def forward(self,
query: Tensor,
key: Tensor,
value: Tensor,
mask: Tensor = None) -> Tensor:
"""
@param query shape -> [batch_size, max_length, emb_size]
@param key shape -> [batch_size, max_length, emb_size]
@param value shape -> [batch_size, max_length, emb_size]
@param mask shape -> [1, max_length, max_length]
@return a tensor with shape -> ??
"""
if mask is not None:
# 1, n, n -> 1, 1, n, n; n is max length of sentence
mask = mask.unsqueeze(1)
batch_size = query.size(0)
# do projection
query, key, value = [
linear_f(x).view(batch_size, -1, self.head_count,
self.model_k_dim).transpose(1, 2)
for linear_f, x in zip(self.linears, (query, key, value))
]
# do attention
x, self.attn = attention(query, key, value, mask, self.dropout)
# do concatenation
x = x.transpose(1, 2).contiguous().view(
batch_size, -1, self.head_count * self.model_k_dim)
return self.linears[-1](x)
def __init__(self, input_size: int, hidden_size: int, bias: bool = True, p: Tuple[float, float] = (0.5, 0.5),
initializer: Callable[[Tensor], None] = None) -> None:
super(VarLSTMCell, self).__init__()
self.input_size: int = input_size
self.hidden_size: int = hidden_size
self.bias: bool = bias
self.weight_ih: Tensor = Parameter(Tensor(4, input_size, hidden_size))
self.weight_hh: Tensor = Parameter(Tensor(4, hidden_size, hidden_size))
if bias:
self.bias_ih: Tensor = Parameter(Tensor(4, hidden_size))
self.bias_hh: Tensor = Parameter(Tensor(4, hidden_size))
else:
self.register_parameter('bias_ih', None)
self.register_parameter('bias_hh', None)
self.initializer: Callable[[Tensor], None] \
= default_initializer(self.hidden_size) if initializer is None else initializer
self.reset_parameters()
p_in, p_hidden = p
if p_in < 0. or p_in > 1.:
raise ValueError("input dropout probability has to be between 0 and 1, "
"but got {}".format(p_in))
if p_hidden < 0. or p_hidden > 1.:
raise ValueError("hidden state dropout probability has to be between 0 and 1, "
"but got {}".format(p_hidden))
self.p_in: float = p_in
self.p_hidden: float = p_hidden
self.noise_in: Tensor = None
self.noise_hidden: Tensor = None
def __init__(self, n_in:int, n_out:int, objective:str, y_range:Optional[Union[Tuple,np.ndarray]]=None, bias_init:Optional[float]=None,
y_mean:Optional[Union[float,List[float],np.ndarray]]=None, y_std:Optional[Union[float,List[float],np.ndarray]]=None,
lookup_init:Callable[[str,Optional[int],Optional[int]],Callable[[Tensor],None]]=lookup_normal_init, freeze:bool=False):
super().__init__(n_in=n_in, n_out=n_out, objective=objective, bias_init=bias_init, lookup_init=lookup_init, freeze=freeze)
self.y_range,self.y_mean,self.y_std = y_range,y_mean,y_std
self.rescale = False
if self.y_range is not None and (self.y_mean is not None or self.y_std is not None):
raise ValueError("Both y_range (sigmoid output and rescaling) and y_mean + y_std (linear output and rescaling) are set. Please only set either.")
if (self.y_mean is None and self.y_std is not None) or (self.y_mean is not None and self.y_std is None):
raise ValueError("Only one of y_mean or y_std is set, but not both. Please set both or neither.")
