def test_stftConfig_multi_res_02():
with make_scope() as session:
layer_name = "stft_layer"
fft_sizes = [400, 200, 800]
frame_sizes = [400, 200, 800]
frame_shift = 160
window = "hanning"
test_input = np.random.normal(0, 0.6, (1, 3200, 2))
num_outputs = int(
np.sum([(int(fft_size / 2) + 1) * test_input.shape[2]
for fft_size in fft_sizes]))
config = Config()
config.update({
"num_outputs": num_outputs,
"num_inputs": test_input.shape[2],
"network": {
layer_name: {
"class": "multichannel_multiresolution_stft_layer",
"frame_shift": frame_shift,
"frame_sizes": frame_sizes,
"window": window,
"fft_sizes": fft_sizes,
"use_rfft": True,
"nr_of_channels": 2,
"is_output_layer": True,
"from": "data:data"
}
}
})
network = TFNetwork(config=config, train_flag=True)
network.construct_from_dict(config.typed_value("network"))
layer = network.layers[layer_name]
test_output = session.run(
layer.output.placeholder,
{network.get_extern_data('data').placeholder: test_input})
assert test_output.shape[2] == num_outputs
comparison_frame = 6
ref00 = _get_ref_output_single_res(test_input, fft_sizes[0],
frame_sizes[0], frame_shift,
window, comparison_frame, 0)
ref01 = _get_ref_output_single_res(test_input, fft_sizes[0],
frame_sizes[0], frame_shift,
window, comparison_frame, 1)
ref10 = _get_ref_output_single_res(test_input, fft_sizes[1],
frame_sizes[1], frame_shift,
window, comparison_frame, 0)
ref11 = _get_ref_output_single_res(test_input, fft_sizes[1],
frame_sizes[1], frame_shift,
window, comparison_frame, 1)
ref20 = _get_ref_output_single_res(test_input, fft_sizes[2],
frame_sizes[2], frame_shift,
window, comparison_frame, 0)
ref21 = _get_ref_output_single_res(test_input, fft_sizes[2],
frame_sizes[2], frame_shift,
window, comparison_frame, 1)
ref = np.concatenate([ref00, ref01, ref10, ref11, ref20, ref21],
axis=0)
resultDiff = np.abs(test_output[0, comparison_frame, :] - ref)
assert np.mean(resultDiff) < 0.06
assert np.max(resultDiff) < 1
def test_melFilterbankLayer():
with make_scope() as session:
n_in, n_out = 257, 3
layer_name = "mel_filterbank_layer"
config = Config()
config.update({
"num_outputs": n_out,
"num_inputs": n_in,
"network": {
layer_name: {
"class": "mel_filterbank",
"fft_size": 512,
"nr_of_filters": n_out,
"n_out": n_out,
"is_output_layer": True,
"from": "data:data"
}
}
})
network = TFNetwork(config=config, train_flag=True)
network.construct_from_dict(config.typed_value("network"))
layer = network.layers[layer_name]
test_out = session.run(layer.output.placeholder,
feed_dict={
network.get_extern_data('data').placeholder:
np.ones((1, 1, 257))
})
assert np.sum(test_out - np.asarray(
[28.27923584, 53.10634232, 99.71585846], dtype=np.float32)) < 1e-5
def test_rfftStftConfig_01():
with make_scope() as session:
layer_name = "stft_layer"
fft_size = 400
frame_size = 400
frame_shift = 160
window = "hanning"
test_input = np.ones((1, 32000, 2), dtype=np.float32)
config = Config()
config.update({
"num_outputs":
int(fft_size / 2) + 1 * test_input.shape[2],
"num_inputs":
test_input.shape[2],
"network": {
layer_name: {
"class": "multichannel_stft_layer",
"frame_shift": frame_shift,
"frame_size": frame_size,
"window": window,
"fft_size": fft_size,
"use_rfft": True,
"nr_of_channels": 2,
"is_output_layer": True,
"from": "data:data"
}
}
})
network = TFNetwork(config=config, train_flag=True)
network.construct_from_dict(config.typed_value("network"))
layer = network.layers[layer_name]
test_output = session.run(
layer.output.placeholder,
{network.get_extern_data('data').placeholder: test_input})
