mmr/tkDNN
Archived
1
0
Fork 0
Code Issues 46 Pull Requests 12 Releases 4 Activity

Pooling and big test, ELU seem not to work

Browse Source
This commit is contained in:
Francesco Gatti
2017-06-29 11:37:29 +00:00
parent cc99347560
commit ba215c35c0
5 changed files with 208 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/simple_dense.py
+26 -12
View File
@@ -2,9 +2,9 @@ import keras
import numpy as np
import pickle
from keras.models import Sequential
from keras.layers import Input, Dense, Activation, Flatten, Dropout, ELU, Reshape
from keras.layers import Input, Dense, Activation, Flatten, Dropout, ELU, Reshape, Lambda
from keras.layers.convolutional import Convolution2D, Convolution3D
from keras.layers.pooling import MaxPooling2D, MaxPooling3D
from keras.layers.pooling import MaxPooling2D, MaxPooling3D, AveragePooling3D
from keras.models import Sequential, Model
from keras.layers import Cropping2D
import keras.backend.tensorflow_backend as KTF
@@ -12,14 +12,24 @@ from weights_exporter import *
def dense_model():
model = Sequential()
model.add(Reshape((10, 10, 4, 1), input_shape=(10, 10, 4)))
model.add(Convolution3D(2, (4, 4, 2), subsample=(2, 2, 1), activation="relu",
bias_initializer='random_uniform'))
model.add(Convolution3D(4, (2, 2, 2), subsample=(1, 1, 1),
bias_initializer='random_uniform'))
model.add(ELU())
model.add(Reshape((100, 100, 4, 1), input_shape=(100, 100, 4)))
model.add(Lambda(lambda x: 2*x - 1.,
batch_input_shape=(1, 100, 100, 4), # 100by100by2
output_shape=(100, 100, 4, 1))) # 100by100by2
model.add(Convolution3D(16, kernel_size=(8, 8, 2), subsample=(4, 4, 1), border_mode="valid",
bias_initializer="random_uniform", activation="relu"))
#model.add(ELU())
model.add(AveragePooling3D(pool_size=(2, 2, 1)))
model.add(Convolution3D(16, kernel_size=(4, 4, 2), subsample=(2, 2, 1), border_mode="valid",
bias_initializer="random_uniform", activation="relu"))
model.add(Convolution3D(24, kernel_size=(3, 3, 2), subsample=(1, 1, 1), border_mode="valid",
bias_initializer="random_uniform", activation="relu"))
#model.add(ELU())
model.add(Flatten())
model.add(Dense(256, activation="relu", bias_initializer="random_uniform"))
model.add(Dense(32, activation="relu", bias_initializer="random_uniform"))
#model.add(ELU())
model.add(Dense(2, bias_initializer="random_uniform"))
sgd = keras.optimizers.Adam(lr=1e-4, decay=1e-8)
model.compile(optimizer=sgd, loss="mse")
@@ -31,10 +41,14 @@ if __name__ == '__main__':
model = dense_model()
wg = model.get_weights()
export_conv3d("conv0", wg[0], wg[1])
export_conv3d("conv1", wg[2], wg[3])
export_conv3d("conv0", wg[0], wg[1])
export_conv3d("conv1", wg[2], wg[3])
export_conv3d("conv2", wg[4], wg[5])
export_dense ("dense3", wg[6], wg[7])
export_dense ("dense4", wg[8], wg[9])
export_dense ("dense5", wg[10], wg[11])
grid = np.random.rand(10,10,4)
grid = np.random.rand(100, 100,4)
X = grid[None,:,:]
i = np.array(grid.flatten(), dtype=np.float32)
