Fashion-MNIST is a dataset of Zalando's article images—consisting of a training set of 60,000 examples and a test set of 10,000 examples. Each example is a 28x28 grayscale image, associated with a label from 10 classes. Fashion-MNIST serves as a direct drop-in replacement for the original MNIST dataset for benchmarking machine learning algorithms. It shares the same image size and structure of training and testing splits.
Here's an example how the data looks (each class takes three-rows):
Each training and test example is assigned to one of the following labels:
| Label | Description |
|---|---|
| 0 | T-shirt/top |
| 1 | Trouser |
| 2 | Pullover |
| 3 | Dress |
| 4 | Coat |
| 5 | Sandal |
| 6 | Shirt |
| 7 | Sneaker |
| 8 | Bag |
| 9 | Ankle boot |
A validation dataset of size 12,000 was deduced from the Training dataset with its size being changed to 48,000. We train the following models for 50 epochs.
- Both models have been initialized with random weights sampled from a normal distribution and bias with 0.
- These parameters have been intialized only for the Linear layers present in both of the models.
- If
nrepresents number of nodes in a Linear Layer, then weights are given as a sample of normal distribution in the range(0,y). Hereyrepresents standard deviation calculated asy=1.0/sqrt(n) - Normal distribution is chosen since the probability of choosing a set of weights closer to zero in the distribution is more than that of the higher values. Unlike in Uniform distribution where probability of choosing any value is equal.
Model - 1 : FFNN
- This
Linear Modeluses 784 nodes at input layer, 512, 256 nodes in the first and second hidden layers respectively, with ouput layer of 10 nodes (10 classes). - The test accuracy is 89.56% (This result uses dropout probability of 20%)
- A
FNet_model.pthfile has been included. With this one can directly load the model state_dict and use for testing.
Model - 2 : CNN
- The
Convolutional Neural Netoworkhas 2 convolution layers and pooling layers with 3 fully connected layers. The first convolution layer takes in a channel of dimension 1 since the images are grayscaled. The kernel size is chosen to be of size 3x3 with stride of 1. The output of this convolution is set to 16 channels which means it will extract 16 feature maps using 16 kernels. We pad the image with a padding size of 1 so that the input and output dimensions are same. The output dimension at this layer will be 16 x 28 x 28. The we apply RelU activation to it followed by a max-pooling layer with kernel size of 2 and stride 2. This down-samples the feature maps to dimension of 16 x 14 x 14. - The second convolution layer will have an input channel size of 16. We choose an output channel size to be 32 which means it will extract 32 feature maps. The kernel size for this layer is 3 with stride 1. We again use a padding size of 1 so that the input and output dimension remain the same. The output dimension at this layer will be 32 x 14 x 14. We then follow up it with a RelU activation and a max-pooling layer with kernel of size 2 and stride 2. This down-samples the feature maps to dimension of 32 x 7 x 7.
- Finally, 3 fully connected layers are used. We will pass a flattened version of the feature maps to the first fully connected layer. The fully connected layers have 1568 nodes at input layer, 512, 256 nodes in the first and second hidden layers respectively, with ouput layer of 10 nodes (10 classes). So we have two fully connected layers of size 1568 x 512 followed up by 512 x 256 and 256 x 10.
- The test accuracy is 91.66% (This result uses dropout probability of 20%)
- A
convNet_model.pthfile has been included. With this one can directly load the model state_dict and use for testing.
