RNN inference time with/without pack_padded_sequence comparison.

pytorch_pack_padded_sequence.ipynb

pytorch_pack_padded_sequence.py

import torch
import torch.nn as nn
from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence

# hyper-parameters
BATCH_SIZE = 2
SEQ_LEN = 3
EMBEDDING_DIM = 5
RNN_HIDDEN_SIZE = 5
PADDING_VALUE = 0.0

# init two embeddings in batch: first full-length and second padded
full_length_idx, padded_idx = 0, 1
lengths = torch.tensor([SEQ_LEN, SEQ_LEN - 1])
embeddings = torch.randn(BATCH_SIZE, SEQ_LEN, EMBEDDING_DIM)

# init recurrent layer
rnn = nn.RNN(input_size=EMBEDDING_DIM, hidden_size=RNN_HIDDEN_SIZE, batch_first=True)

# get recurrent layer output for embeddings
output, last_hidden = rnn(embeddings)
last_hidden = last_hidden.transpose(0, 1)  # transpose to [batch_size, seq_len, emb_dim] (batch_first=True)

# pack embeddings
embeddings_packed = pack_padded_sequence(embeddings, lengths=lengths, batch_first=True)

# get recurrent layer output for packed embeddings
output_packed, last_hidden_packed = rnn(embeddings_packed)
last_hidden_packed = last_hidden_packed.transpose(0, 1)  # transpose to [batch_size, seq_len, emb_dim] (batch_first=True)

# unpack recurrent layer output
output_packed, _ = pad_packed_sequence(output_packed, batch_first=True, padding_value=PADDING_VALUE)


### COMPARISON AND RESULTS ###

# rnn outputs and last hidden state with/without pack_padded_sequence are equal for full-length batch element
comparison_1_1 = torch.allclose(  # True
    output[full_length_idx],
    output_packed[full_length_idx],
)
comparison_1_2 = torch.allclose(  # True
    last_hidden[full_length_idx],
    last_hidden_packed[full_length_idx],
)
print(f"1) full-length batch element outputs equivalence: {comparison_1_1}")
print(f"1) full-length batch element last hidden state equivalence: {comparison_1_2}")

# rnn outputs and last hidden state with/without pack_padded_sequence are different for padded batch element
comparison_2_1 = torch.allclose(  # False
    output[padded_idx],
    output_packed[padded_idx],
)
comparison_2_2 = torch.allclose(  # False
    last_hidden[padded_idx],
    last_hidden_packed[padded_idx],
)
print(f"2) padded batch element outputs equivalence: {comparison_2_1}")
print(f"2) padded batch element last hidden state equivalence: {comparison_2_2}")

# pack_padded_sequence fills output pad indicies with pad value
comparison_3 = torch.allclose(  # True
    output_packed[padded_idx][lengths[padded_idx]],
    torch.full(size=(EMBEDDING_DIM,), fill_value=PADDING_VALUE),
)
print(f"3) pad indicies fills with pad value: {comparison_3}")

# pack_padded_sequence gives last hidden state before padding (non-default behavior)
comparison_4_1 = torch.allclose(  # False
    output[padded_idx][lengths[padded_idx] - 1],
    last_hidden[padded_idx],
)
comparison_4_2 = torch.allclose(  # True
    output_packed[padded_idx][lengths[padded_idx] - 1],
    last_hidden_packed[padded_idx],
)
print(f"4) padded batch element last hidden state and last output before padding equivalence: {comparison_4_1}")
print(f"4) padded batch element last hidden state and last output before padding packed equivalence: {comparison_4_2}")


### CONCLUSION ###

# 1) rnn outputs and last hidden state with/without pack_padded_sequence are equal for full-length batch element
# 2) rnn outputs and last hidden state with/without pack_padded_sequence are different for padded batch element
# 3) pack_padded_sequence fills output pad indicies with pad values
# 4) pack_padded_sequence gives last hidden state before padding (non-default behavior)