Still get inplace operation error when I use clone() - pytorch

I tried to add .clone() at anywhere I think it's necessary.
But I still got inplace operation error.
And this error happened in second epoch, rather than first epoch when the beginning I got this error.
class Model(torch.nn.Module):
def __init__(self, user_M, item_M, feature_M, embedding_M):
super(Model, self).__init__()
self.dimension = embedding_M + feature_M + 1
# Memories
self.user_memory = torch.randn(user_M, 2, 1, self.dimension)
self.item_memory = torch.randn(item_M, 2, 1, self.dimension)
# Writer
self.user_LSTM = nn.LSTMCell(self.dimension, self.dimension)
self.item_LSTM = nn.LSTMCell(self.dimension, self.dimension)
def score(self, user_id, item_id):
user_em = torch.index_select(self.user_memory[:, 1, :].clone(), 0, user_id[:,0].clone())
item_em = torch.index_select(self.item_memory[:, 1, :].clone(), 0, item_id[:,0].clone())
return torch.sigmoid(torch.bmm(user_em, item_em.permute(0,2,1)))
def forward(self, train_data, user_id, item_id):
ui_train = train_data[:, 0:1].type(torch.LongTensor).clone()
ii_train = train_data[:, 1:2].type(torch.LongTensor).clone()
for row in range(len(train_data)):
memory = self.user_memory[ui_train[row][0]].clone()
h, c = self.user_LSTM(train_data[row:row+1].clone(), (memory[0], memory[1]))
self.user_memory[ui_train[row]] = torch.stack((h, c))
for row in range(len(train_data)):
memory = self.item_memory[ii_train[row][0]].clone()
h, c = self.item_LSTM(train_data[row:row+1].clone(), (memory[0], memory[1]))
self.item_memory[ii_train[row]] = torch.stack((h, c))
# Scoring
y = self.score(user_id, item_id)
return y.reshape(-1, 1)
model = Model(user_M, item_M, feature_M, 3)
criterion = torch.nn.BCELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=LEARNING_RATE)
for epoch in range(10):
print(epoch)
output = model(train_data, user_id, item_id)
loss = criterion(output, label)
optimizer.zero_grad()
loss.backward(retain_graph=True)
optimizer.step()
`RuntimeError`: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [176, 704]], which is output 0 of TBackward, is at version 2; expected version 1 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

Related

multi-agent DQN learn single model for all agents

I'm trying to run a DQN for a multi-agent system, so there is one DNN for each agent.
It takes input=state [batch, state size, #time steps, #nodes], while for simplicity we assume #time steps=1. #nodes is number of agents. And output=Q-values for each agent.
The problem is that I test various stuff with this network, but it return not so consistent results. I suspect it has to do with me running separately DQN for each agent, but learning it via the same model. I sum the losses for all agents into one loss, and then it divide by their amount.
I'm not sure it is correct. I'd be grateful for any help.
Here's my code:
class DQN(nn.Module):
def __init__(self, args): #node_size, inputs, outputs, layers=[128, 64, 16]):
# state_size, n_actions = inputs, outputs
super(DQN, self).__init__()
self.model_type = args.model_type
if args.model_type == "seperate_state_DNN":
out_size = args.num_of_actions
self.shared_model = nn.Sequential()
h_sizes = [args.input_state_size] + args.layers
for k in range(len(h_sizes) - 1):
self.shared_model.add_module('k1'+str(k), nn.Linear(h_sizes[k], h_sizes[k + 1]))
self.shared_model.add_module('k2'+str(k), args.activations[args.layers_nl[k]])
self.shared_model.add_module('final', nn.Linear(h_sizes[-1], out_size))
def forward(self, input, i=None):
# input state dimension: [batch, state size, #time steps, #nodes]
if self.model_type == "seperate_state_DNN":
if i is None:
final_output = torch.zeros_like(input)
else:
final_output = self.shared_model(input) # [:, :, :, i].unsqueeze(3))
return final_output
And here is the calling function:
def select_action(self, state, edge_state):
#self.policy_net.eval()
sample = random.random()
if self.configuration == 2:
self.eps_threshold = 0.0 # no exploration at all, only optimal values!
