def test_simple_xor_model(self):
predictor = cntk_to_ell.predictor_from_cntk_model('xorModel1.dnn')
result = predictor.Predict([0, 0])
self.assertAlmostEqual(result[0],
0,
msg='incorrect prediction for [0, 0]')
result = predictor.Predict([0, 1])
self.assertAlmostEqual(result[0],
1,
msg='incorrect prediction for [0, 1]')
result = predictor.Predict([1, 0])
self.assertAlmostEqual(result[0],
1,
msg='incorrect prediction for [1, 0]')
result = predictor.Predict([1, 1])
self.assertAlmostEqual(result[0],
0,
msg='incorrect prediction for [1, 1]')
# create a map and save to file
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
ell_map.Save("xor_test.map")
# create a map and save to file
ell_map = ell_utilities.ell_map_from_float_predictor(
predictor,
step_interval_msec=500,
lag_threshold_msec=750,
function_prefix="XorTest")
ell_map.Save("xor_test_steppable.map")
def verify_compiled(self,
predictor,
input,
expectedOutput,
module_name,
method_name,
precision=5):
map = ell_utilities.ell_map_from_float_predictor(predictor)
# Note: for testing purposes, callback functions assume the "model" namespace
compiled = map.Compile("host",
"model",
method_name,
False,
dtype=np.float32)
compiledResults = compiled.Compute(input, dtype=np.float32)
# Compare compiled results
if precision > 0:
np.testing.assert_array_almost_equal(
expectedOutput, compiledResults, precision,
'results for %s layer do not match ELL compiled output !' %
(module_name))
else:
np.testing.assert_array_equal(
expectedOutput, compiledResults,
'results for %s layer do not match ELL compiled output !' %
(module_name))
def test_simple_xor_model(self):
predictor = cntk_to_ell.predictor_from_cntk_model('xorModel1.dnn')
result = predictor.Predict([0, 0])
self.assertAlmostEqual(result[0],
0,
msg='incorrect prediction for [0, 0]')
result = predictor.Predict([0, 1])
self.assertAlmostEqual(result[0],
1,
msg='incorrect prediction for [0, 1]')
result = predictor.Predict([1, 0])
self.assertAlmostEqual(result[0],
1,
msg='incorrect prediction for [1, 0]')
result = predictor.Predict([1, 1])
self.assertAlmostEqual(result[0],
0,
msg='incorrect prediction for [1, 1]')
# create a map and save to file
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
ell_map.Save("xor_test.map")
# create a steppable map and save to file
ell_steppable_map = ell_utilities.\
ell_steppable_map_from_float_predictor(
predictor, 100, "XorInputCallback", "XorOutputCallback")
ell_steppable_map.Save("xor_steppable_test.map")
def test_darknet_model(self):
# Create synthetic input data
input1 = np.arange(28 * 28, dtype=np.float).reshape(28, 28, 1)
# Create an ELL predictor from the darknet model files
predictor = darknet_to_ell.predictor_from_darknet_model(
'unittest.cfg', 'unittest.weights')
# Feed the input through the model
result1 = predictor.Predict(input1.ravel())
# Verify its what we expect
expectedResult1 = [0.09134083986282349, 0.09748589247465134, 0.09064911305904388, 0.13794259727001190, 0.16832095384597778,
0.08976214379072190, 0.06458559632301330, 0.07894224673509598, 0.12377665191888809, 0.05719388648867607]
np.testing.assert_array_almost_equal(
result1, expectedResult1, 5, 'prediction of first input does not match expected results!')
# Feed the next input through the model
input2 = np.flipud(input1)
result2 = predictor.Predict(input2.ravel())
# Verify its what we expect
expectedResult2 = [0.08052270114421844, 0.08739096671342850, 0.08180813491344452, 0.24630726873874664, 0.12944690883159637,
0.08548084646463394, 0.06091265007853508, 0.07173667103052139, 0.11159289628267288, 0.04480091854929924]
np.testing.assert_array_almost_equal(
result2, expectedResult2, 5, 'prediction of second input does not match expected results!')
