
|
| 223 |
+
},
|
| 224 |
+
{
|
| 225 |
+
"avg_train_loss": 2.897818704395343,
|
| 226 |
+
"train_accuracy": 0.36902,
|
| 227 |
+
"avg_val_loss": 2.7000809681566458,
|
| 228 |
+
"val_accuracy": 0.4271000027656555
|
| 229 |
+
},
|
| 230 |
+
{
|
| 231 |
+
"avg_train_loss": 2.882525757450582,
|
| 232 |
+
"train_accuracy": 0.3717,
|
| 233 |
+
"avg_val_loss": 2.6881412554390822,
|
| 234 |
+
"val_accuracy": 0.4334000051021576
|
| 235 |
+
},
|
| 236 |
+
{
|
| 237 |
+
"avg_train_loss": 2.868310006683135,
|
| 238 |
+
"train_accuracy": 0.37626,
|
| 239 |
+
"avg_val_loss": 2.696658943272844,
|
| 240 |
+
"val_accuracy": 0.4318999946117401
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"avg_train_loss": 2.8593416683509223,
|
| 244 |
+
"train_accuracy": 0.37938,
|
| 245 |
+
"avg_val_loss": 2.6929442973076543,
|
| 246 |
+
"val_accuracy": 0.4361000061035156
|
| 247 |
+
},
|
| 248 |
+
{
|
| 249 |
+
"avg_train_loss": 2.8498354638019183,
|
| 250 |
+
"train_accuracy": 0.38216,
|
| 251 |
+
"avg_val_loss": 2.6896122799643987,
|
| 252 |
+
"val_accuracy": 0.43959999084472656
|
| 253 |
+
},
|
| 254 |
+
{
|
| 255 |
+
"avg_train_loss": 2.841897433066307,
|
| 256 |
+
"train_accuracy": 0.38478,
|
| 257 |
+
"avg_val_loss": 2.6792281911342957,
|
| 258 |
+
"val_accuracy": 0.4350000023841858
|
| 259 |
+
},
|
| 260 |
+
{
|
| 261 |
+
"avg_train_loss": 2.829337977387411,
|
| 262 |
+
"train_accuracy": 0.38544,
|
| 263 |
+
"avg_val_loss": 2.6498478756675237,
|
| 264 |
+
"val_accuracy": 0.44699999690055847
|
| 265 |
+
},
|
| 266 |
+
{
|
| 267 |
+
"avg_train_loss": 2.817666833967809,
|
| 268 |
+
"train_accuracy": 0.38926,
|
| 269 |
+
"avg_val_loss": 2.6291940423506723,
|
| 270 |
+
"val_accuracy": 0.454800009727478
|
| 271 |
+
},
|
| 272 |
+
{
|
| 273 |
+
"avg_train_loss": 2.798290427993326,
|
| 274 |
+
"train_accuracy": 0.39618,
|
| 275 |
+
"avg_val_loss": 2.6448449243473102,
|
| 276 |
+
"val_accuracy": 0.4528999924659729
|
| 277 |
+
},
|
| 278 |
+
{
|
| 279 |
+
"avg_train_loss": 2.7893726972057995,
|
| 280 |
+
"train_accuracy": 0.39842,
|
| 281 |
+
"avg_val_loss": 2.6335499437549448,
|
| 282 |
+
"val_accuracy": 0.447299987077713
|
| 283 |
+
},
|
| 284 |
+
{
|
| 285 |
+
"avg_train_loss": 2.779167408528535,
|
| 286 |
+
"train_accuracy": 0.401,
|
| 287 |
+
"avg_val_loss": 2.625039595591871,
|
| 288 |
+
"val_accuracy": 0.4578000009059906
|
| 289 |
+
},
|
| 290 |
+
{
|
| 291 |
+
"avg_train_loss": 2.769523390723616,
|
| 292 |
+
"train_accuracy": 0.40362,
|
| 293 |
+
"avg_val_loss": 2.606762849831883,
|
| 294 |
+
"val_accuracy": 0.4546999931335449
|
| 295 |
+
},
|
| 296 |
+
{
|
| 297 |
+
"avg_train_loss": 2.7618973175887866,
|
| 298 |
+
"train_accuracy": 0.40552,
|
| 299 |
+
"avg_val_loss": 2.6070912035205698,
|
| 300 |
+
"val_accuracy": 0.4577000141143799
|
| 301 |
+
},
|
| 302 |
+
{
|
| 303 |
+
"avg_train_loss": 2.744235341811119,
|
| 304 |
+
"train_accuracy": 0.4127,
|
| 305 |
+
"avg_val_loss": 2.577495912962322,
|
| 306 |
+
"val_accuracy": 0.4657000005245209
|
| 307 |
+
},
|
| 308 |
+
{
|
| 309 |
+
"avg_train_loss": 2.7439488306679687,
|
| 310 |
+
"train_accuracy": 0.4089,
|
| 311 |
+
"avg_val_loss": 2.610289561597607,
|
| 312 |
+
"val_accuracy": 0.45730000734329224
|
| 313 |
+
},
|
| 314 |
+
{
|
| 315 |
+
"avg_train_loss": 2.726577682263406,
|
| 316 |
+
"train_accuracy": 0.41524,
|
| 317 |
+
"avg_val_loss": 2.5453265226339994,
|
| 318 |
+
"val_accuracy": 0.4781000018119812
|
| 319 |
+
},
|
| 320 |
+
{
|
| 321 |
+
"avg_train_loss": 2.7213859829451423,
|
| 322 |
+
"train_accuracy": 0.41772,
|
| 323 |
+
"avg_val_loss": 2.5277760179736948,
|
| 324 |
+
"val_accuracy": 0.4763000011444092
|
| 325 |
+
},
|
| 326 |
+
{
|
| 327 |
+
"avg_train_loss": 2.712547524810752,
|
| 328 |
+
"train_accuracy": 0.421,
|
| 329 |
+
"avg_val_loss": 2.5841199657585046,
|
| 330 |
+
"val_accuracy": 0.4708000123500824
|
| 331 |
+
},
|
| 332 |
+
{
|
| 333 |
+
"avg_train_loss": 2.6989601530382394,
|
| 334 |
+
"train_accuracy": 0.4234,
|
| 335 |
+
"avg_val_loss": 2.5469041655335247,
|
| 336 |
+
"val_accuracy": 0.47589999437332153
|
| 337 |
+
},
|
| 338 |
+
{
|
| 339 |
+
"avg_train_loss": 2.6853119339174625,
|
| 340 |
+
"train_accuracy": 0.42606,
|
| 341 |
+
"avg_val_loss": 2.5726125210146362,
|
| 342 |
+
"val_accuracy": 0.4747999906539917
|
| 343 |
+
},
|
| 344 |
+
{
|
| 345 |
+
"avg_train_loss": 2.6801007284837612,
|
| 346 |
+
"train_accuracy": 0.42704,
|
| 347 |
+
"avg_val_loss": 2.5191534501087816,
|
| 348 |
+
"val_accuracy": 0.4855000078678131
|
| 349 |
+
},
|
| 350 |
+
{
|
| 351 |
+
"avg_train_loss": 2.673392378765604,
|
| 352 |
+
"train_accuracy": 0.429,
|
| 353 |
+
"avg_val_loss": 2.543144902096519,
|
| 354 |
+
"val_accuracy": 0.46720001101493835
|
| 355 |
+
},
|
| 356 |
+
{
|
| 357 |
+
"avg_train_loss": 2.6638625219959735,
|
| 358 |
+
"train_accuracy": 0.43398,
|
| 359 |
+
"avg_val_loss": 2.537560185299644,
|
| 360 |
+
"val_accuracy": 0.486299991607666
|
| 361 |
+
},
|
| 362 |
+
{
|
| 363 |
+
"avg_train_loss": 2.651780778184876,
|
| 364 |
+
"train_accuracy": 0.43742,
|
| 365 |
+
"avg_val_loss": 2.536473914037777,
|
| 366 |
+
"val_accuracy": 0.4837999939918518
|
| 367 |
+
},
|
| 368 |
+
{
|
| 369 |
+
"avg_train_loss": 2.6511071432581947,
|
| 370 |
+
"train_accuracy": 0.43532,
|
| 371 |
+
"avg_val_loss": 2.520467637460443,
|
| 372 |
+
