Hi Kate, thanks for the Pytorch code of MAML!

I have two questions(in which I suspect a bug?) on your implementation.

Line 10, Algorithm2 from the original paper indicates that meta_update is performed using each D'_i. To do so with your code, I think the function meta_update need access to every task sampled, since each task contains its D'_i in your implementation. https://github.com/katerakelly/pytorch-maml/blob/75907aca148ad053dfaf75fc138319f0d89534a8/src/maml.py#L172 However, it seems that you perform meta_update with a single task, resulting in using only one D'_i of a specific task.

Line 10 also states that meta-loss is calculated with adapted parameters. https://github.com/katerakelly/pytorch-maml/blob/75907aca148ad053dfaf75fc138319f0d89534a8/src/maml.py#L71 You seem to have calculated meta-loss with self.net, which I think is "original parameters"(\theta_i) in stead of adapted parameters.

Am I missing something?

Hi,

Thanks for sharing the code. I have questions about the implementation for inner loop: Is there any reason for the special case of i == 0? Can we just use fast_weights for i == 0?

Thanks!

I have encountered the problem "IndexError: too many indices for array". Is it related to the dataset I use? I just download this file (https://github.com/brendenlake/omniglot/blob/master/python/images_evaluation.zip) and put it under data. The log is as follows:

./train-omniglot-5way-1shot.sh
tee: ../logs/maml-omniglot-5way-1shot-TEST: No such file or directory
exp maml-omniglot-5way-1shot-TEST
dataset omniglot
num_cls 5
num_inst 1
batch 1
m_batch 32
num_updates 15000
num_inner_updates 5
lr 1e-1
meta_lr 1e-3
gpu 0
Setting GPU to 0
init weights
init weights
init weights
Traceback (most recent call last):
  File "maml.py", line 230, in <module>
    main()
  File "/home/maple/programs/miniconda2/lib/python2.7/site-packages/click/core.py", line 722, in __call__
    return self.main(*args, **kwargs)
  File "/home/maple/programs/miniconda2/lib/python2.7/site-packages/click/core.py", line 697, in main
    rv = self.invoke(ctx)
  File "/home/maple/programs/miniconda2/lib/python2.7/site-packages/click/core.py", line 895, in invoke
    return ctx.invoke(self.callback, **ctx.params)
  File "/home/maple/programs/miniconda2/lib/python2.7/site-packages/click/core.py", line 535, in invoke
    return callback(*args, **kwargs)
  File "maml.py", line 227, in main
    learner.train(exp)
  File "maml.py", line 151, in train
    mt_loss, mt_acc, mv_loss, mv_acc = self.test()
  File "maml.py", line 113, in test
    tloss, tacc = evaluate(test_net, train_loader)
  File "/home/maple/pytorch-maml/src/score.py", line 30, in evaluate
    loss += l.data.cpu().numpy()[0]
IndexError: too many indices for array

In the algorithm 2 mentioned in this paper,we have computed adapted parameters θ' in step 7，and I think we have updated the model for task T_i. What does the meta-update mean?(in step 8)

Thanks for your helpful codes.

I want to change the origin net's weights using tasks-average gradients on every task's query set. then using opt.step to uodate it. Inspired by the optim.GSD source code, Can your code: ` hooks = [] for (k,v) in self.net.named_parameters(): def get_closure(): key = k def replace_grad(grad): return gradients[key] return replace_grad hooks.append(v.register_hook(get_closure()))

Compute grads for current step, replace with summed gradients as defined by hook

self.opt.zero_grad() loss.backward()

Update the net parameters with the accumulated gradient according to optimizer

self.opt.step()

Remove the hooks before next training phase

for h in hooks: h.remove() `

be replaced by: for (k,v), (k,g) in zip(self.net.named_parameters(), gradients): v.grad.data = g.data self.opt.step() self.opt.zero_grad()

or just by: for (k,v), (k,g) in zip(self.net.named_parameters(), gradients): v.data.add_(-meta_lr, g.data)

Thanks for your time!

In this line: https://github.com/katerakelly/pytorch-maml/blob/master/src/inner_loop.py#L76 why don't use grads = torch.autograd.grad(loss, fast_weights.values()) ? but self.paramters() which is the original parameters before any inner-update step.

