def forward(self, x):
y = F.dropout(F.relu(self.linears[0](x)), self.training)
for layer in self.linears[1:-1]:
y = F.relu(layer(y))
y = F.dropout(y, self.training)
y = F.log_softmax(self.linears[-1](y))
return y
def forward(self, x):
x = F.relu(self.fc1(x))
x = F.relu(self.fc_1(x))
x = F.relu(self.fc_2(x))
x = F.relu(self.fc_3(x))
y = self.fc2(x)
return y
def forward(self, v, u, d): """ @param v [batch_size, embedding_size] matrix to push @param u [batch_size,] vector of pop signals in (0, 1) @param d [batch_size,] vector of push signals in (0, 1) @return [batch_size, embedding_size] read matrix """ # update V, which is of size [t, bach_size, embedding_size] v = v.view(1, self.batch_size, self.embedding_size) self.V = torch.cat([self.V, v], 0) if len(self.V.data) != 0 else v # TODO initialize queue to fixed size # update s, which is of size [t, batch_size] old_t = self.s.size(0) if self.s.size() else 0 s = Variable(torch.FloatTensor(old_t + 1, self.batch_size)) w = u for i in xrange(old_t): s_ = F.relu(self.s[i,:] - w) w = F.relu(w - self.s[i,:]) s[i,:] = s_ # if len(torch.nonzero(w.data)) == 0: break # TODO does this if work properly now? s[old_t,:] = d self.s = s # calculate r, which is of size [batch_size, embedding_size] r = Variable(torch.zeros([self.batch_size, self.embedding_size])) for i in xrange(old_t + 1): used = torch.sum(self.s[:i,:], 0) if i > 0 else self.zero coeffs = torch.min(self.s[i,:], F.relu(1 - used)) # reformating coeffs into a matrix that can be multiplied element-wise r += coeffs.view(self.batch_size, 1).repeat(1, self.embedding_size) * self.V[i,:,:] return r
def forward(self, x):
x = F.relu(self.conv1(x))
x = self.bn_conv1(x)
x = self.pool(F.relu(self.conv2(x)))
x = self.bn_conv2(x)
x = F.relu(self.conv3(x))
x = self.bn_conv3(x)
x = self.pool(F.relu(self.conv4(x)))
x = self.bn_conv4(x)
'''
x = self.cpdfc1(x)
x = self.bn_2(x)
x = self.cpdfc2(x)
x = self.bn_2(x)
x = self.cpdfc3(x)
x = self.bn_2(x)
x = F.relu(self.cpdfc4(x))
'''
x = F.relu(self.conv2fc1(x))
x = self.bn_4(x)
x = self.conv2fc2(x)
x = self.bn_5(x)
x = x.view(-1, self.num_classes) # Flatten! <---
return x
def forward(self, x, y, y_mask):
"""Input shapes:
x = batch * len1 * h
y = batch * len2 * h
y_mask = batch * len2
Output shapes:
matched_seq = batch * len1 * h
"""
# Project vectors
if self.linear:
x_proj = self.linear(x.view(-1, x.size(2))).view(x.size())
x_proj = F.relu(x_proj)
y_proj = self.linear(y.view(-1, y.size(2))).view(y.size())
y_proj = F.relu(y_proj)
else:
x_proj = x
y_proj = y
# Compute scores
scores = x_proj.bmm(y_proj.transpose(2, 1))
# Mask padding
y_mask = y_mask.unsqueeze(1).expand(scores.size())
scores.data.masked_fill_(y_mask.data, -float('inf'))
# Normalize with softmax
alpha_flat = F.softmax(scores.view(-1, y.size(1)))
alpha = alpha_flat.view(-1, x.size(1), y.size(1))
# Take weighted average
matched_seq = alpha.bmm(y)
return matched_seq
def forward(self, doc_hiddens, question_hiddens):
'''
:param doc_hiddens: batch * input_len * in_channels
:param question_hiddens: batch * input_len * in_channels
:return: batch * output_size
'''
# Concatenate all available relations
num_doc_objects = doc_hiddens.size(1)
num_question_objects = question_hiddens.size(1)
num_total_objects = num_doc_objects * num_question_objects
doc = doc_hiddens.unsqueeze(1)
doc = doc.repeat(1, num_question_objects, 1, 1)
q = question_hiddens.unsqueeze(2).repeat(1, 1, num_doc_objects, 1)
relations = torch.cat([doc, q], 3)
#print('relations :', relations.size())
x_r = relations.view(-1,doc.size(3)+q.size(3))
#print('x_r:',x_r.size())
#res1 = x_r.clone()
#print(self.g_fc1)
x_r = F.relu((self.g_fc1(x_r)))# + res1
x_r = self.ln1.forward(x_r)
#res2 = x_r.clone()
#x_r = F.relu((self.g_fc2(x_r)))# + res2
#x_r = self.ln2.forward(x_r)
#res3 = x_r.clone()
#x_r = F.relu((self.g_fc3(x_r))) + res3
