def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
tf.constant(np.arange(10), dtype=tf.float32, name="x")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_const", test_dir)
def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.placeholder(tf.float32, name="x")
one = tf.constant(1.0, dtype=tf.float32, name="one")
tf.add(x, one, name="y")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_placeholder", test_dir)
def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(0.5 * np.random.randn(5, 5), name="x", dtype=tf.float32)
relu = tf.nn.relu(x, name="relu")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_relu", test_dir)
np_relu = relu.eval()
save_idx(np_relu, os.path.join(test_dir, "output_relu.idx"))
def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.random.rand(5, 3), dtype=tf.float32, name="x")
arg_max = tf.argmax(x, axis=1, name='argmax')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_argmax", test_dir)
np_argmax = arg_max.eval()
save_idx(np_argmax, os.path.join(test_dir, "output_argmax.idx"))
def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.random.rand(2, 2, 2), dtype=tf.float32, name="x")
max_x3_1 = tf.reduce_max(x, axis=2, name="max_x3_1")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_max_3_1", test_dir)
np_max_x3_1 = max_x3_1.eval()
save_idx(np_max_x3_1, os.path.join(test_dir, "output_max_x3_1.idx"))
def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.random.rand(2, 2, 2), dtype=tf.float32, name="x")
min_x = tf.reduce_min(x, axis=1, name="min_x_2")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_min_2", test_dir)
np_min_x = min_x.eval()
save_idx(np_min_x, os.path.join(test_dir, "min_2/output_min_x2.idx"))
def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.random.randn(10), dtype=tf.float32, name='x')
output_x = tf.reshape(x, [5, -1], name="output_x")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, 'test_reshape_5', test_dir)
np_output = output_x.eval()
save_idx(np_output, os.path.join(test_dir, 'output_x.idx'))
def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.array([1, 2, 3, 4]), dtype=tf.float32, name="x")
y = tf.constant(np.array([1, 1, 1, 1]), dtype=tf.float32, name="y")
z = tf.add(x, y, name="z")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_add", test_dir)
np_z = z.eval()
save_idx(np_z, os.path.join(test_dir, "output_z.idx"))
def generate():
""" test 1 for min op (fixed with hotfix/64)
"""
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.random.rand(2, 2, 2), dtype=tf.float32, name="x")
min_x = tf.reduce_min(x, name="min_x_1")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_min_1", test_dir)
np_min_x = min_x.eval()
save_idx(np_min_x, os.path.join(test_dir, "output_min_x1.idx"))
def generate():
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
w = tf.constant(np.arange(1, 10).reshape((3, 3)),
dtype=tf.float32,
name="w")
x = tf.constant(np.ones((3, 1)) / 3, dtype=tf.float32, name="x")
z = tf.matmul(w, x, name='z')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_matmul", test_dir)
np_z = z.eval()
save_idx(np_z, os.path.join(test_dir, "output_z.idx"))
def generate():
"""flatten (fail, since we did not have QuantizedReshape in uTensor)
"""
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.arange(10).reshape(2, 5),
dtype=tf.float32,
name='x')
output_x = tf.reshape(x, shape=[-1], name='x_reshape')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, 'test_reshape_1', test_dir)
np_output = output_x.eval()
save_idx(np_output, os.path.join(test_dir, 'output_x.idx'))
def generate():
"""
VALID padding
small floats
"""
test_dir = os.path.abspath(os.path.dirname(__file__))
graph = tf.Graph()
with graph.as_default():
input_data = np.arange(0, 1.0, 1.0 / (3 * 10 * 10 * 5)).reshape((3, 10, 10, 5))
x = tf.constant(input_data, dtype=tf.float32, name="x")
pool1 = tf.nn.max_pool(x, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1],
padding='VALID', name='pool1')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, 'test_max_pool_1', test_dir)
np_output = pool1.eval()
save_idx(np_output, os.path.join(test_dir, 'output_pool.idx'))
def generate():
"""the reshape op will be used in this case since tensorflow will flatten
the input tensor and find the min/max value for quantized matmul
"""
test_dir = os.path.dirname(__file__)
graph = tf.Graph()
with graph.as_default():
x = tf.placeholder(tf.float32, shape=[None, 5], name='x')
w = tf.constant(0.5 * np.random.randn(5, 3),
dtype=tf.float32,
name='w')
y = tf.matmul(x, w, name='y')
np_x = 0.5 * np.random.randn(1, 5).astype(np.float32)
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, 'test_reshape_3', test_dir)
