def test_convolution(self):
batch = 4
width = 8
height = 8
channels = 32
filter_height = 3
filter_width = 3
num_filters = 128
tf_inputs = tf.Variable(
initializer(shape=[batch, height, width, channels], dtype=self.dtype))
tf_filters = tf.Variable(
initializer(
shape=[filter_height, filter_width, channels, num_filters],
dtype=self.dtype))
tf_results = tf.nn.conv2d(
tf_inputs, tf_filters, strides=[1, 1], padding="SAME",
data_format="NHWC", dilations=None)
inputs = Tensor(data_layout=types_pb2.NHWC, tensor_data=tf_inputs.numpy())
filters = Tensor(
data_layout=types_pb2.NHWC,
tensor_data=np.transpose(tf_filters.numpy(), (3, 0, 1, 2)))
with Graph(name=self.graph_name, backend=self.backend) as graph:
nn_ops.convolution(inputs, filters, stride=[1, 1], padding="same")
self.runAndValidate(graph, tf_results, decimal=2)
def test_use_existing_data_op(self):
with Graph(graph_name, backend) as test_graph:
x = Tensor(data_layout=types_pb2.N, tensor_data=np.array([1]))
y = Tensor(data_layout=types_pb2.N, tensor_data=np.array([1]))
x_ = data_op.input_data(x)
res= math_ops.add(x, y)
self.assertEqual(get_num_data_nodes(test_graph), 2)
def test_shared_data_op(self):
with Graph(graph_name, backend) as test_graph:
x = Tensor(data_layout=types_pb2.N, tensor_data=np.array([1]))
y = Tensor(data_layout=types_pb2.N, tensor_data=np.array([1]))
res = math_ops.add(x, y)
res = math_ops.mul(x, res)
self.assertEqual(get_num_data_nodes(test_graph), 2)
def test_use_existing_data_op_in_parent_graph(self):
with Graph(graph_name, backend) as parent_graph:
x = Tensor(data_layout=types_pb2.N, tensor_data=np.array([1]))
y = Tensor(data_layout=types_pb2.N, tensor_data=np.array([1]))
res = math_ops.mul(x, y)
with Graph(graph_name + "_subgraph", backend) as child_graph:
res = math_ops.add(x, y)
self.assertEqual(get_num_data_nodes(parent_graph), 2)
def build_test_sequential_graph(self, backend):
"""Create a sequential model."""
np_dtype = test_backend_dtypes[backend]
self.expected_dtype = datatypes.np_to_smaug_type[np_dtype]
with Graph(name="test_sequential_graph", backend=backend) as graph:
input_tensor = Tensor(
data_layout=types_pb2.NCHW,
tensor_data=np.random.rand(1, 3, 28, 28).astype(np_dtype))
filter_tensor0 = Tensor(
data_layout=types_pb2.NCHW,
tensor_data=np.random.rand(64, 3, 3, 3).astype(np_dtype))
filter_tensor1 = Tensor(
data_layout=types_pb2.NCHW,
tensor_data=np.random.rand(64, 64, 3, 3).astype(np_dtype))
weight_tensor0 = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(254, 12544).astype(np_dtype))
weight_tensor1 = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(10, 254).astype(np_dtype))
bn_mean_tensor = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(np_dtype))
bn_var_tensor = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(np_dtype))
bn_gamma_tensor = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(np_dtype))
bn_beta_tensor = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(np_dtype))
out = data_op.input_data(input_tensor, "input")
out = nn_ops.convolution(
out, filter_tensor0, stride=[1, 1], padding="same", name="conv0")
out = activation_ops.relu(out, "conv0_relu")
out = nn_ops.batch_norm(
out, bn_mean_tensor, bn_var_tensor, bn_gamma_tensor, bn_beta_tensor,
name="bn")
out = nn_ops.convolution(
out, filter_tensor1, stride=[1, 1], padding="same", name="conv1")
out = activation_ops.relu(out, "conv1_relu")
out = nn_ops.max_pool(out, pool_size=[2, 2], stride=[2, 2], name="pool")
out = array_ops.flatten(out, "flatten")
out = nn_ops.mat_mul(out, weight_tensor0, name="fc0")
out = activation_ops.relu(out, "fc0_relu")
