search

lucidrains / DALLE2-pytorch

Implementation of DALL-E 2, OpenAI's updated text-to-image synthesis neural network, in Pytorch
MIT License
10.97k stars 1.07k forks source link

Why parameter [timesteps] is fixed? #287

Open 4daJKong opened 1 year ago

4daJKong commented 1 year ago

I found that in line 2308, the creating of NoiseScheduler layer depends on timesteps in dalle2_pytorch.py. Did it means I cannot change (increase or decrese) the value of timesteps if I have loaded a pretrained model when testing? 'sampling loot time steps =1000' in decoder is too long for me, and I haven't other parameter, like sample_timesteps in prior, that can control this process.

funmooya commented 1 year ago

I meet the same problem, do you find the answer?

4daJKong commented 1 year ago

I meet the same problem, do you find the answer?

Hi, adding "default_sample_timesteps": [100], after "default_cond_scale": [1.7], in gradio.example.json can adjust the sample_timesteps of decoder.

funmooya commented 1 year ago

I meet the same problem, do you find the answer?

Hi, adding "default_sample_timesteps": [100], after "default_cond_scale": [1.7], in gradio.example.json can adjust the sample_timesteps of decoder.

Thank you very very much.

moonbaby9 commented 7 months ago

I meet the same problem, do you find the answer?

Hi, adding "default_sample_timesteps": [100], after "default_cond_scale": [1.7], in gradio.example.json can adjust the sample_timesteps of decoder.

Hello! I add "default_sample_timesteps": [100], after "default_cond_scale": [1.7], in gradio.example.json, but meet the error below: File "pydantic/main.py", line 341, in pydantic.main.BaseModel.init pydantic.error_wrappers.ValidationError: 3 validation errors for DecoderConfig sample_timesteps value is not a valid integer (type=type_error.integer) sample_timesteps -> 1 none is not an allowed value (type=type_error.none.not_allowed) sample_timesteps wrong tuple length 2, expected 1 (type=value_error.tuple.length; actual_length=2; expected_length=1)

Do you know how to deal with this bug? Thank you very much!