Sequence Classifier

Build a sequence classifier: embedding lookup + RNN + linear classifier.

Pipeline:

1. Embed: Look up embeddings for each token index from embedding_table
2. RNN: Process embedded sequence with a vanilla RNN: $h_t = \tanh(e_t \cdot W_{xh} + h_{t-1} \cdot W_{hh} + b_h)$
3. Classify: Apply a linear layer to the final hidden state: $\text{logits} = h_T \cdot W_{out} + b_{out}$

Input:

• indices: token indices of shape (seq_len,)
• embedding_table: shape (vocab_size, embed_dim)
• W_xh: shape (embed_dim, hidden_dim)
• W_hh: shape (hidden_dim, hidden_dim)
• b_h: shape (hidden_dim,)
• h0: shape (hidden_dim,)
• W_out: shape (hidden_dim, num_classes)
• b_out: shape (num_classes,)

Output: Logits of shape (num_classes,).