def validate(model, df, input_shape, output_shape, n_tiles, n_classes):
dice_coefs = []
for image_path, label_path in zip(df["preprocessed"], df["label"]):
image = load_nifti(image_path)
label = load_nifti(label_path)
output = feedforward(model, image, input_shape, output_shape, n_tiles,
n_classes)
y = np.int32(np.argmax(output, axis=0))
dice_coefs.append(dice_coefficients(y, label, labels=range(n_classes)))
dice_coefs = np.array(dice_coefs)
return np.mean(dice_coefs, axis=0)
def test_prediction(self):
m,_ = self.X.shape
X = np.hstack(( np.ones((m,1)), self.X ))
predictions = feedforward(X,self.theta1,self.theta2)
# because self.y uses 10 for 0, so the vectorized y representation is shifted.
# if y=10, the output layer will look like [0,0,0,0,0,0,0,0,0,1], so argmax == 9 (i.e. 10 in octave/matlab, which represents class 0)
# if y=1 , the output layer will look like [1,0,0,0,0,0,0,0,0,0], so argmax == 0 (i.e. 1 in octave/matlab, which represents class 1)
# so the fix is, we minus 1 on all elements on y, so the argmax will be 0 indexed, which is good for python.
expected = (self.y - 1).reshape(-1)
acc = accuracy(predictions,expected)
self.assertAlmostEqual(acc, 97.5, places=1)
def test_prediction(self):
m, _ = self.X.shape
X = np.hstack((np.ones((m, 1)), self.X))
predictions = feedforward(X, self.theta1, self.theta2)
# because self.y uses 10 for 0, so the vectorized y representation is shifted.
# if y=10, the output layer will look like [0,0,0,0,0,0,0,0,0,1], so argmax == 9 (i.e. 10 in octave/matlab, which represents class 0)
# if y=1 , the output layer will look like [1,0,0,0,0,0,0,0,0,0], so argmax == 0 (i.e. 1 in octave/matlab, which represents class 1)
# so the fix is, we minus 1 on all elements on y, so the argmax will be 0 indexed, which is good for python.
expected = (self.y - 1).reshape(-1)
acc = accuracy(predictions, expected)
self.assertAlmostEqual(acc, 97.5, places=1)
def __call__(self, inputs, mask):
'''
Args:
inputs: sequence embeddings (item_embeddings + pos_embeddings) shape: (batch_size , max_len, embedding_size)
mask: deal with mask shape: (batch_size, max_len, 1)
Return:
Output sequences which has the same shape with inputs
'''
if self.pos_fixed: # use sin /cos positional embedding
position_encoding = self.get_position_encoding(
inputs) # (batch_size, len, num_units)
inputs += position_encoding
inputs *= mask
for i in range(self.num_blocks):
with tf.variable_scope("num_blocks_%d" % i):
# Self-attention
inputs = multihead_attention(
queries=layer_normalization(inputs),
keys=inputs,
num_units=self.num_units,
num_heads=self.num_heads,
dropout_keep_prob=self.dropout_keep_prob,
causality=False,
scope="self_attention")
# Feed forward
inputs = feedforward(
layer_normalization(inputs),
num_units=[self.num_units, self.num_units],
dropout_keep_prob=self.dropout_keep_prob)
inputs *= mask
outputs = layer_normalization(
inputs) # (batch_size, max_len, num_units)
return outputs
def main():
parser = argparse.ArgumentParser(
description="calculate class probabilities with VoxResNet")
parser.add_argument("--input_file",
"-i",
type=str,
help="input json file of test dataset")
parser.add_argument("--output_suffix",
"-o",
type=str,
help="result of the segmentation")
parser.add_argument(
"--model",
"-m",
type=str,
help="a file containing parameters of trained VoxResNet")
parser.add_argument(
"--input_shape",
type=int,
nargs="*",
action="store",
default=[80, 80, 80],
help="input patch shape of VoxResNet, default=[80, 80, 80]")
parser.add_argument(
"--output_shape",
type=int,
nargs="*",
action="store",
default=[60, 60, 60],
help="output patch shape of VoxResNet, default=[60, 60, 60]")
parser.add_argument("--gpu",
"-g",
default=-1,
type=int,
help="negative value indicates no gpu, default=-1")
parser.add_argument("--n_tiles",
type=int,
nargs="*",
action="store",
default=[5, 5, 5],
help="number of tiles along each axis")
args = parser.parse_args()
print(args)
with open(args.input_file) as f:
dataset = json.load(f)
test_df = pd.DataFrame(dataset["data"])
vrn = VoxResNet(dataset["in_channels"], dataset["n_classes"])
chainer.serializers.load_npz(args.model, vrn)
if args.gpu >= 0:
chainer.cuda.get_device_from_id(args.gpu).use()
vrn.to_gpu()
for image_path, subject in zip(test_df["image"], test_df["subject"]):
image, affine = load_nifti(image_path, with_affine=True)
output = feedforward(vrn, image, args.input_shape, args.output_shape,
args.n_tiles, dataset["n_classes"])
output /= np.sum(output, axis=0, keepdims=True)
nib.save(
nib.Nifti1Image(np.float32(output).transpose(1, 2, 3, 0), affine),
os.path.join(os.path.dirname(image_path),
subject + args.output_suffix))
# Dropout
embeds = tf.nn.dropout(embeds, keep_prob=dropout_keep_prob)
enc = embeds
# Blocks
for i in range(conf.num_blocks):
with tf.variable_scope("num_blocks_{}".format(i)):
# Multihead Attention
enc = utils.multihead_attention(queries=enc,
keys=embeds,
num_units=hidden_units,
num_heads=10,
dropout_rate=dropout_keep_prob,
causality=False)
# Feed Forward
enc = utils.feedforward(
enc, num_units=[4 * hidden_units, hidden_units])
text_embeddings = tf.reduce_mean(enc, axis=1)
else:
tf.logging.info("1D Convolution Model")
sizes = range(2, 5)
result_tensors = []
for ngram_size in sizes:
# 256 -> 2,3 best yet.
text_conv1d = tf.layers.conv1d(
inputs=embeds,
filters=256,
kernel_size=ngram_size,
strides=1,
padding='same',
dilation_rate=1,
activation='relu',