if self.y_mean is not None and self.y_std is not None and bias_init is not None:
print("y_mean and y_std are both set, but so is bias_init. Bias init will be set to zero to provide accurate rescaling")
self.bias_init = None
if self.y_range is not None:
if not isinstance(self.y_range, np.ndarray): self.y_range = np.array(self.y_range)
self.y_min = np.array(np.min(self.y_range, axis=-1), dtype='float32')
self.y_diff = np.abs(self.y_range.take([1], axis=-1)-self.y_range.take([0], axis=-1)).ravel()
self.y_min, self.y_diff = to_device(Tensor(self.y_min)), to_device(Tensor(self.y_diff))
elif self.y_mean is not None and self.y_std is not None:
if not hasattr(self.y_mean, 'len'): self.y_mean = [self.y_mean]
if not hasattr(self.y_mean, 'len'): self.y_std = [self.y_std]
self.y_mean,self.y_std = to_device(Tensor(self.y_mean)),to_device(Tensor(self.y_std))
self.rescale = True
self._build_layers()
if self.freeze: self.freeze_layers()
def forward(self, x: Tensor) -> Tensor:
x = self.melspectrogram(x)
x = x.unsqueeze(1)
x = self.norm_input(x)
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = self.maxpool(x)
x = self.drop1(x)
x = F.relu(self.conv3(x))
x = F.relu(self.conv4(x))
x = self.maxpool(x)
x = self.drop2(x)
x = F.relu(self.conv5(x))
x = self.maxpool(x)
x = self.drop2(x)
x = F.relu(self.conv6(x))
x = self.maxpool(x)
x = self.drop2(x)
x = F.relu(self.conv7(x))
x = self.maxpool(x)
x = self.drop2(x)
x = self.fc(x.flatten(start_dim=1))
x = self.fc_norm(x)
return self.linear(x)
def __init__(self,
n_in: int,
n_out: int,
objective: str,
y_range: Optional[Union[Tuple, np.ndarray]] = None,
bias_init: Optional[float] = None,
lookup_init: Callable[[str, Optional[int], Optional[int]],
Callable[[Tensor],
None]] = lookup_normal_init,
freeze: bool = False):
super().__init__(n_in=n_in,
n_out=n_out,
objective=objective,
bias_init=bias_init,
lookup_init=lookup_init,
freeze=freeze)
self.y_range = y_range
if self.y_range is not None:
if not isinstance(self.y_range, np.ndarray):
self.y_range = np.array(self.y_range)
self.y_min = np.array(np.min(self.y_range, axis=-1),
dtype='float32')
self.y_diff = np.abs(
self.y_range.take([1], axis=-1) -
self.y_range.take([0], axis=-1)).ravel()
self.y_min, self.y_diff = to_device(Tensor(self.y_min)), to_device(
Tensor(self.y_diff))
self._build_layers()
if self.freeze: self.freeze_layers()
def multi_class_correct(y_pre: Tensor,
y_true: Tensor,
threshold=0.5,
device='cpu') -> Tensor:
y_pre, y_true = y_pre.argmax(dim=1), y_true.argmax(dim=1)
same = torch.as_tensor(y_pre == y_true, dtype=torch.int).to(device)
return torch.sum(same)
def tolist(paired_wavs: List[Tensor], paired_feature: Tensor):
assert paired_feature.dim() == 3
# (batch_size, max_seq_len, feat_dim)
ratio = max([len(wav) for wav in paired_wavs]) / paired_feature.size(1)
feature_len = [round(len(wav) / ratio) for wav in paired_wavs]
feature = [f[:l] for f, l in zip(paired_feature, feature_len)]
return feature
def runModel(coreTensor, inputSize: int, hiddenSize: int, d, model):
'''
This function runs LSTM for one-way
:param coreTensor:
:return:
'''
# d = D('cuda')
# logging.info('Core Tensor dumping.......')
# pickle.dump(coreTensor, open('coreTensor.pkl', 'wb'))
# logging.info('Core Tensor dumped.')
logging.info('Core Tensor loading.......')
coreTensor = pickle.load(open('coreTensor.pkl', 'rb'))
logging.info('Core Tensor loaded.')
