diff --git a/python/paddle/trainer_config_helpers/default_decorators.py b/python/paddle/trainer_config_helpers/default_decorators.py index ad3efcbf369411b9c42b2a32ed05b04f86bf7de6..2f25579fcdd9793e4c165439c9934a2bccb63617 100644 --- a/python/paddle/trainer_config_helpers/default_decorators.py +++ b/python/paddle/trainer_config_helpers/default_decorators.py @@ -52,6 +52,10 @@ def wrap_param_default(param_names=None, kwargs[name] = default_factory(func) return func(*args, **kwargs) + if hasattr(func, 'argspec'): + __wrapper__.argspec = func.argspec + else: + __wrapper__.argspec = inspect.getargspec(func) return __wrapper__ return __impl__ diff --git a/python/paddle/trainer_config_helpers/layers.py b/python/paddle/trainer_config_helpers/layers.py index 1bb1a01d509e6412c254fce856101137e66b1e12..b68460b6a3ab621904f4dc4e48352044ab265a38 100755 --- a/python/paddle/trainer_config_helpers/layers.py +++ b/python/paddle/trainer_config_helpers/layers.py @@ -14,6 +14,7 @@ import functools import collections +import inspect from paddle.trainer.config_parser import * from .activations import LinearActivation, SigmoidActivation, TanhActivation, \ @@ -316,6 +317,11 @@ def layer_support(*attrs): val.check(method.__name__) return method(*args, **kwargs) + if hasattr(method, 'argspec'): + wrapper.argspec = method.argspec + else: + wrapper.argspec = inspect.getargspec(method) + return wrapper return decorator diff --git a/python/paddle/v2/dataset/cifar.py b/python/paddle/v2/dataset/cifar.py index 2ac71c6effe9d5f1140d1f574db9c9848b56433a..77c54bd268b5d988b0802a3edca91605e56f730e 100644 --- a/python/paddle/v2/dataset/cifar.py +++ b/python/paddle/v2/dataset/cifar.py @@ -1,82 +1,61 @@ """ -CIFAR Dataset. - -URL: https://www.cs.toronto.edu/~kriz/cifar.html - -the default train_creator, test_creator used for CIFAR-10 dataset. +CIFAR dataset: https://www.cs.toronto.edu/~kriz/cifar.html """ import cPickle import itertools -import tarfile - import numpy +import paddle.v2.dataset.common +import tarfile -from config import download - -__all__ = [ - 'cifar_100_train_creator', 'cifar_100_test_creator', 'train_creator', - 'test_creator' -] +__all__ = ['train100', 'test100', 'train10', 'test10'] -CIFAR10_URL = 'https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz' +URL_PREFIX = 'https://www.cs.toronto.edu/~kriz/' +CIFAR10_URL = URL_PREFIX + 'cifar-10-python.tar.gz' CIFAR10_MD5 = 'c58f30108f718f92721af3b95e74349a' -CIFAR100_URL = 'https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz' +CIFAR100_URL = URL_PREFIX + 'cifar-100-python.tar.gz' CIFAR100_MD5 = 'eb9058c3a382ffc7106e4002c42a8d85' -def __read_batch__(filename, sub_name): - def reader(): - def __read_one_batch_impl__(batch): - data = batch['data'] - labels = batch.get('labels', batch.get('fine_labels', None)) - assert labels is not None - for sample, label in itertools.izip(data, labels): - yield (sample / 255.0).astype(numpy.float32), int(label) +def reader_creator(filename, sub_name): + def read_batch(batch): + data = batch['data'] + labels = batch.get('labels', batch.get('fine_labels', None)) + assert labels is not None + for sample, label in itertools.izip(data, labels): + yield (sample / 255.0).astype(numpy.float32), int(label) + def reader(): with tarfile.open(filename, mode='r') as f: names = (each_item.name for each_item in f if sub_name in each_item.name) for name in names: batch = cPickle.load(f.extractfile(name)) - for item in __read_one_batch_impl__(batch): + for item in read_batch(batch): yield item return reader -def