I made a speech recognition Transformer.
But, After the first iteration, the loss becomes NaN.
Another strange thing, however, is that loss is calculated normally in CPU mode.
I will attach the address of GitHub with my code below, my test code and test result.

Transformer source code: KoSpeech/model.py at latest · sooftware/KoSpeech · GitHub

My Test Code
# Copyright (c) 2020, Soohwan Kim. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
import torch.nn as nn
from kospeech.model_builder import build_transformer
batch_size = 4
seq_length = 200
target_length = 10
input_size = 80
cuda = torch.cuda.is_available()
device = torch.device('cuda' if cuda else 'cpu')
transformer = build_transformer(
num_classes=10,
d_model=16,
d_ff=32,
num_heads=2,
input_dim=input_size,
num_encoder_layers=3,
num_decoder_layers=2,
extractor='vgg',
dropout_p=0.1,
device=device,
pad_id=0,
sos_id=1,
eos_id=2,
joint_ctc_attention=False,
max_length=10,
)
criterion = nn.CrossEntropyLoss(ignore_index=0, reduction='mean')
optimizer = torch.optim.Adam(transformer.parameters(), lr=1e04)
for i in range(10):
inputs = torch.FloatTensor(batch_size, seq_length, input_size).to(device)
input_lengths = torch.LongTensor([seq_length, seq_length  10, seq_length  20, seq_length  30])
targets = torch.LongTensor([[1, 3, 3, 3, 3, 3, 4, 5, 6, 2],
[1, 3, 3, 3, 3, 3, 4, 5, 2, 0],
[1, 3, 3, 3, 3, 3, 4, 2, 0, 0],
[1, 3, 3, 3, 3, 3, 4, 2, 0, 0]]).to(device)
outputs, _, _ = transformer(inputs, input_lengths, targets)
loss = criterion(outputs.contiguous().view(1, outputs.size(1)), targets[:, 1:].contiguous().view(1))
loss.backward()
optimizer.step()
print(loss)
 Result in CUDA mode
tensor(10.6171, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
tensor(nan, device='cuda:0', grad_fn=<MeanBackward0>)
 Result in CPU mode
tensor(12.8601, grad_fn=<MeanBackward0>)
tensor(12.5969, grad_fn=<MeanBackward0>)
tensor(12.4775, grad_fn=<MeanBackward0>)
tensor(10.8609, grad_fn=<MeanBackward0>)
tensor(11.9296, grad_fn=<MeanBackward0>)
tensor(9.9925, grad_fn=<MeanBackward0>)
tensor(11.5225, grad_fn=<MeanBackward0>)
tensor(9.3546, grad_fn=<MeanBackward0>)
tensor(11.4693, grad_fn=<MeanBackward0>)
tensor(9.2175, grad_fn=<MeanBackward0>)
Please help me!