Skip to content

model

This abstract class provides support for tiled prediction, and is the base class for all models used in this library.

Model

Bases: ABC

Abstract class for tiled inference models

Source code in src/tcd_pipeline/models/model.py 
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
class Model(ABC):
    """Abstract class for tiled inference models"""

    def __init__(self, config: DictConfig):
        """
        Args:
            config (DictConfig): Configuration dictionary

        """

        self.config = config

        if config.model.device == "cuda" and not torch.cuda.is_available():
            logger.warning("Failed to use CUDA, falling back to CPU")
            self.device = "cpu"
        else:
            self.device = config.model.device

        self.model = None
        self.should_reload = False
        self.post_processor: PostProcessor = None
        self.failed_images = set()
        self.should_exit = False

        logger.info("Device: %s", self.device)
        self.setup()

    @abstractmethod
    def setup(self):
        """Perform any setup actions as needed"""

    @abstractmethod
    def load_model(self):
        """Load the model, defined by subclass"""

    def on_after_predict(self, results: dict) -> None:
        """Append tiled results to the post processor, or cache

        Args:
            results (list): Prediction results from one tile

        """

        t_start = time.time()

        # Invert dict-of-lists to list-of-dicts
        results = [dict(zip(results, t)) for t in zip(*results.values())]

        self.post_processor.add(results)

        self.t_postprocess = time.time() - t_start

    def post_process(self) -> ProcessedResult:
        """Run post-processing to merge results

        Returns:
            ProcessedResult: merged results
        """

        res = self.post_processor.process()

        if self.config.postprocess.cleanup:
            logger.info("Cleaning up post processor")
            self.post_processor.cache.clear()

        return res

    def attempt_reload(self) -> None:
        """Attempts to reload the model."""
        if "cuda" not in self.device:
            return

        del self.model

        if torch.cuda.is_available():
            torch.cuda.synchronize()

        self.load_model()

    def predict_tiled(
        self,
        image: rasterio.DatasetReader,
        skip_empty: Optional[bool] = True,
        warm_start: Optional[bool] = False,
    ) -> ProcessedResult:
        """Run inference on an image using tiling. Outputs a ProcessedResult

        Args:
            image (rasterio.DatasetReader): Image
            skip_empty (bool, optional): Skip empty/all-black images. Defaults to True.
            warm_start (bool, option): Whether or not to continue from where one left off
                                       Defaults to False.

        Returns:
            ProcessedResult: A list of predictions and the bounding boxes for those detections.

        Raises:
            ValueError: If the dataloader is empty
        """

        dataloader = dataloader_from_image(
            image,
            tile_size_px=self.config.data.tile_size,
            overlap_px=self.config.data.tile_overlap,
            gsd_m=self.config.data.gsd,
            batch_size=self.config.model.batch_size,
        )

        if len(dataloader) == 0:
            raise ValueError("No tiles to process")

        if self.post_processor is not None:
            logger.debug("Initialising post processor")
            self.post_processor.initialise(image, warm_start=warm_start)

        progress_bar = tqdm(enumerate(dataloader), total=len(dataloader))
        self.should_exit = False

        if not warm_start:
            self.post_processor.setup_cache()

        # Predict on each tile
        processed_tiles = 0
        for _, batch in progress_bar:
            if self.should_exit:
                break

            if self.should_reload:
                self.attempt_reload()

            # Skip images that are all black or all white - do this first?
            if skip_empty:
                filtered = defaultdict(list)
                for idx, im in enumerate(batch["image"]):
                    img_mean = im.mean()

                    if img_mean <= 1 or img_mean >= 254:
                        continue

                    for key in batch:
                        filtered[key].append(batch[key][idx])

                batch = filtered
                if len(batch["image"]) == 0:
                    progress_bar.set_postfix_str(f"Empty batch, skipping.")
                    continue

            processed_tiles += len(batch["image"])

            if (
                processed_tiles <= self.post_processor.tile_count and warm_start
            ):  # already done
                progress_bar.set_postfix_str(
                    f"Processed batch, skipping - valid tile count {processed_tiles}"
                )
            else:
                predictions = [p.to("cpu") for p in self.predict(batch["image"])]

                # Typically if this happens we hit an OOM...
                if predictions is None:
                    progress_bar.set_postfix_str("Error")
                    logger.error("Failed to run inference on image.")
                    self.failed_images.add(image)
                else:
                    # Run the post-processor.
                    batch["predictions"] = predictions
                    self.on_after_predict(batch)

