def predict_state(state_logit, conv_lens):
batch_size, max_conv_len, num_asv = state_logit.size()
mask = ((state_logit == float("-inf")) |
(state_logit == float("inf")))
pred = (torch.sigmoid(state_logit.masked_fill(
mask, 0)).masked_fill(mask, 0)) > 0.5
pred = utils.to_sparse(pred.view(-1, num_asv))
return utils.DoublyStacked1DTensor(
value=pred.value.view(batch_size, max_conv_len, -1),
lens=conv_lens,
lens1=pred.lens.view(batch_size, max_conv_len))
def test_turn_state_encoder_decoder():
dataset = create_dummy_dataset()
vocabs = list(dataset.vocabs.turn.slot_values.values())
encoder = GenericStateEncoder(
vocabs=vocabs,
output_dim=100,
label_encoder=functools.partial(
EmbeddingLabelEncoder
),
label_layer=feedforward.MultiLayerFeedForward,
label_pooling=pooling.SumPooling,
state_pooling=pooling.MaxPooling,
output_layer=feedforward.MultiLayerFeedForward
)
decoder = GenericStateDecoder(
input_dim=100,
vocabs=vocabs,
input_layer=feedforward.MultiLayerFeedForward,
output_layer=feedforward.MultiLayerFeedForward,
label_emb=EmbeddingLabelEncoder
)
encoder.reset_parameters()
decoder.reset_parameters()
encoder.train(), decoder.train()
params = [p for p in encoder.parameters() if p.requires_grad]
params += [p for p in decoder.parameters() if p.requires_grad]
optimizer = op.Adam(params)
bce = nn.BCEWithLogitsLoss(reduction="none")
vocab_lens = torch.LongTensor(list(map(len, vocabs)))
x_sparse = torch.randint(0, 2, (4, len(vocab_lens), max(vocab_lens))).byte()
x_sparse = x_sparse.masked_fill(~utils.mask(vocab_lens), 0)
x, lens = utils.to_sparse(x_sparse)
x_sparse = x_sparse.masked_fill(~utils.mask(vocab_lens), -1)
lens = torch.randint(0, 3, (4, len(encoder.vocabs))) + 1
for i in range(100):
logits = decoder(encoder(x, lens))
loss = bce(logits, x_sparse.float())
loss = loss.masked_fill(~utils.mask(vocab_lens), 0).sum()
print(i, loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
encoder.eval(), decoder.eval()
logits = decoder(encoder(x, lens))
x_pred = torch.sigmoid(logits) > 0.5
x_pred = x_pred.masked_fill(~utils.mask(vocab_lens), -1)
assert (x_pred == x_sparse).all().item()
def train():
start = time.time()
from lightfm.evaluation import precision_at_k, recall_at_k, auc_score
from utils import clean_data, to_sparse, create_user_dict, create_item_dict, fit_mf_model
from utils import items_to_user, items_to_item, create_item_emdedding_distance_matrix, users_to_item
print('Modules loaded...')
training_metrics={}
data = pd.read_csv(training_data)
piv, cols, interactions_ = to_sparse(data)
interactions_.to_csv(interactions, index=True)
user_dict_ = create_user_dict(interactions=interactions_)
item_dict_ = create_item_dict(df = data, id_col = 'StockCode', name_col = 'Description')
with open(user_dict, 'w') as json_file:
json.dump(user_dict_, json_file)
with open(item_dict, 'w') as json_file:
json.dump(item_dict_, json_file)
print('Data preparations ready...')
mf_model = fit_mf_model(interactions = interactions_,
n_components = 140,
loss = 'warp',
epoch = 10,
n_jobs = 6)
print('Model fit...')
training_metrics["precision_at_3"] = round(precision_at_k(mf_model, piv, k=3).mean()*100)
training_metrics["recall_at_3"] = round(recall_at_k(mf_model, piv, k=3).mean()*100)
training_metrics["auc_score"]=round(auc_score(mf_model, piv).mean()*100)
pickle.dump(mf_model, open(str(model_directory + "/" +"recomender.pkl"), "wb"))
print('Model trained & serialized in %.1f seconds' % (time.time() - start))
return jsonify(training_metrics)
def e_step(votes_ij, activations_j, mean_j, stdv_j, var_j, spatial_routing_matrix):
"""The e-step in EM routing between input capsules (i) and output capsules (j).