out = nn_ops.mat_mul(out, weight_tensor1, name="fc1")
out = array_ops.expand_dims(out, 1, "expand_dims")
out = array_ops.squeeze(out, 1, "squeeze")
out = array_ops.reshape(out, [2, 5], types_pb2.NC, "reshape")
out = array_ops.repeat(out, [4, 2], "repeat")
out = array_ops.stack(out, 4, 1, "stack")
out0, out1, out2, out3 = array_ops.unstack(out, 1, "unstack")
out0 = array_ops.reshape(out0, [1, 1, 8, 10], types_pb2.NCHW, "reshape")
out0 = array_ops.padding(out0, [0, 0, 0, 0, 1, 1, 1, 1], "padding")
self.test_graph, _ = graph.to_proto()
self.backend = backend
self.alignment = global_vars.backend_alignment[backend]
def test_mul(self):
batch = 4
channels = 32
tf_a = tf.Variable(initializer(shape=[batch, channels], dtype=self.dtype))
tf_b = tf.Variable(initializer(shape=[batch, channels], dtype=self.dtype))
tf_result = tf.math.multiply(tf_a, tf_b)
a = Tensor(data_layout=types_pb2.NC, tensor_data=tf_a.numpy())
b = Tensor(data_layout=types_pb2.NC, tensor_data=tf_b.numpy())
with Graph(name=self.graph_name, backend=self.backend) as graph:
math_ops.mul(a, b)
self.runAndValidate(graph, tf_result, decimal=3)
def prepare_states(self):
"""Initialize states as zeros."""
data_type = self.kernel.tensor_data.dtype
num_units = (
self.kernel.shape.dims[0] if self.kernel.shape.layout == types_pb2.NC
else self.wf.shape.dims[1]) // 4
self.h = Tensor(
name=self.name + "/h", data_layout=types_pb2.NC, tensor_data=np.zeros(
(1, num_units), dtype=data_type))
self.c = Tensor(
name=self.name + "/c", data_layout=types_pb2.NC, tensor_data=np.zeros(
(1, num_units), dtype=data_type))
def test_mat_mul(self):
batch = 4
channels = 32
units = 128
tf_a = tf.Variable(initializer(shape=[batch, channels], dtype=self.dtype))
tf_b = tf.Variable(initializer(shape=[units, channels], dtype=self.dtype))
tf_result = tf.linalg.matmul(tf_a, tf_b, transpose_b=True)
a = Tensor(data_layout=types_pb2.NC, tensor_data=tf_a.numpy())
b = Tensor(data_layout=types_pb2.NC, tensor_data=tf_b.numpy())
with Graph(name=self.graph_name, backend=self.backend) as graph:
nn_ops.mat_mul(a, b)
self.runAndValidate(graph, tf_result, decimal=3)
def test_bahdanau_attention(self):
# Build and run an Bahdanau layer in TF.
batch = 2
units = 32
timestep = 8
# Use the Bahdanau attention mechanism.
memory = tf.random.normal([batch, timestep, units], dtype=self.dtype)
attention_mechanism = tfa.seq2seq.BahdanauAttention(units=units,
memory=memory,
dtype=self.dtype)
# Compute attention using the query and state.
tf_cell = tf.keras.layers.LSTMCell(units,
use_bias=False,
unit_forget_bias=False,
dtype=self.dtype)
attention_wrapper = tfa.seq2seq.AttentionWrapper(tf_cell,
attention_mechanism,
output_attention=True,
dtype=self.dtype)
query = tf.random.normal([batch, units], dtype=self.dtype)
tf_initial_state = attention_wrapper.get_initial_state(
batch_size=batch, dtype=self.dtype)
# Perform a step of attention-wrapped RNN.
tf_attention, _ = attention_wrapper(query, tf_initial_state)
# Build the attention model in SMAUG using the tensors from the TF model.
query = Tensor(data_layout=types_pb2.NC, tensor_data=query.numpy())
w, u = recurrent_test.createSmaugWeights(tf_cell)
memory = Tensor(data_layout=types_pb2.NTC, tensor_data=memory.numpy())
weights = attention_mechanism.get_weights()
w_alignment = Tensor(data_layout=types_pb2.NC,
tensor_data=np.expand_dims(weights[0], 0))
w_decoder = Tensor(data_layout=types_pb2.NC,
tensor_data=np.transpose(weights[1]))
w_encoder = Tensor(data_layout=types_pb2.NC,
tensor_data=np.transpose(weights[2]))
with Graph(name=self.graph_name, backend=self.backend) as graph:
# Create an LSTM and an attention, and perform one step.
sg_cell = LSTM([w, u])
sg_attention = BahdanauAttention(memory, w_encoder, w_decoder,
w_alignment)
sg_initial_attention = Tensor(data_layout=types_pb2.NC,
tensor_data=np.zeros(
(batch, units),
dtype=self.dtype))
cell_out, _ = sg_cell.step(concat([query, sg_initial_attention],
axis=1),
timestep=0)
sg_attention(cell_out)
self.runAndValidate(graph, tf_attention, decimal=2)
def build_test_residual_graph(self, backend):
"""Create a residual model.
The graph contains a residual connection, where the output of conv0 and
conv2 is added at the end."""
np_dtype = test_backend_dtypes[backend]
self.expected_dtype = datatypes.np_to_smaug_type[np_dtype]
with Graph(name="test_residual_graph", backend=backend) as graph:
input_tensor = Tensor(
data_layout=types_pb2.NCHW,
tensor_data=np.random.rand(1, 1, 28, 28).astype(np_dtype))
filter_tensor0 = Tensor(
data_layout=types_pb2.NCHW,
tensor_data=np.random.rand(64, 1, 3, 3).astype(np_dtype))
filter_tensor1 = Tensor(
data_layout=types_pb2.NCHW,
tensor_data=np.random.rand(64, 1, 3, 3).astype(np_dtype))
filter_tensor2 = Tensor(
data_layout=types_pb2.NCHW,
tensor_data=np.random.rand(64, 64, 3, 3).astype(np_dtype))
bn_mean_tensor = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(np_dtype))
bn_var_tensor = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(np_dtype))
bn_gamma_tensor = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(np_dtype))
bn_beta_tensor = Tensor(
data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(np_dtype))
act = data_op.input_data(input_tensor, "input")
x = nn_ops.convolution(
act, filter_tensor0, stride=[1, 1], padding="same", name="conv0")
out = nn_ops.convolution(
act, filter_tensor1, stride=[1, 1], padding="same", name="conv1")
out = nn_ops.batch_norm(
out, bn_mean_tensor, bn_var_tensor, bn_gamma_tensor, bn_beta_tensor,
name="bn")
out = activation_ops.relu(out, "relu")
out = nn_ops.convolution(
out, filter_tensor2, stride=[1, 1], padding="same", name="conv2")
out = math_ops.add(x, out, "add")
out = math_ops.mul(x, out, "mul")
# Concatenate the channel dimension of x and out.
axis = 1 if out.shape.layout == types_pb2.NCHW else 3
out = array_ops.concat([x, out], axis, "concat")
# Evenly split the tensor into 4 over the channel dimension.
out0, out1, out2, out3 = array_ops.split(out, 4, axis, "split")
out = math_ops.mul(
math_ops.add(out0, out1, "add1"), math_ops.add(out2, out3, "add2"),
"mul1")
self.test_graph, _ = graph.to_proto()
self.backend = backend
self.alignment = global_vars.backend_alignment[
self.test_graph.backend]
def test_multilayered_lstm(self):
# Build and run an LSTM layer in TF.
tf.keras.backend.set_floatx(
global_vars.backend_datatype[self.backend].__name__)
inputs = tf.random.normal([4, 8, 16],
dtype=global_vars.backend_datatype[self.backend])
model = tf.keras.models.Sequential()
tf_lstm0 = tf.keras.layers.LSTM(
32, return_sequences=True, use_bias=False, unit_forget_bias=False)
# We let TF's LSTM only return the last timestep result, because the SMAUG's
# C++ runtime returns that.
tf_lstm1 = tf.keras.layers.LSTM(
32, return_sequences=False, use_bias=False, unit_forget_bias=False)
model.add(tf_lstm0)
model.add(tf_lstm1)
model.compile()
tf_output = model.predict(inputs)
