Upload folder using huggingface_hub

Chest_Xray_Report_Generator-V2.py

@@ -0,0 +1,307 @@
+import os
+import transformers
+from transformers import pipeline
+import gradio as gr
+import cv2
+import numpy as np
+import pydicom
+
+##### Libraries For Grad-Cam-View
+import os
+import cv2
+import numpy as np
+import torch
+from functools import partial
+from torchvision import transforms
+from pytorch_grad_cam import GradCAM, ScoreCAM, GradCAMPlusPlus, AblationCAM, XGradCAM, EigenCAM, EigenGradCAM, LayerCAM, FullGrad
+from pytorch_grad_cam.utils.image import show_cam_on_image, preprocess_image
+from pytorch_grad_cam.ablation_layer import AblationLayerVit
+from transformers import VisionEncoderDecoderModel
+
+def generate_gradcam(image_path, model_path, output_path, method='gradcam', use_cuda=True, aug_smooth=False, eigen_smooth=False):
+    methods = {
+        "gradcam": GradCAM,
+        "scorecam": ScoreCAM,
+        "gradcam++": GradCAMPlusPlus,
+        "ablationcam": AblationCAM,
+        "xgradcam": XGradCAM,
+        "eigencam": EigenCAM,
+        "eigengradcam": EigenGradCAM,
+        "layercam": LayerCAM,
+        "fullgrad": FullGrad
+    }
+
+    if method not in methods:
+        raise ValueError(f"Method should be one of {list(methods.keys())}")
+
+    model = VisionEncoderDecoderModel.from_pretrained(model_path)
+    model.encoder.eval()
+
+    if use_cuda and torch.cuda.is_available():
+        model.encoder = model.encoder.cuda()
+    else:
+        use_cuda = False
+
+    #target_layers = [model.blocks[-1].norm1]  ## For ViT model
+    #target_layers = model.blocks[-1].norm1    ## For EfficientNet-B7 model
+    target_layers = [model.encoder.encoder.layer[-1].layernorm_before]  ## For ViT-based VisionEncoderDecoder model
+    #target_layers = [model.encoder.encoder.layers[-1].blocks[-1].layernorm_before, model.encoder.encoder.layers[-1].blocks[0].layernorm_before] ## For Swin-based VisionEncoderDecoder mode
+    
+
+    if method == "ablationcam":
+        cam = methods[method](model=model.encoder,
+                              target_layers=target_layers,
+                              use_cuda=use_cuda,
+                              reshape_transform=reshape_transform,
+                              ablation_layer=AblationLayerVit())
+    else:
+        cam = methods[method](model=model.encoder,
+                              target_layers=target_layers,
+                              use_cuda=use_cuda,
+                              reshape_transform=reshape_transform)
+
+    rgb_img = cv2.imread(image_path, 1)[:, :, ::-1]
+    rgb_img = cv2.resize(rgb_img, (224, 224)) ## (224, 224)
+    rgb_img = np.float32(rgb_img) / 255
+    input_tensor = preprocess_image(rgb_img, mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+
+    targets = None
+    cam.batch_size = 16
+
+    grayscale_cam = cam(input_tensor=input_tensor, targets=targets, eigen_smooth=eigen_smooth, aug_smooth=aug_smooth)
+    grayscale_cam = grayscale_cam[0, :]
+
+    cam_image = show_cam_on_image(rgb_img, grayscale_cam)
+    output_file = os.path.join(output_path, 'gradcam_result.png')
+    cv2.imwrite(output_file, cam_image)
+
+
+def reshape_transform(tensor, height=14, width=14):  ### height=14, width=14 for ViT-based Model
+    batch_size, token_number, embed_dim = tensor.size()
+    if token_number < height * width:
+        pad = torch.zeros(batch_size, height * width - token_number, embed_dim, device=tensor.device)
+        tensor = torch.cat([tensor, pad], dim=1)
+    elif token_number > height * width:
+        tensor = tensor[:, :height * width, :]
+
+    result = tensor.reshape(batch_size, height, width, embed_dim)
+    result = result.transpose(2, 3).transpose(1, 2)
+    return result
+
+
+
+
+# Example usage:
+#image_path = "/home/chayan/CGI_Net/images/images/CXR1353_IM-0230-1001.png"
+model_path = "/home/chayan/ViT-GPT2/Mimic_test/"
+output_path = "/home/chayan/ViT-GPT2/CAM-Result/"
+
+
+
+def sentence_case(paragraph):
+    sentences = paragraph.split('. ')
+    formatted_sentences = [sentence.capitalize() for sentence in sentences if sentence]
+    formatted_paragraph = '. '.join(formatted_sentences)
+    return formatted_paragraph
+
+def dicom_to_png(dicom_file, png_file):
+    # Load DICOM file
+    dicom_data = pydicom.dcmread(dicom_file)
+    dicom_data.PhotometricInterpretation = 'MONOCHROME1'
+
+    # Normalize pixel values to 0-255
+    img = dicom_data.pixel_array
+    img = img.astype(np.float32)
+
+    img = cv2.normalize(img, None, 0, 255, cv2.NORM_MINMAX)
+    img = img.astype(np.uint8)
+
+    # Save as PNG
+    cv2.imwrite(png_file, img)
+    return img
+
+
+Image_Captioner = pipeline("image-to-text", model = "/home/chayan/ViT-GPT2/Mimic_test/")
+
+data_dir = '/home/chayan/ViT-GPT2/'
+
+def xray_report_generator(Image_file):
+  if Image_file[-4:] =='.dcm':
+    png_file = 'DCM2PNG.png'
+    dicom_to_png(Image_file, png_file)
+    Image_file = os.path.join(data_dir, png_file)
+    output = Image_Captioner(Image_file, max_new_tokens=512)
+
+  else:
+    output = Image_Captioner(Image_file, max_new_tokens=512)
+
+  result = output[0]['generated_text']
+  output_paragraph = sentence_case(result)
+  
+  generate_gradcam(Image_file, model_path, output_path, method='gradcam', use_cuda=True)
+  
+  grad_cam_image =  output_path + 'gradcam_result.png'
+
+  return Image_file,grad_cam_image, output_paragraph
+  
+
+
+def save_feedback(feedback):
+    feedback_dir = "/home/chayan/ViT-GPT2/Feedback/"  # Update this to your desired directory
+    if not os.path.exists(feedback_dir):
+        os.makedirs(feedback_dir)
+    feedback_file = os.path.join(feedback_dir, "feedback.txt")
+    with open(feedback_file, "a") as f:
+        f.write(feedback + "\n")
+    return "Feedback submitted successfully!"
+
+
+  
+  
+# Custom CSS styles
+custom_css = """
+<style>
+
+#title {
+    color: green;
+    font-size: 36px;
+    font-weight: bold;
+}
+#description {
+    color: green;
+    font-size: 22px;
+}
+
+
+#submit-btn {
+    background-color: #1E90FF; /* DodgerBlue */
+    color: green;
+    padding: 15px 32px;
+    text-align: center;
+    text-decoration: none;
+    display: inline-block;
+    font-size: 20px;
+    margin: 4px 2px;
+    cursor: pointer;
+}
+#submit-btn:hover {
+    background-color: #00FFFF;
+}
+
+.intext textarea {
+    color: green;
+    font-size: 20px;
+    font-weight: bold;
+}
+
+
+.small-button {
+    color: green;
+    padding: 5px 10px;
+    font-size: 20px;
+}
+
+</style>
+"""
+
+# Sample image paths
+sample_images = [
+    "/mnt/data/chayan/MIMIC-CXR-JPG/2.0.0/files/p19565388/s54621108/a9510716-02da91b0-61532c26-a65b2efc-c9dfa6f1.jpg",
+    "/mnt/data/chayan/MIMIC-CXR-JPG/2.0.0/files/p19454978/s52312858/93681764-ec39480e-0518b12c-199850c2-f15118ab.jpg",
+    "/mnt/data/chayan/MIMIC-CXR-JPG/2.0.0/files/p17340686/s55469953/6ff741e9-6ea01eef-1bf10153-d1b6beba-590b6620.jpg"
+    #"sample4.png",
+    #"sample5.png"
+]
+
+def set_input_image(image_path):
+    return gr.update(value=image_path)
+
+
+with gr.Blocks(css = custom_css) as demo:
+
+    #gr.HTML(custom_css)  # Inject custom CSS
+
+    gr.Markdown(
+         """
+        <h1 style="color:blue; font-size: 36px; font-weight: bold">Chest X-ray Report Generator</h1>
+        <p id="description">Upload an X-ray image and get its report with heat-map visualization.</p>
+        """
+    )
+
+    with gr.Row():
+        inputs = gr.File(label="Upload Chest X-ray Image File", type="filepath")
+    
+    with gr.Row():
+        with gr.Column(scale=1, min_width=300):
+            outputs1 = gr.Image(label="Image Viewer")
+        with gr.Column(scale=1, min_width=300):
+            outputs2 = gr.Image(label="Grad_CAM-Visualization")
+        with gr.Column(scale=1, min_width=300):
+            outputs3 = gr.Textbox(label="Generated Report", elem_classes = "intext")
+            
+
+    submit_btn = gr.Button("Generate Report", elem_id="submit-btn")
+    submit_btn.click(
+        fn=xray_report_generator,
+        inputs=inputs,
+        outputs=[outputs1, outputs2, outputs3])
+        
+    
+    gr.Markdown(
+        """
+        <h2 style="color:green; font-size: 24px;">Or choose a sample image:</h2>
+        """
+    )
+
+    with gr.Row():
+        for idx, sample_image in enumerate(sample_images):
+            with gr.Column(scale=1):
+                #sample_image_component = gr.Image(value=sample_image, interactive=False)
+                select_button = gr.Button(f"Select Sample Image {idx+1}")
+                select_button.click(
+                    fn=set_input_image,
+                    inputs=gr.State(value=sample_image),
+                    outputs=inputs
+                )
+
+
+
+      # Feedback section
+    gr.Markdown(
+        """
+        <h2 style="color:green; font-size: 24px;">Provide Your Valuable Feedback:</h2>
+        """
+    )
+
+    with gr.Row():
+        feedback_input = gr.Textbox(label="Your Feedback", lines=4, placeholder="Enter your feedback here...")
+        feedback_submit_btn = gr.Button("Submit Feedback", elem_classes="small-button")
+        feedback_output = gr.Textbox(label="Feedback Status", interactive=False)
+
+    feedback_submit_btn.click(
+        fn=save_feedback,
+        inputs=feedback_input,
+        outputs=feedback_output
+    )
+   
+   
+   
+demo.launch(share=True)
+
+
+# inputs = gr.File(label="Upload Chest X-ray Image File", type="filepath")
+# outputs1 =gr.Image(label="Image Viewer")
+# outputs2 =gr.Image(label="Grad_CAM-Visualization")
+# outputs3 = gr.Textbox(label="Generated Report")
+
+
+# interface = gr.Interface(
+#     fn=xray_report_generator,
+#     inputs=inputs,
+#     outputs=[outputs1, outputs2, outputs3],
+#     title="Chest X-ray Report Generator",
+#     description="Upload an X-ray image and get its report.",
+# )
+
+
+# interface.launch(share=True)

