RTNeural Conv1D Example

conv.py

import numpy as np
import warnings
warnings.filterwarnings("ignore")
import tensorflow as tf
import tensorflow.keras as keras
from tensorflow.keras.layers import Input, Conv1D, PReLU, BatchNormalization
tf.compat.v1.enable_eager_execution()
import matplotlib.pyplot as plt
from model_utils import save_model

input_shape = (128, 1)

x = Input(shape=input_shape,name = "x")
conv1 = Conv1D(filters=12, kernel_size=65, strides=1, dilation_rate=1, activation=None, padding='causal',name = "conv1")(x)
PRelu1 = PReLU(alpha_initializer='glorot_uniform', shared_axes=[1],name = "PRelu1")(conv1)
bn1 = BatchNormalization(momentum=0.0, epsilon=0.01, beta_initializer='random_normal', gamma_initializer='glorot_uniform', moving_mean_initializer="random_normal", moving_variance_initializer="ones",name = "bn1")(PRelu1)
conv2 = Conv1D(filters=8, kernel_size=33, strides=1, dilation_rate=1, activation=None, padding='causal',name = "conv2")(bn1)
PRelu2 = PReLU(alpha_initializer='glorot_uniform', shared_axes=[1],name = "PRelu2")(conv2)
bn2 = BatchNormalization(momentum=0.0, epsilon=0.01, beta_initializer='random_normal', gamma_initializer='glorot_uniform', moving_mean_initializer="random_normal", moving_variance_initializer="ones",name = "bn2")(PRelu2)
conv3 = Conv1D(filters=4, kernel_size=13, strides=1, dilation_rate=1, activation=None, padding='causal',name = "conv3")(bn2)
PRelu3 = PReLU(alpha_initializer='glorot_uniform', shared_axes=[1],name = "PRelu3")(conv3)
bn3 = BatchNormalization(momentum=0.0, epsilon=0.01, beta_initializer='random_normal', gamma_initializer='glorot_uniform', moving_mean_initializer="random_normal", moving_variance_initializer="ones",name = "bn3")(PRelu3)
conv4 = Conv1D(filters=1, kernel_size=5, strides=1, dilation_rate=1, activation="tanh", padding='causal',name = "conv4")(bn3)

model = keras.Model(inputs=x, outputs=conv4)
model.summary()

# construct signals
x_data = 10 * np.sin(np.arange(input_shape[0]) * np.pi * 0.1)
y = model.predict((x_data.reshape((1, -1, 1))))
print(y.shape)
y = y.flatten()

# plot signals
plt.figure()
plt.plot(x_data)
plt.plot(y, '--')
plt.ylim(-1.0, 1.0)
plt.savefig('python/conv.png')

# save signals
np.savetxt('test_data/conv_x_python.csv', x_data, delimiter=',')
np.savetxt('test_data/conv_y_python.csv', y, delimiter=',')

save_model(model, 'models/conv.json')

conv1d_example.cpp

#include "RTNeural/RTNeural.h"
#include "tests/load_csv.hpp"
#include <filesystem>
#include <iostream>

namespace fs = std::filesystem;

std::string getModelFile(fs::path path)
{
    // get path of RTNeural root directory
    while ((--path.end())->string() != "RTNeural")
        path = path.parent_path();

    // get path of model file
    path.append("models/conv.json");

    return path.string();
}

int main(int argc, char* argv[])
{
    std::ifstream modelInputsFile{ "test_data/conv_x_python.csv" };
    std::vector<float> inputs = load_csv::loadFile<float>(modelInputsFile);
    std::cout << "Data with size =  " << inputs.size() <<" are loaded" << std::endl;

    std::ifstream modelOutputsFile{ "test_data/conv_y_python.csv" };
    std::vector<float> referenceOutputs = load_csv::loadFile<float>(modelOutputsFile);

    RTNeural::ModelT<float, 1, 1,
        RTNeural::Conv1DT<float, 1, 12, 65, 1, false>,
        RTNeural::PReLUActivationT<float, 12>,
        RTNeural::BatchNorm1DT<float, 12, true>,
        RTNeural::Conv1DT<float, 12, 8, 33, 1, false>,
        RTNeural::PReLUActivationT<float, 8>,
        RTNeural::BatchNorm1DT<float, 8, true>,
        RTNeural::Conv1DT<float, 8, 4, 13, 1, false>,
        RTNeural::PReLUActivationT<float, 4>,
        RTNeural::BatchNorm1DT<float, 4, true>,
        RTNeural::Conv1DT<float, 4, 1, 5, 1, false>,
        RTNeural::TanhActivationT<float, 1>> modelt;

    auto executablePath = fs::weakly_canonical(fs::path(argv[0]));
    auto modelFilePath = getModelFile(executablePath);

    std::cout << "Loading model from path: " << modelFilePath << std::endl;
    std::ifstream jsonStream(modelFilePath, std::ifstream::binary);

    modelt.parseJson(jsonStream, true);
    modelt.reset();

    std::vector<float> testOutputs;
    testOutputs.resize(referenceOutputs.size(), 0.0f);
    for (size_t i = 0; i < inputs.size(); ++i)
    {
        testOutputs[i] = modelt.forward(inputs.data() + i);
    }

    for (size_t i = 0; i < inputs.size(); ++i)
    {
        std::cout << referenceOutputs[i] << " | " << testOutputs[i] << std::endl;
    }

