Add github files

README.md

@@ -1,12 +1,73 @@
----
-title: CLIP Based NSFW Detector
-emoji: 📊
-colorFrom: gray
-colorTo: indigo
-sdk: gradio
-sdk_version: 3.23.0
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# CLIP-based-NSFW-Detector
+
+This 2 class NSFW-detector is a lightweight Autokeras model that takes CLIP ViT L/14 embbedings as inputs.
+It estimates a value between 0 and 1 (1 = NSFW) and works well with embbedings from images.
+
+DEMO-Colab:
+https://colab.research.google.com/drive/19Acr4grlk5oQws7BHTqNIK-80XGw2u8Z?usp=sharing
+
+The training CLIP V L/14 embbedings can be downloaded here:
+https://drive.google.com/file/d/1yenil0R4GqmTOFQ_GVw__x61ofZ-OBcS/view?usp=sharing (not fully manually annotated so cannot be used as test)
+
+
+The (manually annotated) test set is there https://github.com/LAION-AI/CLIP-based-NSFW-Detector/blob/main/nsfw_testset.zip
+
+https://github.com/rom1504/embedding-reader/blob/main/examples/inference_example.py inference on laion5B
+
+Example of use of the model:
+
+```python
+@lru_cache(maxsize=None)
+def load_safety_model(clip_model):
+    """load the safety model"""
+    import autokeras as ak  # pylint: disable=import-outside-toplevel
+    from tensorflow.keras.models import load_model  # pylint: disable=import-outside-toplevel
+
+    cache_folder = get_cache_folder(clip_model)
+
+    if clip_model == "ViT-L/14":
+        model_dir = cache_folder + "/clip_autokeras_binary_nsfw"
+        dim = 768
+    elif clip_model == "ViT-B/32":
+        model_dir = cache_folder + "/clip_autokeras_nsfw_b32"
+        dim = 512
+    else:
+        raise ValueError("Unknown clip model")
+    if not os.path.exists(model_dir):
+        os.makedirs(cache_folder, exist_ok=True)
+
+        from urllib.request import urlretrieve  # pylint: disable=import-outside-toplevel
+
+        path_to_zip_file = cache_folder + "/clip_autokeras_binary_nsfw.zip"
+        if clip_model == "ViT-L/14":
+            url_model = "https://raw.githubusercontent.com/LAION-AI/CLIP-based-NSFW-Detector/main/clip_autokeras_binary_nsfw.zip"
+        elif clip_model == "ViT-B/32":
+            url_model = (
+                "https://raw.githubusercontent.com/LAION-AI/CLIP-based-NSFW-Detector/main/clip_autokeras_nsfw_b32.zip"
+            )
+        else:
+            raise ValueError("Unknown model {}".format(clip_model))  # pylint: disable=consider-using-f-string
+        urlretrieve(url_model, path_to_zip_file)
+        import zipfile  # pylint: disable=import-outside-toplevel
+
+        with zipfile.ZipFile(path_to_zip_file, "r") as zip_ref:
+            zip_ref.extractall(cache_folder)
+
+    loaded_model = load_model(model_dir, custom_objects=ak.CUSTOM_OBJECTS)
+    loaded_model.predict(np.random.rand(10**3, dim).astype("float32"), batch_size=10**3)
+
+    return loaded_model
+    
+    
+nsfw_values = safety_model.predict(embeddings, batch_size=embeddings.shape[0])
+```
+
+This code and model is released under the MIT license:
+
+Copyright 2022, Christoph Schuhmann
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

clip_autokeras_binary_nsfw.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:905520bb0336accc73ee282c605d926aef90cf14248889d44b9cee50eef71f8a
+size 1094585

clip_autokeras_nsfw_b32.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:78715ab9d64485167e8c9ade3b8e6761a17a16c24d114bd47431ac1dab46462b
+size 231831

