Add option to reuse .bin files
#116
Conversation
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. Please check the settings in the CodeRabbit UI or the You can disable this status message by setting the WalkthroughThe pull request introduces extensive modifications to the configuration documentation and related YAML files for the Changes
Possibly related PRs
Suggested reviewers
Thank you for using CodeRabbit. We offer it for free to the OSS community and would appreciate your support in helping us grow. If you find it useful, would you consider giving us a shout-out on your favorite social media? 🪧 TipsChatThere are 3 ways to chat with CodeRabbit:
Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments. CodeRabbit Commands (Invoked using PR comments)
Other keywords and placeholders
CodeRabbit Configuration File (
|
gitttt-1234
changed the base branch from
main
to
divya/load-head-ckpt-inference
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Caution
Inline review comments failed to post. This is likely due to GitHub's limits when posting large numbers of comments.
Actionable comments posted: 20
🧹 Outside diff range and nitpick comments (35)
tests/assets/minimal_instance_centroid/initial_config.yaml (1)
78-85: Review learning rate scheduler parameters.The ReduceLROnPlateau configuration uses very small values:
threshold: 1.0e-07andmin_lr: 1.0e-08are quite lowpatience: 5withcooldown: 3means LR changes can occur frequentlyConsider if these values are appropriate for your model and training dynamics. Too-frequent LR changes might destabilize training.
Recommendations:
- Consider increasing the threshold to ~1e-4 for more stable training
- You might want to increase cooldown to prevent too-frequent LR changes
- Document why these specific values were chosen
tests/assets/minimal_instance_bottomup/initial_config.yaml (1)
70-71: Document .bin file deletion behavior.Since this PR introduces the option to reuse .bin files, it would be helpful to add a comment or parameter that explicitly indicates whether .bin files will be automatically deleted after training. This would make the behavior more transparent to users.
Consider adding a parameter like
delete_bin_files: truewith appropriate documentation.tests/assets/minimal_instance_centroid/training_config.yaml (1)
89-96: Consider separating LR scheduler changes into a different PRThe learning rate scheduler changes appear to be unrelated to the main objective of this PR (reusing .bin files). While the configuration looks well-tuned, it might be clearer to handle these changes in a separate PR focused on training optimizations.
The current settings are reasonable:
- Reduction factor of 0.5
- Patience of 5 epochs with 3 epochs cooldown
- Minimum learning rate of 1e-8
tests/assets/minimal_instance_bottomup/training_config.yaml (2)
85-85: Document the bin_files_path usage.This parameter enables the reuse of .bin files as described in the PR objectives. Consider:
- Adding a comment explaining the expected path format
- Documenting whether relative paths are supported
- Clarifying what happens when the path is invalid
Add a YAML comment above this line:
+ # Path to existing .bin files for reusing training data chunks. Set to null to generate new chunks. bin_files_path:
99-106: LGTM! Well-structured learning rate scheduler configuration.The ReduceLROnPlateau configuration has sensible defaults:
- Gradual reduction (factor: 0.5)
- Reasonable patience (5 epochs) and cooldown (3 epochs)
- Safe minimum learning rate (1e-8)
Consider adding comments to explain the threshold_mode options (abs vs rel) for future maintainers.
tests/assets/minimal_instance/training_config.yaml (2)
5-6: Document the new data configuration parameters.New parameters have been added without documentation explaining their purpose and impact:
user_instances_onlychunk_sizemin_crop_sizePlease add documentation describing:
- What each parameter controls
- Expected values/ranges
- Default behavior
Also applies to: 15-15
Line range hint
1-103: Consider splitting changes into focused PRs.The current changes mix multiple concerns:
- .bin file reuse functionality (primary objective)
- Learning rate scheduler modifications
- Data preprocessing parameters
This makes the changes harder to review and maintain. Consider:
- Keeping only the .bin file reuse changes in this PR
- Moving LR scheduler and preprocessing changes to separate PRs
- Adding comprehensive documentation for all new parameters
Additionally, please add validation to ensure the
bin_files_pathpoints to a valid directory when provided.docs/config_bottomup.yaml (2)
6-6: Add documentation for thechunk_sizeparameter.Since this parameter is crucial for the new feature of managing training data chunks, please add a comment explaining its purpose, impact on training, and any constraints on its value.
