code
stringlengths
17
6.64M
def run(args, input=None): 'run subproc' args = list(args) print('run:', args) p = Popen(args, stdout=PIPE, stderr=STDOUT, stdin=PIPE, env=build_env()) (out, _) = p.communicate(input=input) print(('Return code is %i' % p.returncode)) print(('std out/err:\n---\n%s\n---\n' % out.decode('utf8'))) if (p.returncode != 0): raise CalledProcessError(cmd=args, returncode=p.returncode, output=out) return out.decode('utf8')
def filter_out(ls): '\n :param list[str] ls:\n :rtype: list[str]\n ' if (not isinstance(ls, list)): ls = list(ls) res = [] i = 0 while (i < len(ls)): s = ls[i] if (('tensorflow/core/' in s) or ('tensorflow/stream_executor/' in s)): i += 1 continue if (((i + 1) < len(ls)) and ls[(i + 1)].startswith(' ')): if (re.match('.*:\\d+: RuntimeWarning: numpy.*', s) or re.match('.*:\\d+: FutureWarning: .*', s)): i += 2 continue if any(((msg in s) for msg in ['Setup TF inter and intra global thread pools', 'Collecting TensorFlow device list', 'CUDA_VISIBLE_DEVICES is not set'])): i += 1 continue if ('Local devices available to TensorFlow:' in s): i += 1 while ((i < len(ls)) and re.match('^ {2}\\d+/\\d+: name:.*', ls[i])): i += 1 while ((i < len(ls)) and re.match('^ {4}.*', ls[i])): i += 1 continue if (s.startswith('systemMemory:') or s.startswith('maxCacheSize:')): i += 1 continue res.append(ls[i]) i += 1 return res
def count_start_with(ls, s): '\n :param list[str] ls:\n :param str s:\n :rtype: int\n ' c = 0 for l in ls: if l.startswith(s): c += 1 return c
def test_filter_out(): s = '\n/home/travis/virtualenv/python2.7.14/lib/python2.7/site-packages/scipy/special/__init__.py:640: RuntimeWarning: numpy.dtype size changed, may indicate binary incompatibility. Expected 96, got 88\n from ._ufuncs import *\n/home/travis/virtualenv/python2.7.14/lib/python2.7/site-packages/h5py/_hl/group.py:22: RuntimeWarning: numpy.dtype size changed, may indicate binary incompatibility. Expected 96, got 88\n from .. import h5g, h5i, h5o, h5r, h5t, h5l, h5p\nRETURNN starting up, version 20180724.141845--git-7865d01, date/time 2018-07-24-13-11-47 (UTC+0000), pid 2196, cwd /home/travis/build/rwth-i6/returnn, Python /home/travis/virtualenv/python2.7.14/bin/python\nfaulthandler import error. No module named faulthandler\nTheano: 0.9.0 (<site-package> in /home/travis/virtualenv/python2.7.14/lib/python2.7/site-packages/theano)\nTask: No-operation\nelapsed: 0:00:00.0001\n' ls = filter(None, s.splitlines()) ls = filter_out(ls) pprint(ls) assert_equal(len(ls), 5)
def test_returnn_startup(): out = run([py, __main_entry__, '-x', 'nop', '++use_tensorflow', '1']) ls = out.splitlines() ls = filter_out(ls) if (not (3 <= len(ls) <= 10)): print(('output:\n%s\n\nNum lines: %i' % ('\n'.join(ls), len(ls)))) raise Exception('unexpected output number of lines') assert_equal(count_start_with(ls, 'RETURNN starting up, version '), 1) assert_equal(count_start_with(ls, 'TensorFlow: '), 1) assert_in('Task: No-operation', ls)
def test_returnn_startup_verbose(): out = run([py, __main_entry__, '-x', 'nop', '++use_tensorflow', '1', '++log_verbosity', '5']) ls = out.splitlines() ls = filter_out(ls) if (not (3 <= len(ls) <= 15)): print(('output:\n%s\n\nNum lines: %i' % ('\n'.join(ls), len(ls)))) raise Exception('unexpected output number of lines') assert_equal(count_start_with(ls, 'RETURNN starting up, version '), 1) assert_equal(count_start_with(ls, 'RETURNN command line options: '), 1) assert_equal(count_start_with(ls, 'TensorFlow: '), 1) assert_in('Task: No-operation', ls) assert_in('Quitting', ls)
def test_simple_log(): code = '\nfrom __future__ import annotations\nprint("hello stdout 1")\nfrom returnn.log import log\nlog.initialize(verbosity=[], logs=[], formatter=[])\nprint("hello stdout 2")\nprint("hello log 1", file=log.v3)\nprint("hello log 2", file=log.v3)\n ' out = run([py], input=code.encode('utf8')) assert_equal(out.splitlines(), ['hello stdout 1', 'hello stdout 2', 'hello log 1', 'hello log 2'])
def test_StreamIO(): import io buf = io.StringIO() assert_equal(buf.getvalue(), '') print(('buf: %r' % buf.getvalue())) buf.write('hello') print(('buf: %r' % buf.getvalue())) assert_equal(buf.getvalue(), 'hello') buf.truncate(0) print(('buf: %r' % buf.getvalue())) assert_equal(buf.getvalue(), '') buf.write('hello') print(('buf: %r' % buf.getvalue())) if PY3: assert_equal(buf.getvalue(), '\x00\x00\x00\x00\x00hello') buf.truncate(0) buf.seek(0) print(('buf: %r' % buf.getvalue())) assert_equal(buf.getvalue(), '') buf.write('hello') print(('buf: %r' % buf.getvalue())) assert_equal(buf.getvalue(), 'hello') buf.truncate(0) buf.seek(0) print(('buf: %r' % buf.getvalue())) assert_equal(buf.getvalue(), '')
def _sig_alarm_handler(signum, frame): raise Exception(f'Alarm (timeout) signal handler')
@contextlib.contextmanager def timeout(seconds=10): '\n :param seconds: when the context is not closed within this time, an exception will be raised\n ' signal.alarm(seconds) try: (yield) finally: signal.alarm(0)
def test_MultiProcDataset_n1_b1_default(): hdf_fn = generate_hdf_from_other({'class': 'Task12AXDataset', 'num_seqs': 23}) hdf_dataset_dict = {'class': 'HDFDataset', 'files': [hdf_fn]} hdf_dataset = init_dataset(hdf_dataset_dict) hdf_dataset_seqs = dummy_iter_dataset(hdf_dataset) with timeout(): mp_dataset = MultiProcDataset(dataset=hdf_dataset_dict, num_workers=1, buffer_size=1) mp_dataset.initialize() mp_dataset_seqs = dummy_iter_dataset(mp_dataset) compare_dataset_seqs(hdf_dataset_seqs, mp_dataset_seqs)
def test_MultiProcDataset_n3_b5_shuffle(): hdf_fn = generate_hdf_from_other({'class': 'Task12AXDataset', 'num_seqs': 23}) hdf_dataset_dict = {'class': 'HDFDataset', 'files': [hdf_fn], 'seq_ordering': 'random'} hdf_dataset = init_dataset(hdf_dataset_dict) hdf_dataset_seqs = dummy_iter_dataset(hdf_dataset) with timeout(): mp_dataset = MultiProcDataset(dataset=hdf_dataset_dict, num_workers=3, buffer_size=5) mp_dataset.initialize() mp_dataset_seqs = dummy_iter_dataset(mp_dataset) compare_dataset_seqs(hdf_dataset_seqs, mp_dataset_seqs)
def test_MultiProcDataset_meta(): hdf_fn = generate_hdf_from_other({'class': 'Task12AXDataset', 'num_seqs': 23}) meta_dataset_dict = {'class': 'MetaDataset', 'data_map': {'classes': ('hdf', 'classes'), 'data': ('hdf', 'data')}, 'datasets': {'hdf': {'class': 'HDFDataset', 'files': [hdf_fn]}}} meta_dataset = init_dataset(meta_dataset_dict) meta_dataset_seqs = dummy_iter_dataset(meta_dataset) with timeout(): mp_dataset = MultiProcDataset(dataset=meta_dataset_dict, num_workers=1, buffer_size=1) mp_dataset.initialize() mp_dataset_seqs = dummy_iter_dataset(mp_dataset) compare_dataset_seqs(meta_dataset_seqs, mp_dataset_seqs)
def test_MultiProcDataset_via_config(): from io import StringIO import textwrap from returnn.config import Config, global_config_ctx config = Config() config.load_file(StringIO(textwrap.dedent(' #!returnn.py\n\n import numpy\n from returnn.datasets.map import MapDatasetBase\n\n class MyCustomMapDatasetInConfig(MapDatasetBase):\n def __init__(self):\n super().__init__(data_types={"data": {"shape": (None, 3)}})\n\n def __len__(self):\n return 2\n\n def __getitem__(self, item):\n return {"data": numpy.zeros((5, 3))}\n '))) with timeout(), global_config_ctx(config): mp_dataset = MultiProcDataset(dataset={'class': 'MapDatasetWrapper', 'map_dataset': config.typed_dict['MyCustomMapDatasetInConfig']}, num_workers=1, buffer_size=1) mp_dataset.initialize() items = dummy_iter_dataset(mp_dataset) assert (len(items) == 2)
class _MyCustomMapDatasetException(Exception): pass
class _MyCustomMapDatasetThrowingExceptionAtInit(MapDatasetBase): def __init__(self): super().__init__() raise _MyCustomMapDatasetException('test exception at init')
class _MyCustomMapDatasetThrowingExceptionAtItem(MapDatasetBase): def __init__(self): super().__init__(data_types={'data': {'shape': (None, 3)}}) def __len__(self): return 2 def __getitem__(self, item): if (item == 0): return {'data': numpy.zeros((5, 3))} raise _MyCustomMapDatasetException('test exception at getitem')
def test_MultiProcDataset_exception_at_init(): with timeout(): mp_dataset = MultiProcDataset(dataset={'class': 'MapDatasetWrapper', 'map_dataset': _MyCustomMapDatasetThrowingExceptionAtInit}, num_workers=1, buffer_size=1) try: mp_dataset.initialize() except Exception as exc: print('Got expected exception:', exc) else: raise Exception('Expected exception')
def test_MultiProcDataset_exception_at_item(): with timeout(): mp_dataset = MultiProcDataset(dataset={'class': 'MapDatasetWrapper', 'map_dataset': _MyCustomMapDatasetThrowingExceptionAtItem}, num_workers=1, buffer_size=1) mp_dataset.initialize() try: dummy_iter_dataset(mp_dataset) except Exception as exc: print('Got expected exception:', exc) else: raise Exception('Expected exception')