coreLSTM = nn.LSTM(input_size=inputSize, hidden_size=hiddenSize).to(d)
batch = 0
logging.info('LSTM Start.(Core-One way)')
retList = []
sum_time = 0
for news_tensor in coreTensor:
time_start = time.time()
inputs_2 = []
for word in news_tensor:
inputs_2.append(model[word])
inputs = Tensor(inputs_2).to(d)
hidden = (torch.rand(1, 1, hiddenSize).to(d), torch.rand(1, 1, hiddenSize).to(d))
iter_time = 0
out = Tensor().to(d)
out_list = []
for i in inputs:
time_per_iter = time.time()
out, hidden = coreLSTM(i.view(1, 1, -1), hidden)
out_list = [out]
# logging.info('\t\t + Iter: %d/%d Time: %.3lf s, remain %d iterations, ETA: %.3lf s' % (
# iter_time + 1, len(inputs), time.time() - time_per_iter, len(inputs) - iter_time - 1,
# (len(inputs) - iter_time - 1) * (time.time() - time_per_iter)))
iter_time = iter_time + 1
sum_time += int(time.time() - time_start)
avg_time = 1.0 * sum_time / (batch + 1)
remainSecTot = 1.0 * (len(coreTensor) - batch - 1) * (avg_time)
remainMin = (int(remainSecTot) % 3600) // 60
remainHour = int(remainSecTot) // 3600
remainSec = int(remainSecTot) % 60
logging.info('* News: %d/%d,Time: %.2lfs, remain %d news, ETA: %d:%d:%d s' % (
batch + 1, len(coreTensor), time.time() - time_start, len(coreTensor) - batch - 1, remainHour, remainMin, remainSec))
fileName = 'processVec/Direct-LSTM-Oneway_Sentence_' + str(batch) + '.pkl'
foutput = open(fileName, 'wb')
pickle.dump(out_list, foutput)
# retList.append(out)
batch = batch + 1
del out
del out_list
del inputs_2
del hidden
logging.info('LSTM Finish.(Core-One way)')
return retList
def evaluate(self,
inputs: Union[Tensor, np.ndarray, Tuple[Tensor, Tensor],
Tuple[np.ndarray, np.ndarray]],
targets: Union[Tensor, np.ndarray],
weights: Optional[Union[Tensor, np.ndarray]] = None,
callbacks: Optional[List[AbsCallback]] = None,
mask_inputs: bool = True) -> float:
r'''
Compute loss on provided data.
Arguments:
inputs: input data
targets: targets
weights: Optional weights
callbacks: list of any callbacks to use during evaluation
mask_inputs: whether to apply input mask if one has been set
Returns:
(weighted) loss of model predictions on provided data
'''
if callbacks is None: callbacks = []
self.model.eval()
if not isinstance(inputs, Tensor):
if isinstance(inputs, tuple):
if not isinstance(inputs[0], Tensor):
inputs = (to_device(Tensor(inputs[0]).float()),
to_device(Tensor(inputs[1]).float()))
else:
inputs = to_device(Tensor(inputs).float())
for c in callbacks:
c.on_eval_begin(inputs=inputs, targets=targets, weights=weights)
if self.input_mask is not None and mask_inputs:
if isinstance(inputs, tuple):
inputs[0] = inputs[0][:, self.input_mask]
else:
inputs = inputs[:, self.input_mask]
y_pred = self.model(inputs)
if not isinstance(targets, Tensor):
targets = to_device(Tensor(targets))
if weights is not None and not isinstance(weights, Tensor):
weights = to_device(Tensor(weights))
if 'multiclass' in self.objective and not isinstance(
targets, torch.LongTensor):
targets = targets.long().squeeze()
elif 'multiclass' not in self.objective and not isinstance(
targets, torch.FloatTensor):
targets = targets.float()
loss = self.loss(weight=weights)(
y_pred, targets) if weights is not None else self.loss()(y_pred,
targets)
for c in callbacks:
c.on_eval_end(loss=loss)
return loss.data.item()
def test_on_batch(self, inputs, outputs, verbose=0):
"""
:param inputs:
:param outputs:
:param verbose:
:return:
"""
grapheme_root = Tensor(outputs[0])
vowel_diacritic = Tensor(outputs[1])
consonant_diacritic = Tensor(outputs[2])
_, grapheme_root_tru = grapheme_root.max(1)
_, vowel_diacritic_tru = vowel_diacritic.max(1)
_, consonant_diacritic_tru = consonant_diacritic.max(1)
grapheme_root_hat, vowel_diacritic_hat, consonant_diacritic_hat = self.predict(inputs)
grapheme_root_hat = grapheme_root_hat.cpu()
vowel_diacritic_hat = vowel_diacritic_hat.cpu()
consonant_diacritic_hat = consonant_diacritic_hat.cpu()
if verbose == 1:
print(grapheme_root_tru[:10])
print(grapheme_root_hat[:10])
print(vowel_diacritic_tru[:10])
print(vowel_diacritic_hat[:10])
print(consonant_diacritic_tru[:10])
print(consonant_diacritic_hat[:10])
grapheme_root = accuracy_score(grapheme_root_tru, grapheme_root_hat)
vowel_diacritic = accuracy_score(vowel_diacritic_tru, vowel_diacritic_hat)
consonant_diacritic = accuracy_score(consonant_diacritic_tru, consonant_diacritic_hat)
print("accuracy: grapheme root {}, vowel diacritic {}, consonant diacritic {}".format(grapheme_root, vowel_diacritic, consonant_diacritic))
def logsumexp(x: Tensor, dim: int) -> Tensor:
"""
Args:
x: A pytorch tensor (any dimension will do)
dim: int, over which to perform the summation.