cifar_100_train_creator(): - fn = download(url=CIFAR100_URL, md5=CIFAR100_MD5) - return __read_batch__(fn, 'train') - - -def cifar_100_test_creator(): - fn = download(url=CIFAR100_URL, md5=CIFAR100_MD5) - return __read_batch__(fn, 'test') - - -def train_creator(): - """ - Default train reader creator. Use CIFAR-10 dataset. - """ - fn = download(url=CIFAR10_URL, md5=CIFAR10_MD5) - return __read_batch__(fn, 'data_batch') +def train100(): + return reader_creator( + paddle.v2.dataset.common.download(CIFAR100_URL, 'cifar', CIFAR100_MD5), + 'train') -def test_creator(): - """ - Default test reader creator. Use CIFAR-10 dataset. - """ - fn = download(url=CIFAR10_URL, md5=CIFAR10_MD5) - return __read_batch__(fn, 'test_batch') +def test100(): + return reader_creator( + paddle.v2.dataset.common.download(CIFAR100_URL, 'cifar', CIFAR100_MD5), + 'test') -def unittest(): - for _ in train_creator()(): - pass - for _ in test_creator()(): - pass +def train10(): + return reader_creator( + paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), + 'data_batch') -if __name__ == '__main__': - unittest() +def test10(): + return reader_creator( + paddle.v2.dataset.common.download(CIFAR10_URL, 'cifar', CIFAR10_MD5), + 'test_batch') diff --git a/python/paddle/v2/dataset/common.py b/python/paddle/v2/dataset/common.py new file mode 100644 index 0000000000000000000000000000000000000000..a5ffe25a116e9be039bdebaaaad435685e23d372 --- /dev/null +++ b/python/paddle/v2/dataset/common.py @@ -0,0 +1,34 @@ +import requests +import hashlib +import os +import shutil + +__all__ = ['DATA_HOME', 'download', 'md5file'] + +DATA_HOME = os.path.expanduser('~/.cache/paddle/dataset') + +if not os.path.exists(DATA_HOME): + os.makedirs(DATA_HOME) + + +def md5file(fname): + hash_md5 = hashlib.md5() + f = open(fname, "rb") + for chunk in iter(lambda: f.read(4096), b""): + hash_md5.update(chunk) + f.close() + return hash_md5.hexdigest() + + +def download(url, module_name, md5sum): + dirname = os.path.join(DATA_HOME, module_name) + if not os.path.exists(dirname): + os.makedirs(dirname) + + filename = os.path.join(dirname, url.split('/')[-1]) + if not (os.path.exists(filename) and md5file(filename) == md5sum): + r = requests.get(url, stream=True) + with open(filename, 'w') as f: + shutil.copyfileobj(r.raw, f) + + return filename diff --git a/python/paddle/v2/dataset/config.py b/python/paddle/v2/dataset/config.py deleted file mode 100644 index 02a009f09c71ccf6a5292a188565adeeb3f875f6..0000000000000000000000000000000000000000 --- a/python/paddle/v2/dataset/config.py +++ /dev/null @@ -1,36 +0,0 @@ -import hashlib -import os -import shutil -import urllib2 - -__all__ = ['DATA_HOME', 'download'] - -DATA_HOME = os.path.expanduser('~/.cache/paddle_data_set') - -if not os.path.exists(DATA_HOME): - os.makedirs(DATA_HOME) - - -def download(url, md5): - filename = os.path.split(url)[-1] - assert DATA_HOME is not None - filepath = os.path.join(DATA_HOME, md5) - if not os.path.exists(filepath): - os.makedirs(filepath) - __full_file__ = os.path.join(filepath, filename) - - def __file_ok__(): - if not os.path.exists(__full_file__): - return False - md5_hash = hashlib.md5() - with open(__full_file__, 'rb') as f: - for chunk in iter(lambda: f.read(4096), b""): - md5_hash.update(chunk) - - return md5_hash.hexdigest() == md5 - - while not __file_ok__(): - response = urllib2.urlopen(url) - with open(__full_file__, mode='wb') as of: - shutil.copyfileobj(fsrc=response, fdst=of) - return __full_file__ diff --git a/python/paddle/v2/dataset/mnist.py