Summary:
There is a module called `2to3` which you can target for future specifically to remove these, the directory of `caffe2` has the most redundant imports:
```2to3 -f future -w caffe2```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/45033
Reviewed By: seemethere
Differential Revision: D23808648
Pulled By: bugra
fbshipit-source-id: 38971900f0fe43ab44a9168e57f2307580d36a38
* Update elementwise ops to support numpy style boradcast
Update elementwise ops to support numpy style boradcast
* Fix sqrt_op
* Fix compare ops
* Fix gradient test
* Fix optimizer legacy broadcast
* Fix legacy broadcast for elementwise ops
* Skip flaky test
* Fix eigen simple binary op
* Fix attention test
* Fix rnn test
* Fix LSTM test
* Fix tan grad
* Fix schema check
Summary: PR 1175 caused a build error because gemmBatched was only under a specific #ifdef. Now put it outside the #ifdef, and things work.
Reviewed By: asaadaldien
Differential Revision: D5834868
fbshipit-source-id: 072a64c8f4b259ff7504104121766115b46b8aa0
Summary:
Implementation of a new variant of attention module, which contains a recurrent decoder state with vectors corresponding to each source-side word and strictly increasing values, thus enabling it to model the degree to which source words have been translated.
The approach is a variant of the approaches described in https://arxiv.org/pdf/1601.04811.pdf. We simply include the sum of all previous attention weights for encoder words as a new recurrent state (coverage_t). A new linear transform on encoder_outputs is used to produce coverage_weights, which has the same dimensionality as encoder_outputs, and implicitly models the fertility of source-side words (and putting this extra information strain on the encoder network).
Thus the encoder output, the decoder state, and the coverage weights have the same dimensionality for a given source word, and attention logits are calculated as v * tanh(coverage * coverage_weights + encoder_output + decoder_state).
Note: the entire coverage state for each translation instance is of shape (encoder_length, coverage_units), but the states for the RecurrentNetwork operator, used to train the decoder, must be flat in the data dimension. This state is therefore initialized with shape (encoder_length * coverage_units) [not shown in the open-source library] and reshaped appropriately within the apply_soft_coverage_attention() function.
Differential Revision: D5593617
fbshipit-source-id: 7d0522b5eb0b26f22e8429e4461a459f2f16ed46
Summary: Use the new SequenceMask op to mask out invalid positions in the attention mechanism rather than using PackSegments and UnpackSegments. This should help us on several fronts, including elision of host<>device copies and using fewer intermediate blobs
Differential Revision: D5619156
fbshipit-source-id: e59c644236cee02f853d8743f9a938fb10adc73b
Summary: Make the command-line arguments pertaining to model architecture the same as between train.py and translate.py. Also use s() scoping function for all intermediate blobs in attention.py (this is for comatibility with multi-headed attention).
Differential Revision: D5594312
fbshipit-source-id: cadf51d854b5a9174ec913f32c655be2abf111e5
Summary:
Implement dot attention as described in https://arxiv.org/abs/1508.04025
This saves the computation of weighted encoder outputs in `rnn_cell.py`
When the encoder and decoder dimensions are different, we apply an FC, which corresponds to the general case below Figure 2.
Refactored unit tests.
Reviewed By: jhcross
Differential Revision: D5486976
fbshipit-source-id: f9e9aea675b3b072fbe631bc004199b90a9d95cb
Summary:
For RNN attention, we should not include the invalid parts of the encoder output (based on encoder_lengths) in the computation. This diff accomplishes that by forcing logits for those positions to be negative infinity.
Note that the this step can be bypassed by passing encoder_lengths=None, which is what we do for beam search, thus incurring no extra overhead for inference.
Reviewed By: jamesr66a
Differential Revision: D5402547
fbshipit-source-id: 1863d6050b5129e4df829c6357f0aa9ded0715dc
Summary: Adding a test to check computational integrity of networks constructed with AttentionCell using UnrolledCell.
Reviewed By: salexspb
Differential Revision: D5306915
fbshipit-source-id: 02acfd1011f7d3ee5fac21cc2778c4a486190c43
Summary: I accidentaly noticed that we were calling the non-CUDNN version of Transpose with attention, and it is super slow. This broke when rnn_cell was changed to use ModelHelper instead of CNNModelHelper in D5062963, but calls to transpose were not "brewed".
Reviewed By: jamesr66a
Differential Revision: D5264248
fbshipit-source-id: b61494ae210f34597245f1195d20547f5b5cd8b5
Summary:
This is going to unblock Nvidia in their work on adding fp16
support to Caffe2. I discussed this with kennyhorror before to make
sure this fits into his work on parameter sharing.