print i
tests/test.cpp
+35 -11
View File
@@ -6,20 +6,35 @@ const char *c0_bin = "../tests/conv0.bin";
const char *c0_bias_bin = "../tests/conv0.bias.bin";
const char *c1_bin = "../tests/conv1.bin";
const char *c1_bias_bin = "../tests/conv1.bias.bin";
const char *d2_bin = "../tests/dense2.bin";
const char *d2_bias_bin = "../tests/dense2.bias.bin";
const char *c2_bin = "../tests/conv2.bin";
const char *c2_bias_bin = "../tests/conv2.bias.bin";
const char *d3_bin = "../tests/dense3.bin";
const char *d3_bias_bin = "../tests/dense3.bias.bin";
const char *d4_bin = "../tests/dense4.bin";
const char *d4_bias_bin = "../tests/dense4.bias.bin";
const char *d5_bin = "../tests/dense5.bin";
const char *d5_bias_bin = "../tests/dense5.bias.bin";
int main() {
// Network layout
tkDNN::Network net;
tkDNN::dataDim_t dim(1, 1, 10, 10, 4);
tkDNN::Conv3d c0 (&net, dim, 2, 4, 4, 2, 2, 2, 1, c0_bin, c0_bias_bin);
tkDNN::dataDim_t dim(1, 1, 100, 100, 4);
tkDNN::MulAdd m0 (&net, dim, 2, -1);
tkDNN::Conv3d c0 (&net, m0.output_dim, 16, 8, 8, 2, 4, 4, 1, c0_bin, c0_bias_bin);
tkDNN::Activation a0 (&net, c0.output_dim, tkDNN::ACTIVATION_RELU);
tkDNN::Conv3d c1 (&net, a0.output_dim, 4, 2, 2, 2, 1, 1, 1, c1_bin, c1_bias_bin);
tkDNN::Activation a1 (&net, c1.output_dim, tkDNN::ACTIVATION_ELU);
tkDNN::Flatten f1 (&net, a1.output_dim);
tkDNN::MulAdd m1 (&net, f1.output_dim, 2, 1);
tkDNN::Pooling p0 (&net, a0.output_dim, 2, 2, 2, 2, tkDNN::POOLING_AVERAGE);
tkDNN::Conv3d c1 (&net, p0.output_dim, 16, 4, 4, 2, 2, 2, 1, c1_bin, c1_bias_bin);
tkDNN::Activation a1 (&net, c1.output_dim, tkDNN::ACTIVATION_RELU);
tkDNN::Conv3d c2 (&net, a1.output_dim, 24, 3, 3, 2, 1, 1, 1, c2_bin, c2_bias_bin);
tkDNN::Activation a2 (&net, c2.output_dim, tkDNN::ACTIVATION_RELU);
tkDNN::Flatten f2 (&net, a2.output_dim);
tkDNN::Dense d3 (&net, f2.output_dim, 256, d3_bin, d3_bias_bin);
tkDNN::Activation a3 (&net, d3.output_dim, tkDNN::ACTIVATION_RELU);
tkDNN::Dense d4 (&net, a3.output_dim, 32, d4_bin, d4_bias_bin);
tkDNN::Activation a4 (&net, d4.output_dim, tkDNN::ACTIVATION_RELU);
tkDNN::Dense d5 (&net, a4.output_dim, 2, d5_bin, d5_bias_bin);
// Load input
value_type *data;
@@ -31,14 +46,23 @@ int main() {
TIMER_START
// Inference
data = m0.infer(dim, data); dim.print();
data = c0.infer(dim, data); dim.print();
data = a0.infer(dim, data); dim.print();
data = p0.infer(dim, data); dim.print();
data = c1.infer(dim, data); dim.print();
data = a1.infer(dim, data); dim.print();
data = f1.infer(dim, data); dim.print();
data = m1.infer(dim, data); dim.print();
data = c2.infer(dim, data); dim.print();
data = a2.infer(dim, data); dim.print();
data = f2.infer(dim, data); dim.print();
data = d3.infer(dim, data); dim.print();
data = a3.infer(dim, data); dim.print();
data = d4.infer(dim, data); dim.print();
data = a4.infer(dim, data); dim.print();
data = d5.infer(dim, data); dim.print();
TIMER_STOP
// Print result
printDeviceVector(dim.tot(), data);
return 0;