"val_accuracy": 0.4832000136375427
|
| 373 |
+
},
|
| 374 |
+
{
|
| 375 |
+
"avg_train_loss": 2.6305635774227056,
|
| 376 |
+
"train_accuracy": 0.44088,
|
| 377 |
+
"avg_val_loss": 2.49658203125,
|
| 378 |
+
"val_accuracy": 0.4912000000476837
|
| 379 |
+
},
|
| 380 |
+
{
|
| 381 |
+
"avg_train_loss": 2.634286204262463,
|
| 382 |
+
"train_accuracy": 0.44186,
|
| 383 |
+
"avg_val_loss": 2.554328242434731,
|
| 384 |
+
"val_accuracy": 0.47690001130104065
|
| 385 |
+
},
|
| 386 |
+
{
|
| 387 |
+
"avg_train_loss": 2.624523153390421,
|
| 388 |
+
"train_accuracy": 0.44464,
|
| 389 |
+
"avg_val_loss": 2.492234821561017,
|
| 390 |
+
"val_accuracy": 0.4921000003814697
|
| 391 |
+
},
|
| 392 |
+
{
|
| 393 |
+
"avg_train_loss": 2.6103764879124243,
|
| 394 |
+
"train_accuracy": 0.44712,
|
| 395 |
+
"avg_val_loss": 2.4733666528629352,
|
| 396 |
+
"val_accuracy": 0.49219998717308044
|
| 397 |
+
},
|
| 398 |
+
{
|
| 399 |
+
"avg_train_loss": 2.597310994592164,
|
| 400 |
+
"train_accuracy": 0.45222,
|
| 401 |
+
"avg_val_loss": 2.4752954410601267,
|
| 402 |
+
"val_accuracy": 0.4984999895095825
|
| 403 |
+
},
|
| 404 |
+
{
|
| 405 |
+
"avg_train_loss": 2.6000741299460914,
|
| 406 |
+
"train_accuracy": 0.45098,
|
| 407 |
+
"avg_val_loss": 2.4956019920638846,
|
| 408 |
+
"val_accuracy": 0.4957999885082245
|
| 409 |
+
},
|
| 410 |
+
{
|
| 411 |
+
"avg_train_loss": 2.585392991295251,
|
| 412 |
+
"train_accuracy": 0.45412,
|
| 413 |
+
"avg_val_loss": 2.4977424718156644,
|
| 414 |
+
"val_accuracy": 0.4909999966621399
|
| 415 |
+
},
|
| 416 |
+
{
|
| 417 |
+
"avg_train_loss": 2.5859423464216538,
|
| 418 |
+
"train_accuracy": 0.45544,
|
| 419 |
+
"avg_val_loss": 2.4699751455572585,
|
| 420 |
+
"val_accuracy": 0.4993000030517578
|
| 421 |
+
},
|
| 422 |
+
{
|
| 423 |
+
"avg_train_loss": 2.5671919788545963,
|
| 424 |
+
"train_accuracy": 0.4598,
|
| 425 |
+
"avg_val_loss": 2.4434667659711233,
|
| 426 |
+
"val_accuracy": 0.5040000081062317
|
| 427 |
+
},
|
| 428 |
+
{
|
| 429 |
+
"avg_train_loss": 2.552455409103647,
|
| 430 |
+
"train_accuracy": 0.4653,
|
| 431 |
+
"avg_val_loss": 2.4661942494066458,
|
| 432 |
+
"val_accuracy": 0.4943000078201294
|
| 433 |
+
},
|
| 434 |
+
{
|
| 435 |
+
"avg_train_loss": 2.5519124113995098,
|
| 436 |
+
"train_accuracy": 0.46254,
|
| 437 |
+
"avg_val_loss": 2.43858723700801,
|
| 438 |
+
"val_accuracy": 0.5048999786376953
|
| 439 |
+
},
|
| 440 |
+
{
|
| 441 |
+
"avg_train_loss": 2.5457511021353096,
|
| 442 |
+
"train_accuracy": 0.4658,
|
| 443 |
+
"avg_val_loss": 2.48284912109375,
|
| 444 |
+
"val_accuracy": 0.5008000135421753
|
| 445 |
+
},
|
| 446 |
+
{
|
| 447 |
+
"avg_train_loss": 2.539912354915648,
|
| 448 |
+
"train_accuracy": 0.46686,
|
| 449 |
+
"avg_val_loss": 2.4563181430478638,
|
| 450 |
+
"val_accuracy": 0.5048999786376953
|
| 451 |
+
},
|
| 452 |
+
{
|
| 453 |
+
"avg_train_loss": 2.5310022376687327,
|
| 454 |
+
"train_accuracy": 0.4681,
|
| 455 |
+
"avg_val_loss": 2.4287455112119263,
|
| 456 |
+
"val_accuracy": 0.5077000260353088
|
| 457 |
+
},
|
| 458 |
+
{
|
| 459 |
+
"avg_train_loss": 2.5193640140011486,
|
| 460 |
+
"train_accuracy": 0.4733,
|
| 461 |
+
"avg_val_loss": 2.4722490914260287,
|
| 462 |
+
"val_accuracy": 0.4941999912261963
|
| 463 |
+
},
|
| 464 |
+
{
|
| 465 |
+
"avg_train_loss": 2.511948669353105,
|
| 466 |
+
"train_accuracy": 0.47668,
|
| 467 |
+
"avg_val_loss": 2.4270836552487145,
|
| 468 |
+
"val_accuracy": 0.510200023651123
|
| 469 |
+
},
|
| 470 |
+
{
|
| 471 |
+
"avg_train_loss": 2.5021252122986346,
|
| 472 |
+
"train_accuracy": 0.47688,
|
| 473 |
+
"avg_val_loss": 2.3986882076987737,
|
| 474 |
+
"val_accuracy": 0.5188000202178955
|
| 475 |
+
},
|
| 476 |
+
{
|
| 477 |
+
"avg_train_loss": 2.4900416806530767,
|
| 478 |
+
"train_accuracy": 0.48144,
|
| 479 |
+
"avg_val_loss": 2.4308195476290546,
|
| 480 |
+
"val_accuracy": 0.5117999911308289
|
| 481 |
+
},
|
| 482 |
+
{
|
| 483 |
+
"avg_train_loss": 2.493807424059914,
|
| 484 |
+
"train_accuracy": 0.48098,
|
| 485 |
+
"avg_val_loss": 2.4242190590387658,
|
| 486 |
+
"val_accuracy": 0.5123999714851379
|
| 487 |
+
},
|
| 488 |
+
{
|
| 489 |
+
"avg_train_loss": 2.4797395354951433,
|
| 490 |
+
"train_accuracy": 0.4871,
|
| 491 |
+
"avg_val_loss": 2.4191836586481408,
|
| 492 |
+
"val_accuracy": 0.5139999985694885
|
| 493 |
+
},
|
| 494 |
+
{
|
| 495 |
+
"avg_train_loss": 2.474879029461795,
|
| 496 |
+
"train_accuracy": 0.48526,
|
| 497 |
+
"avg_val_loss": 2.4089353537257714,
|
| 498 |
+
"val_accuracy": 0.5188000202178955
|
| 499 |
+
},
|
| 500 |
+
{
|
| 501 |
+
"avg_train_loss": 2.4681123287781426,
|
| 502 |
+
"train_accuracy": 0.48606,
|
| 503 |
+
"avg_val_loss": 2.3880093731457674,
|
| 504 |
+
"val_accuracy": 0.5206000208854675
|
| 505 |
+
},
|
| 506 |
+
{
|
| 507 |
+
"avg_train_loss": 2.464694506219586,
|
| 508 |
+
"train_accuracy": 0.48762,
|
| 509 |
+
"avg_val_loss": 2.4599369869956487,
|
| 510 |
+
"val_accuracy": 0.5094000101089478
|
| 511 |
+
},
|
| 512 |
+
{
|
| 513 |
+
"avg_train_loss": 2.4511944366538008,
|
| 514 |
+
"train_accuracy": 0.4928,
|
| 515 |
+
"avg_val_loss": 2.3902290440812894,
|
| 516 |
+
"val_accuracy": 0.5202000141143799
|
| 517 |
+
},
|
| 518 |
+
{
|
| 519 |
+
"avg_train_loss": 2.4458736345896024,
|
| 520 |
+
"train_accuracy": 0.4954,
|
| 521 |
+
"avg_val_loss": 2.3775387534612342,