Mimic_test/config.json

@@ -0,0 +1,184 @@
+{
+  "architectures": [
+    "VisionEncoderDecoderModel"
+  ],
+  "decoder": {
+    "_name_or_path": "gpt2",
+    "activation_function": "gelu_new",
+    "add_cross_attention": true,
+    "architectures": [
+      "GPT2LMHeadModel"
+    ],
+    "attn_pdrop": 0.1,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 50256,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "embd_pdrop": 0.1,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 50256,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "is_decoder": true,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_epsilon": 1e-05,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "gpt2",
+    "n_ctx": 1024,
+    "n_embd": 768,
+    "n_head": 12,
+    "n_inner": null,
+    "n_layer": 12,
+    "n_positions": 1024,
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "reorder_and_upcast_attn": false,
+    "repetition_penalty": 1.0,
+    "resid_pdrop": 0.1,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "scale_attn_by_inverse_layer_idx": false,
+    "scale_attn_weights": true,
+    "sep_token_id": null,
+    "summary_activation": null,
+    "summary_first_dropout": 0.1,
+    "summary_proj_to_labels": true,
+    "summary_type": "cls_index",
+    "summary_use_proj": true,
+    "suppress_tokens": null,
+    "task_specific_params": {
+      "text-generation": {
+        "do_sample": true,
+        "max_length": 50
+      }
+    },
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "vocab_size": 50257
+  },
+  "decoder_start_token_id": 50256,
+  "encoder": {
+    "_name_or_path": "google/vit-base-patch16-224-in21k",
+    "add_cross_attention": false,
+    "architectures": [
+      "ViTModel"
+    ],
+    "attention_probs_dropout_prob": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "encoder_stride": 16,
+    "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "hidden_act": "gelu",
+    "hidden_dropout_prob": 0.0,
+    "hidden_size": 768,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "image_size": 224,
+    "initializer_range": 0.02,
+    "intermediate_size": 3072,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "layer_norm_eps": 1e-12,
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "min_length": 0,
+    "model_type": "vit",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 12,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_channels": 3,
+    "num_hidden_layers": 12,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "patch_size": 16,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "qkv_bias": true,
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false
+  },
+  "eos_token_id": 50256,
+  "is_encoder_decoder": true,
+  "model_type": "vision-encoder-decoder",
+  "pad_token_id": 50256,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float32",
+  "transformers_version": "4.37.1"
+}

Mimic_test/generation_config.json

@@ -0,0 +1,5 @@
+{
+  "bos_token_id": 50256,
+  "eos_token_id": 50256,
+  "transformers_version": "4.37.1"
+}

Mimic_test/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eae9007ca4a208a6a3d523ec5f24ac54b6ae165263243c957a3afe18021483af
+size 956835520

Mimic_test/preprocessor_config.json

@@ -0,0 +1,22 @@
+{
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "feature_extractor_type": "ViTFeatureExtractor",
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "resample": 2,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "height": 224,
+    "width": 224
+  }
+}

Mimic_test/special_tokens_map.json

@@ -0,0 +1,6 @@
+{
+  "bos_token": "<|endoftext|>",
+  "eos_token": "<|endoftext|>",
+  "pad_token": "<|endoftext|>",
+  "unk_token": "<|endoftext|>"
+}

Mimic_test/tokenizer_config.json

@@ -0,0 +1,20 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "50256": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<|endoftext|>",
+  "clean_up_tokenization_spaces": true,
+  "eos_token": "<|endoftext|>",
+  "model_max_length": 1024,
+  "pad_token": "<|endoftext|>",
+  "tokenizer_class": "GPT2Tokenizer",
+  "unk_token": "<|endoftext|>"
+}

Mimic_test/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed01b457b864b39f97e7f2f322bba0c3d52815e5520732e5dab1ea9e98840a8b
+size 4411

