Maze2d Index Question

[erdisayar]

In this line;
https://github.com/Weixy21/SafeDiffuser/blob/maze2d/diffuser/models/diffusion.py#L1048

you are using x[:,2:3] and x has the shape (batch,horizion, (action_dim + observation_dim))

What is the purpose of using 2nd horizion index ?

Shouldn't you use x[:,:,0:1] which is the x position of the x value ?

[Weixy21]

Here x[:, 0:2] are controls along x and y, and x[:,2:4] are the x, y value.

[erdisayar]

Thanks for your quick answer.

But according to this line https://github.com/Weixy21/SafeDiffuser/blob/maze2d/diffuser/models/diffusion.py#L1140

x[:, :, 0:2] are the controls along x and y
and
x[:, :, 2:4] are the x and y position

because x has 3 dimensional vector and its shape is [batch,horizion,transition]

Am I wrong?

[Weixy21]

I mainly checked this during inference, and it is a 2 dimensional vector during inference.

You may check the dimension:

SafeDiffuser/diffuser/models/diffusion.py

Line 1038 in 3a0ede1

xp1 = model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise

During training, please use x = xp1 only, as the current method is for inference. I will fix this

[erdisayar]

I am using this commit hash 3a0ede1049955dcb26ac5f44e340195501728979 as you referenced in the previous comment.
I checked that the dimenison of the xp1 and it is still 3 dimensional

xp1.shape
torch.Size([10, 384, 6])

For more detail,

I am using this command line for the training;

python scripts/train.py --config config.maze2d --dataset maze2d-large-v1

And the inference function is called from this function render_samples(...)

SafeDiffuser/diffuser/utils/training.py

Line 219 in 3a0ede1

samples = self.ema_model.conditional_sample(conditions, return_diffusion = False)

Or what is the strategy to reduce it to 2d case ? just take the first horizon ?

Thanks

[Weixy21]

During training it is 3 dimension as the batch is usually greater than 1, during inference (what I mean here is testing), the batch is only 1 and thus this dimension is omitted.

[Weixy21]

Please use x = xp1 during training, the proposed method is mainly used for inference (testing) now. After you are done with training, you can test all these.

[erdisayar]

I see
Then the problem here is the size of the conditional paremeter ; (as far as I figure out)

what I mean by that;
Here the input variable conditions is a dictionary with 2 keys ( the first timestep 0 and last timestep 383)

In other words, the conditions has this dimension [2,10,4]

SafeDiffuser/diffuser/utils/training.py

Line 219 in 3a0ede1

samples = self.ema_model.conditional_sample(conditions, return_diffusion = False)

then if you go to the line below

SafeDiffuser/diffuser/models/diffusion.py

Line 1138 in 3a0ede1

batch_size = len(cond[0])

batch_size = 10 which would be the conditional second dimension. Because of this reason, xp1 has 3 dimension (batch_size >1)

Should I reduce the conditionals dimension into 1 ?

[Weixy21]

No, you do not have to, just setting x = xp1 in line 1043 and comment out line 1063 of diffusion.py should allow you to train the model.

[erdisayar]

Yes, I comment out the line 1043 and you can check the video;
It has 3 dimension.

2025-04-04.17-59-43.mp4

[Weixy21]

It should be 3 dimension during training. There is no problem with that

[erdisayar]

do you mean that If I use the following command;

python scripts/train.py --config config.maze2d --dataset maze2d-large-v1

these parts should be commented

In order to use Control Barier function case, ( your implementation),

I need to remove the comments from this part and run the following command line?

python scripts/plan_maze2d.py --config config.maze2d --dataset maze2d-large-v1

[Weixy21]

Yes, for now just use CBFs during testing: python scripts/plan_maze2d.py --config config.maze2d --dataset maze2d-large-v1

[erdisayar]

I see

In other words, it basically means that

this command python scripts/train.py --config config.maze2d --dataset maze2d-large-v1 is basically using the original diffuser model. (during training, there is also inference happening)

But to plan the trajectory, I need to commented out your scripts and run the SafeDiffuser ?

But in your scripts there is no data regarding to maze2d-large-v1 model.( also there is no data in the Google Drive)

I think I can use diffuser to train maze2d-large-v1 and save data in SafeDiffuser and then run python scripts/plan_maze2d.py --config config.maze2d --dataset maze2d-large-v1

Thanks for your help

[Weixy21]

Yes, the training is using the original diffuser. During planning, you may use SafeDiffuser.

The data will automatically download when you run the training

[erdisayar]

I managed to make it work, but the result I'm getting is violating the safe space.

There are many options that can be commented out. I tried the following ones — they are currently uncommented and actively used:

        x = xp1
        xr = 2*1/(self.norm_maxs[1] - self.norm_mins[1])
        yr = 2*1/(self.norm_maxs[0] - self.norm_mins[0])
        off_x = 2*(5.8-0.5 - self.norm_mins[1])/(self.norm_maxs[1] - self.norm_mins[1]) - 1
        off_y = 2*(5-0.5 - self.norm_mins[0])/(self.norm_maxs[0] - self.norm_mins[0]) - 1
        b = ((x[:,2:3] - off_y)/yr)**2 + ((x[:,3:4] - off_x)/xr)**2 - 1
        self.safe1 = torch.min(b[:,0])
        xr = 2*1/(self.norm_maxs[1] - self.norm_mins[1])
        yr = 2*1/(self.norm_maxs[0] - self.norm_mins[0])
        off_x = 2*(5.3-0.5 - self.norm_mins[1])/(self.norm_maxs[1] - self.norm_mins[1]) - 1
        off_y = 2*(2-0.5 - self.norm_mins[0])/(self.norm_maxs[0] - self.norm_mins[0]) - 1
        b = ((x[:,2:3] - off_y)/yr)**4 + ((x[:,3:4] - off_x)/xr)**4 - 1
        self.safe2 = torch.min(b[:,0])

x = self.invariance_time(x, xp1, t) 

return xp1

Do you have an example showing which parts of the code I should activate? It would be helpful to understand which specific lines or options are necessary to avoid the safe space violation.