def test_embedding_layer(): random_state = np.random.RandomState(1999) graph = OrderedDict() max_index = 100 proj_dim = 12 fake_str_int = [[1, 5, 7, 1, 6, 2], [2, 3, 6, 2], [3, 3, 3, 3, 3, 3, 3]] minibatch, mask = make_embedding_minibatch( fake_str_int, slice(0, 3)) (emb_slices,), (emb_mask,) = add_embedding_datasets_to_graph( [minibatch], [mask], "emb", graph) emb = embedding_layer(emb_slices, max_index, proj_dim, graph, 'emb', random_state) followup_dim = 17 proj = linear_layer([emb], graph, 'proj', followup_dim, random_state=random_state) f = theano.function(emb_slices, [proj], mode="FAST_COMPILE") out, = f(*minibatch) assert(out.shape[-1] == 17) assert(out.shape[-2] == len(fake_str_int))
def test_embedding_layer(): random_state = np.random.RandomState(1999) graph = OrderedDict() max_index = 100 proj_dim = 12 fake_str_int = [[1, 5, 7, 1, 6, 2], [2, 3, 6, 2], [3, 3, 3, 3, 3, 3, 3]] minibatch, mask = make_embedding_minibatch(fake_str_int, slice(0, 3)) (emb_slices, ), (emb_mask, ) = add_embedding_datasets_to_graph([minibatch], [mask], "emb", graph) emb = embedding_layer(emb_slices, max_index, proj_dim, graph, 'emb', random_state) followup_dim = 17 proj = linear_layer([emb], graph, 'proj', followup_dim, random_state) f = theano.function(emb_slices, [proj], mode="FAST_COMPILE") out, = f(*minibatch) assert (out.shape[-1] == 17) assert (out.shape[-2] == len(fake_str_int))
n_hid = 100 X_story_mb, X_story_mask = make_embedding_minibatch( X_story, slice(0, minibatch_size)) X_query_mb, X_query_mask = make_embedding_minibatch( X_query, slice(0, minibatch_size)) embedding_datasets = [X_story_mb, X_query_mb] masks = [X_story_mask, X_query_mask] r = add_embedding_datasets_to_graph(embedding_datasets, masks, "babi_data", graph) (X_story_syms, X_query_syms), (X_story_mask_sym, X_query_mask_sym) = r y_sym = add_datasets_to_graph([y_answer], ["y"], graph) l1_story = embedding_layer(X_story_syms, vocab_size, n_emb, graph, 'l1_story', random_state) masked_story = X_story_mask_sym.dimshuffle(0, 1, 'x') * l1_story h_story = gru_recurrent_layer([masked_story], X_story_mask_sym, n_hid, graph, 'story_rec', random_state) l1_query = embedding_layer(X_query_syms, vocab_size, n_emb, graph, 'l1_query', random_state) h_query = gru_recurrent_layer([l1_query], X_query_mask_sym, n_hid, graph, 'query_rec', random_state) y_pred = softmax_layer([h_query[-1], h_story[-1]], graph, 'y_pred', y_answer.shape[1], random_state) cost = categorical_crossentropy(y_pred, y_sym).mean() params, grads = get_params_and_grads(graph, cost) opt = adam(params) learning_rate = 0.001
n_hid = 100 X_story_mb, X_story_mask = make_embedding_minibatch( X_story, slice(0, minibatch_size)) X_query_mb, X_query_mask = make_embedding_minibatch( X_query, slice(0, minibatch_size)) embedding_datasets = [X_story_mb, X_query_mb] masks = [X_story_mask, X_query_mask] r = add_embedding_datasets_to_graph(embedding_datasets, masks, "babi_data", graph) (X_story_syms, X_query_syms), (X_story_mask_sym, X_query_mask_sym) = r y_sym = add_datasets_to_graph([y_answer], ["y"], graph) l1_story = embedding_layer(X_story_syms, vocab_size, n_emb, graph, 'l1_story', random_state=random_state) masked_story = X_story_mask_sym.dimshuffle(0, 1, 'x') * l1_story h_story = gru_recurrent_layer([masked_story], X_story_mask_sym, n_hid, graph, 'story_rec', random_state) l1_query = embedding_layer(X_query_syms, vocab_size, n_emb, graph, 'l1_query', random_state) h_query = gru_recurrent_layer([l1_query], X_query_mask_sym, n_hid, graph, 'query_rec', random_state) y_pred = softmax_layer([h_query[-1], h_story[-1]], graph, 'y_pred', y_answer.shape[1], random_state=random_state) cost = categorical_crossentropy(y_pred, y_sym).mean() params, grads = get_params_and_grads(graph, cost) opt = adadelta(params) updates = opt.updates(params, grads)