# Build the model in SMAUG using the tensors from the TF model.
inputs_tensor = Tensor(
data_layout=types_pb2.NTC, tensor_data=inputs.numpy())
w0, u0 = createSmaugWeights(tf_lstm0)
w1, u1 = createSmaugWeights(tf_lstm1)
with Graph(name=self.graph_name, backend=self.backend) as graph:
inputs = input_data(inputs_tensor)
sg_lstm0 = LSTM([w0, u0])
sg_lstm1 = LSTM([w1, u1])
sg_outputs, state = sg_lstm0(inputs)
sg_outputs, state = sg_lstm1(sg_outputs)
self.runAndValidate(graph, tf_output)
def func_true(a, b):
minus_one = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([-1], dtype=self.dtype))
return control_flow_ops.cond(
math_ops.less(a, b),
lambda: math_ops.add(a, math_ops.mul(b, minus_one)),
lambda: math_ops.add(a, b))
def test_attr_smv_padding(self):
"""Test tensor attributes with SMV backend. Additional padding required."""
tensor_data = np.array([[1.1, 2.2, 3.3, 4.4], [5.5, 6.6, 7.7, 8.8]],
dtype=np.float16)
with Graph("test_graph", "SMV") as test_graph:
input_tensor = Tensor(data_layout=types_pb2.NCHW,
tensor_data=tensor_data)
act = input_data(input_tensor, "input")
graph_proto, tensor_data_array = test_graph.to_proto()
self.assertEqual(graph_proto.backend, "SMV")
node = get_node_proto(graph_proto, "input")
self.assertEqual(node.input_tensors[0].data_type, types_pb2.Float16)
self.assertEqual(node.input_tensors[0].shape.dims, [2, 4])
self.assertEqual(node.input_tensors[0].shape.layout, types_pb2.NCHW)
self.assertEqual(node.input_tensors[0].shape.alignment, 8)
tensor_data_proto = get_tensor_data(tensor_data_array,
node.input_tensors[0].name)
self.assertEqualFP16(
tensor_data_proto.half_data,
np.array([
1.1, 2.2, 3.3, 4.4, 0, 0, 0, 0, 5.5, 6.6, 7.7, 8.8, 0, 0, 0, 0
],
dtype=np.float16))
self.assertEqual(len(tensor_data_proto.float_data), 0)
self.assertEqual(len(tensor_data_proto.double_data), 0)
self.assertEqual(len(tensor_data_proto.int_data), 0)
self.assertEqual(len(tensor_data_proto.int64_data), 0)
def test_cond_op_simple_func(self):
with Graph(name=self.graph_name, backend=self.backend) as graph:
x0 = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([2], dtype=self.dtype))
x1 = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([5], dtype=self.dtype))
y = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([10], dtype=self.dtype))
z = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([20], dtype=self.dtype))
expected_res = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([30], dtype=self.dtype))
# res = y + z if x0 < x1 else y * z
res = control_flow_ops.cond(
math_ops.less(x0, x1), lambda: math_ops.add(y, z),
lambda: math_ops.mul(y, z))
self.runAndValidate(graph, expected_res.tensor_data)
def createSmaugWeights(tf_lstm):
""" Extract weights from TF layer and convert them into SMAUG tensors. """
weights = tf_lstm.get_weights()
weights_tensors = []
for w in weights:
weights_tensors.append(
Tensor(data_layout=types_pb2.NC, tensor_data=np.transpose(w)))
return weights_tensors
def func_true(a, b):
minus_one = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([-1], dtype=self.dtype))
res = control_flow_ops.cond(
math_ops.less(a, b),
lambda: math_ops.add(a, math_ops.mul(b, minus_one)),
lambda: math_ops.add(a, b))[0]
# Use the cond results before returning.
return math_ops.mul(res, res)
def test_use_nested_op_result(self):
def func_true(a, b):
minus_one = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([-1], dtype=self.dtype))
res = control_flow_ops.cond(
math_ops.less(a, b),
lambda: math_ops.add(a, math_ops.mul(b, minus_one)),
lambda: math_ops.add(a, b))[0]
# Use the cond results before returning.
return math_ops.mul(res, res)
def func_false(a, b):
two = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([2], dtype=self.dtype))
return control_flow_ops.cond(
math_ops.greater(a, b), lambda: math_ops.mul(a, two),
lambda: math_ops.mul(b, two))
with Graph(name=self.graph_name, backend=self.backend) as graph:
x0 = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([2], dtype=self.dtype))
x1 = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([5], dtype=self.dtype))
y = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([10], dtype=self.dtype))
z = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([20], dtype=self.dtype))
expected_res = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([100],
dtype=self.dtype))
# if x0 < x1:
# if y < z:
# res = (y - z) ^ 2
# else:
# res = y + z
# else:
# if y > z:
# res = 2y
# else:
# res = 2z
res = control_flow_ops.cond(
math_ops.less(x0, x1), lambda: func_true(y, z),
lambda: func_false(y, z))
self.runAndValidate(graph, expected_res.tensor_data)
def test_activation_functions(self):
"""Test activation function attributes."""
