sampling

app.py

@@ -52,12 +52,12 @@ with open("bad_words.txt", "r") as f:
 model, image_processor, tokenizer = create_model_and_transforms(
     clip_vision_encoder_pretrained="openai",
     clip_vision_encoder_path="ViT-L-14",
-    lang_encoder_path="togethercomputer/RedPajama-INCITE-Base-3B-v1",
-    tokenizer_path="togethercomputer/RedPajama-INCITE-Base-3B-v1",
+    lang_encoder_path="anas-awadalla/mpt-7b",
+    tokenizer_path="anas-awadalla/mpt-7b",
     cross_attn_every_n_layers=2,
 )
 
-checkpoint_path = hf_hub_download("openflamingo/OpenFlamingo-4B-vitl-rpj3b", "checkpoint.pt")
+checkpoint_path = hf_hub_download("openflamingo/OpenFlamingo-9B-vitl-mpt7b", "checkpoint.pt")
 model.load_state_dict(torch.load(checkpoint_path), strict=False)
 
 model.eval()
@@ -112,9 +112,7 @@ def generate(
     demo_plus_text += (
         "<image>Output:" if idx != 2 else f"<image>Question: {text.strip()} Answer:"
     )
-    # demo_plus_image = [example_one_image, example_two_image, image]
 
-    # print(demo_plus_image)
     print(demo_plus_text)
 
     lang_x = tokenizer(demo_plus_text, return_tensors="pt")
@@ -132,11 +130,10 @@ def generate(
         lang_x=input_ids,
         attention_mask=attention_mask,
         max_new_tokens=50,
-        num_beams=3,
+        num_beams=5,
         do_sample=True,
-        top_p=0.9,
+        top_p=0.95,
         top_k=0,
-        no_repeat_ngram_size=3,
     )
 
     gen_text = tokenizer.decode(
@@ -178,7 +175,7 @@ with gr.Blocks() as demo:
     Blog posts: #1 [An open-source framework for training vision-language models with in-context learning (like GPT-4!)](https://laion.ai/blog/open-flamingo/) // #2 [OpenFlamingo v2: New Models and Enhanced Training Setup]()\n
     GitHub: [open_flamingo](https://github.com/mlfoundations/open_flamingo)
     
-    In this demo we implement an interactive interface that showcases the in-context learning capabilities of the OpenFlamingo-9B model, a large multimodal model trained on top of LLaMA-7B.
+    In this demo we implement an interactive interface that showcases the in-context learning capabilities of the OpenFlamingo-9B model, a large multimodal model trained on top of MPT-7B.
     The model is trained on an interleaved mixture of text and images and is able to generate text conditioned on sequences of images/text. To safeguard against harmful generations, we detect toxic text in the model output and reject it. However, we understand that this is not a perfect solution and we encourage you to use this demo responsibly. If you find that the model is generating harmful text, please report it using this [form](https://forms.gle/StbcPvyyW2p3Pc7z6).
     """
     )
@@ -186,7 +183,7 @@ with gr.Blocks() as demo:
     with gr.Accordion("See terms and conditions"):
         gr.Markdown("""**Please read the following information carefully before proceeding.**
 [OpenFlamingo-9B](https://huggingface.co/openflamingo/OpenFlamingo-9B-vitl-mpt7b) is a **research prototype** that aims to enable users to interact with AI through both language and images. AI agents equipped with both language and visual understanding can be useful on a larger variety of tasks compared to models that communicate solely via language. By releasing an open-source research prototype, we hope to help the research community better understand the risks and limitations of modern visual-language AI models and accelerate the development of safer and more reliable methods.
-**Limitations.** OpenFlamingo-9B is built on top of the [MPT-7B](https://huggingface.co/mosaicml/mpt-7b) large language model developed by Meta AI. Large language models are trained on mostly unfiltered internet data, and have been shown to be able to produce toxic, unethical, inaccurate, and harmful content. On top of this, OpenFlamingo’s ability to support visual inputs creates additional risks, since it can be used in a wider variety of applications; image+text models may carry additional risks specific to multimodality. Please use discretion when assessing the accuracy or appropriateness of the model’s outputs, and be mindful before sharing its results.
+**Limitations.** OpenFlamingo-9B is built on top of the [MPT-7B](https://huggingface.co/mosaicml/mpt-7b) large language model developed by MosaicML. Large language models are trained on mostly unfiltered internet data, and have been shown to be able to produce toxic, unethical, inaccurate, and harmful content. On top of this, OpenFlamingo’s ability to support visual inputs creates additional risks, since it can be used in a wider variety of applications; image+text models may carry additional risks specific to multimodality. Please use discretion when assessing the accuracy or appropriateness of the model’s outputs, and be mindful before sharing its results.
 **Privacy and data collection.** This demo does NOT store any personal information on its users, and it does NOT store user queries.""")
         read_tc = gr.Checkbox(
             label="I have read and agree to the terms and conditions")