README.md

@@ -1,12 +1,6 @@
 ---
-title: Automatic Chest X-ray Report Generation System
-emoji: 🌖
-colorFrom: red
-colorTo: purple
+title: Automatic_Chest_X-ray_Report_Generation_System
+app_file: Chest_Xray_Report_Generator-V2.py
 sdk: gradio
-sdk_version: 4.41.0
-app_file: app.py
-pinned: false
+sdk_version: 4.32.2
 ---
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

pytorch_grad_cam/Readme.md

@@ -0,0 +1,29 @@
+#### Grad-CAM visualization of any VisionEncoderDecoder model
+
+# Step 1: Open /pytorch_grad_cam folder and make sure that in init.py all the CAM version is imported as the class name not the python file. For example 
+                 from pytorch_grad_cam.grad_cam import GradCAM 
+because when in the main python code (Grad_CAM_Visualization.py) we want to import every Class directly.
+
+# Step2: Open the main Grad-CAM code: Grad_CAM_Visualization.py and edit the following function according to your model. 
+#                             "def reshape_transform(tensor, height=14, width=14):
+                                  result = tensor[:, 1:, :].reshape(tensor.size(0),
+                                      height, width, tensor.size(2))
+                                  result = result.transpose(2, 3).transpose(1, 2)
+#                                 return result"
+here as the resized image tensor was [150,528] which should be equivalent to the reshaped transform of [1,14,14,768]
+## The error message should be like this if any mismatch:
+              RuntimeError: shape '[1, 16, 16, 768]' is invalid for input of size 150528
+
+# Step 3: Choose your desired model from (DeIT_Base16_Pretrained with ImageNeT, Customized VisionTransformer, Dino_Base16_Pretrained with ImageNeT, My customized DeiT-CXR model, My customized EfficientNet model, and ##VisionEncoderDecoder Model)
+
+# Step 4: Open base_cam.py file and go to the "forward" function of Class BaseCAM.
+          Write extra line "outputs = outputs.pooler_output" for ##VisionEncoderDecoder Model as we need to take the tensor of pooler_output of the model configuration. Follow the comment line as well.
+
+# Step 5: Then follow the comments in the Grad_CAM_Visualization.py: 
+              use model.encoder instead of model for ## VisionEncoderDecoder Model
+              use different target_layers for different model
+              target_layers = [model.encoder.encoder.layer[-1].layernorm_before] for ## VisionEncoderDecoder Model
+
+# Step 6: Change the image_path and output_path accordingly
+
+# Step 7: Run python Grad_CAM_Visualization.py --use-cuda --image-path "directory/image_path" --method "any grad-cam method defined in the code"

pytorch_grad_cam/__init__.py

@@ -0,0 +1,20 @@
+from pytorch_grad_cam.grad_cam import GradCAM
+from pytorch_grad_cam.hirescam import HiResCAM
+from pytorch_grad_cam.grad_cam_elementwise import GradCAMElementWise
+from pytorch_grad_cam.ablation_layer import AblationLayer, AblationLayerVit, AblationLayerFasterRCNN
+from pytorch_grad_cam.ablation_cam import AblationCAM
+from pytorch_grad_cam.xgrad_cam import XGradCAM
+from pytorch_grad_cam.grad_cam_plusplus import GradCAMPlusPlus
+from pytorch_grad_cam.score_cam import ScoreCAM
+from pytorch_grad_cam.layer_cam import LayerCAM
+from pytorch_grad_cam.eigen_cam import EigenCAM
+from pytorch_grad_cam.eigen_grad_cam import EigenGradCAM
+from pytorch_grad_cam.random_cam import RandomCAM
+from pytorch_grad_cam.fullgrad_cam import FullGrad
+from pytorch_grad_cam.guided_backprop import GuidedBackpropReLUModel
+from pytorch_grad_cam.activations_and_gradients import ActivationsAndGradients
+from pytorch_grad_cam.feature_factorization.deep_feature_factorization import DeepFeatureFactorization, run_dff_on_image
+import pytorch_grad_cam.utils.model_targets
+import pytorch_grad_cam.utils.reshape_transforms
+import pytorch_grad_cam.metrics.cam_mult_image
+import pytorch_grad_cam.metrics.road

pytorch_grad_cam/ablation_cam.py

@@ -0,0 +1,148 @@
+import numpy as np
+import torch
+import tqdm
+from typing import Callable, List
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.find_layers import replace_layer_recursive
+from pytorch_grad_cam.ablation_layer import AblationLayer
+
+
+""" Implementation of AblationCAM
+https://openaccess.thecvf.com/content_WACV_2020/papers/Desai_Ablation-CAM_Visual_Explanations_for_Deep_Convolutional_Network_via_Gradient-free_Localization_WACV_2020_paper.pdf
+
+Ablate individual activations, and then measure the drop in the target score.
+
+In the current implementation, the target layer activations is cached, so it won't be re-computed.
+However layers before it, if any, will not be cached.
+This means that if the target layer is a large block, for example model.featuers (in vgg), there will
+be a large save in run time.
+
+Since we have to go over many channels and ablate them, and every channel ablation requires a forward pass,
+it would be nice if we could avoid doing that for channels that won't contribute anwyay, making it much faster.
+The parameter ratio_channels_to_ablate controls how many channels should be ablated, using an experimental method
+(to be improved). The default 1.0 value means that all channels will be ablated.
+"""
+
+
+class AblationCAM(BaseCAM):
+    def __init__(self,
+                 model: torch.nn.Module,
+                 target_layers: List[torch.nn.Module],
+                 use_cuda: bool = False,
+                 reshape_transform: Callable = None,
+                 ablation_layer: torch.nn.Module = AblationLayer(),
+                 batch_size: int = 32,
+                 ratio_channels_to_ablate: float = 1.0) -> None:
+
+        super(AblationCAM, self).__init__(model,
+                                          target_layers,
+                                          use_cuda,
+                                          reshape_transform,
+                                          uses_gradients=False)
+        self.batch_size = batch_size
+        self.ablation_layer = ablation_layer
+        self.ratio_channels_to_ablate = ratio_channels_to_ablate
+
+    def save_activation(self, module, input, output) -> None:
+        """ Helper function to save the raw activations from the target layer """
+        self.activations = output
+
+    def assemble_ablation_scores(self,
+                                 new_scores: list,
+                                 original_score: float,
+                                 ablated_channels: np.ndarray,
+                                 number_of_channels: int) -> np.ndarray:
+        """ Take the value from the channels that were ablated,
+            and just set the original score for the channels that were skipped """
+
+        index = 0
+        result = []
+        sorted_indices = np.argsort(ablated_channels)
+        ablated_channels = ablated_channels[sorted_indices]
+        new_scores = np.float32(new_scores)[sorted_indices]
+
+        for i in range(number_of_channels):
+            if index < len(ablated_channels) and ablated_channels[index] == i:
+                weight = new_scores[index]
+                index = index + 1
+            else:
+                weight = original_score
+            result.append(weight)
+
+        return result
+
+    def get_cam_weights(self,
+                        input_tensor: torch.Tensor,
+                        target_layer: torch.nn.Module,
+                        targets: List[Callable],
+                        activations: torch.Tensor,
+                        grads: torch.Tensor) -> np.ndarray:
+
+        # Do a forward pass, compute the target scores, and cache the
+        # activations
+        handle = target_layer.register_forward_hook(self.save_activation)
+        with torch.no_grad():
+            outputs = self.model(input_tensor)
+            handle.remove()
+            original_scores = np.float32(
+                [target(output).cpu().item() for target, output in zip(targets, outputs)])
+
+        # Replace the layer with the ablation layer.
+        # When we finish, we will replace it back, so the original model is
+        # unchanged.
+        ablation_layer = self.ablation_layer
+        replace_layer_recursive(self.model, target_layer, ablation_layer)
+
+        number_of_channels = activations.shape[1]
+        weights = []
+        # This is a "gradient free" method, so we don't need gradients here.
+        with torch.no_grad():
+            # Loop over each of the batch images and ablate activations for it.
+            for batch_index, (target, tensor) in enumerate(
+                    zip(targets, input_tensor)):
+                new_scores = []
+                batch_tensor = tensor.repeat(self.batch_size, 1, 1, 1)
+
+                # Check which channels should be ablated. Normally this will be all channels,
+                # But we can also try to speed this up by using a low
+                # ratio_channels_to_ablate.
+                channels_to_ablate = ablation_layer.activations_to_be_ablated(
+                    activations[batch_index, :], self.ratio_channels_to_ablate)
+                number_channels_to_ablate = len(channels_to_ablate)
+
+                for i in tqdm.tqdm(
+                    range(
+                        0,
+                        number_channels_to_ablate,
+                        self.batch_size)):
+                    if i + self.batch_size > number_channels_to_ablate:
+                        batch_tensor = batch_tensor[:(
+                            number_channels_to_ablate - i)]
+
+                    # Change the state of the ablation layer so it ablates the next channels.
+                    # TBD: Move this into the ablation layer forward pass.
+                    ablation_layer.set_next_batch(
+                        input_batch_index=batch_index,
+                        activations=self.activations,
+                        num_channels_to_ablate=batch_tensor.size(0))
+                    score = [target(o).cpu().item()
+                             for o in self.model(batch_tensor)]
+                    new_scores.extend(score)
+                    ablation_layer.indices = ablation_layer.indices[batch_tensor.size(
+                        0):]
+
+                new_scores = self.assemble_ablation_scores(
+                    new_scores,
+                    original_scores[batch_index],
+                    channels_to_ablate,
+                    number_of_channels)
+                weights.extend(new_scores)
+
+        weights = np.float32(weights)
+        weights = weights.reshape(activations.shape[:2])
+        original_scores = original_scores[:, None]
+        weights = (original_scores - weights) / original_scores
+
+        # Replace the model back to the original state
+        replace_layer_recursive(self.model, ablation_layer, target_layer)
+        return weights