In the original paper, the authors claimed that MAML needs second gradient and Hessian-vector products. Could you explain how do you implement this or Pytorch just do this automatically? Thanks! referenced https://github.com/dragen1860/MAML-Pytorch/issues/18

Hi Kate Rakelly, Thank you for sharing the MAML source code. Here I have a few questions: 1  Since you have obtained the meta_grads in the function “forward()”, why not directly utilize them to update the parameters of self.net. (param - learning rategradient)  (Here, it seems that you want to achieve that by defining a “meta_update()” function. And you make the annotation: “We use a dummy forward / backward pass to get the correct grads into self.net”. for this command: “loss, out = forward_pass(self.net, in_, target)” I’m confused why you compute the loss using val data with respect to original “\theta”.) 2. In the “meta_update()” function, you use this (register_hook) to change the gradients. But the functions are(opt.zero_grad;  loss.backward) also used to compute the gradients. Do these functions(opt.zero_grad;  loss.backward) overwrite the previous gradients (changed by using the register_hook function)? I’m looking forward to your reply.

Hi, thanks for the code.

I am trying test the code on the mini-imagenet dataset. However, my result is very bad, much worse than the reported ones. I wonder if you have tested the implementation on the dataset, and how about your results. If possible, could you possible upload your scripts of experimenting on the mini-imagenet dataset. Your help is much appreciated.

Thanks for the code. I meet the following problem when trying to run your code. I used pytorch 0.2.0, python 3.5.

tee: ../logs/maml-omniglot-5way-1shot-TEST: No such file or directory
exp maml-omniglot-5way-1shot-TEST
dataset omniglot
num_cls 5
num_inst 1
batch 1
m_batch 32
num_updates 15000
num_inner_updates 5
lr 1e-1
meta_lr 1e-3
gpu 0
Setting GPU to 0
init weights
init weights
init weights
> /zdata/users/kaili/code/pytorch-maml/src/maml.py(102)test()
-> for _ in range(10):
(Pdb) c
-------------------------
Meta train: 0.1178786426782608 1.0
Meta val: 1.139253568649292 0.7
-------------------------
inner step 0
inner step 1
Traceback (most recent call last):
  File "maml.py", line 234, in <module>
    main()
  File "/home/mli/kaili/anaconda2/envs/kai_py35_torch02/lib/python3.5/site-packages/click/core.py", line 722, in __call__
    return self.main(*args, **kwargs)
  File "/home/mli/kaili/anaconda2/envs/kai_py35_torch02/lib/python3.5/site-packages/click/core.py", line 697, in main
    rv = self.invoke(ctx)
  File "/home/mli/kaili/anaconda2/envs/kai_py35_torch02/lib/python3.5/site-packages/click/core.py", line 895, in invoke
    return ctx.invoke(self.callback, **ctx.params)
  File "/home/mli/kaili/anaconda2/envs/kai_py35_torch02/lib/python3.5/site-packages/click/core.py", line 535, in invoke
    return callback(*args, **kwargs)
  File "maml.py", line 231, in main
    learner.train(exp)
  File "maml.py", line 166, in train
    metrics, g = self.fast_net.forward(task)
  File "/zdata/users/kaili/code/pytorch-maml/src/inner_loop.py", line 61, in forward
    loss, _ = self.forward_pass(in_, target, fast_weights)
  File "/zdata/users/kaili/code/pytorch-maml/src/inner_loop.py", line 43, in forward_pass
    out = self.net_forward(input_var, weights)
  File "/zdata/users/kaili/code/pytorch-maml/src/inner_loop.py", line 36, in net_forward
    return super(InnerLoop, self).forward(x, weights)
  File "/zdata/users/kaili/code/pytorch-maml/src/omniglot_net.py", line 49, in forward
    x = batchnorm(x, weight = weights['features.bn1.weight'], bias = weights['features.bn1.bias'], momentum=1)
  File "/zdata/users/kaili/code/pytorch-maml/src/layers.py", line 31, in batchnorm
    running_mean = torch.zeros(np.prod(np.array(input.data.size()[1]))).cuda()
TypeError: torch.zeros received an invalid combination of arguments - got (numpy.int64), but expected one of:
 * (int ... size)
      didn't match because some of the arguments have invalid types: (!numpy.int64!)
 * (torch.Size size)
      didn't match because some of the arguments have invalid types: (!numpy.int64!)