# create a map and save to file
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
ell_map.Save("darknet_test.map")
# create a steppable map and save to file
ell_steppable_map = ell_utilities.ell_steppable_map_from_float_predictor(
predictor, 100, "DarknetTestInputCallback", "DarknetTestOutputCallback")
ell_steppable_map.Save("darknet_steppable_test.map")
return
def verify_compiled(self,
predictor,
input,
expectedOutput,
module_name,
precision=1e-4):
map = ell_utilities.ell_map_from_float_predictor(predictor)
# Note: for testing purposes, callback functions assume the "model" namespace
compiled = map.Compile("host",
"model",
"test" + str(self.method_index),
False,
dtype=np.float32)
self.method_index += 1
compiledResults = compiled.Compute(input, dtype=np.float32)
ca = np.array(compiledResults) # convert back to numpy
self.compiled_data = ca
expectedFloats = expectedOutput.astype(dtype=np.float32)
# Compare compiled results
self.compare_arrays(
expectedFloats, ca,
'results for %s layer do not match ELL compiled output !' %
(module_name), precision)
def test_simple_xor_model(self):
predictor = cntk_to_ell.predictor_from_cntk_model('xorModel1.dnn')
result = predictor.Predict([0, 0])
self.assertAlmostEqual(
result[0], 0, msg='incorrect prediction for [0, 0]')
result = predictor.Predict([0, 1])
self.assertAlmostEqual(
result[0], 1, msg='incorrect prediction for [0, 1]')
result = predictor.Predict([1, 0])
self.assertAlmostEqual(
result[0], 1, msg='incorrect prediction for [1, 0]')
result = predictor.Predict([1, 1])
self.assertAlmostEqual(
result[0], 0, msg='incorrect prediction for [1, 1]')
# create a map and save to file
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
ell_map.Save("xor_test.map")
# create a steppable map and save to file
ell_steppable_map = ell_utilities.ell_steppable_map_from_float_predictor(
predictor, 100, "XorInputCallback", "XorOutputCallback")
ell_steppable_map.Save("xor_steppable_test.map")
return
def main(argv):
arg_parser = argparse.ArgumentParser(
"Converts CNTK model to ELL model\n"
"Example:\n"
" cntk_import.py VGG16_ImageNet_Caffe.model\n"
"This outputs 'VGG16_ImageNet_Caffe.ell' and 'VGG16_ImageNet_Caffe_config.json'\n"
)
arg_parser.add_argument(
"cntk_model_file",
help="path to a CNTK model file, or a zip archive of a CNTK model file"
)
arg_parser.add_argument("--zip_ell_model",
help="zips the output ELL model if set",
action="store_true")
model_options = arg_parser.add_argument_group('model_options')
model_options.add_argument(
"--step_interval",
help="produce a steppable ELL model for a millisecond interval",
default=0)
model_options.add_argument(
"--lag_threshold",
help=
"number of step intervals to fall behind before notifying the caller.\n"
"used when step_interval is set\n",
default=5)
args = vars(arg_parser.parse_args(argv))
model_options = args.get('model_options', {})
step_interval = model_options.get('step_interval')
lag_threshold = model_options.get('lag_threshold')
# extract the model if it's in an archive
unzip = ziptools.Extractor(args['cntk_model_file'])
success, filename = unzip.extract_file(".cntk")
if success:
print("extracted: " + filename)
else:
# not a zip archive
filename = args['cntk_model_file']
predictor = cntk_to_ell.predictor_from_cntk_model(filename)
model_file_name = os.path.splitext(filename)[0] + '.ell'
ell_map = ell_utilities.ell_map_from_float_predictor(
predictor, step_interval, lag_threshold)
print("Saving model file: '" + model_file_name + "'")
ell_map.Save(model_file_name)
if args['zip_ell_model']:
print("Zipping model file: '" + model_file_name + ".zip'")
zipper = ziptools.Zipper()
zipper.zip_file(model_file_name, model_file_name + ".zip")
os.remove(model_file_name)
def save_ell_predictor_to_file(self, predictor, filePath, intervalMs = 0):
"""Saves an ELL predictor to file so that it can be compiled to run on a device, with an optional stepInterval in milliseconds"""
name = self.model_name
if (intervalMs > 0):
ell_map = ell_utilities.ell_steppable_map_from_float_predictor(
predictor, intervalMs, name + "InputCallback", name + "OutputCallback")
ell_map.Save(filePath)
else:
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
ell_map.Save(filePath)
def get_predictor_map(self, predictor, intervalMs):
"""Creates an ELL map from an ELL predictor"""
import ell_utilities
name = self.model_name
if (intervalMs > 0):
ell_map = ell_utilities.ell_steppable_map_from_float_predictor(
predictor, intervalMs, name + "InputCallback", name + "OutputCallback")
else:
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