data_config: provider: LabelsReader train_labels_path: minimal_instance.pkg.slp val_labels_path: minimal_instance.pkg.slp user_instances_only: True + # Size of data chunks for training. Controls how many frames are processed + # together when generating .bin files chunk_size: 100
Line range hint
92-92: Document thebin_files_pathparameter.Since this PR introduces the ability to reuse
.binfiles, please add documentation explaining:
- The purpose of this parameter
- Expected path format
- Behavior when the path is empty vs. when it contains a value
save_ckpt: true save_ckpt_path: min_inst_bottomup1 + # Path to existing .bin files for reusing training data chunks + # Leave empty to generate new chunks, or specify path to reuse existing ones bin_files_path:docs/config_topdown_centered_instance.yaml (3)
6-6: Document thechunk_sizeparameter's purpose and impact.The newly added
chunk_sizeparameter aligns with the PR's objective of managing training data chunks, but its purpose and impact should be documented for users.Consider adding a comment explaining:
- What this parameter controls
- How it affects the training process
- Any recommended values or constraints
107-114: LGTM! Well-structured learning rate scheduler configuration.The ReduceLROnPlateau configuration is comprehensive with appropriate parameters for stable training. It pairs well with the early stopping configuration.
Consider adding inline comments explaining the purpose of each parameter, especially
threshold_mode: absvs.rel, to help users customize these values for their needs.
Line range hint
92-93: Add parameter to control .bin files deletion.The PR aims to allow users to control the automatic deletion of .bin files, but this configuration is missing a parameter to enable/disable this feature.
Consider adding a parameter like
retain_bin_files: falsenear thebin_files_pathconfiguration to control this behavior.bin_files_path: + retain_bin_files: false # Controls whether to keep .bin files after training🧰 Tools
🪛 yamllint
[error] 8-8: trailing spaces
(trailing-spaces)
[error] 9-9: trailing spaces
(trailing-spaces)
initial_config.yaml (2)
77-77: Document the bin_files_path parameter.This new parameter aligns with the PR objective to reuse
.binfiles, but needs documentation about:
- Expected directory structure
- File naming conventions
- Any requirements for the binary files
Consider adding a comment above this parameter explaining its usage:
+ # Path to directory containing pre-generated .bin training chunks + # If provided, these chunks will be reused instead of generating new ones bin_files_path: null
73-73: Add warning about GPU determinism.While setting a fixed seed is good for reproducibility, users should be aware that complete determinism requires additional settings when using GPUs.
Consider adding a warning comment:
+ # Note: For complete determinism with GPUs, additional PyTorch settings are required seed: 1000docs/config_centroid.yaml (2)
6-6: Consider documenting chunk_size parameter and its implications.While the
chunk_sizeparameter has been added, its relationship to the.binfiles and data chunking process should be documented in comments.data_config: provider: LabelsReader train_labels_path: minimal_instance.pkg.slp val_labels_path: minimal_instance.pkg.slp user_instances_only: True - chunk_size: 100 + # Size of data chunks used for training. Affects the size of generated .bin files + chunk_size: 100
Line range hint
89-89: Add configuration for.binfile retention.The PR aims to add an option to control
.binfile deletion, but there's no visible configuration parameter for this feature. Consider adding a parameter to control this behavior.save_ckpt: true save_ckpt_path: 'min_inst_centroid' bin_files_path: + # Set to true to retain .bin files after training (default: false) + retain_bin_files: false resume_ckpt_path:🧰 Tools
🪛 yamllint
[error] 8-8: trailing spaces
(trailing-spaces)
[error] 9-9: trailing spaces
(trailing-spaces)
sleap_nn/data/streaming_datasets.py (2)
154-155: Consider adding input validation for crop dimensions.Since the crop dimensions are critical for proper functioning and are now calculated once during initialization, consider adding validation to ensure the scaled dimensions remain valid.
Add input validation:
# Re-crop to original crop size +if not all(x > 0 for x in self.crop_hw): + raise ValueError(f"Invalid crop dimensions after scaling: {self.crop_hw}. Check input_scale: {self.input_scale}") self.crop_hw = [int(x * self.input_scale) for x in self.crop_hw]
154-155: Document the relationship with .bin files.Since this PR focuses on
.binfile reuse, it would be helpful to document how the crop size affects the data chunks stored in.binfiles.Add documentation:
# Re-crop to original crop size +# Note: These dimensions affect the data chunks stored in .bin files. +# When reusing .bin files, ensure consistent crop_hw and input_scale values self.crop_hw = [int(x * self.input_scale) for x in self.crop_hw]docs/config.md (2)
Line range hint
1-188: Documentation missing for.binfiles reuse feature.The PR introduces the ability to reuse
.binfiles, but the documentation only mentions thebin_files_pathparameter for saving them. Please add documentation for:
- Parameters controlling automatic deletion of
.binfiles- How to specify existing
.binfiles for reuse in trainingWould you like me to help draft the documentation for these new parameters?
🧰 Tools
🪛 LanguageTool
[uncategorized] ~179-~179: Loose punctuation mark.