pytorch_grad_cam/ablation_cam_multilayer.py

@@ -0,0 +1,136 @@
+import cv2
+import numpy as np
+import torch
+import tqdm
+from pytorch_grad_cam.base_cam import BaseCAM
+
+
+class AblationLayer(torch.nn.Module):
+    def __init__(self, layer, reshape_transform, indices):
+        super(AblationLayer, self).__init__()
+
+        self.layer = layer
+        self.reshape_transform = reshape_transform
+        # The channels to zero out:
+        self.indices = indices
+
+    def forward(self, x):
+        self.__call__(x)
+
+    def __call__(self, x):
+        output = self.layer(x)
+
+        # Hack to work with ViT,
+        # Since the activation channels are last and not first like in CNNs
+        # Probably should remove it?
+        if self.reshape_transform is not None:
+            output = output.transpose(1, 2)
+
+        for i in range(output.size(0)):
+
+            # Commonly the minimum activation will be 0,
+            # And then it makes sense to zero it out.
+            # However depending on the architecture,
+            # If the values can be negative, we use very negative values
+            # to perform the ablation, deviating from the paper.
+            if torch.min(output) == 0:
+                output[i, self.indices[i], :] = 0
+            else:
+                ABLATION_VALUE = 1e5
+                output[i, self.indices[i], :] = torch.min(
+                    output) - ABLATION_VALUE
+
+        if self.reshape_transform is not None:
+            output = output.transpose(2, 1)
+
+        return output
+
+
+def replace_layer_recursive(model, old_layer, new_layer):
+    for name, layer in model._modules.items():
+        if layer == old_layer:
+            model._modules[name] = new_layer
+            return True
+        elif replace_layer_recursive(layer, old_layer, new_layer):
+            return True
+    return False
+
+
+class AblationCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(AblationCAM, self).__init__(model, target_layers, use_cuda,
+                                          reshape_transform)
+
+        if len(target_layers) > 1:
+            print(
+                "Warning. You are usign Ablation CAM with more than 1 layers. "
+                "This is supported only if all layers have the same output shape")
+
+    def set_ablation_layers(self):
+        self.ablation_layers = []
+        for target_layer in self.target_layers:
+            ablation_layer = AblationLayer(target_layer,
+                                           self.reshape_transform, indices=[])
+            self.ablation_layers.append(ablation_layer)
+            replace_layer_recursive(self.model, target_layer, ablation_layer)
+
+    def unset_ablation_layers(self):
+        # replace the model back to the original state
+        for ablation_layer, target_layer in zip(
+                self.ablation_layers, self.target_layers):
+            replace_layer_recursive(self.model, ablation_layer, target_layer)
+
+    def set_ablation_layer_batch_indices(self, indices):
+        for ablation_layer in self.ablation_layers:
+            ablation_layer.indices = indices
+
+    def trim_ablation_layer_batch_indices(self, keep):
+        for ablation_layer in self.ablation_layers:
+            ablation_layer.indices = ablation_layer.indices[:keep]
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_category,
+                        activations,
+                        grads):
+        with torch.no_grad():
+            outputs = self.model(input_tensor).cpu().numpy()
+            original_scores = []
+            for i in range(input_tensor.size(0)):
+                original_scores.append(outputs[i, target_category[i]])
+        original_scores = np.float32(original_scores)
+
+        self.set_ablation_layers()
+
+        if hasattr(self, "batch_size"):
+            BATCH_SIZE = self.batch_size
+        else:
+            BATCH_SIZE = 32
+
+        number_of_channels = activations.shape[1]
+        weights = []
+
+        with torch.no_grad():
+            # Iterate over the input batch
+            for tensor, category in zip(input_tensor, target_category):
+                batch_tensor = tensor.repeat(BATCH_SIZE, 1, 1, 1)
+                for i in tqdm.tqdm(range(0, number_of_channels, BATCH_SIZE)):
+                    self.set_ablation_layer_batch_indices(
+                        list(range(i, i + BATCH_SIZE)))
+
+                    if i + BATCH_SIZE > number_of_channels:
+                        keep = number_of_channels - i
+                        batch_tensor = batch_tensor[:keep]
+                        self.trim_ablation_layer_batch_indices(self, keep)
+                    score = self.model(batch_tensor)[:, category].cpu().numpy()
+                    weights.extend(score)
+
+        weights = np.float32(weights)
+        weights = weights.reshape(activations.shape[:2])
+        original_scores = original_scores[:, None]
+        weights = (original_scores - weights) / original_scores
+
+        # replace the model back to the original state
+        self.unset_ablation_layers()
+        return weights

