Skip to content

Updating BERT/TF, Transformer-XL and NCF/PyT #392

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
nvpstr merged 3 commits into from
Feb 6, 2020
Merged

Updating BERT/TF, Transformer-XL and NCF/PyT #392

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
cef4bab
Adding BERT/TF Inference with TensorRT; minor README fixes
nvpstr Feb 5, 2020
1def26d
[NCF/PyT] Adding new logging
nvpstr Feb 5, 2020
a38deff
[Transformer-XL/PyT] Large model support; multi-node training; infere…
nvpstr Feb 5, 2020
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
4 changes: 4 additions & 0 deletions PyTorch/LanguageModeling/Transformer-XL/.gitignore
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,3 +1,7 @@
**/.DS_Store
__pycache__/
data/
results/
*.out
*.log
*.json
991 changes: 775 additions & 216 deletions PyTorch/LanguageModeling/Transformer-XL/README.md
View file
Open in desktop

Large diffs are not rendered by default.

3 changes: 3 additions & 0 deletions PyTorch/LanguageModeling/Transformer-XL/pytorch/.dockerignore
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,2 +1,5 @@
LM-TFM*
internal/result*
*.out
*.log
*.json
3 changes: 2 additions & 1 deletion PyTorch/LanguageModeling/Transformer-XL/pytorch/Dockerfile
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,7 +12,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.

ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:19.09-py3
ARG FROM_IMAGE_NAME=nvcr.io/nvidia/pytorch:19.11-py3
FROM ${FROM_IMAGE_NAME}

ENV LANG C.UTF-8
Expand All @@ -26,5 +26,6 @@ WORKDIR /workspace/transformer-xl/pytorch

COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt
RUN pip install --no-cache-dir git+https://github.com/NVIDIA/dllogger.git#egg=dllogger

ADD . /workspace/transformer-xl/pytorch
30 changes: 22 additions & 8 deletions PyTorch/LanguageModeling/Transformer-XL/pytorch/data_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,32 +27,41 @@


class LMOrderedIterator(object):
def __init__(self, data, bsz, bptt, device='cpu', ext_len=None):
def __init__(self, data, bsz, bptt, device='cpu', mem_len=None, ext_len=None, warmup=True):
"""
data -- LongTensor -- the LongTensor is strictly ordered
"""
self.bsz = bsz
self.bptt = bptt
self.ext_len = ext_len if ext_len is not None else 0
self.mem_len = mem_len
self.warmup = warmup

self.device = device

# Work out how cleanly we can divide the dataset into bsz parts.
self.n_step = data.size(0) // bsz
n_step = data.size(0) // bsz

# Trim off any extra elements that wouldn't cleanly fit (remainders).
data = data.narrow(0, 0, self.n_step * bsz)
data = data[:n_step * bsz]

# Evenly divide the data across the bsz batches.
self.data = data.view(bsz, -1).t().contiguous()

if mem_len and warmup:
self.warmup_batches = (mem_len + bptt - 1) // bptt
self.warmup_elems = self.warmup_batches * bptt

warmup_data = self.data.roll((self.warmup_elems, 1), (0, 1))[:self.warmup_elems]
self.data = torch.cat((warmup_data, self.data))

# Partition data for DistributedDataParallel
world_size = utils.distributed.get_world_size()
rank = utils.distributed.get_rank()
self.data = self.data.chunk(world_size, dim=1)[rank].to(device)
self.data = self.data.chunk(world_size, dim=1)[rank]

# Number of mini-batches
self.n_batch = (self.n_step + self.bptt - 1) // self.bptt
self.n_batch = (self.data.size(0) + self.bptt - 1) // self.bptt

def roll(self):
for i in range(self.data.size(1)):
Expand All @@ -70,10 +79,15 @@ def get_batch(self, i, bptt=None):
end_idx = i + seq_len
beg_idx = max(0, i - self.ext_len)

data = self.data[beg_idx:end_idx]
target = self.data[i+1:i+1+seq_len]
data = self.data[beg_idx:end_idx].to(self.device)
target = self.data[i+1:i+1+seq_len].to(self.device)

if self.mem_len and self.warmup:
warm = i >= self.warmup_elems
else:
warm = True

return data, target, seq_len
return data, target, seq_len, warm

def get_fixlen_iter(self, start=0):
for i in range(start, self.data.size(0) - 1, self.bptt):
Expand Down
Loading