Context: ...ReduceLROnPlateau". -step_lr: -step_size`: (int) Period...(UNLIKELY_OPENING_PUNCTUATION)
[uncategorized] ~182-~182: Loose punctuation mark.
Context: ...*: 0.1. -reduce_lr_on_plateau: -threshold: (float) Thre...(UNLIKELY_OPENING_PUNCTUATION)
🪛 Markdownlint
170-170: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
171-171: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
172-172: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
173-173: Expected: 2; Actual: 4
Unordered list indentation(MD007, ul-indent)
174-174: Expected: 2; Actual: 4
Unordered list indentation(MD007, ul-indent)
175-175: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
176-176: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
177-177: Expected: 2; Actual: 4
Unordered list indentation(MD007, ul-indent)
178-178: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
179-179: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
180-180: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
181-181: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
182-182: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
183-183: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
184-184: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
185-185: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
186-186: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
187-187: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
188-188: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
189-189: Expected: 2; Actual: 4
Unordered list indentation(MD007, ul-indent)
190-190: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
191-191: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
184-184: null
Spaces inside emphasis markers(MD037, no-space-in-emphasis)
184-184: null
Spaces inside emphasis markers(MD037, no-space-in-emphasis)
178-188: Enhance scheduler configuration documentation with examples.The scheduler configuration is well-documented but could benefit from practical examples. Consider adding:
- Example configurations for both scheduler types
- Common use cases for each parameter
- Guidelines for choosing between StepLR and ReduceLROnPlateau
Example addition:
Example configurations: ```yaml # StepLR: Reduce learning rate by half every 10 epochs lr_scheduler: scheduler: "StepLR" step_lr: step_size: 10 gamma: 0.5 # ReduceLROnPlateau: Reduce learning rate when validation loss plateaus lr_scheduler: scheduler: "ReduceLROnPlateau" reduce_lr_on_plateau: patience: 5 factor: 0.1 threshold: 1e-4<details> <summary>🧰 Tools</summary> <details> <summary>🪛 LanguageTool</summary> [uncategorized] ~179-~179: Loose punctuation mark. Context: ...ReduceLROnPlateau"`. - `step_lr`: - `step_size`: (int) Period... (UNLIKELY_OPENING_PUNCTUATION) --- [uncategorized] ~182-~182: Loose punctuation mark. Context: ...*: 0.1. - `reduce_lr_on_plateau`: - `threshold`: (float) Thre... (UNLIKELY_OPENING_PUNCTUATION) </details> <details> <summary>🪛 Markdownlint</summary> 178-178: Expected: 4; Actual: 8 Unordered list indentation (MD007, ul-indent) --- 179-179: Expected: 4; Actual: 8 Unordered list indentation (MD007, ul-indent) --- 180-180: Expected: 6; Actual: 12 Unordered list indentation (MD007, ul-indent) --- 181-181: Expected: 6; Actual: 12 Unordered list indentation (MD007, ul-indent) --- 182-182: Expected: 4; Actual: 8 Unordered list indentation (MD007, ul-indent) --- 183-183: Expected: 6; Actual: 12 Unordered list indentation (MD007, ul-indent) --- 184-184: Expected: 6; Actual: 12 Unordered list indentation (MD007, ul-indent) --- 185-185: Expected: 6; Actual: 12 Unordered list indentation (MD007, ul-indent) --- 186-186: Expected: 6; Actual: 12 Unordered list indentation (MD007, ul-indent) --- 187-187: Expected: 6; Actual: 12 Unordered list indentation (MD007, ul-indent) --- 188-188: Expected: 6; Actual: 12 Unordered list indentation (MD007, ul-indent) --- 184-184: null Spaces inside emphasis markers (MD037, no-space-in-emphasis) --- 184-184: null Spaces inside emphasis markers (MD037, no-space-in-emphasis) </details> </details> </blockquote></details> <details> <summary>sleap_nn/inference/predictors.py (3)</summary><blockquote> `94-95`: **Add parameter validation for checkpoint paths.** Consider adding validation to ensure that if `head_ckpt_path` is provided, `backbone_ckpt_path` must also be provided, as loading only head weights without backbone weights might lead to unexpected behavior. ```diff def from_model_paths( cls, model_paths: List[Text], backbone_ckpt_path: Optional[str] = None, head_ckpt_path: Optional[str] = None, + # Add validation + if head_ckpt_path is not None and backbone_ckpt_path is None: + raise ValueError("backbone_ckpt_path must be provided when head_ckpt_path is provided")
1622-1626: Improve parameter documentation in docstring.The docstring for the new parameters could be more descriptive and include examples:
- backbone_ckpt_path: (str) To run inference on any `.ckpt` other than `best.ckpt` - from the `model_paths` dir, the path to the `.ckpt` file should be passed here. - head_ckpt_path: (str) Path to `.ckpt` file if a different set of head layer weights - are to be used. If `None`, the `best.ckpt` from `model_paths` dir is used (or the ckpt - from `backbone_ckpt_path` if provided.) + backbone_ckpt_path: Optional path to a checkpoint file containing backbone weights. + If provided, these weights will be used instead of the backbone weights from + `best.ckpt`. This allows mixing weights from different checkpoints. + Example: "/path/to/backbone_v2.ckpt" + head_ckpt_path: Optional path to a checkpoint file containing head layer weights. + Can only be used if backbone_ckpt_path is also provided. This enables using + different head weights while maintaining the same backbone. + Example: "/path/to/head_specialized.ckpt"
596-597: Add error handling for checkpoint loading.Consider adding try-except blocks when loading checkpoints to handle potential errors gracefully:
- File not found
- Invalid checkpoint format
- Incompatible state dict structure
+def _safe_load_checkpoint(path: str) -> dict: + """Safely load a checkpoint file with error handling.""" + try: + return torch.load(path) + except FileNotFoundError: + raise ValueError(f"Checkpoint file not found: {path}") + except Exception as e: + raise ValueError(f"Error loading checkpoint {path}: {str(e)}")Also applies to: 606-607, 610-611
tests/training/test_model_trainer.py (2)
216-218: Simplify theOmegaConf.updatecall formattingThe
OmegaConf.updatecall can be condensed into a single line for readability, as it doesn't exceed typical line length limits.Apply this diff to improve formatting:
-OmegaConf.update( - config_early_stopping, "trainer_config.lr_scheduler.scheduler", None -) +OmegaConf.update(config_early_stopping, "trainer_config.lr_scheduler.scheduler", None)
330-333: Clarify intentional use of invalid scheduler nameThe scheduler name "ReduceLR" is likely intended to be invalid to test exception handling. Consider adding a comment to clarify this for future maintainability.
Apply this diff to add a clarifying comment:
+ # Intentionally using an invalid scheduler name to test exception handling OmegaConf.update(config, "trainer_config.lr_scheduler.scheduler", "ReduceLR") with pytest.raises(ValueError): trainer = ModelTrainer(config)tests/inference/test_predictors.py (3)
3-3: Remove unused importTextThe
Textclass from thetypingmodule is imported but not used in the code. Removing this unused import will clean up the code.Apply this diff to remove the unused import:
-from typing import Text🧰 Tools
🪛 Ruff
3-3:
typing.Textimported but unusedRemove unused import:
typing.Text(F401)
692-693: Remove unnecessary debug print statementsThe
Apply this diff to remove the print statements:
- print(f"head_layer_ckpt: {head_layer_ckpt}")- print(model_weights)Also applies to: 701-702
692-693: Remove unnecessary debug print statementsThe
Apply this diff:
- print(f"head_layer_ckpt: {head_layer_ckpt}")- print(model_weights)Also applies to: 701-702
sleap_nn/training/model_trainer.py (7)
376-383: Add docstrings for new parameters intrainmethodThe
trainmethod now accepts additional parameters that are not documented. Providing docstrings for these parameters enhances code readability and user understanding.Apply this diff to add the parameter documentation:
def train( self, backbone_trained_ckpts_path: Optional[str] = None, head_trained_ckpts_path: Optional[str] = None, delete_bin_files_after_training: bool = True, train_chunks_dir_path: Optional[str] = None, val_chunks_dir_path: Optional[str] = None, ): + """ + Initiate the training process. + + Args: + backbone_trained_ckpts_path: Path to a trained backbone checkpoint for model initialization. + head_trained_ckpts_path: Path to a trained head checkpoint for model initialization. + delete_bin_files_after_training: Whether to delete `.bin` files after training. Defaults to True. + train_chunks_dir_path: Path to existing training chunks directory. If None, new chunks will be created. + val_chunks_dir_path: Path to existing validation chunks directory. If None, new chunks will be created. + """ logger = []
354-358: Add parameter descriptions to the_initialize_modelmethod docstringThe
_initialize_modelmethod accepts new parameters that aren't documented. Adding descriptions will improve code clarity.Apply this diff to add the parameter documentation:
def _initialize_model( self, backbone_trained_ckpts_path: Optional[str] = None, head_trained_ckpts_path: Optional[str] = None, ): + """ + Initialize the model with optional pretrained checkpoints. + + Args: + backbone_trained_ckpts_path: Path to a pretrained backbone checkpoint. + head_trained_ckpts_path: Path to a pretrained head checkpoint. + """ models = {
528-529: Correct typos in docstring parameter descriptionsIn the docstring, replace "ckpts" with "checkpoints" for clarity.
