def __init__(self, model):
self.model = model
self.dropout = 1.0
self.global_step = tf.Variable(0, name='global_step', trainable=False)
self.input = dataset.Batch(
tf.placeholder(tf.int32, shape=[None, 366, 100, 3]),
tf.placeholder(tf.int32,
shape=[None,
self.model.data.token_sequence_length]))
self.feature_embedding
self.targets
self.convolution
self.decoder
self.logits
self.loss
self.accuracy
self.optimize
def __init__(self, model):
self.model = model
self.dropout = tf.placeholder(tf.float32, shape=[])
self.sparsity_adjustment = 0.9
self.total_loss_ratio = 0.5 # portion of loss coming from autoencoder
self.global_step = tf.Variable(0, name='global_step', trainable=False)
self.input = dataset.Batch(
tf.placeholder(tf.int32, shape=[None, 366, 100, 3]),
tf.placeholder(tf.int32,
shape=[None,
self.model.data.token_sequence_length]))
self.feature_embedding
self.targets
self.convolution
# self.autoencoder
# self.autoencoder_loss
self.encoded
self.decoder
self.logits
self.loss
# self.total_loss
self.accuracy
self.optimize
def test_incremental():
from common import config
import model
from utils import get_batch
net, _ = model.get()
net.eval()
ckpt = torch.load("checkpoints/checkpoint_best_ppl.pth", map_location='cpu')
# reload model parameters
s_dict = {}
for k in ckpt["net"]:
new_k = k[7:]
s_dict[new_k] = ckpt["net"][k]
net.load_state_dict(s_dict)
import dataset
train_iter, _, SRC_TEXT, TGT_TEXT = dataset.get()
#data_iter = iter(train_iter.get_iterator(True, True))
#raw_batch = next(data_iter)
src = np.arange(4, 4+2000).reshape(80, 25)
tgt = np.arange(4, 4+2400).reshape(80, 30)
raw_batch = dataset.Batch(
torch.from_numpy(src).long(),
torch.from_numpy(tgt).long()
)
batch = get_batch(
raw_batch.src, raw_batch.tgt,
SRC_TEXT.vocab, TGT_TEXT.vocab
)
for k, v in batch.items():
try:
print(k, v.size())
except AttributeError:
pass
with torch.no_grad():
enc_out = net.encode(src=batch['src'], src_mask=batch['src_mask'])
# No incremental
logits1 = net.decode(enc_out, batch['src_mask'], batch['tgt'], batch['tgt_mask'])
logits1 = net.generator(logits1, log_prob=True)
# Incremental
print("Incremental encoding finished!")
tlen = batch['tgt'].size(1)
cache = {'cur_len':0}
logits2 = []
for i in range(tlen):
x = batch['tgt'][:, i].unsqueeze(-1)
logit = net.decode(
enc_out, batch['src_mask'], x,
batch['tgt_mask'][:, i, :(i+1)].unsqueeze(-2), cache
)
logit = net.generator(logit, log_prob=True)
if i >= 0:
ref = logits1[:, i, :]
sys = logit.squeeze()
ref_words = torch.topk(ref, 1)[1].squeeze()
sys_words = torch.topk(sys, 1)[1].squeeze()
print("Diff = {}".format(torch.sum(ref - sys).item()))
print("Logits sys size : {}, Logits sys : {}".format(sys.size(), sys.sum().item()))
print("Logits ref size : {}, Logits ref : {}".format(ref.size(), ref.sum().item()))
if (ref_words == sys_words).all() == False:
print("F**k!")
print("\n")
logits2.append(logit)
cache['cur_len'] = i + 1
logits2 = torch.cat(logits2, dim=1).contiguous()
print("Logits1: {}".format(torch.sum(logits1).item()))
print("Logits2: {}".format(torch.sum(logits2).item()))