I want to use an optimizer within the forward pass of a custom defined Function, but it doesn't work. My code is as follows:
class MyFct(Function):
#staticmethod
def forward(ctx, *args):
input, weight, bias = args[0], args[1], args[2]
y = torch.tensor([[0]], dtype=torch.float, requires_grad=True) #initial guess
loss_fn = lambda y_star: (input + weight - y_star)**2
learning_rate = 1e-4
optimizer = torch.optim.Adam([y], lr=learning_rate)
for t in range(5000):
y_star = y
print(y_star)
loss = loss_fn(y_star)
if t % 100 == 99:
print(t, loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
return y_star
And that's my test inputs:
x = torch.tensor([[2]], dtype=torch.float, requires_grad=True)
w = torch.tensor([[2]], dtype=torch.float, requires_grad=True)
y = torch.tensor([[6]], dtype=torch.float)
fct= MyFct.apply
y_hat = fct(x, w, None)
I always get the RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn.
Also, I've tested the optimization outside of the forward and it works, so I guess it's something with the context? According to the documentation "Tensor arguments that track history (i.e., with requires_grad=True) will be converted to ones that don’t track history before the call, and their use will be registered in the graph", see https://pytorch.org/docs/stable/notes/extending.html. Is this the problem? Is there a way to work around it?
I am new to PyTorch and I wonder what I'm overlooking. Any help and explanation is appreciated.
I think I found an answer here: https://github.com/pytorch/pytorch/issues/8847 , i.e. I need to wrap the oprimization with with torch.enable_grad():.
However, I still don't understand why it's necessary to convert the original Tensors to ones that don’t track history in forward().
Related
I have a deep model to train on CIFAR-10. Training works fine with CPU. However, when I use GPU support, it causes gradients for some batches to be NaNs (I checked it using tf.check_numerics) and it happens randomly but early enough. I believe the problem is related to my GPU.
My question is that: is there away not to update if at least one of the gradients has NaNs and force the model to proceed to the next batch ?
Edit: Perhaps I should elaborate more on my problem.
This is how I apply the gradients:
with tf.control_dependencies([tf.check_numerics(grad, message='Gradient %s check failed, possible NaNs' % var.name) for grad, var in grads]):
# Apply the gradients to adjust the shared variables.
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
I have thought of using tf.check_numerics first to verify that there are Nans in the gradients, and, then, if there are Nans (check failed) I can "pass" without using opt.apply_gradients. However, is there a way to catch an error with tf.control_dependencies ?
I could figure it out, albeit not in the most elegant way.
My solution is as follows:
1) check all gradients first
2) if gradients are NaNs-free, apply them
3) otherwise, apply fake update (with zero values), this needs gradient override.
This is my code:
First define custom gradient:
#tf.RegisterGradient("ZeroGrad")
def _zero_grad(unused_op, grad):
return tf.zeros_like(grad)
Then define an exception-handling function:
#this is added for gradient check of NaNs
def check_numerics_with_exception(grad, var):
try:
tf.check_numerics(grad, message='Gradient %s check failed, possible NaNs' % var.name)
except:
return tf.constant(False, shape=())
else:
return tf.constant(True, shape=())
Then create conditional node:
num_nans_grads = tf.Variable(1.0, name='num_nans_grads')
check_all_numeric_op = tf.reduce_sum(tf.cast(tf.stack([tf.logical_not(check_numerics_with_exception(grad, var)) for grad, var in grads]), dtype=tf.float32))
with tf.control_dependencies([tf.assign(num_nans_grads, check_all_numeric_op)]):
# Apply the gradients to adjust the shared variables.
def fn_true_apply_grad(grads, global_step):
apply_gradients_true = opt.apply_gradients(grads, global_step=global_step)
return apply_gradients_true
def fn_false_ignore_grad(grads, global_step):
#print('batch update ignored due to nans, fake update is applied')
g = tf.get_default_graph()
with g.gradient_override_map({"Identity": "ZeroGrad"}):
for (grad, var) in grads:
tf.assign(var, tf.identity(var, name="Identity"))
apply_gradients_false = opt.apply_gradients(grads, global_step=global_step)
return apply_gradients_false
apply_gradient_op = tf.cond(tf.equal(num_nans_grads, 0.), lambda : fn_true_apply_grad(grads, global_step), lambda : fn_false_ignore_grad(grads, global_step))
For the reinforcement learning one usually applies forward pass of the neural network for each step of the episode in order to calculate policy. Afterwards one could calculate parameter gradients using backpropagation. Simplified implementation of my network looks like this:
class AC_Network(object):
def __init__(self, s_size, a_size, scope, trainer, parameters_net):
with tf.variable_scope(scope):
self.is_training = tf.placeholder(shape=[], dtype=tf.bool)
self.inputs = tf.placeholder(shape=[None, s_size], dtype=tf.float32)
