more cleanup

.gitignore

@@ -1,2 +1,5 @@
 *.pt
 *.pth
+*pyc
+*build*
+*.egg*

README.md

@@ -9,30 +9,59 @@ As depicted in following examples, one is able to highlight various discriminati
 
 [Paper Link](https://arxiv.org/abs/2301.08110)
 
+## roadmap
+ - continue to cleanup repo
+ - i.p. remove Explainer class and other overhead
+ - more examples
+ - hf integration?
 
 ## prelim
-This repo includes the XAI methods AtMan, Chefer, and a Captum interface for IG, GradCam etc. for the language-model GPT-J and vision-language model [MAGMA](https://github.com/Aleph-Alpha/magma) and [BLIP](https://colab.research.google.com/github/salesforce/BLIP).
+This repo includes the XAI methods AtMan, Chefer, and a Captum interface for IG, GradCam etc. for the language-model GPT-J and vision-language model [MAGMA](https://github.com/Aleph-Alpha/magma) and [BLIP](https://colab.research.google.com/github/salesforce/BLIP). (Big props to Mayukh Deb.)
 
 To install all required dependencies, run the following command, e.g. in a conda environment with python3.8:
 ```
 bash startup-hook.sh
 ```
+Note: further model-checkpoints will be downloaded when executing for the first time. Sometimes CLIP fails to verify on the first execution -> running again works usually.
 
-# examples
-## image-text/ MAGMA
-TODO: examples for different methods (script + image)
+# examples with MAGMA
 ```
 cd atman-magma
-python example_explain_panda.py
+```
+## image-text/ MAGMA x AtMan
+requires 1 RTX 3090
+
+```
+python example_explain_panda_atman.py
+```
+
+## image-text/ MAGMA x Chefer
+requires 1 A100
+
+```
+python example_explain_panda_chefer.py
 run plot_panda.ipynb
 ```
 
+## image-text/ MAGMA x Captum IxG, ...
+requires 1 A100
+
+```
+python example_explain_panda_captum.py
+run plot_panda.ipynb
+```
+
+## image-text/ rollout
+on it
+
 ## image-text/ BLIP
+on it
 
 ## text/ GPT-J
+on it
 
 
-# more to read
+# Method and Evaluation
 
 ![steering and measuring](figs/fig2.png)
 

atman-magma/example_attention_rollout.py

@@ -1,5 +1,5 @@
 from atman_magma.magma  import Magma
-from atman.attention_rollout import AttentionRollout
+from atman_magma.attention_rollout import AttentionRolloutMagma
 
 print('loading model...')
 model = Magma.from_checkpoint(

atman-magma/example_explain_panda.py → atman-magma/example_explain_panda_atman.py

@@ -2,6 +2,10 @@
 from atman_magma.explainer import Explainer
 from atman_magma.utils import split_str_into_tokens
 from atman_magma.logit_parsing import get_delta_cross_entropies
+import matplotlib.pyplot as plt
+import cv2
+from atman_magma.outputs import DeltaCrossEntropiesOutput
+import numpy as np
 
 
 print('loading model...')
@@ -11,13 +15,10 @@
     device = device
 )
 
-
-'''
-Image example
-'''
 from magma.image_input import ImageInput
 import  PIL.Image as PilImage
 
+
 ex = Explainer(
     model = model,
     device = device,
@@ -26,7 +27,6 @@
     conceptual_suppression_threshold = 0.75
 )
 
-
 prompt =[
     ## supports urls and path/to/image
     ImageInput('',pil=PilImage.open('openimages-panda.jpg')),
@@ -36,19 +36,10 @@
 ## returns a tensor of shape: (1, 149, 4096)
 embeddings = model.preprocess_inputs(prompt.copy())
 
-## returns a list of length embeddings.shape[0] (batch size)
-# output = model.generate(
-#     embeddings = embeddings,
-#     max_steps = 5,
-#     temperature = 0.001,
-#     top_k = 1,
-#     top_p = 0.0,
-# )
-# completion = output[0]
-
+label ='Panda'
 logit_outputs = ex.collect_logits_by_manipulating_attention(
     prompt = prompt.copy(),
-    target = 'Panda',
+    target = label,
     max_batch_size=1,
     # prompt_explain_indices=[i for i in range(10)]
 )
@@ -57,4 +48,20 @@
     output = logit_outputs
 )
 