else:
self.eps_threshold = self.decay_functionn()
self.steps_done += 1
if sample > self.eps_threshold:
self.last_exploration = False
with torch.no_grad():
# t.max(1) will return largest column value of each row.
# second column on max result is index of where max element was
# found, so we pick action with the larger expected reward.
state = state.to(self.device)# torch.from_numpy(state).float().to(self.device) # Convert to tensor.
state = state.unsqueeze(0) # Add batch dimension (also to action below): [batch=1, #time steps, #nodes, state size]
final_output = []
x1 = self.policy_net(state, None)#.detach()
for i in range(self.node_size):
final_output.append(self.policy_net(x1[:, :, -1, i]+state[:, :, -1, i], i).max(1)[1].detach().cpu().view(state.shape[0], -1))
# .to(self.device) # action dimension: [batch=1, #nodes]
return torch.cat(final_output, dim=1)
else:
self.last_exploration = True
return torch.randint(0, self.n_actions, (1, self.node_size))
And this is the main RL training loop:
for epi in range(self.episodes):
print("### Starting Episode: ", epi, ' ### in index=', self.run_index)
state = env.reset(self, heatup=self.sim_heatup) # single step state
done = False
while not done:
action = agent.select_action(state) # .to(device)
next_state1, reward, done = env.do_step(action)
agent.add_to_memory(state, action, next_state, reward)
agent.optimize_model()
state = next_state
agent.curr_episode += 1
# Plot and dump statistics and learning curves.
agent.dump_data_on_episode_end(plot=True)
env.capture_episode()
env.close()
Finally, this is the optimization, executed in "agent.optimize_model()" above, including the functions it uses:
def optimize_model(self):
if len(self.memory) < self.batch_size:
return
transitions = self.memory.sample(self.batch_size)
# This converts batch-array of Transitions
# to Transition of batch-arrays.
batch = Transition(*zip(*transitions))
next_states_batch = torch.stack(batch.next_state).to(self.device)
state_batch = torch.stack(batch.state).to(self.device)
action_batch = torch.cat(batch.action).view(self.batch_size, -1).to(self.device) #torch.stack(batch.action, dim=0).to(self.device)
reward_batch = torch.cat(batch.reward).view(self.batch_size, -1).to(self.device)
# dims: states=[batch, steps, nodes, state size]; action=[batch, nodes]; reward=[batch, nodes]
loss = torch.tensor(0., device=self.device)
self.policy_net.train() # IM NOT SURE IF IT SHOULD BE HERE...
x1 = self.policy_net(state_batch, None)
x2 = self.policy_net(next_states_batch, None)
for i in range(self.node_size):
action_batch1 = action_batch[:,i].unsqueeze(1).reshape(-1, 1) # action=[batchXnodes, 1]
reward_batch1 = reward_batch[:,i].unsqueeze(1).view(-1, 1) # reward=[batchXnodes, 1]
# Compute loss
loss += self._compute_loss(i, x1[:, :, -1, i]+state_batch[:, :, -1, i], edge_state_batch, action_batch1,
x2[:, :, -1, i]+next_states_batch[:, :, -1, i], next_edge_state_batch, reward_batch1)
# Optimize the model
loss.div_(self.node_size)
self.optimizer.zero_grad()
loss.backward()
# clip grad
if self.grad_clip is not None:
for param in self.policy_net.parameters():
param.grad.data.clamp_(-self.grad_clip, self.grad_clip)
# update Policy net weights
self.optimizer.step()
#del loss
self.losses.append(loss.detach().cpu().numpy())
# update Target net weights
self._update_target()
def _compute_loss(self, i, state_batch, edge_state_batch, action_batch, next_states_batch, next_edge_state_batch, reward_batch):
# Q{policy net}(s, a): [batchXnodes, actions] ---gather---> [batchXnodes, 1=q_values according to this policy]
state_action_q_values = self.policy_net(state_batch, i).gather(1, action_batch)
# argmax{a} Q{policy net}(s', a'): [batchXnodes, actions] ---argmax---> [batchXnodes] ---unsqueeze---> [batchXnodes, 1]
next_state_actions = torch.argmax(self.policy_net(next_states_batch, i), dim=1).unsqueeze(1)
# Q{ploicy net}(s', argmax{a} Q{target net}(s', a') ): [batchXnodes, actions] --gather--> [batchXnodes, 1=q_values according to this policy]
next_state_q_values = self.target_net(next_states_batch, i).gather(1, next_state_actions)
# Q* = Disount * Q(s', argmax(..)) + R: [batchXnodes, 1]
expected_state_action_values = (next_state_q_values.detach() * self.discount) + reward_batch
loss = F.smooth_l1_loss(state_action_q_values, expected_state_action_values)
return loss
def _update_target(self):
if self.target_net is None:
# There is nothing to update.
return
# Update the target network, copying all weights and biases in DQN
if self.target_update > 1:
# Hard copy of weights.
if self.steps_done % self.target_update == 0:
self.target_net.load_state_dict(self.policy_net.state_dict())
return
elif self.target_update < 1 and self.target_update > 0:
# polyak averaging:
tau = self.target_update
for target_param, param in zip(self.target_net.parameters(), self.policy_net.parameters()):
target_param.data.copy_(tau * param + (1 - tau) * target_param)
return
else:
raise NotImplementedError
Sorry for the large question, I just wanted to supply all the necessary information.
If more information is needed I'd be happy to give it.
Any suggestion is much appreciated.
Thanks,
Shimon

PyTorch: GRU, one-to-many / many-to-one

I would like to implement a GRU able to encode a sequence of vectors to one vector (many-to-one), and then another GRU able to decode a vector to a sequence of vector (one-to-many). The size of the vectors wouldn't be changed. I would like to have an opinion about what I implemented.
Here is the code:
class AEGRU(nn.Module):
def __init__(self, opt):
super(AEGRU, self).__init__()
self.length = 256
self.latent_space = 256
self.num_layers = 1
self.GRU_enc = nn.GRU(input_size=3, hidden_size=self.latent_space, num_layers=self.num_layers, batch_first=True)
self.fc_enc = nn.Linear(self.latent_space, self.latent_space)
self.GRU_dec = nn.GRU(input_size=self.latent_space, hidden_size=3, num_layers=self.num_layers, batch_first=True)
self.fc_dec = nn.Linear(3, 3)
def enc(self, x):
# x has shape: Batch_size x self.length x 3
h0 = torch.zeros(self.num_layers, x.shape[0], self.latent_space).cuda()
out, _ = self.GRU_enc(x, h0)
out = out[:, -1, :]
out = self.fc_enc(out)
return out
def dec(self, x):
# x has shape: Batch_size x self.latent_space
x = x[:, None, :]
h = torch.zeros(self.num_layers, x.shape[0], 3).cuda()
# method 1 ??
'''outputs = torch.zeros(x.shape[0], self.length, 3).cuda()
for i in range(self.length):
out, h = self.GRU_dec(x, h)
outputs[:, i, :] = out[:, 0, :]'''
# method 2 ??
x = x.repeat(1, self.length, 1)
outputs, _ = self.GRU_dec(x, h)
# linear layer
outputs = self.fc_dec(outputs)
return outputs
def forward(self, x):
self.indices = []
latent = self.enc(x)
output = self.dec(latent)
return output
I am not sure whether this is the good way to do a one-to-many GRU. Could I have some opinions about this?
Thanks for reading!

GPU memory increasing at each batch (PyTorch)

I am trying to build a convolutionnal network using ConvLSTM layer (LSTM cell but with convolutions instead of matrix multiplications), but the problem is that my GPU memory increases at each batch, even if I'm deleting variables, and getting the true value for the loss (and not the graph) for each iteration. I may be doing something wrong but that exact same script ran without issues with another model (with more parameters and also using ConvLSTM layer).