b/python/paddle/v2/dataset/mnist.py index db84f37aa4fc3477b17599a48a4de9b45cfb6c1f..a36c20e3fa3734bdc14c1f47779a61375f298511 100644 --- a/python/paddle/v2/dataset/mnist.py +++ b/python/paddle/v2/dataset/mnist.py @@ -1,39 +1,66 @@ -import sklearn.datasets.mldata -import sklearn.model_selection +""" +MNIST dataset. +""" import numpy -from config import DATA_HOME +import paddle.v2.dataset.common +import subprocess -__all__ = ['train_creator', 'test_creator'] +__all__ = ['train', 'test'] +URL_PREFIX = 'http://yann.lecun.com/exdb/mnist/' +TEST_IMAGE_URL = URL_PREFIX + 't10k-images-idx3-ubyte.gz' +TEST_IMAGE_MD5 = '25e3cc63507ef6e98d5dc541e8672bb6' +TEST_LABEL_URL = URL_PREFIX + 't10k-labels-idx1-ubyte.gz' +TEST_LABEL_MD5 = '4e9511fe019b2189026bd0421ba7b688' +TRAIN_IMAGE_URL = URL_PREFIX + 'train-images-idx3-ubyte.gz' +TRAIN_IMAGE_MD5 = 'f68b3c2dcbeaaa9fbdd348bbdeb94873' +TRAIN_LABEL_URL = URL_PREFIX + 'train-labels-idx1-ubyte.gz' +TRAIN_LABEL_MD5 = 'd53e105ee54ea40749a09fcbcd1e9432' -def __mnist_reader_creator__(data, target): + +def reader_creator(image_filename, label_filename, buffer_size): def reader(): - n_samples = data.shape[0] - for i in xrange(n_samples): - yield (data[i] / 255.0).astype(numpy.float32), int(target[i]) + # According to http://stackoverflow.com/a/38061619/724872, we + # cannot use standard package gzip here. + m = subprocess.Popen(["zcat", image_filename], stdout=subprocess.PIPE) + m.stdout.read(16) # skip some magic bytes - return reader + l = subprocess.Popen(["zcat", label_filename], stdout=subprocess.PIPE) + l.stdout.read(8) # skip some magic bytes + while True: + labels = numpy.fromfile( + l.stdout, 'ubyte', count=buffer_size).astype("int") -TEST_SIZE = 10000 + if labels.size != buffer_size: + break # numpy.fromfile returns empty slice after EOF. -data = sklearn.datasets.mldata.fetch_mldata( - "MNIST original", data_home=DATA_HOME) -X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split( - data.data, data.target, test_size=TEST_SIZE, random_state=0) + images = numpy.fromfile( + m.stdout, 'ubyte', count=buffer_size * 28 * 28).reshape( + (buffer_size, 28 * 28)).astype('float32') + images = images / 255.0 * 2.0 - 1.0 -def train_creator(): - return __mnist_reader_creator__(X_train, y_train) + for i in xrange(buffer_size): + yield images[i, :], int(labels[i]) + m.terminate() + l.terminate() -def test_creator(): - return __mnist_reader_creator__(X_test, y_test) + return reader -def unittest(): - assert len(list(test_creator()())) == TEST_SIZE +def train(): + return reader_creator( + paddle.v2.dataset.common.download(TRAIN_IMAGE_URL, 'mnist', + TRAIN_IMAGE_MD5), + paddle.v2.dataset.common.download(TRAIN_LABEL_URL, 'mnist', + TRAIN_LABEL_MD5), 100) -if __name__ == '__main__': - unittest() +def test(): + return reader_creator( + paddle.v2.dataset.common.download(TEST_IMAGE_URL, 'mnist', + TEST_IMAGE_MD5), + paddle.v2.dataset.common.download(TEST_LABEL_URL, 'mnist', + TEST_LABEL_MD5), 100) diff --git a/python/paddle/v2/dataset/movielens.py b/python/paddle/v2/dataset/movielens.py index 314329e91cadf8a74466ed9f385cd596c0ba6f9f..dcffcff2f58c63d451761d37f14127d730faf621 100644 --- a/python/paddle/v2/dataset/movielens.py +++ b/python/paddle/v2/dataset/movielens.py @@ -1,5 +1,5 @@ import zipfile -from config import download +from common import download import re import random import functools diff --git a/python/paddle/v2/dataset/tests/cifar_test.py b/python/paddle/v2/dataset/tests/cifar_test.py new file mode 100644 index 0000000000000000000000000000000000000000..a2af45ecf508462fe4b596b5d8d6401c5b974eff --- /dev/null +++ b/python/paddle/v2/dataset/tests/cifar_test.py @@ -0,0 +1,42 @@ +import paddle.v2.dataset.cifar +import unittest + + +class TestCIFAR(unittest.TestCase): + def check_reader(self, reader): + sum = 0 + label = 0 + for l in reader(): + self.assertEqual(l[0].size, 3072) + if l[1] > label: + label = l[1] + sum += 1 + return sum, label + + def test_test10(self): + instances, max_label_value = self.check_reader( + paddle.v2.dataset.cifar.test10()) + self.assertEqual(instances, 10000) + self.assertEqual(max_label_value, 9) + + def test_train10(self): + instances, max_label_value = self.check_reader( + paddle.v2.dataset.cifar.train10()) + self.assertEqual(instances, 50000) + self.assertEqual(max_label_value, 9) + + def test_test100(self): + instances, max_label_value = self.check_reader( + paddle.v2.dataset.cifar.test100()) + self.assertEqual(instances, 10000) + self.assertEqual(max_label_value, 99) + + def test_train100(self): + instances, max_label_value = self.check_reader( + paddle.v2.dataset.cifar.train100()) + self.assertEqual(instances, 50000) + self.assertEqual(max_label_value, 99) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/dataset/tests/common_test.py b/python/paddle/v2/dataset/tests/common_test.py new file mode 100644 index 0000000000000000000000000000000000000000..7d8406171b8478e4a8331637c5e867c18d5eb3d8 --- /dev/null +++ b/python/paddle/v2/dataset/tests/common_test.py @@ -0,0 +1,23 @@ +import paddle.v2.dataset.common +import unittest +import tempfile + + +class TestCommon(unittest.TestCase): + def test_md5file(self): + _, temp_path = tempfile.mkstemp() + with open(temp_path, 'w') as f: + f.write("Hello\n") + self.assertEqual('09f7e02f1290be211da707a266f153b3', + paddle.v2.dataset.common.md5file(temp_path)) + + def test_download(self): + yi_avatar = 'https://avatars0.githubusercontent.com/u/1548775?v=3&s=460' + self.assertEqual( + paddle.v2.dataset.common.DATA_HOME + '/test/1548775?v=3&s=460', + paddle.v2.dataset.common.download( + yi_avatar, 'test', 'f75287202d6622414c706c36c16f8e0d')) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/dataset/tests/mnist_test.py b/python/paddle/v2/dataset/tests/mnist_test.py new file mode 100644 index 0000000000000000000000000000000000000000..b4408cc2f590d4d8da4ce5e98213cf7b208cfc15 --- /dev/null +++ b/python/paddle/v2/dataset/tests/mnist_test.py @@ -0,0 +1,30 @@ +import paddle.v2.dataset.mnist +import unittest + + +class TestMNIST(unittest.TestCase): + def check_reader(self, reader): + sum = 0 + label = 0 + for l in reader(): + self.assertEqual(l[0].size, 784) + if l[1] > label: + label = l[1] + sum += 1 + return sum, label + + def test_train(self): + instances, max_label_value = self.check_reader( + paddle.v2.dataset.mnist.train()) + self.assertEqual(instances, 60000) + self.assertEqual(max_label_value, 9) + + def test_test(self): + instances, max_label_value = self.check_reader( + paddle.v2.dataset.mnist.test()) + self.assertEqual(instances, 10000) + self.assertEqual(max_label_value, 9) + + +if __name__ == '__main__': + unittest.main() diff --git a/python/paddle/v2/layer.py b/python/paddle/v2/layer.py index faf5b8bd87d7b8afe653a821607929589c5abc55..19f3c2f998d83c8841535a967c1d6335f13da886 100644 --- a/python/paddle/v2/layer.py +++ b/python/paddle/v2/layer.py @@ -67,6 +67,7 @@ paddle.v2.parameters.create, no longer exposed to users. """ import collections +import inspect import paddle.trainer_config_helpers as conf_helps from paddle.trainer_config_helpers.config_parser_utils