Apply this diff to correct the typos:
model_type: Type of the model. One of `single_instance`, `centered_instance`, `centroid`, `bottomup`. - backbone_trained_ckpts_path: Path to trained ckpts for backbone. - head_trained_ckpts_path: Path to trained ckpts for head layer. + backbone_trained_ckpts_path: Path to trained checkpoints for backbone. + head_trained_ckpts_path: Path to trained checkpoints for head layer. """
743-744: Correct typos in docstring parameter descriptionsIn the docstring of
SingleInstanceModel, replace "ckpts" with "checkpoints" for clarity.Apply this diff to correct the typos:
model_type: Type of the model. One of `single_instance`, `centered_instance`, `centroid`, `bottomup`. - backbone_trained_ckpts_path: Path to trained ckpts for backbone. - head_trained_ckpts_path: Path to trained ckpts for head layer. + backbone_trained_ckpts_path: Path to trained checkpoints for backbone. + head_trained_ckpts_path: Path to trained checkpoints for head layer. """
824-825: Correct typos in docstring parameter descriptionsIn the docstring of
TopDownCenteredInstanceModel, replace "ckpts" with "checkpoints" for clarity.Apply this diff to correct the typos:
model_type: Type of the model. One of `single_instance`, `centered_instance`, `centroid`, `bottomup`. - backbone_trained_ckpts_path: Path to trained ckpts for backbone. - head_trained_ckpts_path: Path to trained ckpts for head layer. + backbone_trained_ckpts_path: Path to trained checkpoints for backbone. + head_trained_ckpts_path: Path to trained checkpoints for head layer. """
905-906: Correct typos in docstring parameter descriptionsIn the docstring of
CentroidModel, replace "ckpts" with "checkpoints" for clarity.Apply this diff to correct the typos:
model_type: Type of the model. One of `single_instance`, `centered_instance`, `centroid`, `bottomup`. - backbone_trained_ckpts_path: Path to trained ckpts for backbone. - head_trained_ckpts_path: Path to trained ckpts for head layer. + backbone_trained_ckpts_path: Path to trained checkpoints for backbone. + head_trained_ckpts_path: Path to trained checkpoints for head layer. """
986-987: Correct typos in docstring parameter descriptionsIn the docstring of
BottomUpModel, replace "ckpts" with "checkpoints" for clarity.Apply this diff to correct the typos:
model_type: Type of the model. One of `single_instance`, `centered_instance`, `centroid`, `bottomup`. - backbone_trained_ckpts_path: Path to trained ckpts for backbone. - head_trained_ckpts_path: Path to trained ckpts for head layer. + backbone_trained_ckpts_path: Path to trained checkpoints for backbone. + head_trained_ckpts_path: Path to trained checkpoints for head layer. """
🛑 Comments failed to post (20)
tests/assets/minimal_instance/training_config.yaml (1)
82-82: 💡 Codebase verification
Based on the search results, I can now generate a final response as I found the implementation details in
sleap_nn/training/model_trainer.py:Clarify bin_files_path behavior in configuration documentation
The implementation shows that
bin_files_pathis used as follows:
- When null (default), it uses the input directory path
- When set, creates a timestamped subdirectory under the specified path (
{bin_files_path}/chunks_{timestamp})- The path can be absolute or relative, and proper directory creation with permissions is handled
Please add these details to the configuration documentation to help users understand:
- The default behavior (null = use input directory)
- The automatic timestamped subdirectory creation
- That both absolute and relative paths are supported
🔗 Analysis chain
Enhance bin_files_path configuration.
The new
bin_files_pathparameter is crucial for the .bin file reuse feature, but needs more clarity:
- Document the expected path format (absolute/relative)
- Clarify what happens when the value is null
- Consider adding path validation
Let's check if there's any documentation or validation for this parameter:
Let me try another search strategy to find any documentation or validation related to
bin_files_path:🏁 Scripts executed
The following scripts were executed for the analysis:
Script:
#!/bin/bash # Search for documentation or validation related to bin_files_path rg -A 5 "bin_files_path"Length of output: 195
Script:
#!/bin/bash # Search for bin_files_path in Python files and YAML files fd -e py -e yaml --exec rg -l "bin_files_path" {} # Get content of files containing bin_files_path fd -e py -e yaml --exec rg -B 5 -A 5 "bin_files_path" {}Length of output: 12827
initial_config.yaml (3)
1-104:
⚠️ Potential issueFix newline characters in the file.
The file uses incorrect newline characters which can cause issues across different operating systems.
Convert all line endings to Unix-style (LF) newlines. You can do this using:
- Git:
git config --global core.autocrlf input- Editor: Most modern editors can handle this automatically
- Command line:
dos2unix initial_config.yaml🧰 Tools
🪛 yamllint
[error] 1-1: wrong new line character: expected \n
(new-lines)
84-84:
⚠️ Potential issueRemove API key field from configuration.