                    # Logging
                    process = psutil.Process(os.getpid())
                    cpu_mem_usage_gb = process.memory_info().rss / 1073741824

                    pbar_string = f"#objs: {len(predictions)}"

                    if "cuda" in self.device and torch.cuda.is_available():
                        _, used_memory_b = torch.cuda.mem_get_info()
                        gpu_mem_usage_gb = used_memory_b / 1073741824
                        pbar_string += f", GPU: {gpu_mem_usage_gb:1.2f}G"

                    pbar_string += f", CPU: {cpu_mem_usage_gb:1.2f}G"
                    pbar_string += f", t_pred: {self.t_predict:1.2f}s"
                    pbar_string += f", t_post: {self.t_postprocess:1.2f}s"

                    progress_bar.set_postfix_str(pbar_string)

        return self.post_process()

    def predict(self, image: List[Union[str, torch.Tensor, DatasetReader]]) -> Any:
        """Run inference on an image file, rasterio dataset or Tensor.

        Args:
            image (Union[str, Tensor, DatasetReader]): List of (Path to image, or, float tensor
                                              in CHW order, un-normalised)

        Returns:
            Any: Raw prediction results

        Raises:
            NotImplementedError: If the image type is not supported


        """

        t_start = time.time()

        if not isinstance(image, list):
            image = [image]

        image_tensor = [image_to_tensor(i) for i in image]

        if self.model is None:
            self.load_model()

        res = self.predict_batch(image_tensor)

        self.t_predict = time.time() - t_start

        return res

    @abstractmethod
    def predict_batch(self, image_tensor: List[torch.Tensor]) -> Any:
        """Run inference on a batch of tensors"""

__init__(config)

Parameters:

Name Type Description Default
config DictConfig

Configuration dictionary

 required
Source code in src/tcd_pipeline/models/model.py 
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
def __init__(self, config: DictConfig):
    """
    Args:
        config (DictConfig): Configuration dictionary

    """

    self.config = config

    if config.model.device == "cuda" and not torch.cuda.is_available():
        logger.warning("Failed to use CUDA, falling back to CPU")
        self.device = "cpu"
    else:
        self.device = config.model.device

    self.model = None
    self.should_reload = False
    self.post_processor: PostProcessor = None
    self.failed_images = set()
    self.should_exit = False

    logger.info("Device: %s", self.device)
    self.setup()

attempt_reload()

Attempts to reload the model.

Source code in src/tcd_pipeline/models/model.py 
 98
 99
100
101
102
103
104
105
106
107
108
def attempt_reload(self) -> None:
    """Attempts to reload the model."""
    if "cuda" not in self.device:
        return

    del self.model

    if torch.cuda.is_available():
        torch.cuda.synchronize()

    self.load_model()

load_model() abstractmethod

Load the model, defined by subclass

Source code in src/tcd_pipeline/models/model.py 
62
63
64
@abstractmethod
def load_model(self):
    """Load the model, defined by subclass"""

on_after_predict(results)

Append tiled results to the post processor, or cache

Parameters:

Name Type Description Default
results list

Prediction results from one tile

 required
Source code in src/tcd_pipeline/models/model.py 
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
def on_after_predict(self, results: dict) -> None:
    """Append tiled results to the post processor, or cache

    Args:
        results (list): Prediction results from one tile

    """

    t_start = time.time()

    # Invert dict-of-lists to list-of-dicts
    results = [dict(zip(results, t)) for t in zip(*results.values())]

    self.post_processor.add(results)

    self.t_postprocess = time.time() - t_start

post_process()

Run post-processing to merge results

Returns:

Name Type Description
ProcessedResult ProcessedResult

merged results
Source code in src/tcd_pipeline/models/model.py 
83
84
85
86
87
88
89
90
91
92
93
94
95
96
def post_process(self) -> ProcessedResult:
    """Run post-processing to merge results

    Returns:
        ProcessedResult: merged results
    """

    res = self.post_processor.process()

    if self.config.postprocess.cleanup:
        logger.info("Cleaning up post processor")
        self.post_processor.cache.clear()

    return res

predict(image)

Run inference on an image file, rasterio dataset or Tensor.

Parameters:

Name Type Description Default
image Union[str, Tensor, DatasetReader]

List of (Path to image, or, float tensor in CHW order, un-normalised)

 required

Returns:

Name Type Description
Any Any

Raw prediction results

Raises:

Type Description
NotImplementedError

If the image type is not supported
Source code in src/tcd_pipeline/models/model.py 
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
def predict(self, image: List[Union[str, torch.Tensor, DatasetReader]]) -> Any:
    """Run inference on an image file, rasterio dataset or Tensor.