class _MyCustomDummyMapDataset(MapDatasetBase): def __init__(self): super().__init__(data_types={'data': {'shape': (None, 3)}}) def __len__(self): return 2 def __getitem__(self, item): return {'data': numpy.zeros((((item * 2) + 5), 3))}
def test_MultiProcDataset_pickle(): import pickle with timeout(): mp_dataset = MultiProcDataset(dataset={'class': 'MapDatasetWrapper', 'map_dataset': _MyCustomDummyMapDataset}, num_workers=1, buffer_size=1) mp_dataset.initialize() mp_dataset_seqs = dummy_iter_dataset(mp_dataset) mp_dataset_serialized = pickle.dumps(mp_dataset) with timeout(): mp_dataset_ = pickle.loads(mp_dataset_serialized) mp_dataset_seqs_ = dummy_iter_dataset(mp_dataset_) compare_dataset_seqs(mp_dataset_seqs, mp_dataset_seqs_)
def test_config_net_dict1(): config = Config() config.update(config_dict) config.typed_dict['network'] = net_dict pretrain = pretrain_from_config(config) assert_equal(pretrain.get_train_num_epochs(), 2) net1_json = pretrain.get_network_json_for_epoch(1) net2_json = pretrain.get_network_json_for_epoch(2) net3_json = pretrain.get_network_json_for_epoch(3) assert_in('hidden_0', net1_json) assert_not_in('hidden_1', net1_json) assert_in('hidden_0', net2_json) assert_in('hidden_1', net2_json) assert_equal(net2_json, net3_json)
def test_config_net_dict2(): config = Config() config.update(config_dict) config.typed_dict['network'] = net_dict2 pretrain = pretrain_from_config(config) assert_equal(pretrain.get_train_num_epochs(), 3)
@contextlib.contextmanager def make_scope(): with tf.Graph().as_default() as graph: with tf_compat.v1.Session(graph=graph) as session: (yield session)
def build_resnet(conv_time_dim): dropout = 0 L2 = 0.1 filter_size = (3, 3) context_window = 1 window = 1 feature_dim = 64 channel_num = 3 num_inputs = ((feature_dim * channel_num) * window) num_outputs = 9001 EpochSplit = 6 cur_feat_dim = feature_dim global _last, network network = {} _last = 'data' def add_sequential_layer(name, d, from_=None): global _last, network assert ('from' not in d) if (from_ is not None): d['from'] = from_ else: d['from'] = [_last] assert (name not in network) network[name] = d _last = name def fixed_padding(prefix, kernel_size, data_format, conv_time_dim): 'Pads the input along the spatial dimensions independently of input size.' pad_total = (kernel_size - 1) feature_pad_beg = (pad_total // 2) feature_pad_end = (pad_total - feature_pad_beg) time_pad_beg = 0 time_pad_end = 0 return add_sequential_layer(('%s_pad' % prefix), {'class': 'pad', 'axes': ('s:0', 's:1'), 'padding': [(time_pad_beg, time_pad_end), (feature_pad_end, feature_pad_end)]}) def conv2d_fixed_padding(prefix, filters, kernel_size, strides, dilation_rate, data_format, conv_time_dim, source=None): 'Strided 2-D convolution with explicit padding.' fixed_padding(('%s_pad' % prefix), kernel_size, data_format, conv_time_dim) padding = 'VALID' strides = ((1, strides) if conv_time_dim else strides) filter_size = (kernel_size, kernel_size) dilation_rate = ((dilation_rate, 1) if conv_time_dim else (1, 1)) if (data_format == 'channels_first'): NCHW = True else: NCHW = False add_sequential_layer(('%s_conv' % prefix), {'class': 'conv', 'n_out': filters, 'filter_size': filter_size, 'auto_use_channel_first': NCHW, 'strides': strides, 'dilation_rate': dilation_rate, 'padding': padding, 'activation': None, 'with_bias': False, 'dropout': 0, 'forward_weights_init': 'xavier', 'L2': L2}, from_=source) return ('%s_conv' % prefix) def _building_block_v2(prefix, filters, projection_shortcut, strides, dilation_rate, dilation_rate_multiplier, kernel_size, data_format, conv_time_dim): 'A single block for ResNet v2, without a bottleneck.\n\n Batch normalization then ReLu then convolution as described by:\n Identity Mappings in Deep Residual Networks\n https://arxiv.org/pdf/1603.05027.pdf\n by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, Jul 2016.\n ' add_sequential_layer(('%s_in' % prefix), {'class': 'copy'}) add_sequential_layer(('%s_relu1' % prefix), {'class': 'activation', 'activation': 'relu', 'batch_norm': False}) if conv_time_dim: conv2d_fixed_padding(prefix=('%s_conv_1' % prefix), filters=filters, kernel_size=kernel_size, strides=1, dilation_rate=dilation_rate, data_format=data_format, conv_time_dim=conv_time_dim) add_sequential_layer(('%s_stride' % prefix), {'class': 'slice', 'axis': 's:1', 'slice_step': strides}) dilation_rate *= dilation_rate_multiplier else: conv2d_fixed_padding(prefix=('%s_conv_1' % prefix), filters=filters, kernel_size=kernel_size, strides=strides, dilation_rate=dilation_rate, data_format=data_format, conv_time_dim=conv_time_dim) add_sequential_layer(('%s_relu2' % prefix), {'class': 'activation', 'activation': 'relu', 'batch_norm': False}) conv = conv2d_fixed_padding(prefix=('%s_conv_2' % prefix), filters=filters, kernel_size=kernel_size, strides=1, dilation_rate=dilation_rate, data_format=data_format, conv_time_dim=conv_time_dim) result = ('%s_conv_2' % prefix) crop_lr = (filter_size[0] - 1) crop_left = (crop_lr // 2) crop_right = (crop_lr - crop_left) if conv_time_dim: if (dilation_rate_multiplier > 1): crop = int((crop_left * ((dilation_rate / dilation_rate_multiplier) + dilation_rate))) else: crop = int(((crop_left * 2) * dilation_rate)) add_sequential_layer(('%s_crop' % prefix), {'class': 'slice', 'axis': 'T', 'slice_start': crop, 'slice_end': (- crop)}, from_=('%s_relu1' % prefix)) shortcut = ('%s_crop' % prefix) if (projection_shortcut is not None): shortcut = projection_shortcut(source=shortcut) else: crop = crop_left add_sequential_layer(('%s_crop_1' % prefix), {'class': 'slice', 'axis': 'T', 'slice_start': crop, 'slice_end': (- crop)}, from_=('%s_relu1' % prefix)) shortcut = ('%s_crop_1' % prefix) if (projection_shortcut is not None): shortcut = projection_shortcut(source=shortcut) add_sequential_layer(('%s_crop_2' % prefix), {'class': 'slice', 'axis': 'T', 'slice_start': crop, 'slice_end': (- crop)}, from_=shortcut) shortcut = ('%s_crop_2' % prefix) add_sequential_layer(('%s_out' % prefix), {'class': 'combine', 'kind': 'add'}, from_=[conv, shortcut]) return def block_layer(prefix, filters, bottleneck, block_fn, blocks, strides, dilation_rate, dilation_rate_multiplier, kernel_size, data_format, conv_time_dim): 'Creates one layer of blocks for the ResNet model.' filters_out = ((filters * 4) if bottleneck else filters) if (not conv_time_dim): strides = (dilation_rate_multiplier, strides) def projection_shortcut(source=None): return conv2d_fixed_padding(prefix=('%s_sc' % prefix), filters=filters_out, kernel_size=1, strides=strides, dilation_rate=1, data_format=data_format, conv_time_dim=conv_time_dim, source=source) block_fn(('%s_0' % prefix), filters, projection_shortcut, strides, dilation_rate, dilation_rate_multiplier, kernel_size, data_format, conv_time_dim) dilation_rate *= dilation_rate_multiplier for i in range(1, blocks): block_fn(('%s_%i' % (prefix, i)), filters, None, 1, dilation_rate, 1, kernel_size, data_format, conv_time_dim) return add_sequential_layer(('%s_out' % prefix), {'class': 'copy'}) resnet_version = 2 conv_time_dim = conv_time_dim bottleneck = False num_filters = 64 first_kernel_size = 5 kernel_size = 3 conv_stride = (2 if conv_time_dim else (1, 2)) first_pool_size = (1, 2) first_pool_stride = (1, 1) last_pool_size = (2, 2) last_pool_stride = ((1, 2) if conv_time_dim else (2, 2)) block_sizes = [2, 2, 2, 2] block_strides = [1, 2, 2, 2] block_dilations = [1, 1, 1, 2] block_fn = _building_block_v2 data_format = 'channels_first' pre_activation = (resnet_version == 2) if (data_format == 'channels_first'): NCHW = True else: NCHW = False if conv_time_dim: multiplier = (1 if bottleneck else 2) building_block_reduction = ((multiplier * 2) * (kernel_size // 2)) total_reduction = (first_kernel_size - 1) dilation_rate_multiplier = 1 total_reduction += (dilation_rate_multiplier * (first_pool_size[0] - 1)) for (i, bs) in enumerate(block_sizes): total_reduction += ((building_block_reduction / multiplier) * dilation_rate_multiplier) dilation_rate_multiplier *= block_dilations[i] total_reduction += ((building_block_reduction / multiplier) * dilation_rate_multiplier) total_reduction += ((building_block_reduction * (bs - 1)) * dilation_rate_multiplier) total_reduction += (dilation_rate_multiplier * (last_pool_size[0] - 1)) dilation_rate_multiplier *= 2 print(total_reduction, dilation_rate_multiplier) total_reduction += (dilation_rate_multiplier * 2) print(total_reduction, dilation_rate_multiplier) time_dim_reduction = total_reduction context_window = int(((2 * (total_reduction // 2)) + 1)) else: time_dim_reduction = 0 print('time_dim_reduction: ', time_dim_reduction) print('context_window: ', context_window) conv2d_fixed_padding(prefix='c_init', filters=num_filters, kernel_size=first_kernel_size, strides=conv_stride, dilation_rate=1, data_format=data_format, conv_time_dim=conv_time_dim) dilation_rate = 1 if (resnet_version == 1): add_sequential_layer('c_init_relu', {'class': 'activation', 'activation': 'relu', 'batch_norm': False}) if first_pool_size: if conv_time_dim: dr = (dilation_rate, 1) else: dr = (1, 1) pad_total = (first_pool_size[1] - 1) pad_beg = (pad_total // 2) pad_end = (pad_total - pad_beg) add_sequential_layer('c_init_pool_pad', {'class': 'pad', 'axes': 's:1', 'padding': (pad_beg, pad_end)}) add_sequential_layer('c_init_pool', {'class': 'pool', 'mode': 'max', 'padding': 'VALID', 'pool_size': first_pool_size, 'strides': first_pool_stride, 'dilation_rate': dr, 'use_channel_first': NCHW}) print('dr: ', dilation_rate) for (i, num_blocks) in enumerate(block_sizes): filters = (num_filters * (2 ** i)) block_layer(prefix=('c_%i' % i), filters=filters, bottleneck=bottleneck, block_fn=block_fn, blocks=num_blocks, strides=block_strides[i], dilation_rate=dilation_rate, dilation_rate_multiplier=block_dilations[i], kernel_size=kernel_size, data_format=data_format, conv_time_dim=conv_time_dim) dilation_rate *= block_dilations[i] print('dr: ', dilation_rate) if pre_activation: add_sequential_layer('c_out_relu', {'class': 'activation', 'activation': 'relu', 'batch_norm': False}) if last_pool_size: pad_total = (last_pool_size[1] - 1) pad_beg = (pad_total // 2) pad_end = (pad_total - pad_beg) add_sequential_layer('c_last_pool_pad', {'class': 'pad', 'axes': 's:1', 'padding': (pad_beg, pad_end)}) if conv_time_dim: dr = (dilation_rate, 1) add_sequential_layer('c_last_pool', {'class': 'pool', 'mode': 'max', 'padding': 'VALID', 'pool_size': last_pool_size, 'strides': (1, 1), 'dilation_rate': dr, 'use_channel_first': NCHW}) add_sequential_layer('c_last_stride', {'class': 'slice', 'axis': 's:1', 'slice_step': last_pool_stride[1]}) dilation_rate *= 2 else: dr = (1, 1) add_sequential_layer('c_last_pool', {'class': 'pool', 'mode': 'max', 'padding': 'VALID', 'pool_size': last_pool_size, 'strides': last_pool_stride, 'dilation_rate': dr, 'use_channel_first': NCHW}) if conv_time_dim: dr = (dilation_rate, 1) else: dr = (1, 1) '\n See https://arxiv.org/pdf/1611.09288.pdf\n Fully connected layers are equivalent to, and can be trivially replaced by,\n convolutional layers with kernel (1×1) (except the first convolution which\n has kernel size matching the output of the conv stack before being flattened\n for the fully connected layers).\n ' add_sequential_layer('fc1', {'class': 'conv', 'n_out': 2048, 'filter_size': (3, 2), 'auto_use_channel_first': NCHW, 'strides': (1, 1), 'dilation_rate': dr, 'padding': 'VALID', 'activation': None, 'with_bias': False, 'dropout': 0, 'forward_weights_init': 'xavier', 'L2': L2}) add_sequential_layer('fc2', {'class': 'conv', 'n_out': 2048, 'filter_size': (1, 1), 'auto_use_channel_first': NCHW, 'strides': (1, 1), 'dilation_rate': (1, 1), 'padding': 'VALID', 'activation': None, 'with_bias': False, 'dropout': 0, 'forward_weights_init': 'xavier', 'L2': L2}) add_sequential_layer('fc3', {'class': 'conv', 'n_out': 2048, 'filter_size': (1, 1), 'auto_use_channel_first': NCHW, 'strides': (1, 1), 'dilation_rate': (1, 1), 'padding': 'VALID', 'activation': None, 'with_bias': False, 'dropout': 0, 'forward_weights_init': 'xavier', 'L2': L2}) add_sequential_layer('fc4', {'class': 'conv', 'n_out': 1024, 'filter_size': (1, 1), 'auto_use_channel_first': NCHW, 'strides': (1, 1), 'dilation_rate': (1, 1), 'padding': 'VALID', 'activation': None, 'with_bias': False, 'dropout': 0, 'forward_weights_init': 'xavier', 'L2': L2}) add_sequential_layer('fc5', {'class': 'conv', 'n_out': num_outputs, 'filter_size': (1, 1), 'auto_use_channel_first': NCHW, 'strides': (1, 1), 'dilation_rate': (1, 1), 'padding': 'VALID', 'activation': None, 'with_bias': False, 'dropout': 0, 'forward_weights_init': 'xavier', 'L2': L2}) add_sequential_layer('merge', {'class': 'merge_dims', 'axes': ('s:0', 's:1')}) add_sequential_layer('swap', {'class': 'swap_axes', 'axis1': 's:0', 'axis2': 'f'}) add_sequential_layer('output', {'class': 'activation', 'activation': 'softmax', 'loss': 'ce'}) return (network, context_window)
def test_ResNet(): 'Test to compare Resnet convolving (window x frequency) vs (time x frequency).\n Batch_norm layers are turned off in oder to compare, since the statistics over the\n windowed input data is a bit different from the plain input (when convolving directing\n over the time dim).\n ' def sliding_window(seq, window_size): import numpy as np import copy it = iter(seq) win = [it.__next__() for cnt in range(window_size)] res_arr = [] res_arr.append(copy.deepcopy(win)) for e in it: win[:(- 1)] = win[1:] win[(- 1)] = e res_arr.append(copy.deepcopy(win)) return np.array(res_arr) with make_scope() as session: import numpy as np import math from tensorflow.python.client import timeline (net_dict_conv_td, window_size) = build_resnet(conv_time_dim=True) (net_dict_windowed, _) = build_resnet(conv_time_dim=False) time_size = (window_size + 1) data_layer_win = Data(name='win', shape=(window_size, 64, 3), dim=3, batch_dim_axis=0, sparse=False) data_layer_win.placeholder = tf_compat.v1.placeholder(shape=(None, window_size, 64, 3), dtype=tf.float32) data_layer_nowin = Data(name='nowin', shape=(time_size, 64, 3), dim=3, batch_dim_axis=0, time_dim_axis=1, sparse=False) data_layer_nowin.placeholder = tf_compat.v1.placeholder(shape=(None, time_size, 64, 3), dtype=tf.float32) extern_data_nowin = ExternData() extern_data_nowin.data['data'] = data_layer_nowin extern_data_win = ExternData() extern_data_win.data['data'] = data_layer_win net_conv_td = TFNetwork(extern_data=extern_data_nowin) net_conv_td.train_flag = True net_conv_td.construct_from_dict(net_dict_conv_td) net_conv_td.initialize_params(session) net_windowed = TFNetwork(extern_data=extern_data_win) net_windowed.train_flag = True net_windowed.construct_from_dict(net_dict_windowed) net_windowed.initialize_params(session) data = np.random.rand(time_size, 64, 3) data_win = sliding_window(data, window_size) data = np.array([data]) feed_dict = {data_layer_nowin.placeholder: data, data_layer_win.placeholder: data_win} (res1, res2) = session.run([net_conv_td.layers['output'].output.placeholder, net_windowed.layers['output'].output.placeholder], feed_dict=feed_dict) print((res1[0][0] - res2[0][0])) print((res1[0][1] - res2[1][0])) assert math.isclose(np.sum((res1[0][0] - res2[0][0])), 0.0, abs_tol=1e-07) assert math.isclose(np.sum((res1[0][1] - res2[1][0])), 0.0, abs_tol=1e-07)
def generate_batch(seq_idx, dataset): batch = Batch() batch.add_frames(seq_idx=seq_idx, seq_start_frame=0, length=dataset.get_seq_length(seq_idx)) return batch
def test_read_all(): config = Config() config.update(dummyconfig_dict) print('Create ExternSprintDataset') python_exec = util.which('python') if (python_exec is None): raise unittest.SkipTest('python not found') num_seqs = 4 dataset = ExternSprintDataset([python_exec, sprintExecPath], ('--*.feature-dimension=2 --*.trainer-output-dimension=3 --*.crnn-dataset=DummyDataset(2,3,num_seqs=%i,seq_len=10)' % num_seqs)) dataset.init_seq_order(epoch=1) seq_idx = 0 while dataset.is_less_than_num_seqs(seq_idx): dataset.load_seqs(seq_idx, (seq_idx + 1)) for key in dataset.get_data_keys(): value = dataset.get_data(seq_idx, key) print(('seq idx %i, data %r: %r' % (seq_idx, key, value))) seq_idx += 1 assert (seq_idx == num_seqs)
def test_assign_dev_data(): config = Config() config.update(dummyconfig_dict) print('Create ExternSprintDataset') dataset = ExternSprintDataset([sys.executable, sprintExecPath], '--*.feature-dimension=2 --*.trainer-output-dimension=3 --*.crnn-dataset=DummyDataset(2,3,num_seqs=4,seq_len=10)') dataset.init_seq_order(epoch=1) assert_true(dataset.is_less_than_num_seqs(0)) recurrent = False batch_generator = dataset.generate_batches(recurrent_net=recurrent, batch_size=5) batches = batch_generator.peek_next_n(2) assert_equal(len(batches), 2)
