def test_sequence_generator(): # Disclaimer: here we only check shapes, not values. output_dim = 1 dim = 20 batch_size = 30 n_steps = 10 transition = GatedRecurrent( name="transition", activation=Tanh(), dim=dim, weights_init=Orthogonal()) generator = SequenceGenerator( LinearReadout(readout_dim=output_dim, source_names=["states"], emitter=TestEmitter(name="emitter"), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.initialize() y = tensor.tensor3('y') mask = tensor.matrix('mask') costs = generator.cost(y, mask) assert costs.ndim == 2 costs_val = theano.function([y, mask], [costs])( numpy.zeros((n_steps, batch_size, output_dim), dtype=floatX), numpy.ones((n_steps, batch_size), dtype=floatX))[0] assert costs_val.shape == (n_steps, batch_size) states, outputs, costs = [variable.eval() for variable in generator.generate( iterate=True, batch_size=batch_size, n_steps=n_steps)] assert states.shape == (n_steps, batch_size, dim) assert outputs.shape == (n_steps, batch_size, output_dim) assert costs.shape == (n_steps, batch_size)
def test_sequence_generator(): # Disclaimer: here we only check shapes, not values. output_dim = 1 dim = 20 batch_size = 30 n_steps = 10 transition = GatedRecurrent( name="transition", activation=Tanh(), dim=dim, weights_init=Orthogonal()) generator = SequenceGenerator( LinearReadout(readout_dim=output_dim, source_names=["states"], emitter=TestEmitter(name="emitter"), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.initialize() y = tensor.tensor3('y') mask = tensor.matrix('mask') costs = generator.cost(y, mask) assert costs.ndim == 2 costs_val = theano.function([y, mask], [costs])( numpy.zeros((n_steps, batch_size, output_dim), dtype=floatX), numpy.ones((n_steps, batch_size), dtype=floatX))[0] assert costs_val.shape == (n_steps, batch_size) states, outputs, costs = [variable.eval() for variable in generator.generate( iterate=True, batch_size=batch_size, n_steps=n_steps)] assert states.shape == (n_steps, batch_size, dim) assert outputs.shape == (n_steps, batch_size, output_dim) assert costs.shape == (n_steps, batch_size)
def test_integer_sequence_generator(): # Disclaimer: here we only check shapes, not values. readout_dim = 5 feedback_dim = 3 dim = 20 batch_size = 30 n_steps = 10 transition = GatedRecurrent(name="transition", activation=Tanh(), dim=dim, weights_init=Orthogonal()) generator = SequenceGenerator(LinearReadout( readout_dim=readout_dim, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback(readout_dim, feedback_dim), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.initialize() y = tensor.lmatrix('y') mask = tensor.matrix('mask') costs = generator.cost(y, mask) assert costs.ndim == 2 costs_val = theano.function([y, mask], [costs])(numpy.zeros((n_steps, batch_size), dtype='int64'), numpy.ones((n_steps, batch_size), dtype=floatX))[0] assert costs_val.shape == (n_steps, batch_size) states, outputs, costs = generator.generate(iterate=True, batch_size=batch_size, n_steps=n_steps) states_val, outputs_val, costs_val = theano.function( [], [states, outputs, costs], updates=costs.owner.inputs[0].owner.tag.updates)() assert states_val.shape == (n_steps, batch_size, dim) assert outputs_val.shape == (n_steps, batch_size) assert outputs_val.dtype == 'int64' assert costs_val.shape == (n_steps, batch_size)
def test_integer_sequence_generator(): # Disclaimer: here we only check shapes, not values. readout_dim = 5 feedback_dim = 3 dim = 20 batch_size = 30 n_steps = 10 transition = GatedRecurrent( name="transition", activation=Tanh(), dim=dim, weights_init=Orthogonal()) generator = SequenceGenerator( LinearReadout(readout_dim=readout_dim, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback(readout_dim, feedback_dim), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.initialize() y = tensor.lmatrix('y') mask = tensor.matrix('mask') costs = generator.cost(y, mask) assert costs.ndim == 2 costs_val = theano.function([y, mask], [costs])( numpy.zeros((n_steps, batch_size), dtype='int64'), numpy.ones((n_steps, batch_size), dtype=floatX))[0] assert costs_val.shape == (n_steps, batch_size) states, outputs, costs = generator.generate( iterate=True, batch_size=batch_size, n_steps=n_steps) states_val, outputs_val, costs_val = theano.function( [], [states, outputs, costs], updates=costs.owner.inputs[0].owner.tag.updates)() assert states_val.shape == (n_steps, batch_size, dim) assert outputs_val.shape == (n_steps, batch_size) assert outputs_val.dtype == 'int64' assert costs_val.shape == (n_steps, batch_size)
def test_recurrentstack_sequence_generator(): """Test RecurrentStack behaviour inside a SequenceGenerator. """ floatX = theano.config.floatX rng = numpy.random.RandomState(1234) output_dim = 1 dim = 20 batch_size = 30 n_steps = 10 depth=2 transitions = [LSTM(dim=dim) for _ in range(depth)] transition = RecurrentStack(transitions,fast=True, weights_init=Constant(2), biases_init=Constant(0)) generator = SequenceGenerator( Readout(readout_dim=output_dim, source_names=["states_%d"%(depth-1)], emitter=TestEmitter()), transition, weights_init=IsotropicGaussian(0.1), biases_init=Constant(0.0), seed=1234) generator.initialize() y = tensor.tensor3('y') cost = generator.cost(y) # Check that all states can be accessed and not just the state connected # to readout. cg = ComputationGraph(cost) from blocks.roles import INPUT, OUTPUT dropout_target = VariableFilter(roles=[INNER_OUTPUT], # bricks=transitions, # name_regex='*' )(cg.variables) assert_equal(len(dropout_target), depth)
def main(): logging.basicConfig( level=logging.DEBUG, format="%(asctime)s: %(name)s: %(levelname)s: %(message)s") parser = argparse.ArgumentParser( "Case study of generating simple 1d sequences with RNN.", formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument( "mode", choices=["train", "plot"], help="The mode to run. Use `train` to train a new model" " and `plot` to plot a sequence generated by an" " existing one.") parser.add_argument( "prefix", default="sine", help="The prefix for model, timing and state files") parser.add_argument( "--input-noise", type=float, default=0.0, help="Adds Gaussian noise of given intensity to the " " training sequences.") parser.add_argument( "--function", default="lambda a, x: numpy.sin(a * x)", help="An analytical description of the sequence family to learn." " The arguments before the last one are considered parameters.") parser.add_argument( "--steps", type=int, default=100, help="Number of steps to plot") parser.add_argument( "--params", help="Parameter values for plotting") args = parser.parse_args() function = eval(args.function) num_params = len(inspect.getargspec(function).args) - 1 class Emitter(TrivialEmitter): @application def cost(self, readouts, outputs): """Compute MSE.""" return ((readouts - outputs) ** 2).sum(axis=readouts.ndim - 1) transition = GatedRecurrent( name="transition", activation=Tanh(), dim=10, weights_init=Orthogonal()) with_params = AddParameters(transition, num_params, "params", name="with_params") generator = SequenceGenerator( LinearReadout(readout_dim=1, source_names=["states"], emitter=Emitter(name="emitter"), name="readout"), with_params, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.allocate() logger.debug("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in Selector(generator).get_params().items()], width=120)) if args.mode == "train": seed = 1 rng = numpy.random.RandomState(seed) batch_size = 10 generator.initialize() cost = ComputationGraph( generator.cost(tensor.tensor3('x'), params=tensor.matrix("params")).sum()) cost = apply_noise(cost, cost.inputs, args.input_noise) gh_model = GroundhogModel(generator, cost) state = GroundhogState(args.prefix, batch_size, learning_rate=0.0001).as_dict() data = SeriesIterator(rng, function, 100, batch_size) trainer = SGD(gh_model, state, data) main_loop = MainLoop(data, None, None, gh_model, trainer, state, None) main_loop.load() main_loop.main() elif args.mode == "plot": load_params(generator, args.prefix + "model.npz") params = tensor.matrix("params") sample = theano.function([params], generator.generate( params=params, n_steps=args.steps, batch_size=1)) param_values = numpy.array(map(float, args.params.split()), dtype=floatX) states, outputs, _ = sample(param_values[None, :]) actual = outputs[:, 0, 0] desired = numpy.array([function(*(list(param_values) + [T])) for T in range(args.steps)]) print("MSE: {}".format(((actual - desired) ** 2).sum())) pyplot.plot(numpy.hstack([actual[:, None], desired[:, None]])) pyplot.show() else: assert False
class Decoder(Initializable): def __init__(self, vocab_size, embedding_dim, state_dim, representation_dim, **kwargs): super(Decoder, self).__init__(**kwargs) self.vocab_size = vocab_size self.embedding_dim = embedding_dim self.state_dim = state_dim self.representation_dim = representation_dim readout = Readout( source_names=['states', 'feedback', 'readout_context'], readout_dim=self.vocab_size, emitter=SoftmaxEmitter(), feedback_brick=LookupFeedback(vocab_size, embedding_dim), post_merge=InitializableFeedforwardSequence([ Bias(dim=1000).apply, Maxout(num_pieces=2).apply, Linear(input_dim=state_dim / 2, output_dim=100, use_bias=False).apply, Linear(input_dim=100).apply ]), merged_dim=1000) self.transition = GatedRecurrentWithContext(Tanh(), dim=state_dim, name='decoder') # Readout will apply the linear transformation to 'readout_context' # with a Merge brick, so no need to fork it here self.fork = Fork([ name for name in self.transition.apply.contexts + self.transition.apply.states if name != 'readout_context' ], prototype=Linear()) self.tanh = Tanh() self.sequence_generator = SequenceGenerator( readout=readout, transition=self.transition, fork_inputs=[ name for name in self.transition.apply.sequences if name != 'mask' ], ) self.children = [self.fork, self.sequence_generator, self.tanh] def _push_allocation_config(self): self.fork.input_dim = self.representation_dim self.fork.output_dims = [ self.state_dim for _ in self.fork.output_names ] @application( inputs=['representation', 'target_sentence_mask', 'target_sentence'], outputs=['cost']) def cost(self, representation, target_sentence, target_sentence_mask): target_sentence = target_sentence.dimshuffle(1, 0) target_sentence_mask = target_sentence_mask.T # The initial state and contexts, all functions of the representation contexts = { key: value.dimshuffle('x', 0, 1) if key not in self.transition.apply.states else value for key, value in self.fork.apply(representation, as_dict=True).items() } contexts['states'] = self.tanh.apply(contexts['states']) cost = self.sequence_generator.cost(**merge( contexts, { 'mask': target_sentence_mask, 'outputs': target_sentence, 'readout_context': representation.dimshuffle('x', 0, 1) })) return (cost * target_sentence_mask).sum() / target_sentence_mask.shape[1]
