def test_image_quality_classification(self):
model_path = ck_data.initialize_best_focus_model()
# Make sure to use default op TF config since GPU memory errors during testing will
# occur otherwise (i.e. when gpu_options.allow_growth is not set to True)
miqest = prediction.ImageQualityClassifier(
model_path, DEFAULT_PATCH_SIZE, DEFAULT_N_CLASSES, session_config=op.get_tf_config(None))
img1 = np.zeros((DEFAULT_PATCH_SIZE*2, DEFAULT_PATCH_SIZE*2))
img1[6:10, 6:10] = 1
img2 = filters.gaussian(img1, 5)
# Get class prediction for images (0 = best quality)
max1 = np.argmax(miqest.predict(img1).probabilities)
max2 = np.argmax(miqest.predict(img2).probabilities)
# Assert that probabilities (for ~11 classes) for original image have
# max at first element (i.e. the highest quality class)
self.assertEqual(
max1, 0,
msg='Expecting highest quality class prediction for static image (got class {})'
.format(max1)
)
# Ensure that max class for blurred image (where higher class = lower quality) is greater than original
self.assertTrue(
max2 > max1,
msg='Expecting quality class of original image ({}) to be less than quality class of blurred image ({})'
.format(max1, max2)
)
def initialize(self):
model_path = cytokit_data.initialize_best_focus_model()
self.graph = tf.Graph()
self.mqiest = prediction.ImageQualityClassifier(
model_path,
self.patch_size,
self.n_classes,
graph=self.graph,
session_config=get_tf_config(self))
return self
def _run(self, tile, **kwargs):
if not np.issubdtype(tile.dtype, np.unsignedinteger):
raise ValueError('Only unsigned integer images supported; '
'type given = {}'.format(tile.dtype))
# Tile should have shape (cycles, z, channel, height, width)
ncyc, nz, nch, nh, nw = tile.shape
# Ensure that given tile has same number of channels as required in configuration
# since PSF generation is specific for each channel
if nch != self.config.n_channels_per_cycle:
raise AssertionError(
'Given tile with shape {} ({} channels) does not have expected number of channels {}'
.format(tile.shape, nch, self.config.n_channels_per_cycle))
img_cyc = []
for icyc in range(ncyc):
img_ch = []
for ich in range(nch):
img = tile[icyc, :, ich, :, :]
# Skip deconvolution if channel was configured to be ignored
if self.channels is not None and ich not in self.channels:
img_ch.append(img)
continue
acq = fd_data.Acquisition(img, kernel=self.psfs[ich])
logger.debug(
'Running deconvolution for cycle {}, channel {} [dtype = {}]'
.format(icyc, ich, acq.data.dtype))
res = self.algo.run(acq,
self.n_iter,
session_config=get_tf_config(self)).data
# Restore mean intensity if a scale factor was given
if self.scale_factor is not None:
res, mean_ratio = rescale_stack(acq.data, res,
self.scale_factor)
self.record({
'mean_ratio': mean_ratio,
'cycle': icyc,
'channel': ich
})
# Clip float32 and convert to type of original image (i.e. w/ no scaling)
res = np_utils.arr_to_uint(res, acq.data.dtype)
img_ch.append(res)
img_cyc.append(np.stack(img_ch, 1))
return np.stack(img_cyc, 0)
def _run(self, tile, **kwargs):
if not np.issubdtype(tile.dtype, np.unsignedinteger):
raise ValueError('Only unsigned integer images supported; '
'type given = {}'.format(tile.dtype))
# Tile should have shape (cycles, z, channel, height, width)
dims = self.config.tile_dims
if dims != tile.shape:
raise AssertionError(
'Given tile with shape {} does not match expected shape {}'.
format(tile.shape, dims))
ncyc, nz, nch, nh, nw = dims
img_cyc = []
for icyc in range(ncyc):
img_ch = []
for ich in range(nch):
acq = fd_data.Acquisition(tile[icyc, :, ich, :, :],
kernel=self.psfs[ich])
logger.debug(
'Running deconvolution for cycle {}, channel {} [dtype = {}]'
.format(icyc, ich, acq.data.dtype))
res = self.algo.run(acq,
self.n_iter,
session_config=get_tf_config(self)).data
# Restore mean intensity if a scale factor was given
if self.scale_factor is not None:
res, mean_ratio = rescale_stack(acq.data, res,
self.scale_factor)
self.record({
'mean_ratio': mean_ratio,
'cycle': icyc,
'channel': ich
})
# Clip float32 and convert to type of original image (i.e. w/ no scaling)
res = np_utils.arr_to_uint(res, acq.data.dtype)
img_ch.append(res)
img_cyc.append(np.stack(img_ch, 1))
return np.stack(img_cyc, 0)
def set_keras_session(op):
import keras.backend.tensorflow_backend as KTF
import tensorflow as tf
tf_config = cytokit_op.get_tf_config(op)
KTF.set_session(tf.Session(config=tf_config))