x_r = F.relu((self.g_fc4(x_r)))
x_r = self.ln4.forward(x_r)
x_g = x_r.view(relations.size(0), relations.size(1) * relations.size(2), -1)
x_g = x_g.sum(1).squeeze(1)/num_total_objects
return x_g
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2(x), 2))
x = x.view(-1, 320)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return x
def forward(self, x):
out = F.relu(self.bn1(x))
shortcut = self.shortcut(out) if hasattr(self, 'shortcut') else x
out = self.conv1(out)
out = self.conv2(F.relu(self.bn2(out)))
out += shortcut
return out
def forward(self, x):
x = self.fc1(x).view(-1, self.channels, self.rows, self.rows)
x = F.relu(self.batch_norm1(x))
x = F.relu(self.batch_norm2(self.conv1(x)))
x = F.relu(self.batch_norm3(self.conv2(x)))
x = F.relu(self.batch_norm4(self.conv3(x)))
return F.tanh(self.conv4(x))
def forward(self, x):
x = x.view(-1, 784) # Flatten the data (n, 1, 28, 28)-> (n, 784)
x = F.relu(self.l1(x))
x = F.relu(self.l2(x))
x = F.relu(self.l3(x))
x = F.relu(self.l4(x))
return self.l5(x)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = F.relu(self.bn2(self.conv2(out)))
out = self.bn3(self.conv3(out))
out += self.shortcut(x)
out = F.relu(out)
return out
def forward(self, x_left, x_right):
x_left = F.relu(x_left)
x_left = self.conv_left(x_left)
x_left = self.bn_left(x_left)
x_right = F.relu(x_right)
x_right = self.conv_right(x_right)
x_right = self.bn_right(x_right)
x_comb_iter_0_left = self.comb_iter_0_left(x_right)
x_comb_iter_0_right = self.comb_iter_0_right(x_left)
x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
x_comb_iter_1_left = self.comb_iter_1_left(x_left)
x_comb_iter_1_right = self.comb_iter_1_right(x_right)
x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
x_comb_iter_2_left = self.comb_iter_2_left(x_right)
x_comb_iter_2_right = self.comb_iter_2_right(x_left)
x_comb_iter_2 = x_comb_iter_2_left + x_comb_iter_2_right
x_comb_iter_3_left = self.comb_iter_3_left(x_comb_iter_0)
x_comb_iter_3 = x_comb_iter_3_left + x_comb_iter_1
x_comb_iter_4_left = self.comb_iter_4_left(x_comb_iter_0)
x_comb_iter_4_right = self.comb_iter_4_right(x_right)
x_comb_iter_4 = x_comb_iter_4_left + x_comb_iter_4_right
return torch.cat([x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], 1)
def adpW(self,x):
'''
calculate the pairwise_att of everypair of inputs
output_size: (x.size(0),x.size(1)/2)
'''
x = x.detach()
x = self.adp_metric_embedding1(x)
x = self.adp_metric_embedding1_bn(x)
x = F.relu(x)
x = self.adp_metric_embedding2(x)
x = self.adp_metric_embedding2_bn(x)
x = F.relu(x)
x = self.adp_metric_embedding3(x)
x = self.adp_metric_embedding3_bn(x)
x = F.relu(x)
pairwise_att = F.sigmoid(self.adp_metric_embedding4(x))
# x = self.adp_metric_embedding2_bn(x)
diag_matrix1 = []
diag_matrix2 = []
for i in range(x.size(0)):
diag_matrix1.append(torch.diag(pairwise_att[i, :x.size(1)/2]))
for i in range(x.size(0)):
diag_matrix2.append(torch.diag(pairwise_att[i, x.size(1)/2:]))
pairwise_att1 = torch.stack(diag_matrix1)
pairwise_att2 = torch.stack(diag_matrix1)
return pairwise_att1, pairwise_att2
def forward(self, x_left, x_right):
x_left = F.relu(x_left)
x_left_0 = self.pool_left_0(x_left)
x_left_0 = self.conv_left_0(x_left_0)
x_left_1 = self.pool_left_1(x_left) # TODO: strange padding + stride operation
x_left_1 = self.conv_left_1(x_left_1)
x_left = torch.cat([x_left_0, x_left_1], 1)
x_left = self.bn_left(x_left)
x_right = F.relu(x_right)
x_right = self.conv_right(x_right)
x_right = self.bn_right(x_right)
x_comb_iter_0_left = self.comb_iter_0_left(x_right)
x_comb_iter_0_right = self.comb_iter_0_right(x_left)
x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
x_comb_iter_1_left = self.comb_iter_1_left(x_left)
x_comb_iter_1_right = self.comb_iter_1_right(x_left)
x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
x_comb_iter_2_left = self.comb_iter_2_left(x_right)