Storing API keys in configuration files, even if empty, is a security risk as it might encourage users to commit sensitive information.
Consider:
- Removing the API key field entirely and using environment variables
- Adding a warning comment about not committing API keys
- api_key: '' + # Configure API key using WANDB_API_KEY environment variable📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.# Configure API key using WANDB_API_KEY environment variable
3-4:
⚠️ Potential issueReplace hardcoded Windows paths with platform-independent paths.
The current paths have several issues:
- Windows-specific backslashes make it non-portable
- Absolute paths are not suitable for version control
- Using the same file for both training and validation could lead to overfitting
Consider using:
- Relative paths with forward slashes
- Different datasets for training and validation
- train_labels_path: C:\Users\TalmoLab\Desktop\Divya\sleap-nn\tests\assets/minimal_instance.pkg.slp - val_labels_path: C:\Users\TalmoLab\Desktop\Divya\sleap-nn\tests\assets/minimal_instance.pkg.slp + train_labels_path: data/train/minimal_instance.pkg.slp + val_labels_path: data/val/minimal_instance.pkg.slp📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.train_labels_path: data/train/minimal_instance.pkg.slp val_labels_path: data/val/minimal_instance.pkg.slpdocs/config.md (1)
178-188:
⚠️ Potential issueFix markdown formatting issues.
There are several formatting inconsistencies in the scheduler configuration section:
- List indentation is inconsistent
- Unnecessary spaces in emphasis markers
Apply these formatting fixes:
- - `scheduler`: (str) Name of the scheduler to use. Valid schedulers: `"StepLR"`, `"ReduceLROnPlateau"`. - - `step_lr`: - - `step_size`: (int) Period of learning rate decay. If `step_size`=10, then every 10 epochs, learning rate will be reduced by a factor of `gamma`. + - `scheduler`: (str) Name of the scheduler to use. Valid schedulers: `"StepLR"`, `"ReduceLROnPlateau"`. + - `step_lr`: + - `step_size`: (int) Period of learning rate decay. If `step_size`=10, then every 10 epochs, learning rate will be reduced by a factor of `gamma`.Committable suggestion skipped: line range outside the PR's diff.
🧰 Tools
🪛 LanguageTool
[uncategorized] ~179-~179: Loose punctuation mark.
Context: ...ReduceLROnPlateau". -step_lr: -step_size`: (int) Period...(UNLIKELY_OPENING_PUNCTUATION)
[uncategorized] ~182-~182: Loose punctuation mark.
Context: ...*: 0.1. -reduce_lr_on_plateau: -threshold: (float) Thre...(UNLIKELY_OPENING_PUNCTUATION)
🪛 Markdownlint
178-178: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
179-179: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
180-180: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
181-181: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
182-182: Expected: 4; Actual: 8
Unordered list indentation(MD007, ul-indent)
183-183: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
184-184: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
185-185: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
186-186: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
187-187: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
188-188: Expected: 6; Actual: 12
Unordered list indentation(MD007, ul-indent)
184-184: null
Spaces inside emphasis markers(MD037, no-space-in-emphasis)
184-184: null
Spaces inside emphasis markers(MD037, no-space-in-emphasis)
sleap_nn/inference/predictors.py (3)
594-617: 🛠️ Refactor suggestion
Refactor duplicated weight loading logic.
The weight loading logic is duplicated between the centroid and confmap models. Consider extracting this into a helper function to improve maintainability and reduce code duplication.