|
| 522 |
+
"val_accuracy": 0.5228999853134155
|
| 523 |
+
},
|
| 524 |
+
{
|
| 525 |
+
"avg_train_loss": 2.4391189106285114,
|
| 526 |
+
"train_accuracy": 0.4964,
|
| 527 |
+
"avg_val_loss": 2.3863923278035997,
|
| 528 |
+
"val_accuracy": 0.5250999927520752
|
| 529 |
+
},
|
| 530 |
+
{
|
| 531 |
+
"avg_train_loss": 2.431982190102872,
|
| 532 |
+
"train_accuracy": 0.49786,
|
| 533 |
+
"avg_val_loss": 2.399633093725277,
|
| 534 |
+
"val_accuracy": 0.5228000283241272
|
| 535 |
+
},
|
| 536 |
+
{
|
| 537 |
+
"avg_train_loss": 2.4256825097991377,
|
| 538 |
+
"train_accuracy": 0.49946,
|
| 539 |
+
"avg_val_loss": 2.4045251773882517,
|
| 540 |
+
"val_accuracy": 0.515500009059906
|
| 541 |
+
},
|
| 542 |
+
{
|
| 543 |
+
"avg_train_loss": 2.4132598466275597,
|
| 544 |
+
"train_accuracy": 0.50316,
|
| 545 |
+
"avg_val_loss": 2.38846037659464,
|
| 546 |
+
"val_accuracy": 0.5232999920845032
|
| 547 |
+
},
|
| 548 |
+
{
|
| 549 |
+
"avg_train_loss": 2.4151184967411754,
|
| 550 |
+
"train_accuracy": 0.50272,
|
| 551 |
+
"avg_val_loss": 2.3768584818779668,
|
| 552 |
+
"val_accuracy": 0.5232999920845032
|
| 553 |
+
},
|
| 554 |
+
{
|
| 555 |
+
"avg_train_loss": 2.405930356906198,
|
| 556 |
+
"train_accuracy": 0.50574,
|
| 557 |
+
"avg_val_loss": 2.378061415273932,
|
| 558 |
+
"val_accuracy": 0.5267000198364258
|
| 559 |
+
},
|
| 560 |
+
{
|
| 561 |
+
"avg_train_loss": 2.4047722526828346,
|
| 562 |
+
"train_accuracy": 0.50608,
|
| 563 |
+
"avg_val_loss": 2.3851123278654076,
|
| 564 |
+
"val_accuracy": 0.5238000154495239
|
| 565 |
+
},
|
| 566 |
+
{
|
| 567 |
+
"avg_train_loss": 2.3914314154773724,
|
| 568 |
+
"train_accuracy": 0.50882,
|
| 569 |
+
"avg_val_loss": 2.3767078254796283,
|
| 570 |
+
"val_accuracy": 0.5264999866485596
|
| 571 |
+
},
|
| 572 |
+
{
|
| 573 |
+
"avg_train_loss": 2.3860856683357903,
|
| 574 |
+
"train_accuracy": 0.51156,
|
| 575 |
+
"avg_val_loss": 2.376488021657437,
|
| 576 |
+
"val_accuracy": 0.5235000252723694
|
| 577 |
+
},
|
| 578 |
+
{
|
| 579 |
+
"avg_train_loss": 2.386218143546063,
|
| 580 |
+
"train_accuracy": 0.51266,
|
| 581 |
+
"avg_val_loss": 2.3630600941332083,
|
| 582 |
+
"val_accuracy": 0.5235000252723694
|
| 583 |
+
},
|
| 584 |
+
{
|
| 585 |
+
"avg_train_loss": 2.3744330151611583,
|
| 586 |
+
"train_accuracy": 0.51428,
|
| 587 |
+
"avg_val_loss": 2.3920840492731408,
|
| 588 |
+
"val_accuracy": 0.5213000178337097
|
| 589 |
+
},
|
| 590 |
+
{
|
| 591 |
+
"avg_train_loss": 2.3711826228119834,
|
| 592 |
+
"train_accuracy": 0.51532,
|
| 593 |
+
"avg_val_loss": 2.3973187796677213,
|
| 594 |
+
"val_accuracy": 0.5235000252723694
|
| 595 |
+
},
|
| 596 |
+
{
|
| 597 |
+
"avg_train_loss": 2.3610368445706182,
|
| 598 |
+
"train_accuracy": 0.51854,
|
| 599 |
+
"avg_val_loss": 2.3672910279865507,
|
| 600 |
+
"val_accuracy": 0.5357000231742859
|
| 601 |
+
},
|
| 602 |
+
{
|
| 603 |
+
"avg_train_loss": 2.359229836165143,
|
| 604 |
+
"train_accuracy": 0.52028,
|
| 605 |
+
"avg_val_loss": 2.3615512123590783,
|
| 606 |
+
"val_accuracy": 0.5253999829292297
|
| 607 |
+
},
|
| 608 |
+
{
|
| 609 |
+
"avg_train_loss": 2.3491924016372017,
|
| 610 |
+
"train_accuracy": 0.52172,
|
| 611 |
+
"avg_val_loss": 2.3583934156200552,
|
| 612 |
+
"val_accuracy": 0.5288000106811523
|
| 613 |
+
},
|
| 614 |
+
{
|
| 615 |
+
"avg_train_loss": 2.336290584348352,
|
| 616 |
+
"train_accuracy": 0.52592,
|
| 617 |
+
"avg_val_loss": 2.3765073365803007,
|
| 618 |
+
"val_accuracy": 0.5221999883651733
|
| 619 |
+
},
|
| 620 |
+
{
|
| 621 |
+
"avg_train_loss": 2.3395893063081803,
|
| 622 |
+
"train_accuracy": 0.52642,
|
| 623 |
+
"avg_val_loss": 2.377288818359375,
|
| 624 |
+
"val_accuracy": 0.5306000113487244
|
| 625 |
+
},
|
| 626 |
+
{
|
| 627 |
+
"avg_train_loss": 2.3326609063026544,
|
| 628 |
+
"train_accuracy": 0.52884,
|
| 629 |
+
"avg_val_loss": 2.3460967631279668,
|
| 630 |
+
"val_accuracy": 0.5358999967575073
|
| 631 |
+
},
|
| 632 |
+
{
|
| 633 |
+
"avg_train_loss": 2.3343486081608726,
|
| 634 |
+
"train_accuracy": 0.5261,
|
| 635 |
+
"avg_val_loss": 2.3548485478268395,
|
| 636 |
+
"val_accuracy": 0.5310999751091003
|
| 637 |
+
},
|
| 638 |
+
{
|
| 639 |
+
"avg_train_loss": 2.319910573532514,
|
| 640 |
+
"train_accuracy": 0.53014,
|
| 641 |
+
"avg_val_loss": 2.3654568829113924,
|
| 642 |
+
"val_accuracy": 0.5289999842643738
|
| 643 |
+
},
|
| 644 |
+
{
|
| 645 |
+
"avg_train_loss": 2.3231105095590165,
|
| 646 |
+
"train_accuracy": 0.53024,
|
| 647 |
+
"avg_val_loss": 2.3426397782337816,
|
| 648 |
+
"val_accuracy": 0.5408999919891357
|
| 649 |
+
},
|
| 650 |
+
{
|
| 651 |
+
"avg_train_loss": 2.3172146889864638,
|
| 652 |
+
"train_accuracy": 0.52916,
|
| 653 |
+
"avg_val_loss": 2.3707584429390822,
|
| 654 |
+
"val_accuracy": 0.5284000039100647
|
| 655 |
},
|
| 656 |
{
|
| 657 |
+
"avg_train_loss": 2.3137913414889284,
|
| 658 |
+
"train_accuracy": 0.5338,
|
| 659 |
+
"avg_val_loss": 2.378921701938291,
|
| 660 |
+
"val_accuracy": 0.5299000144004822
|
| 661 |
},
|
| 662 |
{
|
| 663 |
+
"avg_train_loss": 2.292580285188182,
|
| 664 |
+
"train_accuracy": 0.5417,
|
| 665 |
+
"avg_val_loss": 2.3381581366816655,
|
| 666 |
+
"val_accuracy": 0.5406000018119812
|
| 667 |
},
|
| 668 |
{
|
| 669 |
+
"avg_train_loss": 2.2915546366625734,
|
| 670 |
+
"train_accuracy": 0.53912,
|
| 671 |
+