docs/train_h14_nsfw.md

@@ -0,0 +1,176 @@
+This is a basic guide that outlines what is necessary to train a NSFW detector
+on top of a CLIP model.
+
+Note that this is a general guide and meant to be used as a guideline.
+
+
+## Dataset Prep
+You will need to obtain thousands of NSFW and SFW images to train a good model.
+
+Once the images are gathered, you can place them in a directory and leverage
+[clip-retrieval](https://github.com/rom1504/clip-retrieval#clip-inference)
+to easily create a numpy array of embeddings. (For more details read the docs)
+
+Finally, you will need to create target values which can be done easily in the
+python interpreter.
+
+### Target Values & Dataset Combination
+
+```python
+# find out how many positive samples you have
+pos_x = np.load(path/to_positive/samples.npy)
+neg_x = np.load(path/to_negative/samples.npy)
+
+num_pos = pos_x.shape[0]
+num_neg = neg_x.shape[0]
+
+# create target values
+pos_y = np.ones((num_pos, 1))
+neg_y = np.zeros((num_neg, 1))
+
+# combine the x samples
+# NOTE: we will rely on torch dataloader shuffling to break the ordering here
+x = np.vstack((pos_x, neg_x))
+y = np.vstack((pos_y, neg_y))
+
+# save the dataset x & y
+
+np.save("train_x.npy", x)
+np.save("train_y.npy", y)
+```
+
+## Model Training
+
+Thankfully it is possible to use a very simple linear model to train the NSFW
+detector.
+
+For the purposes of this guide we will reference [this repo](https://github.com/christophschuhmann/improved-aesthetic-predictor)
+and its model architecture.
+
+> NOTE: It is also possible to utilize the training script provided in that repo
+> as boilerplate code, provided you have `.npy` files for your dataset's x & y
+
+### Model Architecture
+
+Feel free to tweak the model architecture here, but the important thing to
+remember is that your input dimension should match the dimension of your CLIP
+embeddings, and your output dimension should be 1.
+
+```python
+import torch.nn as nn
+
+class H14_NSFW_Detector(nn.Module):
+    def __init__(self, input_size=1024):
+        super().__init__()
+        self.input_size = input_size
+        self.layers = nn.Sequential(
+            nn.Linear(self.input_size, 1024),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(1024, 2048),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(2048, 1024),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(1024, 256),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(256, 128),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(128, 16),
+            nn.Linear(16, 1)
+        )
+
+    def forward(self, x):
+        return self.layers(x)
+```
+
+### Training Snippets
+
+Below is a list of snippets that should walk you through the general steps
+necessary to train a MLP in PyTorch, however, it is completely fine to replace
+the existing MLP in [this-repo](https://github.com/christophschuhmann/improved-aesthetic-predictor)
+with the model provided above & begin training.
+
+For those of you who wish to build out custom code, the following snippets should
+get the ball rolling...
+
+#### Import the necessary libraries:
+
+```python
+import torch
+import torch.nn as nn
+from torch.optim import Adam
+from torch.utils.data import TensorDataset, DataLoader
+```
+
+#### Define the dataset and data loaders:
+
+```python
+# Define the dataset
+x = torch.from_numpy(np.load("train_x.npy"))
+y = torch.from_numpy(np.load("train_y.npy"))
+train_dataset = TensorDataset(x,y)
+
+# Define the data loaders
+train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=32, shuffle=True)
+```
+
+#### Initialize the model
+```python
+model = H14_NSFW_Detector()
+```
+
+#### Define the loss function
+```python
+criterion = nn.MSELoss()
+```
+
+#### Define the optimizer
+```python
+# Define the optimizer
+optimizer = Adam(model.parameters())
+```
+
+#### Define the training loop
+```python
+# Define the number of epochs
+num_epochs = 10
+
+# Training loop
+for epoch in range(num_epochs):
+    for inputs, labels in train_loader:
+        # Clear gradients
+        optimizer.zero_grad()
+
+        # Forward pass
+        outputs = model(inputs)
+
+        # Compute the loss
+        loss = criterion(outputs, labels)
+
+        # Backward pass and optimization
+        loss.backward()
+        optimizer.step()
+```
+
+#### Define an evaluation loop
+```python
+# Evaluation loop
+model.eval()
+with torch.no_grad():
+    for inputs, labels in val_loader:
+        # Forward pass
+        outputs = model(inputs)
+
+        # Compute the loss
+        loss = criterion(outputs, labels)
+```
+
+Note that this is a basic guide and you may need to add additional functionality
+such as model saving and loading, early stopping, etc.
+You may want to adjust the learning rate, 
+batch size, and number of epochs as well.
+

