get_gap_t produces DivisionByZero Error

[s-weissbach]

Hello,

I am trying to train DeepCAD-RT on a custom dataset. However, I always get a division by zero error in the get_gap_t function. It is working when changing self.gap_t = math.floor((self.whole_t - self.patch_t * 2) / (s_num - 1)) to self.gap_t = math.floor((self.whole_t - self.patch_t * 2) / (s_num )).

Noise image shape -----> (2100, 800, 800)

Thank you in advance

[cabooster]

Hi, I think you can use a larger overlap_factor or smaller train_datasets_size to avoid this problem

[s-weissbach]

Even when setting overlap_factor=0.9 and train_dataset_size=1 the error persists.

Training parameters -----> 
{'overlap_factor': 0.9, 'datasets_path': '/run/media/stephan/Elements/igludata_tiff', 'n_epochs': 5, 'fmap': 16, 'output_dir': './results', 'pth_dir': '/home/stephan/Desktop/deepcardrt', 'onnx_dir': './onnx', 'batch_size': 1, 'patch_t': 10, 'patch_x': 100, 'patch_y': 100, 'gap_y': 9, 'gap_x': 9, 'gap_t': 0, 'lr': 5e-05, 'b1': 0.5, 'b2': 0.999, 'GPU': '0', 'ngpu': 1, 'num_workers': 1, 'scale_factor': 1, 'train_datasets_size': 1, 'select_img_num': 1000000, 'test_datasize': 400, 'visualize_images_per_epoch': False, 'save_test_images_per_epoch': False, 'colab_display': False, 'result_display': ''}
Image list for training -----> 
Total stack number ----->  333
Noise image name ----->  30105_CS1-1_37a+47_iGlu_.tiff
Noise image shape ----->  (2100, 800, 800)
---------------------------------------------------------------------------
ZeroDivisionError                         Traceback (most recent call last)
Cell In[16], [line 2](vscode-notebook-cell:?execution_count=16&line=2)
      [1](vscode-notebook-cell:?execution_count=16&line=1) tc = training_class(train_dict)
----> [2](vscode-notebook-cell:?execution_count=16&line=2) tc.run()

File ~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:77, in training_class.run(self)
     [75](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:75) self.prepare_file()
     [76](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:76) # crop input tiff file into 3D patches
---> [77](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:77) self.train_preprocess_lessMemoryMulStacks()
     [78](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:78) # save some essential training parameters in para.yaml
     [79](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:79) self.save_yaml_train()

File ~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:189, in training_class.train_preprocess_lessMemoryMulStacks(self)
    [187](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:187) print('Noise image shape -----> ', noise_im.shape)
    [188](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:188) # Calculate real gap_t
--> [189](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:189) self.get_gap_t()
    [190](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:190) # No preprocessing
    [191](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:191) # noise_im = noise_im.astype(np.float32) / self.scale_factor
    [192](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:192) # Minus mean before training
    [193](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:193) noise_im = noise_im.astype(np.float32)/self.scale_factor

File ~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:155, in training_class.get_gap_t(self)
    [153](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:153) h_num = math.floor((self.whole_y - self.patch_y) / self.gap_y) + 1
    [154](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:154) s_num = math.ceil(self.train_datasets_size / w_num / h_num / self.stack_num)
--> [155](https://vscode-remote+ssh-002dremote-002bromolus.vscode-resource.vscode-cdn.net/home/stephan/Desktop/deepcadrt/~/mambaforge/envs/deepcadrt/lib/python3.9/site-packages/deepcad/train_collection.py:155) self.gap_t = math.floor((self.whole_t - self.patch_t * 2) / (s_num - 1))

ZeroDivisionError: division by zero

[cabooster]

Could you tell me the full hyperparameters you used? You can find them here

[joshbgarfinkel]