"avg_val_loss": 2.36516909056072,
|
| 672 |
+
"val_accuracy": 0.5367000102996826
|
| 673 |
},
|
| 674 |
{
|
| 675 |
+
"avg_train_loss": 2.291307615197223,
|
| 676 |
+
"train_accuracy": 0.53922,
|
| 677 |
+
"avg_val_loss": 2.3603519487984572,
|
| 678 |
+
"val_accuracy": 0.535099983215332
|
| 679 |
},
|
| 680 |
{
|
| 681 |
+
"avg_train_loss": 2.2937697252958937,
|
| 682 |
+
"train_accuracy": 0.5405,
|
| 683 |
+
"avg_val_loss": 2.3516428500791138,
|
| 684 |
+
"val_accuracy": 0.5401999950408936
|
| 685 |
},
|
| 686 |
{
|
| 687 |
+
"avg_train_loss": 2.2855114741703435,
|
| 688 |
+
"train_accuracy": 0.53978,
|
| 689 |
+
"avg_val_loss": 2.3426988818977454,
|
| 690 |
+
"val_accuracy": 0.536300003528595
|
| 691 |
},
|
| 692 |
{
|
| 693 |
+
"avg_train_loss": 2.2704710070129552,
|
| 694 |
+
"train_accuracy": 0.54678,
|
| 695 |
+
"avg_val_loss": 2.3620895192592957,
|
| 696 |
+
"val_accuracy": 0.5358999967575073
|
| 697 |
},
|
| 698 |
{
|
| 699 |
+
"avg_train_loss": 2.2660938359587393,
|
| 700 |
+
"train_accuracy": 0.5457,
|
| 701 |
+
"avg_val_loss": 2.3297165496439876,
|
| 702 |
+
"val_accuracy": 0.5368000268936157
|
| 703 |
},
|
| 704 |
{
|
| 705 |
+
"avg_train_loss": 2.2636455373690865,
|
| 706 |
+
"train_accuracy": 0.54736,
|
| 707 |
+
"avg_val_loss": 2.370776212668117,
|
| 708 |
+
"val_accuracy": 0.5335999727249146
|
| 709 |
}
|
| 710 |
]
|
performance_plot.png
CHANGED
|
|
plots.py
CHANGED
|
@@ -5,7 +5,7 @@ with open("performance.json", "r") as f:
|
|
| 5 |
performance = json.load(f)
|
| 6 |
|
| 7 |
# Extract values from the performance list
|
| 8 |
-
epochs = range(1, len(performance) + 1)
|
| 9 |
train_losses = [epoch["avg_train_loss"] for epoch in performance]
|
| 10 |
val_losses = [epoch["avg_val_loss"] for epoch in performance]
|
| 11 |
train_accuracies = [epoch["train_accuracy"] for epoch in performance]
|
|
@@ -22,7 +22,7 @@ plt.xlabel("Epochs")
|
|
| 22 |
plt.ylabel("Loss")
|
| 23 |
plt.title("Training and Validation Loss")
|
| 24 |
plt.legend()
|
| 25 |
-
plt.xticks(epochs)
|
| 26 |
|
| 27 |
# Subplot for Accuracy
|
| 28 |
plt.subplot(1, 2, 2)
|
|
@@ -32,7 +32,7 @@ plt.xlabel("Epochs")
|
|
| 32 |
plt.ylabel("Accuracy")
|
| 33 |
plt.title("Training and Validation Accuracy")
|
| 34 |
plt.legend()
|
| 35 |
-
plt.xticks(epochs)
|
| 36 |
|
| 37 |
plt.tight_layout()
|
| 38 |
|
|
|
|
| 5 |
performance = json.load(f)
|
| 6 |
|
| 7 |
# Extract values from the performance list
|
| 8 |
+
epochs = list(range(1, len(performance) + 1))
|
| 9 |
train_losses = [epoch["avg_train_loss"] for epoch in performance]
|
| 10 |
val_losses = [epoch["avg_val_loss"] for epoch in performance]
|
| 11 |
train_accuracies = [epoch["train_accuracy"] for epoch in performance]
|
|
|
|
| 22 |
plt.ylabel("Loss")
|
| 23 |
plt.title("Training and Validation Loss")
|
| 24 |
plt.legend()
|
| 25 |
+
plt.xticks([1] + epochs[9::10] + [epochs[-1]])
|
| 26 |
|
| 27 |
# Subplot for Accuracy
|
| 28 |
plt.subplot(1, 2, 2)
|
|
|
|
| 32 |
plt.ylabel("Accuracy")
|
| 33 |
plt.title("Training and Validation Accuracy")
|
| 34 |
plt.legend()
|
| 35 |
+
plt.xticks([1] + epochs[9::10] + [epochs[-1]])
|
| 36 |
|
| 37 |
plt.tight_layout()
|
| 38 |
|
predictions.csv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
train.py
DELETED
|
@@ -1,394 +0,0 @@
|
|
| 1 |
-
#!/usr/bin/env python3
|
| 2 |
-
import os
|
| 3 |
-
import csv
|
| 4 |
-
import json
|
| 5 |
-
from tqdm import tqdm
|
| 6 |
-
import torch
|
| 7 |
-
import argparse
|
| 8 |
-
from PIL import Image
|
| 9 |
-
from torchvision import transforms
|
| 10 |
-
from torch.utils.data import DataLoader, Dataset
|
| 11 |
-
from model import MyModel
|
| 12 |
-
import numpy as np
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
class MiniPlaces(Dataset):
|
| 16 |
-
def __init__(self, root_dir, split, transform=None, label_dict=None):
|
| 17 |
-
"""
|
| 18 |
-
Initialize the MiniPlaces dataset with the root directory for the images,
|
| 19 |
-
the split (train/val/test), an optional data transformation,
|
| 20 |
-
and an optional label dictionary.
|
| 21 |
-
|
| 22 |
-
Args:
|
| 23 |
-
root_dir (str): Root directory for the MiniPlaces images.
|
| 24 |
-
split (str): Split to use ('train', 'val', or 'test').
|
| 25 |
-
transform (callable, optional): Optional data transformation to apply to the images.
|
| 26 |
-
label_dict (dict, optional): Optional dictionary mapping integer labels to class names.
|
| 27 |
-
"""
|
| 28 |
-
assert split in ['train', 'val', 'test']
|
| 29 |
-
self.root_dir = root_dir
|
| 30 |
-
self.split = split
|
| 31 |
-
self.transform = transform
|
| 32 |
-
self.filenames = []
|
| 33 |
-
self.labels = []
|
| 34 |
-
|
| 35 |
-
self.label_dict = label_dict if label_dict is not None else {}
|
| 36 |
-
|
| 37 |
-
with open(os.path.join(self.root_dir, self.split + '.txt')) as r:
|
| 38 |
-
lines = r.readlines()
|
| 39 |
-
for line in lines:
|
| 40 |
-
line = line.split()
|
| 41 |
-
self.filenames.append(line[0])
|
| 42 |
-
if split == 'test':
|
| 43 |
-
label = line[0]
|
| 44 |
-
else:
|
| 45 |
-
label = int(line[1])
|
| 46 |
-
self.labels.append(label)
|
| 47 |
-
if split == 'train':
|
| 48 |
-
text_label = line[0].split('/')[2]
|
| 49 |
-
self.label_dict[label] = text_label
|
| 50 |
-
|
| 51 |
-
def __len__(self):
|
| 52 |
-
"""
|
| 53 |
-
Return the number of images in the dataset.