x_comb_iter_2 = x_comb_iter_2_left + x_left
x_comb_iter_3_left = self.comb_iter_3_left(x_left) # TODO: those two avgPool look similar
x_comb_iter_3_right = self.comb_iter_3_right(x_left)
x_comb_iter_3 = x_comb_iter_3_left + x_comb_iter_3_right
x_comb_iter_4_left = self.comb_iter_4_left(x_right)
x_comb_iter_4 = x_comb_iter_4_left + x_right
return torch.cat([x_left, x_comb_iter_0, x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], 1)
def forward(self, x_left, x_right):
x_left = F.relu(x_left)
x_left = self.conv_left(x_left)
x_left = self.bn_left(x_left)
x_right = F.relu(x_right)
x_right = self.conv_right(x_right)
x_right = self.bn_right(x_right)
x_comb_iter_0_left = self.comb_iter_0_left(x_right)
x_comb_iter_0_right = self.comb_iter_0_right(x_left)
x_comb_iter_0 = x_comb_iter_0_left + x_comb_iter_0_right
x_comb_iter_1_left = self.comb_iter_1_left(x_left)
x_comb_iter_1_right = self.comb_iter_1_right(x_right)
x_comb_iter_1 = x_comb_iter_1_left + x_comb_iter_1_right
x_comb_iter_2_left = self.comb_iter_2_left(x_right)
x_comb_iter_2 = x_comb_iter_2_left + x_left
x_comb_iter_3_left = self.comb_iter_3_left(x_left) # TODO: those two avgPool look similar
x_comb_iter_3_right = self.comb_iter_3_right(x_left)
x_comb_iter_3 = x_comb_iter_3_left + x_comb_iter_3_right
x_comb_iter_4_left = self.comb_iter_4_left(x_right)
x_comb_iter_4 = x_comb_iter_4_left + x_right
return torch.cat([x_left, x_comb_iter_0, x_comb_iter_1, x_comb_iter_2, x_comb_iter_3, x_comb_iter_4], 1)
def forward(self, x):
# print("aaaaa")
x_no_static = self.embed_no_static(x)
# x_no_static = self.dropout(x_no_static)
x_static = self.embed_static(x)
# fix the embedding
# x_static = Variable(x_static.data)
# x_static = self.dropout(x_static)
x = torch.stack([x_static, x_no_static], 1)
# x = x.unsqueeze(1) # (N,Ci,W,D)
x = self.dropout(x)
if self.args.batch_normalizations is True:
x = [F.relu(self.convs1_bn(conv(x))).squeeze(3) for conv in self.convs1] #[(N,Co,W), ...]*len(Ks)
x = [F.max_pool1d(i, i.size(2)).squeeze(2) for i in x] #[(N,Co), ...]*len(Ks)
else:
x = [F.relu(conv(x)).squeeze(3) for conv in self.convs1] #[(N,Co,W), ...]*len(Ks)
x = [F.max_pool1d(i, i.size(2)).squeeze(2) for i in x] # [(N,Co), ...]*len(Ks)
x = torch.cat(x, 1)
'''
x1 = self.conv_and_pool(x,self.conv13) #(N,Co)
x2 = self.conv_and_pool(x,self.conv14) #(N,Co)
x3 = self.conv_and_pool(x,self.conv15) #(N,Co)
x = torch.cat((x1, x2, x3), 1) # (N,len(Ks)*Co)
'''
x = self.dropout(x) # (N,len(Ks)*Co)
if self.args.batch_normalizations is True:
x = self.fc1(x)
logit = self.fc2(F.relu(x))
else:
x = self.fc1(x)
logit = self.fc2(F.relu(x))
return logit
def forward(self, x):
x = F.relu(self.fc1(x))
x = F.relu(self.fc11(x))
x = F.relu(self.fc2(x))
x = F.relu(self.fc3(x))
x = F.tanh(self.out(x))
return x
def forward(self, sequence, graph): """ Apply self-attention to the sequence, ignores the graph """ sequence = sequence.squeeze(1) #get the dimension n, d = sequence.size() #project the sequence into key, value, and query sequences keySeq = f.relu(self.keyProj(sequence)) valueSeq = f.relu(self.valueProj(sequence)) querySeq = f.relu(self.queryProj(sequence)) #combine query with each key #a_ijh = softmax( (q_ih^T k_jh) / sqrt(d) ) #the result is, row i is the importance of the sequence for key i importance = f.softmax(t.matmul(querySeq, keySeq.permute(1,0)) * math.sqrt(d),0).permute(1,0) #apply the importance weights to the value sequence attention = t.matmul(valueSeq.permute(1,0), importance).permute(1,0) #sum the sequence for a complete representation final = t.sum(attention, 0) return attention.unsqueeze(1), final
def forward(self, x):
"""
In the forward function we accept a Variable of input data and we must return
a Variable of output data. We can use Modules defined in the constructor as
well as arbitrary operators on Variables.