def test_window(): input_dim = 2 output_dim = 3 num_seqs = 2 seq_len = 5 window = 3 dataset_kwargs = dict(sprintTrainerExecPath=[sys.executable, sprintExecPath], sprintConfigStr=' '.join([('--*.feature-dimension=%i' % input_dim), ('--*.trainer-output-dimension=%i' % output_dim), ('--*.crnn-dataset=DummyDataset(input_dim=%i,output_dim=%i,num_seqs=%i,seq_len=%i)' % (input_dim, output_dim, num_seqs, seq_len))])) dataset1 = ExternSprintDataset(**dataset_kwargs) dataset2 = ExternSprintDataset(window=window, **dataset_kwargs) try: dataset1.init_seq_order(epoch=1) dataset2.init_seq_order(epoch=1) dataset1.load_seqs(0, 1) dataset2.load_seqs(0, 1) assert_equal(dataset1.get_data_dim('data'), input_dim) assert_equal(dataset2.get_data_dim('data'), (input_dim * window)) data1 = dataset1.get_data(0, 'data') data2 = dataset2.get_data(0, 'data') assert_equal(data1.shape, (seq_len, input_dim)) assert_equal(data2.shape, (seq_len, (window * input_dim))) data2a = data2.reshape(seq_len, window, input_dim) print('data1:') print(data1) print('data2:') print(data2) print('data1[0]:') print(data1[0]) print('data2[0]:') print(data2[0]) print('data2a[0,0]:') print(data2a[(0, 0)]) assert_equal(list(data2a[(0, 0)]), ([0] * input_dim)) assert_equal(list(data2a[(0, 1)]), list(data1[0])) assert_equal(list(data2a[(0, 2)]), list(data1[1])) assert_equal(list(data2a[(1, 0)]), list(data1[0])) assert_equal(list(data2a[(1, 1)]), list(data1[1])) assert_equal(list(data2a[(1, 2)]), list(data1[2])) assert_equal(list(data2a[((- 1), 2)]), ([0] * input_dim)) finally: dataset1._exit_handler() dataset2._exit_handler()
def install_sigint_handler(): import signal def signal_handler(signal, frame): print('\nSIGINT at:') better_exchook.print_tb(tb=frame, file=sys.stdout) print('') if (getattr(sys, 'exited_frame', None) is not None): print('interrupt_main via:') better_exchook.print_tb(tb=sys.exited_frame, file=sys.stdout) print('') sys.exited_frame = None raise Exception('Got SIGINT!') else: print('\nno sys.exited_frame\n') if old_action: old_action() else: raise KeyboardInterrupt old_action = signal.signal(signal.SIGINT, signal_handler)
def test_forward(): tmpdir = mkdtemp('returnn-test-sprint') olddir = os.getcwd() os.chdir(tmpdir) from returnn.datasets.generating import DummyDataset seq_len = 5 n_data_dim = 2 n_classes_dim = 3 train_data = DummyDataset(input_dim=n_data_dim, output_dim=n_classes_dim, num_seqs=4, seq_len=seq_len) train_data.init_seq_order(epoch=1) cv_data = DummyDataset(input_dim=n_data_dim, output_dim=n_classes_dim, num_seqs=2, seq_len=seq_len) cv_data.init_seq_order(epoch=1) config = '\n'.join(['#!rnn.py', 'use_tensorflow = True', ("model = '%s/model'" % tmpdir), ('num_outputs = %i' % n_classes_dim), ('num_inputs = %i' % n_data_dim), 'network = {"output": {"class": "softmax", "loss": "ce", "from": "data:data"}}', 'num_epochs = 2']) open('config_write', 'w').write(config) open('config', 'w').write((config + ("\nload= '%s/model'" % tmpdir))) config = Config() config.load_file('config_write') engine = Engine(config=config) engine.init_train_from_config(config=config, train_data=train_data, dev_data=cv_data, eval_data=None) engine.epoch = 1 engine.save_model(engine.get_epoch_model_filename()) Engine._epoch_model = None from returnn.util.basic import BackendEngine BackendEngine.selected_engine = None inputDim = 2 outputDim = 3 SprintAPI.init(inputDim=inputDim, outputDim=outputDim, config='action:forward,configfile:config,epoch:1', targetMode='forward-only') assert isinstance(SprintAPI.engine, Engine) print('used data keys via net:', SprintAPI.engine.network.get_used_data_keys()) features = numpy.array([[0.1, 0.2], [0.2, 0.3], [0.3, 0.4], [0.4, 0.5]]) seq_len = features.shape[0] posteriors = SprintAPI._forward('segment1', features.T).T assert_equal(posteriors.shape, (seq_len, outputDim)) SprintAPI.exit() os.chdir(olddir) shutil.rmtree(tmpdir)
@contextlib.contextmanager def make_scope(): with tf.Graph().as_default() as graph: with tf_compat.v1.Session(graph=graph) as session: (yield session)
class DummyLoss(Loss): need_target = False def get_value(self): assert (self.layer and isinstance(self.layer, DummyLayer)) return self.layer._get_loss_value() def get_error(self): return None
class DummyLayer(LayerBase): def __init__(self, initial_value=0.0, loss_value_factor=1.0, **kwargs): super(DummyLayer, self).__init__(**kwargs) self.loss_value_factor = loss_value_factor self.x = self.add_param(tf.Variable(initial_value)) self.output.placeholder = self.x def _get_loss_value(self): return (self.loss_value_factor * self.x) @classmethod def get_losses(cls, name, network, output, loss=None, reduce_func=None, layer=None, **kwargs): assert (not loss) loss = DummyLoss(base_network=network) return super(DummyLayer, cls).get_losses(name=name, network=network, output=output, loss=loss, reduce_func=reduce_func, layer=layer, **kwargs) @classmethod def get_out_data_from_opts(cls, name, **kwargs): from returnn.tf.util.basic import Data return Data(name=('%s_output' % name), batch_dim_axis=None, shape=(), dtype='float32')
def test_Updater_GradientDescent(): with make_scope() as session: from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config config = Config() network = TFNetwork(extern_data=ExternData(), train_flag=True) network.add_layer(name='output', layer_class=DummyLayer, initial_value=5.0, loss_value_factor=3.0) network.initialize_params(session=session) updater = Updater(config=config, network=network) updater.set_learning_rate(1.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) session.run(updater.get_optim_op()) assert_almost_equal(session.run(network.get_default_output_layer().output.placeholder), 2.0)
def test_Updater_CustomUpdate(): with make_scope() as session: from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config from returnn.tf.util.basic import CustomUpdate config = Config() network = TFNetwork(extern_data=ExternData(), train_flag=True) layer = network.add_layer(name='output', layer_class=DummyLayer, initial_value=4.0) assert isinstance(layer, DummyLayer) network.initialize_params(session=session) class CustomUpdateAdd13(CustomUpdate): def update_var(self, var): return tf_compat.v1.assign_add(var, 13.0) CustomUpdateAdd13().set_on_var(layer.x) updater = Updater(config=config, network=network) updater.set_learning_rate(1000.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) session.run(updater.get_optim_op()) assert_almost_equal(session.run(network.get_default_output_layer().output.placeholder), 17.0)
def test_add_check_numerics_ops(): with make_scope() as session: x = tf.constant(3.0, name='x') y = tf_compat.v1.log((x * 3), name='y') assert isinstance(y, tf.Tensor) assert_almost_equal(session.run(y), numpy.log(9.0)) check = add_check_numerics_ops([y]) session.run(check) z1 = tf_compat.v1.log((x - 3), name='z1') assert_equal(str(session.run(z1)), '-inf') z2 = tf_compat.v1.log((x - 4), name='z2') assert_equal(str(session.run(z2)), 'nan') check1 = add_check_numerics_ops([z1]) try: session.run(check1) except tf.errors.InvalidArgumentError as exc: print(('Expected exception: %r' % exc)) else: assert False, 'should have raised an exception' check2 = add_check_numerics_ops([z2]) try: session.run(check2) except tf.errors.InvalidArgumentError as exc: print(('Expected exception: %r' % exc)) else: assert False, 'should have raised an exception'
def test_grad_add_check_numerics_ops(): with make_scope() as session: x = tf.Variable(initial_value=0.0, name='x') session.run(x.initializer) y = (1.0 / x) grad_x = tf.gradients(y, x)[0] print('grad_x:', grad_x.eval()) assert_equal(str(float('-inf')), '-inf') assert_equal(str(grad_x.eval()), '-inf') session.run(x.assign(1.0)) opt = tf_compat.v1.train.GradientDescentOptimizer(learning_rate=1.0) train_op = opt.minimize(y, var_list=[x]) check = add_check_numerics_ops([train_op]) session.run(check) session.run(x.assign(0.0)) try: session.run(check) except tf.errors.InvalidArgumentError as exc: print(('Expected exception: %r' % exc)) else: assert False, 'should have raised an exception'
def test_Updater_add_check_numerics_ops(): class _Layer(DummyLayer): def _get_loss_value(self): return tf_compat.v1.log(self.x) from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config with make_scope() as session: config = Config() config.set('debug_add_check_numerics_ops', True) network = TFNetwork(extern_data=ExternData(), train_flag=True) network.add_layer(name='output', layer_class=_Layer, initial_value=1.0) network.initialize_params(session=session) updater = Updater(config=config, network=network) updater.set_learning_rate(1.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) session.run(updater.get_optim_op()) assert_almost_equal(session.run(network.get_default_output_layer().output.placeholder), 0.0) try: session.run(updater.get_optim_op()) except tf.errors.InvalidArgumentError as exc: print(('Expected exception: %r' % exc)) else: assert False, 'should have raised an exception'