ref0 = _get_ref_output(test_input, fft_size, frame_size,
frame_shift, window, 0, 0)
# np.fft.rfft and tensorflow.python.ops.rfft differ a little bit in their
# results, thus an error margin is allowed in the result
resultDiff = np.abs(test_output[0, 0, 0:(int(fft_size / 2) + 1)] -
ref0)
assert np.mean(resultDiff) < 0.02
assert np.max(resultDiff) < 1
pass
def test_stftConfig_single_res_01():
with make_scope() as session:
layer_name = "stft_layer"
fft_sizes = [400]
frame_sizes = [400]
frame_shift = 160
window = "hanning"
test_input = np.ones((1, 32000, 2), dtype=np.float32)
num_outputs = (int(fft_sizes[0] / 2) + 1) * test_input.shape[2]
config = Config()
config.update({
"num_outputs": num_outputs,
"num_inputs": test_input.shape[2],
"network": {
layer_name: {
"class": "multichannel_multiresolution_stft_layer",
"frame_shift": frame_shift,
"frame_sizes": frame_sizes,
"window": window,
"fft_sizes": fft_sizes,
"use_rfft": True,
"nr_of_channels": 2,
"is_output_layer": True,
"from": "data:data"
}
}
})
network = TFNetwork(config=config, train_flag=True)
network.construct_from_dict(config.typed_value("network"))
layer = network.layers[layer_name]
test_output = session.run(
layer.output.placeholder,
{network.get_extern_data('data').placeholder: test_input})
ref0 = _get_ref_output_single_res(test_input, fft_sizes[0],
frame_sizes[0], frame_shift,
window, 0, 0)
resultDiff = np.abs(test_output[0, 0, 0:(int(fft_sizes[0] / 2) +
1)] - ref0)
assert test_output.shape[2] == num_outputs
assert np.mean(resultDiff) < 0.02
assert np.max(resultDiff) < 1
def test_complexLinearProjectionLayer():
with make_scope() as session:
n_in, n_out = 514, 128
layer_name = "clp_layer"
config = Config()
config.update({
"num_outputs": n_out,
"num_inputs": n_in,
"network": {
layer_name: {
"class": "complex_linear_projection",
"nr_of_filters": n_out,
"n_out": n_out,
"is_output_layer": True,
"from": "data:data"
}
}
})
network = TFNetwork(config=config, train_flag=True)
network.construct_from_dict(config.typed_value("network"))
layer = network.layers[layer_name]
assert isinstance(layer, ComplexLinearProjectionLayer)
i_r = np.ones((1, n_in // 2))
i_i = np.ones((1, n_in // 2)) * 0.5
test_input = np.expand_dims(
np.reshape(
np.transpose(
np.reshape(np.concatenate([i_r, i_i], axis=1),
(1, 2, 257)), [0, 2, 1]), (1, 514)), 0)
test_clp_kernel = np.ones((2, n_in // 2, 128))
network.initialize_params(session)
test_clp_output = session.run(
layer.output.placeholder,
feed_dict={
network.get_extern_data('data').placeholder: test_input,
layer._clp_kernel: test_clp_kernel
})
assert test_clp_output[0, 0, 0] - 6.00722122 < 1e-5
def benchmark(lstm_unit, use_gpu):
"""
:param str lstm_unit: e.g. "LSTMBlock", one of LstmCellTypes
:param bool use_gpu:
:return: runtime in seconds of the training itself, excluding initialization
:rtype: float
"""
device = {True: "GPU", False: "CPU"}[use_gpu]
key = "%s:%s" % (device, lstm_unit)
print(">>> Start benchmark for %s." % key)
config = Config()
config.update(make_config_dict(lstm_unit=lstm_unit, use_gpu=use_gpu))
dataset_kwargs = config.typed_value("train")
Dataset.kwargs_update_from_config(config, dataset_kwargs)
dataset = init_dataset(dataset_kwargs)
engine = Engine(config=config)
engine.init_train_from_config(config=config, train_data=dataset)
print(">>> Start training now for %s." % key)
start_time = time.time()
engine.train()
runtime = time.time() - start_time
print(">>> Runtime of %s: %s" % (key, hms_fraction(runtime)))
engine.finalize()
return runtime
def demo():
"""
Demo.