Each batch is composed of num_batch x 3 images (grayscale) and I'm trying to predict the difference |Im(t+1)-Im(t)| with the input Im(t)
def main():
config = Config()
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=config.batch_size, num_workers=0, shuffle=True, drop_last=True)
nb_img = len(train_dataset)
util.clear_progress_dir()
step_tensorboard = 0
###################################
# Model Setup #
###################################
model = fully_convLSTM()
if torch.cuda.is_available():
model = model.float().cuda()
lr = 0.001
optimizer = torch.optim.Adam(model.parameters(),lr=lr)
util.enumerate_params([model])
###################################
# Training Loop #
###################################
model.train() #Put model in training mode
train_loss_recon = []
train_loss_recon2 = []
for epoch in tqdm(range(config.num_epochs)):
running_loss1 = 0.0
running_loss2 = 0.0
for i, (inputs, outputs) in enumerate(train_dataloader, 0):
print(i)
torch.cuda.empty_cache()
gc.collect()
# if torch.cuda.is_available():
inputs = autograd.Variable(inputs.float()).cuda()
outputs = autograd.Variable(outputs.float()).cuda()
im1 = inputs[:,0,:,:,:]
im2 = inputs[:,1,:,:,:]
im3 = inputs[:,2,:,:,:]
diff1 = torch.abs(im2 - im1).cuda().float()
diff2 = torch.abs(im3 - im2).cuda().float()
model.initialize_hidden()
optimizer.zero_grad()
pred1 = model.forward(im1)
loss = reconstruction_loss(diff1, pred1)
loss.backward()
# optimizer.step()
model.update_hidden()
optimizer.zero_grad()
pred2 = model.forward(im2)
loss2 = reconstruction_loss(diff2, pred2)
loss2.backward()
optimizer.step()
model.update_hidden()
## print statistics
running_loss1 += loss.detach().data
running_loss2 += loss2.detach().data
if i==0:
with torch.no_grad():
img_grid_diff_true = (diff2).cpu()
img_grid_diff_pred = (pred2).cpu()
f, axes = plt.subplots(2, 4, figsize=(48,48))
for l in range(4):
axes[0, l].imshow(img_grid_diff_true[l].squeeze(0).squeeze(0), cmap='gray')
axes[1, l].imshow(img_grid_diff_pred[l].squeeze(0).squeeze(0), cmap='gray')
plt.show()
plt.close()
writer_recon_loss.add_scalar('Reconstruction loss', running_loss1, step_tensorboard)
writer_recon_loss2.add_scalar('Reconstruction loss2', running_loss2, step_tensorboard)
step_tensorboard += 1
del pred1
del pred2
del im1
del im2
del im3
del diff1
del diff2#, im1_noised, im2_noised
del inputs
del outputs
del loss
del loss2
for obj in gc.get_objects():
if torch.is_tensor(obj) :
del obj
torch.cuda.empty_cache()
gc.collect()
epoch_loss = running_loss1 / len(train_dataloader.dataset)
epoch_loss2 = running_loss2/ len(train_dataloader.dataset)
print(f"Epoch {epoch} loss reconstruction1: {epoch_loss:.6f}")
print(f"Epoch {epoch} loss reconstruction2: {epoch_loss2:.6f}")
train_loss_recon.append(epoch_loss)
train_loss_recon2.append(epoch_loss2)
del running_loss1, running_loss2, epoch_loss, epoch_loss2
Here is the model used :
class ConvLSTMCell(nn.Module):
def __init__(self, input_channels, hidden_channels, kernel_size):
super(ConvLSTMCell, self).__init__()
# assert hidden_channels % 2 == 0
self.input_channels = input_channels
self.hidden_channels = hidden_channels
self.kernel_size = kernel_size
# self.num_features = 4
self.padding = 1
self.Wxi = nn.Conv2d(self.input_channels, self.hidden_channels, self.kernel_size, 1, self.padding, bias=True)
self.Whi = nn.Conv2d(self.hidden_channels, self.hidden_channels, self.kernel_size, 1, self.padding, bias=False)