def test_attention_transition(): inp_dim = 2 inp_len = 10 attended_dim = 3 attended_len = 11 batch_size = 4 n_steps = 30 transition = TestTransition(dim=inp_dim, attended_dim=attended_dim, name="transition") attention = SequenceContentAttention(transition.apply.states, match_dim=inp_dim, name="attention") mixer = Mixer( [name for name in transition.apply.sequences if name != 'mask'], attention.take_look.outputs[0], name="mixer") att_trans = AttentionTransition(transition, attention, mixer, name="att_trans") att_trans.weights_init = IsotropicGaussian(0.01) att_trans.biases_init = Constant(0) att_trans.initialize() attended = tensor.tensor3("attended") attended_mask = tensor.matrix("attended_mask") inputs = tensor.tensor3("inputs") inputs_mask = tensor.matrix("inputs_mask") states, glimpses, weights = att_trans.apply(input_=inputs, mask=inputs_mask, attended=attended, attended_mask=attended_mask) assert states.ndim == 3 assert glimpses.ndim == 3 assert weights.ndim == 3 input_vals = numpy.zeros((inp_len, batch_size, inp_dim), dtype=floatX) input_mask_vals = numpy.ones((inp_len, batch_size), dtype=floatX) attended_vals = numpy.zeros((attended_len, batch_size, attended_dim), dtype=floatX) attended_mask_vals = numpy.ones((attended_len, batch_size), dtype=floatX) func = theano.function([inputs, inputs_mask, attended, attended_mask], [states, glimpses, weights]) states_vals, glimpses_vals, weight_vals = func(input_vals, input_mask_vals, attended_vals, attended_mask_vals) assert states_vals.shape == input_vals.shape assert glimpses_vals.shape == (inp_len, batch_size, attended_dim) assert weight_vals.shape == (inp_len, batch_size, attended_len) # Test SequenceGenerator using AttentionTransition generator = SequenceGenerator(LinearReadout( readout_dim=inp_dim, source_names=["state"], emitter=TestEmitter(name="emitter"), name="readout"), transition=transition, attention=attention, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") outputs = tensor.tensor3('outputs') costs = generator.cost(outputs, attended=attended, attended_mask=attended_mask) costs_vals = costs.eval({ outputs: input_vals, attended: attended_vals, attended_mask: attended_mask_vals }) assert costs_vals.shape == (inp_len, batch_size) results = (generator.generate(n_steps=n_steps, batch_size=attended.shape[1], attended=attended, attended_mask=attended_mask)) assert len(results) == 5 states_vals, outputs_vals, glimpses_vals, weights_vals, costs_vals = ( theano.function([attended, attended_mask], results)(attended_vals, attended_mask_vals)) assert states_vals.shape == (n_steps, batch_size, inp_dim) assert states_vals.shape == outputs_vals.shape assert glimpses_vals.shape == (n_steps, batch_size, attended_dim) assert weights_vals.shape == (n_steps, batch_size, attended_len) assert costs_vals.shape == (n_steps, batch_size)
def test_sequence_generator(): """Test a sequence generator with no contexts and continuous outputs. Such sequence generators can be used to model e.g. dynamical systems. """ rng = numpy.random.RandomState(1234) output_dim = 1 dim = 20 batch_size = 30 n_steps = 10 transition = SimpleRecurrent(activation=Tanh(), dim=dim, weights_init=Orthogonal()) generator = SequenceGenerator( Readout(readout_dim=output_dim, source_names=["states"], emitter=TestEmitter()), transition, weights_init=IsotropicGaussian(0.1), biases_init=Constant(0.0), seed=1234) generator.initialize() # Test 'cost_matrix' method y = tensor.tensor3('y') mask = tensor.matrix('mask') costs = generator.cost_matrix(y, mask) assert costs.ndim == 2 y_test = rng.uniform(size=(n_steps, batch_size, output_dim)).astype(floatX) m_test = numpy.ones((n_steps, batch_size), dtype=floatX) costs_val = theano.function([y, mask], [costs])(y_test, m_test)[0] assert costs_val.shape == (n_steps, batch_size) assert_allclose(costs_val.sum(), 115.593, rtol=1e-5) # Test 'cost' method cost = generator.cost(y, mask) assert cost.ndim == 0 cost_val = theano.function([y, mask], [cost])(y_test, m_test) assert_allclose(cost_val, 3.8531, rtol=1e-5) # Test 'AUXILIARY' variable 'per_sequence_element' in 'cost' method cg = ComputationGraph([cost]) var_filter = VariableFilter(roles=[AUXILIARY]) aux_var_name = '_'.join([generator.name, generator.cost.name, 'per_sequence_element']) cost_per_el = [el for el in var_filter(cg.variables) if el.name == aux_var_name][0] assert cost_per_el.ndim == 0 cost_per_el_val = theano.function([y, mask], [cost_per_el])(y_test, m_test) assert_allclose(cost_per_el_val, 0.38531, rtol=1e-5) # Test 'generate' method states, outputs, costs = [variable.eval() for variable in generator.generate( states=rng.uniform( size=(batch_size, dim)).astype(floatX), iterate=True, batch_size=batch_size, n_steps=n_steps)] assert states.shape == (n_steps, batch_size, dim) assert outputs.shape == (n_steps, batch_size, output_dim) assert costs.shape == (n_steps, batch_size) assert_allclose(outputs.sum(), -0.33683, rtol=1e-5) assert_allclose(states.sum(), 15.7909, rtol=1e-5) # There is no generation cost in this case, since generation is # deterministic assert_allclose(costs.sum(), 0.0)
def main(name, epochs, batch_size, learning_rate, dim, mix_dim, old_model_name, max_length, bokeh, GRU, dropout, depth, max_grad, step_method, epsilon, sample, skip, uniform, top): #---------------------------------------------------------------------- datasource = name def shnum(x): """ Convert a positive float into a short tag-usable string E.g.: 0 -> 0, 0.005 -> 53, 100 -> 1-2 """ return '0' if x <= 0 else '%s%d' % (("%e"%x)[0], -np.floor(np.log10(x))) jobname = "%s-%dX%dm%dd%dr%sb%de%s" % (datasource, depth, dim, mix_dim, int(dropout*10), shnum(learning_rate), batch_size, shnum(epsilon)) if max_length != 600: jobname += '-L%d'%max_length if GRU: jobname += 'g' if max_grad != 5.: jobname += 'G%g'%max_grad if step_method != 'adam': jobname += step_method if skip: jobname += 'D' assert depth > 1 if top: jobname += 'T' assert depth > 1 if uniform > 0.: jobname += 'u%d'%int(uniform*100) if debug: jobname += ".debug" if sample: print("Sampling") else: print("\nRunning experiment %s" % jobname) if old_model_name: print("starting from model %s"%old_model_name) #---------------------------------------------------------------------- transitions = [GatedRecurrent(dim=dim) if GRU else LSTM(dim=dim) for _ in range(depth)] if depth > 1: transition = RecurrentStack(transitions, name="transition", fast=True, skip_connections=skip or top) if skip: source_names=['states'] + ['states_%d'%d for d in range(1,depth)] else: source_names=['states_%d'%(depth-1)] else: transition = transitions[0] transition.name = "transition" source_names=['states'] emitter = SketchEmitter(mix_dim=mix_dim, epsilon=epsilon, name="emitter") readout = Readout( readout_dim=emitter.get_dim('inputs'), source_names=source_names, emitter=emitter, name="readout") normal_inputs = [name for name in transition.apply.sequences if 'mask' not in name] fork = Fork(normal_inputs, prototype=Linear(use_bias=True)) generator = SequenceGenerator(readout=readout, transition=transition, fork=fork) # Initialization settings if uniform > 0.: generator.weights_init = Uniform(width=uniform*2.) else: generator.weights_init = OrthogonalGlorot() generator.biases_init = Constant(0) # Build the cost computation graph [steps, batch_size, 3] x = T.tensor3('features', dtype=floatX) if debug: x.tag.test_value = np.ones((max_length,batch_size,3)).astype(floatX) x = x[:max_length,:,:] # has to be after setting test_value cost = generator.cost(x) cost.name = "sequence_log_likelihood" # Give an idea of what's going on model = Model(cost) params = model.get_params() logger.info("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in params.items()], width=120)) model_size = 0 for v in params.itervalues(): s = v.get_value().shape model_size += s[0] * (s[1] if len(s) > 1 else 1) logger.info("Total number of parameters %d"%model_size) #------------------------------------------------------------ extensions = [] if old_model_name == 'continue': extensions.append(LoadFromDump(jobname)) elif old_model_name: # or you can just load the weights without state using: old_params = LoadFromDump(old_model_name).manager.load_parameters() model.set_param_values(old_params) else: # Initialize parameters for brick in model.get_top_bricks(): brick.initialize() if sample: assert old_model_name and old_model_name != 'continue' Sample(generator, steps=max_length, path=old_model_name).do(None) exit(0) #------------------------------------------------------------ # Define the training algorithm. cg = ComputationGraph(cost) if dropout > 0.: from blocks.roles import INPUT, OUTPUT dropout_target = VariableFilter(roles=[OUTPUT], bricks=transitions, name_regex='states')(cg.variables) print('# dropout %d' % len(dropout_target)) cg = apply_dropout(cg, dropout_target, dropout) opt_cost = cg.outputs[0] else: opt_cost = cost if step_method == 'adam': step_rule = Adam(learning_rate) elif step_method == 'rmsprop': step_rule = RMSProp(learning_rate, decay_rate=0.95) elif step_method == 'adagrad': step_rule = AdaGrad(learning_rate) elif step_method == 'adadelta': step_rule = AdaDelta() elif step_method == 'scale': step_rule = Scale(learning_rate) else: raise Exception('Unknown sttep method %s'%step_method) step_rule = CompositeRule([StepClipping(max_grad), step_rule]) algorithm = GradientDescent( cost=opt_cost, params=cg.parameters, step_rule=step_rule) #------------------------------------------------------------ observables = [cost] # Fetch variables useful for debugging (energies,) = VariableFilter( applications=[generator.readout.readout], name_regex="output")(cg.variables) min_energy = named_copy(energies.min(), "min_energy") max_energy = named_copy(energies.max(), "max_energy") observables += [min_energy, max_energy] # (activations,) = VariableFilter( # applications=[generator.transition.apply], # name=generator.transition.apply.states[0])(cg.variables) # mean_activation = named_copy(abs(activations).mean(), # "mean_activation") # observables.append(mean_activation) observables += [algorithm.total_step_norm, algorithm.total_gradient_norm] for name, param in params.items(): observables.append(named_copy( param.norm(2), name + "_norm")) observables.append(named_copy( algorithm.gradients[param].norm(2), name + "_grad_norm")) #------------------------------------------------------------ datasource_fname = os.path.join(fuel.config.data_path, datasource, datasource+'.hdf5') train_ds = H5PYDataset(datasource_fname, #max_length=max_length, which_set='train', sources=('features',), load_in_memory=True) train_stream = DataStream(train_ds, iteration_scheme=ShuffledScheme( train_ds.num_examples, batch_size)) test_ds = H5PYDataset(datasource_fname, #max_length=max_length, which_set='test', sources=('features',), load_in_memory=True) test_stream = DataStream(test_ds, iteration_scheme=SequentialScheme( test_ds.num_examples, batch_size)) train_stream = Mapping(train_stream, _transpose) test_stream = Mapping(test_stream, _transpose) def stream_stats(ds, label): itr = ds.get_epoch_iterator(as_dict=True) batch_count = 0 examples_count = 0 for batch in itr: batch_count += 1 examples_count += batch['features'].shape[1] print('%s #batch %d #examples %d' % (label, batch_count, examples_count)) stream_stats(train_stream, 'train') stream_stats(test_stream, 'test') extensions += [Timing(every_n_batches=10), TrainingDataMonitoring( observables, prefix="train", every_n_batches=10), DataStreamMonitoring( [cost], # without dropout test_stream, prefix="test", on_resumption=True, after_epoch=False, # by default this is True every_n_batches=100), # all monitored data is ready so print it... # (next steps may take more time and we want to see the # results as soon as possible so print as soon as you can) Printing(every_n_batches=10), # perform multiple dumps at different intervals # so if one of them breaks (has nan) we can hopefully # find a model from few batches ago in the other Dump(jobname, every_n_batches=11), Dump(jobname+'.test', every_n_batches=100), Sample(generator, steps=max_length, path=jobname+'.test', every_n_batches=100), ProgressBar(), FinishAfter(after_n_epochs=epochs) # This shows a way to handle NaN emerging during # training: simply finish it. .add_condition("after_batch", _is_nan), ] if bokeh: from blocks.extensions.plot import Plot extensions.append(Plot( 'sketch', channels=[ ['cost'],])) # Construct the main loop and start training! main_loop = MainLoop( model=model, data_stream=train_stream, algorithm=algorithm, extensions=extensions ) main_loop.run()
def train(): if os.path.isfile('trainingdata.tar'): with open('trainingdata.tar', 'rb') as f: main = load(f) else: hidden_size = 512 filename = 'warpeace.hdf5' encoder = HDF5CharEncoder('warpeace_input.txt', 1000) encoder.write(filename) alphabet_len = encoder.length x = theano.tensor.lmatrix('x') readout = Readout( readout_dim=alphabet_len, feedback_brick=LookupFeedback(alphabet_len, hidden_size, name='feedback'), source_names=['states'], emitter=RandomSoftmaxEmitter(), name='readout' ) transition = GatedRecurrent( activation=Tanh(), dim=hidden_size) transition.weights_init = IsotropicGaussian(0.01) gen = SequenceGenerator(readout=readout, transition=transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name='sequencegenerator') gen.push_initialization_config() gen.initialize() cost = gen.cost(outputs=x) cost.name = 'cost' cg = ComputationGraph(cost) algorithm = GradientDescent(cost=cost, parameters=cg.parameters, step_rule=Scale(0.5)) train_set = encoder.get_dataset() train_stream = DataStream.default_stream( train_set, iteration_scheme=SequentialScheme( train_set.num_examples, batch_size=128)) main = MainLoop( model=Model(cost), data_stream=train_stream, algorithm=algorithm, extensions=[ FinishAfter(), Printing(), Checkpoint('trainingdata.tar', every_n_epochs=10), ShowOutput(every_n_epochs=10) ]) main.run()
feedback_brick=LookupFeedback(vocab_size, feedback_dim, name='feedback'), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.push_initialization_config() transition.weights_init = Orthogonal() generator.initialize() # Build the cost computation graph. x = tensor.lmatrix('inchar') cost = generator.cost(outputs=x) cost.name = "sequence_cost" algorithm = GradientDescent( cost=cost, parameters=list(Selector(generator).get_parameters().values()), step_rule=Adam(), # because we want use all the stuff in the training data on_unused_sources='ignore') main_loop = MainLoop(algorithm=algorithm, data_stream=DataStream( train_data, iteration_scheme=SequentialScheme( train_data.num_examples, batch_size=20)), model=Model(cost), extensions=[
def test_sequence_generator_with_lm(): floatX = theano.config.floatX rng = numpy.random.RandomState(1234) readout_dim = 5 feedback_dim = 3 dim = 20 batch_size = 30 n_steps = 10 transition = GatedRecurrent(dim=dim, activation=Tanh(), weights_init=Orthogonal()) language_model = SequenceGenerator( Readout(readout_dim=readout_dim, source_names=["states"], emitter=SoftmaxEmitter(theano_seed=1234), feedback_brick=LookupFeedback(readout_dim, dim, name='feedback')), SimpleRecurrent(dim, Tanh()), name='language_model') generator = SequenceGenerator( Readout(readout_dim=readout_dim, source_names=["states", "lm_states"], emitter=SoftmaxEmitter(theano_seed=1234), feedback_brick=LookupFeedback(readout_dim, feedback_dim)), transition, language_model=language_model, weights_init=IsotropicGaussian(0.1), biases_init=Constant(0), seed=1234) generator.initialize() # Test 'cost_matrix' method y = tensor.lmatrix('y') y.tag.test_value = numpy.zeros((15, batch_size), dtype='int64') mask = tensor.matrix('mask') mask.tag.test_value = numpy.ones((15, batch_size)) costs = generator.cost_matrix(y, mask) assert costs.ndim == 2 costs_fun = theano.function([y, mask], [costs]) y_test = rng.randint(readout_dim, size=(n_steps, batch_size)) m_test = numpy.ones((n_steps, batch_size), dtype=floatX) costs_val = costs_fun(y_test, m_test)[0] assert costs_val.shape == (n_steps, batch_size) assert_allclose(costs_val.sum(), 483.153, rtol=1e-5) # Test 'cost' method cost = generator.cost(y, mask) assert cost.ndim == 0 cost_val = theano.function([y, mask], cost)(y_test, m_test) assert_allclose(cost_val, 16.105, rtol=1e-5) # Test 'AUXILIARY' variable 'per_sequence_element' in 'cost' method cg = ComputationGraph([cost]) var_filter = VariableFilter(roles=[AUXILIARY]) aux_var_name = '_'.join([generator.name, generator.cost.name, 'per_sequence_element']) cost_per_el = [el for el in var_filter(cg.variables) if el.name == aux_var_name][0] assert cost_per_el.ndim == 0 cost_per_el_val = theano.function([y, mask], [cost_per_el])(y_test, m_test) assert_allclose(cost_per_el_val, 1.61051, rtol=1e-5) # Test generate states, outputs, lm_states, costs = generator.generate( iterate=True, batch_size=batch_size, n_steps=n_steps) cg = ComputationGraph([states, outputs, costs]) states_val, outputs_val, costs_val = theano.function( [], [states, outputs, costs], updates=cg.updates)() assert states_val.shape == (n_steps, batch_size, dim) assert outputs_val.shape == (n_steps, batch_size) assert outputs_val.dtype == 'int64' assert costs_val.shape == (n_steps, batch_size) assert_allclose(states_val.sum(), -4.88367, rtol=1e-5) assert_allclose(costs_val.sum(), 486.681, rtol=1e-5) assert outputs_val.sum() == 627 # Test masks agnostic results of cost cost1 = costs_fun([[1], [2]], [[1], [1]])[0] cost2 = costs_fun([[3, 1], [4, 2], [2, 0]], [[1, 1], [1, 1], [1, 0]])[0] assert_allclose(cost1.sum(), cost2[:, 1].sum(), rtol=1e-5)
), transition=ContextSimpleRecurrent( name='image', activation=Tanh(), dim=context_dim, weights_init=initialization.IsotropicGaussian(0.1) ), weights_init=initialization.IsotropicGaussian(0.1), biases_init=initialization.Constant(0.0)) generator.initialize() x = tensor.matrix('image') y = tensor.lmatrix('sequence') cost = generator.cost(y, context=x) cg = ComputationGraph([cost]) algorithm = GradientDescent( cost=cost, parameters=cg.parameters, step_rule=Adam() ) from blocks.extensions import Timing, FinishAfter, Printing, ProgressBar from blocks.extensions.monitoring import TrainingDataMonitoring from blocks.extensions.saveload import Checkpoint from fuel.streams import DataStream from fuel.schemes import SequentialScheme
def main(): logging.basicConfig( level=logging.DEBUG, format="%(asctime)s: %(name)s: %(levelname)s: %(message)s") parser = argparse.ArgumentParser( "Case study of language modeling with RNN", formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument( "mode", choices=["train", "sample"], help="The mode to run. Use `train` to train a new model" " and `sample` to sample a sequence generated by an" " existing one.") parser.add_argument("prefix", default="sine", help="The prefix for model, timing and state files") parser.add_argument("state", nargs="?", default="", help="Changes to Groundhog state") parser.add_argument("--path", help="Path to a language dataset") parser.add_argument("--dict", help="Path to the dataset dictionary") parser.add_argument("--restart", help="Start anew") parser.add_argument("--reset", action="store_true", default=False, help="Reset the hidden state between batches") parser.add_argument("--steps", type=int, default=100, help="Number of steps to plot for the 'sample' mode" " OR training sequence length for the 'train' mode.") args = parser.parse_args() logger.debug("Args:\n" + str(args)) dim = 200 num_chars = 50 transition = GatedRecurrent(name="transition", activation=Tanh(), dim=dim, weights_init=Orthogonal()) generator = SequenceGenerator(LinearReadout( readout_dim=num_chars, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback(num_chars, dim, name='feedback'), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.allocate() logger.debug("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in Selector(generator).get_params().items()], width=120)) if args.mode == "train": batch_size = 1 seq_len = args.steps generator.initialize() # Build cost computation graph that uses the saved hidden states. # An issue: for Groundhog this is completely transparent, that's # why it does not carry the hidden state over the period when # validation in done. We should find a way to fix in the future. x = tensor.lmatrix('x') init_states = shared_floatx_zeros((batch_size, dim), name='init_states') reset = tensor.scalar('reset') cost = ComputationGraph( generator.cost(x, states=init_states * reset).sum()) # TODO: better search routine states = [ v for v in cost.variables if hasattr(v.tag, 'application_call') and v.tag.application_call.brick == generator.transition and (v.tag.application_call.application == generator.transition.apply) and v.tag.role == VariableRole.OUTPUT and v.tag.name == 'states' ] assert len(states) == 1 states = states[0] gh_model = GroundhogModel(generator, cost) gh_model.properties.append( ('bpc', cost.outputs[0] * numpy.log(2) / seq_len)) gh_model.properties.append(('mean_init_state', init_states.mean())) gh_model.properties.append(('reset', reset)) if not args.reset: gh_model.updates.append((init_states, states[-1])) state = GroundhogState(args.prefix, batch_size, learning_rate=0.0001).as_dict() changes = eval("dict({})".format(args.state)) state.update(changes) def output_format(x, y, reset): return dict(x=x[:, None], reset=reset) train, valid, test = [ LMIterator(batch_size=batch_size, use_infinite_loop=mode == 'train', path=args.path, seq_len=seq_len, mode=mode, chunks='chars', output_format=output_format, can_fit=True) for mode in ['train', 'valid', 'test'] ] trainer = SGD(gh_model, state, train) state['on_nan'] = 'warn' state['cutoff'] = 1. main_loop = MainLoop(train, valid, None, gh_model, trainer, state, None) if not args.restart: main_loop.load() main_loop.main() elif args.mode == "sample": load_params(generator, args.prefix + "model.npz") chars = numpy.load(args.dict)['unique_chars'] sample = ComputationGraph( generator.generate(n_steps=args.steps, batch_size=10, iterate=True)).function() states, outputs, costs = sample() for i in range(10): print("Generation cost: {}".format(costs[:, i].sum())) print("".join([chars[o] for o in outputs[:, i]])) else: assert False