"""
x = self.first_conv_layer(x)
x = self.second_conv_layer(x)
x = self.third_conv_layer(x)
x = self.fourth_conv_layer(x)
x = self.fifth_conv_layer(x)
x = self.last_conv_layer(x)
# auxiliary classifier branch (for 9 colors)
x = x.view(BATCH_SIZE, 4 * 4 * 10)
x_aux = F.relu(self.fc_aux(x))
class_out = nn.functional.softmax(x_aux)
#print 'class_out =', class_out
x_disc = F.relu(self.fc_disc(x))
disc_out = nn.functional.sigmoid(x_disc)
#print 'disc_out =', disc_out
return disc_out, class_out
def forward(self, x):
x = F.relu(self.linear1(x))
x = F.dropout(x, 0.8)
x = F.relu(self.linear2(x))
x = F.dropout(x, 0.8)
x = F.log_softmax(self.linear3(x))
return x
def forward(self, x):
x = x.float()
x = func.relu(self.fc1(x))
x = func.relu(self.fc2(x))
x = func.relu(self.fc3(x))
x = self.fc4(x).double()
return x
def forward(self, x):
out = F.relu(F.max_pool2d(self.conv1(x), 2))
out = F.relu(F.max_pool2d(self.conv2(out), 2))
out = out.view(-1, 320)
out = F.relu(self.fc1(out))
out = self.fc2(out)
return F.log_softmax(out, dim=1)
def forward(self, x):
x_no_static = self.embed_no_static(x)
# x_no_static = self.dropout(x_no_static)
x_static = self.embed_static(x)
# fix the embedding
x_static = Variable(x_static.data)
# x_static = self.dropout(x_static)
x = torch.stack([x_static, x_no_static], 1)
one_layer = x # (N,W,D) # torch.Size([64, 43, 300])
# print("one_layer {}".format(one_layer.size()))
# one_layer = self.dropout(one_layer)
# one_layer = one_layer.unsqueeze(1) # (N,Ci,W,D) # torch.Size([64, 1, 43, 300])
# one layer
one_layer = [torch.transpose(F.relu(conv(one_layer)).squeeze(3), 1, 2).unsqueeze(1) for conv in self.convs1] # torch.Size([64, 100, 36])
# one_layer = [F.relu(conv(one_layer)).squeeze(3).unsqueeze(1) for conv in self.convs1] # torch.Size([64, 100, 36])
# print(one_layer[0].size())
# print(one_layer[1].size())
# two layer
two_layer = [F.relu(conv(one_layer)).squeeze(3) for (conv, one_layer) in zip(self.convs2, one_layer)]
# print("two_layer {}".format(two_layer[0].size()))
# print("two_layer {}".format(two_layer[1].size()))
# pooling
output = [F.max_pool1d(i, i.size(2)).squeeze(2) for i in two_layer] # torch.Size([64, 100]) torch.Size([64, 100])
output = torch.cat(output, 1) # torch.Size([64, 300])
# dropout
output = self.dropout(output)
# linear
output = self.fc1(output)
logit = self.fc2(F.relu(output))
return logit
def forward(self, x):
x1_1 = F.relu(self.cpm1(x), inplace=True)
x1_2 = F.relu(self.cpm2(x), inplace=True)
x2_1 = F.relu(self.cpm3(x1_2), inplace=True)
x2_2 = F.relu(self.cpm4(x1_2), inplace=True)
x3_1 = F.relu(self.cpm5(x2_2), inplace=True)
return torch.cat([x1_1, x2_1, x3_1] , 1)
def forward(self, x):
out = F.relu(self.bn1(x))
shortcut = self.shortcut(out)
out = self.conv1(out)
out = self.conv2(F.relu(self.bn2(out)))
out += shortcut
return out
def forward(self, word_inputs, feature_inputs, word_seq_lengths, char_inputs, char_seq_lengths, char_seq_recover):
"""
input:
word_inputs: (batch_size, sent_len)
word_seq_lengths: list of batch_size, (batch_size,1)
char_inputs: (batch_size*sent_len, word_length)
char_seq_lengths: list of whole batch_size for char, (batch_size*sent_len, 1)
char_seq_recover: variable which records the char order information, used to recover char order
output:
Variable(batch_size, sent_len, hidden_dim)
"""
word_represent = self.wordrep(word_inputs,feature_inputs, word_seq_lengths, char_inputs, char_seq_lengths, char_seq_recover)
## word_embs (batch_size, seq_len, embed_size)
if self.word_feature_extractor == "CNN":
word_in = F.tanh(self.word2cnn(word_represent)).transpose(2,1).contiguous()
for idx in range(self.cnn_layer):
if idx == 0:
cnn_feature = F.relu(self.cnn_list[idx](word_in))
else:
cnn_feature = F.relu(self.cnn_list[idx](cnn_feature))
cnn_feature = self.cnn_drop_list[idx](cnn_feature)
cnn_feature = self.cnn_batchnorm_list[idx](cnn_feature)
feature_out = cnn_feature.transpose(2,1).contiguous()
else:
packed_words = pack_padded_sequence(word_represent, word_seq_lengths.cpu().numpy(), True)
hidden = None
lstm_out, hidden = self.lstm(packed_words, hidden)
lstm_out, _ = pad_packed_sequence(lstm_out)
## lstm_out (seq_len, seq_len, hidden_size)
feature_out = self.droplstm(lstm_out.transpose(1,0))
## feature_out (batch_size, seq_len, hidden_size)
outputs = self.hidden2tag(feature_out)
return outputs
def forward(self, x):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
before_pool = x
if self.pooling:
x = self.pool(x)
return x, before_pool
def forward(self, x):
# flatten input
if len(x.size()) > 2:
x = x.view(-1, int(np.prod(x.size()[1:])))