self.Wxf = nn.Conv2d(self.input_channels, self.hidden_channels, self.kernel_size, 1, self.padding, bias=True)
self.Whf = nn.Conv2d(self.hidden_channels, self.hidden_channels, self.kernel_size, 1, self.padding, bias=False)
self.Wxc = nn.Conv2d(self.input_channels, self.hidden_channels, self.kernel_size, 1, self.padding, bias=True)
self.Whc = nn.Conv2d(self.hidden_channels, self.hidden_channels, self.kernel_size, 1, self.padding, bias=False)
self.Wxo = nn.Conv2d(self.input_channels, self.hidden_channels, self.kernel_size, 1, self.padding, bias=True)
self.Who = nn.Conv2d(self.hidden_channels, self.hidden_channels, self.kernel_size, 1, self.padding, bias=False)
self.Wci = None
self.Wcf = None
self.Wco = None
def forward(self, x, h, c): ## Equation (3) dans Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting
ci = torch.sigmoid(self.Wxi(x) + self.Whi(h) + c * self.Wci)
cf = torch.sigmoid(self.Wxf(x) + self.Whf(h) + c * self.Wcf)
cc = cf * c + ci * torch.tanh(self.Wxc(x) + self.Whc(h)) ###gt= tanh(cc)
co = torch.sigmoid(self.Wxo(x) + self.Who(h) + cc * self.Wco) ##channel out = hidden channel
ch = co * torch.tanh(cc)
return ch, cc #short memory, long memory
def init_hidden(self, batch_size, hidden, shape):
if self.Wci is None:
self.Wci = nn.Parameter(torch.zeros(1, hidden, shape[0], shape[1])).cuda()
self.Wcf = nn.Parameter(torch.zeros(1, hidden, shape[0], shape[1])).cuda()
self.Wco = nn.Parameter(torch.zeros(1, hidden, shape[0], shape[1])).cuda()
else:
assert shape[0] == self.Wci.size()[2], 'Input Height Mismatched!'
assert shape[1] == self.Wci.size()[3], 'Input Width Mismatched!'
return (autograd.Variable(torch.zeros(batch_size, hidden, shape[0], shape[1])).cuda(),
autograd.Variable(torch.zeros(batch_size, hidden, shape[0], shape[1])).cuda())
class fully_convLSTM(nn.Module):
def __init__(self):
super(fully_convLSTM, self).__init__()
layers = []
self.hidden_list = [1,32,32,1]#,32,64,32,
for k in range(len(self.hidden_list)-1): # Define blocks of [ConvLSTM,BatchNorm,Relu]
name_conv = "self.convLSTM" +str(k)
cell_conv = ConvLSTMCell(self.hidden_list[k],self.hidden_list[k+1],3)
setattr(self, name_conv, cell_conv)
name_batchnorm = "self.batchnorm"+str(k)
batchnorm=nn.BatchNorm2d(self.hidden_list[k+1])
setattr(self, name_batchnorm, batchnorm)
name_relu =" self.relu"+str(k)
relu=nn.ReLU()
setattr(self, name_relu, relu)
self.sigmoid = nn.Sigmoid()
self.internal_state=[]
def initialize_hidden(self):
for k in range(len(self.hidden_list)-1):
name_conv = "self.convLSTM" +str(k)
(h,c) = getattr(self,name_conv).init_hidden(config.batch_size, self.hidden_list[k+1],(256,256))
self.internal_state.append((h,c))
self.internal_state_new=[]
def update_hidden(self):
for i, hidden in enumerate(self.internal_state_new):
self.internal_state[i] = (hidden[0].detach(), hidden[1].detach())
self.internal_state_new = []
def forward(self, input):
x = input
for k in range(len(self.hidden_list)-1):
name_conv = "self.convLSTM" +str(k)
name_batchnorm = "self.batchnorm"+str(k)
name_relu =" self.relu"+str(k)
x, c = getattr(self,name_conv)(x, self.internal_state[k][1], self.internal_state[k][0])
self.internal_state_new.append((x.detach(),c.detach()))
x = getattr(self,name_batchnorm)(x)
if k!= len(self.hidden_list)-2:
x = getattr(self,name_relu)(x)
else :
x = self.sigmoid(x)
return x
So my question is, what in my code is causing memory to accumulate during the training phase?
