def test_wrong_model_name(mock_data, mock_model, use_cluster):
"""User requests to run a model that is not there"""
data = mock_data.create_data(1)
model = mock_model.create_torch_cnn()
c = Client(None, use_cluster)
c.set_model("simple_cnn", model, "TORCH", "CPU")
c.put_tensor("input", data[0])
with pytest.raises(RedisReplyError):
c.run_model("wrong_cnn", ["input"], ["output"])
def test_wrong_model_name_from_file(mock_data, mock_model, use_cluster):
"""User requests to run a model that is not there
that was loaded from file."""
try:
data = mock_data.create_data(1)
mock_model.create_torch_cnn(filepath="./torch_cnn.pt")
c = Client(None, use_cluster)
c.set_model_from_file("simple_cnn_from_file", "./torch_cnn.pt",
"TORCH", "CPU")
c.put_tensor("input", data[0])
with pytest.raises(RedisReplyError):
c.run_model("wrong_cnn", ["input"], ["output"])
finally:
os.remove("torch_cnn.pt")
def test_torch_inference(mock_model, use_cluster):
# get model and set into database
model = mock_model.create_torch_cnn()
c = Client(None, use_cluster)
c.set_model("torch_cnn", model, "TORCH")
# setup input tensor
data = torch.rand(1, 1, 3, 3).numpy()
c.put_tensor("torch_cnn_input", data)
# run model and get output
c.run_model("torch_cnn",
inputs=["torch_cnn_input"],
outputs=["torch_cnn_output"])
out_data = c.get_tensor("torch_cnn_output")
assert out_data.shape == (1, 1, 1, 1)
# Connect a SmartRedis client
db_address = "127.0.0.1:6379"
client = Client(address=db_address, cluster=False)
try:
net = Net()
example_forward_input = torch.rand(1, 1, 3, 3)
# Trace a module (implicitly traces `forward`) and construct a
# `ScriptModule` with a single `forward` method
module = torch.jit.trace(net, example_forward_input)
# Save the traced model to a file
torch.jit.save(module, "./torch_cnn.pt")
# Set the model in the Redis database from the file
client.set_model_from_file("file_cnn", "./torch_cnn.pt", "TORCH", "CPU")
# Put a tensor in the database as a test input
data = torch.rand(1, 1, 3, 3).numpy()
client.put_tensor("torch_cnn_input", data)
# Run model and retrieve the output
client.run_model("file_cnn",
inputs=["torch_cnn_input"],
outputs=["torch_cnn_output"])
out_data = client.get_tensor("torch_cnn_output")
finally:
os.remove("torch_cnn.pt")
c.set_data_source("producer_1" if args.exchange else "producer_0")
data = torch.ones(1, 1, 3, 3).numpy()
data_other = -torch.ones(1, 1, 3, 3).numpy()
elif keyout == "producer_1":
c.set_data_source("producer_0" if args.exchange else "producer_1")
data = -torch.ones(1, 1, 3, 3).numpy()
data_other = torch.ones(1, 1, 3, 3).numpy()
# setup input tensor
c.put_tensor("torch_cnn_input", data)
input_exists = c.poll_tensor("torch_cnn_input", 100, 100)
assert input_exists
other_input = c.get_tensor("torch_cnn_input")
if args.exchange:
assert np.all(other_input == data_other)
else:
assert np.all(other_input == data)
# run model and get output
c.run_model("torch_cnn",
inputs=["torch_cnn_input"],
outputs=["torch_cnn_output"])
output_exists = c.poll_tensor("torch_cnn_output", 100, 100)
assert output_exists
out_data = c.get_tensor("torch_cnn_output")
assert out_data.shape == (1, 1, 1, 1)