# corrupt input
x = self.input_corrupt(x)
corrupted = x
# encode
for layer in self.encode_layers:
x = layer(x)
x = F.relu(x)
# decode
if self.tied_weights:
for i, (layer, bias) in enumerate(self.decode_params):
x = F.linear(x, weight=layer.weight.t(), bias=bias)
if i == len(self.decode_params)-1:
x = self.visible_act(x)
else:
x = F.relu(x)
else:
for i, layer in enumerate(self.decode_layers):
x = layer(x)
if i == len(self.decode_layers)-1:
x = self.visible_act(x)
else:
x = F.relu(x)
return x, corrupted
def forward(self, x):
x = F.relu(self.bn1(self.conv1(x)), True)
x = F.max_pool2d(self.conv2(x), 2)
x = self.conv3(x)
x = self.conv4(x)
previous = x
outputs = []
for i in range(self.num_modules):
hg = self._modules['m' + str(i)](previous)
ll = hg
ll = self._modules['top_m_' + str(i)](ll)
ll = F.relu(self._modules['bn_end' + str(i)]
(self._modules['conv_last' + str(i)](ll)), True)
# Predict heatmaps
tmp_out = self._modules['l' + str(i)](ll)
outputs.append(tmp_out)
if i < self.num_modules - 1:
ll = self._modules['bl' + str(i)](ll)
tmp_out_ = self._modules['al' + str(i)](tmp_out)
previous = previous + ll + tmp_out_
return outputs
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = F.relu(self.bn2(self.conv2(out)))
out = self.bn3(self.conv3(out))
out += self.downsample(x)
out = F.relu(out)
return out
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out += self.shortcut(x)
out = F.relu(out)
return out
def forward(self, x):
x = x.view(-1, 84 * 84 * 3)
x = F.relu((self.fc1(x)))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return F.log_softmax(x)
lr=lr,
momentum=0.9,
nesterov=True)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=sch)
for epoch in range(epochs):
attack_model.train()
print("On Epoch, ", epoch)
for x1, x2, m1, m2 in train_dl:
# Forward pass
y_pred = attack_model(x1, x2, m1, m2)
#y_pred_above_0 = Functional.relu(y_pred)
y_6 = 6 - y_pred
y_less = Functional.relu(y_6)
yp = 6 - y_less
# Compute loss
loss = -1 * torch.sum(yp)
# Zero gradients, backward pass, update weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
print("running avg: ", (-1 * loss / bs))
# Keep weights below barrier
clip_params(attack_model, e)
# Save the trained model
def forward(self, obs: torch.Tensor) -> torch.Tensor:
x = F.relu(self.linear1(obs))
x = F.relu(self.linear2(x))
x = torch.tanh(self.linear3(x))
return x
def forward(self, x):
out = self.conv1(F.relu(self.bn1(x)))
out = F.avg_pool2d(out, 2)
return out
def forward(self, x):
out = self.conv1(F.relu(self.bn1(x)))
out = torch.cat((x, out), 1)
return out
def forward(self, sentence_1, sentence_2):
matrix_x, size = DynamicPool.cal_similar_matrix(sentence_1, sentence_2)
matrix_x.view(size, -1, 56, 56)
matrix_x = F.relu(self.conv1(matrix_x))
# 暂时用不到的动态pooling
# origin_size1 = len(matrix_x[0][0])
# origin_size2 = len(matrix_x[0][0][0])
# # 填充卷积输出
# # print(matrix_x)
# while origin_size1 < self.target_pool[0]:
# matrix_x = torch.cat([matrix_x, matrix_x[:, :, :origin_size1, :]], dim=2)
# if len(matrix_x[0][0]) >= self.target_pool[0]:
# break
#
# while origin_size2 < self.target_pool[1]:
# matrix_x = torch.cat([matrix_x, matrix_x[:, :, :, :origin_size2]], dim=3)
# if len(matrix_x[0][0][0]) >= self.target_pool[1]:
# break
#
# dynamic_pool_size1 = len(matrix_x[0][0])
# dynamic_pool_size2 = len(matrix_x[0][0][0])
# get_index = DynamicPool(self.target_pool[0], self.target_pool[1], dynamic_pool_size1, dynamic_pool_size2)
# index, pool_size = get_index.d_pool_index()
# m, n, high_judge, weight_judge = get_index.cal(index)
# stride = pool_size[0]
# stride1 = pool_size[1]
#
# matrix_x1 = matrix_x[:, :, :m, :n]
# matrix_x1 = F.max_pool2d(matrix_x1, (stride, stride1))
#
# if high_judge > 0:
# matrix_x2 = matrix_x[:, :, m:, :n]
# matrix_x2 = F.max_pool2d(matrix_x2, (stride + 1, stride1))
# if weight_judge > 0:
# matrix_x3 = matrix_x[:, :, :m, n:]
# matrix_x3 = F.max_pool2d(matrix_x3, (stride, stride1 + 1))
# if high_judge > 0 and weight_judge > 0:
# matrix_x4 = matrix_x[:, :, m:, n:]
# matrix_x4 = F.max_pool2d(matrix_x4, (stride + 1, stride1 + 1))
#
# if high_judge == 0 and weight_judge == 0:
# matrix_x = matrix_x1
# elif high_judge > 0 and weight_judge == 0:
# matrix_x = torch.cat([matrix_x1, matrix_x2], dim=2)
# elif high_judge == 0 and weight_judge > 0:
# matrix_x = torch.cat([matrix_x1, matrix_x3], dim=3)
# else:
# matrix_x_1 = torch.cat([matrix_x1, matrix_x2], dim=2)
# matrix_x_2 = torch.cat([matrix_x3, matrix_x4], dim=2)
# matrix_x = torch.cat([matrix_x_1, matrix_x_2], dim=3)
matrix_x = F.max_pool2d(matrix_x, (3, 3))
need_matrix = matrix_x[:, 0, :, :].view(size, 1, CONV_TARGET,