def test_Updater_simple_batch(): with make_scope() as session: from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config from returnn.datasets.generating import Task12AXDataset dataset = Task12AXDataset() dataset.init_seq_order(epoch=1) extern_data = ExternData() extern_data.init_from_dataset(dataset) config = Config() network = TFNetwork(extern_data=extern_data, train_flag=True) network.construct_from_dict({'layer1': {'class': 'linear', 'activation': 'tanh', 'n_out': 13, 'from': 'data:data'}, 'layer2': {'class': 'linear', 'activation': 'tanh', 'n_out': 13, 'from': ['layer1']}, 'output': {'class': 'softmax', 'loss': 'ce', 'target': 'classes', 'from': ['layer2']}}) network.initialize_params(session=session) updater = Updater(config=config, network=network) updater.set_learning_rate(1.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) from returnn.tf.data_pipeline import FeedDictDataProvider batches = dataset.generate_batches(recurrent_net=network.recurrent, batch_size=100, max_seqs=10, max_seq_length=sys.maxsize, used_data_keys=network.used_data_keys) data_provider = FeedDictDataProvider(extern_data=extern_data, data_keys=network.used_data_keys, dataset=dataset, batches=batches) (feed_dict, _) = data_provider.get_feed_dict(single_threaded=True) session.run(updater.get_optim_op(), feed_dict=feed_dict)
def test_Updater_decouple_constraints(): with make_scope() as session: from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config from returnn.datasets.generating import Task12AXDataset dataset = Task12AXDataset() dataset.init_seq_order(epoch=1) extern_data = ExternData() extern_data.init_from_dataset(dataset) config = Config({'decouple_constraints': True}) network = TFNetwork(extern_data=extern_data, train_flag=True) network.construct_from_dict({'layer1': {'class': 'linear', 'activation': 'tanh', 'n_out': 13, 'L2': 0.01, 'from': 'data:data'}, 'layer2': {'class': 'linear', 'activation': 'tanh', 'n_out': 13, 'L2': 0.01, 'from': 'layer1'}, 'output': {'class': 'softmax', 'loss': 'ce', 'target': 'classes', 'from': 'layer2'}}) network.initialize_params(session=session) updater = Updater(config=config, network=network) updater.set_learning_rate(1.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) update_op = updater.get_optim_op() assert updater.decouple_constraints from returnn.tf.data_pipeline import FeedDictDataProvider batches = dataset.generate_batches(recurrent_net=network.recurrent, batch_size=100, max_seqs=10, max_seq_length=sys.maxsize, used_data_keys=network.used_data_keys) data_provider = FeedDictDataProvider(extern_data=extern_data, data_keys=network.used_data_keys, dataset=dataset, batches=batches) (feed_dict, _) = data_provider.get_feed_dict(single_threaded=True) session.run(update_op, feed_dict=feed_dict)
def test_Updater_decouple_constraints_simple_graph(): with make_scope() as session: from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config config = Config({'decouple_constraints': True, 'decouple_constraints_factor': 1.0}) extern_data = ExternData({'data': dict(shape=(), dtype='float32')}) network = TFNetwork(config=config, extern_data=extern_data, train_flag=True) network.construct_from_dict({'var': {'class': 'variable', 'shape': (), 'init': 1.0, 'L2': 0.01}, 'loss': {'class': 'eval', 'from': ['data', 'var'], 'eval': 'source(0, auto_convert=False) * source(1, auto_convert=False)', 'loss': 'as_is'}}) network.initialize_params(session=session) var = network.get_layer('var').output.placeholder assert_equal(session.run(var), 1.0) updater = Updater(config=config, network=network, initial_learning_rate=1.0) updater.set_learning_rate(1.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) update_op = updater.get_optim_op() assert updater.decouple_constraints tf_util.print_graph_output(update_op) session.run(update_op, feed_dict={extern_data.data['data'].placeholder: [0.0, 0.0, 0.0], extern_data.get_batch_info().dim: 3}) assert_almost_equal(session.run(var), (1.0 - (2.0 * 0.01)))
def test_Updater_decouple_constraints_simple_graph_grad_accum(): with make_scope() as session: from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config config = Config({'decouple_constraints': True, 'decouple_constraints_factor': 1.0, 'accum_grad_multiple_step': 2}) extern_data = ExternData({'data': dict(shape=(), dtype='float32')}) network = TFNetwork(config=config, extern_data=extern_data, train_flag=True) network.construct_from_dict({'var': {'class': 'variable', 'shape': (), 'init': 1.0, 'L2': 0.01}, 'loss': {'class': 'eval', 'from': ['data', 'var'], 'eval': 'source(0, auto_convert=False) * source(1, auto_convert=False)', 'loss': 'as_is'}}) network.initialize_params(session=session) var = network.get_layer('var').output.placeholder assert_equal(session.run(var), 1.0) updater = Updater(config=config, network=network, initial_learning_rate=1.0) updater.set_learning_rate(1.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) update_op = updater.get_optim_op() assert updater.decouple_constraints tf_util.print_graph_output(update_op) assert_equal(session.run(updater.global_train_step), 0) expected_var = 1.0 for i in range(10): print('Run step', i) session.run(update_op, feed_dict={extern_data.data['data'].placeholder: [0.0, 0.0, 0.0], extern_data.get_batch_info().dim: 3}) assert_equal(session.run(updater.global_train_step), (i + 1)) var_value = session.run(var) print('var:', var_value) if ((i % 2) == 1): expected_var -= ((2.0 * 0.01) * expected_var) assert_almost_equal(var_value, expected_var)
def test_Updater_multiple_optimizers(): with make_scope() as session: from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config from returnn.datasets.generating import Task12AXDataset dataset = Task12AXDataset() dataset.init_seq_order(epoch=1) extern_data = ExternData() extern_data.init_from_dataset(dataset) config = Config() network = TFNetwork(extern_data=extern_data, train_flag=True) network.construct_from_dict({'layer1': {'class': 'linear', 'activation': 'tanh', 'n_out': 13, 'from': 'data:data', 'updater_opts': {'optimizer': {'class': 'Adam'}}}, 'layer2': {'class': 'linear', 'activation': 'tanh', 'n_out': 13, 'from': ['layer1'], 'updater_opts': {'optimizer': {'class': 'Adagrad'}}}, 'output': {'class': 'softmax', 'loss': 'ce', 'target': 'classes', 'from': ['layer2']}}) network.initialize_params(session=session) updater = Updater(config=config, network=network) updater.set_learning_rate(1.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) optim_op = updater.get_optim_op() assert (len(updater.optimizers) == 3) from returnn.tf.data_pipeline import FeedDictDataProvider batches = dataset.generate_batches(recurrent_net=network.recurrent, batch_size=100, max_seqs=10, max_seq_length=sys.maxsize, used_data_keys=network.used_data_keys) data_provider = FeedDictDataProvider(extern_data=extern_data, data_keys=network.used_data_keys, dataset=dataset, batches=batches) (feed_dict, _) = data_provider.get_feed_dict(single_threaded=True) session.run(optim_op, feed_dict=feed_dict)
def test_Updater_multiple_optimizers_and_opts(): with make_scope() as session: from returnn.tf.network import TFNetwork, ExternData from returnn.config import Config from returnn.datasets.generating import Task12AXDataset dataset = Task12AXDataset() dataset.init_seq_order(epoch=1) extern_data = ExternData() extern_data.init_from_dataset(dataset) config = Config() network = TFNetwork(extern_data=extern_data, train_flag=True) network.construct_from_dict({'layer1': {'class': 'linear', 'activation': 'tanh', 'n_out': 13, 'from': 'data:data', 'updater_opts': {'optimizer': {'class': 'Adam'}, 'accum_grad_multiple_step': 2}}, 'layer2': {'class': 'linear', 'activation': 'tanh', 'n_out': 13, 'from': ['layer1'], 'updater_opts': {'optimizer': {'class': 'Adagrad', 'learning_rate_multiplier': 3}, 'gradient_noise': 0.1}}, 'output': {'class': 'softmax', 'loss': 'ce', 'target': 'classes', 'from': ['layer2']}}) network.initialize_params(session=session) updater = Updater(config=config, network=network) updater.set_learning_rate(1.0, session=session) updater.set_trainable_vars(network.get_trainable_params()) updater.init_optimizer_vars(session=session) optim_op = updater.get_optim_op() assert (len(updater.optimizers) == 3) from returnn.tf.data_pipeline import FeedDictDataProvider batches = dataset.generate_batches(recurrent_net=network.recurrent, batch_size=100, max_seqs=10, max_seq_length=sys.maxsize, used_data_keys=network.used_data_keys) data_provider = FeedDictDataProvider(extern_data=extern_data, data_keys=network.used_data_keys, dataset=dataset, batches=batches) (feed_dict, _) = data_provider.get_feed_dict(single_threaded=True) session.run(optim_op, feed_dict=feed_dict)
def load_data(): from returnn.__main__ import load_data (dev_data, _) = load_data(config, 0, 'dev', chunking=config.value('chunking', ''), seq_ordering='sorted', shuffle_frames_of_nseqs=0) (eval_data, _) = load_data(config, 0, 'eval', chunking=config.value('chunking', ''), seq_ordering='sorted', shuffle_frames_of_nseqs=0) (train_data, _) = load_data(config, 0, 'train') return (dev_data, eval_data, train_data)