|
| 54 |
-
|
| 55 |
-
Returns:
|
| 56 |
-
int: Number of images in the dataset.
|
| 57 |
-
"""
|
| 58 |
-
return len(self.labels)
|
| 59 |
-
|
| 60 |
-
def __getitem__(self, idx):
|
| 61 |
-
"""
|
| 62 |
-
Return a single image and its corresponding label when given an index.
|
| 63 |
-
|
| 64 |
-
Args:
|
| 65 |
-
idx (int): Index of the image to retrieve.
|
| 66 |
-
|
| 67 |
-
Returns:
|
| 68 |
-
tuple: Tuple containing the image and its label.
|
| 69 |
-
"""
|
| 70 |
-
if self.transform is not None:
|
| 71 |
-
image = self.transform(
|
| 72 |
-
Image.open(os.path.join(self.root_dir, "images", self.filenames[idx])))
|
| 73 |
-
else:
|
| 74 |
-
image = Image.open(os.path.join(self.root_dir, "images", self.filenames[idx]))
|
| 75 |
-
label = self.labels[idx]
|
| 76 |
-
return image, label
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
def create_train_transform():
|
| 80 |
-
"""
|
| 81 |
-
Create training data transformation with augmentation
|
| 82 |
-
"""
|
| 83 |
-
image_net_mean = torch.Tensor([0.485, 0.456, 0.406])
|
| 84 |
-
image_net_std = torch.Tensor([0.229, 0.224, 0.225])
|
| 85 |
-
|
| 86 |
-
return transforms.Compose([
|
| 87 |
-
transforms.RandomResizedCrop(128, scale=(0.8, 1.0)),
|
| 88 |
-
transforms.RandomHorizontalFlip(p=0.5),
|
| 89 |
-
transforms.ColorJitter(
|
| 90 |
-
brightness=0.4,
|
| 91 |
-
contrast=0.4,
|
| 92 |
-
saturation=0.4,
|
| 93 |
-
hue=0.1
|
| 94 |
-
),
|
| 95 |
-
transforms.RandomAffine(
|
| 96 |
-
degrees=15, # rotation
|
| 97 |
-
translate=(0.1, 0.1), # horizontal/vertical translation
|
| 98 |
-
scale=(0.9, 1.1), # scale
|
| 99 |
-
),
|
| 100 |
-
transforms.ToTensor(),
|
| 101 |
-
transforms.Resize((128, 128)),
|
| 102 |
-
transforms.Normalize(image_net_mean, image_net_std)
|
| 103 |
-
])
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
def create_val_transform():
|
| 107 |
-
"""
|
| 108 |
-
Create validation/test data transformation without augmentation
|
| 109 |
-
"""
|
| 110 |
-
image_net_mean = torch.Tensor([0.485, 0.456, 0.406])
|
| 111 |
-
image_net_std = torch.Tensor([0.229, 0.224, 0.225])
|
| 112 |
-
|
| 113 |
-
return transforms.Compose([
|
| 114 |
-
transforms.ToTensor(),
|
| 115 |
-
transforms.Resize((128, 128)),
|
| 116 |
-
transforms.Normalize(image_net_mean, image_net_std)
|
| 117 |
-
])
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
def evaluate(model, test_loader, criterion, device):
|
| 121 |
-
"""
|
| 122 |
-
Evaluate the CNN classifier on the validation set.
|
| 123 |
-
|
| 124 |
-
Args:
|
| 125 |
-
model (CNN): CNN classifier to evaluate.
|
| 126 |
-
test_loader (torch.utils.data.DataLoader): Data loader for the test set.
|
| 127 |
-
criterion (callable): Loss function to use for evaluation.
|
| 128 |
-
device (torch.device): Device to use for evaluation.
|
| 129 |
-
|
| 130 |
-
Returns:
|
| 131 |
-
float: Average loss on the test set.
|
| 132 |
-
float: Accuracy on the test set.
|
| 133 |
-
"""
|
| 134 |
-
model.eval() # Set model to evaluation mode
|
| 135 |
-
|
| 136 |
-
with torch.no_grad():
|
| 137 |
-
total_loss = 0.0
|
| 138 |
-
num_correct = 0
|
| 139 |
-
num_samples = 0
|
| 140 |
-
|
| 141 |
-
for inputs, labels in test_loader:
|
| 142 |
-
# Move inputs and labels to device
|
| 143 |
-
inputs = inputs.to(device)
|
| 144 |
-
labels = labels.to(device)
|
| 145 |
-
|
| 146 |
-
# Compute the logits and loss
|
| 147 |
-
logits = model(inputs)
|
| 148 |
-
loss = criterion(logits, labels)
|
| 149 |
-
total_loss += loss.item()
|
| 150 |
-
|
| 151 |
-
# Compute the accuracy
|
| 152 |
-
_, predictions = torch.max(logits, dim=1)
|
| 153 |
-
num_correct += (predictions == labels).sum().item()
|
| 154 |
-
num_samples += len(inputs)
|
| 155 |
-
|
| 156 |
-
# Evaluate the model on the validation set
|
| 157 |
-
avg_loss = total_loss / len(test_loader)
|
| 158 |
-
accuracy = num_correct / num_samples
|
| 159 |
-
|
| 160 |
-
return avg_loss, accuracy
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
def train(model, train_loader, val_loader, optimizer, criterion, device,
|
| 164 |
-
num_epochs):
|
| 165 |
-
"""
|
| 166 |
-
Train the CNN classifer on the training set and evaluate it on the validation set every epoch.
|
| 167 |
-
|
| 168 |
-
Args:
|
| 169 |
-
model (CNN): CNN classifier to train.
|
| 170 |
-
train_loader (torch.utils.data.DataLoader): Data loader for the training set.
|
| 171 |
-
val_loader (torch.utils.data.DataLoader): Data loader for the validation set.
|
| 172 |
-
optimizer (torch.optim.Optimizer): Optimizer to use for training.
|
| 173 |
-
criterion (callable): Loss function to use for training.
|
| 174 |
-
device (torch.device): Device to use for training.
|
| 175 |
-
num_epochs (int): Number of epochs to train the model.