def main(mode, save_path, steps, num_batches): num_states = MarkovChainDataset.num_states if mode == "train": # Experiment configuration rng = numpy.random.RandomState(1) batch_size = 50 seq_len = 100 dim = 10 feedback_dim = 8 # Build the bricks and initialize them transition = GatedRecurrent(name="transition", activation=Tanh(), dim=dim) generator = SequenceGenerator(LinearReadout( readout_dim=num_states, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback(num_states, feedback_dim, name='feedback'), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.push_initialization_config() transition.weights_init = Orthogonal() generator.initialize() # Give an idea of what's going on. logger.info("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in Selector(generator).get_params().items()], width=120)) logger.info("Markov chain entropy: {}".format( MarkovChainDataset.entropy)) logger.info("Expected min error: {}".format( -MarkovChainDataset.entropy * seq_len)) # Build the cost computation graph. x = tensor.lmatrix('data') cost = aggregation.mean(generator.cost(x[:, :]).sum(), x.shape[1]) cost.name = "sequence_log_likelihood" algorithm = GradientDescent( cost=cost, params=list(Selector(generator).get_params().values()), step_rule=Scale(0.001)) main_loop = MainLoop(algorithm=algorithm, data_stream=DataStream( MarkovChainDataset(rng, seq_len), iteration_scheme=ConstantScheme(batch_size)), model=Model(cost), extensions=[ FinishAfter(after_n_batches=num_batches), TrainingDataMonitoring( [cost], prefix="this_step", after_every_batch=True), TrainingDataMonitoring([cost], prefix="average", every_n_batches=100), SerializeMainLoop(save_path, every_n_batches=500), Printing(every_n_batches=100) ]) main_loop.run() elif mode == "sample": main_loop = cPickle.load(open(save_path, "rb")) generator = main_loop.model sample = ComputationGraph( generator.generate(n_steps=steps, batch_size=1, iterate=True)).get_theano_function() states, outputs, costs = [data[:, 0] for data in sample()] numpy.set_printoptions(precision=3, suppress=True) print("Generation cost:\n{}".format(costs.sum())) freqs = numpy.bincount(outputs).astype(floatX) freqs /= freqs.sum() print("Frequencies:\n {} vs {}".format(freqs, MarkovChainDataset.equilibrium)) trans_freqs = numpy.zeros((num_states, num_states), dtype=floatX) for a, b in zip(outputs, outputs[1:]): trans_freqs[a, b] += 1 trans_freqs /= trans_freqs.sum(axis=1)[:, None] print("Transition frequencies:\n{}\nvs\n{}".format( trans_freqs, MarkovChainDataset.trans_prob)) else: assert False
def main(config): vocab_src, _ = text_to_dict([config['train_src'], config['dev_src'], config['test_src']]) vocab_tgt, cabvo = text_to_dict([config['train_tgt'], config['dev_tgt']]) # Create Theano variables logger.info('Creating theano variables') source_sentence = tensor.lmatrix('source') source_sentence_mask = tensor.matrix('source_mask') target_sentence = tensor.lmatrix('target') target_sentence_mask = tensor.matrix('target_mask') source_sentence.tag.test_value = [[13, 20, 0, 20, 0, 20, 0], [1, 4, 8, 4, 8, 4, 8],] source_sentence_mask.tag.test_value = [[0, 1, 0, 1, 0, 1, 0], [1, 0, 1, 0, 1, 0, 1],] target_sentence.tag.test_value = [[0,1,1,5], [2,0,1,0],] target_sentence_mask.tag.test_value = [[0,1,1,0], [1,1,1,0],] logger.info('Building RNN encoder-decoder') ### Building Encoder embedder = LookupTable( length=len(vocab_src), dim=config['embed_src'], weights_init=IsotropicGaussian(), biases_init=Constant(0.0), name='embedder') transformer = Linear( config['embed_src'], config['hidden_src']*4, weights_init=IsotropicGaussian(), biases_init=Constant(0.0), name='transformer') lstminit = np.asarray([0.0,]*config['hidden_src']+[0.0,]*config['hidden_src']+[1.0,]*config['hidden_src']+[0.0,]*config['hidden_src']) encoder = Bidirectional( LSTM( dim=config['hidden_src'], weights_init=IsotropicGaussian(0.01), biases_init=Constant(lstminit)), name='encoderBiLSTM' ) encoder.prototype.weights_init = Orthogonal() ### Building Decoder lstminit = np.asarray([0.0,]*config['hidden_tgt']+[0.0,]*config['hidden_tgt']+[1.0,]*config['hidden_tgt']+[0.0,]*config['hidden_tgt']) transition = LSTM2GO( attended_dim=config['hidden_tgt'], dim=config['hidden_tgt'], weights_init=IsotropicGaussian(0.01), biases_init=Constant(lstminit), name='decoderLSTM') attention = SequenceContentAttention( state_names=transition.apply.states, # default activation is Tanh state_dims=[config['hidden_tgt']], attended_dim=config['hidden_src']*2, match_dim=config['hidden_tgt'], name="attention") readout = Readout( source_names=['states', 'feedback', attention.take_glimpses.outputs[0]], readout_dim=len(vocab_tgt), emitter = SoftmaxEmitter( name='emitter'), feedback_brick = LookupFeedback( num_outputs=len(vocab_tgt), feedback_dim=config['embed_tgt'], name='feedback'), post_merge=InitializableFeedforwardSequence([ Bias(dim=config['hidden_tgt'], name='softmax_bias').apply, Linear(input_dim=config['hidden_tgt'], output_dim=config['embed_tgt'], use_bias=False, name='softmax0').apply, Linear(input_dim=config['embed_tgt'], name='softmax1').apply]), merged_dim=config['hidden_tgt']) decoder = SequenceGenerator( readout=readout, transition=transition, attention=attention, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator", fork=Fork( [name for name in transition.apply.sequences if name != 'mask'], prototype=Linear()), add_contexts=True) decoder.transition.weights_init = Orthogonal() #printchildren(encoder, 1) # Initialize model logger.info('Initializing model') embedder.initialize() transformer.initialize() encoder.initialize() decoder.initialize() # Apply model embedded = embedder.apply(source_sentence) tansformed = transformer.apply(embedded) encoded = encoder.apply(tansformed)[0] generated = decoder.generate( n_steps=2*source_sentence.shape[1], batch_size=source_sentence.shape[0], attended = encoded.dimshuffle(1,0,2), attended_mask=tensor.ones(source_sentence.shape).T ) print 'Generated: ', generated # generator_generate_outputs #samples = generated[1] # For GRU samples = generated[2] # For LSTM samples.name = 'samples' #samples_cost = generated[4] # For GRU samples_cost = generated[5] # For LSTM samples_cost = 'sampling_cost' cost = decoder.cost( mask = target_sentence_mask.T, outputs = target_sentence.T, attended = encoded.dimshuffle(1,0,2), attended_mask = source_sentence_mask.T) cost.name = 'target_cost' cost.tag.aggregation_scheme = TakeLast(cost) model = Model(cost) logger.info('Creating computational graph') cg = ComputationGraph(cost) # apply dropout for regularization if config['dropout'] < 1.0: # dropout is applied to the output of maxout in ghog logger.info('Applying dropout') dropout_inputs = [x for x in cg.intermediary_variables if x.name == 'maxout_apply_output'] cg = apply_dropout(cg, dropout_inputs, config['dropout']) ######## # Print shapes shapes = [param.get_value().shape for param in cg.parameters] logger.info("Parameter shapes: ") for shape, count in Counter(shapes).most_common(): logger.info(' {:15}: {}'.format(shape, count)) logger.info("Total number of parameters: {}".format(len(shapes))) printchildren(embedder, 1) printchildren(transformer, 1) printchildren(encoder, 1) printchildren(decoder, 1) # Print parameter names # enc_dec_param_dict = merge(Selector(embedder).get_parameters(), Selector(encoder).get_parameters(), Selector(decoder).get_parameters()) # enc_dec_param_dict = merge(Selector(decoder).get_parameters()) # logger.info("Parameter names: ") # for name, value in enc_dec_param_dict.items(): # logger.info(' {:15}: {}'.format(value.get_value().shape, name)) # logger.info("Total number of parameters: {}".format(len(enc_dec_param_dict))) ########## # Training data train_stream = get_train_stream(config, [config['train_src'],], [config['train_tgt'],], vocab_src, vocab_tgt) dev_stream = get_dev_stream( [config['dev_src'],], [config['dev_tgt'],], vocab_src, vocab_tgt) test_stream = get_test_stream([config['test_src'],], vocab_src) # Set extensions logger.info("Initializing extensions") extensions = [ FinishAfter(after_n_batches=config['finish_after']), ProgressBar(), TrainingDataMonitoring([cost], prefix="tra", after_batch=True), DataStreamMonitoring(variables=[cost], data_stream=dev_stream, prefix="dev", after_batch=True), Sampler( model=Model(samples), data_stream=dev_stream, vocab=cabvo, saveto=config['saveto']+'dev', every_n_batches=config['save_freq']), Sampler( model=Model(samples), data_stream=test_stream, vocab=cabvo, saveto=config['saveto']+'test', after_n_batches=1, on_resumption=True, before_training=True), Plotter(saveto=config['saveto'], after_batch=True), Printing(after_batch=True), Checkpoint( path=config['saveto'], parameters = cg.parameters, save_main_loop=False, every_n_batches=config['save_freq'])] if BOKEH_AVAILABLE: Plot('Training cost', channels=[['target_cost']], after_batch=True) if config['reload']: extensions.append(Load(path=config['saveto'], load_iteration_state=False, load_log=False)) else: with open(config['saveto']+'.txt', 'w') as f: pass # Set up training algorithm logger.info("Initializing training algorithm") algorithm = GradientDescent(cost=cost, parameters=cg.parameters, step_rule=CompositeRule([StepClipping(config['step_clipping']), eval(config['step_rule'])()]) ) # Initialize main loop logger.info("Initializing main loop") main_loop = MainLoop( model=model, algorithm=algorithm, data_stream=train_stream, extensions=extensions) main_loop.run()
