Update README.md

README.md

@@ -1,204 +1,93 @@
 ---
 library_name: peft
 base_model: mistralai/Mistral-7B-Instruct-v0.1
+license: apache-2.0
+datasets:
+- liuhaotian/LLaVA-Instruct-150K
 ---
 
 # Model Card for Model ID
 
-<!-- Provide a quick summary of what the model is/does. -->
+![Text Meme](meme.jpg)
 
+Is text really all you need? Probably not, but the least we can do is try. This repo contains a QLoRA fine-tune of Mistral-7B on the original Llava-150K-Instruct dataset; however, each image is encoded as a base64 representation. With enough data, can a LLM learn to "see" just from text? Early results say absolutely not, but I am committed to burning my GPU credits regardless of how bad the result.
 
+I do believe in the future we will see a "simplification" of architectures designed to work for multiple modalities. LLaVA, for example, combines a vision encoder with a pre-trained LLM. Perhaps models of the future will have a joint-representation for both images and text, and not have to rely on splicing 2 models together. For example, perhaps [Token-Free Models](https://arxiv.org/html/2401.13660v1) could be trained on multi-modal byte representations of inputs. Of course, this would be extremely computationally expensive compared to modern vision models, but maybe 10-20 years down the line it's not that big of a deal?
 
-## Model Details
-
-### Model Description
-
-<!-- Provide a longer summary of what this model is. -->
-
-
-
-- **Developed by:** [More Information Needed]
-- **Funded by [optional]:** [More Information Needed]
-- **Shared by [optional]:** [More Information Needed]
-- **Model type:** [More Information Needed]
-- **Language(s) (NLP):** [More Information Needed]
-- **License:** [More Information Needed]
-- **Finetuned from model [optional]:** [More Information Needed]
-
-### Model Sources [optional]
-
-<!-- Provide the basic links for the model. -->
-
-- **Repository:** [More Information Needed]
-- **Paper [optional]:** [More Information Needed]
-- **Demo [optional]:** [More Information Needed]
-
-## Uses
-
-<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
-
-### Direct Use
-
-<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
-
-[More Information Needed]
-
-### Downstream Use [optional]
-
-<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
-
-[More Information Needed]
-
-### Out-of-Scope Use
-
-<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
-
-[More Information Needed]
-
-## Bias, Risks, and Limitations
-
-<!-- This section is meant to convey both technical and sociotechnical limitations. -->
-
-[More Information Needed]
-
-### Recommendations
-
-<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
-
-Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
-
-## How to Get Started with the Model
-
-Use the code below to get started with the model.
-
-[More Information Needed]
-
-## Training Details
-
-### Training Data
-
-<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
-
-[More Information Needed]
-
-### Training Procedure 
-
-<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
-
-#### Preprocessing [optional]
-
-[More Information Needed]
-
+To use this model, you can load the base Mistral model and the adapter:
 
-#### Training Hyperparameters
+```python
+import torch
+from peft import PeftModel
+from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
 
-- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
+BASE_MODEL = "mistralai/Mistral-7B-Instruct-v0.1"
+ADAPTER_MODEL = "seanmor5/mistral-7b-instruct-vision-64-qlora"
+MAX_SEQ_LEN = 2048
 
-#### Speeds, Sizes, Times [optional]
+device = "cuda"
 
-<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
+bnb_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_use_double_quant=True,
+    bnb_4bit_quant_type="nf4",
+    bnb_4bit_compute_dtype=torch.bfloat16,
+)
 
-[More Information Needed]
+model = AutoModelForCausalLM.from_pretrained(BASE_MODEL)
+model = PeftModel.from_pretrained(model, ADAPTER_MODEL)
 
-## Evaluation
+tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, model_max_length=MAX_SEQ_LEN)
+tokenizer.pad_token = tokenizer.eos_token
+```
 
-<!-- This section describes the evaluation protocols and provides the results. -->
+One challenge with this approach is sequence length. High resolution images are large, and when encoded in base64 create prohibitively large sequences. To naively overcome this we aggressively resize and downsample the image:
 
-### Testing Data, Factors & Metrics
+```python
+import base64
+from io import BytesIO
+from PIL import Image
 
-#### Testing Data
+TARGET_SIZE = (224, 168)
+TARGET_QUALITY = 5
 
-<!-- This should link to a Dataset Card if possible. -->
+def downsample(path):
+    img = Image.open(path)
+    img = img.resize(TARGET_SIZE, Image.ANTIALIAS)
+    buf = BytesIO()
+    img.save(buf, optimize=True, quality=5, format="JPEG")
+    return f"<image>{base64.b64encode(buf.getvalue()).decode()}</image>"
+```
 
-[More Information Needed]
+Then we can use the default Mistral chat output, ensuring our images are encoded properly within the text:
 
-#### Factors
+```python
+def replace_image(seq, img):
+    return seq.replace("<image>", downsample(img))
 
-<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
+prompt = (
+    "<image>\nWhat is the dog doing in this photo?"
+)
+prompt = replace_image(prompt, "dog.jpg")
+print(prompt)
 
-[More Information Needed]
+messages = [{"role": "user", "content": prompt}]
 
-#### Metrics
+encodeds = tokenizer.apply_chat_template(messages, return_tensors="pt")
 
-<!-- These are the evaluation metrics being used, ideally with a description of why. -->
+model_inputs = encodeds.to(device)
+model.to(device)
 
-[More Information Needed]
+generated_ids = model.generate(
+    input_ids=model_inputs, max_new_tokens=1000, do_sample=True
+)
+decoded = tokenizer.batch_decode(generated_ids)
+print(decoded[0])
+```
 
-### Results
+Even with this aggressive downsampling, some images result in sequences that are too large. Tough luck. I also did not do this experiment with any other format but JPEG images, and I did not consider the effect that the image format may have had on the model's performance.
 
-[More Information Needed]
-
-#### Summary
-
-
-
-## Model Examination [optional]
-
-<!-- Relevant interpretability work for the model goes here -->
-
-[More Information Needed]
-
-## Environmental Impact
-
-<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
-
-Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
-
-- **Hardware Type:** [More Information Needed]
-- **Hours used:** [More Information Needed]
-- **Cloud Provider:** [More Information Needed]
-- **Compute Region:** [More Information Needed]
-- **Carbon Emitted:** [More Information Needed]
-
-## Technical Specifications [optional]
-
-### Model Architecture and Objective
-
-[More Information Needed]
-
-### Compute Infrastructure
-
-[More Information Needed]
-
-#### Hardware
-
-[More Information Needed]
-
-#### Software
-
-[More Information Needed]
-
-## Citation [optional]
-
-<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
-
-**BibTeX:**
-
-[More Information Needed]
-
-**APA:**
-
-[More Information Needed]
-
-## Glossary [optional]
-
-<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
-
-[More Information Needed]
-
-## More Information [optional]
-
-[More Information Needed]
-
-## Model Card Authors [optional]
-
-[More Information Needed]
-
-## Model Card Contact
-
-[More Information Needed]
-
-
-### Framework versions
+## Model Details
 
-- PEFT 0.8.2
+- **Developed by:** Sean Moriarity
+- **License:** Apache 2.0