"""
print("SprintDataset demo.")
from argparse import ArgumentParser
from returnn.util.basic import progress_bar_with_time
from returnn.log import log
from returnn.config import Config
from returnn.datasets.basic import init_dataset
arg_parser = ArgumentParser()
arg_parser.add_argument("--config", help="config with ExternSprintDataset", required=True)
arg_parser.add_argument("--sprint_cache_dataset", help="kwargs dict for SprintCacheDataset", required=True)
arg_parser.add_argument("--max_num_seqs", default=sys.maxsize, type=int)
arg_parser.add_argument("--action", default="compare", help="compare or benchmark")
args = arg_parser.parse_args()
log.initialize(verbosity=[4])
sprint_cache_dataset_kwargs = eval(args.sprint_cache_dataset)
assert isinstance(sprint_cache_dataset_kwargs, dict)
sprint_cache_dataset = SprintCacheDataset(**sprint_cache_dataset_kwargs)
print("SprintCacheDataset: %r" % sprint_cache_dataset)
config = Config()
config.load_file(args.config)
dataset = init_dataset(config.typed_value("train"))
print("Dataset via config: %r" % dataset)
assert sprint_cache_dataset.num_inputs == dataset.num_inputs
assert tuple(sprint_cache_dataset.num_outputs["classes"]) == tuple(dataset.num_outputs["classes"])
sprint_cache_dataset.init_seq_order(epoch=1)
if args.action == "compare":
print("Iterating through dataset...")
seq_idx = 0
dataset.init_seq_order(epoch=1)
while seq_idx < args.max_num_seqs:
if not dataset.is_less_than_num_seqs(seq_idx):
break
dataset.load_seqs(seq_idx, seq_idx + 1)
tag = dataset.get_tag(seq_idx)
assert not tag.startswith("seq-"), "dataset does not provide tag-names for seqs"
dataset_seq = sprint_cache_dataset.get_dataset_seq_for_name(tag)
data = dataset.get_data(seq_idx, "data")
targets = dataset.get_data(seq_idx, "classes")
assert data.shape == dataset_seq.features["data"].shape
assert targets.shape == dataset_seq.features["classes"].shape
assert numpy.allclose(data, dataset_seq.features["data"])
assert numpy.allclose(targets, dataset_seq.features["classes"])
seq_idx += 1
progress_bar_with_time(dataset.get_complete_frac(seq_idx))
print("Finished through dataset. Num seqs: %i" % seq_idx)
print("SprintCacheDataset has num seqs: %i." % sprint_cache_dataset.num_seqs)
elif args.action == "benchmark":
print("Iterating through dataset...")
start_time = time.time()
seq_tags = []
seq_idx = 0
dataset.init_seq_order(epoch=1)
while seq_idx < args.max_num_seqs:
if not dataset.is_less_than_num_seqs(seq_idx):
break
dataset.load_seqs(seq_idx, seq_idx + 1)
tag = dataset.get_tag(seq_idx)
assert not tag.startswith("seq-"), "dataset does not provide tag-names for seqs"
seq_tags.append(tag)
dataset.get_data(seq_idx, "data")
dataset.get_data(seq_idx, "classes")
seq_idx += 1
progress_bar_with_time(dataset.get_complete_frac(seq_idx))
print("Finished through dataset. Num seqs: %i, time: %f" % (seq_idx, time.time() - start_time))
print("SprintCacheDataset has num seqs: %i." % sprint_cache_dataset.num_seqs)
if hasattr(dataset, "exit_handler"):
dataset.exit_handler()
else:
print("No way to stop any background tasks.")
del dataset
start_time = time.time()
print("Iterating through SprintCacheDataset...")