|
| 176 |
-
"""
|
| 177 |
-
|
| 178 |
-
# Place the model on device
|
| 179 |
-
model = model.to(device)
|
| 180 |
-
|
| 181 |
-
# Define early stopping parameters
|
| 182 |
-
patience = 5 # Number of epochs to wait for improvement
|
| 183 |
-
best_val_accuracy = 0.0 # Best validation accuracy so far
|
| 184 |
-
epochs_without_improvement = 0 # Counter for epochs without improvement
|
| 185 |
-
best_model_state = None # To store the state of the best model
|
| 186 |
-
|
| 187 |
-
# Performance tracking
|
| 188 |
-
performance = []
|
| 189 |
-
|
| 190 |
-
for epoch in range(num_epochs):
|
| 191 |
-
model.train() # Set model to training mode
|
| 192 |
-
|
| 193 |
-
running_loss = 0.0 # Track cumulative loss for averaging
|
| 194 |
-
correct_predictions = 0
|
| 195 |
-
total_samples = 0
|
| 196 |
-
|
| 197 |
-
with tqdm(total=len(train_loader),
|
| 198 |
-
desc=f'Epoch {epoch + 1}/{num_epochs}',
|
| 199 |
-
position=0,
|
| 200 |
-
leave=True) as pbar:
|
| 201 |
-
for inputs, labels in train_loader:
|
| 202 |
-
# Move inputs and labels to device
|
| 203 |
-
inputs = inputs.to(device)
|
| 204 |
-
labels = labels.to(device)
|
| 205 |
-
|
| 206 |
-
# Zero the gradients
|
| 207 |
-
optimizer.zero_grad()
|
| 208 |
-
|
| 209 |
-
# Compute the logits and loss
|
| 210 |
-
logits = model(inputs)
|
| 211 |
-
loss = criterion(logits, labels)
|
| 212 |
-
|
| 213 |
-
# Backward pass: Compute gradients
|
| 214 |
-
loss.backward()
|
| 215 |
-
|
| 216 |
-
# Optimize model parameters
|
| 217 |
-
optimizer.step()
|
| 218 |
-
|
| 219 |
-
# Track running loss
|
| 220 |
-
running_loss += loss.item()
|
| 221 |
-
|
| 222 |
-
# Track accuracy
|
| 223 |
-
_, predicted = logits.max(1)
|
| 224 |
-
correct_predictions += (predicted == labels).sum().item()
|
| 225 |
-
total_samples += labels.size(0)
|
| 226 |
-
|
| 227 |
-
# Update the progress bar
|
| 228 |
-
pbar.update(1)
|
| 229 |
-
pbar.set_postfix(loss=loss.item())
|
| 230 |
-
|
| 231 |
-
# Calculate average loss and accuracy
|
| 232 |
-
avg_train_loss = running_loss / len(train_loader)
|
| 233 |
-
train_accuracy = correct_predictions / total_samples
|
| 234 |
-
avg_val_loss, val_accuracy = evaluate(model, val_loader, criterion, device)
|
| 235 |
-
|
| 236 |
-
performance.append({
|
| 237 |
-
"avg_train_loss": avg_train_loss,
|
| 238 |
-
"train_accuracy": train_accuracy,
|
| 239 |
-
"avg_val_loss": avg_val_loss,
|
| 240 |
-
"val_accuracy": val_accuracy
|
| 241 |
-
})
|
| 242 |
-
print(
|
| 243 |
-
f"Train Loss: {avg_train_loss:.4f}, Accuracy: {train_accuracy:.4f} "
|
| 244 |
-
f"Validation Loss: {avg_val_loss:.4f}, Validation Accuracy: {val_accuracy:.4f}"
|
| 245 |
-
)
|
| 246 |
-
|
| 247 |
-
# Check for early stopping
|
| 248 |
-
if val_accuracy > best_val_accuracy:
|
| 249 |
-
best_val_accuracy = val_accuracy
|
| 250 |
-
epochs_without_improvement = 0 # Reset counter if there's an improvement
|
| 251 |
-
|
| 252 |
-
# Save the model checkpoint for the best model
|
| 253 |
-
best_model_state = {
|
| 254 |
-
'model_state_dict': model.module.state_dict(),
|
| 255 |
-
'optimizer_state_dict': optimizer.state_dict(),
|
| 256 |
-
'epoch': epoch,
|
| 257 |
-
}
|
| 258 |
-
else:
|
| 259 |
-
epochs_without_improvement += 1
|
| 260 |
-
|
| 261 |
-
# Early stopping condition
|
| 262 |
-
if epochs_without_improvement >= patience:
|
| 263 |
-
print(f"Early stopping at epoch {epoch + 1}.")
|
| 264 |
-
break # Stop training if no improvement for 'patience' epochs
|
| 265 |
-
|
| 266 |
-
# Save the performance list to a JSON file
|
| 267 |
-
with open("performance.json", "w") as f:
|
| 268 |
-
json.dump(performance, f, indent=4)
|
| 269 |
-
torch.save(best_model_state, 'model.ckpt')
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
def test(model, test_loader, device):
|
| 273 |
-
"""
|
| 274 |
-
Get predictions for the test set.
|
| 275 |
-
|
| 276 |
-
Args:
|
| 277 |
-
model (CNN): classifier to evaluate.
|
| 278 |
-
test_loader (torch.utils.data.DataLoader): Data loader for the test set.
|
| 279 |
-
device (torch.device): Device to use for evaluation.
|
| 280 |
-
|
| 281 |
-
Returns:
|
| 282 |
-
float: Average loss on the test set.
|
| 283 |
-
float: Accuracy on the test set.
|
| 284 |
-
"""
|
| 285 |
-
model = model.to(device)
|
| 286 |
-
model.eval() # Set model to evaluation mode
|
| 287 |
-
|
| 288 |
-
with torch.no_grad():
|
| 289 |
-
all_preds = []
|
| 290 |
-
|
| 291 |
-
for inputs, labels in test_loader:
|
| 292 |
-
# Move inputs and labels to device
|
| 293 |
-
inputs = inputs.to(device)
|
| 294 |
-
|
| 295 |
-
logits = model(inputs)
|
| 296 |
-
|
| 297 |
-
_, predictions = torch.max(logits, dim=1)
|
| 298 |
-
preds = list(zip(labels, predictions.tolist()))
|
| 299 |
-
all_preds.extend(preds)
|
| 300 |
-
|
| 301 |
-
return all_preds
|
| 302 |
-
|
| 303 |
-
|
| 304 |
-
def write_predictions(preds, filename):
|
| 305 |
-
with open(filename, 'w') as f:
|
| 306 |
-
writer = csv.writer(f, delimiter=',')
|
| 307 |
-
for im, pred in preds:
|
| 308 |
-
writer.writerow((im, pred))
|
| 309 |
-
|
| 310 |
-
|
| 311 |
-
def main(args):
|
| 312 |
-
image_net_mean = torch.Tensor([0.485, 0.456, 0.406])
|
| 313 |
-
image_net_std = torch.Tensor([0.229, 0.224, 0.225])
|
| 314 |
-
|
| 315 |
-
# Define data transformation
|
| 316 |
-
data_transform = transforms.Compose([
|
| 317 |
-
transforms.ToTensor(),
|
| 318 |
-
transforms.Resize((128, 128)),
|
| 319 |
-
transforms.Normalize(image_net_mean, image_net_std),
|
| 320 |
-
])
|
| 321 |
-
|
| 322 |
-
# Separate transforms for training and validation
|
| 323 |
-
train_transform = create_train_transform()
|
| 324 |
-
val_transform = create_val_transform()
|
| 325 |
-
|
| 326 |
-
# Create datasets
|
| 327 |
-
data_root = 'data'
|
| 328 |
-
miniplaces_train = MiniPlaces(data_root,
|
| 329 |
-
split='train',
|
| 330 |
-
transform=data_transform)
|
| 331 |
-
miniplaces_val = MiniPlaces(data_root,
|
| 332 |
-
split='val',
|
| 333 |
-
transform=data_transform,
|
| 334 |
-
label_dict=miniplaces_train.label_dict)
|
| 335 |
-
|
| 336 |
-
# Create the dataloaders
|
| 337 |
-
|
| 338 |
-
# Define the batch size and number of workers
|
| 339 |
-
batch_size = int(args.batch_size)
|
| 340 |
-
num_workers = 2
|
| 341 |
-
|
| 342 |
-
# Create DataLoader for training and validation sets
|
| 343 |
-
train_loader = DataLoader(miniplaces_train,
|
| 344 |
-
batch_size=batch_size,
|
| 345 |
-
num_workers=num_workers,
|
| 346 |
-
shuffle=True)
|
| 347 |
-
val_loader = DataLoader(miniplaces_val,
|
| 348 |
-
batch_size=batch_size,
|
| 349 |
-
num_workers=num_workers,
|
| 350 |
-
shuffle=False)
|
| 351 |
