Assign2: Transcription Factor Binding via LSTMs and CNNs
Due Date: Feb 9, before midnight (EST 11:59:59PM)
This assignment follows up on the assignment 1: CSCI4969-6969 Assign1. The input data and task remain the same, but you'll try out the LSTM and CNN models.
Deep Learning Models
In this assignment you'll write two models, namely an LSTM and a CNN, to the
predict the binding sites for the JUND transcription factor protein.
You'll compute the weighted loss, and include accessibility information in your model, as before.
See description in assign 1.
For the LSTM model, use the encoder followed by a two layer MLP approach. That is, pass the input sequence (batch) through the LSTM and use the last hidden layer as the representation or embedding vector for the sequence. You can choose the dimensionality of the hidden layer. Next, use this vector as input to a two fully connected MLP layers -- the first connects the input vector to the hidden layer (again you can choose the size of the hidden layer), and the second connects the hidden to the output neuron. Use dropout and relu as appropriate.
Keep in mind that for the input to the LSTM module in pytorch use batch_first=True. This means that the batch dimension comes first, so the input is $(B \times 101 \times 4)$, which is how the input data is structured. Make note of the output of the LSTM layer so that you store the last hidden layer as the representation, to be used as input to the MLP layers.
Also, before feeding the output of the hidden layer to the output layer, you must concatenate the accessibility value. So if you are using hidden dimension of 128, then after concatenating the accessibility value, it will become a 129d vector, which should be fed to the final output layer of size 1, since we have a binary class/label.
You should use binary_cross_entropy_with_logits with weight set to the weights per input element.
You need to train the model on the training data, and use the validation data to select how many epochs you want to use and to choose the hidden dimension. Use the weighted prediction accuracy as the evaluation metric. That is, sum of the weights of the correct predictions divided by the total weight across all the input elements. Finally, report the weighted accuracy on the test data.
For the CNN model, you have to use the 1D convolution module. The in_channels will be 4, one per DNA base. You can decide what number of out_channels and kernel size you want to use. Note the dimensions of the input required for the 1D convolution -- $(B \times 4 \times L)$. That is the input sequence has to be $4\times 101$ and not $101\times 4$ as in the data. So you should use torch.swapaxes function to swap the last two axes (not the batch axis) in the forward function.
After the Conv1d, you can apply a relu activation, do dropout, and then try a maxpooling layer (1d). You can try more than one convolution layer too.
Finally, flatten out the last convolution layer and use as input to a two layer MLP with the same setup as done for the LSTM above.
What to submit
- Upload your jupyter notebooks on submitty, one for the LSTM and another for the CNN.
- The notebook must have output values for the final test accuracy. You can also show the training/val losses and accuracy values.
- Do not submit the data file or directories.