for i, tag in enumerate(seq_tags):
sprint_cache_dataset.get_dataset_seq_for_name(tag)
progress_bar_with_time(float(i) / len(seq_tags))
print("Finished through SprintCacheDataset. time: %f" % (time.time() - start_time,))
else:
raise Exception("invalid action: %r" % args.action)
def test_multi_target_init():
config = Config()
config.update({
"multiprocessing": False,
"blocking": True,
"device": "cpu",
"num_epochs": 1,
"num_inputs": 3,
"num_outputs": {
"t1": 4,
"t2": 5
},
"learning_rate": 1.0,
})
config.network_topology_json = """
{
"fw0": {"class": "hidden", "activation": "identity", "n_out": 3},
"out1": {"class": "softmax", "loss": "ce", "target": "t1", "from": ["fw0"]},
"out2": {"class": "softmax", "loss": "ce", "target": "t2", "from": ["fw0"]}
}
"""
device = Device("cpu", config=config, blocking=True)
assert_true(device.trainnet, "train network initialized")
assert_true(device.testnet, "test network initialized")
param_vars = device.trainnet.get_all_params_vars()
print("params:", param_vars)
assert_equal(len(param_vars), 6, "W, b vars for each out, and fw")
num_params = get_num_params(param_vars)
assert_equal(num_params, (3 * 3 + 3) + (3 * 4 + 4) + (3 * 5 + 5),
"W, b for each out, and fw")
assert_in("fw0", device.testnet.hidden)
assert_in("out1", device.testnet.output)
assert_in("out2", device.testnet.output)
assert_is(device.testnet.j["t1"], device.testnet.output["out1"].index)
assert_true(device.updater)
update_list = device.updater.getUpdateList()
print("update list:")
pprint(update_list)
update_dict = dict(update_list)
assert_equal(len(update_dict), len(update_list),
"all params in update list only once")
assert_in("fw0", device.trainnet.hidden)
assert_equal(len(device.trainnet.hidden), 1)
assert_in("W_in_data_fw0", device.trainnet.hidden["fw0"].params)
assert_in("b_fw0", device.trainnet.hidden["fw0"].params)
assert_equal(len(device.trainnet.hidden["fw0"].params), 2)
assert_in("out1", device.trainnet.output)
assert_equal(len(device.trainnet.output), 2)
assert_in("W_in_fw0_out1", device.trainnet.output["out1"].params)
assert_in("b_out1", device.trainnet.output["out1"].params)
assert_equal(len(device.trainnet.output["out1"].params), 2)
assert_in(device.trainnet.hidden["fw0"].params["W_in_data_fw0"],
update_dict)
assert_in(device.trainnet.hidden["fw0"].params["b_fw0"], update_dict)
assert_in(device.trainnet.output["out1"].params["W_in_fw0_out1"],
update_dict)
assert_in(device.trainnet.output["out1"].params["b_out1"], update_dict)
assert_in(device.trainnet.output["out2"].params["W_in_fw0_out2"],
update_dict)
assert_in(device.trainnet.output["out2"].params["b_out2"], update_dict)
# assert_equal(len(update_dict), 6) # updater adds other stuff...
# Set net params.
net_params = {
"fw0": {
"W_in_data_fw0": numpy.identity(3, dtype="float32"),
"b_fw0": numpy.zeros((3, ), dtype="float32")
},
"out1": {
"W_in_fw0_out1":
numpy.arange(0.0, 1.2, 0.1, dtype="float32").reshape((3, 4)),
"b_out1":
numpy.arange(0.0, 4, dtype="float32")
},
"out2": {
"W_in_fw0_out2":
numpy.arange(0.0, 1.5, 0.1, dtype="float32").reshape((3, 5)),
"b_out2":
numpy.arange(0.0, 5, dtype="float32")
}
}
device.trainnet.set_params_by_dict(net_params)
device.testnet.set_params_by_dict(net_params)
# Show params.
for p in param_vars:
print("init %s:" % p)
pprint(p.get_value())
# Init dataset.
dataset = StaticDataset(data=[{
"data":
numpy.array([[0.1, 0.2, -0.3]], dtype="float32"),
"t1":
numpy.array([2]),
"t2":
numpy.array([4])
}],
output_dim=config.typed_value("num_outputs"))
dataset.init_seq_order()
assert_equal(dataset.is_data_sparse("data"), False)
assert_equal(dataset.is_data_sparse("t1"), True)
assert_equal(dataset.is_data_sparse("t2"), True)