Any clue to solve this? Thanks.

Thanks for your nice implementation. But I meet this mistake when I sh train-omniglot-20way-1shot.sh ,is there any dimension dismatch? I don't know how to fix it. It's much appreciated if anyone can help me. Traceback (most recent call last): File "maml.py", line 230, in main() File "/usr/local/lib/python3.6/dist-packages/click/core.py", line 829, in call return self.main(*args, **kwargs) File "/usr/local/lib/python3.6/dist-packages/click/core.py", line 782, in main rv = self.invoke(ctx) File "/usr/local/lib/python3.6/dist-packages/click/core.py", line 1066, in invoke return ctx.invoke(self.callback, **ctx.params) File "/usr/local/lib/python3.6/dist-packages/click/core.py", line 610, in invoke return callback(*args, **kwargs) File "maml.py", line 227, in main learner.train(exp) File "maml.py", line 151, in train mt_loss, mt_acc, mv_loss, mv_acc = self.test() File "maml.py", line 113, in test tloss, tacc = evaluate(test_net, train_loader) File "/home/src-plot/score.py", line 30, in evaluate loss += l.data.cpu().numpy()[0] IndexError: too many indices for array

Thanks for your good implementation of MAML, however, I think that maybe use state_dict() and load_stat_dict() is much faster than modifying the weights (in omniglot_net.py 43), can I first deepcopy the net parameters(state_dict()) and use the fast weights (also use a optimizer to update), then load the origin parameters back to update the meta learner? Thanks.

Hi Kate,

Under split='train' in inner loop, the data loader always output the same data no matter under what batch_cutoff. I know that this is the setting in the MAML paper, but I am wondering is there a way to make batch_cutoff work? Many thanks!

Hi kelly, I read the code and find that the get_data_loader function get the same batch samples every iteration, is that you want? (sorry, I'm not familiar with the maml algorithm.)

below is the samples got from the get_data_loader funtion.