-
|
| 352 |
-
device = torch.device(f"cuda:{args.gpu}" if torch.cuda.is_available() else 'cpu') # TODO: check cuda
|
| 353 |
-
|
| 354 |
-
model = MyModel(num_classes=len(miniplaces_train.label_dict))
|
| 355 |
-
|
| 356 |
-
# optimizer = torch.optim.Adam(model.parameters(), lr=0.001, betas=(0.9, 0.999), eps=1e-08, weight_decay=1e-4, amsgrad=False)
|
| 357 |
-
optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9, dampening=0, weight_decay=1e-4, nesterov=True)
|
| 358 |
-
|
| 359 |
-
print("PARAMS NUM:", sum(p.numel() for p in model.parameters() if p.requires_grad))
|
| 360 |
-
|
| 361 |
-
if args.checkpoint:
|
| 362 |
-
checkpoint = torch.load(args.checkpoint)
|
| 363 |
-
model.load_state_dict(checkpoint['model_state_dict'])
|
| 364 |
-
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
|
| 365 |
-
|
| 366 |
-
criterion = torch.nn.CrossEntropyLoss(reduction='mean', label_smoothing=0.1)
|
| 367 |
-
|
| 368 |
-
if not args.test:
|
| 369 |
-
train(model, train_loader, val_loader, optimizer, criterion,
|
| 370 |
-
device, num_epochs=int(args.epochs))
|
| 371 |
-
|
| 372 |
-
else:
|
| 373 |
-
miniplaces_test = MiniPlaces(data_root,
|
| 374 |
-
split='test',
|
| 375 |
-
transform=data_transform)
|
| 376 |
-
test_loader = DataLoader(miniplaces_test,
|
| 377 |
-
batch_size=batch_size,
|
| 378 |
-
num_workers=num_workers,
|
| 379 |
-
shuffle=False)
|
| 380 |
-
checkpoint = torch.load(args.checkpoint, weights_only=True)
|
| 381 |
-
model.load_state_dict(checkpoint['model_state_dict'])
|
| 382 |
-
preds = test(model, test_loader, device)
|
| 383 |
-
write_predictions(preds, 'predictions.csv')
|
| 384 |
-
|
| 385 |
-
|
| 386 |
-
if __name__ == "__main__":
|
| 387 |
-
parser = argparse.ArgumentParser()
|
| 388 |
-
parser.add_argument('--test', action='store_true')
|
| 389 |
-
parser.add_argument('--checkpoint')
|
| 390 |
-
parser.add_argument('--gpu', default=0)
|
| 391 |
-
parser.add_argument('--epochs', default=100)
|
| 392 |
-
parser.add_argument('--batch_size', default=32)
|
| 393 |
-
args = parser.parse_args()
|
| 394 |
-
main(args)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
train_dist.py
CHANGED
|
@@ -2,11 +2,13 @@
|
|
| 2 |
import os
|
| 3 |
import csv
|
| 4 |
import json
|
|
|
|
| 5 |
from tqdm import tqdm
|
| 6 |
import torch
|
| 7 |
import torch.distributed as dist
|
| 8 |
import torch.multiprocessing as mp
|
| 9 |
from torch.nn.parallel import DistributedDataParallel as DDP
|
|
|
|
| 10 |
from torch.utils.data.distributed import DistributedSampler
|
| 11 |
import argparse
|
| 12 |
from PIL import Image
|
|
@@ -36,6 +38,7 @@ def cleanup():
|
|
| 36 |
if dist.is_initialized():
|
| 37 |
dist.barrier() # Synchronize all processes before destroying process group
|
| 38 |
dist.destroy_process_group()
|
|
|
|
| 39 |
|
| 40 |
|
| 41 |
class MiniPlaces(Dataset):
|
|
@@ -161,6 +164,7 @@ def evaluate(model, test_loader, criterion, device):
|
|
| 161 |
with torch.no_grad():
|
| 162 |
total_loss = 0.0
|
| 163 |
num_correct = 0
|
|
|
|
| 164 |
num_samples = 0
|
| 165 |
|
| 166 |
for inputs, labels in test_loader:
|
|
@@ -173,22 +177,29 @@ def evaluate(model, test_loader, criterion, device):
|
|
| 173 |
|
| 174 |
_, predictions = torch.max(logits, dim=1)
|
| 175 |
num_correct += (predictions == labels).sum().item()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 176 |
num_samples += len(inputs)
|
| 177 |
|
| 178 |
# Gather metrics from all processes
|
| 179 |
world_size = dist.get_world_size()
|
| 180 |
total_loss = torch.tensor(total_loss).to(device)
|
| 181 |
num_correct = torch.tensor(num_correct).to(device)
|
|
|
|
| 182 |
num_samples = torch.tensor(num_samples).to(device)
|
| 183 |
|
| 184 |
dist.all_reduce(total_loss, op=dist.ReduceOp.SUM)
|
| 185 |
dist.all_reduce(num_correct, op=dist.ReduceOp.SUM)
|
|
|
|
| 186 |
dist.all_reduce(num_samples, op=dist.ReduceOp.SUM)
|
| 187 |
|
| 188 |
avg_loss = (total_loss / world_size).item() / len(test_loader)
|
| 189 |
accuracy = (num_correct / num_samples).item()
|
|
|
|
| 190 |
|
| 191 |
-
return avg_loss, accuracy
|
| 192 |
|
| 193 |
|
| 194 |
def train_worker(rank, world_size, args):
|
|
@@ -201,15 +212,18 @@ def train_worker(rank, world_size, args):
|
|
| 201 |
args (argparse.Namespace): Command-line arguments.
|
| 202 |
"""
|
| 203 |
try:
|
|
|
|
| 204 |
setup(rank, world_size, args.port)
|
| 205 |
device = torch.device(f'cuda:{rank}')
|
| 206 |
|
| 207 |
# Define early stopping parameters
|
| 208 |
-
patience =
|
| 209 |
best_val_accuracy = 0.0 # Best validation accuracy so far
|
| 210 |
epochs_without_improvement = 0 # Counter for epochs without improvement
|
| 211 |
best_model_state = None # To store the state of the best model
|
| 212 |
|
|
|
|
|
|
|
| 213 |
# Separate transforms for training and validation
|
| 214 |
train_transform = create_train_transform()
|
| 215 |
val_transform = create_val_transform()
|
|
@@ -233,7 +247,7 @@ def train_worker(rank, world_size, args):
|
|
| 233 |
pin_memory=True)
|
| 234 |
|
| 235 |
# Create model and move to GPU
|
| 236 |
-
model = MyModel(num_classes=len(miniplaces_train.label_dict))
|
| 237 |
model = model.to(device)
|
| 238 |
model = DDP(model, device_ids=[rank])
|
| 239 |
|
|
@@ -247,6 +261,9 @@ def train_worker(rank, world_size, args):
|
|
| 247 |
model.module.load_state_dict(checkpoint['model_state_dict'])
|
| 248 |
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
|
| 249 |
|
|
|
|
|
|
|
|
|
|
| 250 |
if not args.test:
|
| 251 |
# Training loop
|
| 252 |
performance = []
|
|
@@ -288,7 +305,14 @@ def train_worker(rank, world_size, args):
|
|
| 288 |
# Evaluate and log metrics
|
| 289 |
avg_train_loss = running_loss / len(train_loader)
|
| 290 |
train_accuracy = correct_predictions / total_samples
|
| 291 |
-
avg_val_loss, val_accuracy = evaluate(model, val_loader, criterion, device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 292 |
|
| 293 |
if rank == 0: # Only save metrics on rank 0
|
| 294 |
performance.append({
|
|
@@ -327,16 +351,25 @@ def train_worker(rank, world_size, args):
|
|
| 327 |
torch.save(best_model_state, 'model.ckpt')
|
| 328 |
|
| 329 |
else: # Testing mode
|
| 330 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 331 |
test_loader = DataLoader(miniplaces_test, batch_size=args.batch_size, num_workers=2, shuffle=False)
|
| 332 |
checkpoint = torch.load(args.checkpoint, map_location=device)
|
| 333 |
model.module.load_state_dict(checkpoint['model_state_dict'])
|
|
|
|
| 334 |
preds = test(model, test_loader, device)
|
| 335 |
if rank == 0: # Only write predictions on rank 0
|
| 336 |