h14_nsfw.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed87167db4dd6676f921d37f3f1b26a1f32275724cb5e7925017277f435ca78e
+size 22182219

h14_nsfw_model.py

@@ -0,0 +1,28 @@
+import torch.nn as nn
+
+class H14_NSFW_Detector(nn.Module):
+    def __init__(self, input_size=1024):
+        super().__init__()
+        self.input_size = input_size
+        self.layers = nn.Sequential(
+            nn.Linear(self.input_size, 1024),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(1024, 2048),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(2048, 1024),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(1024, 256),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(256, 128),
+            nn.ReLU(),
+            nn.Dropout(0.2),
+            nn.Linear(128, 16),
+            nn.Linear(16, 1)
+        )
+
+    def forward(self, x):
+        return self.layers(x)

license.md

@@ -0,0 +1,9 @@
+This code and model is released under the MIT license:
+
+Copyright 2022, Christoph Schuhmann
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

nsfw-clip.py

@@ -0,0 +1,75 @@
+
+import time
+
+import autokeras as ak
+
+import tensorflow as tf
+
+
+
+
+import numpy as np
+
+neutral = np.load ("./neutral/img_emb/img_emb_0.npy")
+print(neutral.shape)
+
+porn = np.load ("./porn/img_emb/img_emb_0.npy")
+print(porn.shape)
+
+drawings = np.load ("./drawings/img_emb/img_emb_0.npy")
+print(drawings.shape)
+
+hentai = np.load ("./hentai/img_emb/img_emb_0.npy")
+print(hentai.shape)
+
+sexy = np.load ("./sexy/img_emb/img_emb_0.npy")
+print(sexy.shape)
+
+
+
+x_t =np.concatenate((porn,sexy),axis = 0)
+x_t =np.concatenate((x_t,hentai),axis = 0)
+nsfw_t_len=x_t.shape[0]
+print(nsfw_t_len)
+x_t =np.concatenate((x_t,neutral),axis = 0)
+x_t =np.concatenate((x_t,drawings),axis = 0)
+y_t = np.zeros(x_t.shape[0], dtype = np.uint8)
+sfw_t_len=x_t.shape[0] - nsfw_t_len
+print(sfw_t_len)
+
+for i in range(nsfw_t_len):
+  y_t[i]=1
+from sklearn.utils import shuffle
+x_train, y_train = shuffle(x_t, y_t)
+
+
+print(y_t)
+print(y_train)
+
+
+
+
+
+
+
+x_train = x_train.astype(float) #[100:-100]
+y_train = y_train.astype(int)#[100:-100]
+
+
+#x_test = x_test.astype(float) #[100:-100]
+#y_test = y_test.astype(int)#[100:-100]
+
+
+# It tries 10 different models.
+clf = ak.StructuredDataClassifier(overwrite=True, max_trials=5)
+
+# Feed the structured data classifier with training data.
+clf.fit(x_train, y_train, epochs=10, validation_split=0.1)
+
+model = clf.export_model()
+model.summary()
+
+
+
+
+model.save("clip_autokeras_nsfw")

nsfw_testset.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c70cf196d2653ea5512e78fb8773102c9886ea8b698905310fcb1ccba56fe67f
+size 4576563

safety_settings.yml

@@ -0,0 +1,28 @@
+nsfw:
+  concepts:
+    sexual: 0.2
+    nude: 0.20
+    sex: 0.206
+    18+: 0.21
+    naked: 0.195
+    nsfw: 0.2
+    porn: 0.2
+    dick: 0.19
+    vagina: 0.19
+    naked child: 0.22
+    explicit content: 0.19
+    uncensored: 0.2
+    fuck: 0.2
+    nipples: 0.2
+    visible nipples: 0.21
+    naked breasts: 0.214
+    areola: 0.2
+  action: obfuscate
+special:
+  concepts:
+    little girl: 0.21
+    little boy: 0.21
+    young child: 0.21
+    young girl: 0.22
+    toddler: 0.21
+    baby: 0.22

violence_detection_vit_b_32.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76909cca3e77abbbb21313895756773a9c6fb26c9cc743f3ef0625f025741982
+size 4224

violence_detection_vit_l_14.npy

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc08f81ef50b1ef71452ddc87285fe6fbbc01b6fe8f8b8c24482c386ef2ab7e1
+size 6272