For the original issue (@s-weissbach):
s_num = ceil(train_datasets_size / (w_num * h_num * stack_num))
where:
w_num = floor((x_whole_size - x_patch_size) / x_gap_size) + 1
h_num = floor((y_whole_size - y_patch_size) / y_gap_size) + 1
x_gap_size = x_patch_size * (1 - overlap_factor)
y_gap_size = y_patch_size * (1 - overlap_factor)
In your case, x_whole_size = y_whole_size = 800, so:
w_num = h_num = floor[ ( 800-100 ) / ( 100*(1-0.9) ) ] + 1
= floor[ 700 / 90 ] + 1
= floor[ 7.77 ] + 1
= 7 + 1
= 8
The default train_datasets_size = 2000, but you specified train_datasets_size = 1.
So, with stack_num = 333,
s_num = ceil[ train_datasets_size / ( w_num * h_num * stack_num ) ]
= ceil[ 1 / ( 64 * 333 ) ]
= ceil[ 4.7E-5 ]
= 1
Because whole_t = image_shape[0] = 2100,
gap_t = floor[ ( whole_t - 2*patch_t ) / ( s_num - 1 ) ]
= floor[ ( 2100 - 20 ) / (1 - 1) ]
= floor[ 2080 / 0 ]
And there lies the issue.
If you were to raise train_datasets_size to 10,000, it would still be too little to overcome the error. Therefore, the best solution for your dataset is likely to lower w_num and h_num by increasing the xy patch size (maybe to 150), decreasing the patch overlap to 0.25 (for training), and increasing the train_datasets_size to about 2000. This would give a w_num of 5, an s_num of 3, and a gap_t of 250, solving the error.
You can choose other combinations so long as s_num > 1.
If you were to use s_num alone, instead of s_num-1 in the denominator, the last patch would extend beyond the temporal range of your data (because s_num - 1 represents the number of gaps separating each of the s_num patches).

On a another note, I ran into a similar issue as well. The quality of DeepCAD's output for my 4D dataset is heavily reliant upon a large temporal patch size. My smallest dataset returned an error on the patch_size I've been using for the majority of my data:

Image list for training -----> 
Total stack number ----->  11
Noise image name ----->  NameOfSampleFile.tif
Noise image shape ----->  (200, 512, 512)

---------------------------------------------------------------------------
ZeroDivisionError                         Traceback (most recent call last)
Cell In[10], line 1
----> 1 tc.run()

File /vf/users/garfinkeljb/CloudCode/Models/DeepCAD/deepcad/train_collection.py:78, in training_class.run(self)
     76 self.prepare_file()
     77 # crop input tiff file into 3D patches
---> 78 self.train_preprocess_lessMemoryMulStacks()
     79 # save some essential training parameters in para.yaml
     80 self.save_yaml_train()

File /vf/users/garfinkeljb/CloudCode/Models/DeepCAD/deepcad/train_collection.py:212, in training_class.train_preprocess_lessMemoryMulStacks(self)
    210 for x in range(0, int((self.whole_y - self.patch_y + self.gap_y) / self.gap_y)):
    211     for y in range(0, int((self.whole_x - self.patch_x + self.gap_x) / self.gap_x)):
--> 212         for z in range(0, int((self.whole_t - patch_t2 + self.gap_t) / self.gap_t)):
    213             single_coordinate = {'init_h': 0, 'end_h': 0, 'init_w': 0, 'end_w': 0, 'init_s': 0, 'end_s': 0}
    214             init_h = self.gap_y * x

ZeroDivisionError: division by zero

Because I wanted the maximum temporal patch size for a chosen overlap_factor, patch_xy, and train_datasets_size, I wrote the following code snippet into my training notebook:

patch_t_abs = 91      # this sets the minimum patch_t for the adjustments made by the condition function

n_epochs = 20 
alpha = 200                                  # alpha is just some scale factor I use to scale a constant train_datasets_size
train_datasets_size = alpha * Z
patch_xy = 150                            # P_XY = patch_xy
patch_t = 150                               # P_T = patch_t
overlap_factor = 0.5                    # O = overlap_factor
batch_size = 8


def condition(P_T, alpha, T, X, P_XY, O):
    num = T - 2*P_T
    G_XY = floor(P_XY * (1 - O))
    N_XY = floor((X - P_XY) / G_XY)
    denom = ceil(alpha * N_XY**-2) - 1
    denom = floor(num / denom)
    t = 0
    P0 = P_T
    while denom <= 0:
        P_T -= 1
        num = T - 2*P_T
        G_XY = floor(P_XY * (1 - O))
        denom = ceil(alpha * N_XY**-2) - 1
        denom = floor(num / denom)
        t += 1
        if P_T < patch_t_abs_min: raise warnings.warn("The adjusted temporal patch size goes below the limit."); denom = 1
    if t > 0: print(f'Temporal patch size changed from {P0} to {P_T} to ensure DeepCAD works.')
    return P_T

patch_t = condition(patch_t, alpha, T, X, patch_xy, overlap_factor)

The attached PDF shows the specific logic and problem a little more explicitly.

Patch Size Calculation.pdf