def test_train_iters(tf_graph, mocker, capfd):
params = get_params()
pd1 = PerceptualDiscrimination(dt=params['dt'],
tau=params['tau'],
T=2000,
N_batch=params['N_batch'])
gen1 = pd1.batch_generator()
for remainder in [0, 1]:
train_params = {}
train_params['training_iters'] = params[
'N_batch'] + remainder # number of iterations to train for Default: 10000
train_params[
'loss_epoch'] = 1 # Compute and record loss every 'loss_epoch' epochs. Default: 10
train_params['verbosity'] = True
rnn = RNN(params)
mocker.patch.object(RNN, 'forward_pass')
RNN.forward_pass.return_value = tf.fill(
[params['N_batch'], params['N_steps'], params['N_out']],
float('nan')), tf.fill(
[params['N_batch'], params['N_steps'], params['N_rec']],
float('nan'))
rnn.train(gen1, train_params)
rnn.destruct()
out, _ = capfd.readouterr()
N_epochs = int(
ceil(train_params['training_iters'] / (params['N_batch'] * 1.0)))
assert "Iter " + str(N_epochs * params['N_batch']) in out
assert "Iter " + str((N_epochs + 1) * params['N_batch']) not in out
def test_train(tf_graph, mocker, capfd):
rdm = rd.RDM(dt=10, tau=100, T=2000, N_batch=128)
gen = rdm.batch_generator()
params = get_params()
rnn = RNN(params)
with pytest.raises(UserWarning) as excinfo:
rnn.train(gen)
assert 'build' in str(excinfo.value)
mocker.patch.object(RNN, 'forward_pass')
RNN.forward_pass.return_value = tf.fill(
[params['N_batch'], params['N_steps'], params['N_out']],
float('nan')), tf.fill(
[params['N_batch'], params['N_steps'], params['N_rec']],
float('nan'))
rnn.build()
rdm1 = rd.RDM(dt=params['dt'],
tau=params['tau'],
T=2000,
N_batch=params['N_batch'])
gen1 = rdm1.batch_generator()
assert rnn.is_initialized is False
rnn.train(gen1)
assert rnn.is_initialized is True
out, _ = capfd.readouterr()
assert out != ""
def test_train(tf_graph, mocker, capfd):
pd = PerceptualDiscrimination(dt=10, tau=100, T=2000, N_batch=128)
gen = pd.batch_generator()
params = get_params()
rnn = RNN(params)
mocker.patch.object(RNN, 'forward_pass')
RNN.forward_pass.return_value = tf.fill(
[params['N_batch'], params['N_steps'], params['N_out']],
float('nan')), tf.fill(
[params['N_batch'], params['N_steps'], params['N_rec']],
float('nan'))
pd1 = PerceptualDiscrimination(dt=params['dt'],
tau=params['tau'],
T=2000,
N_batch=params['N_batch'])
gen1 = pd1.batch_generator()
assert rnn.is_initialized is False
assert rnn.is_built is False
rnn.train(gen1)
assert rnn.is_initialized is True
assert rnn.is_built is True
out, _ = capfd.readouterr()
assert out != ""
def test_save(tf_graph, mocker, tmpdir):
save_weights_path = str(tmpdir.dirpath("save_weights.npz"))
params = get_params()
rnn = RNN(params)
with pytest.raises(UserWarning) as excinfo:
rnn.save(save_weights_path)
assert "No weights to return" in str(excinfo.value)
mocker.patch.object(RNN, 'forward_pass')
RNN.forward_pass.return_value = tf.fill(
[params['N_batch'], params['N_steps'], params['N_out']],
float('nan')), tf.fill(
[params['N_batch'], params['N_steps'], params['N_rec']],
float('nan'))
rnn.build()
rdm1 = rd.RDM(dt=params['dt'],
tau=params['tau'],
T=2000,
N_batch=params['N_batch'])
gen1 = rdm1.batch_generator()
rnn.train(gen1)
assert not tmpdir.dirpath("save_weights.npz").check(exists=1)
rnn.save(save_weights_path)
assert tmpdir.dirpath("save_weights.npz").check(exists=1)
tmpdir.dirpath("save_weights.npz").remove()
def test_get_weights(tf_graph, mocker):
params = get_params()
rnn = RNN(params)
with pytest.raises(UserWarning) as excinfo:
rnn.get_weights()
assert 'No weights to return yet -- model has not yet been initialized.' in str(
excinfo.value)
mocker.patch.object(RNN, 'forward_pass')
RNN.forward_pass.return_value = tf.fill(
[params['N_batch'], params['N_steps'], params['N_out']],
float('nan')), tf.fill(
[params['N_batch'], params['N_steps'], params['N_rec']],
float('nan'))
rnn.build()
with pytest.raises(UserWarning) as excinfo:
rnn.get_weights()
assert 'No weights to return yet -- model has not yet been initialized.' in str(
excinfo.value)
rdm1 = rd.RDM(dt=params['dt'],
tau=params['tau'],
T=2000,
N_batch=params['N_batch'])
gen1 = rdm1.batch_generator()
rnn.train(gen1)
assert type(rnn.get_weights()) is dict
def test_load_weights_path_rnn(tf_graph, mocker, tmpdir, capfd):
params = get_params()
pd1 = PerceptualDiscrimination(dt=params['dt'],
tau=params['tau'],
