Do batch prediction.

Do batch prediction inside the `predict` method instead of calling
`predict` once for each image.

app.py

@@ -5,6 +5,7 @@ from dotenv import load_dotenv
 from autotagger import Autotagger
 from base64 import b64encode
 from flask import Flask, request, render_template, jsonify
+from fastai.vision.core import PILImage
 
 load_dotenv()
 model_path = getenv("MODEL_PATH", "models/model.pth")
@@ -21,15 +22,21 @@ def index():
 @app.route("/evaluate", methods=["POST"])
 def evaluate():
     files = request.files.getlist("file")
+    images = [PILImage.create(file) for file in files]
     threshold = float(request.form.get("threshold", 0.1))
     output = request.form.get("format", "html")
     limit = int(request.form.get("limit", 50))
 
+    predictions = autotagger.predict(images, threshold=threshold, limit=limit)
+
     if output == "html":
-        predictions = [{ "data": b64encode(data).decode(), "tags": autotagger.predict(data, threshold=threshold, limit=limit) } for data in (file.stream.read() for file in files)]
-        return render_template("evaluate.html", predictions=predictions)
+        for file in files:
+            file.seek(0)
+
+        base64data = [b64encode(file.read()).decode() for file in files]
+        return render_template("evaluate.html", predictions=zip(base64data, predictions))
     elif output == "json":
-        predictions = [{ "filename": file.filename, "tags": autotagger.predict(file.read(), threshold=threshold, limit=limit) } for file in files]
+        predictions = [{ "filename": file.filename, "tags": tags } for file, tags in zip(files, predictions)]
         return jsonify(predictions)
     else:
         return 400

autotag

@@ -12,8 +12,12 @@ from fastai.vision.core import PILImage
 @click.argument("file", nargs=-1, type=click.File("rb"), required=True)
 def main(file, threshold, limit, model):
     autotagger = Autotagger(model)
-    predictions = [{ "filename": f.name, "tags": autotagger.predict(PILImage.create(f), threshold=threshold, limit=limit) } for f in file]
-    click.echo(json.dumps(predictions, indent=2))
+    images = [PILImage.create(f) for f in file]
+    predictions = autotagger.predict(images, threshold=threshold, limit=limit)
+
+    for i, tags in enumerate(predictions):
+        data = { "filename": file[i].name, "tags": tags }
+        click.echo(json.dumps(data))
 
 if __name__ == "__main__":
     main()

autotagger/autotagger.py

@@ -1,5 +1,5 @@
 from fastbook import *
-from pandas import read_csv
+from pandas import DataFrame, read_csv
 import timm
 
 class Autotagger:
@@ -26,10 +26,13 @@ class Autotagger:
 
         return learn
 
-    def predict(self, path, threshold=0.01, limit=50):
+    def predict(self, images, threshold=0.01, limit=50, bs=64):
         with self.learn.no_bar(), self.learn.no_logging():
-            pred = self.learn.predict(path)
-            scores = [score.item() for score in pred[2]]
-            results = { tag : score for (tag, score) in zip(self.learn.dls.vocab, scores) if score >= threshold }
-            results = sorted(results.items(), key = lambda x: -x[1])
-            return dict(results[:limit])
+            dl = self.learn.dls.test_dl(images, bs=bs)
+            batch, _ = self.learn.get_preds(dl=dl)
+
+            for scores in batch:
+                df = DataFrame({ "tag": self.learn.dls.vocab, "score": scores })
+                df = df[df.score >= threshold].sort_values("score", ascending=False).head(limit)
+                tags = dict(zip(df.tag, df.score))
+                yield tags

templates/evaluate.html

@@ -9,15 +9,15 @@
     <a class="text-xs text-sky-600 hover:text-sky-500 mr-4" href="/">&lt; Back</a>
 
     <div class="mt-4">
-      {% for prediction in predictions %}
+      {% for base64data, tags in predictions %}
         <div class="flex flex-col p-2 gap-2 border rounded md:flex-row md:max-h-[80vh]">
           <div class="flex-1 flex items-center justify-center">
-            <img class="max-w-full max-h-full h-auto" src="data:image/jpg;base64,{{ prediction["data"] | safe }}">
+            <img class="max-w-full max-h-full h-auto" src="data:image/jpg;base64,{{ base64data | safe }}">
           </div>
 
           <div class="flex-0 overflow-scroll md:pr-2">
             <table class="w-full leading-4">
-              {% for tag, score in prediction["tags"].items() %}
+              {% for tag, score in tags.items() %}
               <tr>
                 <td>
                   <a class="text-sky-600 hover:text-sky-500" href="https://danbooru.donmai.us/wiki_pages/{{ tag | urlencode }}">?</a>
@@ -28,7 +28,7 @@
               {% endfor %}
             </table>
 
-            <textarea class="w-full text-gray-500 mt-2" rows="4">{{ " ".join(prediction["tags"].keys()) }}</textarea>
+            <textarea class="w-full text-gray-500 mt-2" rows="4">{{ " ".join(tags.keys()) }}</textarea>
           </div>
         </div>
       {% endfor %}