+def _load_model_weights(model, backbone_ckpt_path: Optional[str], head_ckpt_path: Optional[str]) -> None: + """Load backbone and head weights into the model. + + Args: + model: The model to load weights into + backbone_ckpt_path: Path to backbone checkpoint + head_ckpt_path: Path to head checkpoint + """ + if backbone_ckpt_path is not None and head_ckpt_path is not None: + print(f"Loading backbone weights from `{backbone_ckpt_path}` ...") + ckpt = torch.load(backbone_ckpt_path) + ckpt["state_dict"] = { + k: v for k, v in ckpt["state_dict"].items() + if ".backbone" in k + } + model.load_state_dict(ckpt["state_dict"], strict=False) + elif backbone_ckpt_path is not None: + print(f"Loading weights from `{backbone_ckpt_path}` ...") + ckpt = torch.load(backbone_ckpt_path) + model.load_state_dict(ckpt["state_dict"], strict=False) + + if head_ckpt_path is not None: + print(f"Loading head weights from `{head_ckpt_path}` ...") + ckpt = torch.load(head_ckpt_path) + ckpt["state_dict"] = { + k: v for k, v in ckpt["state_dict"].items() + if ".head_layers" in k + } + model.load_state_dict(ckpt["state_dict"], strict=False)Then use this helper function:
- if backbone_ckpt_path is not None and head_ckpt_path is not None: - print(f"Loading backbone weights from `{backbone_ckpt_path}` ...") - ckpt = torch.load(backbone_ckpt_path) - ckpt["state_dict"] = { - k: ckpt["state_dict"][k] - for k in ckpt["state_dict"].keys() - if ".backbone" in k - } - centroid_model.load_state_dict(ckpt["state_dict"], strict=False) - - elif backbone_ckpt_path is not None: - print(f"Loading weights from `{backbone_ckpt_path}` ...") - ckpt = torch.load(backbone_ckpt_path) - centroid_model.load_state_dict(ckpt["state_dict"], strict=False) - - if head_ckpt_path is not None: - print(f"Loading head weights from `{head_ckpt_path}` ...") - ckpt = torch.load(head_ckpt_path) - ckpt["state_dict"] = { - k: ckpt["state_dict"][k] - for k in ckpt["state_dict"].keys() - if ".head_layers" in k - } - centroid_model.load_state_dict(ckpt["state_dict"], strict=False) + _load_model_weights(centroid_model, backbone_ckpt_path, head_ckpt_path)Also applies to: 636-659
🧰 Tools
🪛 Ruff
599-599: Use
key in dictinstead ofkey in dict.keys()Remove
.keys()(SIM118)
614-614: Use
key in dictinstead ofkey in dict.keys()Remove
.keys()(SIM118)
1349-1372: 🛠️ Refactor suggestion
Apply consistent improvements across all predictor classes.
The same improvements suggested for other predictor classes should be applied here:
- Simplify dictionary key checks
- Use the shared weight loading helper function
This ensures consistency across the codebase and reduces maintenance overhead.
🧰 Tools
🪛 Ruff
1354-1354: Use
key in dictinstead ofkey in dict.keys()Remove
.keys()(SIM118)
1369-1369: Use
key in dictinstead ofkey in dict.keys()Remove
.keys()(SIM118)
988-1011: 🛠️ Refactor suggestion
Simplify dictionary key checks and apply consistent weight loading pattern.
- Simplify the dictionary key checks by removing unnecessary
.keys()calls- Use the same helper function suggested for TopDownPredictor to handle weight loading
- k: ckpt["state_dict"][k] - for k in ckpt["state_dict"].keys() + k: v for k, v in ckpt["state_dict"].items() if ".backbone" in kApply the same
_load_model_weightshelper function here to maintain consistency and reduce code duplication.Committable suggestion skipped: line range outside the PR's diff.
🧰 Tools
🪛 Ruff
993-993: Use
key in dictinstead ofkey in dict.keys()Remove
.keys()(SIM118)
1008-1008: Use
key in dictinstead ofkey in dict.keys()Remove
.keys()(SIM118)
tests/training/test_model_trainer.py (2)
338-376: 🛠️ Refactor suggestion
Avoid hardcoded indices when accessing model parameters
Directly accessing model parameters with hardcoded indices like
[0, 0, :]can lead to maintenance issues if the model architecture changes. Consider using parameter names or iterating over the parameters for a more robust approach.
379-424:
⚠️ Potential issueAdd assertions to verify reuse of
.binfilesThe test
test_reuse_bin_filessets up for reusing.binfiles but lacks assertions to confirm that the files are indeed reused. Adding assertions will strengthen the test by ensuring the files are not regenerated.Consider adding these assertions:
assert os.path.exists(trainer1.train_input_dir) assert os.path.exists(trainer1.val_input_dir) assert os.path.exists(trainer2.train_input_dir) assert os.path.exists(trainer2.val_input_dir) assert trainer1.train_input_dir == trainer2.train_input_dir assert trainer1.val_input_dir == trainer2.val_input_dirtests/inference/test_predictors.py (5)
188-218: 🛠️ Refactor suggestion
Refactor duplicate test code into a helper function
The code block from lines 188 to 218 appears to be duplicated in other test functions (e.g., lines 447-496 and 670-704). Refactoring this repeated code into a helper function would enhance maintainability and reduce code duplication.
447-496: 🛠️ Refactor suggestion
Refactor duplicate test code into a helper function
The code from lines 447 to 496 is similar to code in other test sections. Refactoring this repeated logic into a shared helper function can improve code maintainability and readability.
670-704: 🛠️ Refactor suggestion
Refactor duplicate test code into a helper function
The code block from lines 670 to 704 is repeated in other tests. Refactoring into a helper function will reduce redundancy and simplify future maintenance.
