Introduced Fine tuning for the models

app.py

@@ -2,13 +2,23 @@ import gradio as gr
 import torch
 from PIL import Image
 
-from model import GitBaseCocoModel
+from model import GitBaseCocoModel, BlipBaseModel
 
+MODELS = {
+	"Git-Base-COCO": GitBaseCocoModel,
+	"Blip Base": BlipBaseModel,
+}
 
 def generate_captions(
 	image,
-	max_len,
 	num_captions,
+	max_length,
+	temperature,
+	top_k,
+	top_p,
+	repetition_penalty,
+	diversity_penalty,
+	model_name,
 	):
 	"""
 	Generates captions for the given image.
@@ -28,14 +38,23 @@ def generate_captions(
 	"""
 
 	device = "cuda" if torch.cuda.is_available() else "cpu"
-	checkpoint = "microsoft/git-base-coco"
 	
-	model = GitBaseCocoModel(device, checkpoint)
+	model = MODELS[model_name](device)
+
+	captions = model.generate(
+		image,
+		max_length,
+		num_captions,
+		temperature,
+		top_k,
+		top_p,
+		repetition_penalty,
+		diversity_penalty,
+	)
 
-	caption = model.generate(image, max_len, num_captions)
 	# Convert list to a single string separated by newlines.
-	caption = "\n".join(caption)
-	return caption
+	captions = "\n".join(captions)
+	return captions
 
 title = "Git-Base-COCO Image Captioning"
 description = "A model for generating captions for images."
@@ -44,8 +63,14 @@ interface = gr.Interface(
 	fn=generate_captions,
 	inputs=[
 		gr.inputs.Image(type="pil", label="Image"),
-		gr.inputs.Slider(minimum=20, maximum=100, step=5, default=50, label="Maximum Caption Length"),
 		gr.inputs.Slider(minimum=1, maximum=10, step=1, default=1, label="Number of Captions to Generate"),
+		gr.inputs.Slider(minimum=20, maximum=100, step=5, default=50, label="Maximum Caption Length"),
+		gr.inputs.Slider(minimum=0.1, maximum=10.0, step=0.1, default=1.0, label="Temperature"),
+		gr.inputs.Slider(minimum=1, maximum=100, step=1, default=50, label="Top K"),
+		gr.inputs.Slider(minimum=-5.0, maximum=5.0, step=0.1, default=1.0, label="Top P"),
+		gr.inputs.Slider(minimum=1.0, maximum=10.0, step=0.1, default=1.0, label="Repetition Penalty"),
+		gr.inputs.Slider(minimum=0.0, maximum=10.0, step=0.1, default=0.0, label="Diversity Penalty"),
+		gr.Inputs.Dropdown(MODELS.keys(), label="Model"),
 	],
 	outputs=[
 		gr.outputs.Textbox(label="Caption"),

model.py

@@ -1,7 +1,81 @@
-from transformers import AutoProcessor, AutoModelForCausalLM
+from transformers import AutoProcessor, AutoModelForCausalLM, BlipForConditionalGeneration
 
