def setUp(self):
self._layers = pickle.load(
open(os.path.join(_TEST_DATA_DIR, 'layers.pickle')))
self._data = pickle.load(
open(os.path.join(_TEST_DATA_DIR, 'data', 'data_batch_5')))
self._decaf_data = translator.imgs_cudaconv_to_decaf(
self._data['data'][:_BATCH_SIZE], 32, 3)
#self._decaf_labels = self._data['labels'].flatten()[:_BATCH_SIZE]
#self._decaf_labels = self._decaf_labels.astype(np.int)
self._output_shapes = {'data': (32, 32, 3), 'labels': -1}
self._net = translator.translate_cuda_network(self._layers,
self._output_shapes)
self._net.predict(data=self._decaf_data)
def setUp(self):
self._layers = pickle.load(open(os.path.join(_TEST_DATA_DIR,
'layers.pickle')))
self._data = pickle.load(open(os.path.join(_TEST_DATA_DIR,
'data',
'data_batch_5')))
self._decaf_data = translator.imgs_cudaconv_to_decaf(
self._data['data'][:_BATCH_SIZE], 32, 3)
#self._decaf_labels = self._data['labels'].flatten()[:_BATCH_SIZE]
#self._decaf_labels = self._decaf_labels.astype(np.int)
self._output_shapes = {'data': (32, 32, 3), 'labels': -1}
self._net = translator.translate_cuda_network(
self._layers, self._output_shapes)
self._net.predict(data=self._decaf_data)
def _testSingleLayer(self, decaf_name, cuda_name, reshape_size=0,
reshape_channels=0, decimal=6):
output = self._net.feature(self._net.provides[decaf_name][0])
self.assertEqual(output.shape[1:], self._output_shapes[decaf_name])
ref_data = pickle.load(open(
os.path.join(_TEST_DATA_DIR, cuda_name, 'data_batch_5')))
ref_data = ref_data['data'][:_BATCH_SIZE]
if reshape_size:
ref_data = translator.imgs_cudaconv_to_decaf(
ref_data, reshape_size, reshape_channels)
# We rescale the data so that the decimal specified would also count
# the original scale of the data.
maxval = ref_data.max()
ref_data /= maxval
output /= maxval
#print 'data range: [%f, %f], max diff: %f' % (
# ref_data.min(), ref_data.max(), np.abs(ref_data - output).max())
np.testing.assert_array_almost_equal(ref_data, output, decimal)
def _testSingleLayer(self,
decaf_name,
cuda_name,
reshape_size=0,
reshape_channels=0,
decimal=6):
output = self._net.feature(self._net.provides[decaf_name][0])
self.assertEqual(output.shape[1:], self._output_shapes[decaf_name])
ref_data = pickle.load(
open(os.path.join(_TEST_DATA_DIR, cuda_name, 'data_batch_5')))
ref_data = ref_data['data'][:_BATCH_SIZE]
if reshape_size:
ref_data = translator.imgs_cudaconv_to_decaf(
ref_data, reshape_size, reshape_channels)
# We rescale the data so that the decimal specified would also count
# the original scale of the data.
maxval = ref_data.max()
ref_data /= maxval
output /= maxval
#print 'data range: [%f, %f], max diff: %f' % (
# ref_data.min(), ref_data.max(), np.abs(ref_data - output).max())
np.testing.assert_array_almost_equal(ref_data, output, decimal)