T=2000,
N_batch=params['N_batch'])
gen1 = pd1.batch_generator()
mocker.patch.object(RNN, 'forward_pass')
RNN.forward_pass.return_value = tf.fill(
[params['N_batch'], params['N_steps'], params['N_out']],
float('nan')), tf.fill(
[params['N_batch'], params['N_steps'], params['N_rec']],
float('nan'))
rnn = RNN(params)
rnn.build()
train_params = {}
train_params['save_weights_path'] = str(
tmpdir.dirpath("save_weights.npz")
) # Where to save the model after training. Default: None
train_params['verbosity'] = False
### save out some weights to test with and destroy the rnn that created them
assert not tmpdir.dirpath("save_weights.npz").check(exists=1)
rnn.train(gen1, train_params)
assert rnn.is_initialized is True
out, _ = capfd.readouterr()
print(out)
assert out == ""
assert tmpdir.dirpath("save_weights.npz").check(exists=1)
rnn.destruct()
### Make sure loading weights fails on nonexistent file
params['load_weights_path'] = "nonexistent"
with pytest.raises(EnvironmentError) as excinfo:
rnn = RNN(params)
assert "No such file" in str(excinfo.value)
rnn.destruct()
### Ensure success when loading weights created previously
params['load_weights_path'] = str(tmpdir.dirpath("save_weights.npz"))
rnn = RNN(params)
tmpdir.dirpath("save_weights.npz").remove()
def test_save(tf_graph, mocker, tmpdir):
save_weights_path = str(tmpdir.dirpath("save_weights.npz"))
params = get_params()
rnn = RNN(params)
mocker.patch.object(RNN, 'forward_pass')
RNN.forward_pass.return_value = tf.fill([params['N_batch'], params['N_steps'], params['N_out']], float('nan')), tf.fill([params['N_batch'], params['N_steps'], params['N_rec']], float('nan'))
pd1 = PerceptualDiscrimination(dt = params['dt'], tau = params['tau'], T = 2000, N_batch = params['N_batch'])
gen1 = pd1.batch_generator()
rnn.train(gen1)
assert not tmpdir.dirpath("save_weights.npz").check(exists=1)
rnn.save(save_weights_path)
assert tmpdir.dirpath("save_weights.npz").check(exists=1)
tmpdir.dirpath("save_weights.npz").remove()
def test_train_train_params_file_creation(tf_graph, mocker, tmpdir, capfd):
params = get_params()
pd1 = PerceptualDiscrimination(dt = params['dt'], tau = params['tau'], T = 2000, N_batch = params['N_batch'])
gen1 = pd1.batch_generator()
pd2 = PerceptualDiscrimination(dt = params['dt'], tau = params['tau'], T = 1000, N_batch = params['N_batch'])
gen2 = pd2.batch_generator()
mocker.patch.object(RNN, 'forward_pass')
RNN.forward_pass.return_value = tf.fill([params['N_batch'], params['N_steps'], params['N_out']], float('nan')), tf.fill([params['N_batch'], params['N_steps'], params['N_rec']], float('nan'))
rnn =RNN(params)
rnn.build()
train_params = {}
train_params['save_weights_path'] = str(tmpdir.dirpath("save_weights.npz")) # Where to save the model after training. Default: None
train_params['training_iters'] = 1000 # number of iterations to train for Default: 10000
train_params['learning_rate'] = .01 # Sets learning rate if use default optimizer Default: .001
train_params['loss_epoch'] = 20 # Compute and recopd loss every 'loss_epoch' epochs. Default: 10
train_params['verbosity'] = False
train_params['save_training_weights_epoch'] = 10 # save training weights every 'save_training_weights_epoch' epochs. Default: 100
train_params['training_weights_path'] = str(tmpdir.dirpath("training_weights")) # where to save training weights as training progresses. Default: None
train_params['generator_function'] = gen2 # replaces trial_batch_generator with the generator_function when not none. Default: None
train_params['optimizer'] = tf.compat.v1.train.AdamOptimizer(learning_rate=train_params['learning_rate']) # What optimizer to use to compute gradients. Default: tf.train.AdamOptimizer(learning_rate=train_params['learning_rate'])
train_params['clip_grads'] = False # If true, clip gradients by norm 1. Default: True
assert not tmpdir.dirpath("save_weights.npz").check(exists=1)
assert not tmpdir.dirpath("training_weights" + str(train_params['save_training_weights_epoch'])).check(exists=1)
rnn.train(gen1, train_params)
assert rnn.is_initialized is True
out, _ = capfd.readouterr()
print(out)
assert out==""
assert tmpdir.dirpath("save_weights.npz").check(exists=1)
assert tmpdir.dirpath("training_weights" + str(train_params['save_training_weights_epoch'])+ ".npz").check(exists=1)