def main(mode, save_path, steps, num_batches): num_states = MarkovChainDataset.num_states if mode == "train": # Experiment configuration rng = numpy.random.RandomState(1) batch_size = 50 seq_len = 100 dim = 10 feedback_dim = 8 # Build the bricks and initialize them transition = GatedRecurrent(name="transition", activation=Tanh(), dim=dim) generator = SequenceGenerator( LinearReadout(readout_dim=num_states, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback( num_states, feedback_dim, name='feedback'), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.push_initialization_config() transition.weights_init = Orthogonal() generator.initialize() # Give an idea of what's going on. logger.info("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in Selector(generator).get_params().items()], width=120)) logger.info("Markov chain entropy: {}".format( MarkovChainDataset.entropy)) logger.info("Expected min error: {}".format( -MarkovChainDataset.entropy * seq_len)) # Build the cost computation graph. x = tensor.lmatrix('data') cost = aggregation.mean(generator.cost(x[:, :]).sum(), x.shape[1]) cost.name = "sequence_log_likelihood" algorithm = GradientDescent( cost=cost, params=list(Selector(generator).get_params().values()), step_rule=Scale(0.001)) main_loop = MainLoop( algorithm=algorithm, data_stream=DataStream( MarkovChainDataset(rng, seq_len), iteration_scheme=ConstantScheme(batch_size)), model=Model(cost), extensions=[FinishAfter(after_n_batches=num_batches), TrainingDataMonitoring([cost], prefix="this_step", after_every_batch=True), TrainingDataMonitoring([cost], prefix="average", every_n_batches=100), SerializeMainLoop(save_path, every_n_batches=500), Printing(every_n_batches=100)]) main_loop.run() elif mode == "sample": main_loop = cPickle.load(open(save_path, "rb")) generator = main_loop.model sample = ComputationGraph(generator.generate( n_steps=steps, batch_size=1, iterate=True)).get_theano_function() states, outputs, costs = [data[:, 0] for data in sample()] numpy.set_printoptions(precision=3, suppress=True) print("Generation cost:\n{}".format(costs.sum())) freqs = numpy.bincount(outputs).astype(floatX) freqs /= freqs.sum() print("Frequencies:\n {} vs {}".format(freqs, MarkovChainDataset.equilibrium)) trans_freqs = numpy.zeros((num_states, num_states), dtype=floatX) for a, b in zip(outputs, outputs[1:]): trans_freqs[a, b] += 1 trans_freqs /= trans_freqs.sum(axis=1)[:, None] print("Transition frequencies:\n{}\nvs\n{}".format( trans_freqs, MarkovChainDataset.trans_prob)) else: assert False
def __init__(self, config, vocab_size): context = tensor.imatrix('context') context_mask = tensor.imatrix('context_mask') answer = tensor.imatrix('answer') answer_mask = tensor.imatrix('answer_mask') bricks = [] context = context.dimshuffle(1, 0) context_mask = context_mask.dimshuffle(1, 0) answer = answer.dimshuffle(1, 0) answer_mask = answer_mask.dimshuffle(1, 0) context_bag = to_bag(context, vocab_size) # Embed questions and context embed = LookupTable(vocab_size, config.embed_size, name='embed') embed.weights_init = IsotropicGaussian(0.01) #embeddings_initial_value = init_embedding_table(filename='embeddings/vocab_embeddings.txt') #embed.weights_init = Constant(embeddings_initial_value) # Calculate context encoding (concatenate layer1) cembed = embed.apply(context) clstms, chidden_list = make_bidir_lstm_stack( cembed, config.embed_size, context_mask.astype(theano.config.floatX), config.ctx_lstm_size, config.ctx_skip_connections, 'ctx') bricks = bricks + clstms if config.ctx_skip_connections: cenc_dim = 2 * sum(config.ctx_lstm_size) #2 : fw & bw cenc = tensor.concatenate(chidden_list, axis=2) else: cenc_dim = 2 * config.ctx_lstm_size[-1] cenc = tensor.concatenate(chidden_list[-2:], axis=2) cenc.name = 'cenc' # Build the encoder bricks transition = GatedRecurrent(activation=Tanh(), dim=config.generator_lstm_size, name="transition") attention = SequenceContentAttention( state_names=transition.apply.states, attended_dim=cenc_dim, match_dim=config.generator_lstm_size, name="attention") readout = Readout(readout_dim=vocab_size, source_names=[ transition.apply.states[0], attention.take_glimpses.outputs[0] ], emitter=MaskedSoftmaxEmitter(context_bag=context_bag, name='emitter'), feedback_brick=LookupFeedback( vocab_size, config.feedback_size), name="readout") generator = SequenceGenerator(readout=readout, transition=transition, attention=attention, name="generator") cost = generator.cost(answer, answer_mask.astype(theano.config.floatX), attended=cenc, attended_mask=context_mask.astype( theano.config.floatX), name="cost") self.predictions = generator.generate( n_steps=7, batch_size=config.batch_size, attended=cenc, attended_mask=context_mask.astype(theano.config.floatX), iterate=True)[1] # Apply dropout cg = ComputationGraph([cost]) if config.w_noise > 0: noise_vars = VariableFilter(roles=[WEIGHT])(cg) cg = apply_noise(cg, noise_vars, config.w_noise) if config.dropout > 0: cg = apply_dropout(cg, chidden_list, config.dropout) [cost_reg] = cg.outputs # Other stuff cost.name = 'cost' cost_reg.name = 'cost_reg' self.sgd_cost = cost_reg self.monitor_vars = [[cost_reg]] self.monitor_vars_valid = [[cost_reg]] # initialize new stuff manually (change!) generator.weights_init = IsotropicGaussian(0.01) generator.biases_init = Constant(0) generator.push_allocation_config() generator.push_initialization_config() transition.weights_init = Orthogonal() generator.initialize() # Initialize bricks embed.initialize() for brick in bricks: brick.weights_init = config.weights_init brick.biases_init = config.biases_init brick.initialize()
def main(): logging.basicConfig( level=logging.DEBUG, format="%(asctime)s: %(name)s: %(levelname)s: %(message)s") parser = argparse.ArgumentParser( "Case study of generating a Markov chain with RNN.", formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument( "mode", choices=["train", "sample"], help="The mode to run. Use `train` to train a new model" " and `sample` to sample a sequence generated by an" " existing one.") parser.add_argument("prefix", default="sine", help="The prefix for model, timing and state files") parser.add_argument("--steps", type=int, default=100, help="Number of steps to plot") args = parser.parse_args() dim = 10 num_states = ChainIterator.num_states feedback_dim = 8 transition = GatedRecurrent(name="transition", activation=Tanh(), dim=dim) generator = SequenceGenerator(LinearReadout( readout_dim=num_states, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback(num_states, feedback_dim, name='feedback'), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.allocate() logger.debug("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in Selector(generator).get_params().items()], width=120)) if args.mode == "train": rng = numpy.random.RandomState(1) batch_size = 50 generator.push_initialization_config() transition.weights_init = Orthogonal() generator.initialize() logger.debug("transition.weights_init={}".format( transition.weights_init)) cost = generator.cost(tensor.lmatrix('x')).sum() gh_model = GroundhogModel(generator, cost) state = GroundhogState(args.prefix, batch_size, learning_rate=0.0001).as_dict() data = ChainIterator(rng, 100, batch_size) trainer = SGD(gh_model, state, data) main_loop = MainLoop(data, None, None, gh_model, trainer, state, None) main_loop.main() elif args.mode == "sample": load_params(generator, args.prefix + "model.npz") sample = ComputationGraph( generator.generate(n_steps=args.steps, batch_size=1, iterate=True)).function() states, outputs, costs = [data[:, 0] for data in sample()] numpy.set_printoptions(precision=3, suppress=True) print("Generation cost:\n{}".format(costs.sum())) freqs = numpy.bincount(outputs).astype(floatX) freqs /= freqs.sum() print("Frequencies:\n {} vs {}".format(freqs, ChainIterator.equilibrium)) trans_freqs = numpy.zeros((num_states, num_states), dtype=floatX) for a, b in zip(outputs, outputs[1:]): trans_freqs[a, b] += 1 trans_freqs /= trans_freqs.sum(axis=1)[:, None] print("Transition frequencies:\n{}\nvs\n{}".format( trans_freqs, ChainIterator.trans_prob)) else: assert False
def test_attention_transition(): inp_dim = 2 inp_len = 10 attended_dim = 3 attended_len = 11 batch_size = 4 n_steps = 30 transition = TestTransition(dim=inp_dim, attended_dim=attended_dim, name="transition") attention = SequenceContentAttention(transition.apply.states, match_dim=inp_dim, name="attention") mixer = Mixer([name for name in transition.apply.sequences if name != 'mask'], attention.take_look.outputs[0], name="mixer") att_trans = AttentionTransition(transition, attention, mixer, name="att_trans") att_trans.weights_init = IsotropicGaussian(0.01) att_trans.biases_init = Constant(0) att_trans.initialize() attended = tensor.tensor3("attended") attended_mask = tensor.matrix("attended_mask") inputs = tensor.tensor3("inputs") inputs_mask = tensor.matrix("inputs_mask") states, glimpses, weights = att_trans.apply( input_=inputs, mask=inputs_mask, attended=attended, attended_mask=attended_mask) assert states.ndim == 3 assert glimpses.ndim == 3 assert weights.ndim == 3 input_vals = numpy.zeros((inp_len, batch_size, inp_dim), dtype=floatX) input_mask_vals = numpy.ones((inp_len, batch_size), dtype=floatX) attended_vals = numpy.zeros((attended_len, batch_size, attended_dim), dtype=floatX) attended_mask_vals = numpy.ones((attended_len, batch_size), dtype=floatX) func = theano.function([inputs, inputs_mask, attended, attended_mask], [states, glimpses, weights]) states_vals, glimpses_vals, weight_vals = func( input_vals, input_mask_vals, attended_vals, attended_mask_vals) assert states_vals.shape == input_vals.shape assert glimpses_vals.shape == (inp_len, batch_size, attended_dim) assert weight_vals.shape == (inp_len, batch_size, attended_len) # Test SequenceGenerator using AttentionTransition generator = SequenceGenerator( LinearReadout(readout_dim=inp_dim, source_names=["state"], emitter=TestEmitter(name="emitter"), name="readout"), transition=transition, attention=attention, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") outputs = tensor.tensor3('outputs') costs = generator.cost(outputs, attended=attended, attended_mask=attended_mask) costs_vals = costs.eval({outputs: input_vals, attended: attended_vals, attended_mask: attended_mask_vals}) assert costs_vals.shape == (inp_len, batch_size) results = ( generator.generate(n_steps=n_steps, batch_size=attended.shape[1], attended=attended, attended_mask=attended_mask)) assert len(results) == 5 states_vals, outputs_vals, glimpses_vals, weights_vals, costs_vals = ( theano.function([attended, attended_mask], results) (attended_vals, attended_mask_vals)) assert states_vals.shape == (n_steps, batch_size, inp_dim) assert states_vals.shape == outputs_vals.shape assert glimpses_vals.shape == (n_steps, batch_size, attended_dim) assert weights_vals.shape == (n_steps, batch_size, attended_len) assert costs_vals.shape == (n_steps, batch_size)