# (...)
layer = slim.fully_connected(self.inputs,
layer_size,
activation_fn=tf.nn.relu,
biases_initializer=None)
layer = tf.contrib.layers.dropout(inputs=layer, keep_prob=parameters_net["dropout_keep_prob"],
is_training=self.is_training)
self.policy = slim.fully_connected(layer, a_size,
activation_fn=tf.nn.softmax,
biases_initializer=None)
self.actions = tf.placeholder(shape=[None], dtype=tf.int32)
self.advantages = tf.placeholder(shape=[None], dtype=tf.float32)
actions_onehot = tf.one_hot(self.actions, a_size, dtype=tf.float32)
responsible_outputs = tf.reduce_sum(self.policy * actions_onehot, [1])
self.policy_loss = - policy_loss_multiplier * tf.reduce_mean(tf.log(responsible_outputs) * self.advantages)
local_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope)
self.gradients = tf.gradients(self.policy_loss, local_vars)
Now during training I will fist rollout the episode by consecutive forward passes (again, simplified version):
s = self.local_env.reset() # list of input variables for the first step
while done == False:
a_dist = sess.run([self.policy],
feed_dict = {self.local_AC.inputs: [s],
self.is_training: True})
a = np.argmax(a_dist)
s, r, done, extra_stat = self.local_env.step(a)
# (...)
and in the end I will calculate gradients by backward pass:
p_l, grad = sess.run([self.policy_loss,
self.gradients],
feed_dict={self.inputs: np.vstack(comb_observations),
self.is_training: True,
self.actions: np.hstack(comb_actions),})
(please note that I could have made a mistake somewhere above trying to remove as much as possible of the original code irrelevant to the issue in question)
So finally the question: Is there a way of ensuring that all the consecutive calls to the sess.run() will generate the same dropout structure? Ideally I would like to have exactly the same dropout structure within each episode and only change it between episodes. Things seem to work well as they are but I continue to wonder.
I found a peculiar property of lstm cell(not limited to lstm but I only examined with this) of tensorflow which has not been reported as far as I know.
I don't know whether it actually has, so I left this post in SO. Below is a toy code for this problem:
import tensorflow as tf
import numpy as np
import time
def network(input_list):
input,init_hidden_c,init_hidden_m = input_list
cell = tf.nn.rnn_cell.BasicLSTMCell(256, state_is_tuple=True)
init_hidden = tf.nn.rnn_cell.LSTMStateTuple(init_hidden_c, init_hidden_m)
states, hidden_cm = tf.nn.dynamic_rnn(cell, input, dtype=tf.float32, initial_state=init_hidden)
net = [v for v in tf.trainable_variables()]
return states, hidden_cm, net
def action(x, h_c, h_m):
t0 = time.time()
outputs, output_h = sess.run([rnn_states[:,-1:,:], rnn_hidden_cm], feed_dict={
rnn_input:x,
rnn_init_hidden_c: h_c,
rnn_init_hidden_m: h_m
})
dt = time.time() - t0
return outputs, output_h, dt
rnn_input = tf.placeholder("float", [None, None, 512])
rnn_init_hidden_c = tf.placeholder("float", [None,256])
rnn_init_hidden_m = tf.placeholder("float", [None,256])
rnn_input_list = [rnn_input, rnn_init_hidden_c, rnn_init_hidden_m]
rnn_states, rnn_hidden_cm, rnn_net = network(rnn_input_list)
feed_input = np.random.uniform(low=-1.,high=1.,size=(1,1,512))
feed_init_hidden_c = np.zeros(shape=(1,256))
feed_init_hidden_m = np.zeros(shape=(1,256))
sess = tf.Session()
sess.run(tf.global_variables_initializer())
for i in range(10000):
_, output_hidden_cm, deltat = action(feed_input, feed_init_hidden_c, feed_init_hidden_m)
if i % 10 == 0:
print 'Running time: ' + str(deltat)
(feed_init_hidden_c, feed_init_hidden_m) = output_hidden_cm
feed_input = np.random.uniform(low=-1.,high=1.,size=(1,1,512))
[Not important]What this code does is to generate an output from 'network()' function containing LSTM where the input's temporal dimension is 1, so output's is also 1, and pull in&out initial state for each step of running.