Returns: The result of the log(sum(exp(...))) operation.
"""
xmax, _ = x.max(dim, keepdim=True)
xmax_, _ = x.max(dim)
return xmax_ + torch.log(torch.exp(x - xmax).sum(dim))
def nests_loss(self, energy: Tensor, target: Tensor) -> Tensor:
"""
Args:
energy: Tensor
the energy tensor with shape = [length, num_label, num_label]
target: Tensor
the tensor of target labels with shape [length]
Returns: Tensor
A 0D tensor for minus log likelihood loss
"""
length, _, _ = energy.size()
num_label_3 = self.indices_is.size(0)
indices_3 = energy.new_empty((length, num_label_3)).long()
indices_3[0, :] = self.indices_bs
if length > 2:
indices_3[1:length - 1, :] = self.indices_is.repeat(
(length - 2, 1))
indices_3[length - 1, :] = self.indices_es
# shape = [num_label]
partition_1 = None
partition_3 = None
# shape = []
prev_label = self.index_bos
tgt_energy = 0
for t in range(length):
# shape = [num_label, num_label]
curr_energy = energy[t]
if t == 0:
partition_1 = curr_energy[self.index_bos, :]
partition_3 = energy.new_full((num_label_3, ), -1e4)
else:
# shape = [num_label]
partition = partition_1.clone()
partition[indices_3[t - 1]] = partition_3
partition_1 = logsumexp(curr_energy + partition_1.unsqueeze(1),
dim=0)
partition_3 = logsumexp(curr_energy[:, indices_3[t]] +
partition.unsqueeze(1),
dim=0)
label = target[t]
tgt_energy += curr_energy[prev_label, label]
prev_label = label
t = length - 1
curr_energy = self.trans_matrix.data[:, self.index_eos]
partition = curr_energy + partition_1
partition[indices_3[t]] = curr_energy[indices_3[t]] + partition_3
return logsumexp(partition, dim=0) - tgt_energy
def arange(start, end=None, step=1, out=None, **kwargs):
if end is None: # arange(N)
end = start
start = 0
if isinstance(start, int) and isinstance(end, int) and isinstance(step, int):
dtype = torch.intDefault
else:
dtype = torch.floatDefault
tensor = Tensor(int((end - start) / step))
tensor.dtype = dtype
LibCall.torch.copyOut(tensor, out)
return tensor
def get_inputs(
self,
on_device: bool = False) -> Union[Tensor, Tuple[Tensor, Tensor]]:
if on_device:
if self.matrix_inputs is None:
return to_device(Tensor(self.inputs))
else:
return (to_device(Tensor(self.inputs)),
to_device(Tensor(self.matrix_inputs)))
else:
if self.matrix_inputs is None: return self.inputs
else: return (self.inputs, self.matrix_inputs)
def calc_loss(self, q_values: Tensor, target_q_values: Tensor,
actions: Tensor, rewards: Tensor,
done_mask: Tensor) -> Tensor:
"""
Calculate the MSE loss of this step.
The loss for an example is defined as:
Q_samp(s) = r if done
= r + gamma * max_a' Q_target(s', a') otherwise
loss = (Q_samp(s) - Q(s, a))^2
Args:
q_values: (torch tensor) shape = (batch_size, num_actions)
The Q-values that your current network estimates (i.e. Q(s, a') for all a')
target_q_values: (torch tensor) shape = (batch_size, num_actions)
The Target Q-values that your target network estimates (i.e. (i.e. Q_target(s', a') for all a')
actions: (torch tensor) shape = (batch_size,)
The actions that you actually took at each step (i.e. a)
rewards: (torch tensor) shape = (batch_size,)
The rewards that you actually got at each step (i.e. r)
done_mask: (torch tensor) shape = (batch_size,)
A boolean mask of examples where we reached the terminal state
Hint:
You may find the following functions useful
- torch.max
- torch.sum
- torch.nn.functional.one_hot
- torch.nn.functional.mse_loss
You can treat `done_mask` as a 0 and 1 where 0 is not done and 1 is done using torch.type as
done below
To extract Q(a) for a specific "a" you can use the torch.sum and torch.nn.functional.one_hot.
