6.4.1 RNN Roadmap: Process Sequences Step by Step

RNNs are built for ordered data: text, time series, clicks, sensor readings, and any input where earlier steps affect later steps.

Look at the Sequence Flow First

Concept	First meaning
sequence length	how many time steps
input size	features per step
hidden state	rolling memory
LSTM / GRU	gated memory control
batch first	shape style [batch, seq_len, features]

Run One GRU Shape Check

Create rnn_first_loop.py and run it after installing torch.

import torch

sequence = torch.randn(2, 3, 5)
gru = torch.nn.GRU(input_size=5, hidden_size=4, batch_first=True)
outputs, hidden = gru(sequence)

print("sequence_shape:", tuple(sequence.shape))
print("outputs_shape:", tuple(outputs.shape))
print("hidden_shape:", tuple(hidden.shape))

Expected output:

sequence_shape: (2, 3, 5)
outputs_shape: (2, 3, 4)
hidden_shape: (1, 2, 4)

Read this as two sequences, three steps each, five features per step. The GRU returns a hidden representation of size 4.

Learn in This Order

Order	Read	What to practice
1	6.4.2 RNN Basics	sequence input, hidden state, shape
2	6.4.3 LSTM and GRU	gates, long dependency, memory control
3	6.4.4 Sequence Practice	sliding windows, train/eval loop

Pass Check

You pass this roadmap when you can read [batch, seq_len, features], explain hidden state as rolling memory, and know why LSTM/GRU were introduced for longer dependencies.