-class GitBaseCocoModel:
-	def __init__(self, device, checkpoint="microsoft/git-base-coco"):
+class ImageCaptionModel:
+	def __init__(
+		self,
+		device,
+		processor,
+		model,
+	) -> None:
+		self.device = device
+		self.processor = processor
+		self.model = model
+		self.model.to(self.device)
+	
+	def generate(
+		self,
+		image,
+		num_captions=1,
+		max_length=50,
+		num_beam_groups=1,
+		temperature=1.0,
+		top_k=50,
+		top_p=1.0,
+		repetition_penalty=1.0,
+		diversity_penalty=0.0,
+	):
+		"""
+		Generates captions for the given image.
+		
+		-----
+		Parameters:
+		preprocessor: transformers.PreTrainedTokenizerFast
+			The preprocessor to use for the model.
+		model: transformers.PreTrainedModel	
+			The model to use for generating captions.
+		image: PIL.Image
+			The image to generate captions for.
+		num_captions: int
+			The number of captions to generate.
+		num_beam_groups: int
+			The number of beam groups to use for beam search in order to maintain diversity. Must be between 1 and num_beams. 1 means no group_beam_search..
+		temperature: float
+			The temperature to use for sampling. The value used to module the next token probabilities that will be used by default in the generate method of the model. Must be strictly positive. Defaults to 1.0.
+		top_k: int
+			The number of highest probability vocabulary tokens to keep for top-k-filtering. A large value of top_k will keep more probabilities for each token leading to a better but slower generation. Defaults to 50.
+		top_p: float
+			The value that will be used by default in the generate method of the model for top_p. If set to float < 1, only the most probable tokens with probabilities that add up to top_p or higher are kept for generation.
+		repetition_penalty: float
+			The parameter for repetition penalty. 1.0 means no penalty. Defaults to 1.0.
+		diversity_penalty: float
+			The parameter for diversity penalty. 0.0 means no penalty. Defaults to 0.0.
+		
+		"""
+		pixel_values = self.processor(images=image, return_tensors="pt").pixel_values.to(self.device)
+			
+		if diversity_penalty != 0.0:
+			num_beam_groups = 2
+			num_captions = num_captions if num_captions % 2 == 0 else num_captions + 1
+
+		generated_ids = self.model.generate(
+			pixel_values=pixel_values,
+			max_length=max_length,
+			num_beams=num_captions,
+			num_beam_groups=num_beam_groups,
+			num_return_sequences=num_captions,
+			temperature=temperature,
+			top_k=top_k,
+			top_p=top_p,
+			repetition_penalty=repetition_penalty,
+			diversity_penalty=diversity_penalty,
+		)
+
+		generated_caption = self.processor.batch_decode(generated_ids, skip_special_tokens=True)
+
+		return generated_caption[:num_captions]
+
+class GitBaseCocoModel(ImageCaptionModel):
+	def __init__(self, device):
 		"""
 		A wrapper class for the Git-Base-COCO model. It is a pretrained model for image captioning.
 
@@ -16,12 +90,12 @@ class GitBaseCocoModel:
 		Returns:
 		None
 		"""
-		self.checkpoint = checkpoint
-		self.device = device
-		self.processor = AutoProcessor.from_pretrained(self.checkpoint)
-		self.model = AutoModelForCausalLM.from_pretrained(self.checkpoint).to(self.device)
+		checkpoint = "microsoft/git-base-coco"
+		processor = AutoProcessor.from_pretrained(checkpoint)
+		model = AutoModelForCausalLM.from_pretrained(checkpoint)
+		super().__init__(device, processor, model)
 
-	def generate(self, image, max_len=50, num_captions=1):
+	def generate(self, image, max_length=50, num_captions=1, **kwargs):
 		"""
 		Generates captions for the given image.
 
@@ -34,14 +108,29 @@ class GitBaseCocoModel:
 		num_captions: int
 			The number of captions to generate.
 		"""
-		pixel_values = self.processor(
-			images=image, return_tensors="pt"
-			).pixel_values.to(self.device)
-		generated_ids = self.model.generate(
-			pixel_values=pixel_values,
-			max_length=max_len,
-			num_beams=num_captions,
-    		num_return_sequences=num_captions,
-		)
-		return self.processor.batch_decode(generated_ids, skip_special_tokens=True)
-	
+		captions = super().generate(image, max_length, num_captions, **kwargs)
+		return captions
+	
+
+class BlipBaseModel(ImageCaptionModel):
+	def __init__(self, device):
+		self.checkpoint = "Salesforce/blip-image-captioning-base"
+		processor = AutoProcessor.from_pretrained(self.checkpoint)
+		model = BlipForConditionalGeneration.from_pretrained(self.checkpoint)
+		super().__init__(device, processor, model)
+
+	def generate(self, image, max_length=50, num_captions=1, **kwargs):
+		"""
+		Generates captions for the given image.
+
+		-----
+		Parameters:
+		image: PIL.Image
+			The image to generate captions for.
+		max_len: int
+			The maximum length of the caption.
+		num_captions: int
+			The number of captions to generate.
+		"""
+		captions = super().generate(image, max_length, num_captions, **kwargs)
+		return captions