return ell_map
def save_ell_predictor_to_file(self, predictor, filePath, intervalMs=0):
"""Saves an ELL predictor to file so that it can be compiled to run on a device, with an optional stepInterval in milliseconds"""
name = self.model_name
if (intervalMs > 0):
ell_map = ell_utilities.ell_steppable_map_from_float_predictor(
predictor, intervalMs, name + "InputCallback",
name + "OutputCallback")
ell_map.Save(filePath)
else:
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
ell_map.Save(filePath)
def run(self):
predictor = darknet_to_ell.predictor_from_darknet_model(
self.config_file, self.weights_file)
weights_directory, weights_filename = os.path.split(self.weights_file)
if self.output_directory:
output_directory = self.output_directory
else:
output_directory = weights_directory
filename_base = os.path.splitext(weights_filename)[0]
model_file_name = filename_base + '.ell'
model_file_path = os.path.join(output_directory, model_file_name)
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
print("Saving model file: '" + model_file_name + "'")
ell_map.Save(model_file_path)
def verify_compiled(self, predictor, input, expectedOutput, module_name,
method_name, precision=5):
# now run same over ELL compiled model
map = ell_utilities.ell_map_from_float_predictor(predictor)
compiled = map.Compile("host", module_name, method_name)
compiledResults = compiled.ComputeFloat(input)
# Compare compiled results
if precision > 0:
np.testing.assert_array_almost_equal(
expectedOutput, compiledResults, precision,
'results for %s layer do not match ELL compiled output !' %
(module_name))
else:
np.testing.assert_array_equal(
expectedOutput, compiledResults,
'results for %s layer do not match ELL compiled output !' %
(module_name))
def verify_compiled(
self, predictor, inputData, expectedOutput, module_name):
"""Takes the input data and passes it through a compiled ELL model
and compare it against the CNTK output, `expectedOutput`.
"""
# now run same over ELL compiled model
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
compiled = ell_map.Compile("host", module_name, "test{}".format(
self.method_index))
self.method_index += 1
compiledResults = np.array(compiled.ComputeFloat(inputData))
# Compare compiled results
np.testing.assert_array_almost_equal(
expectedOutput, compiledResults, decimal=4, err_msg=(
'results for {} layer do not match ELL compiled output!'
).format(module_name))
def main(argv):
arg_parser = argparse.ArgumentParser(
"Converts CNTK model to ELL model\n"
"Example:\n"
" cntk_import.py VGG16_ImageNet_Caffe.model\n"
"This outputs 'VGG16_ImageNet_Caffe.ell' and 'VGG16_ImageNet_Caffe_config.json'\n"
)
arg_parser.add_argument(
"cntk_model_file",
help="path to a CNTK model file, or a zip archive of a CNTK model file"
)
arg_parser.add_argument("--zip_ell_model",
help="zips the output ELL model if set",
action="store_true")
args = arg_parser.parse_args(argv)
# extract the model if it's in an archive
unzip = ziptools.Extractor(args.cntk_model_file)
success, filename = unzip.extract_file(".cntk")
if (success):
print("extracted: " + filename)
else:
# not a zip archive
filename = args.cntk_model_file
predictor = cntk_to_ell.predictor_from_cntk_model(filename)
input_shape = predictor.GetInputShape()
output_shape = predictor.GetOutputShape()
model_file_name = os.path.splitext(filename)[0] + '.ell'
head, tail = os.path.split(model_file_name)
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
print("Saving model file: '" + model_file_name + "'")
ell_map.Save(model_file_name)
if (args.zip_ell_model):
print("Zipping model file: '" + model_file_name + ".zip'")
zipper = ziptools.Zipper()
zipper.zip_file(model_file_name, model_file_name + ".zip")
os.remove(model_file_name)
def verify_compiled(self,
predictor,
input,
expectedOutput,
module_name,
precision=1e-4):
# now run same over ELL compiled model
map = ell_utilities.ell_map_from_float_predictor(predictor)
compiled = map.Compile("host", module_name,
"test" + str(self.method_index))
self.method_index += 1
compiledResults = compiled.ComputeFloat(input)
ca = np.array(compiledResults) # convert back to numpy
self.compiled_data = ca
expectedFloats = expectedOutput.astype(dtype=np.float32)
# Compare compiled results
self.compare_arrays(
expectedFloats, ca,
'results for %s layer do not match ELL compiled output !' %
(module_name), precision)
def verify_compiled(self, predictor, inputData, expectedOutput,
module_name):
"""Takes the input data and passes it through a compiled ELL model
and compare it against the CNTK output, `expectedOutput`.