import \ @@ -74,26 +75,14 @@ from paddle.trainer_config_helpers.config_parser_utils import \ from paddle.trainer_config_helpers.default_decorators import wrap_name_default from paddle.trainer_config_helpers.default_decorators import wrap_act_default -from paddle.trainer_config_helpers.default_decorators import wrap_bias_attr_default +from paddle.trainer_config_helpers.default_decorators import \ + wrap_bias_attr_default from paddle.trainer_config_helpers.layers import layer_support import data_type import activation -import attr - -__all__ = [ - 'parse_network', 'data', 'fc', 'conv_shift', 'img_conv', 'img_pool', 'spp', - 'maxout', 'img_cmrnorm', 'batch_norm', 'sum_to_one_norm', 'recurrent', - 'lstmemory', 'grumemory', 'pool', 'last_seq', 'first_seq', 'concat', - 'seq_concat', 'block_expand', 'expand', 'repeat', 'seq_reshape', 'addto', - 'linear_comb', 'interpolation', 'bilinear_interp', 'power', 'scaling', - 'slope_intercept', 'tensor', 'cos_sim', 'trans', 'max_id', 'sampling_id', - 'pad', 'classification_cost', 'cross_entropy_cost', - 'cross_entropy_with_selfnorm_cost', 'regression_cost', - 'multi_binary_label_cross_entropy_cost', 'rank_cost', 'lambda_cost', - 'sum_cost', 'huber_cost', 'crf', 'crf_decoding', 'ctc', 'warp_ctc', 'nce', - 'hsigmoid', 'eos' -] + +__all__ = ['parse_network', 'data'] __projection_names__ = filter(lambda x: x.endswith('_projection'), dir(conf_helps)) @@ -289,83 +278,51 @@ data = DataLayerV2 AggregateLevel = conf_helps.layers.AggregateLevel ExpandLevel = conf_helps.layers.ExpandLevel -layer_list = [ - # [V2LayerImpl, V1_method_name, parent_names] - # fully connected layers - ['fc', 'fc_layer', ['input']], - # conv layers - ['conv_shift', 'conv_shift_layer', ['a', 'b']], - ['img_conv', 'img_conv_layer', ['input']], - # image pooling layers - ['img_pool', 'img_pool_layer', ['input']], - ['spp', 'spp_layer', ['input']], - ['maxout', 'maxout_layer', ['input']], - # norm layers - ['img_cmrnorm', 'img_cmrnorm_layer', ['input']], - ['batch_norm', 'batch_norm_layer', ['input']], - ['sum_to_one_norm', 'sum_to_one_norm_layer', ['input']], - # recurrent layers - ['recurrent', 'recurrent_layer', ['input']], - ['lstmemory', 'lstmemory', ['input']], - ['grumemory', 'grumemory', ['input']], - # aggregate layers - ['pool', 'pooling_layer', ['input']], - ['last_seq', 'last_seq', ['input']], - ['first_seq', 'first_seq', ['input']], - ['concat', 'concat_layer', ['input']], - ['seq_concat', 'seq_concat_layer', ['a', 'b']], - # reshaping layers - ['block_expand', 'block_expand_layer', ['input']], - ['expand', 'expand_layer', ['input', 'expand_as']], - ['repeat', 'repeat_layer', ['input']], - ['rotate', 'rotate_layer', ['input']], - ['seq_reshape', 'seq_reshape_layer', ['input']], - # math layers - ['addto', 'addto_layer', ['input']], - ['linear_comb', 'linear_comb_layer', ['weights', 'vectors']], - ['interpolation', 'interpolation_layer', ['input', 'weight']], - ['bilinear_interp', 'bilinear_interp_layer', ['input']], - ['power', 'power_layer', ['input', 'weight']], - ['scaling', 'scaling_layer', ['input', 'weight']], - ['slope_intercept', 'slope_intercept_layer', ['input']], - ['tensor', 'tensor_layer', ['a', 'b']], - ['cos_sim', 'cos_sim', ['a', 'b']], - ['trans', 'trans_layer', ['input']], - # sampling layers - ['max_id', 'maxid_layer', ['input']], - ['sampling_id', 'sampling_id_layer', ['input']], - # slicing and joining layers - ['pad', 'pad_layer', ['input']], - # cost layers - [ - 'classification_cost', 'classification_cost', - ['input', 'label', 