def test_determinism_of_vanillalstm(): def create_engine(): (dev_data, eval_data, train_data) = load_data() engine = Engine() engine.init_train_from_config(config, train_data, dev_data, eval_data) engine.init_train_epoch() engine.train_batches = engine.train_data.generate_batches(recurrent_net=engine.network.recurrent, batch_size=engine.batch_size, max_seqs=engine.max_seqs, max_seq_length=engine.max_seq_length, seq_drop=engine.seq_drop, shuffle_batches=engine.shuffle_batches, used_data_keys=engine.network.used_data_keys) engine.updater.set_learning_rate(engine.learning_rate, session=engine.tf_session) engine.updater.init_optimizer_vars(session=engine.tf_session) return engine def train_engine_fetch_vars(engine): data_provider = engine._get_data_provider(dataset=engine.train_data, batches=engine.train_batches, feed_dict=True) (feed_dict, _) = data_provider.get_feed_dict(single_threaded=True) trainer = Runner(engine=engine, dataset=engine.train_data, batches=engine.train_batches, train=True) (feed_dict, _) = data_provider.get_feed_dict(single_threaded=True) trainer.run(report_prefix='One Run') return [e.eval(engine.tf_session) for e in engine.network.get_params_list()] e1 = create_engine() r1 = train_engine_fetch_vars(e1) e2 = create_engine() r2 = train_engine_fetch_vars(e2) assert_array_equal(r1, r2)
def pickle_dumps(obj): sio = BytesIO() p = Pickler(sio) p.dump(obj) return sio.getvalue()
def pickle_loads(s): p = Unpickler(BytesIO(s)) return p.load()
def test_pickle_anon_new_class(): class Foo(object): a = 'class' b = 'foo' def __init__(self): self.a = 'hello' def f(self, a): return a s = pickle_dumps(Foo) Foo2 = pickle_loads(s) assert inspect.isclass(Foo2) assert (Foo is not Foo2) assert (Foo2.a == 'class') assert (Foo2.b == 'foo') inst = Foo2() assert (inst.a == 'hello') assert (inst.b == 'foo') assert (inst.f(42) == 42)
def test_pickle_anon_old_class(): class Foo(): a = 'class' b = 'foo' def __init__(self): self.a = 'hello' def f(self, a): return a s = pickle_dumps(Foo) Foo2 = pickle_loads(s) assert inspect.isclass(Foo2) assert (Foo is not Foo2) assert (Foo2.a == 'class') assert (Foo2.b == 'foo') inst = Foo2() assert (inst.a == 'hello') assert (inst.b == 'foo') assert (inst.f(42) == 42)
def test_pickle_inst_anon_class(): class Foo(object): a = 'class' b = 'foo' def __init__(self): self.a = 'hello' def f(self, a): return a s = pickle_dumps(Foo()) inst = pickle_loads(s) assert (inst.a == 'hello') assert (inst.b == 'foo') assert (inst.f(42) == 42)
class DemoClass(): def method(self): return 42
def test_pickle(): obj = DemoClass() s = pickle_dumps(obj.method) inst = pickle_loads(s) assert_equal(inst(), 42)
def pickle_dumps(obj): sio = BytesIO() p = Pickler(sio) p.dump(obj) return sio.getvalue()
def pickle_loads(s): p = Unpickler(BytesIO(s)) return p.load()
def find_numpy_shared_by_shmid(shmid): for sh in SharedNumpyArray.ServerInstances: assert isinstance(sh, SharedNumpyArray) assert (sh.mem is not None) assert (sh.mem.shmid > 0) if (sh.mem.shmid == shmid): return sh return None
def have_working_shmget(): global _have_working_shmget if (_have_working_shmget is None): _have_working_shmget = SharedMem.is_shmget_functioning() print('shmget functioning:', _have_working_shmget) return _have_working_shmget
@unittest.skipIf((not have_working_shmget()), 'shmget does not work') def test_shmget_functioning(): assert SharedMem.is_shmget_functioning()
@unittest.skipIf((not have_working_shmget()), 'shmget does not work') def test_pickle_numpy(): m = numpy.random.randn(10, 10) p = pickle_dumps(m) m2 = pickle_loads(p) assert isinstance(m2, numpy.ndarray) assert numpy.allclose(m, m2) assert isinstance(m2.base, SharedNumpyArray) shared_client = m2.base assert (not shared_client.is_server) shared_server = find_numpy_shared_by_shmid(m2.base.mem.shmid) assert shared_server.is_server assert numpy.allclose(m, shared_server.create_numpy_array()) assert numpy.allclose(m, shared_client.create_numpy_array()) assert shared_server.is_in_use() assert shared_client.is_in_use() numpy_set_unused(m2) assert (not shared_server.is_in_use()) assert (shared_client.mem is None)
@unittest.skipIf((not have_working_shmget()), 'shmget does not work') def test_pickle_numpy_scalar(): m = numpy.array([numpy.random.randn()]) assert isinstance(m, numpy.ndarray) assert (m.shape == (1,)) assert (m.nbytes >= 1) p = pickle_dumps(m) m2 = pickle_loads(p) assert isinstance(m2, numpy.ndarray) assert numpy.allclose(m, m2) assert isinstance(m2.base, SharedNumpyArray) shared_client = m2.base assert (not shared_client.is_server) shared_server = find_numpy_shared_by_shmid(m2.base.mem.shmid) assert shared_server.is_server assert numpy.allclose(m, shared_server.create_numpy_array()) assert numpy.allclose(m, shared_client.create_numpy_array()) assert shared_server.is_in_use() assert shared_client.is_in_use() numpy_set_unused(m2) assert (not shared_server.is_in_use()) assert (shared_client.mem is None)
@unittest.skipIf((not have_working_shmget()), 'shmget does not work') def test_pickle_gc_aggressive(): m = numpy.random.randn(10, 10) p = pickle_dumps(m) m2 = pickle_loads(p) assert isinstance(m2, numpy.ndarray) assert numpy.allclose(m, m2) assert isinstance(m2.base, SharedNumpyArray) print(('refcount: %i' % sys.getrefcount(m2.base))) gc.collect() gc.collect() print(('refcount: %i' % sys.getrefcount(m2.base))) assert m2.base.is_in_use() server = find_numpy_shared_by_shmid(m2.base.mem.shmid) m2 = None gc.collect()
@unittest.skipIf((not have_working_shmget()), 'shmget does not work') def test_pickle_multiple(): for i in range(20): ms = [numpy.random.randn(10, 10) for i in range(((i % 3) + 1))] p = pickle_dumps(ms) ms2 = pickle_loads(p) assert (len(ms) == len(ms2)) for (m, m2) in zip(ms, ms2): assert numpy.allclose(m, m2) assert isinstance(m2.base, SharedNumpyArray)
@unittest.skipIf((not have_working_shmget()), 'shmget does not work') def test_pickle_unpickle_auto_unused(): old_num_servers = None for i in range(10): m = numpy.random.randn(((i * 2) + 1), ((i * 3) + 2)) p = pickle_dumps((m, m, m)) new_num_servers = len(SharedNumpyArray.ServerInstances) if (old_num_servers is not None): assert (old_num_servers == new_num_servers) old_num_servers = new_num_servers (m2, m3, m4) = pickle_loads(p) assert numpy.allclose(m, m2) assert numpy.allclose(m, m3) assert numpy.allclose(m, m4) assert (not m4.base.is_server) m4.base._get_in_use_flag_ref().value = 42 assert (m4.base._get_in_use_flag_ref().value == 42) assert (find_numpy_shared_by_shmid(m4.base.mem.shmid)._get_in_use_flag_ref().value == 42) assert numpy.allclose(m, m4) ss = list([find_numpy_shared_by_shmid(_m.base.mem.shmid) for _m in (m2, m3, m4)]) _m = None m2 = m3 = m4 = None gc.collect() for s in ss: assert isinstance(s, SharedNumpyArray) assert s.is_server assert (not s.is_in_use())
def create_vocabulary(text): '\n :param str text: any natural text\n :return: mapping of words in the text to ids, as well as the inverse mapping\n :rtype: (dict[str, int], dict[int, str])\n ' vocabulary = {word: index for (index, word) in enumerate(set(text.strip().split()))} inverse_vocabulary = {index: word for (word, index) in vocabulary.items()} return (vocabulary, inverse_vocabulary)
def word_ids_to_sentence(word_ids, vocabulary): '\n :param list[int] word_ids:\n :param dict[int, str] vocabulary: mapping from word ids to words\n :return: concatenation of all words\n :rtype: str\n ' words = [vocabulary[word_id] for word_id in word_ids] return ' '.join(words)
def test_translation_dataset(): '\n Checks whether a dummy translation dataset can be read and whether the returned word indices are correct.\n We create the necessary corpus and vocabulary files on the fly.\n ' dummy_dataset = tempfile.mkdtemp() source_file_name = os.path.join(dummy_dataset, 'source.test.gz') target_file_name = os.path.join(dummy_dataset, 'target.test') with gzip.open(source_file_name, 'wb') as source_file: source_file.write(dummy_source_text.encode('utf-8')) with open(target_file_name, 'wb') as target_file: target_file.write(dummy_target_text.encode('utf-8')) for postfix in ['', ' </S>']: dummy_target_text_with_unk = dummy_target_text.replace('TranslationDatasets', '<UNK>') (source_vocabulary, inverse_source_vocabulary) = create_vocabulary((dummy_source_text + postfix)) (target_vocabulary, inverse_target_vocabulary) = create_vocabulary((dummy_target_text_with_unk + postfix)) source_vocabulary_file_name = os.path.join(dummy_dataset, 'source.vocab.pkl') target_vocabulary_file_name = os.path.join(dummy_dataset, 'target.vocab.pkl') with open(source_vocabulary_file_name, 'wb') as source_vocabulary_file: pickle.dump(source_vocabulary, source_vocabulary_file) with open(target_vocabulary_file_name, 'wb') as target_vocabulary_file: pickle.dump(target_vocabulary, target_vocabulary_file) translation_dataset = TranslationDataset(path=dummy_dataset, file_postfix='test', source_postfix=postfix, target_postfix=postfix, unknown_label={'classes': '<UNK>'}) translation_dataset.init_seq_order(epoch=1) translation_dataset.load_seqs(0, 10) num_seqs = len(dummy_source_text.splitlines()) assert_equal(translation_dataset.num_seqs, num_seqs) for sequence_index in range(num_seqs): source_word_ids = translation_dataset.get_data(sequence_index, 'data') source_sentence = word_ids_to_sentence(source_word_ids, inverse_source_vocabulary) assert_equal(source_sentence, (dummy_source_text.splitlines()[sequence_index] + postfix)) target_word_ids = translation_dataset.get_data(sequence_index, 'classes') target_sentence = word_ids_to_sentence(target_word_ids, inverse_target_vocabulary) assert_equal(target_sentence, (dummy_target_text_with_unk.splitlines()[sequence_index] + postfix)) shutil.rmtree(dummy_dataset)