pytorch_grad_cam/ablation_layer.py

@@ -0,0 +1,155 @@
+import torch
+from collections import OrderedDict
+import numpy as np
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+
+class AblationLayer(torch.nn.Module):
+    def __init__(self):
+        super(AblationLayer, self).__init__()
+
+    def objectiveness_mask_from_svd(self, activations, threshold=0.01):
+        """ Experimental method to get a binary mask to compare if the activation is worth ablating.
+            The idea is to apply the EigenCAM method by doing PCA on the activations.
+            Then we create a binary mask by comparing to a low threshold.
+            Areas that are masked out, are probably not interesting anyway.
+        """
+
+        projection = get_2d_projection(activations[None, :])[0, :]
+        projection = np.abs(projection)
+        projection = projection - projection.min()
+        projection = projection / projection.max()
+        projection = projection > threshold
+        return projection
+
+    def activations_to_be_ablated(
+            self,
+            activations,
+            ratio_channels_to_ablate=1.0):
+        """ Experimental method to get a binary mask to compare if the activation is worth ablating.
+            Create a binary CAM mask with objectiveness_mask_from_svd.
+            Score each Activation channel, by seeing how much of its values are inside the mask.
+            Then keep the top channels.
+
+        """
+        if ratio_channels_to_ablate == 1.0:
+            self.indices = np.int32(range(activations.shape[0]))
+            return self.indices
+
+        projection = self.objectiveness_mask_from_svd(activations)
+
+        scores = []
+        for channel in activations:
+            normalized = np.abs(channel)
+            normalized = normalized - normalized.min()
+            normalized = normalized / np.max(normalized)
+            score = (projection * normalized).sum() / normalized.sum()
+            scores.append(score)
+        scores = np.float32(scores)
+
+        indices = list(np.argsort(scores))
+        high_score_indices = indices[::-
+                                     1][: int(len(indices) *
+                                              ratio_channels_to_ablate)]
+        low_score_indices = indices[: int(
+            len(indices) * ratio_channels_to_ablate)]
+        self.indices = np.int32(high_score_indices + low_score_indices)
+        return self.indices
+
+    def set_next_batch(
+            self,
+            input_batch_index,
+            activations,
+            num_channels_to_ablate):
+        """ This creates the next batch of activations from the layer.
+            Just take corresponding batch member from activations, and repeat it num_channels_to_ablate times.
+        """
+        self.activations = activations[input_batch_index, :, :, :].clone(
+        ).unsqueeze(0).repeat(num_channels_to_ablate, 1, 1, 1)
+
+    def __call__(self, x):
+        output = self.activations
+        for i in range(output.size(0)):
+            # Commonly the minimum activation will be 0,
+            # And then it makes sense to zero it out.
+            # However depending on the architecture,
+            # If the values can be negative, we use very negative values
+            # to perform the ablation, deviating from the paper.
+            if torch.min(output) == 0:
+                output[i, self.indices[i], :] = 0
+            else:
+                ABLATION_VALUE = 1e7
+                output[i, self.indices[i], :] = torch.min(
+                    output) - ABLATION_VALUE
+
+        return output
+
+
+class AblationLayerVit(AblationLayer):
+    def __init__(self):
+        super(AblationLayerVit, self).__init__()
+
+    def __call__(self, x):
+        output = self.activations
+        output = output.transpose(1, len(output.shape) - 1)
+        for i in range(output.size(0)):
+
+            # Commonly the minimum activation will be 0,
+            # And then it makes sense to zero it out.
+            # However depending on the architecture,
+            # If the values can be negative, we use very negative values
+            # to perform the ablation, deviating from the paper.
+            if torch.min(output) == 0:
+                output[i, self.indices[i], :] = 0
+            else:
+                ABLATION_VALUE = 1e7
+                output[i, self.indices[i], :] = torch.min(
+                    output) - ABLATION_VALUE
+
+        output = output.transpose(len(output.shape) - 1, 1)
+
+        return output
+
+    def set_next_batch(
+            self,
+            input_batch_index,
+            activations,
+            num_channels_to_ablate):
+        """ This creates the next batch of activations from the layer.
+            Just take corresponding batch member from activations, and repeat it num_channels_to_ablate times.
+        """
+        repeat_params = [num_channels_to_ablate] + \
+            len(activations.shape[:-1]) * [1]
+        self.activations = activations[input_batch_index, :, :].clone(
+        ).unsqueeze(0).repeat(*repeat_params)
+
+
+class AblationLayerFasterRCNN(AblationLayer):
+    def __init__(self):
+        super(AblationLayerFasterRCNN, self).__init__()
+
+    def set_next_batch(
+            self,
+            input_batch_index,
+            activations,
+            num_channels_to_ablate):
+        """ Extract the next batch member from activations,
+            and repeat it num_channels_to_ablate times.
+        """
+        self.activations = OrderedDict()
+        for key, value in activations.items():
+            fpn_activation = value[input_batch_index,
+                                   :, :, :].clone().unsqueeze(0)
+            self.activations[key] = fpn_activation.repeat(
+                num_channels_to_ablate, 1, 1, 1)
+
+    def __call__(self, x):
+        result = self.activations
+        layers = {0: '0', 1: '1', 2: '2', 3: '3', 4: 'pool'}
+        num_channels_to_ablate = result['pool'].size(0)
+        for i in range(num_channels_to_ablate):
+            pyramid_layer = int(self.indices[i] / 256)
+            index_in_pyramid_layer = int(self.indices[i] % 256)
+            result[layers[pyramid_layer]][i,
+                                          index_in_pyramid_layer, :, :] = -1000
+        return result

pytorch_grad_cam/activations_and_gradients.py

@@ -0,0 +1,46 @@
+class ActivationsAndGradients:
+    """ Class for extracting activations and
+    registering gradients from targetted intermediate layers """
+
+    def __init__(self, model, target_layers, reshape_transform):
+        self.model = model
+        self.gradients = []
+        self.activations = []
+        self.reshape_transform = reshape_transform
+        self.handles = []
+        for target_layer in target_layers:
+            self.handles.append(
+                target_layer.register_forward_hook(self.save_activation))
+            # Because of https://github.com/pytorch/pytorch/issues/61519,
+            # we don't use backward hook to record gradients.
+            self.handles.append(
+                target_layer.register_forward_hook(self.save_gradient))
+
+    def save_activation(self, module, input, output):
+        activation = output
+
+        if self.reshape_transform is not None:
+            activation = self.reshape_transform(activation)
+        self.activations.append(activation.cpu().detach())
+
+    def save_gradient(self, module, input, output):
+        if not hasattr(output, "requires_grad") or not output.requires_grad:
+            # You can only register hooks on tensor requires grad.
+            return
+
+        # Gradients are computed in reverse order
+        def _store_grad(grad):
+            if self.reshape_transform is not None:
+                grad = self.reshape_transform(grad)
+            self.gradients = [grad.cpu().detach()] + self.gradients
+
+        output.register_hook(_store_grad)
+
+    def __call__(self, x):
+        self.gradients = []
+        self.activations = []
+        return self.model(x)
+
+    def release(self):
+        for handle in self.handles:
+            handle.remove()

