Natsu6767 / temperature-reading

Temperature Reading Prediction

Model Architectures

1. Small_CNN: Conv(5, 5, 64) -> Conv(3, 3, 64) -> Conv(3, 3, 128) -> Conv(3, 3, 256) -> Conv(3, 3, 512) -> Conv(3, 3, 512) -> Linear(8192, 101) During all convolutions, the input is zero padded to make input-output spatial dimensions same and stride is 1. Each convolution layer is followed by a Batch Normalization layer, ReLU activation and 2x2 MaxPooling layer.

2. Large_CNN: Conv(5, 5, 32) -> Conv(3, 3, 64) -> Conv(3, 3, 128) -> Conv(3, 3, 256) -> Conv(3, 3, 512) -> Conv(3, 3, 512) -> Linear(524288, 101) During all convolutions, the input is zero padded to make input-output spatial dimensions same and stride is 1. Each convolution layer is followed by a Batch Normalization layer, ReLU activation. 2x2 MaxPooling layer is added after every 2 convolution layers.

The Adam optimizer with a learning rate of 1e-3 and default values of beta is used. Batch size of 32 is used during training.

Data Pre-Processing

Input images are resized to 256 x 256 dimensions using bilinear interpolation.

Loss Function Used

The loss function used is the Cross Entropy Loss. The problem is framed as a 101 class (0, 1, 2, . . ., 100) classification problem. For the Small_CNN model the MSE Loss is also tried. In this case the problem is framed as a regression problem.

Metric Used

The accuracy is used as the metric for the problem. When doing classification, taking the argmax of the models predictions gives the predicted class. During regression, the models prediction is calculated by rounding the model output to the nearest integer.

Results

All results are averaged over 3 random seeds.

Small CNN Training Curve

Large CNN Training Curve

Small CNN with MSE Training Curve

Final Accuracy on Validation Set (Averaged over 3 seeds)

	Small CNN	Large CNN	Small CNN MSE
Mean	91.99	30.10	44.68
Error	+/- 0.32	+/- 12.71	+/- 30.63

Final Accuracy on Test Set (Averaged over 3 seeds)

	Small CNN	Large CNN	Small CNN MSE
Mean	91.86	30.63	43.78
Error	+/- 0.93	+/- 13.68	+/- 30.08

Visualizations of Results

Small CNN

Correct Results

Prediction	34	47	70
Ground Truth	34	47	70

Incorrect Results

Prediction	22	65	76
Ground Truth	20	64	77

Large CNN

Correct Results

Prediction	26	45	50
Ground Truth	26	45	50

Incorrect Results

Prediction	51	56	85
Ground Truth	36	55	84

Discussion

From the results we notice that when the model makes an error, it is not much far from the ground truth. In fact, the difference is so small it is not easy to distinguish by human inspection. The Smaller model performs better. This could be due to the smaller dataset size which can cause the larger model to overfit. Furthermore, the models were only trained for 10 epochs. This may not have been enough to fully converge the larger model. A future next step is to try data augmentation. Some augmentations that could be useful are vertical flipping and grayscaling the image. Another extension can be to use an off-the shelf digit detector to detect the minimum and maximum readings of the thermometers. This can be used as input to the model and will help provide more information.