Skip to content

[ONNX] support onnx export rnn with batch_first=True #19766

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
BowenBao wants to merge 1 commit into from
Closed

[ONNX] support onnx export rnn with batch_first=True #19766

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 10 additions & 12 deletions test/onnx/test_pytorch_onnx_caffe2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -288,11 +288,13 @@ def _dispatch_rnn_test(self, name, *args, **kwargs):

def _elman_rnn_test(self, layers, nonlinearity, bidirectional,
initial_state, packed_sequence, dropout):
batch_first = True if packed_sequence == 2 else False
model = nn.RNN(RNN_INPUT_SIZE, RNN_HIDDEN_SIZE,
layers,
nonlinearity=nonlinearity,
bidirectional=bidirectional,
dropout=dropout)
dropout=dropout,
batch_first=batch_first)

if packed_sequence == 1:
model = RnnModelWithPackedSequence(model, False)
Expand All @@ -303,9 +305,7 @@ def make_input(batch_size):
seq_lengths = np.random.randint(1, RNN_SEQUENCE_LENGTH + 1, size=batch_size)
seq_lengths = list(reversed(sorted(map(int, seq_lengths))))
inputs = [torch.randn(l, RNN_INPUT_SIZE) for l in seq_lengths]
inputs = rnn_utils.pad_sequence(inputs)
if packed_sequence == 2:
inputs = inputs.transpose(0, 1)
inputs = rnn_utils.pad_sequence(inputs, batch_first=batch_first)
inputs = [inputs]

directions = 2 if bidirectional else 1
Expand All @@ -331,9 +331,10 @@ def make_input(batch_size):

def _lstm_test(self, layers, bidirectional, initial_state,
packed_sequence, dropout):
batch_first = True if packed_sequence == 2 else False
model = LstmFlatteningResult(
RNN_INPUT_SIZE, RNN_HIDDEN_SIZE, layers,
bidirectional=bidirectional, dropout=dropout)
bidirectional=bidirectional, dropout=dropout, batch_first=batch_first)
if packed_sequence == 1:
model = RnnModelWithPackedSequence(model, False)
if packed_sequence == 2:
Expand All @@ -343,9 +344,7 @@ def make_input(batch_size):
seq_lengths = np.random.randint(1, RNN_SEQUENCE_LENGTH + 1, size=batch_size)
seq_lengths = list(reversed(sorted(map(int, seq_lengths))))
inputs = [torch.randn(l, RNN_INPUT_SIZE) for l in seq_lengths]
inputs = rnn_utils.pad_sequence(inputs)
if packed_sequence == 2:
inputs = inputs.transpose(0, 1)
inputs = rnn_utils.pad_sequence(inputs, batch_first=batch_first)
inputs = [inputs]

directions = 2 if bidirectional else 1
Expand All @@ -372,8 +371,9 @@ def make_input(batch_size):

def _gru_test(self, layers, bidirectional, initial_state,
packed_sequence, dropout):
batch_first = True if packed_sequence == 2 else False
model = nn.GRU(RNN_INPUT_SIZE, RNN_HIDDEN_SIZE, layers,
bidirectional=bidirectional, dropout=dropout)
bidirectional=bidirectional, dropout=dropout, batch_first=batch_first)
if packed_sequence == 1:
model = RnnModelWithPackedSequence(model, False)
if packed_sequence == 2:
Expand All @@ -383,9 +383,7 @@ def make_input(batch_size):
seq_lengths = np.random.randint(1, RNN_SEQUENCE_LENGTH + 1, size=batch_size)
seq_lengths = list(reversed(sorted(map(int, seq_lengths))))
inputs = [torch.randn(l, RNN_INPUT_SIZE) for l in seq_lengths]
inputs = rnn_utils.pad_sequence(inputs)
if packed_sequence == 2:
inputs = inputs.transpose(0, 1)
inputs = rnn_utils.pad_sequence(inputs, batch_first=batch_first)
inputs = [inputs]

directions = 2 if bidirectional else 1
Expand Down
6 changes: 5 additions & 1 deletion torch/onnx/symbolic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1470,7 +1470,8 @@ def _generic_rnn(g, variant, input, initial_states, all_weights, has_biases,
assert len(all_weights) == num_layers * weights_per_layer * (1 + bidirectional)
layer_weights = [all_weights[i:i + weights_per_layer] for i in range(0, len(all_weights), weights_per_layer)]
if batch_first:
return _unimplemented("RNN/GRU/LSTM", "batch_first")
# batch, seq, feat -> seq, batch, feat
input = g.op('Transpose', input, perm_i=[1, 0, 2])
if dropout and train:
return _unimplemented("RNN/GRU/LSTM", "dropout in training mode")

Expand Down Expand Up @@ -1572,6 +1573,9 @@ def retrieve_state(x, start, end):
h_outs.append(h_out)
if variant == 'LSTM':
c_outs.append(c_out)
if batch_first:
# seq, batch, num_directions * hidden_size -> batch, seq, num_directions * hidden_size
prev_output = g.op('Transpose', prev_output, perm_i=[1, 0, 2])
h_outs = h_out if num_layers == 1 else g.op('Concat', *h_outs, axis_i=0)
if variant == 'RNN' or variant == 'GRU':
return prev_output, h_outs
Expand Down