-results.save('output.json')
+image_filename = 'openimages-panda.jpg'
+
+label_tokens =  model.tokenizer.encode(label)
+
+image = np.zeros((12,12))
+for i in range(len(label_tokens)):
+    image += results.show_image(image_token_start_idx = 0, target_token_idx= i) **2
+
+# image[image<0.6]=1.0
+fig, ax = plt.subplots(nrows=1, ncols=2, figsize = (15 , 6))
+title = f''
+fig.suptitle(title)
+ax[0].imshow(cv2.cvtColor(cv2.imread(image_filename), cv2.COLOR_BGR2RGB))
+ax[1].imshow(image)
+
+fig.savefig('panda-explained-atman.jpg')
+print('panda-explained-atman.jpg')

atman-magma/example_explain_panda_captum.py

@@ -0,0 +1,68 @@
+from PIL import Image
+import matplotlib.pyplot as plt
+
+from atman_magma.captum_helper import (
+    CaptumMagma,
+)
+from multimodal_explain_eval.utils import check_if_a_or_an_and_get_prefix
+import numpy as np
+from atman_magma.magma  import Magma
+from magma.image_input import ImageInput
+from  PIL import Image
+
+from captum.attr import IntegratedGradients, InputXGradient, GuidedGradCam
+from captum.attr import LayerGradCam
+
+
+targets = 'Panda'
+final_img = Image.open('openimages-panda.jpg')
+
+
+print('loading model...')
+model = Magma.from_checkpoint(
+    checkpoint_path =  "./mp_rank_00_model_states.pt",
+    device = 'cuda:0'
+)
+
+cmagma = CaptumMagma(magma = model)
+# captum_tool = IntegratedGradients(cmagma)
+# captum_tool = GuidedGradCam(cmagma, layer = cmagma.magma.lm.transformer.h[0].ln_1) #cmagma.magma.image_prefix.enc.layer4[-1].conv3) #, layer = cmagma.magma.image_prefix.enc.layer4[-1].conv3)
+captum_tool = InputXGradient(cmagma)
+#captum_tool = IntegratedGradients(cmagma) #! set n_steps below
+
+
+cmagma.mode='text' #hack- leave it as it is - just passes below's image embeddings thru ...
+
+
+label_tokens =  model.tokenizer.encode(targets)
+
+att_combined = np.zeros((12,12))
+for i in range(len(label_tokens)):
+
+    text_prompt = f"This is a picture of {check_if_a_or_an_and_get_prefix(targets.lower())} "
+    if i >= 1:
+        text_prompt += model.tokenizer.decode(label_tokens[:i])
+
+
+    prompt = [
+        ImageInput(None, pil=final_img),
+        text_prompt
+    ]
+
+    embeddings = cmagma.magma.preprocess_inputs(prompt)
+
+    attribution = captum_tool.attribute(
+        embeddings,
+        target=label_tokens[i],
+        #n_steps = 1 #integ gradients parameters !
+    )
+
+    att = attribution[0].abs().sum(dim = 1).cpu().detach().numpy()[:144].reshape(12,12)
+
+    att_combined += att/att.max()
+
+
+fig = plt.figure()
+plt.imshow(att_combined)
+fig.savefig('panda-explained-captum.jpg')
+print('panda-explained-captum.jpg')

atman-magma/example_chefer.py → atman-magma/example_explain_panda_chefer.py

@@ -3,6 +3,7 @@
 from magma.image_input import ImageInput
 from atman_magma.chefer.method import CheferMethod
 from atman_magma.chefer.chefer_magma.magma import CheferMagma
+import  PIL.Image as PilImage
 
 device = 'cuda:0'
 model = CheferMagma.from_checkpoint(
@@ -17,17 +18,18 @@
 
 prompt =[
     ## supports urls and path/to/image
-    ImageInput('https://www.art-prints-on-demand.com/kunst/thomas_cole/woods_hi.jpg'),
+    #ImageInput('https://www.art-prints-on-demand.com/kunst/thomas_cole/woods_hi.jpg'),
+    ImageInput('',pil=PilImage.open('openimages-panda.jpg')),
     'This is a picture of a'
 ]
 embeddings = model.preprocess_inputs(prompt)
 
 relevance_maps = cm.run(
-    embeddings = embeddings, 
-    target = ' cabin in the woods'
+    embeddings = embeddings,
+    target = 'Panda'
 )
 
 fig = plt.figure()
 plt.imshow(relevance_maps[0]['relevance_map'].reshape(12,12))
-plt.show()
-fig.savefig('chefer.jpg')
+fig.savefig('panda-explained-chefer.jpg')
+print('panda-explained-chefer.jpg')