Construected a dataset of binary sequences and stored them as tuples of (sequence, count) pairs. Examples are :-
After this, split the dataset into training, validation and test sets with the following splits - (0.8, 0.1, 0.1)
Tried out various types of hyperparameters and found the best one to be with the following combination: -
- Used layer normalization
- hidden_size = 64
- num_layers = 2
- dropout = 0.2
- lr = 0.0001
- criterion = nn.L1Loss()
Trained the model to get the final loss on training and validation sets as follows:
- Training Loss: 0.3485
- Validation Loss: 0.0866
Plotted the training and validation losses across all epochs on a graph.
Random Baseline MAE: 2.55304
Explanation
-
The lowest MAE at 15 bits reflects that this is the length with the most training experience for the model.
-
Shorter sequences have fewer bits to process, making the input distribution significantly different from the sequences the model encountered during training due to padding.
-
For sequence lengths greater than 16 bits, the model's MAE increases again because the sequences are out-of-distribution, meaning the model hasn’t seen examples with these lengths during training. Longer sequences require the model to handle more complex dependencies over a larger time horizon, which the current architecture might struggle with due to a mismatch in training and test distributions.
Generated the dataset of 100k words. 100 examples of these are present in dataset_images_ocr
The class is present in rnn/ocr.py
The model achieves an accuracy of 67 % on the evaluation metric of average number of correct characters.
| Labels | Decoded Words |
|---|---|
| aardvark | aardvaar |
| aardwolf | aardwolg |
| Aaron | Aarno |
| Aaronic | Aarnoci |
| Aaronical | Aaronucal |
| Aaronite | Aaranite |
| Aaronitic | Aaronttic |
| capkin | capkint |
| capless | capliss |
| caplin | caplni |
| capmaker | capmkarer |
| capmaking | capmanking |
| capman | capmna |
| capmint | capnint |
| deflate | defflate |
| deflation | deflatoin |
| deflationary | deflatinoary |
| deflationist | deflatiionist |
| deflator | deflattor |