Update the assignment weights using in routung. The output capsules (j)
compete for the input capsules (i).
See Hinton et al. "Matrix Capsules with EM Routing" for detailed description
of e-step.
Author:
Ashley Gritzman 19/10/2018
Args:
votes_ij:
votes from capsules in layer i to capsules in layer j
For conv layer:
(N, OH, OW, kh*kw*i, o, 4x4)
(64, 6, 6, 9*8, 32, 16)
For FC layer:
The kernel dimensions are equal to the spatial dimensions of the input
layer i, and the spatial dimensions of the output layer j are 1x1.
(N, 1, 1, child_space*child_space*i, output_classes, 4x4)
(64, 1, 1, 4*4*16, 5, 16)
activations_j:
activations of capsules in layer j (L+1)
(N, OH, OW, 1, o, 1)
(64, 6, 6, 1, 32, 1)
mean_j:
mean of each channel in capsules of layer j (L+1)
(N, OH, OW, 1, o, n_channels)
(24, 6, 6, 1, 32, 16)
stdv_j:
standard deviation of each channel in capsules of layer j (L+1)
(N, OH, OW, 1, o, n_channels)
(24, 6, 6, 1, 32, 16)
var_j:
variance of each channel in capsules of layer j (L+1)
(N, OH, OW, 1, o, n_channels)
(24, 6, 6, 1, 32, 16)
spatial_routing_matrix: ???
Returns:
rr:
assignment weights between capsules in layer i and layer j
(N, OH, OW, kh*kw*i, o, 1)
(64, 6, 6, 9*8, 16, 1)
"""
with tf.variable_scope("e_step") as scope:
# AG 26/06/2018: changed stdv_j to var_j
o_p_unit0 = - tf.reduce_sum(
tf.square(votes_ij - mean_j, name="num") / (2 * var_j),
axis=-1,
keepdims=True,
name="o_p_unit0")
o_p_unit2 = - 0.5 * tf.reduce_sum(
tf.log(2*np.pi * var_j),
axis=-1,
keepdims=True,
name="o_p_unit2"
)
# (24, 6, 6, 288, 32, 1)
o_p = o_p_unit0 + o_p_unit2
zz = tf.log(activations_j + FLAGS.epsilon) + o_p
# AG 13/11/2018: New implementation of normalising across parents
#----- Start -----#
zz_shape = zz.get_shape().as_list()
batch_size = zz_shape[0]
parent_space = zz_shape[1]
kh_kw_i = zz_shape[3]
parent_caps = zz_shape[4]
kk = int(np.sum(spatial_routing_matrix[:,0]))
child_caps = int(kh_kw_i / kk)
zz = tf.reshape(zz, [batch_size, parent_space, parent_space, kk,
child_caps, parent_caps])
"""
# In un-log space
with tf.variable_scope("to_sparse_unlog") as scope:
zz_unlog = tf.exp(zz)
#zz_sparse_unlog = utl.to_sparse(zz_unlog, spatial_routing_matrix,
# sparse_filler=1e-15)
zz_sparse_unlog = utl.to_sparse(
zz_unlog,
spatial_routing_matrix,
sparse_filler=0.0)
# maybe this value should be even lower 1e-15
zz_sparse_log = tf.log(zz_sparse_unlog + 1e-15)
zz_sparse = zz_sparse_log
"""
# In log space
with tf.variable_scope("to_sparse_log") as scope:
# Fill the sparse matrix with the smallest value in zz (at least -100)
sparse_filler = tf.minimum(tf.reduce_min(zz), -100)
# sparse_filler = -100
zz_sparse = utl.to_sparse(
zz,
spatial_routing_matrix,
sparse_filler=sparse_filler)
with tf.variable_scope("softmax_across_parents") as scope:
rr_sparse = utl.softmax_across_parents(zz_sparse, spatial_routing_matrix)
with tf.variable_scope("to_dense") as scope:
rr_dense = utl.to_dense(rr_sparse, spatial_routing_matrix)
rr = tf.reshape(
rr_dense,
[batch_size, parent_space, parent_space, kh_kw_i, parent_caps, 1])