CONV_TARGET)
mat2_1 = self.conv2_1(need_matrix)
need_matrix = matrix_x[:, 1, :, :].view(size, 1, CONV_TARGET,
CONV_TARGET)
mat2_2 = self.conv2_1(need_matrix)
need_matrix = matrix_x[:, 2, :, :].view(size, 1, CONV_TARGET,
CONV_TARGET)
mat2_3 = self.conv2_1(need_matrix)
reshape_matrix = mat2_1 + mat2_2 + mat2_3
reshape_matrix = self.dropout(reshape_matrix)
reshape_matrix = F.max_pool2d(reshape_matrix, (2, 2))
reshape_matrix = reshape_matrix.view(
-1, self.num_flat_features(reshape_matrix))
reshape_matrix = F.tanh(self.fc1(reshape_matrix))
reshape_matrix = F.sigmoid(self.fc2(reshape_matrix))
# matrix_x = self.fc3(matrix_x)
reshape_matrix = reshape_matrix
return reshape_matrix
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
return F.relu(x, inplace=True)
def forward(self, x):
x = F.relu(self.conv(x), inplace=True)
x = F.relu(self.deconv(x), inplace=True)
return x
def forward(self, x):
x = self.convs(x) #convolutions maxpull pass
x = F.relu(
self.fc1(x)) #pass the convolution through fully connected linear
x = self.fc2(x) # bc this is our output layer. No activation here.
return F.softmax(x, dim=1)
def generate(self,
tensor_score,
tensor_metadata,
constraints_location,
temperature=1.):
self.eval()
tensor_score = to_cuda_variable(tensor_score)
num_voices, chorale_length, num_metadatas = tensor_metadata.size()
# generated chorale
gen_chorale = self.dataset.empty_score_tensor(chorale_length)
m = to_cuda_variable(tensor_metadata[None, :, :, :])
m = self.embed_metadata(m, tensor_score[None, :, :],
constraints_location=constraints_location[None, :, :])
output_constraints, activations_constraints = self.output_lstm_constraints(m)
hidden = init_hidden_lstm(
num_layers=self.num_layers,
batch_size=1,
lstm_hidden_size=self.num_lstm_generation_units
)
print(tensor_score)
# 1 bar of start symbols
# todo check
for tick_index in range(self.dataset.num_voices
* 4
* self.dataset.subdivision - 1):
voice_index = tick_index % self.dataset.num_voices
# notes
time_slice = gen_chorale[voice_index, 0]
time_slice = torch.from_numpy(np.array([time_slice]))[None, :]
note = self.note_embeddings[voice_index](
to_cuda_variable(time_slice)
)
time_slice = note
# concat with first metadata
# todo wrong
time_slice_cat = torch.cat(
(time_slice, output_constraints[:,
tick_index + 1: tick_index + 2, :]
), 2)
(output_gen, hidden), activations_generation = lstm_with_activations(
lstm_list=self.lstm_generation,
input=time_slice_cat,
hidden=hidden
)
h, c = hidden
hidden = h[:, 0, :, :], c[:, 0, :, :]
# generation:
for tick_index in range(-1, chorale_length * num_voices - 1):
voice_index = tick_index % num_voices
time_index = (tick_index - voice_index) // num_voices
next_voice_index = (tick_index + 1) % num_voices
next_time_index = (tick_index + 1 - next_voice_index) // num_voices
if tick_index == -1:
last_start_symbol = 0 # gen_chorale[-1, 0]
last_start_symbol = torch.from_numpy(np.array([last_start_symbol]))[None, :]
time_slice = self.note_embeddings[-1](
to_cuda_variable(last_start_symbol)
)
else:
time_slice = gen_chorale[voice_index, time_index]
time_slice = torch.from_numpy(np.array([time_slice]))[None, :]
note = self.note_embeddings[voice_index](
to_cuda_variable(time_slice)
)
time_slice = note
time_slice_cat = torch.cat(
(time_slice, output_constraints[:,
tick_index + 1: tick_index + 2, :]
), 2
)
(output_gen, hidden), activations_generation = lstm_with_activations(
lstm_list=self.lstm_generation,
input=time_slice_cat, hidden=hidden)
h, c = hidden
hidden = h[:, 0, :, :], c[:, 0, :, :]
weights = F.relu(self.linear_1(output_gen[:, 0, :]), inplace=True)
weights = self.linear_ouput_notes[next_voice_index](weights)
# compute predictions
# temperature
weights = weights * temperature
preds = F.softmax(weights)
# first batch element
preds = to_numpy(preds[0])
new_pitch_index = np.random.choice(np.arange(
self.num_notes_per_voice[next_voice_index]
), p=preds)
# new_pitch_index = np.argmax(preds)
gen_chorale[next_voice_index, next_time_index] = int(new_pitch_index)
# # show original chorale
#score_original = self.dataset.tensor_to_score(
# tensor_score.cpu())
#score_original.show()
print(gen_chorale)
score = self.dataset.tensor_to_score(gen_chorale)
return score, gen_chorale, tensor_metadata
def stage2_pre(self, image):
x = self.pool1_stage2(F.relu(self.conv1_stage2(image)))
x = self.pool2_stage2(F.relu(self.conv2_stage2(x)))
x = self.pool3_stage2(F.relu(self.conv3_stage2(x)))
return x
def forward(self, state, action):
"""Build a critic (value) network that maps (state, action) pairs -> Q-values."""