# Copy to device allocation.
success = assign_dev_data_single_seq(device, dataset, 0)
assert_true(success, "failed to allocate & assign data")
# Check allocated data.
assert_equal(device.targets["data"].shape,
(1, 1, 3)) # input shape. (time,batch,dim)
assert_in("t1", device.targets)
assert_in("t2", device.targets)
assert_equal(device.targets["t1"].shape, (1, 1))
assert_equal(device.targets["t2"].shape, (1, 1))
assert_equal(device.output_index["data"].shape, (1, 1))
numpy.testing.assert_equal(device.output_index["data"], numpy.array([[1]]))
assert_equal(device.output_index["t1"].shape, (1, 1))
numpy.testing.assert_equal(device.output_index["t1"], numpy.array([[1]]))
# Forward test.
device.update_data()
device.testnet.costs["out1"].name = "out1_cost" # nice in the func graph
out_i1 = device.testnet.output["out1"].index
out_i1_nonzero = device.testnet.output["out1"].i
nll1, pcx1 = T.nnet.crossentropy_softmax_1hot(
x=device.testnet.output["out1"].y_m[out_i1_nonzero],
y_idx=device.testnet.output["out1"].y_data_flat[out_i1_nonzero])
forward_func = theano.function(
inputs=[device.block_start, device.block_end],
outputs=[
device.testnet.j["t1"], out_i1, out_i1_nonzero[0], nll1, pcx1,
device.testnet.costs["out1"],
device.testnet.output["out1"].p_y_given_x,
device.testnet.costs["out2"],
device.testnet.output["out2"].p_y_given_x
],
givens=device.make_givens(device.testnet),
no_default_updates=True,
on_unused_input='warn',
name="forward")
#print "forward func:"
#theano.printing.debugprint(forward_func)
net_j1, out_i1_val, out_i1_nz_val, nll1_val, pcx1_val, t1_cost, t1_y, t2_cost, t2_y = forward_func(
0, 1)
print("forward results:")
pprint(net_j1)
pprint(out_i1_val)
pprint(out_i1_nz_val)
pprint(nll1_val)
pprint(pcx1_val)
pprint(t1_cost)
pprint(t1_y)
pprint(t2_cost)
pprint(t2_y)
assert_equal(net_j1, numpy.array([[1]]))
assert_equal(out_i1_val, numpy.array([[1]]))
assert_equal(out_i1_nz_val, numpy.array([0]))
assert_almost_equal(nll1_val, numpy.array([t1_cost]))
numpy.testing.assert_almost_equal(t1_y, pcx1_val[None, ...])
assert_almost_equal(t1_cost, 1.440189698561195, places=6)
assert_almost_equal(t2_cost, 0.45191439593759336, places=6)
numpy.testing.assert_almost_equal(
t1_y,
numpy.array([[[0.0320586, 0.08714432, 0.23688282, 0.64391426]]]),
decimal=6)
numpy.testing.assert_almost_equal(t2_y,
numpy.array([[[
0.01165623, 0.03168492, 0.08612854,
0.23412166, 0.63640865
]]]),
decimal=6)
# One train step.
device.set_learning_rate(config.typed_value("learning_rate"))
device.run("train")
output_list, outputs_format = device.result()
assert_is_instance(output_list, list)
assert_true(outputs_format, "for train, we should always get the format")
outputs = Device.make_result_dict(output_list, outputs_format)
pprint(outputs)
assert_in("cost:out1", outputs)
assert_greater(outputs["cost:out1"], 0)
assert_almost_equal(outputs["cost:out1"], t1_cost)