pytorch_grad_cam/base_cam.py

@@ -0,0 +1,205 @@
+import numpy as np
+import torch
+import ttach as tta
+from typing import Callable, List, Tuple
+from pytorch_grad_cam.activations_and_gradients import ActivationsAndGradients
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+from pytorch_grad_cam.utils.image import scale_cam_image
+from pytorch_grad_cam.utils.model_targets import ClassifierOutputTarget
+
+
+class BaseCAM:
+    def __init__(self,
+                 model: torch.nn.Module,
+                 target_layers: List[torch.nn.Module],
+                 use_cuda: bool = False,
+                 reshape_transform: Callable = None,
+                 compute_input_gradient: bool = False,
+                 uses_gradients: bool = True) -> None:
+        self.model = model.eval()
+        self.target_layers = target_layers
+        self.cuda = use_cuda
+        if self.cuda:
+            self.model = model.cuda()
+        self.reshape_transform = reshape_transform
+        self.compute_input_gradient = compute_input_gradient
+        self.uses_gradients = uses_gradients
+        self.activations_and_grads = ActivationsAndGradients(
+            self.model, target_layers, reshape_transform)
+
+    """ Get a vector of weights for every channel in the target layer.
+        Methods that return weights channels,
+        will typically need to only implement this function. """
+
+    def get_cam_weights(self,
+                        input_tensor: torch.Tensor,
+                        target_layers: List[torch.nn.Module],
+                        targets: List[torch.nn.Module],
+                        activations: torch.Tensor,
+                        grads: torch.Tensor) -> np.ndarray:
+        raise Exception("Not Implemented")
+
+    def get_cam_image(self,
+                      input_tensor: torch.Tensor,
+                      target_layer: torch.nn.Module,
+                      targets: List[torch.nn.Module],
+                      activations: torch.Tensor,
+                      grads: torch.Tensor,
+                      eigen_smooth: bool = False) -> np.ndarray:
+
+        weights = self.get_cam_weights(input_tensor,
+                                       target_layer,
+                                       targets,
+                                       activations,
+                                       grads)
+        weighted_activations = weights[:, :, None, None] * activations
+        if eigen_smooth:
+            cam = get_2d_projection(weighted_activations)
+        else:
+            cam = weighted_activations.sum(axis=1)
+        return cam
+
+    def forward(self,
+                input_tensor: torch.Tensor,
+                targets: List[torch.nn.Module],
+                eigen_smooth: bool = False) -> np.ndarray:
+
+        if self.cuda:
+            input_tensor = input_tensor.cuda()
+
+        if self.compute_input_gradient:
+            input_tensor = torch.autograd.Variable(input_tensor,
+                                                   requires_grad=True)
+
+        outputs = self.activations_and_grads(input_tensor)
+        outputs = outputs.pooler_output  # Only for ViT-GPT2 or any other VisionEncoderDecoder model
+        print(outputs)
+        if targets is None:
+            target_categories = np.argmax(outputs.cpu().data.numpy(), axis=-1)  #np.argmax(outputs.cpu().data.numpy(), axis=-1)
+            targets = [ClassifierOutputTarget(
+                category) for category in target_categories]
+
+        if self.uses_gradients:
+            self.model.zero_grad()
+            loss = sum([target(output)
+                       for target, output in zip(targets, outputs)])
+            loss.backward(retain_graph=True)
+
+        # In most of the saliency attribution papers, the saliency is
+        # computed with a single target layer.
+        # Commonly it is the last convolutional layer.
+        # Here we support passing a list with multiple target layers.
+        # It will compute the saliency image for every image,
+        # and then aggregate them (with a default mean aggregation).
+        # This gives you more flexibility in case you just want to
+        # use all conv layers for example, all Batchnorm layers,
+        # or something else.
+        cam_per_layer = self.compute_cam_per_layer(input_tensor,
+                                                   targets,
+                                                   eigen_smooth)
+        return self.aggregate_multi_layers(cam_per_layer)
+
+    def get_target_width_height(self,
+                                input_tensor: torch.Tensor) -> Tuple[int, int]:
+        width, height = input_tensor.size(-1), input_tensor.size(-2)
+        return width, height
+
+    def compute_cam_per_layer(
+            self,
+            input_tensor: torch.Tensor,
+            targets: List[torch.nn.Module],
+            eigen_smooth: bool) -> np.ndarray:
+        activations_list = [a.cpu().data.numpy()
+                            for a in self.activations_and_grads.activations]
+        grads_list = [g.cpu().data.numpy()
+                      for g in self.activations_and_grads.gradients]
+        target_size = self.get_target_width_height(input_tensor)
+
+        cam_per_target_layer = []
+        # Loop over the saliency image from every layer
+        for i in range(len(self.target_layers)):
+            target_layer = self.target_layers[i]
+            layer_activations = None
+            layer_grads = None
+            if i < len(activations_list):
+                layer_activations = activations_list[i]
+            if i < len(grads_list):
+                layer_grads = grads_list[i]
+
+            cam = self.get_cam_image(input_tensor,
+                                     target_layer,
+                                     targets,
+                                     layer_activations,
+                                     layer_grads,
+                                     eigen_smooth)
+            cam = np.maximum(cam, 0)
+            scaled = scale_cam_image(cam, target_size)
+            cam_per_target_layer.append(scaled[:, None, :])
+
+        return cam_per_target_layer
+
+    def aggregate_multi_layers(
+            self,
+            cam_per_target_layer: np.ndarray) -> np.ndarray:
+        cam_per_target_layer = np.concatenate(cam_per_target_layer, axis=1)
+        cam_per_target_layer = np.maximum(cam_per_target_layer, 0)
+        result = np.mean(cam_per_target_layer, axis=1)
+        return scale_cam_image(result)
+
+    def forward_augmentation_smoothing(self,
+                                       input_tensor: torch.Tensor,
+                                       targets: List[torch.nn.Module],
+                                       eigen_smooth: bool = False) -> np.ndarray:
+        transforms = tta.Compose(
+            [
+                tta.HorizontalFlip(),
+                tta.Multiply(factors=[0.9, 1, 1.1]),
+            ]
+        )
+        cams = []
+        for transform in transforms:
+            augmented_tensor = transform.augment_image(input_tensor)
+            cam = self.forward(augmented_tensor,
+                               targets,
+                               eigen_smooth)
+
+            # The ttach library expects a tensor of size BxCxHxW
+            cam = cam[:, None, :, :]
+            cam = torch.from_numpy(cam)
+            cam = transform.deaugment_mask(cam)
+
+            # Back to numpy float32, HxW
+            cam = cam.numpy()
+            cam = cam[:, 0, :, :]
+            cams.append(cam)
+
+        cam = np.mean(np.float32(cams), axis=0)
+        return cam
+
+    def __call__(self,
+                 input_tensor: torch.Tensor,
+                 targets: List[torch.nn.Module] = None,
+                 aug_smooth: bool = False,
+                 eigen_smooth: bool = False) -> np.ndarray:
+
+        # Smooth the CAM result with test time augmentation
+        if aug_smooth is True:
+            return self.forward_augmentation_smoothing(
+                input_tensor, targets, eigen_smooth)
+
+        return self.forward(input_tensor,
+                            targets, eigen_smooth)
+
+    def __del__(self):
+        self.activations_and_grads.release()
+
+    def __enter__(self):
+        return self
+
+    def __exit__(self, exc_type, exc_value, exc_tb):
+        self.activations_and_grads.release()
+        if isinstance(exc_value, IndexError):
+            # Handle IndexError here...
+            print(
+                f"An exception occurred in CAM with block: {exc_type}. Message: {exc_value}")
+            return True

pytorch_grad_cam/cam_mult_image.py

@@ -0,0 +1,37 @@
+import torch
+import numpy as np
+from typing import List, Callable
+from pytorch_grad_cam.metrics.perturbation_confidence import PerturbationConfidenceMetric
+
+
+def multiply_tensor_with_cam(input_tensor: torch.Tensor,
+                             cam: torch.Tensor):
+    """ Multiply an input tensor (after normalization)
+        with a pixel attribution map
+    """
+    return input_tensor * cam
+
+
+class CamMultImageConfidenceChange(PerturbationConfidenceMetric):
+    def __init__(self):
+        super(CamMultImageConfidenceChange,
+              self).__init__(multiply_tensor_with_cam)
+
+
+class DropInConfidence(CamMultImageConfidenceChange):
+    def __init__(self):
+        super(DropInConfidence, self).__init__()
+
+    def __call__(self, *args, **kwargs):
+        scores = super(DropInConfidence, self).__call__(*args, **kwargs)
+        scores = -scores
+        return np.maximum(scores, 0)
+
+
+class IncreaseInConfidence(CamMultImageConfidenceChange):
+    def __init__(self):
+        super(IncreaseInConfidence, self).__init__()
+
+    def __call__(self, *args, **kwargs):
+        scores = super(IncreaseInConfidence, self).__call__(*args, **kwargs)
+        return np.float32(scores > 0)

pytorch_grad_cam/eigen_cam.py

@@ -0,0 +1,23 @@
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+# https://arxiv.org/abs/2008.00299
+
+
+class EigenCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(EigenCAM, self).__init__(model,
+                                       target_layers,
+                                       use_cuda,
+                                       reshape_transform,
+                                       uses_gradients=False)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        return get_2d_projection(activations)

pytorch_grad_cam/eigen_grad_cam.py

@@ -0,0 +1,21 @@
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+# Like Eigen CAM: https://arxiv.org/abs/2008.00299
+# But multiply the activations x gradients
+
+
+class EigenGradCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(EigenGradCAM, self).__init__(model, target_layers, use_cuda,
+                                           reshape_transform)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        return get_2d_projection(grads * activations)

