A question related to batch size and training speed

[paidaxinbao]

Hi!

The code in this repository has helped me a lot!

I found that as the batch size increases, the training time increases dramatically. When I set the batch size to 4 (the dataset has 25k images) the training time is about 2 days, but when the batch size is set to 128, the training time increases to 800 hours!

I don't know much about this.

My training configuration is as follows:
model = Unet(
dim=64,
out_dim=1,
dim_mults=(1, 2, 4, 8),
channels=2
)

diffusion = GaussianDiffusion(
model,
image_size=128,
timesteps=1000, # number of steps
sampling_timesteps=250,
# number of sampling timesteps (using ddim for faster inference [see citation for ddim paper])
)

trainer = Trainer(
diffusion,
'/home/pxy/ML_work/train_picset/',
train_batch_size=4,
train_lr=8e-5,
train_num_steps=700000, # total training steps
gradient_accumulate_every=4, # gradient accumulation steps
ema_decay=0.995, # exponential moving average decay
amp=True, # turn on mixed precision
calculate_fid = False
)

trainer.train()

[gggah]

Hi!

The code in this repository has helped me a lot!

I found that as the batch size increases, the training time increases dramatically. When I set the batch size to 4 (the dataset has 25k images) the training time is about 2 days, but when the batch size is set to 128, the training time increases to 800 hours!

I don't know much about this.

My training configuration is as follows: model = Unet( dim=64, out_dim=1, dim_mults=(1, 2, 4, 8), channels=2 )

diffusion = GaussianDiffusion( model, image_size=128, timesteps=1000, # number of steps sampling_timesteps=250, # number of sampling timesteps (using ddim for faster inference [see citation for ddim paper]) )

trainer = Trainer( diffusion, '/home/pxy/ML_work/train_picset/', train_batch_size=4, train_lr=8e-5, train_num_steps=700000, # total training steps gradient_accumulate_every=4, # gradient accumulation steps ema_decay=0.995, # exponential moving average decay amp=True, # turn on mixed precision calculate_fid = False )

trainer.train()

May I ask what is your data format, why can't I recognize it, and the error should be greater than 100, while mine is 1200 pictures

[paidaxinbao]

Hi!
The code in this repository has helped me a lot!
I found that as the batch size increases, the training time increases dramatically. When I set the batch size to 4 (the dataset has 25k images) the training time is about 2 days, but when the batch size is set to 128, the training time increases to 800 hours!
I don't know much about this.
My training configuration is as follows: model = Unet( dim=64, out_dim=1, dim_mults=(1, 2, 4, 8), channels=2 )
diffusion = GaussianDiffusion( model, image_size=128, timesteps=1000, # number of steps sampling_timesteps=250, # number of sampling timesteps (using ddim for faster inference [see citation for ddim paper]) )
trainer = Trainer( diffusion, '/home/pxy/ML_work/train_picset/', train_batch_size=4, train_lr=8e-5, train_num_steps=700000, # total training steps gradient_accumulate_every=4, # gradient accumulation steps ema_decay=0.995, # exponential moving average decay amp=True, # turn on mixed precision calculate_fid = False )
trainer.train()

May I ask what is your data format, why can't I recognize it, and the error should be greater than 100, while mine is 1200 pictures

Hi, my data is a grayscale map and then I used the strategy in SR3 to use the condition and concatenate the original grayscale image as an input to Unet. I didn't understand what you mean by error, do you mean loss?