def test_tf_dataset_load_prepare_fashion_mnist_incompatible_shape():
""" Load Fashion MNIST dataset (quick), attempt to prepare for input with less dimensions """
test_data = DataSet()
load_test_data('fashion_mnist', test_data, 128)
with pytest.raises(ModelError):
test_data.x_format = keras_prepare_input(np.float, np.array([None,
32]),
test_data.x)
def test_keras_list_models():
model_sequence = KerasModelSequence(DataSet())
model_paths = model_sequence.list_models(['./dnnviewer-data/models'])
assert len(model_paths) > 0
def test_tf_dataset_load_prepare_fashion_mnist():
""" Load Fashion MNIST dataset (quick), prepare for model with 32x32 input """
test_data = DataSet()
load_test_data('fashion_mnist', test_data, 128)
assert test_data.mode is DataSet.MODE_FILESET
assert np.array_equal(test_data.x.shape, [128, 28, 28, 1])
assert test_data.x.dtype is np.dtype('|u1'), "uint8"
assert np.array_equal(test_data.y.shape, [128])
assert test_data.y.dtype is np.dtype('<i8'), "int64"
test_data.x_format = keras_prepare_input(np.float,
np.array([None, 32, 32, 1]),
test_data.x)
assert np.array_equal(test_data.x_format.shape, [128, 32, 32, 1])
assert test_data.x_format.dtype is np.dtype('<f8'), "float32"
def test_tf_dataset_load_mnist():
""" Load MNIST dataset (quick)"""
test_data = DataSet()
load_test_data('mnist', test_data, 128)
assert test_data.mode is DataSet.MODE_FILESET
assert np.array_equal(test_data.x.shape, [128, 28, 28, 1])
assert np.array_equal(test_data.y.shape, [128])
def test_keras_list_model_sequences():
model_sequence = KerasModelSequence(DataSet())
model_paths = model_sequence.list_models(
['./dnnviewer-data/models/FashionMNIST_checkpoints'],
'{model}_{epoch}')
assert len(model_paths) == 3
assert model_paths[
0] == 'dnnviewer-data/models/FashionMNIST_checkpoints/model1_{epoch}'
assert model_paths[
1] == 'dnnviewer-data/models/FashionMNIST_checkpoints/model2_{epoch}'
assert model_paths[
2] == 'dnnviewer-data/models/FashionMNIST_checkpoints/model3_{epoch}'
def test_tf_keras_prepare_labels_unneeded():
""" Prepare labels in case unneeded """
test_data = DataSet()
load_test_data('fashion_mnist', test_data, 128)
assert np.array_equal(test_data.y.shape, [128])
assert test_data.y.dtype is np.dtype('<i8'), "int32"
model = keras.models.Sequential([keras.layers.Dense(10)])
model.compile(loss=keras.losses.BinaryCrossentropy())
prepared_labels = keras_prepare_labels(model, test_data.y)
assert np.array_equal(prepared_labels.shape,
[128]), "unchanged from input labels"
def test_tf_keras_prepare_labels():
""" Prepare labels in case of categorical crossentropy loss """
test_data = DataSet()
load_test_data('fashion_mnist', test_data, 128)
assert np.array_equal(test_data.y.shape, [128])
assert test_data.y.dtype is np.dtype('<i8'), "int32"
model = keras.models.Sequential([keras.layers.Dense(10)])
model.compile(loss=keras.losses.CategoricalCrossentropy(from_logits=False))
prepared_labels = keras_prepare_labels(model, test_data.y)
assert np.array_equal(prepared_labels.shape, [128, 10])
assert prepared_labels.dtype is np.dtype('<f4'), "float32"
async def run_app(args):
""" Run the app """
logger = logging.getLogger('dnnviewer')
# create console handler
console_handler = logging.StreamHandler()
formatter = logging.Formatter(
'%(asctime)s - %(name)s - %(levelname)s - %(message)s')
console_handler.setFormatter(formatter)
logger.addHandler(console_handler)
numeric_level = getattr(logging, args.log_level.upper(), None)
if isinstance(numeric_level, int):
logger.setLevel(numeric_level)
console_handler.setLevel(numeric_level)
# Create App, set stylesheets
app = dash.Dash(
__name__,
assets_folder="assets",
external_stylesheets=[dbc.themes.BOOTSTRAP, font_awesome.CDN_CSS_URL])
app.title = 'DNN Viewer'
# Parameters for model selection
model_selection = {}
# Test data
test_data = DataSet()
model_selection['test_dataset'] = args.test_dataset
model_selection['test_mode'] = DataSet.MODE_FILESET # Default in any case
# Model sequence : currently only supporting from Keras
model_sequence = KerasModelSequence(test_data)
# Initialize the model sequence
model_selection['model'] = args.model_keras
model_selection['sequence'] = args.sequence_keras
preselection = True and model_selection['model'] or model_selection[
'sequence']
# Model selection
if args.model_directories:
model_selection['directories'] = args.model_directories.split(',')
else:
model_selection['directories'] = []
model_selection['pattern'] = args.sequence_pattern
# Top level pages
grapher = Grapher()
router = Router()
router.add_route(
'/',
MainModelSelection(app, model_selection, model_sequence, test_data,
grapher))
router.add_route(
'/network-view',
MainNetworkView(app, model_sequence, test_data, grapher,
not preselection))
def main_layout():
return router.layout()
# Top level layout : manage URL + current page content
app.layout = main_layout
# Install the callbacks
[p.callbacks() for p in router.pages.values()]
router.callbacks(app)
# Run the server, will call the layout
app.run_server(debug=args.debug)
app.serve_routes()