with Graph("test_graph", "SMV") as test_graph:
tensor_data = np.random.rand(*self.tensor_shape).astype(np.float16)
input_tensor = Tensor(data_layout=types_pb2.NHWC,
tensor_data=tensor_data)
act = data_op.input_data(input_tensor, "input")
act = activation_ops.relu(act, "relu")
act = activation_ops.lrelu(act, slope=0.5, name="lrelu")
act = activation_ops.elu(act, alpha=0.2, name="elu")
act = activation_ops.selu(act,
alpha=0.4,
lambda_param=0.8,
name="selu")
act = activation_ops.tanh(act, "tanh")
act = activation_ops.hard_tanh(act,
min=-1.5,
max=1.5,
name="hard_tanh")
act = activation_ops.sigmoid(act, "sigmoid")
# Softmax expects NC format, so reorder NHWC to NC.
act = array_ops.reorder(act,
target_layout=types_pb2.NC,
name="reorder")
act = activation_ops.softmax(act, "softmax")
# ReLU
self.do_basic_test(test_graph, "relu", types_pb2.ReLU)
# LReLU
node = self.do_basic_test(test_graph, "lrelu", types_pb2.LReLU)
self.assertAlmostEqual(node.params.act_params.lrelu_params.slope, 0.5)
# ELU
node = self.do_basic_test(test_graph, "elu", types_pb2.ELU)
self.assertAlmostEqual(node.params.act_params.elu_params.alpha, 0.2)
# SELU
node = self.do_basic_test(test_graph, "selu", types_pb2.SELU)
self.assertAlmostEqual(node.params.act_params.elu_params.alpha, 0.4)
self.assertAlmostEqual(node.params.act_params.elu_params.lambda_param,
0.8)
# Tanh
self.do_basic_test(test_graph, "tanh", types_pb2.Tanh)
# HardTanh
node = self.do_basic_test(test_graph, "hard_tanh", types_pb2.HardTanh)
self.assertAlmostEqual(node.params.act_params.hard_tanh_params.min,
-1.5)
self.assertAlmostEqual(node.params.act_params.hard_tanh_params.max,
1.5)
# Sigmoid
self.do_basic_test(test_graph, "sigmoid", types_pb2.Sigmoid)
# Softmax
self.do_basic_test(test_graph,
"softmax",
types_pb2.Softmax,
tensor_shape=[2, 32768])
def add_node(self,
name,
op,
input_tensors,
output_tensors_dims,
output_tensor_layout=types_pb2.NCHW,
output_tensor_dtype=types_pb2.Float32,
output_tensor_dformat=types_pb2.Uncompressed,
params=None):
"""Create a node and add it to graph.
Args:
name: Name of the node. If the name is already used by another node, a
"_N" suffix will be added.
op: Operator type.
input_tensors: A list of input tensors of the node.
output_tensors_dims: A list of dims of the output tensors.
output_tensor_layout: Layout of the output tensor.
output_tensor_dtype: Data type of the output tensor.
output_tensor_dformat: Storage format of the output tensor.
params: The parameters of the node.
Returns:
The output tensor of the added node.