def main(mode, save_path, num_batches, from_dump): if mode == "train": # Experiment configuration dimension = 100 readout_dimension = len(char2code) # Data processing pipeline data_stream = DataStreamMapping( mapping=lambda data: tuple(array.T for array in data), data_stream=PaddingDataStream( BatchDataStream( iteration_scheme=ConstantScheme(10), data_stream=DataStreamMapping( mapping=reverse_words, add_sources=("targets", ), data_stream=DataStreamFilter( predicate=lambda data: len(data[0]) <= 100, data_stream=OneBillionWord( "training", [99], char2code, level="character", preprocess=str.lower).get_default_stream()))))) # Build the model chars = tensor.lmatrix("features") chars_mask = tensor.matrix("features_mask") targets = tensor.lmatrix("targets") targets_mask = tensor.matrix("targets_mask") encoder = Bidirectional(GatedRecurrent(dim=dimension, activation=Tanh()), weights_init=Orthogonal()) encoder.initialize() fork = Fork([ name for name in encoder.prototype.apply.sequences if name != 'mask' ], weights_init=IsotropicGaussian(0.1), biases_init=Constant(0)) fork.input_dim = dimension fork.fork_dims = {name: dimension for name in fork.fork_names} fork.initialize() lookup = LookupTable(readout_dimension, dimension, weights_init=IsotropicGaussian(0.1)) lookup.initialize() transition = Transition(activation=Tanh(), dim=dimension, attended_dim=2 * dimension, name="transition") attention = SequenceContentAttention( state_names=transition.apply.states, match_dim=dimension, name="attention") readout = LinearReadout(readout_dim=readout_dimension, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback( readout_dimension, dimension), name="readout") generator = SequenceGenerator(readout=readout, transition=transition, attention=attention, weights_init=IsotropicGaussian(0.1), biases_init=Constant(0), name="generator") generator.push_initialization_config() transition.weights_init = Orthogonal() generator.initialize() bricks = [encoder, fork, lookup, generator] # Give an idea of what's going on params = Selector(bricks).get_params() logger.info("Parameters:\n" + pprint.pformat([(key, value.get_value().shape) for key, value in params.items()], width=120)) # Build the cost computation graph batch_cost = generator.cost( targets, targets_mask, attended=encoder.apply(**dict_union(fork.apply( lookup.lookup(chars), return_dict=True), mask=chars_mask)), attended_mask=chars_mask).sum() batch_size = named_copy(chars.shape[1], "batch_size") cost = aggregation.mean(batch_cost, batch_size) cost.name = "sequence_log_likelihood" logger.info("Cost graph is built") # Fetch variables useful for debugging max_length = named_copy(chars.shape[0], "max_length") cost_per_character = named_copy( aggregation.mean(batch_cost, batch_size * max_length), "character_log_likelihood") cg = ComputationGraph(cost) energies = unpack(VariableFilter(application=readout.readout, name="output")(cg.variables), singleton=True) min_energy = named_copy(energies.min(), "min_energy") max_energy = named_copy(energies.max(), "max_energy") (activations, ) = VariableFilter( application=generator.transition.apply, name="states")(cg.variables) mean_activation = named_copy(activations.mean(), "mean_activation") # Define the training algorithm. algorithm = GradientDescent(cost=cost, step_rule=CompositeRule([ GradientClipping(10.0), SteepestDescent(0.01) ])) observables = [ cost, min_energy, max_energy, mean_activation, batch_size, max_length, cost_per_character, algorithm.total_step_norm, algorithm.total_gradient_norm ] for name, param in params.items(): observables.append(named_copy(param.norm(2), name + "_norm")) observables.append( named_copy(algorithm.gradients[param].norm(2), name + "_grad_norm")) main_loop = MainLoop( model=bricks, data_stream=data_stream, algorithm=algorithm, extensions=([LoadFromDump(from_dump)] if from_dump else []) + [ Timing(), TrainingDataMonitoring(observables, after_every_batch=True), TrainingDataMonitoring( observables, prefix="average", every_n_batches=10), FinishAfter(after_n_batches=num_batches).add_condition( "after_batch", lambda log: math.isnan( log.current_row.total_gradient_norm)), Plot(os.path.basename(save_path), [["average_" + cost.name], ["average_" + cost_per_character.name]], every_n_batches=10), SerializeMainLoop(save_path, every_n_batches=500, save_separately=["model", "log"]), Printing(every_n_batches=1) ]) main_loop.run() elif mode == "test": with open(save_path, "rb") as source: encoder, fork, lookup, generator = dill.load(source) logger.info("Model is loaded") chars = tensor.lmatrix("features") generated = generator.generate( n_steps=3 * chars.shape[0], batch_size=chars.shape[1], attended=encoder.apply(**dict_union( fork.apply(lookup.lookup(chars), return_dict=True))), attended_mask=tensor.ones(chars.shape)) sample_function = ComputationGraph(generated).get_theano_function() logging.info("Sampling function is compiled") while True: # Python 2-3 compatibility line = input("Enter a sentence\n") batch_size = int(input("Enter a number of samples\n")) encoded_input = [ char2code.get(char, char2code["<UNK>"]) for char in line.lower().strip() ] encoded_input = ([char2code['<S>']] + encoded_input + [char2code['</S>']]) print("Encoder input:", encoded_input) target = reverse_words((encoded_input, ))[0] print("Target: ", target) states, samples, glimpses, weights, costs = sample_function( numpy.repeat(numpy.array(encoded_input)[:, None], batch_size, axis=1)) messages = [] for i in range(samples.shape[1]): sample = list(samples[:, i]) try: true_length = sample.index(char2code['</S>']) + 1 except ValueError: true_length = len(sample) sample = sample[:true_length] cost = costs[:true_length, i].sum() message = "({})".format(cost) message += "".join(code2char[code] for code in sample) if sample == target: message += " CORRECT!" messages.append((cost, message)) messages.sort(key=lambda tuple_: -tuple_[0]) for _, message in messages: print(message)
def test_sequence_generator_with_lm(): floatX = theano.config.floatX rng = numpy.random.RandomState(1234) readout_dim = 5 feedback_dim = 3 dim = 20 batch_size = 30 n_steps = 10 transition = GatedRecurrent(dim=dim, activation=Tanh(), weights_init=Orthogonal()) language_model = SequenceGenerator(Readout( readout_dim=readout_dim, source_names=["states"], emitter=SoftmaxEmitter(theano_seed=1234), feedback_brick=LookupFeedback(readout_dim, dim, name='feedback')), SimpleRecurrent(dim, Tanh()), name='language_model') generator = SequenceGenerator(Readout( readout_dim=readout_dim, source_names=["states", "lm_states"], emitter=SoftmaxEmitter(theano_seed=1234), feedback_brick=LookupFeedback(readout_dim, feedback_dim)), transition, language_model=language_model, weights_init=IsotropicGaussian(0.1), biases_init=Constant(0), seed=1234) generator.initialize() # Test 'cost_matrix' method y = tensor.lmatrix('y') y.tag.test_value = numpy.zeros((15, batch_size), dtype='int64') mask = tensor.matrix('mask') mask.tag.test_value = numpy.ones((15, batch_size)) costs = generator.cost_matrix(y, mask) assert costs.ndim == 2 costs_fun = theano.function([y, mask], [costs]) y_test = rng.randint(readout_dim, size=(n_steps, batch_size)) m_test = numpy.ones((n_steps, batch_size), dtype=floatX) costs_val = costs_fun(y_test, m_test)[0] assert costs_val.shape == (n_steps, batch_size) assert_allclose(costs_val.sum(), 483.153, rtol=1e-5) # Test 'cost' method cost = generator.cost(y, mask) assert cost.ndim == 0 cost_val = theano.function([y, mask], cost)(y_test, m_test) assert_allclose(cost_val, 16.105, rtol=1e-5) # Test 'AUXILIARY' variable 'per_sequence_element' in 'cost' method cg = ComputationGraph([cost]) var_filter = VariableFilter(roles=[AUXILIARY]) aux_var_name = '_'.join( [generator.name, generator.cost.name, 'per_sequence_element']) cost_per_el = [ el for el in var_filter(cg.variables) if el.name == aux_var_name ][0] assert cost_per_el.ndim == 0 cost_per_el_val = theano.function([y, mask], [cost_per_el])(y_test, m_test) assert_allclose(cost_per_el_val, 1.61051, rtol=1e-5) # Test generate states, outputs, lm_states, costs = generator.generate( iterate=True, batch_size=batch_size, n_steps=n_steps) cg = ComputationGraph([states, outputs, costs]) states_val, outputs_val, costs_val = theano.function( [], [states, outputs, costs], updates=cg.updates)() assert states_val.shape == (n_steps, batch_size, dim) assert outputs_val.shape == (n_steps, batch_size) assert outputs_val.dtype == 'int64' assert costs_val.shape == (n_steps, batch_size) assert_allclose(states_val.sum(), -4.88367, rtol=1e-5) assert_allclose(costs_val.sum(), 486.681, rtol=1e-5) assert outputs_val.sum() == 627 # Test masks agnostic results of cost cost1 = costs_fun([[1], [2]], [[1], [1]])[0] cost2 = costs_fun([[3, 1], [4, 2], [2, 0]], [[1, 1], [1, 1], [1, 0]])[0] assert_allclose(cost1.sum(), cost2[:, 1].sum(), rtol=1e-5)
def main(): logging.basicConfig( level=logging.DEBUG, format="%(asctime)s: %(name)s: %(levelname)s: %(message)s") parser = argparse.ArgumentParser( "Case study of generating a Markov chain with RNN.", formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument( "mode", choices=["train", "sample"], help="The mode to run. Use `train` to train a new model" " and `sample` to sample a sequence generated by an" " existing one.") parser.add_argument( "prefix", default="sine", help="The prefix for model, timing and state files") parser.add_argument( "--steps", type=int, default=100, help="Number of steps to plot") args = parser.parse_args() dim = 10 num_states = ChainIterator.num_states feedback_dim = 8 transition = GatedRecurrent(name="transition", activation=Tanh(), dim=dim) generator = SequenceGenerator( LinearReadout(readout_dim=num_states, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback( num_states, feedback_dim, name='feedback'), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.allocate() logger.debug("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in Selector(generator).get_params().items()], width=120)) if args.mode == "train": rng = numpy.random.RandomState(1) batch_size = 50 generator.push_initialization_config() transition.weights_init = Orthogonal() generator.initialize() logger.debug("transition.weights_init={}".format( transition.weights_init)) cost = generator.cost(tensor.lmatrix('x')).sum() gh_model = GroundhogModel(generator, cost) state = GroundhogState(args.prefix, batch_size, learning_rate=0.0001).as_dict() data = ChainIterator(rng, 100, batch_size) trainer = SGD(gh_model, state, data) main_loop = MainLoop(data, None, None, gh_model, trainer, state, None) main_loop.main() elif args.mode == "sample": load_params(generator, args.prefix + "model.npz") sample = ComputationGraph(generator.generate( n_steps=args.steps, batch_size=1, iterate=True)).function() states, outputs, costs = [data[:, 0] for data in sample()] numpy.set_printoptions(precision=3, suppress=True) print("Generation cost:\n{}".format(costs.sum())) freqs = numpy.bincount(outputs).astype(floatX) freqs /= freqs.sum() print("Frequencies:\n {} vs {}".format(freqs, ChainIterator.equilibrium)) trans_freqs = numpy.zeros((num_states, num_states), dtype=floatX) for a, b in zip(outputs, outputs[1:]): trans_freqs[a, b] += 1 trans_freqs /= trans_freqs.sum(axis=1)[:, None] print("Transition frequencies:\n{}\nvs\n{}".format( trans_freqs, ChainIterator.trans_prob)) else: assert False