pytorch_grad_cam/feature_factorization/deep_feature_factorization.py

@@ -0,0 +1,131 @@
+import numpy as np
+from PIL import Image
+import torch
+from typing import Callable, List, Tuple, Optional
+from sklearn.decomposition import NMF
+from pytorch_grad_cam.activations_and_gradients import ActivationsAndGradients
+from pytorch_grad_cam.utils.image import scale_cam_image, create_labels_legend, show_factorization_on_image
+
+
+def dff(activations: np.ndarray, n_components: int = 5):
+    """ Compute Deep Feature Factorization on a 2d Activations tensor.
+
+    :param activations: A numpy array of shape batch x channels x height x width
+    :param n_components: The number of components for the non negative matrix factorization
+    :returns: A tuple of the concepts (a numpy array with shape channels x components),
+              and the explanation heatmaps (a numpy arary with shape batch x height x width)
+    """
+
+    batch_size, channels, h, w = activations.shape
+    reshaped_activations = activations.transpose((1, 0, 2, 3))
+    reshaped_activations[np.isnan(reshaped_activations)] = 0
+    reshaped_activations = reshaped_activations.reshape(
+        reshaped_activations.shape[0], -1)
+    offset = reshaped_activations.min(axis=-1)
+    reshaped_activations = reshaped_activations - offset[:, None]
+
+    model = NMF(n_components=n_components, init='random', random_state=0)
+    W = model.fit_transform(reshaped_activations)
+    H = model.components_
+    concepts = W + offset[:, None]
+    explanations = H.reshape(n_components, batch_size, h, w)
+    explanations = explanations.transpose((1, 0, 2, 3))
+    return concepts, explanations
+
+
+class DeepFeatureFactorization:
+    """ Deep Feature Factorization: https://arxiv.org/abs/1806.10206
+        This gets a model andcomputes the 2D activations for a target layer,
+        and computes Non Negative Matrix Factorization on the activations.
+
+        Optionally it runs a computation on the concept embeddings,
+        like running a classifier on them.
+
+        The explanation heatmaps are scalled to the range [0, 1]
+        and to the input tensor width and height.
+     """
+
+    def __init__(self,
+                 model: torch.nn.Module,
+                 target_layer: torch.nn.Module,
+                 reshape_transform: Callable = None,
+                 computation_on_concepts=None
+                 ):
+        self.model = model
+        self.computation_on_concepts = computation_on_concepts
+        self.activations_and_grads = ActivationsAndGradients(
+            self.model, [target_layer], reshape_transform)
+
+    def __call__(self,
+                 input_tensor: torch.Tensor,
+                 n_components: int = 16):
+        batch_size, channels, h, w = input_tensor.size()
+        _ = self.activations_and_grads(input_tensor)
+
+        with torch.no_grad():
+            activations = self.activations_and_grads.activations[0].cpu(
+            ).numpy()
+
+        concepts, explanations = dff(activations, n_components=n_components)
+
+        processed_explanations = []
+
+        for batch in explanations:
+            processed_explanations.append(scale_cam_image(batch, (w, h)))
+
+        if self.computation_on_concepts:
+            with torch.no_grad():
+                concept_tensors = torch.from_numpy(
+                    np.float32(concepts).transpose((1, 0)))
+                concept_outputs = self.computation_on_concepts(
+                    concept_tensors).cpu().numpy()
+            return concepts, processed_explanations, concept_outputs
+        else:
+            return concepts, processed_explanations
+
+    def __del__(self):
+        self.activations_and_grads.release()
+
+    def __exit__(self, exc_type, exc_value, exc_tb):
+        self.activations_and_grads.release()
+        if isinstance(exc_value, IndexError):
+            # Handle IndexError here...
+            print(
+                f"An exception occurred in ActivationSummary with block: {exc_type}. Message: {exc_value}")
+            return True
+
+
+def run_dff_on_image(model: torch.nn.Module,
+                     target_layer: torch.nn.Module,
+                     classifier: torch.nn.Module,
+                     img_pil: Image,
+                     img_tensor: torch.Tensor,
+                     reshape_transform=Optional[Callable],
+                     n_components: int = 5,
+                     top_k: int = 2) -> np.ndarray:
+    """ Helper function to create a Deep Feature Factorization visualization for a single image.
+        TBD: Run this on a batch with several images.
+    """
+    rgb_img_float = np.array(img_pil) / 255
+    dff = DeepFeatureFactorization(model=model,
+                                   reshape_transform=reshape_transform,
+                                   target_layer=target_layer,
+                                   computation_on_concepts=classifier)
+
+    concepts, batch_explanations, concept_outputs = dff(
+        img_tensor[None, :], n_components)
+
+    concept_outputs = torch.softmax(
+        torch.from_numpy(concept_outputs),
+        axis=-1).numpy()
+    concept_label_strings = create_labels_legend(concept_outputs,
+                                                 labels=model.config.id2label,
+                                                 top_k=top_k)
+    visualization = show_factorization_on_image(
+        rgb_img_float,
+        batch_explanations[0],
+        image_weight=0.3,
+        concept_labels=concept_label_strings)
+
+    result = np.hstack((np.array(img_pil), visualization))
+    return result

pytorch_grad_cam/fullgrad_cam.py

@@ -0,0 +1,95 @@
+import numpy as np
+import torch
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.find_layers import find_layer_predicate_recursive
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+from pytorch_grad_cam.utils.image import scale_accross_batch_and_channels, scale_cam_image
+
+# https://arxiv.org/abs/1905.00780
+
+
+class FullGrad(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        if len(target_layers) > 0:
+            print(
+                "Warning: target_layers is ignored in FullGrad. All bias layers will be used instead")
+
+        def layer_with_2D_bias(layer):
+            bias_target_layers = [torch.nn.Conv2d, torch.nn.BatchNorm2d]
+            if type(layer) in bias_target_layers and layer.bias is not None:
+                return True
+            return False
+        target_layers = find_layer_predicate_recursive(
+            model, layer_with_2D_bias)
+        super(
+            FullGrad,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform,
+            compute_input_gradient=True)
+        self.bias_data = [self.get_bias_data(
+            layer).cpu().numpy() for layer in target_layers]
+
+    def get_bias_data(self, layer):
+        # Borrowed from official paper impl:
+        # https://github.com/idiap/fullgrad-saliency/blob/master/saliency/tensor_extractor.py#L47
+        if isinstance(layer, torch.nn.BatchNorm2d):
+            bias = - (layer.running_mean * layer.weight
+                      / torch.sqrt(layer.running_var + layer.eps)) + layer.bias
+            return bias.data
+        else:
+            return layer.bias.data
+
+    def compute_cam_per_layer(
+            self,
+            input_tensor,
+            target_category,
+            eigen_smooth):
+        input_grad = input_tensor.grad.data.cpu().numpy()
+        grads_list = [g.cpu().data.numpy() for g in
+                      self.activations_and_grads.gradients]
+        cam_per_target_layer = []
+        target_size = self.get_target_width_height(input_tensor)
+
+        gradient_multiplied_input = input_grad * input_tensor.data.cpu().numpy()
+        gradient_multiplied_input = np.abs(gradient_multiplied_input)
+        gradient_multiplied_input = scale_accross_batch_and_channels(
+            gradient_multiplied_input,
+            target_size)
+        cam_per_target_layer.append(gradient_multiplied_input)
+
+        # Loop over the saliency image from every layer
+        assert(len(self.bias_data) == len(grads_list))
+        for bias, grads in zip(self.bias_data, grads_list):
+            bias = bias[None, :, None, None]
+            # In the paper they take the absolute value,
+            # but possibily taking only the positive gradients will work
+            # better.
+            bias_grad = np.abs(bias * grads)
+            result = scale_accross_batch_and_channels(
+                bias_grad, target_size)
+            result = np.sum(result, axis=1)
+            cam_per_target_layer.append(result[:, None, :])
+        cam_per_target_layer = np.concatenate(cam_per_target_layer, axis=1)
+        if eigen_smooth:
+            # Resize to a smaller image, since this method typically has a very large number of channels,
+            # and then consumes a lot of memory
+            cam_per_target_layer = scale_accross_batch_and_channels(
+                cam_per_target_layer, (target_size[0] // 8, target_size[1] // 8))
+            cam_per_target_layer = get_2d_projection(cam_per_target_layer)
+            cam_per_target_layer = cam_per_target_layer[:, None, :, :]
+            cam_per_target_layer = scale_accross_batch_and_channels(
+                cam_per_target_layer,
+                target_size)
+        else:
+            cam_per_target_layer = np.sum(
+                cam_per_target_layer, axis=1)[:, None, :]
+
+        return cam_per_target_layer
+
+    def aggregate_multi_layers(self, cam_per_target_layer):
+        result = np.sum(cam_per_target_layer, axis=1)
+        return scale_cam_image(result)