def test_translation_factors_dataset(): '\n Similar to test_translation_dataset(), but using translation factors.\n ' source_text_per_factor = [dummy_source_text_factor_0, dummy_source_text_factor_1] target_text_per_factor = [dummy_target_text_factor_0, dummy_target_text_factor_1, dummy_target_text_factor_2] source_vocabulary_names = ['source.vocab.pkl', 'source_factor1.vocab.pkl'] target_vocabulary_names = ['target.vocab.pkl', 'target_factor1.vocab.pkl', 'target_factor2.vocab.pkl'] source_data_keys = ['data', 'source_factor1'] target_data_keys = ['classes', 'target_factor1', 'target_factor2'] dummy_dataset = tempfile.mkdtemp() source_file_name = os.path.join(dummy_dataset, 'source.test.gz') target_file_name = os.path.join(dummy_dataset, 'target.test') with gzip.open(source_file_name, 'wb') as source_file: source_file.write(dummy_source_text_factored_format.encode('utf8')) with open(target_file_name, 'wb') as target_file: target_file.write(dummy_target_text_factored_format.encode('utf8')) for postfix in ['', ' </S>']: (vocabularies, inverse_vocabularies) = ([], []) for dummy_text in (source_text_per_factor + target_text_per_factor): (vocabulary, inverse_vocabulary) = create_vocabulary((dummy_text + postfix)) vocabularies.append(vocabulary) inverse_vocabularies.append(inverse_vocabulary) vocabulary_names = (source_vocabulary_names + target_vocabulary_names) for (index, vocabulary) in enumerate(vocabularies): with open(os.path.join(dummy_dataset, vocabulary_names[index]), 'wb') as vocabulary_file: pickle.dump(vocabulary, vocabulary_file) translation_dataset = TranslationFactorsDataset(path=dummy_dataset, file_postfix='test', factor_separator='|', source_factors=source_data_keys[1:], target_factors=target_data_keys[1:], source_postfix=postfix, target_postfix=postfix) translation_dataset.init_seq_order(epoch=1) translation_dataset.load_seqs(0, 10) num_seqs = len(dummy_target_text_factored_format.splitlines()) assert_equal(translation_dataset.num_seqs, num_seqs) data_keys = (source_data_keys + target_data_keys) texts_per_factor = (source_text_per_factor + target_text_per_factor) for (index, text) in enumerate(texts_per_factor): for sequence_index in range(num_seqs): word_ids = translation_dataset.get_data(sequence_index, data_keys[index]) sentence = word_ids_to_sentence(word_ids, inverse_vocabularies[index]) assert_equal(sentence, (text.splitlines()[sequence_index] + postfix)) shutil.rmtree(dummy_dataset)
def build_env(env_update=None): '\n :param dict[str,str]|None env_update:\n :return: env dict for Popen\n :rtype: dict[str,str]\n ' env_update_ = os.environ.copy() if env_update: env_update_.update(env_update) return env_update_
def run(*args, env_update=None, print_stdout=False): args = list(args) print('run:', args) p = Popen(args, stdout=PIPE, stderr=STDOUT, env=build_env(env_update=env_update)) (out, _) = p.communicate() out = out.decode('utf8') if (p.returncode != 0): print(('Return code is %i' % p.returncode)) print(('std out/err:\n---\n%s\n---\n' % out)) raise CalledProcessError(cmd=args, returncode=p.returncode, output=out) if print_stdout: print(('std out/err:\n---\n%s\n---\n' % out)) return out
def parse_last_fer(out: str) -> float: '\n :param out:\n :return: FER\n ' parsed_fer = None for line in out.splitlines(): m = re.match('epoch [0-9]+ score: .* dev: .* error ([0-9.]+)\\s?', line) if (not m): m = re.match('dev: score .* error ([0-9.]+)\\s?', line) if (not m): continue parsed_fer = float(m.group(1)) err_msg = 'ERROR: No epoch dev errors found in output' assert (parsed_fer is not None), ('%s.\nOutput:\n\n%s\n\n%s.' % (err_msg, out, err_msg)) return parsed_fer
def run_and_parse_last_fer(*args, **kwargs): out = run(*args, **kwargs) return parse_last_fer(out)
def run_config_get_fer(config_filename, *args, env_update=None, log_verbosity=5, print_stdout=False, pre_cleanup=True, post_cleanup=True): if pre_cleanup: cleanup_tmp_models(config_filename) fer = run_and_parse_last_fer(py, 'rnn.py', config_filename, '++log_verbosity', str(log_verbosity), *args, env_update=env_update, print_stdout=print_stdout) print('FER:', fer) if post_cleanup: cleanup_tmp_models(config_filename) return fer
def get_model_filename(config_filename: str) -> str: assert os.path.exists(config_filename) from returnn.config import Config config = Config() config.load_file(config_filename) model_filename = config.value('model', '') assert model_filename assert model_filename.startswith('/tmp/') return model_filename
def cleanup_tmp_models(config_filename: str): model_filename = get_model_filename(config_filename) for f in glob((model_filename + '.*')): os.remove(f)
@unittest.skipIf((not tf), 'no TF') def test_demo_tf_task12ax(): fer = run_config_get_fer('demos/demo-tf-native-lstm.12ax.config', print_stdout=True) assert_less(fer, 0.015)
@unittest.skipIf((not tf), 'no TF') def test_demo_tf_task12ax_eval(): cfg_filename = 'demos/demo-tf-native-lstm.12ax.config' train_dataset_repr = '{"class": "Task12AXDataset", "num_seqs": 10}' dev_dataset_repr = '{"class": "Task12AXDataset", "num_seqs": 10}' fer1 = run_config_get_fer(cfg_filename, '++num_epochs', '2', '++train', train_dataset_repr, '++dev', dev_dataset_repr, print_stdout=True, post_cleanup=False) fer2 = run_config_get_fer(cfg_filename, '++task', 'eval', '++load_epoch', '2', '++train', 'None', '++dev', dev_dataset_repr, print_stdout=True, pre_cleanup=False, post_cleanup=False) assert (fer1 == fer2) model_filename = get_model_filename(cfg_filename) ep2_files = glob((model_filename + '.002.*')) assert ep2_files, f'No model files found for epoch 2, {model_filename}' for fn in ep2_files: shutil.copy(fn, fn.replace('.002.', '.003.')) fer3 = run_config_get_fer(cfg_filename, '++task', 'eval', '++load_epoch', '3', '++train', 'None', '++dev', dev_dataset_repr, print_stdout=True, pre_cleanup=False) assert (fer3 != fer2)
@unittest.skipIf((not torch), 'no PyTorch') def test_demo_torch_task12ax(): cleanup_tmp_models('demos/demo-torch.config') out = run(py, 'rnn.py', 'demos/demo-torch.config', print_stdout=True) fer = parse_last_fer(out) assert_less(fer, 0.02)
def _test_torch_export_to_onnx(cfg_filename: str) -> str: '\n Executes the demo passed as a parameter and returns the ONNX exported model as a filename.\n\n :param cfg_filename: Demo filename, either "demos/demo-rf.config" or "demos/demo-torch.config".\n :return: Filename representing the ONNX model location.\n ' cleanup_tmp_models(cfg_filename) out = run(py, 'rnn.py', cfg_filename, '--num_epochs', '1', '--device', 'cpu') out_pt_model = re.search('\\nSave model under (.*)\\n', out, re.MULTILINE) assert out_pt_model, ('Could not find the model filename in the output:\n%s' % out) out_pt_model = out_pt_model.group(1) print('*** PT model:', out_pt_model) out_onnx_model = out_pt_model.replace('.pt', '.onnx') run(py, 'tools/torch_export_to_onnx.py', cfg_filename, out_pt_model, out_onnx_model, print_stdout=True) return out_onnx_model
def _test_torch_onnx_inference_no_seq_lens(out_onnx_model: str): '\n Tests the inference of the torch demo with an ONNX model passed as parameter.\n ' import onnxruntime as ort torch.manual_seed(42) dummy_data = torch.randn([3, 50, 9]) session = ort.InferenceSession(out_onnx_model) outputs_onnx = session.run(None, {'data': dummy_data.numpy()}) print('*** Result:', torch.FloatTensor(outputs_onnx[0])) print('*** Batch size:', torch.IntTensor(outputs_onnx[1])) print('*** Sequence lengths:', torch.IntTensor(outputs_onnx[2]))
def _test_torch_onnx_inference_seq_lens_in_out(out_onnx_model: str): '\n Tests the inference of the torch demo with an ONNX model passed as parameter.\n ' print(out_onnx_model) import onnxruntime as ort torch.manual_seed(42) dummy_data = torch.randn([3, 50, 9]) dummy_seq_lens = torch.tensor([27, 50, 43], dtype=torch.int32) session = ort.InferenceSession(out_onnx_model) outputs_onnx = session.run(None, {'data': dummy_data.numpy(), 'data:size1': dummy_seq_lens.numpy()}) out_result = torch.FloatTensor(outputs_onnx[0]) out_seq_lens = torch.IntTensor(outputs_onnx[1]) print('*** Result:', out_result) print('*** Sequence lengths:', out_seq_lens) assert (out_result.shape == torch.Size([3, 50, 2]))
