Suboptimal convergence when compared with TensorFlow model

The paper says:

Note that the efficiency of algorithm 1 can, at the expense of clarity, be improved upon by changing the order of computation, e.g. by replacing the last three lines in the loop with the following lines…

So I’m surprised that it should make a noticeable difference, but maybe that is the case. @christianperone would you mind trying the altered version of Adam on your problem? Fingers crossed this might be the solution.

The bit about sparse updates with TensorFlow Adam I would assume don’t matter in this case. I don’t know how PyTorch deals with sparse modules wrt gradient updates, but what TF claims to do sounds like the correct approach.


It was this part that made me think it could lead to noticeable difference–

Tensorflow Adam – “    The sparse implementation of this algorithm (used when the gradient is an
    IndexedSlices object, typically because of `tf.gather` or an embedding
    lookup in the forward pass) does apply momentum to variable slices even if
    they were not used in the forward pass (meaning they have a gradient equal to zero. Momentum decay (beta1) is also applied to the entire momentum
accumulator. This means that the sparse behavior is equivalent to the dense
behavior (in contrast to some momentum implementations which ignore momentum
unless a variable slice was actually used).”

As I see a lot of training embedding models in pytorch and would be comparing to tensorflow I bet a lot of these performance differences stem from that as it auto applies the momentum decay and we would have default not too.

Anyways I have always been able to get just as good or better than tensorflow performance but I usually use custom stuff most the time but the underlying framework has shown no insuffiency in performance for me and usually find quite the opposite

Small minute differences in hyperparameters do often show unproportional performance differences in my experience

I’ll test it. If someone has the code change in hands that would help a lot, otherwise I’ll have to come back to this in near future due to my time constraints. Thanks for the help !


I also experienced suboptimal behaviour with Adam compared to SGD in PyTorch. Similar code in Tensorflow performed the other way around, i.e. optimizing with Adam was much easier. I have also used an Embedding layer.


I thought I was the only one! Same problem here: RNN and Adam: slower convergence than Keras

When I’ll have time I’ll try with other optimizers.

EDIT: same situation with RMSProp.

Up… Shouldn’t this problem be investigated?


Same issue here, same model architecture in Keras that is trained using Adam, gives better result comparing with Pytorch.

1 Like

There are many factors that can cause differences. Some people have reported things to try here.

Same problem here. Cannot replicate TF Adam optimizer success in Pytorch.

Edit: Disregard. I’m actually getting better loss in Pytorch over TF with Adam now that I’m actually taking the mean of my losses.
size_average=False found in jcjohnson’s github examples can make for a long night for a newbie.

I also have the same problem.
I implemented AE and VAE on both Keras(Tensorflow) and Pytorch.
Using Adadelta gave me different loss values and Pytorch did the worst thing on my network.
I spent 2 weeks to double check my codes untill I found this post.
Thank you guys that I am not the only one who experiences this issue.


Same problem here!

More specifically, it turns out that Pytorch training with Adam will stuck at a worse level (in terms of both loss and accuracy) than Tensorflow with exactly the same setting. I came across this issue in two process:

(1) standard training of a VGG-16 model with CIFAR-10 as dataset.
(2) generating CW L2 attack. See for details. I reproduce this attack method to test my model trained with Pytorch. The loss also stuck at a undesirable level for some images, and the adversarial counterparts couldn’t be generated.

Interestingly, I solved these issues by manually letting the learning rate decay to its half at scheduled step (e.g. lr = 0.5 * lr, every 20 epochs). After doing so, Pytorch could reach comparable results as Tensorflow (without decaying its learning rate), and everything works fine for me.

However, I think that actually Adam should adjust its learning rate automatically. So I still don’t know the true reason for this.


In general, a whole learning system consists of:

  1. data loading (including train/val/test split, data augmentation, batching, etc)
  2. prediction model (your neural network)
  3. loss computation
  4. gradient computation
  5. model initialization
  6. optimization
  7. metric (accuracy, precision, etc) computation

In my experience, double check every aspect of you code before concluding it is an optimizer-related issue (Most of the time, it’s not…).

Specifically, you can do the followings to check the correctness of your code:

  • [easy check] switch optimizers (SGD, SGD + momentum, etc.) and check if the performance gap persists
  • [easy check] disable more advanced techniques like BatchNorm, Dropout and check the final performance
  • use the same dataloader (therefore, both tensorflow and pytorch will get the same inputs for every batch) and check the final performance
  • use the same inputs, check both the forward and backward outputs

Good Luck.


Can anyone from the PyTorch Dev team address this issue? @ptrblck @smth

@bily’s suggestions seem very reasonable.
If you still have some issues getting approx. the same results, I would like to dig a bit deeper.
Also, it would help if you could provide executable scripts for both implementations.

Also, since the loss function is non-convex, random weight initialization can make huge difference. I recommend repeating the experiment with ~5 different random seeds in both frameworks (TensorFlow, PyTorch and then compare the top ~1-3 results.



I’m having the same problem, and spent long time to double check all what @bily suggested.
Here are two projects, one is the original Tensorflow code of a paper called “Fast-Slow Recurrent Neural Networks”, which had state of the art results in Language Model task.
The second is my Pytorch implementation.
I got poor results using Adam optimizer. I also tried different optimizers on both implementations, but still got poor results. It seems that no matter what optimizer I choose, the Pytorch loss stack at some level where TF loss keep getting smaller.

Here are the links for both project in my github account:
Pytorch implementation:
TF implementation:

I removed the fancy optimizations in both implementatinos (like zoneout and layer normalization) but still got poor results in Pytorch compared to TF.

The architecture is not complicated at all, its only 3 LSTM cells. Just look at the forward method to understand it.

I’ll appreciate you response on it.

1 Like

What I usually do at that point:

  • Do you get the same outputs for the same inputs (I usually save a batch from TF in numpy format when I do this)?
  • If so, do you get the same gradients?
    (again, I usually save the TF gradients in numpy to compare)

Best regards


Hi Tom,

No, I didn’t do one to one comparison. I would have to export the initial weights for that manner from TF classifier to the Pytorch one and then run the network.
And also make sure that the input is the same and in the same order of course.

What I did do is I checked that each batch contains the same samples in both implementations. It does.
But the batches don’t come at the same order which shouldn’t be a problem.

I also checked that the gradients are pretty much on the same scale during the run. Means that after each batch I printed the gradients and look at the numbers. So in the first batches, the gradients are big and then getting lower during the epochs. Same scale in both implementations.

I’ll consider your advice about trying to replicate the results of the TF network with my Pytorch one.

Another thing to consider is that I think Tf and PyTorch use different default weight initialization schemes, which may also have an effect (and will also effect the learning rate etc etc)