# Get net params.
params = device.get_net_train_params(device.trainnet)
references_params = {
"W_in_data_fw0":
numpy.array([[1.00055406e+00, 5.54056978e-04, 5.54056978e-04],
[1.10811396e-03, 1.00110811e+00, 1.10811396e-03],
[-1.66217093e-03, -1.66217093e-03, 9.98337829e-01]]),
"b_fw0":
numpy.array([0.00554057, 0.00554057, 0.00554057]),
"W_in_fw0_out1":
numpy.array([[-0.00320586, 0.09128557, 0.27631172, 0.23560857],
[0.39358828, 0.48257114, 0.75262344, 0.57121715],
[0.80961758, 0.9261433, 0.77106485, 1.29317428]]),
"b_out1":
numpy.array([-0.0320586, 0.91285568, 2.76311718, 2.35608574]),
"W_in_fw0_out2":
numpy.array([[
-1.16562310e-03, 9.68315079e-02, 1.91387146e-01, 2.76587834e-01,
4.36359135e-01
],
[
4.97668754e-01, 5.93663016e-01, 6.82774291e-01,
7.53175669e-01, 9.72718271e-01
],
[
1.00349687e+00, 1.10950548e+00, 1.22583856e+00,
1.37023650e+00, 1.29092259e+00
]]),
"b_out2":
numpy.array(
[-0.01165623, 0.96831508, 1.91387146, 2.76587834, 4.36359135])
}
assert_equal(len(param_vars), len(params))
for p, v in zip(param_vars, params):
print("%s:" % p)
pprint(v)
assert_true(p.name)
numpy.testing.assert_almost_equal(references_params[p.name],
v,
decimal=6)
def test_combi_auto_enc_longer():
config = Config()
config.update({
"multiprocessing": False,
"blocking": True,
"device": "cpu",
"num_epochs": 1,
"num_inputs": 3,
"num_outputs": {
"classes": 2
},
"learning_rate": 1.0,
"adadelta": True,
"network": {
"output": {
"class": "softmax",
"loss": "ce",
"target": "classes"
},
"auto-enc": {
"class": "softmax",
"loss": "sse",
"dtype": "float32",
"target": "data"
}
}
})
device = Device("cpu", config=config, blocking=True)
# Set net params.
def get_net_params(with_auto_enc=True):
d = {
"output": {
"W_in_data_output":
numpy.arange(0.1, 0.7, 0.1, dtype="float32").reshape((3, 2)),
"b_output":
numpy.arange(0.0, 2, dtype="float32")
}
}
if with_auto_enc:
d["auto-enc"] = {
"W_in_data_auto-enc":
numpy.arange(0.1, 1.0, 0.1, dtype="float32").reshape((3, 3)),
"b_auto-enc":
numpy.arange(0.0, 3, dtype="float32")
}
return d
device.trainnet.set_params_by_dict(get_net_params())
device.testnet.set_params_by_dict(get_net_params())
# Show params.
for p in device.trainnet.get_all_params_vars():
print("init %s:" % p)
pprint(p.get_value())
# Init dataset.
dataset = DummyDataset(input_dim=config.typed_value("num_inputs"),
output_dim=config.typed_value("num_outputs"),
num_seqs=10)
dataset.init_seq_order()
cost_output_sum = 0.0
for seq_idx in range(dataset.num_seqs):
# Copy to device allocation.
success = assign_dev_data_single_seq(device, dataset, seq_idx)
assert_true(success, "failed to allocate & assign data")
# One train step.
device.set_learning_rate(config.typed_value("learning_rate"))
device.run("train")
output_list, outputs_format = device.result()
assert_is_instance(output_list, list)
assert_true(outputs_format,
"for train, we should always get the format")
outputs = Device.make_result_dict(output_list, outputs_format)
print(("seq %i" % seq_idx))
pprint(outputs)
assert_in("cost:output", outputs)
assert_in("cost:auto-enc", outputs)
cost_output_sum += outputs["cost:output"]
# Now, drop the auto-enc from the network, and redo the same thing.
del config.typed_value("network")["auto-enc"]
device = Device("cpu", config=config, blocking=True)
device.trainnet.set_params_by_dict(get_net_params(with_auto_enc=False))
device.testnet.set_params_by_dict(get_net_params(with_auto_enc=False))
for p in device.trainnet.get_all_params_vars():
print("second run, init %s:" % p)
pprint(p.get_value())
dataset.init_seq_order() # reset
cost2_output_sum = 0.0
for seq_idx in range(dataset.num_seqs):
# Copy to device allocation.
success = assign_dev_data_single_seq(device, dataset, seq_idx)
assert_true(success, "failed to allocate & assign data")