"""
name = self.create_unique_name(name)
node = Node(name, op, params)
self._nodes.append(node)
# Add every input tensor to the node.
for i, tensor in enumerate(input_tensors):
if tensor.name == None:
tensor.name = node.name + "/input%d" % i
node.add_input(tensor)
tensor.targets.append(node)
# Create the output tensor (with the node as its source), and add it to the
# node.
output_tensors = []
for i, d in enumerate(output_tensors_dims):
output_tensor = Tensor(dims=d,
name="%s/output%d" % (node.name, i),
data_layout=output_tensor_layout,
data_type=output_tensor_dtype,
data_format=output_tensor_dformat,
source=node,
source_index=i,
alignment=self._alignment)
node.add_output(output_tensor)
output_tensors.append(output_tensor)
return output_tensors
def test_fp16_odd(self):
"""Test float16 packing when tensor's last dimension is of odd size"""
tensor_data = np.random.rand(4, 3).astype(np.float16)
with Graph("test_graph", "Reference") as test_graph:
input_tensor = Tensor(tensor_data=tensor_data)
act = input_data(input_tensor, "input")
graph_proto, tensor_data_array = test_graph.to_proto()
node = get_node_proto(graph_proto, "input")
self.assertEqual(node.input_tensors[0].data_type, types_pb2.Float16)
tensor_data_proto = get_tensor_data(tensor_data_array,
node.input_tensors[0].name)
self.assertEqualFP16(tensor_data_proto.half_data,
tensor_data.flatten())
def test_cond_op_func_call(self):
def func(a, b):
minus_three = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([-3], dtype=self.dtype))
return math_ops.add(a, math_ops.mul(b, minus_three))
with Graph(name=self.graph_name, backend=self.backend) as graph:
x0 = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([2], dtype=self.dtype))
x1 = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([5], dtype=self.dtype))
y = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([10], dtype=self.dtype))
z = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([20], dtype=self.dtype))
expected_res = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([-50],
dtype=self.dtype))
# res = y - 3z if x0 < x1 else y * z
res = control_flow_ops.cond(
math_ops.less(x0, x1), lambda: func(y, z), lambda: math_ops.mul(y, z))
self.runAndValidate(graph, expected_res.tensor_data)
def test_bidirectional_lstm(self):
# Build and run an BidirectionalLSTM layer in TF.
tf.keras.backend.set_floatx(
global_vars.backend_datatype[self.backend].__name__)
inputs = tf.random.normal([1, 8, 32],
dtype=global_vars.backend_datatype[self.backend])
tf_lstm = tf.keras.layers.LSTM(32, use_bias=False, unit_forget_bias=False)
tf_bilstm = tf.keras.layers.Bidirectional(layer = tf_lstm)
tf_output = tf_bilstm(inputs)
# Build the model in SMAUG using the tensors from the TF model.
input_tensor = Tensor(data_layout=types_pb2.NTC, tensor_data=inputs.numpy())
fwd_w, fwd_u, bwd_w, bwd_u = createSmaugWeights(tf_bilstm)
with Graph(name=self.graph_name, backend=self.backend) as graph:
inputs = input_data(input_tensor)
sg_bilstm = BidirectionalLSTM([fwd_w, fwd_u], [bwd_w, bwd_u])
sg_bilstm(inputs)
self.runAndValidate(graph, tf_output)
def test_attr_reference(self):
"""Test tensor attributes with reference backend."""
tensor_data = np.random.rand(2, 2, 4, 4).astype(np.float32)
with Graph("test_graph", "Reference") as test_graph:
input_tensor = Tensor(data_layout=types_pb2.NHWC,
tensor_data=tensor_data)
act = input_data(input_tensor, "input")
graph_proto, tensor_data_array = test_graph.to_proto()
self.assertEqual(graph_proto.backend, "Reference")
node = get_node_proto(graph_proto, "input")
self.assertEqual(node.input_tensors[0].data_type, types_pb2.Float32)
self.assertEqual(node.input_tensors[0].shape.dims, [2, 2, 4, 4])
self.assertEqual(node.input_tensors[0].shape.layout, types_pb2.NHWC)
self.assertEqual(node.input_tensors[0].shape.alignment, 0)
tensor_data_proto = get_tensor_data(tensor_data_array,
node.input_tensors[0].name)
self.assertEqual(tensor_data_proto.float_data,
list(tensor_data.flatten()))
self.assertEqual(len(tensor_data_proto.half_data), 0)
self.assertEqual(len(tensor_data_proto.double_data), 0)
self.assertEqual(len(tensor_data_proto.int_data), 0)
self.assertEqual(len(tensor_data_proto.int64_data), 0)
def test_attr_smv_no_padding(self):
"""Test tensor attributes with SMV backend. No padding is required."""