Think about how.
"""
# you may need this variable
num_actions = self.env.action_space.n
gamma = self.config.gamma
done_mask = done_mask.type(torch.int)
actions = actions.type(torch.int64)
##############################################################
##################### YOUR CODE HERE - 3-5 lines #############
target_q = torch.reshape(
torch.max(target_q_values, dim=1, keepdim=True).values, (-1, ))
q_val1 = rewards + (1 - done_mask) * gamma * target_q
q_val2 = torch.sum(
q_values *
torch.nn.functional.one_hot(actions, self.env.action_space.n),
dim=1)
loss = torch.nn.functional.mse_loss(q_val1, q_val2)
##############################################################
######################## END YOUR CODE #######################
return loss
def forward(self, img: Tensor, labels: int):
"""Forward pass of the Discriminator.
Args:
img: the image that should be classified in fake or real.
labels: the label of the image.
Returns:
"""
d_in = torch.cat(
(img.view(img.size(0), -1), self.label_embedding(labels)), -1)
score = self.model(d_in)
return score
def forward(self, X: Tensor, initial_states=None):
#if self.init_states is None:
self.init_states = torch.zeros(
self.gru_hidden_layers * self.num_directions,
X.size(self.batch_index), self.hidden_dimensions)
# self.init_states = self.init_states.to(util.device)
# TODO
if X.shape[self.batch_index] != self.init_states.shape[1]:
pass
#
output_gru, initial_states = self.gru_encoder(X, self.init_states)
# TODO: if batchnorm handle differently
# if self.use_batchnorm:
# pass
# TODO: if birdirectional handle differently [Note?]: This task should not need bidirectional RNN
# Remember that initial states will be [(self.gru_hidden_layers * self.num_directions) x (X.shape[self.batch_index]) x (self.hidden_dimensions)]
#
# initial_states[-self.num_directions:, :, :] # output_gru[:,-1, :self.hidden_dimensions].view(1, -1, self.hidden_dimensions)
return output_gru[:, -1, :self.hidden_dimensions].view(
1, -1, self.hidden_dimensions) # initial_states[-1, :, :]
def forward(self, input: Tensor, mask: Tensor = None) -> Tensor:
"""
Args:
input: Tensor
the input tensor with shape = [batch, length, input_size]
mask: Tensor or None
the mask tensor with shape = [batch, length]
Returns: Tensor
the energy tensor with shape = [batch, length, num_label, num_label]
"""
batch, length, _ = input.size()
# compute out_s by tensor dot [batch, length, input_size] * [input_size, num_label]
# thus out_s should be [batch, length, num_label] --> [batch, length, 1, num_label]
out_s = self.state_nn(input)
if mask is not None:
out_s[:, :, self.index_eos] += (mask == 0).float() * 2e4
# [batch, length, num_label, num_label]
output = self.trans_matrix + out_s.unsqueeze(2)
return output
def biLSTM(all_corpus: list, model, d=None) -> (Tensor, object):
if d is None:
d = D('cpu')
else:
d = D('cuda')
MatrixRet = []
size = 0
totalLength = len(all_corpus)
cmtt = 0
for topic in all_corpus:
word_list_vec = []
for news in topic:
if size != 0:
break
for sentence in news:
if size != 0:
break
word_list = sentence.split(' ')
for word in word_list:
if word != '' and word != '\n' and word != ' ' and word != ' ':
vec = model[word]
word_list_vec.append(word)
size = len(vec)
break
# word_list_vec = FT(word_list_vec)
MatrixRet.append(word_list_vec)
cmtt = cmtt + 1