def main(): logging.basicConfig( level=logging.DEBUG, format="%(asctime)s: %(name)s: %(levelname)s: %(message)s") parser = argparse.ArgumentParser( "Case study of generating simple 1d sequences with RNN.", formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument( "mode", choices=["train", "plot"], help="The mode to run. Use `train` to train a new model" " and `plot` to plot a sequence generated by an" " existing one.") parser.add_argument("prefix", default="sine", help="The prefix for model, timing and state files") parser.add_argument("--input-noise", type=float, default=0.0, help="Adds Gaussian noise of given intensity to the " " training sequences.") parser.add_argument( "--function", default="lambda a, x: numpy.sin(a * x)", help="An analytical description of the sequence family to learn." " The arguments before the last one are considered parameters.") parser.add_argument("--steps", type=int, default=100, help="Number of steps to plot") parser.add_argument("--params", help="Parameter values for plotting") args = parser.parse_args() function = eval(args.function) num_params = len(inspect.getargspec(function).args) - 1 class Emitter(TrivialEmitter): @application def cost(self, readouts, outputs): """Compute MSE.""" return ((readouts - outputs)**2).sum(axis=readouts.ndim - 1) transition = GatedRecurrent(name="transition", activation=Tanh(), dim=10, weights_init=Orthogonal()) with_params = AddParameters(transition, num_params, "params", name="with_params") generator = SequenceGenerator(LinearReadout( readout_dim=1, source_names=["states"], emitter=Emitter(name="emitter"), name="readout"), with_params, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.allocate() logger.debug("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in Selector(generator).get_params().items()], width=120)) if args.mode == "train": seed = 1 rng = numpy.random.RandomState(seed) batch_size = 10 generator.initialize() cost = ComputationGraph( generator.cost(tensor.tensor3('x'), params=tensor.matrix("params")).sum()) cost = apply_noise(cost, cost.inputs, args.input_noise) gh_model = GroundhogModel(generator, cost) state = GroundhogState(args.prefix, batch_size, learning_rate=0.0001).as_dict() data = SeriesIterator(rng, function, 100, batch_size) trainer = SGD(gh_model, state, data) main_loop = MainLoop(data, None, None, gh_model, trainer, state, None) main_loop.load() main_loop.main() elif args.mode == "plot": load_params(generator, args.prefix + "model.npz") params = tensor.matrix("params") sample = theano.function([params], generator.generate(params=params, n_steps=args.steps, batch_size=1)) param_values = numpy.array(map(float, args.params.split()), dtype=floatX) states, outputs, _ = sample(param_values[None, :]) actual = outputs[:, 0, 0] desired = numpy.array( [function(*(list(param_values) + [T])) for T in range(args.steps)]) print("MSE: {}".format(((actual - desired)**2).sum())) pyplot.plot(numpy.hstack([actual[:, None], desired[:, None]])) pyplot.show() else: assert False
def main(): logging.basicConfig( level=logging.DEBUG, format="%(asctime)s: %(name)s: %(levelname)s: %(message)s") parser = argparse.ArgumentParser( "Case study of language modeling with RNN", formatter_class=argparse.ArgumentDefaultsHelpFormatter) parser.add_argument( "mode", choices=["train", "sample"], help="The mode to run. Use `train` to train a new model" " and `sample` to sample a sequence generated by an" " existing one.") parser.add_argument( "prefix", default="sine", help="The prefix for model, timing and state files") parser.add_argument( "state", nargs="?", default="", help="Changes to Groundhog state") parser.add_argument("--path", help="Path to a language dataset") parser.add_argument("--dict", help="Path to the dataset dictionary") parser.add_argument("--restart", help="Start anew") parser.add_argument( "--reset", action="store_true", default=False, help="Reset the hidden state between batches") parser.add_argument( "--steps", type=int, default=100, help="Number of steps to plot for the 'sample' mode" " OR training sequence length for the 'train' mode.") args = parser.parse_args() logger.debug("Args:\n" + str(args)) dim = 200 num_chars = 50 transition = GatedRecurrent( name="transition", activation=Tanh(), dim=dim, weights_init=Orthogonal()) generator = SequenceGenerator( LinearReadout(readout_dim=num_chars, source_names=["states"], emitter=SoftmaxEmitter(name="emitter"), feedbacker=LookupFeedback( num_chars, dim, name='feedback'), name="readout"), transition, weights_init=IsotropicGaussian(0.01), biases_init=Constant(0), name="generator") generator.allocate() logger.debug("Parameters:\n" + pprint.pformat( [(key, value.get_value().shape) for key, value in Selector(generator).get_params().items()], width=120)) if args.mode == "train": batch_size = 1 seq_len = args.steps generator.initialize() # Build cost computation graph that uses the saved hidden states. # An issue: for Groundhog this is completely transparent, that's # why it does not carry the hidden state over the period when # validation in done. We should find a way to fix in the future. x = tensor.lmatrix('x') init_states = shared_floatx_zeros((batch_size, dim), name='init_states') reset = tensor.scalar('reset') cost = ComputationGraph( generator.cost(x, states=init_states * reset).sum()) # TODO: better search routine states = [v for v in cost.variables if hasattr(v.tag, 'application_call') and v.tag.application_call.brick == generator.transition and (v.tag.application_call.application == generator.transition.apply) and v.tag.role == VariableRole.OUTPUT and v.tag.name == 'states'] assert len(states) == 1 states = states[0] gh_model = GroundhogModel(generator, cost) gh_model.properties.append( ('bpc', cost.outputs[0] * numpy.log(2) / seq_len)) gh_model.properties.append(('mean_init_state', init_states.mean())) gh_model.properties.append(('reset', reset)) if not args.reset: gh_model.updates.append((init_states, states[-1])) state = GroundhogState(args.prefix, batch_size, learning_rate=0.0001).as_dict() changes = eval("dict({})".format(args.state)) state.update(changes) def output_format(x, y, reset): return dict(x=x[:, None], reset=reset) train, valid, test = [ LMIterator(batch_size=batch_size, use_infinite_loop=mode == 'train', path=args.path, seq_len=seq_len, mode=mode, chunks='chars', output_format=output_format, can_fit=True) for mode in ['train', 'valid', 'test']] trainer = SGD(gh_model, state, train) state['on_nan'] = 'warn' state['cutoff'] = 1. main_loop = MainLoop(train, valid, None, gh_model, trainer, state, None) if not args.restart: main_loop.load() main_loop.main() elif args.mode == "sample": load_params(generator, args.prefix + "model.npz") chars = numpy.load(args.dict)['unique_chars'] sample = ComputationGraph(generator.generate( n_steps=args.steps, batch_size=10, iterate=True)).function() states, outputs, costs = sample() for i in range(10): print("Generation cost: {}".format(costs[:, i].sum())) print("".join([chars[o] for o in outputs[:, i]])) else: assert False
class Decoder(Initializable): def __init__(self, vocab_size, embedding_dim, state_dim, representation_dim, **kwargs): super(Decoder, self).__init__(**kwargs) self.vocab_size = vocab_size self.embedding_dim = embedding_dim self.state_dim = state_dim self.representation_dim = representation_dim readout = Readout( source_names=['states', 'feedback', 'readout_context'], readout_dim=self.vocab_size, emitter=SoftmaxEmitter(), feedback_brick=LookupFeedback(vocab_size, embedding_dim), post_merge=InitializableFeedforwardSequence( [Bias(dim=1000).apply, Maxout(num_pieces=2).apply, Linear(input_dim=state_dim / 2, output_dim=100, use_bias=False).apply, Linear(input_dim=100).apply]), merged_dim=1000) self.transition = GatedRecurrentWithContext(Tanh(), dim=state_dim, name='decoder') # Readout will apply the linear transformation to 'readout_context' # with a Merge brick, so no need to fork it here self.fork = Fork([name for name in self.transition.apply.contexts + self.transition.apply.states if name != 'readout_context'], prototype=Linear()) self.tanh = Tanh() self.sequence_generator = SequenceGenerator( readout=readout, transition=self.transition, fork_inputs=[name for name in self.transition.apply.sequences if name != 'mask'], ) self.children = [self.fork, self.sequence_generator, self.tanh] def _push_allocation_config(self): self.fork.input_dim = self.representation_dim self.fork.output_dims = [self.state_dim for _ in self.fork.output_names] @application(inputs=['representation', 'target_sentence_mask', 'target_sentence'], outputs=['cost']) def cost(self, representation, target_sentence, target_sentence_mask): target_sentence = target_sentence.dimshuffle(1, 0) target_sentence_mask = target_sentence_mask.T # The initial state and contexts, all functions of the representation contexts = {key: value.dimshuffle('x', 0, 1) if key not in self.transition.apply.states else value for key, value in self.fork.apply(representation, as_dict=True).items()} contexts['states'] = self.tanh.apply(contexts['states']) cost = self.sequence_generator.cost(**merge( contexts, {'mask': target_sentence_mask, 'outputs': target_sentence, 'readout_context': representation.dimshuffle('x', 0, 1)} )) return (cost * target_sentence_mask).sum() / target_sentence_mask.shape[1]