tensor_data = np.random.rand(2, 2, 4, 8).astype(np.float16)
with Graph("test_graph", "SMV") as test_graph:
input_tensor = Tensor(data_layout=types_pb2.NCHW,
tensor_data=tensor_data)
act = input_data(input_tensor, "input")
graph_proto, tensor_data_array = test_graph.to_proto()
self.assertEqual(graph_proto.backend, "SMV")
node = get_node_proto(graph_proto, "input")
self.assertEqual(node.input_tensors[0].data_type, types_pb2.Float16)
self.assertEqual(node.input_tensors[0].shape.dims, [2, 2, 4, 8])
self.assertEqual(node.input_tensors[0].shape.layout, types_pb2.NCHW)
self.assertEqual(node.input_tensors[0].shape.alignment, 8)
tensor_data_proto = get_tensor_data(tensor_data_array,
node.input_tensors[0].name)
self.assertEqualFP16(tensor_data_proto.half_data,
tensor_data.flatten())
self.assertEqual(len(tensor_data_proto.float_data), 0)
self.assertEqual(len(tensor_data_proto.double_data), 0)
self.assertEqual(len(tensor_data_proto.int_data), 0)
self.assertEqual(len(tensor_data_proto.int64_data), 0)
def func(a, b):
minus_three = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([-3], dtype=self.dtype))
return math_ops.add(a, math_ops.mul(b, minus_three))
#!/usr/bin/env python
""" This tests the unique name creation in the Graph class."""
import unittest
import numpy as np
from smaug.core import types_pb2
from smaug.python.tensor import Tensor
from smaug.python.graph import Graph
from smaug.python.ops import math_ops
graph_name = "test_graph"
backend = "Reference"
x = Tensor(data_layout=types_pb2.N,
tensor_data=np.random.rand(4).astype(np.float32))
y = Tensor(data_layout=types_pb2.N,
tensor_data=np.random.rand(4).astype(np.float32))
def get_node_names(graph):
return set(
[node.name for node in graph.get_nodes() if node.op != types_pb2.Data])
class TestUniqueName(unittest.TestCase):
def test_default_names(self):
with Graph(graph_name, backend) as test_graph:
res = math_ops.add(x, y)
res = math_ops.mul(x, res)
self.assertEqual(get_node_names(test_graph), {"add", "mul"})
def test_fusing_activation_functions(self):
"""Test activation function when they are fused with other operators."""
with Graph("test_graph", "SMV") as test_graph:
input_tensor = Tensor(data_layout=types_pb2.NHWC,
tensor_data=np.random.rand(
*self.tensor_shape).astype(np.float16))
filter_tensor = Tensor(data_layout=types_pb2.NHWC,
tensor_data=np.random.rand(
32, 3, 3, 32).astype(np.float16))
weight_tensor = Tensor(data_layout=types_pb2.NC,
tensor_data=np.random.rand(
10, 32768).astype(np.float16))
bn_mean_tensor = Tensor(data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(
np.float16))
bn_var_tensor = Tensor(data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(
np.float16))
bn_gamma_tensor = Tensor(data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(
np.float16))
bn_beta_tensor = Tensor(data_layout=types_pb2.NC,
tensor_data=np.random.rand(1, 64).astype(
np.float16))
act = data_op.input_data(input_tensor, "input")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation=None,
name="conv_none")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation="relu",
name="conv_relu")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation="lrelu",
name="conv_lrelu")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation="elu",
name="conv_elu")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation="selu",
name="conv_selu")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation="tanh",
name="conv_tanh")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation="hard_tanh",
name="conv_hard_tanh")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation="sigmoid",
name="conv_sigmoid")
act = nn_ops.convolution(act,
filter_tensor,
stride=[1, 1],
padding="same",
activation="softmax",
name="conv_softmax")
act = nn_ops.batch_norm(act,