A few quick notes about training code:
torch.Variable is deprecated since at least 8 minor versions (see here), don't use it
gc.collect() has no point, PyTorch does the garbage collector on it's own
Don't use torch.cuda.empty_cache() for each batch, as PyTorch reserves some GPU memory (doesn't give it back to OS) so it doesn't have to allocate it for each batch once again. It will make your code slow, don't use this function at all tbh, PyTorch handles this.
Don't spam random memory cleaning, that's most probably not where the error is
Model
Yes, this is probably the case (although it's hard to read this model's code).
Take notice of self.internal_state list and self.internal_state_new list also.
Each time you call model.initialize_hidden() a new set of tensor is added to this list (and never cleaned as far as I can tell)
self.internal_state_new seems to be cleaned in update_hidden, maybe self.internal_state should be also?
In essence, check out this self.internal_state property of your model, the list grows indefinitely from what I see. Initializing with zeros everywhere is quite strange, there is probably no need to do that (e.g. PyTorch's RNN is initialized with zeros by default, this is probably similar).

Why tensorflow1.1 gets slower and slower when training? Is it memory leak or queue starvation?

I trained a ESPCN in tensorflow1.1, the costed time per patch increase nearly linearly when training. The first 100 epoch takes only 4-5 seconds, but the 70th epoch takes about half a minute. See the training result below:
I've searched the same question on Google and Stack-overflow, and tried the solutions below, but seemed no work:
1.add tf.reset_default_graph() after every sess.run();
2.add time.sleep(5) to prevent queue starvation;
I know the general idea, that is to reduce the operations in Session(). But how? Anyone have the solution?
Here's part of my code:
L3, var_w_list, var_b_list = model_train(IN, FLAGS)
cost = tf.reduce_mean(tf.reduce_sum(tf.square(OUT - L3), reduction_indices=0))
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(FLAGS.base_lr, global_step * FLAGS.batch_size, FLAGS.decay_step, 0.96, staircase=True)
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(cost, global_step = global_step, var_list = var_w_list + var_b_list)
# optimizer = tf.train.MomentumOptimizer(learning_rate, 0.9).minimize(cost, var_list = var_w_list + var_b_list)
cnt = 0
with tf.Session() as sess:
init_op = tf.initialize_all_variables()
sess.run(init_op)
saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
print('\n\n\n =========== All initialization finished, now training begins ===========\n\n\n')
t_start = time.time()
t1 = t_start
for i in range(1, FLAGS.max_Epoch + 1):
LR_batch, HR_batch = batch.__next__()
global_step += 1
[_, cost1] = sess.run([optimizer, cost], feed_dict = {IN: LR_batch, OUT: HR_batch})
# tf.reset_default_graph()
if i % 100 == 0 or i == 1:
print_step = i
print_loss = cost1 / FLAGS.batch_size
test_LR_batch, test_HR_batch = test_batch.__next__()
test_SR_batch = test_HR_batch.copy()
test_SR_batch[:,:,:,0:3] = sess.run(L3, feed_dict = {IN: test_LR_batch[:,:,:,0:3]})
# tf.reset_default_graph()
psnr_tmp = 0.0
ssim_tmp = 0.0
for k in range(test_SR_batch.shape[0]):
com1 = test_SR_batch[k, :, :, 0]
com2 = test_HR_batch[k, :, :, 0]
psnr_tmp += get_psnr(com1, com2, FLAGS.HR_size, FLAGS.HR_size)
ssim_tmp += get_ssim(com1, com2, FLAGS.HR_size, FLAGS.HR_size)
psnr[cnt] = psnr_tmp / test_SR_batch.shape[0]
ssim[cnt] = ssim_tmp / test_SR_batch.shape[0]
ep[cnt] = print_step
t2 = time.time()
print_time = t2 - t1
t1 = t2
print(("[Epoch] : {0:d} [Current cost] : {1:5.8f} \t [Validation PSNR] : {2:5.8f} \t [Duration time] : {3:10.8f} s \n").format(print_step, print_loss, psnr[cnt], print_time))
# tf.reset_default_graph()
cnt += 1
if i % 1000 == 0:
L3_test = model_test(IN_TEST, var_w_list, var_b_list, FLAGS)
output_img = single_HR.copy()
output_img[:,:,:,0:3] = sess.run(L3_test, feed_dict = {IN_TEST:single_LR[:,:,:,0:3]})