@unittest.skipIf((not torch), 'no PyTorch') def test_demo_torch_export_to_onnx(): out_onnx_model = _test_torch_export_to_onnx('demos/demo-torch.config') _test_torch_onnx_inference_seq_lens_in_out(out_onnx_model)
@unittest.skipIf((not torch), 'no PyTorch') def test_demo_rf_export_to_onnx(): out_onnx_model = _test_torch_export_to_onnx('demos/demo-rf.config') _test_torch_onnx_inference_seq_lens_in_out(out_onnx_model)
@unittest.skipIf((not torch), 'no PyTorch') def test_demo_rf_torch_task12ax(): cleanup_tmp_models('demos/demo-rf.config') out = run(py, 'rnn.py', 'demos/demo-rf.config', print_stdout=True) fer = parse_last_fer(out) assert_less(fer, 0.02)
@unittest.skipIf((not tf), 'no TF') def test_demo_rf_tf_task12ax(): cleanup_tmp_models('demos/demo-rf.config') out = run(py, 'rnn.py', 'demos/demo-rf.config', '++backend', 'tensorflow-net-dict', print_stdout=True) fer = parse_last_fer(out) assert_less(fer, 0.02)
def test_demo_iter_dataset_task12ax(): cleanup_tmp_models('demos/demo-tf-vanilla-lstm.12ax.config') out = run(py, 'demos/demo-iter-dataset.py', 'demos/demo-tf-vanilla-lstm.12ax.config') assert_in('Epoch 5.', out.splitlines())
@unittest.skipIf((not tf), 'no TF') def test_demo_returnn_as_framework(): print('Prepare.') import subprocess import shutil from glob import glob from returnn.util.basic import get_login_username subprocess.check_call(['echo', 'travis_fold:start:test_demo_returnn_as_framework']) assert os.path.exists('setup.py') if glob('dist/*.tar.gz'): for fn in glob('dist/*.tar.gz'): os.remove(fn) if os.path.exists('MANIFEST'): os.remove('MANIFEST') if os.path.exists('docs/crnn'): os.remove('docs/crnn') if os.path.exists('docs/returnn'): os.remove('docs/returnn') tmp_model_dir = ('/tmp/%s/returnn-demo-as-framework' % get_login_username()) if os.path.exists(tmp_model_dir): shutil.rmtree(tmp_model_dir, ignore_errors=True) print('setup.py sdist, to create package.') subprocess.check_call([py, 'setup.py', 'sdist']) dist_fns = glob('dist/*.tar.gz') assert (len(dist_fns) == 1) dist_fn = os.path.abspath(dist_fns[0]) pip_path = which_pip() print('Pip install Returnn.') in_virtual_env = hasattr(sys, 'real_prefix') cmd = [py, pip_path, 'install'] if (not in_virtual_env): cmd += ['--user'] cmd += ['-v', dist_fn] print(('$ %s' % ' '.join(cmd))) subprocess.check_call(cmd, cwd='/') print('Running demo now.') subprocess.check_call([py, 'demo-returnn-as-framework.py'], cwd='demos') print('Success.') subprocess.check_call(['echo', 'travis_fold:end:test_demo_returnn_as_framework'])
@unittest.skipIf((not tf), 'no TF') def test_demo_sprint_interface(): import subprocess subprocess.check_call(['echo', 'travis_fold:start:test_demo_sprint_interface']) subprocess.check_call([py, os.path.abspath('demos/demo-sprint-interface.py')], cwd='/') subprocess.check_call(['echo', 'travis_fold:end:test_demo_sprint_interface'])
def test_returnn_as_framework_TaskSystem(): import subprocess subprocess.check_call(['echo', 'travis_fold:start:test_returnn_as_framework_TaskSystem']) subprocess.check_call([py, os.path.abspath('tests/returnn-as-framework.py'), 'test_TaskSystem_Pickler()'], cwd='/') subprocess.check_call(['echo', 'travis_fold:end:test_returnn_as_framework_TaskSystem'])
@unittest.skipIf((not tf), 'no TF') def test_returnn_as_framework_old_style_crnn_TFUtil(): "\n Check that old-style `import crnn.TFUtil` works.\n\n It's not so much about TFUtil, it also could be some other module.\n It's about the old-style module names.\n This is the logic in __old_mod_loader__.\n " import subprocess subprocess.check_call(['echo', 'travis_fold:start:test_returnn_as_framework_old_style_crnn_TFUtil']) subprocess.check_call([py, os.path.abspath('tests/returnn-as-framework.py'), '--old-style', '--returnn-package-name', 'crnn', 'test_old_style_import_crnn_TFUtil()'], cwd='/') subprocess.check_call(['echo', 'travis_fold:end:test_returnn_as_framework_old_style_crnn_TFUtil'])
@unittest.skipIf((not tf), 'no TF') def test_returnn_as_framework_old_style_TFUtil(): '\n Check that old-style `import TFUtil` works.\n See also :func:`test_returnn_as_framework_old_style_crnn_TFUtil`.\n ' import subprocess subprocess.check_call(['echo', 'travis_fold:start:test_returnn_as_framework_old_style_TFUtil']) subprocess.check_call([py, os.path.abspath('tests/returnn-as-framework.py'), '--old-style', 'test_old_style_import_TFUtil()'], cwd='/') subprocess.check_call(['echo', 'travis_fold:end:test_returnn_as_framework_old_style_TFUtil'])
class CLibAtForkDemo(): def __init__(self): self._load_lib() def _load_lib(self): from returnn.util.basic import NativeCodeCompiler native = NativeCodeCompiler(base_name='test_fork_exec', code_version=1, code=c_code_at_fork_demo, is_cpp=False, ld_flags=['-lpthread']) self._lib = native.load_lib_ctypes() print('loaded lib:', native.get_lib_filename()) import ctypes self._lib.register_hello_from_child.restype = None self._lib.register_hello_from_child.argtypes = () self._lib.register_hello_from_fork_prepare.restype = None self._lib.register_hello_from_fork_prepare.argtypes = () self._lib.set_magic_number.restype = None self._lib.set_magic_number.argtypes = (ctypes.c_long,) def set_magic_number(self, i): self._lib.set_magic_number(i) def register_hello_from_child(self): self._lib.register_hello_from_child() def register_hello_from_fork_prepare(self): self._lib.register_hello_from_fork_prepare()
def demo_hello_from_fork(): print('Hello.') sys.stdout.flush() clib_at_fork_demo.set_magic_number(3) clib_at_fork_demo.register_hello_from_child() clib_at_fork_demo.register_hello_from_fork_prepare() pid = os.fork() if (pid == 0): print('Hello from child after fork.') sys.exit() print('Hello from parent after fork.') os.waitpid(pid, 0) print('Bye.')
def demo_start_subprocess(): print('Hello.') sys.stdout.flush() clib_at_fork_demo.set_magic_number(5) clib_at_fork_demo.register_hello_from_child() clib_at_fork_demo.register_hello_from_fork_prepare() from subprocess import check_call check_call('echo Hello from subprocess.', shell=True) print('Bye.')
def run_demo_check_output(name): '\n :param str name: e.g. "demo_hello_from_fork"\n :return: lines of stdout of the demo\n :rtype: list[str]\n ' from subprocess import check_output output = check_output([sys.executable, __file__, name]) return output.decode('utf8').splitlines()
def filter_demo_output(ls): '\n :param list[str] ls:\n :rtype: list[str]\n ' ls = [l for l in ls if (not l.startswith('Executing: '))] ls = [l for l in ls if (not l.startswith('Compiler call: '))] ls = [l for l in ls if (not l.startswith('loaded lib: '))] ls = [l for l in ls if (not l.startswith('dlopen: '))] ls = [l for l in ls if (('installLibSigSegfault' not in l) and ('libSegFault' not in l))] ls = [l for l in ls if ('faulthandler' not in l)] found_hello = False for (i, l) in enumerate(ls): if (l in ['Ignoring pthread_atfork call!', 'Ignoring __register_atfork call!']): continue assert (l == 'Hello.') ls.pop(i) found_hello = True break assert found_hello, ('no Hello: %r' % (ls,)) assert (ls[(- 1)] == 'Bye.') ls = ls[:(- 1)] return ls
def test_demo_hello_from_fork(): ls = run_demo_check_output('demo_hello_from_fork') pprint(ls) ls = filter_demo_output(ls) pprint(ls) assert_equal(set(ls), {'Hello from child after fork.', 'Hello from child atfork, magic number 3.', 'Hello from atfork prepare, magic number 3.', 'Hello from parent after fork.'})
def test_demo_start_subprocess(): ls = run_demo_check_output('demo_start_subprocess') pprint(ls) ls = filter_demo_output(ls) pprint(ls) assert ('Hello from subprocess.' in ls) import platform print('Python impl:', platform.python_implementation()) print('Python version:', sys.version_info[:3]) if (platform.python_implementation() == 'CPython'): if (sys.version_info[0] == 2): old = (sys.version_info[:3] <= (2, 7, 12)) else: old = (sys.version_info[:3] <= (3, 6, 1)) else: old = False if old: print('atfork handler should have been called.') assert ('Hello from child atfork, magic number 5.' in ls) assert ('Hello from atfork prepare, magic number 5.' in ls) else: print('Not checking for atfork handler output.')
def patched_check_demo_start_subprocess(): '\n Just like test_demo_start_subprocess(), but here we assert that no atfork handlers are executed.\n ' assert_equal(os.environ.get('__RETURNN_ATFORK_PATCHED'), '1') ls = run_demo_check_output('demo_start_subprocess') pprint(ls) ls = filter_demo_output(ls) pprint(ls) assert ('Hello from subprocess.' in ls) ls = [l for l in ls if (l not in ['Ignoring pthread_atfork call!', 'Ignoring __register_atfork call!'])] pprint(ls) assert_equal(ls, ['Hello from subprocess.'])