more progress

5_clean_ground_truth.py

@@ -1,4 +1,10 @@
 import fiftyone as fo
 
+"""
+This almost all happens in the client (at least it did the first time). Will need to do this again so will
+update doc here as appropriate
+"""
+
+
 if __name__ == '__main__':
     fo.load_dataset("labeled_dataset")

7_new_predictions.py

@@ -1,7 +1,7 @@
 import fiftyone as fo
 import tempfile
 import torch
-from safetensors.torch import save_model
+from fiftyone import Classification
 from ultralytics import YOLO
 
 """
@@ -33,8 +33,11 @@ def load_data(datasetname) -> fo.Dataset:
 
 def split_again(dataset_with_orig, first_training) -> fo.Dataset:
     # Take a random sample from the original data, excluding the first set of sample data
-    # We
-    return dataset_with_orig.exclude(first_training).take(SAMPLES_TO_TAKE)
+    # sample ID is not generally a good unique ID. For 51, filepath is what you should use when working
+    # with the same dataset over and over again
+    if "second_play_photos" in fo.list_datasets():
+        fo.delete_dataset("second_play_photos")
+    return dataset_with_orig.exclude_by("filepath", first_training.values("filepath")).take(SAMPLES_TO_TAKE).clone("second_play_photos", persistent=True)
 
 def run_predictions(dataset):
     model = YOLO(MODEL_LOCATION)  # load a custom model
@@ -50,7 +53,7 @@ def run_predictions(dataset):
         device=get_torch_device(),
         imgsz=IMAGE_SIZE,
         stream=True,
-        save=True,
+        # save=True,
         project="predictions_round2",
         name="write_something"
     )
@@ -68,19 +71,36 @@ def run_predictions(dataset):
     # return [results, naive_results]
     return results
 
+def extract_orig_path(dataset):
+    sample = dataset.first()
+    result = sample.filepath.replace(sample.filename, "")
+    return result
+
 
 if __name__ == '__main__':
     print("starting")
     whole_dataset = load_data("photo_album")
+    original_path = extract_orig_path(whole_dataset)
     first_training = load_data("labeled_dataset")
     candidate_data = split_again(whole_dataset,first_training)
     print("about to predict")
     results = run_predictions(candidate_data)
-    for r in results:
-        # we save the data to a 51 dataset. Now that we are done exploring, I think we might want to clone above.
-       r.save()
-    # for r2 in results[1]:
-    #     r2.save()
+
+    # Convert the results list into a dict so we can iterate through the dataset rather than
+    # the results. This will allow for faster saves to the dataset rather than individual sample by sample saves
+    # Using the images name without the path as the key
+    results_dict = {x.path[x.path.rfind("/")+1:]:x for x in results}
+
+    # Now for each sample in the new dataset, add the prediction
+    for sample in candidate_data.iter_samples(progress=True, autosave=True):
+        filename = sample.filename
+        res = results_dict[filename]
+        predicted_class = Classification(label=res.names[res.probs.top1], confidence=round(res.probs.top1conf.item(), 2))
+        sample["prediction"] = predicted_class
+
+    # Display new dataset and hold it open with a wait()
+    session = fo.launch_app(candidate_data)
+    session.wait()
     print("Done")
 
 

8_cleaning_ground_truth_round2.py

@@ -0,0 +1,56 @@
+import fiftyone as fo
+import argparse
+
+""" 
+This script needs to accept arguments because we don't want to do all the tasks every time we run it.
+* at the end means it happens in the app, not here in code
+So what we do is:
+0. load in the new dataset - "second_play_photos"
+1. Create a new field for ground truth
+2. One at a time, go through each of the categories and just view that category
+3. If there is a mislabel, then fix it in the ground truth field. 
+4. Once all the images have been fixed for that category, head on to the next category
+5. When finished all the categories, anything that has a blank ground truth field should have the ground
+truth field set = to the prediction field. 
+Then we are ready to do our final training/test
+Then final predictions
+
+To do the ground_truth updates use the console
+import fiftyone as fo
+dataset = fo.load_dataset("second_play_photos")
+session = fo.launch_app(dataset)
+
+"""
+DATASET_NAME = "second_play_photos"
+
+if __name__ == '__main__':
+    print("starting")
+
+    dataset = fo.load_dataset(DATASET_NAME)
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-g", help="specify to create ground truth field", action="store_true")
+    parser.add_argument("-p", help="specify to move correct predictions to ground truth field",
+                        action="store_true")
+    args = parser.parse_args()
+
+    if args.g:
+        print("make the field")
+        if dataset.has_field("ground_truth"):
+            dataset.delete_sample_field("ground_truth")
+        dataset.add_sample_field(
+            "ground_truth",
+            fo.EmbeddedDocumentField,
+            embedded_doc_type=fo.Classification,
+        )
+        dataset.save()
+    elif args.p:
+        for sample in dataset.iter_samples(autosave=True):
+            if sample["ground_truth"] is None:
+                sample["ground_truth"] = sample["prediction"]
+
+    else:
+        print("you need to specify either -g or -p")
+
+    print("done")
+

README.md

@@ -13,4 +13,24 @@ Here are the general steps we want to accomplish
 
 ## Order to run things
 1. First run make_dataset.py to make the dataset in FiftyOne
-2. Now run make_embeddings.py to create all the embeddings
+2. Now run make_embeddings.py to create all the embeddings
+
+#### The classes I am going for
+"boy"
+"girl"
+"man"
+"woman"
+"people"
+"dog"
+"cat"
+"bird"
+"insect"
+"monkey"
+"crustacean",
+"fish"
+"animal"
+"plant"
+"flower"
+"landscape"
+"architecture"
+"not an animal, plant, landscape, person, or building"