tf.reset_default_graph()
subname = FLAGS.img_save_dir + '/' + str(i) + ".jpg"
img_gen(output_img[0,:,:,:], subname)
print(('================= Saving model to {}/model.ckpt ================= \n').format(FLAGS.checkpoint_dir))
time.sleep(5)
# saver.save(sess, FLAGS.checkpoint_dir + '/model.ckpt', print_step)
t_tmp = time.time() - t_start
My configuration is: windows10 + tf1.1 + python3.5 + cuda8.0 + cudnn5.1
================================================================
Besides, I used pixel-shuffle(PS) layer instead of deconvolution in the last layer. I copied the PS code from others, which is shown below:
def _phase_shift(I, r):
bsize, a, b, c = I.get_shape().as_list()
bsize = tf.shape(I)[0] # Handling Dimension(None) type for undefined batch dim
X = tf.reshape(I, (bsize, a, b, r, r))
X = tf.transpose(X, (0, 1, 2, 4, 3)) # bsize, a, b, 1, 1
X = tf.split(X, a, 1) # a, [bsize, b, r, r]
X = tf.concat([tf.squeeze(x, axis=1) for x in X], 2) # bsize, b, a*r, r
X = tf.split(X, b, 1) # b, [bsize, a*r, r]
X = tf.concat([tf.squeeze(x, axis=1) for x in X], 2) # bsize, a*r, b*r
return tf.reshape(X, (bsize, a*r, b*r, 1))
def PS(X, r, color=False):
if color:
Xc = tf.split(X, 3, 3)
X = tf.concat([_phase_shift(x, r) for x in Xc], 3)
else:
X = _phase_shift(X, r)
return X
Which X is the 4-dimensional image tensor, r means the up-scaling factor, color determine whether the channel of images is 3(Ycbcr format) or 1(Grayscale format).
To use the layer is very simple, just like the tf.nn.relu() does:
L3_ps = PS(L3, scale, True)
Now I'm wondering whether this layer caused the slowing-down, because the program goes well when using deconvolution layer. Using deconvolution layer may be a solution, but I have to use PS layer for some reason.
I suspect this line is causing a memory leak (although without seeing the code, I can't say for certain):
L3_test = model_test(IN_TEST, var_w_list, var_b_list, FLAGS)
L3_test seems to be a tf.Tensor (because you later pass it to sess.run(), so it seems likely that model_test() is adding new nodes to the graph each time it is called (every 1000 steps), which causes more work to be done over time.
The solution is quite simple though: since model_test() does not depend on anything calculated in the training loop, you can move the call to outside the training loop, so it is only called once.

TensorFlow, losses after training the model are different than losses printed during the last Epoch of Stochastic Gradient Descent.

I'm trying to do binary classification on two spirals. For testing, I am feeding my neural network the exact spiral data with no noise, and the model seems to work as the losses near 0 during SGD. However, after using my model to infer the exact same data points after SGD has completed, I get completely different losses than what was printed during the last epoch of SGD.
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
np.set_printoptions(threshold=np.nan)
# get the spiral points
t_p = np.linspace(0, 4, 1000)
x1_p = t_p * np.cos(t_p*2*np.pi)
y1_p = t_p * np.sin(t_p*2*np.pi)
x2_p = t_p * np.cos(t_p*2*np.pi + np.pi)
y2_p = t_p * np.sin(t_p*2*np.pi + np.pi)
plt.plot(x1_p, y1_p, x2_p, y2_p)
# generate data points
x1_dat = x1_p
y1_dat = y1_p
x2_dat = x2_p
y2_dat = y2_p
def model_variable(shape, name, initializer):
variable = tf.get_variable(name=name,
dtype=tf.float32,
shape=shape,
initializer=initializer
)
tf.add_to_collection('model_variables', variable)
return variable
class Model():
#layer specifications includes bias nodes
def __init__(self, sess, data, nEpochs, learning_rate, layer_specifications):
self.sess = sess
self.data = data
self.nEpochs = nEpochs
self.learning_rate = learning_rate
if layer_specifications[0] != 2 or layer_specifications[-1] != 1:
raise ValueError('First layer only two nodes, last layer only 1 node')
else:
self.layer_specifications = layer_specifications
self.build_model()
def build_model(self):