# One train step.
device.set_learning_rate(config.typed_value("learning_rate"))
device.run("train")
output_list, outputs_format = device.result()
assert_is_instance(output_list, list)
assert_true(outputs_format,
"for train, we should always get the format")
outputs = Device.make_result_dict(output_list, outputs_format)
print(("seq %i" % seq_idx))
pprint(outputs)
assert_in("cost:output", outputs)
assert_not_in("cost:auto-enc", outputs)
cost2_output_sum += outputs["cost:output"]
assert_equal(cost_output_sum, cost2_output_sum)
assert_almost_equal(cost_output_sum, 16.028842568397522, places=6)
def test_combi_auto_enc():
config = Config()
config.update({
"multiprocessing": False,
"blocking": True,
"device": "cpu",
"num_epochs": 1,
"num_inputs": 3,
"num_outputs": {
"classes": 2
},
"learning_rate": 1.0,
"network": {
"output": {
"class": "softmax",
"loss": "ce",
"target": "classes"
},
"auto-enc": {
"class": "softmax",
"loss": "sse",
"dtype": "float32",
"target": "data"
}
}
})
device = Device("cpu", config=config, blocking=True)
# Set net params.
def get_net_params(with_auto_enc=True):
d = {
"output": {
"W_in_data_output":
numpy.arange(0.1, 0.7, 0.1, dtype="float32").reshape((3, 2)),
"b_output":
numpy.arange(0.0, 2, dtype="float32")
}
}
if with_auto_enc:
d["auto-enc"] = {
"W_in_data_auto-enc":
numpy.arange(0.1, 1.0, 0.1, dtype="float32").reshape((3, 3)),
"b_auto-enc":
numpy.arange(0.0, 3, dtype="float32")
}
return d
device.trainnet.set_params_by_dict(get_net_params())
device.testnet.set_params_by_dict(get_net_params())
# Show params.
for p in device.trainnet.get_all_params_vars():
print("init %s:" % p)
pprint(p.get_value())
# Init dataset.
dataset = StaticDataset(data=[{
"data":
numpy.array([[0.1, 0.2, -0.3]], dtype="float32"),
"classes":
numpy.array([1]),
}],
output_dim=config.typed_value("num_outputs"))
dataset.init_seq_order()
# Copy to device allocation.
success = assign_dev_data_single_seq(device, dataset, 0)
assert_true(success, "failed to allocate & assign data")
# One train step.
device.set_learning_rate(config.typed_value("learning_rate"))
device.run("train")
output_list, outputs_format = device.result()
assert_is_instance(output_list, list)
assert_true(outputs_format, "for train, we should always get the format")
outputs = Device.make_result_dict(output_list, outputs_format)
pprint(outputs)
assert_in("cost:output", outputs)
assert_in("cost:auto-enc", outputs)
expected_cost_output = 0.3132616877555847
assert_almost_equal(outputs["cost:output"], expected_cost_output, places=6)
exact_cost_output = outputs["cost:output"]
assert_almost_equal(outputs["cost:auto-enc"], 1.7544001340866089, places=6)
# Now, drop the auto-enc from the network, and redo the same thing.
del config.typed_value("network")["auto-enc"]
device = Device("cpu", config=config, blocking=True)
device.trainnet.set_params_by_dict(get_net_params(with_auto_enc=False))
device.testnet.set_params_by_dict(get_net_params(with_auto_enc=False))
for p in device.trainnet.get_all_params_vars():
print("second run, init %s:" % p)
pprint(p.get_value())
dataset.init_seq_order() # reset. probably not needed
success = assign_dev_data_single_seq(device, dataset, 0)
assert_true(success, "failed to allocate & assign data")
device.set_learning_rate(config.typed_value("learning_rate"))
device.run("train")
output_list, outputs_format = device.result()
assert_is_instance(output_list, list)
assert_true(outputs_format, "for train, we should always get the format")
outputs = Device.make_result_dict(output_list, outputs_format)
pprint(outputs)
assert_in("cost:output", outputs)
assert_not_in("cost:auto-enc", outputs)
assert_almost_equal(outputs["cost:output"], expected_cost_output, places=6)
assert_equal(outputs["cost:output"], exact_cost_output)