np_output = y.eval(feed_dict={'x:0': np_x})
save_idx(np_x, os.path.join(test_dir, 'input_x.idx'))
save_idx(np_output, os.path.join(test_dir, 'output_y.idx'))
def generate():
"""
add test 3 (broadcasting)
"""
test_dir = os.path.dirname(__file__)
np.random.seed(1234)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.random.randn(3, 3),
dtype=tf.float32,
name='x')
b = tf.constant(1, dtype=tf.float32, name='b')
z = tf.add(b, x, name='z')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_add_3", test_dir)
np_z = z.eval()
save_idx(np_z, os.path.join(test_dir, "output_z.idx"))
def generate():
test_dir = os.path.dirname(__file__)
np.random.seed(1234)
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.random.random((3, 5, 5, 3)),
dtype=tf.float32,
name="x")
w_filter = tf.constant(np.random.random((3, 3, 3, 2)),
dtype=tf.float32,
name="w_filter")
out_conv = tf.nn.conv2d(x, w_filter,
strides=[1, 2, 2, 1],
padding='VALID',
name="out_conv")
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_conv", test_dir)
np_out = sess.run(out_conv)
save_idx(np_out, os.path.join(test_dir, "output_conv.idx"))
def generate():
test_dir = os.path.dirname(__file__)
X = np.stack([0.5 * np.random.randn(50), 0.1 * np.random.randn(50)],
axis=1)
bias = 0.5
W = np.array([0.29, 0.18])
Y = X.dot(W) + bias + np.random.randn(50) * 0.01
train_graph = tf.Graph()
with train_graph.as_default():
tf_x = tf.constant(X, dtype=tf.float32, name="X")
tf_y = tf.constant(Y, dtype=tf.float32, name="Y")
tf_w = tf.Variable(np.random.randn(2, 1), dtype=tf.float32, name="W")
tf_b = tf.Variable(0, dtype=tf.float32, name="b")
tf_yhat = tf.add(tf.matmul(tf_x, tf_w), tf_b, name="yhat")
loss = tf.reduce_mean(tf.pow(tf_yhat - tf_y, 2), name="loss")
train_op = tf.train.AdamOptimizer(0.001).minimize(loss,
name="train_op")
with tf.Session(graph=train_graph) as sess:
tf.global_variables_initializer().run()
for step in range(1, 10001):
_, l = sess.run([train_op, loss])
if step % 1000 == 0:
print("step:", step)
print("loss:", l)
const_graphdef = graph_util.convert_variables_to_constants(
sess, train_graph.as_graph_def(), ["yhat"])
graph = tf.Graph()
with graph.as_default():
tf.import_graph_def(const_graphdef, name='')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, "test_linreg", test_dir)
pred = graph.get_tensor_by_name("yhat:0")
np_pred = pred.eval()
save_idx(np_pred, os.path.join(test_dir, "output_yhat.idx"))
def generate():
"""
VALID padding
large floats
"""
test_dir = os.path.abspath(os.path.dirname(__file__))
graph = tf.Graph()
with graph.as_default():
x = tf.constant(np.arange(3 * 4 * 4 * 5).reshape((3, 4, 4, 5)),
dtype=tf.float32,
name='x')
pool = tf.nn.max_pool(x,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='VALID',
name='pool2')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, 'test_max_pool_2', test_dir)
out_pool = pool.eval()
save_idx(out_pool, os.path.join(test_dir, 'output_max_pool_2.idx'))
def generate():
"""Test SAME paddding
expecting small error
"""
test_dir = os.path.abspath(os.path.dirname(__file__))
graph = tf.Graph()
np.random.seed(3690)
with graph.as_default():
input_data = np.random.randint(0, 256, (3, 4, 4, 5), dtype=np.uint8)
x = tf.constant(input_data, dtype=tf.float32, name="x")
pool5 = tf.nn.max_pool(x,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool5')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, 'test_max_pool_5', test_dir)
np_output = pool5.eval()
save_idx(np_output, os.path.join(test_dir, 'output_max_pool_5.idx'))
def generate():
"""Test SAME paddding
pad_top == 0
pad_left == 0
larger floats
"""
test_dir = os.path.abspath(os.path.dirname(__file__))
graph = tf.Graph()
with graph.as_default():
input_data = np.arange(0, 1, 1 / (3 * 4 * 4 * 5)).reshape((3, 4, 4, 5))
x = tf.constant(input_data, dtype=tf.float32, name="x")
pool3 = tf.nn.max_pool(x,
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME',
name='pool3')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, 'test_max_pool_3', test_dir)
np_output = pool3.eval()
save_idx(np_output, os.path.join(test_dir, 'output_max_pool_3.idx'))
def generate():
"""Test SAME paddding
pad_left != 0
pad_top != 0
"""
test_dir = os.path.abspath(os.path.dirname(__file__))
graph = tf.Graph()
np.random.seed(3690)
with graph.as_default():
input_data = np.random.randint(0, 256, (3, 13, 13, 5))
x = tf.constant(input_data, dtype=tf.float32, name="x")
pool6 = tf.nn.max_pool(x,
ksize=[1, 7, 7, 1],
strides=[1, 3, 3, 1],
padding='SAME',
name='pool6')
with tf.Session(graph=graph) as sess:
save_consts(sess, test_dir)
save_graph(graph, 'test_max_pool_6', test_dir)
np_output = pool6.eval()
save_idx(np_output, os.path.join(test_dir, 'output_max_pool_6.idx'))