[Important] Looking the 'sess.run()' part. For some reasons in my real code, I happened to put [:,-1:,:] for 'rnn_states'. What is happening is then the time spent for each 'sess.run()' increases. For some inspection by my own, I found this slow down stems from that [:,-1:,:]. I just wanted to get the output at the last time step. If you do 'outputs, output_h = sess.run([rnn_states, rnn_hidden_cm], feed_dict{~' w/o [:,-1:,:] and take 'last_output = outputs[:,-1:,:]' after the 'sess.run()', then the slow down does not occur.
I do not know why this exponential increment of time happens with that [:,-1:,:] running. Is this the nature of tensorflow hasn't been documented but particularly slows down(may be adding more graph by its own?)?
Thank you, and hope this mistake not happen for other users by this post.
I encountered the same problem, with TensorFlow slowing down for each iteration I ran it, and found this question while trying to debug it. Here's a short description of my situation and how I solved it for future reference. Hopefully it can point someone in the right direction and save them some time.
In my case the problem was mainly that I didn't make use of feed_dict to supply the network state when executing sess.run(). Instead I redeclared outputs, final_state and prediction every iteration. The answer at https://github.com/tensorflow/tensorflow/issues/1439#issuecomment-194405649 made me realize how stupid that was... I was constantly creating new graph nodes in every iteration, making it all slower and slower. The problematic code looked something like this:
# defining the network
lstm_layer = rnn.BasicLSTMCell(num_units, forget_bias=1)
outputs, final_state = rnn.static_rnn(lstm_layer, input, initial_state=rnn_state, dtype='float32')
prediction = tf.nn.softmax(tf.matmul(outputs[-1], out_weights)+out_bias)
for input_data in data_seq:
# redeclaring, stupid stupid...
outputs, final_state = rnn.static_rnn(lstm_layer, input, initial_state=rnn_state, dtype='float32')
prediction = tf.nn.softmax(tf.matmul(outputs[-1], out_weights)+out_bias)
p, rnn_state = sess.run((prediction, final_state), feed_dict={x: input_data})
The solution was of course to only declare the nodes once in the beginning, and supply the new data with feed_dict. The code went from being half slow (> 15 ms in the beginning) and becoming slower for every iteration, to execute every iteration in around 1 ms. My new code looks something like this:
out_weights = tf.Variable(tf.random_normal([num_units, n_classes]), name="out_weights")
out_bias = tf.Variable(tf.random_normal([n_classes]), name="out_bias")
# placeholder for the network state
state_placeholder = tf.placeholder(tf.float32, [2, 1, num_units])
rnn_state = tf.nn.rnn_cell.LSTMStateTuple(state_placeholder[0], state_placeholder[1])
x = tf.placeholder('float', [None, 1, n_input])
input = tf.unstack(x, 1, 1)
# defining the network
lstm_layer = rnn.BasicLSTMCell(num_units, forget_bias=1)
outputs, final_state = rnn.static_rnn(lstm_layer, input, initial_state=rnn_state, dtype='float32')
prediction = tf.nn.softmax(tf.matmul(outputs[-1], out_weights)+out_bias)
# actual network state, which we input with feed_dict
_rnn_state = tf.nn.rnn_cell.LSTMStateTuple(np.zeros((1, num_units), dtype='float32'), np.zeros((1, num_units), dtype='float32'))
it = 0
for input_data in data_seq:
encl_input = [[input_data]]
p, _rnn_state = sess.run((prediction, final_state), feed_dict={x: encl_input, rnn_state: _rnn_state})
print("{} - {}".format(it, p))
it += 1
Moving the declaration out from the for loop also got rid of the problem which the OP sdr2002 had, doing a slice outputs[-1] in sess.run() inside the for loop.
As mentioned above, no sliced output for 'sess.run()' is much appreciated for this case.
def action(x, h_c, h_m):
t0 = time.time()
outputs, output_h = sess.run([rnn_states, rnn_hidden_cm], feed_dict={
rnn_input:x,
rnn_init_hidden_c: h_c,
rnn_init_hidden_m: h_m
})
outputs = outputs[:,-1:,:]
dt = time.time() - t0
return outputs, output_h, dt
I want to use maxout activation function in tensorflow, but I don't know which function should use.