491-491: 🛠️ Refactor suggestion
Use
np.testing.assert_allclosefor numerical assertionsFor consistency and better error handling, use
np.testing.assert_allcloseinstead of manual assertions when comparing numerical arrays.Apply this diff:
-assert np.all(np.abs(head_layer_ckpt - model_weights) < 1e-6) +np.testing.assert_allclose(head_layer_ckpt, model_weights, atol=1e-6)📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.np.testing.assert_allclose(head_layer_ckpt, model_weights, atol=1e-6)
218-218: 🛠️ Refactor suggestion
Use
np.testing.assert_allclosefor numerical assertionsInstead of manually checking numerical closeness with
assert np.all(np.abs(...)) < tolerance, consider usingnp.testing.assert_allclosefor better readability and error messages.Apply this diff to improve the assertion:
-assert np.all(np.abs(head_layer_ckpt - model_weights) < 1e-6) +np.testing.assert_allclose(head_layer_ckpt, model_weights, atol=1e-6)📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.np.testing.assert_allclose(head_layer_ckpt, model_weights, atol=1e-6)sleap_nn/training/model_trainer.py (5)
380-380:
⚠️ Potential issueClarify default behavior of
delete_bin_files_after_trainingparameterThe parameter
delete_bin_files_after_trainingdefaults toTrue, meaning.binfiles will be deleted after training. This could be unexpected for users who want to reuse these files. Consider setting the default toFalseor clearly documenting this behavior.
240-240:
⚠️ Potential issueUse
raise ... from ewhen re-raising exceptions to preserve tracebackWhen re-raising an exception, include the original exception using
from eto maintain the traceback.Apply this diff to modify the raise statement:
- raise Exception(f"Error while creating the `.bin` files... {e}") + raise Exception("Error while creating the `.bin` files...") from e📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.raise Exception("Error while creating the `.bin` files...") from e🧰 Tools
🪛 Ruff
240-240: Within an
exceptclause, raise exceptions withraise ... from errorraise ... from Noneto distinguish them from errors in exception handling(B904)
224-226:
⚠️ Potential issueUse
raise ... from ewhen re-raising exceptions to preserve tracebackWhen raising a new exception within an
exceptblock, it's best practice to useraise ... from eto maintain the original traceback.Apply this diff to modify the raise statement:
- raise OSError( - f"Cannot create a new folder in {self.bin_files_path}. Check the permissions to the given Checkpoint directory. \n {e}" - ) + raise OSError( + f"Cannot create a new folder in {self.bin_files_path}. Check the permissions to the given Checkpoint directory." + ) from e📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.raise OSError( f"Cannot create a new folder in {self.bin_files_path}. Check the permissions to the given Checkpoint directory." ) from e🧰 Tools
🪛 Ruff
224-226: Within an
exceptclause, raise exceptions withraise ... from errorraise ... from Noneto distinguish them from errors in exception handling(B904)
613-613:
⚠️ Potential issueSimplify dictionary key iteration by removing
.keys()When iterating over a dictionary's keys, you can omit
.keys()for simplicity.Apply this diff to simplify the comprehension:
ckpt["state_dict"] = { k: ckpt["state_dict"][k] - for k in ckpt["state_dict"].keys() + for k in ckpt["state_dict"] if ".backbone" in k }Committable suggestion skipped: line range outside the PR's diff.
🧰 Tools
🪛 Ruff
613-613: Use
key in dictinstead ofkey in dict.keys()Remove
.keys()(SIM118)
624-624:
⚠️ Potential issueSimplify dictionary key iteration by removing
.keys()When iterating over a dictionary's keys, you can omit
.keys()for simplicity.Apply this diff to simplify the comprehension:
ckpt["state_dict"] = { k: ckpt["state_dict"][k] - for k in ckpt["state_dict"].keys() + for k in ckpt["state_dict"] if ".head_layers" in k }Committable suggestion skipped: line range outside the PR's diff.
🧰 Tools
🪛 Ruff
624-624: Use
key in dictinstead ofkey in dict.keys()Remove
.keys()(SIM118)
Codecov ReportAttention: Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## divya/load-head-ckpt-inference #116 +/- ##
==================================================================
+ Coverage 97.34% 97.51% +0.17%
==================================================================
Files 38 38
Lines 3761 3778 +17
==================================================================
+ Hits 3661 3684 +23
+ Misses 100 94 -6 ☔ View full report in Codecov by Sentry. |
This PR introduces an option to enable/ disable the auto-deletion of
.binfiles (data chunks for training) generated byld.optimie. Additionally, this provides the flexibility to load exitsing train and validation chunks into any training process, by passing the paths to the.binfolder to theModelTrainer.train()function.Summary by CodeRabbit
New Features
chunk_sizeanduser_instances_onlyacross various configuration files.Bug Fixes
Tests