def test_integer_sequence_generator(): """Test a sequence generator with integer outputs. Such sequence generators can be used to e.g. model language. """ rng = numpy.random.RandomState(1234) readout_dim = 5 feedback_dim = 3 dim = 20 batch_size = 30 n_steps = 10 transition = GatedRecurrent(dim=dim, activation=Tanh(), weights_init=Orthogonal()) generator = SequenceGenerator( Readout(readout_dim=readout_dim, source_names=["states"], emitter=SoftmaxEmitter(theano_seed=1234), feedback_brick=LookupFeedback(readout_dim, feedback_dim)), transition, weights_init=IsotropicGaussian(0.1), biases_init=Constant(0), seed=1234) generator.initialize() # Test 'cost_matrix' method y = tensor.lmatrix('y') mask = tensor.matrix('mask') costs = generator.cost_matrix(y, mask) assert costs.ndim == 2 costs_fun = theano.function([y, mask], [costs]) y_test = rng.randint(readout_dim, size=(n_steps, batch_size)) m_test = numpy.ones((n_steps, batch_size), dtype=floatX) costs_val = costs_fun(y_test, m_test)[0] assert costs_val.shape == (n_steps, batch_size) assert_allclose(costs_val.sum(), 482.827, rtol=1e-5) # Test 'cost' method cost = generator.cost(y, mask) assert cost.ndim == 0 cost_val = theano.function([y, mask], [cost])(y_test, m_test) assert_allclose(cost_val, 16.0942, rtol=1e-5) # Test 'AUXILIARY' variable 'per_sequence_element' in 'cost' method cg = ComputationGraph([cost]) var_filter = VariableFilter(roles=[AUXILIARY]) aux_var_name = '_'.join([generator.name, generator.cost.name, 'per_sequence_element']) cost_per_el = [el for el in var_filter(cg.variables) if el.name == aux_var_name][0] assert cost_per_el.ndim == 0 cost_per_el_val = theano.function([y, mask], [cost_per_el])(y_test, m_test) assert_allclose(cost_per_el_val, 1.60942, rtol=1e-5) # Test generate states, outputs, costs = generator.generate( iterate=True, batch_size=batch_size, n_steps=n_steps) cg = ComputationGraph(states + outputs + costs) states_val, outputs_val, costs_val = theano.function( [], [states, outputs, costs], updates=cg.updates)() assert states_val.shape == (n_steps, batch_size, dim) assert outputs_val.shape == (n_steps, batch_size) assert outputs_val.dtype == 'int64' assert costs_val.shape == (n_steps, batch_size) assert_allclose(states_val.sum(), -17.91811, rtol=1e-5) assert_allclose(costs_val.sum(), 482.863, rtol=1e-5) assert outputs_val.sum() == 630 # Test masks agnostic results of cost cost1 = costs_fun([[1], [2]], [[1], [1]])[0] cost2 = costs_fun([[3, 1], [4, 2], [2, 0]], [[1, 1], [1, 1], [1, 0]])[0] assert_allclose(cost1.sum(), cost2[:, 1].sum(), rtol=1e-5)
def main(name, epochs, batch_size, learning_rate, dim, mix_dim, old_model_name, max_length, bokeh, GRU, dropout, depth, max_grad, step_method, epsilon, sample): #---------------------------------------------------------------------- datasource = name def shnum(x): """ Convert a positive float into a short tag-usable string E.g.: 0 -> 0, 0.005 -> 53, 100 -> 1-2 """ return '0' if x <= 0 else '%s%d' % ( ("%e" % x)[0], -np.floor(np.log10(x))) jobname = "%s-%dX%dm%dd%dr%sb%de%s" % ( datasource, depth, dim, mix_dim, int( dropout * 10), shnum(learning_rate), batch_size, shnum(epsilon)) if max_length != 600: jobname += '-L%d' % max_length if GRU: jobname += 'g' if max_grad != 5.: jobname += 'G%g' % max_grad if step_method != 'adam': jobname += step_method if sample: print("Sampling") else: print("\nRunning experiment %s" % jobname) #---------------------------------------------------------------------- if depth > 1: transition = LSTMstack(dim=dim, depth=depth, name="transition", lstm_name="transition") assert not GRU elif GRU: transition = GatedRecurrent(dim=dim, name="transition") else: transition = LSTM(dim=dim, name="transition") emitter = SketchEmitter(mix_dim=mix_dim, epsilon=epsilon, name="emitter") readout = Readout(readout_dim=emitter.get_dim('inputs'), source_names=['states'], emitter=emitter, name="readout") normal_inputs = [ name for name in transition.apply.sequences if 'mask' not in name ] fork = Fork(normal_inputs, prototype=Linear(use_bias=True)) generator = SequenceGenerator(readout=readout, transition=transition, fork=fork) # Initialization settings generator.weights_init = OrthogonalGlorot() generator.biases_init = Constant(0) # Build the cost computation graph [steps,batch_size, 3] x = T.tensor3('features', dtype=floatX)[:max_length, :, :] x.tag.test_value = np.ones((max_length, batch_size, 3)).astype(np.float32) cost = generator.cost(x) cost.name = "sequence_log_likelihood" # Give an idea of what's going on model = Model(cost) params = model.get_params() logger.info("Parameters:\n" + pprint.pformat([(key, value.get_value().shape) for key, value in params.items()], width=120)) model_size = 0 for v in params.itervalues(): s = v.get_value().shape model_size += s[0] * (s[1] if len(s) > 1 else 1) logger.info("Total number of parameters %d" % model_size) #------------------------------------------------------------ extensions = [] if old_model_name == 'continue': extensions.append(LoadFromDump(jobname)) elif old_model_name: # or you can just load the weights without state using: old_params = LoadFromDump(old_model_name).manager.load_parameters() model.set_param_values(old_params) else: # Initialize parameters for brick in model.get_top_bricks(): brick.initialize() if sample: assert old_model_name and old_model_name != 'continue' Sample(generator, steps=max_length, path='.').do(None) exit(0) #------------------------------------------------------------ # Define the training algorithm. cg = ComputationGraph(cost) if dropout > 0.: from blocks.roles import INPUT, OUTPUT dropout_target = VariableFilter(roles=[OUTPUT], bricks=[transition], name_regex='states')(cg.variables) cg = apply_dropout(cg, dropout_target, dropout) cost = cg.outputs[0] if step_method == 'adam': step_rule = Adam(learning_rate) elif step_method == 'rmsprop': step_rule = RMSProp(learning_rate, decay_rate=0.95) elif step_method == 'adagrad': step_rule = AdaGrad(learning_rate) elif step_method == 'adadelta': step_rule = AdaDelta() elif step_method == 'scale': step_rule = Scale(learning_rate=0.1) else: raise Exception('Unknown sttep method %s' % step_method) step_rule = CompositeRule([StepClipping(max_grad), step_rule]) algorithm = GradientDescent(cost=cost, params=cg.parameters, step_rule=step_rule) #------------------------------------------------------------ observables = [cost] # Fetch variables useful for debugging (energies, ) = VariableFilter(applications=[generator.readout.readout], name_regex="output")(cg.variables) (activations, ) = VariableFilter( applications=[generator.transition.apply], name=generator.transition.apply.states[0])(cg.variables) min_energy = named_copy(energies.min(), "min_energy") max_energy = named_copy(energies.max(), "max_energy") mean_activation = named_copy(abs(activations).mean(), "mean_activation") observables += [min_energy, max_energy, mean_activation] observables += [algorithm.total_step_norm, algorithm.total_gradient_norm] for name, param in params.items(): observables.append(named_copy(param.norm(2), name + "_norm")) observables.append( named_copy(algorithm.gradients[param].norm(2), name + "_grad_norm")) #------------------------------------------------------------ datasource_fname = os.path.join(fuel.config.data_path, datasource, datasource + '.hdf5') train_ds = H5PYDataset( datasource_fname, #max_length=max_length, which_set='train', sources=('features', ), load_in_memory=True) train_stream = DataStream(train_ds, iteration_scheme=ShuffledScheme( train_ds.num_examples, batch_size)) test_ds = H5PYDataset( datasource_fname, #max_length=max_length, which_set='test', sources=('features', ), load_in_memory=True) test_stream = DataStream(test_ds, iteration_scheme=SequentialScheme( test_ds.num_examples, batch_size)) train_stream = Mapping(train_stream, _transpose) test_stream = Mapping(test_stream, _transpose) def stream_stats(ds, label): itr = ds.get_epoch_iterator(as_dict=True) batch_count = 0 examples_count = 0 for batch in itr: batch_count += 1 examples_count += batch['features'].shape[1] print('%s #batch %d #examples %d' % (label, batch_count, examples_count)) stream_stats(train_stream, 'train') stream_stats(test_stream, 'test') extensions += [ Timing(every_n_batches=10), TrainingDataMonitoring(observables, prefix="train", every_n_batches=10), DataStreamMonitoring( [cost], test_stream, prefix="test", on_resumption=True, after_epoch=False, # by default this is True every_n_batches=100), # all monitored data is ready so print it... # (next steps may take more time and we want to see the # results as soon as possible so print as soon as you can) Printing(every_n_batches=10), # perform multiple dumps at different intervals # so if one of them breaks (has nan) we can hopefully # find a model from few batches ago in the other Dump(jobname, every_n_batches=11), Dump(jobname + '.test', every_n_batches=100), Sample(generator, steps=max_length, path=jobname + '.test', every_n_batches=100), ProgressBar(), FinishAfter(after_n_epochs=epochs) # This shows a way to handle NaN emerging during # training: simply finish it. .add_condition("after_batch", _is_nan), ] if bokeh: extensions.append(Plot('sketch', channels=[ ['cost'], ])) # Construct the main loop and start training! main_loop = MainLoop(model=model, data_stream=train_stream, algorithm=algorithm, extensions=extensions) main_loop.run()
name = 'feedback' ), name = "readout"), transition, weights_init = IsotropicGaussian(0.01), biases_init = Constant(0), name = "generator" ) generator.push_initialization_config() transition.weights_init = Orthogonal() generator.initialize() # Build the cost computation graph. x = tensor.lmatrix('inchar') cost = generator.cost(outputs=x) cost.name = "sequence_cost" algorithm = GradientDescent( cost = cost, parameters = list(Selector(generator).get_parameters().values()), step_rule = Adam(), # because we want use all the stuff in the training data on_unused_sources = 'ignore' ) main_loop = MainLoop( algorithm=algorithm, data_stream = DataStream( train_data, iteration_scheme = SequentialScheme( train_data.num_examples,
def train(): if os.path.isfile('trainingdata.tar'): with open('trainingdata.tar', 'rb') as f: main = load(f) else: hidden_size = 512 train_dataset = dataset.T_H5PYDataset( 'dataset/wikifonia-seqlen-100.txt.hdf5', which_sets=('train', )) alphabet_len = train_dataset.vocab_size() x = theano.tensor.lmatrix('inchar') recurrent_block = LSTM(dim=hidden_size, activation=Tanh()) recurrent_block2 = LSTM(dim=hidden_size, activation=Tanh()) recurrent_block3 = LSTM(dim=hidden_size, activation=Tanh()) transition = RecurrentStack( [recurrent_block, recurrent_block2, recurrent_block3]) readout = Readout(readout_dim=alphabet_len, feedback_brick=LookupFeedback(alphabet_len, hidden_size, name='feedback'), source_names=[ thing for thing in transition.apply.states if "states" in thing ], emitter=RandomSoftmaxEmitter(), name='readout') gen = SequenceGenerator(readout=readout, transition=transition, weights_init=Uniform(width=0.02), biases_init=Uniform(width=0.0001), name='sequencegenerator') gen.push_initialization_config() gen.initialize() cost = gen.cost(outputs=x) cost.name = 'cost' cg = ComputationGraph(cost) step_rules = [Adam(), StepClipping(1.0)] algorithm = GradientDescent(cost=cost, parameters=cg.parameters, step_rule=CompositeRule(step_rules), on_unused_sources='ignore') train_stream = DataStream.default_stream( train_dataset, iteration_scheme=SequentialScheme(train_dataset.num_examples, batch_size=20)) main = MainLoop(model=Model(cost), data_stream=train_stream, algorithm=algorithm, extensions=[ FinishAfter(), Printing(), Checkpoint('trainingdata.tar', every_n_epochs=10), ShowOutput(every_n_epochs=10) ]) main.run()