#----- End -----#
# AG 02/11/2018
# In response to a question on OpenReview, Hinton et al. wrote the
# following:
# "The gradient flows through EM algorithm. We do not use stop gradient. A
# routing of 3 is like a 3 layer network where the weights of layers are
# shared."
# https://openreview.net/forum?id=HJWLfGWRb¬eId=S1eo2P1I3Q
return rr
def test_dense_sparse():
x = torch.randint(0, 2, (3, 4, 5)).byte()
y = utils.to_dense(*utils.to_sparse(x))
assert (x == y).all()
if ARG_do_bounce == 'true':
do_bounce = True
else:
do_bounce = False
if ARG_type == 'seq':
ratings_str = ','.join([str(o.rating) for o in seq])
vec_str = ','.join([':'.join(str(s) for s in o.vec) for o in seq])
ret = []
ret.append(qid)
ret.append(str(seq_len))
ret.append(ratings_str)
ret.append(','.join([str(o[0]) for o in fb_seq]))
ret.append(','.join([str(o[1]) for o in fb_seq]))
ret.append(vec_str)
print ';'.join(ret)
elif ARG_type == 'single':
for index, o in enumerate(seq):
ret = []
ret.append(str(fb_seq[index][0]))
ret.append(qid)
ret.append(utils.to_sparse(o.vec))
ret.append('#')
ret.append('rating=' + str(o.rating))
print ' '.join(ret)
if do_bounce and fb_seq[index][1] == 1:
break
parser.add_argument('--dropout', type=float, default=0.15)
parser.add_argument('--lr', type=float, default=0.01)
parser.add_argument('--infusion', type=str, default='inner')
parser.add_argument('--dataset', type=str, default='cora')
parser.add_argument('--sparse', dest='sparse', action='store_true')
parser.add_argument('--no-sparse', dest='sparse', action='store_false')
parser.set_defaults(sparse=True)
args = parser.parse_args()
# Load data
adj_1, features, labels, idx_train, idx_val, idx_test = load_data(args.dataset)
adj_3 = onp.linalg.matrix_power(adj_1, 3)
adj_5 = onp.linalg.matrix_power(adj_1, 5)
if args.sparse:
adj_1 = to_sparse(adj_1) # custom format
adj_3 = to_sparse(adj_3) # custom format
adj_5 = to_sparse(adj_5) # custom format
adj = (adj_3, adj_3) # the k-hop adj used in each layer
rng_key = random.PRNGKey(args.seed)
dropout = args.dropout
step_size = args.lr
hidden = args.hidden
num_epochs = args.epochs
n_nodes = features.shape[0]
n_feats = features.shape[1]
infusion = args.infusion
init_fun, predict_fun = GHNet(nhid=hidden,
if __name__ == "__main__":
dilation = [1, 1, 1, 1]
seq_length = [9, 9, 9, 9]
transform_fp = os.path.join("data", "CoMA", "transform.pkl")
with open(transform_fp, 'rb') as f:
tmp = pickle.load(f, encoding='latin1')
spiral_indices_list = [
utils.preprocess_spiral(tmp['face'][idx], seq_length[idx],
tmp['vertices'][idx], dilation[idx]).to(device)
for idx in range(len(tmp['face']) - 1)
]
down_transform_list = [
utils.to_sparse(down_transform).to(device)
for down_transform in tmp['down_transform']
]
up_transform_list = [
utils.to_sparse(up_transform).to(device)
for up_transform in tmp['up_transform']
]
meshdata = MeshData("data/CoMA",
"data/CoMA/template/template.obj",
split="interpolation",
test_exp="bareteeth")
mean = meshdata.mean
std = meshdata.std