I sent a pull request for maxout, here is the link:
https://github.com/tensorflow/tensorflow/pull/5528
Code is as follows:
def maxout(inputs, num_units, axis=None):
shape = inputs.get_shape().as_list()
if axis is None:
# Assume that channel is the last dimension
axis = -1
num_channels = shape[axis]
if num_channels % num_units:
raise ValueError('number of features({}) is not a multiple of num_units({})'
.format(num_channels, num_units))
shape[axis] = -1
shape += [num_channels // num_units]
outputs = tf.reduce_max(tf.reshape(inputs, shape), -1, keep_dims=False)
return outputs
Here is how it works:
I don't think there is a maxout activation but there is nothing stopping yourself from making it yourself. You could do something like the following.
with tf.variable_scope('maxout'):
layer_input = ...
layer_output = None
for i in range(n_maxouts):
W = tf.get_variable('W_%d' % d, (n_input, n_output))
b = tf.get_variable('b_%d' % i, (n_output,))
y = tf.matmul(layer_input, W) + b
if layer_output is None:
layer_output = y
else:
layer_output = tf.maximum(layer_output, y)
Note that this is code I just wrote in my browser so there may be syntax errors but you should get the general idea. You simply perform a number of linear transforms and take the maximum across all the transforms.
How about this code?
This seems to work in my test.
def max_out(input_tensor,output_size):
shape = input_tensor.get_shape().as_list()
if shape[1] % output_size == 0:
return tf.transpose(tf.reduce_max(tf.split(input_tensor,output_size,1),axis=2))
else:
raise ValueError("Output size or input tensor size is not fine. Please check it. Reminder need be zero.")
I refer the diagram in the following page.
From version 1.4 on you can use tf.contrib.layers.maxout.
Maxout is a layer such that it calculates N*M output for a N*1 input, and then it returns the maximum value across the column, i.e., the final output has shape N*1 as well. Basically it uses multiple linear fittings to mimic a complex function.
The documentation is not quite clear about this. I suppose the gradients one can obtain by opt.compute_gradients(E, [v]) contain the ∂E/∂x = g(x) for each element x of the tensor that v stores. Does opt.apply_gradients(grads_and_vars) essentially execute x ← -η·g(x), where η is the learning rate? That would imply that if I want to add a positive additive change p to the variable, I would need to need to change g(x) ← g(x) - (1/η)p, e.g. like this:
opt = tf.train.GradientDescentOptimizer(learning_rate=l)
grads_and_vars = opt.compute_gradients(loss, var_list)
for l, gv in enumerate(grads_and_vars):
grads_and_vars[l] = (gv[0] - (1/l) * p, gv[1])
train_op = opt.apply_gradients(grads_and_vars)
Is there a better way to do this?
The update rule that the apply_gradients method actually applies depends on the specific optimizer. Take a look at the implementation of apply_gradients in the tf.train.Optimizer class here. It relies on the derived classes implementing the update rule in the methods _apply_dense and _apply_spares. The update rule you are referring to is implemented by the GradientDescentOptimizer.
Regarding your desired positive additive update: If what you are calling opt is an instantiation of GradientDescentOptimizer, then you could indeed achieve what you want to do by
grads_and_vars = opt.compute_gradients(E, [v])
eta = opt._learning_rate
my_grads_and_vars = [(g-(1/eta)*p, v) for g, v in grads_and_vars]
opt.apply_gradients(my_grads_and_vars)
The more elegant way to do this is probably to write a new optimizer (inheriting from tf.train.Optimizer) that implements your desired update rule directly.
You can also use eager execution API.
import tensorflow as tf
tf.enable_eager_execution()
tfe = tf.contrib.eager
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
grad = tfe.implicit_gradients(loss)
optimizer.apply_gradients(grad(model_fn, val_list))
I will make an instance for it as follow:
import tensorflow as tf
tf.enable_eager_exeuction()
tfe = tf.contrib.eager
W = tfe.Variable(np.random.randn())
b = tfe.Variable(np.random.randn())
def linear_regression(inputs):
return inputs * W + b;
def MSE(model_fn, inputs, labels):
return tf.reduce_sum(tf.pow(model_fn(inputs) - labels, 2)) / (2 * n_samples)
optimizer = tf.train.GradientDescentOptimizer(learning_rate = 0.001)
grad = tfe.implicit_gradients(MSE)
optimizer.apply_gradients(grad(linear_regression, train_X, train_Y)) # train_X and train_Y are your input data and label