    return 0;
}

output

/Users/jatin/ChowDSP/Research/RTNeural/RTNeural/build-examples-eigen/examples/conv1d_example/Debug/conv1d_example
Data with size =  128 are loaded
Loading model from path: /Users/jatin/ChowDSP/Research/RTNeural/RTNeural/models/conv.json
# dimensions: 1
Layer: conv1d
  Dims: 12
Layer: prelu
  Dims: 12
Layer: batchnorm
  Dims: 12
Layer: conv1d
  Dims: 8
Layer: prelu
  Dims: 8
Layer: batchnorm
  Dims: 8
Layer: conv1d
  Dims: 4
Layer: prelu
  Dims: 4
Layer: batchnorm
  Dims: 4
Layer: conv1d
  Dims: 1
  activation: tanh
0.0227485 | 0.0227485
0.0469352 | 0.0469352
0.0793129 | 0.0793129
0.108585 | 0.108585
0.0864172 | 0.0864172
0.0816107 | 0.0816108
0.0844359 | 0.0844359
0.0852961 | 0.0852961
0.0859339 | 0.0859338
0.088233 | 0.088233
0.0864823 | 0.0864823
0.0889542 | 0.0889542
0.0874944 | 0.0874944
0.084971 | 0.084971
0.0836835 | 0.0836835
0.0837997 | 0.0837997
0.0825718 | 0.0825717
0.0824659 | 0.0824659
0.082222 | 0.082222
0.0807985 | 0.0807985
0.0798836 | 0.0798836
0.0821496 | 0.0821496
0.0758316 | 0.0758316
0.0753655 | 0.0753656
0.0557327 | 0.0557327
0.0489168 | 0.0489168
0.0440596 | 0.0440596
0.0354367 | 0.0354367
0.0408674 | 0.0408674
0.0411478 | 0.0411478
0.0226464 | 0.0226464
0.0226619 | 0.0226619
0.0246905 | 0.0246904
0.028442 | 0.028442
0.0355418 | 0.0355418
0.0497209 | 0.0497209
0.0642134 | 0.0642134
0.065164 | 0.065164
0.0547634 | 0.0547634
0.0498201 | 0.0498201
0.0443614 | 0.0443614
0.0450946 | 0.0450946
0.035899 | 0.035899
0.0136339 | 0.0136339
-0.00703312 | -0.00703312
-0.0271763 | -0.0271763
-0.0496456 | -0.0496455
-0.0619817 | -0.0619817
-0.0646335 | -0.0646335
-0.0618597 | -0.0618597
-0.0656286 | -0.0656286
-0.0572471 | -0.0572471
-0.0551374 | -0.0551374
-0.0566866 | -0.0566865
-0.0438252 | -0.0438252
-0.0247437 | -0.0247437
-0.0167545 | -0.0167544
-0.00840331 | -0.0084033
-0.00356988 | -0.00356986
-0.0074762 | -0.00747616
-0.00947504 | -0.00947505
-0.0142049 | -0.0142048
-0.00737881 | -0.0073788
-0.00884268 | -0.00884265
-0.0198353 | -0.0198353
-0.0218407 | -0.0218407
-0.0508656 | -0.0508656
-0.0679643 | -0.0679643
-0.0943454 | -0.0943454
-0.115273 | -0.115273
-0.114673 | -0.114673
-0.104984 | -0.104984
-0.123324 | -0.123324
-0.111403 | -0.111403
-0.067003 | -0.0670029
-0.038416 | -0.0384161
-0.0239184 | -0.0239184
-0.0259809 | -0.0259809
-0.0385573 | -0.0385573
-0.0476143 | -0.0476144
-0.0580758 | -0.0580757
-0.0515314 | -0.0515313
-0.0652689 | -0.0652689
-0.0908117 | -0.0908117
-0.0975168 | -0.0975168
-0.105118 | -0.105118
-0.130663 | -0.130663
-0.152049 | -0.152049
-0.167665 | -0.167665
-0.159497 | -0.159497
-0.136552 | -0.136552
-0.115465 | -0.115465
-0.0956429 | -0.095643
-0.0714352 | -0.0714352
-0.0286225 | -0.0286224
-0.00121808 | -0.00121809
0.00518463 | 0.00518466
0.00864545 | 0.00864546
0.00701354 | 0.00701358
-0.00885276 | -0.00885276
-0.00656808 | -0.00656806
-0.0277943 | -0.0277942
-0.0605838 | -0.0605837
-0.0877284 | -0.0877284
-0.11282 | -0.11282
-0.137201 | -0.137201
-0.158516 | -0.158516
-0.172455 | -0.172455
-0.179945 | -0.179945
-0.167146 | -0.167146
-0.137652 | -0.137652
-0.11634 | -0.11634
-0.0851963 | -0.0851962
-0.0660128 | -0.0660128
-0.0213864 | -0.0213864
-0.00102227 | -0.00102228
0.0017092 | 0.00170925
0.00731937 | 0.00731936
-0.00124542 | -0.0012454
-0.0143371 | -0.014337
-0.0119931 | -0.0119931
-0.03146 | -0.03146
-0.061953 | -0.061953
-0.0881181 | -0.0881181
-0.112118 | -0.112118
-0.137004 | -0.137004
-0.157973 | -0.157973
-0.171742 | -0.171742

Process finished with exit code 0