"""
# Note: for testing purposes, callback functions assume the "model" namespace
ell_map = ell_utilities.ell_map_from_float_predictor(predictor)
compiled = ell_map.Compile("host",
"model",
"test{}".format(self.method_index),
False,
dtype=np.float32)
self.method_index += 1
compiledResults = np.array(
compiled.Compute(inputData, dtype=np.float32))
# Compare compiled results
np.testing.assert_array_almost_equal(
expectedOutput,
compiledResults,
decimal=4,
err_msg=('results for {} layer do not match ELL compiled output!'
).format(module_name))
def model_test_impl(self, modelName):
with self.subTest(modelName=modelName):
_logger.info('Testing {0}.cntk vs ELL ({0})'.format(modelName))
# Load the cntk model
cntkModel = load_model(modelName + '.cntk')
# Import the model into an ELL map live, without unarchiving
predictor = cntk_to_ell.predictor_from_cntk_model(modelName +
'.cntk')
ellMap = ell_utilities.ell_map_from_float_predictor(predictor)
# Load the map from archive
ellMapFromArchive = ell.model.Map(modelName + '.ell')
inputShape = ellMap.GetInputShape()
outputShape = ellMap.GetOutputShape()
# Compile the live map
# Note: for testing purposes, callback functions assume the "model" namespace
ellCompiledMap = ellMap.Compile('host',
'model',
'predict',
False,
dtype=np.float32)
# Compile the unarchived map
# Note: for testing purposes, callback functions assume the "model" namespace
ellCompiledMapFromArchive = ellMapFromArchive.Compile(
'host', 'model', 'predict', False, dtype=np.float32)
cntkInput = np.random.uniform(
high=255,
size=(inputShape.channels, inputShape.rows,
inputShape.columns)).astype(np.float)
ellOrderedInput = cntk_converters.\
get_float_vector_from_cntk_array(cntkInput)
cntkInput = np.ascontiguousarray(cntkInput)
# Get the CNTK results
_, out = cntkModel.forward({
cntkModel.arguments[0]: [cntkInput],
cntkModel.arguments[1]: [list(range(len(self.categories)))]
})
for output in cntkModel.outputs:
if (output.shape == (len(self.categories), )):
out = out[output]
cntkResults = softmax(out[0]).eval()
_logger.info('Comparing predictor output (reference)')
sys.stdout.flush()
ellPredictorResults = predictor.Predict(ellOrderedInput)
# Verify CNTK results and predictor results match
np.testing.assert_array_almost_equal(
cntkResults,
ellPredictorResults,
decimal=5,
err_msg=('results for CNTK and ELL predictor (' + modelName +
') do not match!'))
_logger.info('Comparing map output (reference)')
sys.stdout.flush()
ellMapResults = ellMap.Compute(ellOrderedInput, dtype=np.float32)
# Verify CNTK results and ELL map results match
np.testing.assert_array_almost_equal(
cntkResults,
ellMapResults,
decimal=5,
err_msg=('results for CNTK and ELL map reference (' +
modelName + ') do not match!'))
_logger.info('Comparing unarchived map output (reference)')
sys.stdout.flush()
ellMapFromArchiveResults = self.compute_ell_map(
ellMapFromArchive, ellOrderedInput, cntkResults, modelName)
_logger.info('Comparing map output (compiled)')
sys.stdout.flush()
ellCompiledMapResults = ellCompiledMap.Compute(ellOrderedInput,
dtype=np.float32)
# Verify CNTK results and unarchived ELL model results match
np.testing.assert_array_almost_equal(
cntkResults,
ellCompiledMapResults,
decimal=5,
err_msg=('results for CNTK and ELL map compiled (' +
modelName + ') do not match!'))
_logger.info('Comparing unarchived map output (compiled)')
sys.stdout.flush()
ellCompiledMapFromArchiveResults = ellCompiledMapFromArchive.\
Compute(ellOrderedInput, dtype=np.float32)
# Verify CNTK results and unarchived ELL model results match
np.testing.assert_array_almost_equal(
cntkResults,
ellCompiledMapFromArchiveResults,
decimal=5,
err_msg=('results for CNTK and ELL unarchived map compiled (' +
modelName + ') do not match!'))
_logger.info(('Testing output of {0}.cntk vs ELL for model {0}'
' passed!').format(modelName))
def get_predictor_map(self, predictor, intervalMs):
"""Creates an ELL map from an ELL predictor"""
import ell_utilities
name = self.model_name
return ell_utilities.ell_map_from_float_predictor(predictor)
