clean up some code

evaluate_model_2.py

@@ -9,12 +9,10 @@ from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, Tenso
 import logging
 from transformers import get_scheduler, AdamW
 from t5_model import preprocess
-#from accelerate import Accelerator
 import argparse
 logging.basicConfig(filename='log_t5.log',  level=logging.DEBUG)
 device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
-#accelerator=Accelerator()
-#device=accelerator.device
+
 def run():
 	print(args.saved_model)
 	print(args.tokenizer)
@@ -22,22 +20,14 @@ def run():
 	logging.info("performing evaluation")
 	logging.info("loading saved model")
 	
-	#checkpoint=torch.load(args.saved_model)
-	#model.load_state_dict(checkpoint)
-	model=BioGptForCausalLM.from_pretrained(args.saved_model)
-	#model=AutoModelForCausalLM.from_pretrained(args.saved_model)
-	#model=T5ForConditionalGeneration.from_pretrained("t5-small")
-	#model.load_state_dict(torch.load(args.saved_model))
-	#model=accelerator.load("./t5_small_deepspeed_train_test_2.pt")
-	#model=accelerator.load_state("./t5_small_deepspee_train_test.pt")
+	#model=BioGptForCausalLM.from_pretrained(args.saved_model)
+	model=AutoModelForCausalLM.from_pretrained(args.saved_model)
+
 	tokenizer=AutoTokenizer.from_pretrained(args.tokenizer)
 	tokenizer.padding_side="left"
 	tokenizer.pad_token=tokenizer.eos_token
 	model.config.pad_token_id=model.config.eos_token_id
 	test_dataset=preprocess(tokenizer, args.test_dataset)
-	#with open(args.test_dataset, "rb") as f:
-	#	data=pkl.load(f)
-	#test_dataset=DiagnosesDataset(data, tokenizer)
 	logging.info("running evaluation")
 	res=evaluate_model_loop(model, args.config_name, test_dataset, args.batch_size,tokenizer, args.topk, args.temp,  args.num_beams, args.early_stopping, args.no_rep_ngram, args.num_return_sequences, args.metrics, args.do_sample, args.generative, args.icd_codes)
 	logging.info(res)
@@ -62,25 +52,9 @@ class DiagnosesDataset(torch.utils.data.Dataset):
 		label_instruction=self.tokenizer(labels)
 
 		i=len(tokenized_instruction["input_ids"])-1
-		#while 1<len(item["input_ids"])
-		#print("Len of item labels before ", len(item["labels"]))
-		item["labels"][i:]=label_instruction["input_ids"]
-		#item.pop("token_type_ids")
-		#print(item["labels"])
-		#we now need to pad to 2048 
 
+		item["labels"][i:]=label_instruction["input_ids"]
 
-		#print("Len labels: ", len(item["labels"]))
-		#print("Len input ids: ", len(item["input_ids"]))
-		#print("\n\n")
-		#try:		
-		#	assert len(item["labels"]) == len(item["input_ids"])
-		#except AssertionError:
-		#	print(len(item["labels"]))
-		#	print(len(item["input_ids"]))
-		#	print(len(tokenized_instruction["input_ids"]))
-		#	print("\n\n")
-		#	break
 		return item
 
 	def tokenize(self, prompt):
@@ -90,24 +64,7 @@ class DiagnosesDataset(torch.utils.data.Dataset):
 				max_length=1024,
 				padding=False,
 				return_tensors=None)
-		
-		#print(type(result_prompt))
-		#print(len(result_prompt["input_ids"]))
-
-		#result_labels=self.tokenizer(labels,
-		#		truncation=True,
-		#		max_length=1024,
-		#		padding=False,
-		#		return_tensors=None)
-		#old_labels=result_labels["input_ids"].copy()
-		#result_prompt["labels"]=[-100 for i in result_prompt["input_ids"]] + result_labels["input_ids"]
-		#result_prompt["input_ids"]=result_prompt["input_ids"] + old_labels
-		#print(result_prompt["input_ids"]
-		#result_prompt["labels"] = [-100 for i in result_prompt["input_ids"]] + result_labels["input_ids"]
-		#print(len(result_prompt["labels"]))
-		#assert len(result_prompt["input_ids"]) == len(result_prompt["labels"])
 		result_prompt["labels"]=[-100]*len(result_prompt["input_ids"])
-		#print(result_prompt["labels"])	
 		return result_prompt
 
 	def __len__(self):
@@ -137,10 +94,6 @@ def evaluate_model_loop(model,config_name, test_dataset, batch_size, tokenizer,
 	print("num_return_sequences: ", num_return_sequences)
 	print("metrics: ", metrics)
 	print("generative? ", generative) 
-
-	#config=GenerationConfig.from_pretrained(config_name, top_k=top_k, temperature=temp, num_beams=num_beams, early_stopping=early_stopping, no_repeat_ngram_size=no_rep, num_return_sequences=num_return_sequences, max_length=512)
-	#print(config.num_return_sequences)
-	#print(tokenizer.max_length)
 	eval_sampler=SequentialSampler(test_dataset)
 	eval_dataloader=DataLoader(test_dataset, sampler=eval_sampler, batch_size=batch_size)
 	accuracies=[]
@@ -149,35 +102,15 @@ def evaluate_model_loop(model,config_name, test_dataset, batch_size, tokenizer,
 	precs=[]
 	for index, batch in tqdm(enumerate(eval_dataloader)):
 		with torch.no_grad():
-
-			#print(batch.to(device))
-			#if index == 20:
-			#	break
 			print(len(batch["input_ids"]))
-			#input_ids=torch.tensor(batch[0]).unsqueeze(0).to(device)
-			
-
 			input_ids=batch["input_ids"]
-			#print(input_ids.size())
 			attention_mask=batch["attention_mask"]
 			labels=batch["labels"]
 			labels_len=len(batch[2][batch[2] != tokenizer.pad_token_id])
 			input_ids_len=len(batch[0][0])
 			print("Len input ids: ", input_ids_len)
 			print("Len labels: ", labels_len)
-			
-			#attention_mask=torch.tensor(batch[1]).unsqueeze(0).to(device)
-			#labels=torch.tensor(batch[2]).unsqueeze(0).to(device)
-			
-
-
-			#last_inp_token_index=batch[2][::-1].index(-100)
-			#last_occurrence=len(batch[2]) - last_inp_token_index
-			#print(last_occurrence)
-			#label_length=len(batch["input_ids"][last_occurrence])
-			#print(label_length)
-			#outputs=model.generate(input_ids=input_ids, attention_mask=attention_mask, max_length=512, top_k=top_k, temperature=temp, num_beams=num_beams, early_stopping=early_stopping, no_repeat_ngram_size=no_rep, num_return_sequences=num_return_sequences)
-			
+					
 			outputs=model.generate(input_ids=input_ids, attention_mask=attention_mask, top_k=top_k, temperature=temp, num_beams=num_beams, early_stopping=early_stopping, no_repeat_ngram_size=no_rep, num_return_sequences=num_return_sequences, max_new_tokens=labels_len+8)# , length_penalty=-0.8)
 			#print("Outputs: ",len(outputs[0]))
 			#print("Len input ids: ", labels_len)