LSTM benchmark: tensorflow, eager, pytorch

lstm_test_eager.py

#!/usr/bin/env python
import tensorflow as tf
import tensorflow.contrib.eager as tfe

# use 1 CPU
conf=tf.ConfigProto(
    intra_op_parallelism_threads=1,
    inter_op_parallelism_threads=1)

tfe.enable_eager_execution(conf)

n_iter = 100

n_layers = 2

batch_size = 32
seq_len = 1000
input_dim = 7
data = tf.random_uniform((batch_size, seq_len, input_dim))

cells = [tf.contrib.rnn.LSTMCell(input_dim) for _ in range(n_layers)]
multicell = tf.contrib.rnn.MultiRNNCell(cells)


for _ in range(n_iter):
    tf.nn.dynamic_rnn(multicell, data, dtype=tf.float32)

lstm_test_pytorch.py

#!/usr/bin/env python
import torch
import torch.nn as nn
import torch.autograd as autograd

n_iter = 100

n_layers = 2

batch_size = 32
seq_len = 1000
input_dim = 7
x = autograd.Variable(torch.rand(batch_size, seq_len, input_dim))

lstm = nn.LSTM(input_dim, input_dim, n_layers, batch_first=True)

for _ in range(n_iter):
    lstm(x)

lstm_test_tensorflow.py

#!/usr/bin/env python
import numpy as np
import tensorflow as tf


# use 1 CPU
conf=tf.ConfigProto(
    intra_op_parallelism_threads=1,
    inter_op_parallelism_threads=1)

n_iter = 100

n_layers = 2

batch_size = 32
seq_len = 1000
input_dim = 7
data = np.random.uniform(size=(batch_size, seq_len, input_dim))

x = tf.placeholder(tf.float32, shape=(batch_size, seq_len, input_dim))

cells = [tf.contrib.rnn.LSTMCell(input_dim) for _ in range(n_layers)]
multicell = tf.contrib.rnn.MultiRNNCell(cells)
rnn_outputs, final_state = tf.nn.dynamic_rnn(multicell, x, dtype=tf.float32)

init = tf.global_variables_initializer()

with tf.Session(config=conf) as sess:
    sess.run(init)
    for _ in range(n_iter):
        sess.run(rnn_outputs, {x: data})

I'd be curious to see the results of this

On gpu that is