bn_mean_tensor,
bn_var_tensor,
bn_gamma_tensor,
bn_beta_tensor,
activation="relu",
name="bn_relu")
out = nn_ops.mat_mul(act,
weight_tensor,
activation="relu",
name="fc_relu")
out = nn_ops.mat_mul(act,
weight_tensor,
activation="lrelu",
activation_params={"slope": 0.1},
name="fc_lrelu")
out = nn_ops.mat_mul(act,
weight_tensor,
activation="elu",
activation_params={"alpha": 0.3},
name="fc_elu")
out = nn_ops.mat_mul(act,
weight_tensor,
activation="selu",
activation_params={
"alpha": 1.0,
"lambda_param": 1.0
},
name="fc_selu")
out = nn_ops.mat_mul(act,
weight_tensor,
activation="hard_tanh",
activation_params={
"min": -2.0,
"max": 2.0
},
name="fc_hard_tanh")
graph_proto, _ = test_graph.to_proto()
# None
node = get_node_proto(graph_proto, "conv_none")
self.assertEqual(node.params.act_params.activation,
types_pb2.UnknownOp)
# ReLU
node = get_node_proto(graph_proto, "conv_relu")
self.assertEqual(node.params.act_params.activation, types_pb2.ReLU)
# LReLU (default slope = 0.2)
node = get_node_proto(graph_proto, "conv_lrelu")
self.assertEqual(node.params.act_params.activation, types_pb2.LReLU)
self.assertAlmostEqual(node.params.act_params.lrelu_params.slope, 0.2)
# ELU (default alpha = 0.1)
node = get_node_proto(graph_proto, "conv_elu")
self.assertEqual(node.params.act_params.activation, types_pb2.ELU)
self.assertAlmostEqual(node.params.act_params.elu_params.alpha, 0.1)
# SELU (default alpha = 1.6733, lambda = 1.0507)
node = get_node_proto(graph_proto, "conv_selu")
self.assertEqual(node.params.act_params.activation, types_pb2.SELU)
self.assertAlmostEqual(node.params.act_params.elu_params.alpha, 1.6733,
5)
self.assertAlmostEqual(node.params.act_params.elu_params.lambda_param,
1.0507, 5)
# Tanh
node = get_node_proto(graph_proto, "conv_tanh")
self.assertEqual(node.params.act_params.activation, types_pb2.Tanh)
# HardTanh (default min = -1, max = 1)
node = get_node_proto(graph_proto, "conv_hard_tanh")
self.assertEqual(node.params.act_params.activation, types_pb2.HardTanh)
self.assertAlmostEqual(node.params.act_params.hard_tanh_params.min, -1)
self.assertAlmostEqual(node.params.act_params.hard_tanh_params.max, 1)
# Sigmoid
node = get_node_proto(graph_proto, "conv_sigmoid")
self.assertEqual(node.params.act_params.activation, types_pb2.Sigmoid)
# Softmax
node = get_node_proto(graph_proto, "conv_softmax")
self.assertEqual(node.params.act_params.activation, types_pb2.Softmax)
# Fusion with inner products and batch norms.
node = get_node_proto(graph_proto, "bn_relu")
self.assertEqual(node.params.act_params.activation, types_pb2.ReLU)
node = get_node_proto(graph_proto, "fc_relu")
self.assertEqual(node.params.act_params.activation, types_pb2.ReLU)
node = get_node_proto(graph_proto, "fc_lrelu")
self.assertEqual(node.params.act_params.activation, types_pb2.LReLU)
self.assertAlmostEqual(node.params.act_params.lrelu_params.slope, 0.1)
node = get_node_proto(graph_proto, "fc_elu")
self.assertEqual(node.params.act_params.activation, types_pb2.ELU)
self.assertAlmostEqual(node.params.act_params.elu_params.alpha, 0.3)
node = get_node_proto(graph_proto, "fc_selu")
self.assertEqual(node.params.act_params.activation, types_pb2.SELU)
self.assertAlmostEqual(node.params.act_params.elu_params.alpha, 1.0)
self.assertAlmostEqual(node.params.act_params.elu_params.lambda_param,
1.0)
node = get_node_proto(graph_proto, "fc_hard_tanh")
self.assertEqual(node.params.act_params.activation, types_pb2.HardTanh)
self.assertAlmostEqual(node.params.act_params.hard_tanh_params.min,
-2.0)
self.assertAlmostEqual(node.params.act_params.hard_tanh_params.max,
2.0)
def func_false(a, b):
two = Tensor(
data_layout=types_pb2.N, tensor_data=np.array([2], dtype=self.dtype))
return control_flow_ops.cond(
math_ops.greater(a, b), lambda: math_ops.mul(a, two),
lambda: math_ops.mul(b, two))