Skip to content

Cherry pick/fix transformer #16620

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

Merged
JiabinYang merged 3 commits into from
Apr 8, 2019
Merged

Cherry pick/fix transformer #16620

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6080c9b
Imperative deep-first backward process (#16605)
velconia Apr 2, 2019
9727eab
test=develop, cherry-pick fix for transformer in dygraph
JiabinYang Apr 2, 2019
7e39250
test=develop, fix transformer in dygraph
JiabinYang Apr 3, 2019
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
8 changes: 4 additions & 4 deletions paddle/fluid/imperative/layer.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -122,14 +122,14 @@ class Autograd {
std::map<std::string, std::vector<VarBase*>> input_grads =
ready_op->ApplyGrad();

for (auto it : input_grads) {
const std::vector<VarBase*>& ingrads = it.second;
for (auto it = input_grads.rbegin(); it != input_grads.rend(); ++it) {
const std::vector<VarBase*>& ingrads = it->second;
for (size_t i = 0; i < ingrads.size(); ++i) {
if (!ingrads[i]) continue;
if (ready_op->input_vars_[it.first][i]->IsStopGradient()) {
if (ready_op->input_vars_[it->first][i]->IsStopGradient()) {
continue;
}
OpBase* pre_op = ready_op->pre_ops_[it.first][i];
OpBase* pre_op = ready_op->pre_ops_[it->first][i];
if (!pre_op) continue;

dep_counts[pre_op] -= 1;
Expand Down
3 changes: 2 additions & 1 deletion python/paddle/fluid/framework.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -493,7 +493,8 @@ def _backward(self):
self._ivar._run_backward()

def _gradient(self):
return np.array(self._ivar._grad_value())
new_ivar = self._ivar._grad_ivar()._copy_to(core.CPUPlace(), True)
return np.array(new_ivar.value().get_tensor())

def _clear_gradient(self):
self._ivar._clear_gradient()
Expand Down
42 changes: 12 additions & 30 deletions python/paddle/fluid/tests/unittests/test_imperative_transformer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -302,8 +302,11 @@ def make_all_inputs(input_fields):
# if we run sync mode
sync = False

# how many batches we use
batch_num = 2
if not core.is_compiled_with_cuda():
# how many batches we use
batch_num = 50
else:
batch_num = 5

np.random.seed = 1
src_word_np = np.random.randint(
Expand Down Expand Up @@ -335,24 +338,6 @@ def make_all_inputs(input_fields):
dtype='int64')
lbl_weight_np = np.random.randn(batch_size * seq_len, 1).astype('float32')

# np.random.seed = 1
# src_word_np = np.arange(0, 10).reshape([batch_size, seq_len, 1]).astype('int64')
# src_pos_np = np.random.randint(
# 1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
# src_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
# seq_len, seq_len).astype('float32')
#
# trg_word_np = np.arange(0, 10).reshape([batch_size, seq_len, 1]).astype('int64')
# trg_pos_np = np.random.randint(
# 1, seq_len, size=(batch_size, seq_len, 1), dtype='int64')
# trg_slf_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
# seq_len, seq_len).astype('float32')
# trg_src_attn_bias_np = np.random.randn(batch_size, ModelHyperParams.n_head,
# seq_len, seq_len).astype('float32')
#
# lbl_word_np = np.arange(0, 10).reshape([batch_size * seq_len, 1]).astype('int64')
# lbl_weight_np = np.random.randn(batch_size * seq_len, 1).astype('float32')
#
pos_inp1 = position_encoding_init(ModelHyperParams.max_length,
ModelHyperParams.d_model)
pos_inp2 = position_encoding_init(ModelHyperParams.max_length,
Expand Down Expand Up @@ -739,7 +724,7 @@ def forward(self, dec_input, enc_output, slf_attn_bias, dec_enc_attn_bias):
enc_attn_output_pp = self._multihead_attention_layer2(
pre_process_rlt2, enc_output, enc_output, dec_enc_attn_bias)
enc_attn_output = self._post_process_layer2(
slf_attn_output, enc_attn_output_pp, self._postprocess_cmd,
slf_attn_output_pp, enc_attn_output_pp, self._postprocess_cmd,
self._prepostprcess_dropout)
pre_process_rlt3 = self._pre_process_layer3(None, enc_attn_output,
self._preprocess_cmd,
Expand Down Expand Up @@ -1076,20 +1061,17 @@ def test_transformer_float32(self):
4]] = out[k]

self.assertTrue(
np.allclose(static_avg_cost_value, dy_avg_cost._numpy()))
np.array_equal(static_avg_cost_value, dy_avg_cost._numpy()))
self.assertTrue(
np.allclose(static_sum_cost_value, dy_sum_cost._numpy()))
np.array_equal(static_sum_cost_value, dy_sum_cost._numpy()))
self.assertTrue(
np.allclose(
static_predict_value, dy_predict._numpy(), atol=1e-5))
np.array_equal(static_predict_value, dy_predict._numpy()))
self.assertTrue(
np.allclose(static_token_num_value, dy_token_num._numpy()))
np.array_equal(static_token_num_value, dy_token_num._numpy()))
for key, value in six.iteritems(static_param_init):
self.assertTrue(np.allclose(value, dy_param_init[key]))
self.assertTrue(np.array_equal(value, dy_param_init[key]))
for key, value in six.iteritems(static_param_updated):
self.assertTrue(
np.allclose(
value, dy_param_updated[key], atol=1e-4))
self.assertTrue(np.array_equal(value, dy_param_updated[key]))


if __name__ == '__main__':
Expand Down