pytorch_grad_cam/grad_cam.py

@@ -0,0 +1,22 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+
+
+class GradCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(
+            GradCAM,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_layer,
+                        target_category,
+                        activations,
+                        grads):
+        return np.mean(grads, axis=(2, 3))

pytorch_grad_cam/grad_cam_elementwise.py

@@ -0,0 +1,30 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+
+class GradCAMElementWise(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(
+            GradCAMElementWise,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        elementwise_activations = np.maximum(grads * activations, 0)
+
+        if eigen_smooth:
+            cam = get_2d_projection(elementwise_activations)
+        else:
+            cam = elementwise_activations.sum(axis=1)
+        return cam

pytorch_grad_cam/grad_cam_plusplus.py

@@ -0,0 +1,32 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+
+# https://arxiv.org/abs/1710.11063
+
+
+class GradCAMPlusPlus(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(GradCAMPlusPlus, self).__init__(model, target_layers, use_cuda,
+                                              reshape_transform)
+
+    def get_cam_weights(self,
+                        input_tensor,
+                        target_layers,
+                        target_category,
+                        activations,
+                        grads):
+        grads_power_2 = grads**2
+        grads_power_3 = grads_power_2 * grads
+        # Equation 19 in https://arxiv.org/abs/1710.11063
+        sum_activations = np.sum(activations, axis=(2, 3))
+        eps = 0.000001
+        aij = grads_power_2 / (2 * grads_power_2 +
+                               sum_activations[:, :, None, None] * grads_power_3 + eps)
+        # Now bring back the ReLU from eq.7 in the paper,
+        # And zero out aijs where the activations are 0
+        aij = np.where(grads != 0, aij, 0)
+
+        weights = np.maximum(grads, 0) * aij
+        weights = np.sum(weights, axis=(2, 3))
+        return weights

pytorch_grad_cam/guided_backprop.py

@@ -0,0 +1,100 @@
+import numpy as np
+import torch
+from torch.autograd import Function
+from pytorch_grad_cam.utils.find_layers import replace_all_layer_type_recursive
+
+
+class GuidedBackpropReLU(Function):
+    @staticmethod
+    def forward(self, input_img):
+        positive_mask = (input_img > 0).type_as(input_img)
+        output = torch.addcmul(
+            torch.zeros(
+                input_img.size()).type_as(input_img),
+            input_img,
+            positive_mask)
+        self.save_for_backward(input_img, output)
+        return output
+
+    @staticmethod
+    def backward(self, grad_output):
+        input_img, output = self.saved_tensors
+        grad_input = None
+
+        positive_mask_1 = (input_img > 0).type_as(grad_output)
+        positive_mask_2 = (grad_output > 0).type_as(grad_output)
+        grad_input = torch.addcmul(
+            torch.zeros(
+                input_img.size()).type_as(input_img),
+            torch.addcmul(
+                torch.zeros(
+                    input_img.size()).type_as(input_img),
+                grad_output,
+                positive_mask_1),
+            positive_mask_2)
+        return grad_input
+
+
+class GuidedBackpropReLUasModule(torch.nn.Module):
+    def __init__(self):
+        super(GuidedBackpropReLUasModule, self).__init__()
+
+    def forward(self, input_img):
+        return GuidedBackpropReLU.apply(input_img)
+
+
+class GuidedBackpropReLUModel:
+    def __init__(self, model, use_cuda):
+        self.model = model
+        self.model.eval()
+        self.cuda = use_cuda
+        if self.cuda:
+            self.model = self.model.cuda()
+
+    def forward(self, input_img):
+        return self.model(input_img)
+
+    def recursive_replace_relu_with_guidedrelu(self, module_top):
+
+        for idx, module in module_top._modules.items():
+            self.recursive_replace_relu_with_guidedrelu(module)
+            if module.__class__.__name__ == 'ReLU':
+                module_top._modules[idx] = GuidedBackpropReLU.apply
+        print("b")
+
+    def recursive_replace_guidedrelu_with_relu(self, module_top):
+        try:
+            for idx, module in module_top._modules.items():
+                self.recursive_replace_guidedrelu_with_relu(module)
+                if module == GuidedBackpropReLU.apply:
+                    module_top._modules[idx] = torch.nn.ReLU()
+        except BaseException:
+            pass
+
+    def __call__(self, input_img, target_category=None):
+        replace_all_layer_type_recursive(self.model,
+                                         torch.nn.ReLU,
+                                         GuidedBackpropReLUasModule())
+
+        if self.cuda:
+            input_img = input_img.cuda()
+
+        input_img = input_img.requires_grad_(True)
+
+        output = self.forward(input_img)
+
+        if target_category is None:
+            target_category = np.argmax(output.cpu().data.numpy())
+
+        loss = output[0, target_category]
+        loss.backward(retain_graph=True)
+
+        output = input_img.grad.cpu().data.numpy()
+        output = output[0, :, :, :]
+        output = output.transpose((1, 2, 0))
+
+        replace_all_layer_type_recursive(self.model,
+                                         GuidedBackpropReLUasModule,
+                                         torch.nn.ReLU())
+
+        return output

pytorch_grad_cam/hirescam.py

@@ -0,0 +1,32 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+
+class HiResCAM(BaseCAM):
+    def __init__(self, model, target_layers, use_cuda=False,
+                 reshape_transform=None):
+        super(
+            HiResCAM,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        elementwise_activations = grads * activations
+
+        if eigen_smooth:
+            print(
+                "Warning: HiResCAM's faithfulness guarantees do not hold if smoothing is applied")
+            cam = get_2d_projection(elementwise_activations)
+        else:
+            cam = elementwise_activations.sum(axis=1)
+        return cam

pytorch_grad_cam/layer_cam.py

@@ -0,0 +1,36 @@
+import numpy as np
+from pytorch_grad_cam.base_cam import BaseCAM
+from pytorch_grad_cam.utils.svd_on_activations import get_2d_projection
+
+# https://ieeexplore.ieee.org/document/9462463
+
+
+class LayerCAM(BaseCAM):
+    def __init__(
+            self,
+            model,
+            target_layers,
+            use_cuda=False,
+            reshape_transform=None):
+        super(
+            LayerCAM,
+            self).__init__(
+            model,
+            target_layers,
+            use_cuda,
+            reshape_transform)
+
+    def get_cam_image(self,
+                      input_tensor,
+                      target_layer,
+                      target_category,
+                      activations,
+                      grads,
+                      eigen_smooth):
+        spatial_weighted_activations = np.maximum(grads, 0) * activations
+
+        if eigen_smooth:
+            cam = get_2d_projection(spatial_weighted_activations)
+        else:
+            cam = spatial_weighted_activations.sum(axis=1)
+        return cam