def run_debug_test(self,
model,
train,
batch_size,
state_dict=None,
input=None,
use_gpu=True):
"""
# TODO: remove this from the final release version
This test is for our debugging only for the case where
embed_params=False
"""
model.train(train)
if state_dict is not None:
model.load_state_dict(state_dict)
# Either user specified input or random (deterministic) input
if input is None:
input = Variable(torch.randn(batch_size, 3, 224, 224),
requires_grad=True)
if use_gpu:
model, input = self.convert_cuda(model, input)
onnxir, torch_out = do_export(model,
input,
export_params=self.embed_params,
verbose=False)
if isinstance(torch_out, torch.autograd.Variable):
torch_out = (torch_out, )
caffe2_out = run_embed_params(onnxir, model, input, state_dict,
use_gpu)
for i, (x, y) in enumerate(zip(torch_out, caffe2_out)):
np.testing.assert_almost_equal(x.data.cpu().numpy(), y, decimal=3)
def run_debug_test(self, model, train, batch_size, state_dict=None,
input=None, use_gpu=True):
"""
# TODO: remove this from the final release version
This test is for our debugging only for the case where
embed_params=False
"""
model.train(train)
if state_dict is not None:
model.load_state_dict(state_dict)
# Either user specified input or random (deterministic) input
if input is None:
input = Variable(torch.randn(batch_size, 3, 224, 224),
requires_grad=True)
if use_gpu:
model, input = self.convert_cuda(model, input)
onnxir, torch_out = do_export(model, input, export_params=self.embed_params, verbose=False)
if isinstance(torch_out, torch.autograd.Variable):
torch_out = (torch_out,)
caffe2_out = run_embed_params(onnxir, model, input, state_dict, use_gpu)
for i, (x, y) in enumerate(zip(torch_out, caffe2_out)):
np.testing.assert_almost_equal(x.data.cpu().numpy(), y, decimal=3)
def _lstm_test(self, layers, bidirectional, initial_state, packed_sequence,
dropout):
model = LstmFlatteningResult(RNN_INPUT_SIZE,
RNN_HIDDEN_SIZE,
layers,
bidirectional=bidirectional,
dropout=dropout)
if packed_sequence == 1:
model = RnnModelWithPackedSequence(model, False)
if packed_sequence == 2:
model = RnnModelWithPackedSequence(model, True)
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 = [
Variable(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 = [inputs]
directions = 2 if bidirectional else 1
if initial_state:
h0 = Variable(
torch.randn(directions * layers, batch_size,
RNN_HIDDEN_SIZE))
c0 = Variable(
torch.randn(directions * layers, batch_size,
RNN_HIDDEN_SIZE))
inputs.append((h0, c0))
if packed_sequence != 0:
inputs.append(Variable(torch.IntTensor(seq_lengths)))
if len(inputs) == 1:
input = inputs[0]
else:
input = tuple(inputs)
return input
input = make_input(RNN_BATCH_SIZE)
self.run_model_test(model,
train=False,
batch_size=RNN_BATCH_SIZE,
input=input,
use_gpu=False)
# test that the model still runs with a different batch size
onnxir, _ = do_export(model, input)
other_input = make_input(RNN_BATCH_SIZE + 1)
_ = run_embed_params(onnxir, model, other_input, use_gpu=False)
def _elman_rnn_test(self, layers, nonlinearity, bidirectional,
initial_state, packed_sequence, dropout):
model = nn.RNN(RNN_INPUT_SIZE, RNN_HIDDEN_SIZE,
layers,
nonlinearity=nonlinearity,
bidirectional=bidirectional,
dropout=dropout)
if packed_sequence == 1:
model = RnnModelWithPackedSequence(model, False)
if packed_sequence == 2:
model = RnnModelWithPackedSequence(model, True)
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 = [Variable(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 = [inputs]
directions = 2 if bidirectional else 1
if initial_state:
h0 = Variable(torch.randn(directions * layers, batch_size, RNN_HIDDEN_SIZE))
inputs.append(h0)
if packed_sequence != 0:
inputs.append(Variable(torch.IntTensor(seq_lengths)))
if len(inputs) == 1:
input = inputs[0]
else:
input = tuple(inputs)
return input
input = make_input(RNN_BATCH_SIZE)
self.run_model_test(model, train=False, batch_size=RNN_BATCH_SIZE, input=input, use_gpu=False, atol=1e-7)
# test that the model still runs with a different batch size
onnxir, _ = do_export(model, input)
other_input = make_input(RNN_BATCH_SIZE + 1)
_ = run_embed_params(onnxir, model, other_input, use_gpu=False)