Reviewed By: kennyhorror
Differential Revision: D5127797
fbshipit-source-id: 4db155d320b1862570c23b77c4252bdacbf2296f
Summary:
Update rnn_cell.py and char_rnn.py example with new `brew` model.
- Deprecated CNNModelHelper
- replace all helper functions with brew helper functions
- Use `model.net.<SingleOp>` format to create bare bone Operator for better clarity.
Reviewed By: salexspb
Differential Revision: D5062963
fbshipit-source-id: 254f7b9059a29621027d2b09e932f3f81db2e0ce
Summary: We can avoid this extra Reshape.
Reviewed By: jamesr66a
Differential Revision: D5032874
fbshipit-source-id: 92bd568bc6bec53d7f81a64cfa96d2c610823f8c
Summary: decoder_hidden_encoder_outputs_sum_tmp is tiny after D5010109, no need to recompute it.
Reviewed By: akyrola
Differential Revision: D5014335
fbshipit-source-id: cc9e8f91372889d10bd99c79366018cb3943a435
Summary: Fixes unit test test_seq2seq_caffe2_model_cnn_one_stack_encoder, broken by D4905003. (Also some commas.)
Differential Revision: D4920699
fbshipit-source-id: 2fe501095e3e26a475d666afcae8e48c953f2eef
Summary:
This diff adds an option to recurrent_net to define some cell blobs to be recomputed on backward step, and thus they don't need to be stored in the step workspace. This is done by modifying the backward step to automatically include all operators that are needed to produce the output that is to be recomputed, and by storing those blobs in a shared workspace. To enable the shared workspace, i had to modify the stepworkspaces blob to also store a forward shared workspace. Making it a class field won't work since the lifecycle of the blob does not match the lifecycle of the operator.
For basic LSTM, the performance hit is quite modest (about 15% with one setting, but your mileage might vary. For Attention models, I am sure this is beneficial as computing the attention blobs is not expensive.
For basic LSTM, the memory saving is wonderful: each forward workspace only has 4 bytes (for timestep).
I also modified the neural_mt LSTM Cells, but there is no test available, so I am not 100% sure I did it correctly. Please have a look.
Added options to LSTM, MILSTM and LSTMAttention to enable memory mode.
Reviewed By: urikz
Differential Revision: D4853890
fbshipit-source-id: d8d0e0e75a5330d174fbfa39b96d8e4e8c446baa
Summary: The caffe2 implementation of bare Softmax() has a race condition that wipes out the numerical stability trick. Use the CUDNN implementation instead
Reviewed By: urikz
Differential Revision: D4831298
fbshipit-source-id: d11b1de700e3954629e7ed43225a2416c27b3252
Summary:
Uses the cudnnTransformTensor function. It works by shuffling the strides according to the transpose axis. Significant speedup over current GPU version .
+ moves the transpose test under utility_ops, because hypothesis_test is too big
Reviewed By: jamesr66a
Differential Revision: D4810993
fbshipit-source-id: 82577c4ced1389e70bd5992820ae4d8297a3817f
Summary: Apart from copying gradient blobs for inputs with initial_cell_input, we needed to perform a similar operation for external parameters used by the step net
Reviewed By: salexspb
Differential Revision: D4752259
fbshipit-source-id: 13ee48cf583ed86221a4cc1cc9f57f5c3a7d2450
Summary: D4734505 part 2. Remove more instances of the batch_size parameter
Reviewed By: urikz
Differential Revision: D4736906
fbshipit-source-id: fc9d374e9308017d61c427890364c5ab9cec2edf
Summary: Reshape based on tensor shapes in the graph rather than based on a passed-in batch_size parameter
Reviewed By: urikz
Differential Revision: D4734505
fbshipit-source-id: d9c23d85be84f61124106e752ef2b4f6945e2a07
Summary: Reshape based on tensor shapes in the graph rather than based on a passed-in batch_size parameter
Reviewed By: urikz
Differential Revision: D4702086
fbshipit-source-id: c4c1d8425cd36c1e86695918eaba2667c27e9601
Summary: Super rough implementation of recurrent attention. Planning to factor out the common code between the two functions as well as train and eval. I want to get this out and get eyes on it sooner rather than later
Differential Revision: D4647837
fbshipit-source-id: 54bc4e8ed0df6f04c86c425926decbe89f73b068
Summary:
Implementation of ##LSTMWithAttention##
Still TBD:
1. There are problems with back propagation, because gradient is not implemented for ops with broadcasting
2. I need to make initial_recurrent_state to be of shape [dim] rather than [1, batch_size, dim], so one doesn't need to provide batch_size to LSTMWithAttention
Differential Revision: D4298735
fbshipit-source-id: 8903fcff4d6a66647ee6d45a6ef28803fc3091e5