write_predictions(preds, 'predictions.csv')
|
|
|
|
|
|
|
| 337 |
finally:
|
| 338 |
cleanup()
|
| 339 |
-
#
|
| 340 |
torch.cuda.synchronize()
|
| 341 |
if dist.is_initialized():
|
| 342 |
dist.barrier()
|
|
@@ -403,7 +436,7 @@ if __name__ == "__main__":
|
|
| 403 |
parser.add_argument('--test', action='store_true')
|
| 404 |
parser.add_argument('--checkpoint')
|
| 405 |
parser.add_argument('--epochs', type=int, default=100)
|
| 406 |
-
parser.add_argument('--batch_size', type=int, default=
|
| 407 |
parser.add_argument('--port', type=int, default=4224)
|
| 408 |
args = parser.parse_args()
|
| 409 |
main(args)
|
|
|
|
| 2 |
import os
|
| 3 |
import csv
|
| 4 |
import json
|
| 5 |
+
import warnings
|
| 6 |
from tqdm import tqdm
|
| 7 |
import torch
|
| 8 |
import torch.distributed as dist
|
| 9 |
import torch.multiprocessing as mp
|
| 10 |
from torch.nn.parallel import DistributedDataParallel as DDP
|
| 11 |
+
from torch.optim.lr_scheduler import ReduceLROnPlateau
|
| 12 |
from torch.utils.data.distributed import DistributedSampler
|
| 13 |
import argparse
|
| 14 |
from PIL import Image
|
|
|
|
| 38 |
if dist.is_initialized():
|
| 39 |
dist.barrier() # Synchronize all processes before destroying process group
|
| 40 |
dist.destroy_process_group()
|
| 41 |
+
torch.cuda.synchronize()
|
| 42 |
|
| 43 |
|
| 44 |
class MiniPlaces(Dataset):
|
|
|
|
| 164 |
with torch.no_grad():
|
| 165 |
total_loss = 0.0
|
| 166 |
num_correct = 0
|
| 167 |
+
num_correct_top5 = 0
|
| 168 |
num_samples = 0
|
| 169 |
|
| 170 |
for inputs, labels in test_loader:
|
|
|
|
| 177 |
|
| 178 |
_, predictions = torch.max(logits, dim=1)
|
| 179 |
num_correct += (predictions == labels).sum().item()
|
| 180 |
+
|
| 181 |
+
_, top5_predictions = torch.topk(logits, 5, dim=1)
|
| 182 |
+
num_correct_top5 += (top5_predictions == labels.unsqueeze(1)).any(dim=1).sum().item()
|
| 183 |
+
|
| 184 |
num_samples += len(inputs)
|
| 185 |
|
| 186 |
# Gather metrics from all processes
|
| 187 |
world_size = dist.get_world_size()
|
| 188 |
total_loss = torch.tensor(total_loss).to(device)
|
| 189 |
num_correct = torch.tensor(num_correct).to(device)
|
| 190 |
+
num_correct_top5 = torch.tensor(num_correct_top5).to(device)
|
| 191 |
num_samples = torch.tensor(num_samples).to(device)
|
| 192 |
|
| 193 |
dist.all_reduce(total_loss, op=dist.ReduceOp.SUM)
|
| 194 |
dist.all_reduce(num_correct, op=dist.ReduceOp.SUM)
|
| 195 |
+
dist.all_reduce(num_correct_top5, op=dist.ReduceOp.SUM)
|
| 196 |
dist.all_reduce(num_samples, op=dist.ReduceOp.SUM)
|
| 197 |
|
| 198 |
avg_loss = (total_loss / world_size).item() / len(test_loader)
|
| 199 |
accuracy = (num_correct / num_samples).item()
|
| 200 |
+
top5_accuracy = (num_correct_top5 / num_samples).item()
|
| 201 |
|
| 202 |
+
return avg_loss, accuracy, top5_accuracy
|
| 203 |
|
| 204 |
|
| 205 |
def train_worker(rank, world_size, args):
|
|
|
|
| 212 |
args (argparse.Namespace): Command-line arguments.
|
| 213 |
"""
|
| 214 |
try:
|
| 215 |
+
warnings.filterwarnings("ignore")
|
| 216 |
setup(rank, world_size, args.port)
|
| 217 |
device = torch.device(f'cuda:{rank}')
|
| 218 |
|
| 219 |
# Define early stopping parameters
|
| 220 |
+
patience = 10 # Number of epochs to wait for improvement
|
| 221 |
best_val_accuracy = 0.0 # Best validation accuracy so far
|
| 222 |
epochs_without_improvement = 0 # Counter for epochs without improvement
|
| 223 |
best_model_state = None # To store the state of the best model
|
| 224 |
|
| 225 |
+
last_lr = 0
|
| 226 |
+
|
| 227 |
# Separate transforms for training and validation
|
| 228 |
train_transform = create_train_transform()
|
| 229 |
val_transform = create_val_transform()
|
|
|
|
| 247 |
pin_memory=True)
|
| 248 |
|
| 249 |
# Create model and move to GPU
|
| 250 |
+
model = MyModel(num_classes=len(miniplaces_train.label_dict), dropout_rate=0.2)
|
| 251 |
model = model.to(device)
|
| 252 |
model = DDP(model, device_ids=[rank])
|
| 253 |
|
|
|
|
| 261 |
model.module.load_state_dict(checkpoint['model_state_dict'])
|
| 262 |
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
|
| 263 |
|
| 264 |
+
# Initialize the ReduceLROnPlateau scheduler
|
| 265 |
+
scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=4)
|
| 266 |
+
|
| 267 |
if not args.test:
|
| 268 |
# Training loop
|
| 269 |
performance = []
|
|
|
|
| 305 |
# Evaluate and log metrics
|
| 306 |
avg_train_loss = running_loss / len(train_loader)
|
| 307 |
train_accuracy = correct_predictions / total_samples
|
| 308 |
+
avg_val_loss, val_accuracy, val_top5_accuracy = evaluate(model, val_loader, criterion, device)
|
| 309 |
+
|
| 310 |
+
# Step the scheduler with the validation loss
|
| 311 |
+
scheduler.step(avg_val_loss)
|
| 312 |
+
if scheduler.get_last_lr()[0] != last_lr:
|
| 313 |
+
last_lr = scheduler.get_last_lr()[0]
|
| 314 |
+
if epoch != 0:
|
| 315 |
+
print(f"New learning rate: {scheduler.get_last_lr()[0]}")
|
| 316 |
|
| 317 |
if rank == 0: # Only save metrics on rank 0
|
| 318 |
performance.append({
|
|
|
|
| 351 |
torch.save(best_model_state, 'model.ckpt')
|
| 352 |
|
| 353 |
else: # Testing mode
|
| 354 |
+
avg_val_loss, val_accuracy, val_top5_accuracy = evaluate(model, val_loader, criterion, device)
|
| 355 |
+
if rank == 0:
|
| 356 |
+
print(f"\nValidation Loss: {avg_val_loss:.4f}\n"
|
| 357 |
+
f"Validation Accuracy: {val_accuracy:.4f}\n"
|
| 358 |
+
f"Validation Top-5 Accuracy: {val_top5_accuracy:.4f}\n")
|
| 359 |
+
|
| 360 |
+
miniplaces_test = MiniPlaces(data_root, split='test', transform=val_transform)
|
| 361 |
test_loader = DataLoader(miniplaces_test, batch_size=args.batch_size, num_workers=2, shuffle=False)
|
| 362 |
checkpoint = torch.load(args.checkpoint, map_location=device)
|
| 363 |
model.module.load_state_dict(checkpoint['model_state_dict'])
|
| 364 |
+
|
| 365 |
preds = test(model, test_loader, device)
|
| 366 |
if rank == 0: # Only write predictions on rank 0
|
| 367 |
write_predictions(preds, 'predictions.csv')
|
| 368 |
+
print("Predictions saved to predictions.csv\n")
|
| 369 |
+
|
| 370 |
finally:
|
| 371 |
cleanup()
|
| 372 |
+
# Explicit synchronization before exiting
|
| 373 |
torch.cuda.synchronize()
|
| 374 |
if dist.is_initialized():
|
| 375 |
dist.barrier()
|
|
|
|
| 436 |
parser.add_argument('--test', action='store_true')
|
| 437 |
parser.add_argument('--checkpoint')
|
| 438 |
parser.add_argument('--epochs', type=int, default=100)
|
| 439 |
+
parser.add_argument('--batch_size', type=int, default=64)
|
| 440 |
parser.add_argument('--port', type=int, default=4224)
|
| 441 |
args = parser.parse_args()
|
| 442 |
main(args)
|