the grad of cpu ,cuda are not same

[lcylcy]

import torch
import numpy as np
from warprnnt_pytorch import RNNTLoss

acts = np.random.rand(2,2,3,5)
labels =  np.array([[1, 2],[2,2]])
act_length = np.array([2,2])
label_length = np.array([2,2])
blank = 1
reduction = 'mean'

print('----------------------------------------')
rnnt_loss = RNNTLoss(blank,reduction)
acts_t = torch.tensor(acts,dtype=torch.float)
acts_t.requires_grad = True
labels_t = torch.tensor(labels,dtype=torch.int)
act_length_t = torch.tensor(act_length,dtype=torch.int)
label_length_t = torch.tensor(label_length,dtype=torch.int)
loss = rnnt_loss(acts_t,labels_t,act_length_t,label_length_t)
loss.backward()
print(loss)
print(acts_t.grad)
print('-----------------------------------------')
rnnt_loss = RNNTLoss(blank,reduction).cuda()
acts_t = torch.tensor(acts,dtype=torch.float).cuda()
acts_t.requires_grad = True
labels_t = torch.tensor(labels,dtype=torch.int).cuda()
act_length_t = torch.tensor(act_length,dtype=torch.int).cuda()
label_length_t = torch.tensor(label_length,dtype=torch.int).cuda()
loss = rnnt_loss(acts_t,labels_t,act_length_t,label_length_t)
loss.backward()
print(loss)
print(acts_t.grad)
print('-----------------------------------------')

----------------------------------------
tensor([4.6379], grad_fn=<_RNNTBackward>)
tensor([[[[ 0.0550, -0.2406,  0.0716,  0.0589,  0.0550],
          [ 0.0492, -0.1081, -0.0603,  0.0419,  0.0774],
          [ 0.0290, -0.1254,  0.0351,  0.0291,  0.0322]],

         [[ 0.0346, -0.1494,  0.0226,  0.0539,  0.0383],
          [ 0.0726,  0.0889, -0.2694,  0.0452,  0.0627],
          [ 0.0730, -0.4038,  0.1470,  0.0579,  0.1259]]],

        [[[ 0.0749, -0.0353, -0.2435,  0.1090,  0.0948],
          [ 0.0529, -0.1224, -0.0426,  0.0533,  0.0587],
          [ 0.0290, -0.1383,  0.0383,  0.0461,  0.0249]],

         [[ 0.0163,  0.0226, -0.0751,  0.0122,  0.0240],
          [ 0.0655,  0.0462, -0.2269,  0.0494,  0.0658],
          [ 0.0755, -0.3531,  0.1155,  0.0999,  0.0623]]]])
-----------------------------------------
tensor([4.6379], device='cuda:0', grad_fn=<_RNNTBackward>)
tensor([[[[ 0.0885, -0.3869,  0.1152,  0.0947,  0.0885],
          [ 0.0492, -0.1081, -0.0603,  0.0419,  0.0774],
          [ 0.0290, -0.1254,  0.0351,  0.0291,  0.0322]],

         [[ 0.0346, -0.1494,  0.0226,  0.0539,  0.0383],
          [ 0.0726,  0.0889, -0.2694,  0.0452,  0.0627],
          [ 0.0730, -0.4038,  0.1470,  0.0579,  0.1259]]],

        [[[ 0.0749, -0.0353, -0.2435,  0.1090,  0.0948],
          [ 0.0529, -0.1224, -0.0426,  0.0533,  0.0587],
          [ 0.0290, -0.1383,  0.0383,  0.0461,  0.0249]],

         [[ 0.0163,  0.0226, -0.0751,  0.0122,  0.0240],
          [ 0.0655,  0.0462, -0.2269,  0.0494,  0.0658],
          [ 0.0755, -0.3531,  0.1155,  0.0999,  0.0623]]]], device='cuda:0')
-----------------------------------------

[csukuangfj]

Please have a try with https://github.com/csukuangfj/optimized_transducer

It has a consistent gradient for CPU and CUDA.