num_layers=num_layers,
logit_class_weights=[1, 100 * c],
count_class_weights=[1, 3])
batch_size = 64
open_every = 10
save_every = 500
log_every = 10
name = "{}-{}_{}".format(str(net), env_type, s)
c2 = 1
#%%
for i in range(10000000):
if i % open_every == 0:
X, Y, GN = open_file(env_type=env_type, open_GN=True, open_A=False)
SS = np.concatenate([X, GN], 2).transpose((1, 0, 2))
n = SS.shape[1]
idx = np.random.choice(n, size=batch_size)
inputs = SS[:, idx, :]
ytrue = Y[idx]
lens = inputs.any(2).sum(0)
avg_ones = np.hstack([Y[k, :l, 0] for k, l in zip(idx, lens)]).mean()
if avg_ones > 0:
w = c * 1 / avg_ones
net.logit_loss_fn = nn.CrossEntropyLoss(reduce=False,
weight=torch.FloatTensor(
dropout_prob = .4)
name = "{}-{}_{}".format(
str(net),
env_type,
s)
c = 1
c2 = .25
#%%
for i in range(int(1e8)):
if i % open_every == 0:
print("new file!")
if use_gn:
A, X, GN, Y = open_file(env_type = env_type,
open_A = True,
open_GN = True)
Z = np.concatenate([X, GN], 2)
else:
A, X, Y = open_file(env_type = env_type,
open_A = True,
open_GN = False)
Z = X
n = A.shape[0]
idx = np.random.choice(n, size = batch_size)
inputs = (A[idx], Z[idx])
ytrue = Y[idx]
lens = inputs[1].any(2).sum(1)
avg_ones = np.hstack([Y[k,:l,0] for k,l in zip(idx, lens)]).mean()
decoder_hidden_size = 100
bidirectional = True
num_layers = 1
batch_size = 32
rnn = torch.load("results/policy_function_lstm")
batch_size = 32
open_every = 100
log_every = 10
save_every = 500
#%%
for i_iter in range(1000):
if i % open_ == 0:
X, Y, GN = open_file(open_GN=True, open_A=False)
SS = np.concatenate([X, GN], 2).transpose((1, 0, 2))
n = SS.shape[1]
num_ys = None
optimizer.zero_grad()
idx = np.random.choice(n, size=batch_size)
subset = SS[:, idx, :]
lens = np.argmax(subset.any(2) == 0, 0)
inputs = Variable(torch.FloatTensor(subset), requires_grad=False)
order = np.flip(np.argsort(lens), 0).astype(int)
order_t = torch.LongTensor(order)