Adding protests analyser (#156)

* Topic/tutorial (#147)

* complete tutorial 2

* complete tutorial 3, restructure main readme

* lint

* update commands

* Adding prostests analyser

* Fixing requirements.txt and useless imports

* Linted core and test scripts

* Fixed requirements syntax

* Fixed syntax errors in .yaml, moved utils.py into core.

* Added absolute path for model

* Return predictions as float instead of float32

* Linted core

* Compatibility with imagededup

* Automatic download of the model

* Deleted the now invalid api key

* Update README.md

* Update README.md

* Update README.md

* Update README.md

* Adding prostests analyser

* Fixing requirements.txt and useless imports

* Linted core and test scripts

* Fixed requirements syntax

* Fixed syntax errors in .yaml, moved utils.py into core.

* Added absolute path for model

* Return predictions as float instead of float32

* Linted core

* Compatibility with imagededup

* Automatic download of the model

* Deleted the now invalid api key

* fix import from commit

* rm empty spaces

* black fmt fix

* black fmt fix 2

Co-authored-by: Lachlan Kermode <[email protected]>

.env.example

@@ -1 +1 @@
-GOOGLE_API_KEY=my_key_from_gcp_project
+GOOGLE_API_KEY=

src/lib/analysers/ProtestsPretrained/core.py

@@ -0,0 +1,70 @@
+import numpy as np
+import sys
+import json
+import os
+import torch
+from torch.autograd import Variable
+from PIL import Image
+
+from lib.common.analyser import Analyser
+from lib.common.etypes import Etype, Union, Array
+from lib.analysers.ProtestsPretrained.utils import transform, modified_resnet50, decode
+
+PTH_TAR = "/mtriage/model.pth.tar"
+
+# TODO cuda ?
+
+
+class ProtestsPretrained(Analyser):
+    def pre_analyse(self, config):
+        """
+        Init the logging, etc
+        Init the model
+        """
+        rLabels = config["labels"]
+        self.THRESH = 0.0
+
+        t = transform()
+        model = modified_resnet50()
+        model.load_state_dict(
+            torch.load(
+                PTH_TAR,
+                map_location=torch.device("cpu"),
+            )["state_dict"]
+        )
+        model.eval()
+
+        def get_preds(img_path):
+            """
+            Gives labelds and probabilities for a single image
+            This is were we preprocess the image, using a function defined in the model class
+            """
+            # load image
+            img = Image.open(img_path).convert("RGB")
+            # process it
+            x = t(img)
+            # get in in the right format
+            x = Variable(x).unsqueeze(0)
+            # predictions
+            output = model(x)
+            # decode
+            output = decode(output.cpu().data.numpy()[0])
+            # filter
+            output = [(x[0], x[1]) for x in output if x[0] in rLabels]
+            output = [(x[0], float(x[1])) for x in output if x[1] >= self.THRESH]
+
+            return output
+
+        self.get_preds = get_preds
+
+    def analyse_element(
+        self, element: Union(Array(Etype.Image), Etype.Json), _
+    ) -> Etype.Json:
+        self.logger(f"Running inference on frames in {element.id}...")
+        val = Etype.CvJson.from_preds(element, self.get_preds)
+        self.logger(f"Wrote predictions JSON for {element.id}.")
+        self.disk.delete_local_on_write = True
+        return val
+
+
+module = ProtestsPretrained

src/lib/analysers/ProtestsPretrained/info.yaml

@@ -0,0 +1,6 @@
+desc: Classify the presence of protests and violence in images.
+args:
+   - name: labels
+     desc: Filter results to a limited array of labels.
+     required: true
+     input: whitelist

src/lib/analysers/ProtestsPretrained/partial.Dockerfile

@@ -0,0 +1 @@
+RUN wget -O /mtriage/model.pth.tar https://www.dropbox.com/s/vgh2nwxrzembxpw/model.pth.tar?dl=0

src/lib/analysers/ProtestsPretrained/requirements.txt

@@ -0,0 +1,4 @@
+torch==1.4
+torchvision==0.5
+pillow==6.1.0
+numpy<1.17

src/lib/analysers/ProtestsPretrained/test.py

@@ -0,0 +1,48 @@
+import numpy as np
+import os
+import torch
+from torch.autograd import Variable
+from PIL import Image
+from utils import transform, modified_resnet50, decode
+
+
+def pre_analyse():
+    """
+    Init the logging, etc
+    Init the model
+    Same as KerasPretrained
+    """
+    t = transform()
+    model = modified_resnet50()
+    model.load_state_dict(
+        torch.load(
+            "model.pth.tar",
+            map_location=torch.device("cpu"),
+        )["state_dict"]
+    )
+    model.eval()
+
+    def get_preds(img_path):
+        """
+        Gives labelds and probabilities for a single image
+        This is were we preprocess the image, using a function defined in the model class
+        """
+        # load image
+        img = Image.open(img_path).convert("RGB")
+        # process it
+        x = t(img)
+        # get in in the right format
+        x = Variable(x).unsqueeze(0)
+        # predictions
+        output = model(x)
+        # decode
+        output = decode(output.cpu().data.numpy()[0])
+
+        # filter
+        # return pred, proba
+        return output
+
+    return get_preds("image.jpg")
+
+
+print(pre_analyse())

src/lib/analysers/ProtestsPretrained/utils.py

@@ -0,0 +1,56 @@
+"""
+created by: Donghyeon Won
+"""
+import torch
+import torch.nn as nn
+import torchvision.transforms as transforms
+import torchvision.models as models
+
+
+def transform():
+    return transforms.Compose(
+        [
+            transforms.Resize(256),
+            transforms.CenterCrop(224),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
+        ]
+    )
+
+
+def decode(preds):
+    classes = [
+        "protest",
+        "violence",
+        "sign",
+        "photo",
+        "fire",
+        "police",
+        "children",
+        "group_20",
+        "group_100",
+        "flag",
+        "night",
+        "shouting",
+    ]
+    return [(x, preds[c]) for c, x in enumerate(classes)]
+
+
+class FinalLayer(nn.Module):
+    """modified last layer for resnet50 for our dataset"""
+
+    def __init__(self):
+        super(FinalLayer, self).__init__()
+        self.fc = nn.Linear(2048, 12)
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x):
+        out = self.fc(x)
+        out = self.sigmoid(out)
+        return out
+
+
+def modified_resnet50():
+    model = models.resnet50(pretrained=True)
+    model.fc = FinalLayer()
+    return model

src/lib/common/analyser.py

@@ -1,6 +1,7 @@
 import glob
 import os
 import shutil
+import traceback
 from abc import ABC, abstractmethod
 from pathlib import Path
 from typing import Generator, List, Union, Tuple
@@ -196,3 +197,4 @@ def __attempt_analyse(self, attempts, element):
                 raise e
             else:
                 self.error_logger(f"{str(e)}: skipping element", element)
+                print(traceback.format_exc())