diff options
| -rw-r--r-- | gentext.py | 61 | ||||
| -rwxr-xr-x | train.py | 60 |
2 files changed, 66 insertions, 55 deletions
diff --git a/gentext.py b/gentext.py new file mode 100644 index 0000000..b8a27bf --- /dev/null +++ b/gentext.py @@ -0,0 +1,61 @@ +import sys + +import numpy + +import theano +from theano import tensor + +from blocks.extensions import SimpleExtension +from blocks.graph import ComputationGraph + +class GenText(SimpleExtension): + def __init__(self, model, init_text, max_bytes, sample_temperature, **kwargs): + super(GenText, self).__init__(**kwargs) + + self.init_text = init_text + self.max_bytes = max_bytes + + out = model.out[:, -1, :] / numpy.float32(sample_temperature) + prob = tensor.nnet.softmax(out) + + cg = ComputationGraph([prob]) + assert(len(cg.inputs) == 1) + assert(cg.inputs[0].name == 'bytes') + + state_vars = [theano.shared(v[0:1, :].zeros_like().eval(), v.name+'-gen') + for v, _ in model.states] + givens = [(v, x) for (v, _), x in zip(model.states, state_vars)] + updates= [(x, upd) for x, (_, upd) in zip(state_vars, model.states)] + + self.f = theano.function(inputs=cg.inputs, outputs=[prob], + givens=givens, updates=updates) + self.reset_states = theano.function(inputs=[], outputs=[], + updates=[(v, v.zeros_like()) for v in state_vars]) + + def do(self, which_callback, *args): + + print "Sample:" + print "-------" + + self.reset_states() + + v = numpy.array([ord(i) for i in self.init_text], + dtype='int16')[None, :] + prob, = self.f(v) + + sys.stdout.write(self.init_text) + while v.shape[1] < self.max_bytes: + prob = prob / 1.00001 + pred = numpy.random.multinomial(1, prob[0, :]).nonzero()[0][0] + + v = numpy.concatenate([v, pred[None, None]], axis=1) + sys.stdout.write(chr(int(pred))) + sys.stdout.flush() + + prob, = self.f(pred[None, None]) + print + print "-------" + print + + + @@ -22,7 +22,7 @@ from blocks.model import Model from blocks.algorithms import GradientDescent import datastream -# from apply_model import Apply +import gentext logging.basicConfig(level='INFO') logger = logging.getLogger(__name__) @@ -36,57 +36,6 @@ if __name__ == "__main__": model_name = sys.argv[1] config = importlib.import_module('%s' % model_name) - -class GenText(SimpleExtension): - def __init__(self, model, init_text, max_bytes, **kwargs): - super(GenText, self).__init__(**kwargs) - - self.init_text = init_text - self.max_bytes = max_bytes - - out = model.out[:, -1, :] / numpy.float32(config.sample_temperature) - prob = tensor.nnet.softmax(out) - - cg = ComputationGraph([prob]) - assert(len(cg.inputs) == 1) - assert(cg.inputs[0].name == 'bytes') - - state_vars = [theano.shared(v[0:1, :].zeros_like().eval(), v.name+'-gen') - for v, _ in model.states] - givens = [(v, x) for (v, _), x in zip(model.states, state_vars)] - updates= [(x, upd) for x, (_, upd) in zip(state_vars, model.states)] - - self.f = theano.function(inputs=cg.inputs, outputs=[prob], - givens=givens, updates=updates) - self.reset_states = theano.function(inputs=[], outputs=[], - updates=[(v, v.zeros_like()) for v in state_vars]) - - def do(self, which_callback, *args): - - print "Sample:" - print "-------" - - self.reset_states() - - v = numpy.array([ord(i) for i in self.init_text], - dtype='int16')[None, :] - prob, = self.f(v) - - sys.stdout.write(self.init_text) - while v.shape[1] < self.max_bytes: - prob = prob / 1.00001 - pred = numpy.random.multinomial(1, prob[0, :]).nonzero()[0][0] - - v = numpy.concatenate([v, pred[None, None]], axis=1) - sys.stdout.write(chr(int(pred))) - sys.stdout.flush() - - prob, = self.f(pred[None, None]) - print - print "-------" - print - - class ResetStates(SimpleExtension): def __init__(self, state_vars, **kwargs): super(ResetStates, self).__init__(**kwargs) @@ -142,9 +91,10 @@ def train_model(m, train_stream, dump_path=None): server_url='http://eos21:4201/', every_n_epochs=1, after_epoch=False), - GenText(m, '\nalex\ttu crois ?\n', config.sample_len, - every_n_epochs=config.sample_freq, - after_epoch=False, before_training=True), + gentext.GenText(m, '\nalex\ttu crois ?\n', + config.sample_len, config.sample_temperature, + every_n_epochs=config.sample_freq, + after_epoch=False, before_training=True), ResetStates([v for v, _ in m.states], after_epoch=True) ] ) |