LSTM cudnn test

tests/simple/test_model.py

@@ -1,43 +1,49 @@
 import keras
 import numpy as np
 from keras.models import Sequential
-from keras.layers import Input, Dense, Activation, Flatten, Dropout, ELU, Reshape, Lambda
+from keras.layers import Input, Dense, Activation, Flatten, Dropout, ELU, Reshape, Lambda, Conv1D
 from keras.layers.convolutional import Convolution2D, Convolution3D
 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
+import struct
+from keras.models import Sequential, Model
 
-def dense_model():
-    model = Sequential()
+def bin_write(f, data):
+    data = data.flatten()
+    fmt = 'f'*len(data)
+    bin = struct.pack(fmt, *data)
+    f.write(bin)
+
+def create_model():
+    x1 = Input((6, 16), name='x1')
+    conv = Conv1D(4, 2)(x1)
+    model = Model([x1], [conv])
+    model.summary()
 
-    model.add(Reshape((10, 10, 1), input_shape=(10, 10)))
-    model.add(Convolution2D(2, (4, 4), subsample=(2, 2), 
-                           bias_initializer='random_uniform', activation="relu"))
-    model.add(Convolution2D(4, (2, 2), subsample=(1, 1), 
-                           bias_initializer='random_uniform', activation="relu"))
-    model.add(Flatten())
-    model.add(Dense(4, bias_initializer='random_uniform', activation="relu"))    
-    sgd = keras.optimizers.Adam(lr=1e-4, decay=1e-8)
-    model.compile(optimizer=sgd, loss="mse")
     return model
 
-
 if __name__ == '__main__':
-    print "DATA FORMAT: ", keras.backend.image_data_format()
+    print ("DATA FORMAT: ", keras.backend.image_data_format())
 
-    model = dense_model()
-    model.save("net.h5")
+    model = create_model()
+    model.save("net.hdf5")
 
-    grid = np.random.rand(10,10)
-    X = grid[None,:,:]
-    i = np.array(grid.flatten(), dtype=np.float32)
-    print i
-    i.tofile("input.bin", format="f")
-    print "Input: ", X    
+    x = np.random.rand(1,1,6,16)
+    r = model.predict( x[0], batch_size=1)
+    r = np.array([r])
+
+    x =  x.transpose(0, 3, 1, 2)
+    r =  r.transpose(0, 3, 1, 2)
+    print("in: ", np.shape(x))
+    print("out: ", np.shape(r))
+
+    x = np.array(x.flatten(), dtype=np.float32)
+    f = open("input.bin", mode='wb')
+    bin_write(f, x)
+
+    r = np.array(r.flatten(), dtype=np.float32)
+    f = open("output.bin", mode='wb')
+    bin_write(f, r)
 
-    r = model.predict( X, batch_size=1)
-    print np.shape(r)  
-    print "Result: ", r
-    print "Result shape: ", np.shape(r) 
-    r.tofile("output.bin", format="f")

tests/simple/test_simple.cpp

@@ -2,23 +2,15 @@
 #include "tkdnn.h"
 
 const char *input_bin   = "../tests/simple/input.bin";
-const char *c0_bin      = "../tests/simple/layers/c0.bin";
-const char *c1_bin      = "../tests/simple/layers/c1.bin";
-const char *d2_bin      = "../tests/simple/layers/d2.bin";
+const char *c0_bin      = "../tests/simple/layers/conv1d_1.bin";
 const char *output_bin  = "../tests/simple/output.bin";
 
 int main() {
 
     // Network layout
-    tk::dnn::dataDim_t dim(1, 1, 10, 10, 1);
+    tk::dnn::dataDim_t dim(1, 16, 1, 6);
     tk::dnn::Network net(dim);
-    tk::dnn::Conv2d     l0(&net, 2, 4, 4, 2, 2, 0, 0, c0_bin);
-    tk::dnn::Activation l1(&net, CUDNN_ACTIVATION_RELU);
-    tk::dnn::Conv2d     l2(&net, 4, 2, 2, 1, 1, 0, 0, c1_bin);
-    tk::dnn::Activation l3(&net, CUDNN_ACTIVATION_RELU);
-    tk::dnn::Flatten    l4(&net);
-    tk::dnn::Dense      l5(&net, 4, d2_bin);
-    tk::dnn::Activation l6(&net, CUDNN_ACTIVATION_RELU);
+    tk::dnn::Conv2d     l0(&net, 4, 1, 2, 1, 1, 0, 0, c0_bin);
 
     // Load input
     dnnType *data;