logging.info('\t * News: %d/%d' % (cmtt, totalLength))
ret1st = runModel(MatrixRet, size, size, d, model)
return (Tensor([0,0]), ret1st)
def forward(self, input: Tensor, mask: Tensor = None, hx: Tuple[Tensor, Tensor] = None) -> Tuple[Tensor, Tensor]:
batch_size = input.size(0) if self.batch_first else input.size(1)
if hx is None:
num_directions = 2 if self.bidirectional else 1
hx = input.new_zeros((self.num_layers * num_directions, batch_size, self.hidden_size))
hx = (hx, hx)
func = rnn_f.autograd_var_masked_rnn(num_layers=self.num_layers,
batch_first=self.batch_first,
bidirectional=self.bidirectional,
lstm=True)
self.reset_noise(batch_size)
output, hidden = func(input, self.all_cells, hx, None if mask is None else mask.view(mask.size() + (1,)))
return output, hidden
def forward(self, x: Tensor) -> Tensor:
x = self.conv(x)
x = x.view(-1, 512 * 2 * 2)
x = self.linear(x)
return x
def predict_recursively(preds: Tensor, energy: Tensor,
offset: int) -> NestedSequenceLabel:
length = preds.size(0)
nested_preds_list = []
index = 0
while index < length:
id = preds[index]
if id == eos_id:
break
if id != o_id:
if id == b_id: # B-XXX
start_tmp = index
index += 1
if index == length:
break
id = preds[index]
while id == i_id: # I-XXX
index += 1
if index == length:
break
id = preds[index]
if id == e_id: # E-XXX
end_tmp = index + 1
nested_preds = decode_nest(
energy[start_tmp:end_tmp, :, :])
nested_preds_list.append(
predict_recursively(
nested_preds,
energy[start_tmp:end_tmp, :, :],
start_tmp + offset))
index += 1
return NestedSequenceLabel(offset, length + offset, preds,
nested_preds_list)
def forward(self, logits: Tensor, mask: Tensor) -> Tensor:
"""Adds the loss functions, weighted by the prefactor."""
loss = logits.new_zeros(())
for loss_fn, prefact in self.loss_fns.items():
if prefact != 0:
loss += prefact * loss_fn(logits, mask)
return loss
def loss(self,
input: Tensor,
target: Tensor,
mask: Tensor = None) -> Tuple[Tensor, Tensor]:
"""
Args:
input: Tensor
the input tensor with shape = [batch, length, input_size]
target: Tensor
the tensor of target labels with shape [batch, length]
mask: Tensor or None
the mask tensor with shape = [batch, length]
Returns: Tensor
A 1D tensor for minus log likelihood loss
"""
batch, length, _ = input.size()
energy = self.forward(input, mask=mask)
# shape = [length, batch, num_label, num_label]
energy_transpose = energy.transpose(0, 1)
# shape = [length, batch]
target_transpose = target.transpose(0, 1)
# shape = [batch, num_label]
partition = None
# shape = [batch]
batch_index = torch.arange(0, batch).type_as(input).long()
prev_label = input.new_full((batch, ), self.index_bos).long()
tgt_energy = input.new_zeros(batch)
for t in range(length):
# shape = [batch, num_label, num_label]
curr_energy = energy_transpose[t]
if t == 0:
partition = curr_energy[:, self.index_bos, :]
else:
# shape = [batch, num_label]
partition = logsumexp(curr_energy + partition.unsqueeze(2),
dim=1)
label = target_transpose[t]
tgt_energy += curr_energy[batch_index, prev_label, label]
prev_label = label
return logsumexp(
self.trans_matrix.data[:, self.index_eos].unsqueeze(0) + partition,
dim=1) - tgt_energy, energy
def slice_last_dim(d: Tensor, length: int = 160) -> Tensor:
"""
Slice last dimention if length is too much.
If input is shorter than `length`, error is thrown.