'weight'] - ], - ['regression_cost', 'regression_cost', ['input', 'label', 'weight']], - ['cross_entropy_cost', 'cross_entropy', ['input', 'label']], - [ - 'cross_entropy_with_selfnorm_cost', 'cross_entropy_with_selfnorm', - ['input', 'label'] - ], - [ - 'multi_binary_label_cross_entropy_cost', - 'multi_binary_label_cross_entropy', ['input', 'label'] - ], - ['rank_cost', 'rank_cost', ['left', 'right', 'label', 'weight']], - ['lambda_cost', 'lambda_cost', ['input', 'score']], - ['sum_cost', 'sum_cost', ['input']], - ['huber_cost', 'huber_cost', ['input', 'label']], - ['crf', 'crf_layer', ['input', 'label']], - ['crf_decoding', 'crf_decoding_layer', ['input']], - ['ctc', 'ctc_layer', ['input', 'label']], - ['warp_ctc', 'warp_ctc_layer', ['input', 'label']], - ['nce', 'nce_layer', ['input', 'label']], - ['hsigmoid', 'hsigmoid', ['input', 'label']], - # check layers - ['eos', 'eos_layer', ['input']] -] -for l in layer_list: - globals()[l[0]] = __convert_to_v2__(l[1], l[2]) + +def __layer_name_mapping__(inname): + if inname in ['data_layer', 'memory', 'mixed_layer']: + # Do Not handle these layers + return + elif inname == 'maxid_layer': + return 'max_id' + elif inname.endswith('memory') or inname.endswith( + '_seq') or inname.endswith('_sim') or inname == 'hsigmoid': + return inname + elif inname in [ + 'cross_entropy', 'multi_binary_label_cross_entropy', + 'cross_entropy_with_selfnorm' + ]: + return inname + "_cost" + elif inname.endswith('_cost'): + return inname + elif inname.endswith("_layer"): + return inname[:-len("_layer")] + + +def __layer_name_mapping_parent_names__(inname): + all_args = getattr(conf_helps, inname).argspec.args + return filter( + lambda x: x in ['input1', 'input2','label', 'input', 'a', 'b', 'expand_as', + 'weights', 'vectors', 'weight', 'score', 'left', 'right'], + all_args) + + +def __convert_layer__(_new_name_, _old_name_, _parent_names_): + global __all__ + __all__.append(_new_name_) + globals()[new_name] = __convert_to_v2__(_old_name_, _parent_names_) + + +for each_layer_name in dir(conf_helps): + new_name = __layer_name_mapping__(each_layer_name) + if new_name is not None: + parent_names = __layer_name_mapping_parent_names__(each_layer_name) + assert len(parent_names) != 0, each_layer_name + __convert_layer__(new_name, each_layer_name, parent_names) + +del parent_names +del new_name +del each_layer_name # convert projection for prj in __projection_names__: diff --git a/python/paddle/v2/tests/test_layer.py b/python/paddle/v2/tests/test_layer.py index bb0099ea2fbb78b0a05eedf23af95a02e8849015..b138ddbbe6c0a431393fef165b4eaebf7bfa81e4 100644 --- a/python/paddle/v2/tests/test_layer.py +++ b/python/paddle/v2/tests/test_layer.py @@ -11,17 +11,13 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -import difflib import unittest -import paddle.trainer_config_helpers as conf_helps import paddle.v2.activation as activation import paddle.v2.attr as attr import paddle.v2.data_type as data_type import paddle.v2.layer as layer import paddle.v2.pooling as pooling -from paddle.trainer_config_helpers.config_parser_utils import \ - parse_network_config as parse_network pixel = layer.data(name='pixel', type=data_type.dense_vector(128)) label = layer.data(name='label', type=data_type.integer_value(10)) @@ -70,7 +66,7 @@ class ImageLayerTest(unittest.TestCase): class AggregateLayerTest(unittest.TestCase): def test_aggregate_layer(self): - pool = layer.pool( + pool = layer.pooling( input=pixel, pooling_type=pooling.Avg(), agg_level=layer.AggregateLevel.EACH_SEQUENCE)