xs = F.relu(self.fcs1(state))
x = torch.cat((xs, action), dim=1)
x = F.relu(self.fc2(x))
return self.fc3(x)
def forward(self, image):
"""
Forward propagation.
:param image: images, a tensor of dimensions (N, 3, 300, 300)
:return: lower-level feature maps conv4_3 and conv7
"""
out = F.relu(self.conv1_1(image)) # (N, 64, 300, 300)
out = F.relu(self.conv1_2(out)) # (N, 64, 300, 300)
out = self.pool1(out) # (N, 64, 150, 150)
out = F.relu(self.conv2_1(out)) # (N, 128, 150, 150)
out = F.relu(self.conv2_2(out)) # (N, 128, 150, 150)
out = self.pool2(out) # (N, 128, 75, 75)
out = F.relu(self.conv3_1(out)) # (N, 256, 75, 75)
out = F.relu(self.conv3_2(out)) # (N, 256, 75, 75)
out = F.relu(self.conv3_3(out)) # (N, 256, 75, 75)
out = self.pool3(out) # (N, 256, 38, 38), it would have been 37 if not for ceil_mode = True
out = F.relu(self.conv4_1(out)) # (N, 512, 38, 38)
out = F.relu(self.conv4_2(out)) # (N, 512, 38, 38)
out = F.relu(self.conv4_3(out)) # (N, 512, 38, 38)
conv4_3_feats = out
# g1 = out # (N, 512, 38, 38)
out = self.pool4(out) # (N, 512, 19, 19)
out = F.relu(self.conv5_1(out)) # (N, 512, 19, 19)
out = F.relu(self.conv5_2(out)) # (N, 512, 19, 19)
out = F.relu(self.conv5_3(out)) # (N, 512, 19, 19)
out = self.pool5(out) # (N, 512, 19, 19), pool5 does not reduce dimensions
out = F.relu(self.conv6(out)) # (N, 1024, 19, 19)
conv7_feats = F.relu(self.conv7(out))
# g2 = F.relu(self.conv7(out)) # (N, 1024, 19, 19)
# Lower-level feature maps
return conv4_3_feats, conv7_feats
def forward(self, x): return self.w_2(self.dropout(F.relu(self.w_1(x))))
def forward_inpaint(self, score_tensor: Variable, metadata_tensor: Variable, constraints_loc, start_tick, end_tick):
batch_size, num_voices, chorale_length = score_tensor.size()
sequence_length = num_voices * chorale_length
gen_chorale = self.dataset.empty_score_tensor(chorale_length).unsqueeze(0)
gen_chorale = gen_chorale.repeat(batch_size, 1, 1)
gen_chorale[:, :, :start_tick] = score_tensor[:, :, :start_tick]
gen_chorale[:, :, end_tick:] = score_tensor[:, :, end_tick:]
# === embed as wrapped sequence ===
# --- chorale
x = self.embed_tensor_score(score_tensor)
# --- metadata
m = self.embed_metadata(
metadata_tensor,
score_tensor,
constraints_location=constraints_loc
)
# === LSTM on constraints ===
output_constraints, activations_constraint = self.output_lstm_constraints(m)
hidden = init_hidden_lstm(
num_layers=self.num_layers,
batch_size=batch_size,
lstm_hidden_size=self.num_lstm_generation_units
)
offset_seq = torch.cat(
[to_cuda_variable(torch.zeros(batch_size, 1, self.note_embedding_dim)),
x[:, :sequence_length - 1, :]
], 1)