[..., L>160] => [..., L==160]
"""
start = torch.randint(0, d.size()[-1] - (length - 1), (1, )).item()
return torch.narrow(d, -1, start, length)
def predict(self, inputs):
"""
Predits with the model based on the given input feature values
:param inputs: array
Input feature values
:return: (Tensor, Tensor, Tensor)
Indices for grapheme_root, vowel_diacritic, consonant_diacritic
"""
inputs = Tensor(inputs)
inputs = inputs.to(self._device)
grapheme_root_hat, vowel_diacritic_hat, consonant_diacritic_hat = self(inputs)
_, grapheme_root_indices = grapheme_root_hat.max(1)
_, vowel_diacritic_indices = vowel_diacritic_hat.max(1)
_, consonant_diacritic_indices = consonant_diacritic_hat.max(1)
return grapheme_root_indices, vowel_diacritic_indices, consonant_diacritic_indices
def read_image(image_path):
"""读出图片数据"""
img = transform(cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_RGB2BGR))
img = np.reshape(img, (3, 32, 32))
# debug: 输出图片
# img = to_pil_image(img)
# img.show()
return Tensor(img).unsqueeze(0)
def predict_array(self,
arr: Union[np.ndarray, Tuple[np.ndarray, np.ndarray]],
n_models: Optional[int] = None,
parent_bar: Optional[master_bar] = None,
display: bool = True,
callbacks: Optional[List[AbsCallback]] = None,
bs: Optional[int] = None) -> np.ndarray:
r'''
Apply ensemble to Numpy array and get predictions. If an output pipe has been added to the ensemble, then the predictions will be deprocessed.
Inputs are expected to be preprocessed; i.e. any input pipe added to the ensemble is not used.
Arguments:
arr: input data
n_models: number of models to use in predictions as ranked by the metric which was used when constructing the :class:`~lumin.nn.ensemble.ensemble.Ensemble`.
By default, entire ensemble is used.
parent_bar: not used when calling the method directly
display: whether to display a progress bar for model evaluations
callbacks: list of any callbacks to use during evaluation
bs: if not `None`, will run prediction in batches of specified size to save of memory
Returns:
Numpy array of predictions
Examples::
>>> preds = ensemble.predict_array(inputs)
'''
n_models = len(self.models) if n_models is None else n_models
models = self.models[:n_models]
weights = self.weights[:n_models]
weights = weights / weights.sum()
if isinstance(arr, tuple):
arr = (to_device(Tensor(arr[0])), to_device(Tensor(arr[1])))
pred = np.zeros((len(arr[0]), self.n_out))
else:
arr = to_device(Tensor(arr))
pred = np.zeros((len(arr), self.n_out))
for i, m in enumerate(
progress_bar(models, parent=parent_bar, display=display)):
tmp_pred = m.predict(arr, callbacks=callbacks, bs=bs)
if self.output_pipe is not None:
tmp_pred = self.output_pipe.inverse_transform(Xt=tmp_pred)
pred += weights[i] * tmp_pred
return pred
def eye(n, m=None, out=None, dtype=None, **kwargs):
if dtype is None:
dtype = torch.floatDefault
if m is None:
m = n
tensor = Tensor(n, m, dtype=dtype)
LibCall.torch.copyOut(tensor, out)
return tensor
def calc_loss(self, q_values: Tensor, target_q_values: Tensor,
actions: Tensor, rewards: Tensor,
done_mask: Tensor) -> Tensor:
"""
Calculate the MSE loss of this step.
The loss for an example is defined as:
Q_samp(s) = r if done
= r + gamma * max_a' Q_target(s', a')
loss = (Q_samp(s) - Q(s, a))^2
Args:
q_values: (torch tensor) shape = (batch_size, num_actions)
The Q-values that your current network estimates (i.e. Q(s, a') for all a')
target_q_values: (torch tensor) shape = (batch_size, num_actions)
The Target Q-values that your target network estimates (i.e. (i.e. Q_target(s', a') for all a')
actions: (torch tensor) shape = (batch_size,)
The actions that you actually took at each step (i.e. a)
rewards: (torch tensor) shape = (batch_size,)
The rewards that you actually got at each step (i.e. r)
done_mask: (torch tensor) shape = (batch_size,)
A boolean mask of examples where we reached the terminal state
Hint:
You may find the following functions useful
- torch.max
- torch.sum
- torch.nn.functional.one_hot
- torch.nn.functional.mse_loss
"""
# you may need this variable
num_actions = self.env.action_space.n
gamma = self.config.gamma
##############################################################
##################### YOUR CODE HERE - 3-5 lines #############
notdone = 1 - done_mask.to(torch.int64)
current_q = torch.max(
q_values *
torch.nn.functional.one_hot(actions.to(torch.int64), num_actions),
1).values # elementwise product to get reward for each batch
target_q = rewards + notdone * gamma * torch.max(target_q_values,
1).values
loss = torch.nn.functional.mse_loss(current_q, target_q)
return loss