('test train loader: ', ['../data/omniglot/images_evaluation/Kannada/character01/1205_20.png', '../data/omniglot/images_evaluation/Manipuri/character05/1323_07.png', '../data/omniglot/images_evaluation/Angelic/character06/0970_19.png', '../data/omniglot/images_evaluation/ULOG/character14/1611_15.png', '../data/omniglot/images_evaluation/Keble/character02/1247_11.png', '../data/omniglot/images_evaluation/Kannada/character05/1209_20.png', '../data/omniglot/images_evaluation/Sylheti/character05/1507_12.png', '../data/omniglot/images_evaluation/Manipuri/character15/1333_01.png', '../data/omniglot/images_evaluation/Kannada/character27/1231_08.png', '../data/omniglot/images_evaluation/Mongolian/character30/1388_20.png', '../data/omniglot/images_evaluation/Keble/character07/1252_12.png', '../data/omniglot/images_evaluation/Tibetan/character05/1560_19.png', '../data/omniglot/images_evaluation/Glagolitic/character14/1128_04.png', '../data/omniglot/images_evaluation/Ge_ez/character06/1094_07.png', '../data/omniglot/images_evaluation/Avesta/character05/1067_17.png', '../data/omniglot/images_evaluation/Malayalam/character05/1276_16.png', '../data/omniglot/images_evaluation/Aurek-Besh/character03/1039_18.png', '../data/omniglot/images_evaluation/Syriac_(Serto)/character14/1493_07.png', '../data/omniglot/images_evaluation/Keble/character16/1261_09.png', '../data/omniglot/images_evaluation/Aurek-Besh/character13/1049_15.png']) ('test train loader: ', ['../data/omniglot/images_evaluation/Syriac_(Serto)/character14/1493_07.png', '../data/omniglot/images_evaluation/Sylheti/character05/1507_12.png', '../data/omniglot/images_evaluation/Keble/character16/1261_09.png', '../data/omniglot/images_evaluation/Ge_ez/character06/1094_07.png', '../data/omniglot/images_evaluation/Tibetan/character05/1560_19.png', '../data/omniglot/images_evaluation/Manipuri/character05/1323_07.png', '../data/omniglot/images_evaluation/Kannada/character05/1209_20.png', '../data/omniglot/images_evaluation/Manipuri/character15/1333_01.png', '../data/omniglot/images_evaluation/Avesta/character05/1067_17.png', '../data/omniglot/images_evaluation/ULOG/character14/1611_15.png', '../data/omniglot/images_evaluation/Glagolitic/character14/1128_04.png', '../data/omniglot/images_evaluation/Malayalam/character05/1276_16.png', '../data/omniglot/images_evaluation/Aurek-Besh/character13/1049_15.png', '../data/omniglot/images_evaluation/Kannada/character01/1205_20.png', '../data/omniglot/images_evaluation/Keble/character02/1247_11.png', '../data/omniglot/images_evaluation/Angelic/character06/0970_19.png', '../data/omniglot/images_evaluation/Aurek-Besh/character03/1039_18.png', '../data/omniglot/images_evaluation/Mongolian/character30/1388_20.png', '../data/omniglot/images_evaluation/Kannada/character27/1231_08.png', '../data/omniglot/images_evaluation/Keble/character07/1252_12.png']) ('test train loader: ', ['../data/omniglot/images_evaluation/Kannada/character05/1209_20.png', '../data/omniglot/images_evaluation/Malayalam/character05/1276_16.png', '../data/omniglot/images_evaluation/Manipuri/character05/1323_07.png', '../data/omniglot/images_evaluation/Tibetan/character05/1560_19.png', '../data/omniglot/images_evaluation/Kannada/character27/1231_08.png', '../data/omniglot/images_evaluation/Ge_ez/character06/1094_07.png', '../data/omniglot/images_evaluation/Keble/character16/1261_09.png', '../data/omniglot/images_evaluation/Kannada/character01/1205_20.png', '../data/omniglot/images_evaluation/Glagolitic/character14/1128_04.png', '../data/omniglot/images_evaluation/Angelic/character06/0970_19.png', '../data/omniglot/images_evaluation/Aurek-Besh/character03/1039_18.png', '../data/omniglot/images_evaluation/ULOG/character14/1611_15.png', '../data/omniglot/images_evaluation/Aurek-Besh/character13/1049_15.png', '../data/omniglot/images_evaluation/Keble/character02/1247_11.png', '../data/omniglot/images_evaluation/Sylheti/character05/1507_12.png', '../data/omniglot/images_evaluation/Avesta/character05/1067_17.png', '../data/omniglot/images_evaluation/Manipuri/character15/1333_01.png', '../data/omniglot/images_evaluation/Syriac_(Serto)/character14/1493_07.png', '../data/omniglot/images_evaluation/Mongolian/character30/1388_20.png', '../data/omniglot/images_evaluation/Keble/character07/1252_12.png']) ('test train loader: ', ['../data/omniglot/images_evaluation/Keble/character02/1247_11.png', '../data/omniglot/images_evaluation/ULOG/character14/1611_15.png', '../data/omniglot/images_evaluation/Tibetan/character05/1560_19.png', '../data/omniglot/images_evaluation/Kannada/character27/1231_08.png', '../data/omniglot/images_evaluation/Kannada/character05/1209_20.png', '../data/omniglot/images_evaluation/Syriac_(Serto)/character14/1493_07.png', '../data/omniglot/images_evaluation/Aurek-Besh/character03/1039_18.png', '../data/omniglot/images_evaluation/Avesta/character05/1067_17.png', '../data/omniglot/images_evaluation/Glagolitic/character14/1128_04.png', '../data/omniglot/images_evaluation/Malayalam/character05/1276_16.png', '../data/omniglot/images_evaluation/Mongolian/character30/1388_20.png', '../data/omniglot/images_evaluation/Manipuri/character15/1333_01.png', '../data/omniglot/images_evaluation/Manipuri/character05/1323_07.png', '../data/omniglot/images_evaluation/Aurek-Besh/character13/1049_15.png', '../data/omniglot/images_evaluation/Sylheti/character05/1507_12.png', '../data/omniglot/images_evaluation/Kannada/character01/1205_20.png', '../data/omniglot/images_evaluation/Ge_ez/character06/1094_07.png', '../data/omniglot/images_evaluation/Keble/character07/1252_12.png', '../data/omniglot/images_evaluation/Keble/character16/1261_09.png', '../data/omniglot/images_evaluation/Angelic/character06/0970_19.png']) ('test train loader: ',