# x is the two nodes that will be layer one, will input an x, y coordinate
# and need to classify which spiral is it on, the non phase shifted or the phase
# shifted one.
# y is the output of the model
self.x = tf.placeholder(tf.float32, shape=[2, 1])
self.y = tf.placeholder(tf.float32, shape=[])
self.thetas = []
self.biases = []
for i in range(1, len(self.layer_specifications)):
self.thetas.append(model_variable([self.layer_specifications[i], self.layer_specifications[i-1]], 'theta'+str(i), tf.random_normal_initializer(stddev=0.1)))
self.biases.append(model_variable([self.layer_specifications[i], 1], 'bias'+str(i), tf.constant_initializer()))
#forward propagation
intermediate = self.x
for i in range(0, len(self.layer_specifications)-1):
if i != (len(self.layer_specifications) - 2):
intermediate = tf.nn.elu(tf.add(tf.matmul(self.thetas[i], intermediate), self.biases[i]))
else:
intermediate = tf.add(tf.matmul(self.thetas[i], intermediate), self.biases[i])
self.yhat = tf.squeeze(intermediate)
self.loss = tf.nn.sigmoid_cross_entropy_with_logits(self.yhat, self.y);
def train_init(self):
model_variables = tf.get_collection('model_variables')
self.optim = (
tf.train.GradientDescentOptimizer(learning_rate=self.learning_rate)
.minimize(self.loss, var_list=model_variables)
)
self.check = tf.add_check_numerics_ops()
self.sess.run(tf.initialize_all_variables())
# here is where x and y combine to get just x in tf with shape [2, 1] and where label becomes y in tf
def train_iter(self, x, y):
loss, _, _ = sess.run([self.loss, self.optim, self.check],
feed_dict = {self.x: x, self.y: y})
print('loss: {0} on:{1}'.format(loss, x))
# here x and y are still x and y coordinates, label is separate
def train(self):
for _ in range(self.nEpochs):
for x, y, label in self.data():
print(label)
self.train_iter([[x], [y]], label)
print("NEW ONE:\n")
# here x and y are still x and y coordinates, label is separate
def infer(self, x, y, label):
return self.sess.run((tf.sigmoid(self.yhat), self.loss), feed_dict={self.x : [[x], [y]], self.y : label})
def data():
#so first spiral is label 0, second is label 1
for _ in range(len(x1_dat)-1, -1, -1):
for dat in range(2):
if dat == 0:
yield x1_dat[_], y1_dat[_], 0
else:
yield x2_dat[_], y2_dat[_], 1
layer_specifications = [2, 100, 100, 100, 1]
sess = tf.Session()
model = Model(sess, data, nEpochs=10, learning_rate=1.1e-2, layer_specifications=layer_specifications)
model.train_init()
model.train()
inferrences_1 = []
inferrences_2 = []
losses = 0
for i in range(len(t_p)-1, -1, -1):
infer, loss = model.infer(x1_p[i], y1_p[i], 0)
if infer >= 0.5:
print('loss: {0} on point {1}, {2}'.format(loss, x1_p[i], y1_p[i]))
losses = losses + 1
inferrences_1.append('r')
else:
inferrences_1.append('g')
for i in range(len(t_p)-1, -1, -1):
infer, loss = model.infer(x2_p[i], y2_p[i], 1)
if infer >= 0.5:
inferrences_2.append('r')
else:
print('loss: {0} on point {1}, {2}'.format(loss, x2_p[i], y2_p[i]))
losses = losses + 1
inferrences_2.append('g')
print('total losses: {}'.format(losses))
plt.scatter(x1_p, y1_p, c=inferrences_1)
plt.scatter(x2_p, y2_p, c=inferrences_2)
plt.show()

Resources