    Args:
        image (Union[str, Tensor, DatasetReader]): List of (Path to image, or, float tensor
                                          in CHW order, un-normalised)

    Returns:
        Any: Raw prediction results

    Raises:
        NotImplementedError: If the image type is not supported


    """

    t_start = time.time()

    if not isinstance(image, list):
        image = [image]

    image_tensor = [image_to_tensor(i) for i in image]

    if self.model is None:
        self.load_model()

    res = self.predict_batch(image_tensor)

    self.t_predict = time.time() - t_start

    return res

predict_batch(image_tensor) abstractmethod

Run inference on a batch of tensors

Source code in src/tcd_pipeline/models/model.py 
250
251
252
@abstractmethod
def predict_batch(self, image_tensor: List[torch.Tensor]) -> Any:
    """Run inference on a batch of tensors"""

predict_tiled(image, skip_empty=True, warm_start=False)

Run inference on an image using tiling. Outputs a ProcessedResult

Parameters:

Name Type Description Default
image DatasetReader

Image

 required
skip_empty bool

Skip empty/all-black images. Defaults to True.

 True
warm_start (bool, option)

Whether or not to continue from where one left off Defaults to False.

 False

Returns:

Name Type Description
ProcessedResult ProcessedResult

A list of predictions and the bounding boxes for those detections.

Raises:

Type Description
ValueError

If the dataloader is empty
Source code in src/tcd_pipeline/models/model.py 
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
def predict_tiled(
    self,
    image: rasterio.DatasetReader,
    skip_empty: Optional[bool] = True,
    warm_start: Optional[bool] = False,
) -> ProcessedResult:
    """Run inference on an image using tiling. Outputs a ProcessedResult

    Args:
        image (rasterio.DatasetReader): Image
        skip_empty (bool, optional): Skip empty/all-black images. Defaults to True.
        warm_start (bool, option): Whether or not to continue from where one left off
                                   Defaults to False.

    Returns:
        ProcessedResult: A list of predictions and the bounding boxes for those detections.

    Raises:
        ValueError: If the dataloader is empty
    """

    dataloader = dataloader_from_image(
        image,
        tile_size_px=self.config.data.tile_size,
        overlap_px=self.config.data.tile_overlap,
        gsd_m=self.config.data.gsd,
        batch_size=self.config.model.batch_size,
    )

    if len(dataloader) == 0:
        raise ValueError("No tiles to process")

    if self.post_processor is not None:
        logger.debug("Initialising post processor")
        self.post_processor.initialise(image, warm_start=warm_start)

    progress_bar = tqdm(enumerate(dataloader), total=len(dataloader))
    self.should_exit = False

    if not warm_start:
        self.post_processor.setup_cache()

    # Predict on each tile
    processed_tiles = 0
    for _, batch in progress_bar:
        if self.should_exit:
            break

        if self.should_reload:
            self.attempt_reload()

        # Skip images that are all black or all white - do this first?
        if skip_empty:
            filtered = defaultdict(list)
            for idx, im in enumerate(batch["image"]):
                img_mean = im.mean()

                if img_mean <= 1 or img_mean >= 254:
                    continue

                for key in batch:
                    filtered[key].append(batch[key][idx])

            batch = filtered
            if len(batch["image"]) == 0:
                progress_bar.set_postfix_str(f"Empty batch, skipping.")
                continue

        processed_tiles += len(batch["image"])

        if (
            processed_tiles <= self.post_processor.tile_count and warm_start
        ):  # already done
            progress_bar.set_postfix_str(
                f"Processed batch, skipping - valid tile count {processed_tiles}"
            )
        else:
            predictions = [p.to("cpu") for p in self.predict(batch["image"])]

            # Typically if this happens we hit an OOM...
            if predictions is None:
                progress_bar.set_postfix_str("Error")
                logger.error("Failed to run inference on image.")
                self.failed_images.add(image)
            else:
                # Run the post-processor.
                batch["predictions"] = predictions
                self.on_after_predict(batch)

                # Logging
                process = psutil.Process(os.getpid())
                cpu_mem_usage_gb = process.memory_info().rss / 1073741824

                pbar_string = f"#objs: {len(predictions)}"

                if "cuda" in self.device and torch.cuda.is_available():
                    _, used_memory_b = torch.cuda.mem_get_info()
                    gpu_mem_usage_gb = used_memory_b / 1073741824
                    pbar_string += f", GPU: {gpu_mem_usage_gb:1.2f}G"

                pbar_string += f", CPU: {cpu_mem_usage_gb:1.2f}G"
                pbar_string += f", t_pred: {self.t_predict:1.2f}s"
                pbar_string += f", t_post: {self.t_postprocess:1.2f}s"

                progress_bar.set_postfix_str(pbar_string)

    return self.post_process()

setup() abstractmethod

Perform any setup actions as needed

Source code in src/tcd_pipeline/models/model.py 
58
59
60
@abstractmethod
def setup(self):
    """Perform any setup actions as needed"""