# todo dropout only on offset_seq?
offset_seq = self.drop_input(offset_seq)
input_seq = torch.cat([offset_seq, output_constraints], 2)
past_input = input_seq[:, :start_tick, :]
(_, hidden), activations_gen = lstm_with_activations(
lstm_list=self.lstm_generation,
input=past_input,
hidden=hidden)
h, c = hidden
hidden = h[:, 0, :, :], c[:, 0, :, :]
final_weights = [[] for i in range(num_voices)]
for tick_index in range(start_tick-1, end_tick-1):
voice_index = tick_index % num_voices
time_index = (tick_index - voice_index) // num_voices
next_voice_index = (tick_index + 1) % num_voices
next_time_index = (tick_index + 1 - next_voice_index) // num_voices
if tick_index == -1:
last_start_symbol = 0 # gen_chorale[-1, 0]
last_start_symbol = torch.from_numpy(np.array([last_start_symbol]))[None, :]
time_slice = self.note_embeddings[-1](
to_cuda_variable(last_start_symbol)
)
else:
time_slice = gen_chorale[:, voice_index:voice_index+1, time_index]
# time_slice = torch.from_numpy(np.array([time_slice]))[None, :]
note = self.note_embeddings[voice_index](
to_cuda_variable(time_slice)
)
time_slice = note
time_slice_cat = torch.cat((time_slice, output_constraints[:, tick_index+1:tick_index+2, :]), 2)
(output_gen, hidden), activations_generation = lstm_with_activations(
lstm_list=self.lstm_generation,
input=time_slice_cat, hidden=hidden)
h, c = hidden
hidden = h[:, 0, :, :], c[:, 0, :, :]
weights = F.relu(self.linear_1(output_gen[:, 0, :]), inplace=True)
weights = self.linear_ouput_notes[next_voice_index](weights)
# compute predictions
# temperature
weights = weights
preds = F.softmax(weights)
final_weights[voice_index].append(weights.unsqueeze(1))
# first batch element
preds = to_numpy(preds[0])
new_pitch_index = np.argmax(preds)
gen_chorale[:, next_voice_index, next_time_index] = int(new_pitch_index)
for i in range(num_voices):
final_weights[i] = torch.cat(final_weights[i], 1)
return final_weights, gen_chorale
def forward(self, x):
x = x.view(-1, 28 * 28)
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
x = self.fc3(x)
return x
def forward(self, state):
"""Build an actor (policy) network that maps states -> actions."""
x = F.relu(self.fc1(state))
x = F.relu(self.fc2(x))
return torch.tanh(self.fc3(x))
def forward(self, c5):
p6 = self.p6(c5)
p7 = self.p7(F.relu(p6))
return [p6, p7]
def forward(self, X, A_hat): ### 1-layer GCN architecture
X = torch.mm(X, self.weight)
if self.bias is not None:
X = (X + self.bias)
X = F.relu(torch.mm(A_hat, X))
return X
def encode(self, x: Variable) -> (Variable, Variable):
h1 = F.relu(self.fc1(x)) # h1 is 400-dimensional vector
return self.fc21(h1), self.fc22(h1) # this returns two 20-dimensional vectors
def forward(self, x):
x = F.relu(self.conv1_bn(self.conv1(x)))
x = F.relu(self.conv2_bn(self.conv2(x)))
x = F.max_pool2d(x, (2, 2))
x = self.do1(x)
x = F.relu(self.conv3_bn(self.conv3(x)))
x = F.relu(self.conv4_bn(self.conv4(x)))
x = F.max_pool2d(x, (2, 2))
x = self.do2(x)
x = F.relu(self.conv5_bn(self.conv5(x)))
x = self.do3(x)
x = F.relu(self.conv6_bn(self.conv6(x)))
x = self.do4(x)
x = F.relu(self.conv7_bn(self.conv7(x)))
x = self.do5(x)
x = F.relu(self.conv8_bn(self.conv8(x)))
x = F.max_pool2d(x, (2, 2))
x = self.do6(x)
x = F.relu(self.conv9_bn(self.conv9(x)))
x = F.relu(self.conv10_bn(self.conv10(x)))
x = F.max_pool2d(x, (1, 2))
x = self.do7(x)
x = F.relu(self.conv11_bn(self.conv11(x)))
x = F.relu(self.conv12_bn(self.conv12(x)))
x = F.max_pool2d(x, (1, 2))
x = self.do8(x)
x = F.relu(self.conv13_bn(self.conv13(x)))
x = self.do9(x)
x = F.relu(self.conv14_bn(self.conv14(x)))
x = self.do10(x)
x = self.conv15_bn(self.conv15(x))
size_to_pool = x.size()
x = F.avg_pool2d(x, (size_to_pool[2], size_to_pool[3]))
x = x.view(-1, 41)
return x
def encode(self, x):
h = F.relu(self.fc1(x))
mu = self.fc_encoder_mu(h)
logvar = self.fc_encoder_logvar(h)
return mu, logvar
def decode(self, z: Variable) -> Variable:
h3 = F.relu(self.fc3(z)) # z can be used to generate new samples
return torch.sigmoid(self.fc4(h3)) # remember that z ~ N(mu, var)
def G(z):
h = nn.relu(z @ Wzh + bzh.repeat(z.size(0), 1))
X = nn.sigmoid(h @ Whx + bhx.repeat(h.size(0), 1))
return X
def forward(self, z):
h = F.relu(self.fc_decoder(z))
x_reconst = self.softmax(self.fc2(h))
return x_reconst
def forward(self, x):
x = F.relu(self.fc1(x))
x = F.relu(self.fc2(x))
return self.fc3(x)
def D(X):
h = nn.relu(X @ Wxh + bxh.repeat(X.size(0), 1))
y = nn.sigmoid(h @ Why + bhy.repeat(h.size(0), 1))
return y
def forward(self, x):
x = F.relu(self.bn1(self.lin1(x)))
x = F.relu(self.bn2(self.lin2(x)))
x = F.relu(self.bn3(self.lin3(x)))
return torch.sigmoid(self.head(x))
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = F.relu(self.bn2(self.conv2(out)))
out = self.bn3(self.conv3(out))
out = out + self.shortcut(x) if self.stride==1 else out
return out
