update bash files

download_all_data.sh

@@ -0,0 +1 @@
+TBD

low_nlu/run_glue.py

@@ -114,6 +114,7 @@
 parser.add_argument('--attn_proj_vary', action="store_true")
 parser.add_argument('--save_all', action="store_true", help="save full parameters of the model")
 
+cache_dir = '/home/tuhq/.cache/torch/transformers'
 
 def tokenize(input_example, tokenizer, device, max_length):
     x_tokenized = tokenizer(input_example['text_a'], padding=True, truncation=True,
@@ -214,9 +215,9 @@ def train(args):
             torch.cuda.manual_seed_all(seed_i)
 
         config = GPT2Config()
-        tokenizer = GPT2Tokenizer.from_pretrained('gpt2', cache_dir='/home/tuhq/.cache/torch/transformers')
+        tokenizer = GPT2Tokenizer.from_pretrained('gpt2', cache_dir=cache_dir)
         tokenizer.pad_token = tokenizer.eos_token
-        gpt2_model = GPT2LMHeadModel.from_pretrained('gpt2', cache_dir='/home/tuhq/.cache/torch/transformers')
+        gpt2_model = GPT2LMHeadModel.from_pretrained('gpt2', cache_dir=cache_dir)
         ada_config = AdapterConfig(hidden_size=768,
                                    adapter_size=args.adapter_size,
                                    adapter_act='relu',
@@ -325,7 +326,7 @@ def train(args):
         if args.dataset == "yelp":
             prefix_data_path = "../data/yelp_polarity"
         else:
-            prefix_data_path = f"./glue_data/{args.dataset.upper()}"
+            prefix_data_path = f"../glue_data/{args.dataset.upper()}"
         train_loader = DataLoader(
             DictDataset(processor.get_train_examples(prefix_data_path, args.percentage_per_label, args.sample_per_label)),
             batch_size=args.batch_sizes[0],

src/run_manipulation.sh

@@ -0,0 +1,12 @@
+GPU=0
+DATA=yelp_polarity
+MODE=interpolation # or analogy
+EXPERIMENT= [the pretrained model folder using run_lm.sh]
+
+CUDA_VISIBLE_DEVICES=$GPU python test.py \
+        --experiment $EXPERIMENT
+        --mode $MODE \
+        --weighted_sample \
+        --add_attn --latent_size 768 \
+        --max_length 50 \
+        --batch_size 10

src/test.py

@@ -28,8 +28,7 @@
 # Default parameters are set based on single GPU training
 parser.add_argument("--seed", type=int, default=42)
 
-parser.add_argument('--model_type', type=str, default='cvae', choices=['cvae'])
-parser.add_argument('--dataset', type=str, default='yelp_polarity', choices=['yelp_polarity, imdb_polariry'],
+parser.add_argument('--dataset', type=str, default='yelp_polarity',
                     help="Dataset to use for training")
 
 ## mode options
@@ -723,9 +722,9 @@ def test_short(args):
     AdaVAE.eval()
     ## load ckpt
     prefix = "/data/tuhq/PTMs/bert_adapter/src/out"
-    # save_folder = os.path.join(prefix, "snli_data_iter15000_as128_scalar1.0_cycle-auto_prenc-start_wsTrue_lg-averaged_attn_add_attn_beta1.0_reg-kld_attn_mode-none_ffn_option-parallel_ffn_enc_layer-8_dec_layer-12_zdim-768_optFalse_ftFalse_zrate-0.1_fb-1sd-42_3.25")
-    save_folder = os.path.join(prefix,
-                               "yelp_polarity_iter15000_as128_scalar1.0_cycle-auto_prenc-start_wsTrue_lg-averaged_attn_add_attn_beta1.0_reg-kld_attn_mode-none_ffn_option-parallel_ffn_enc_layer-8_dec_layer-12_zdim-768_optFalse_ftFalse_zrate-0.1_fb-1sd-42_3.25")
+    save_folder = os.path.join(prefix,args.experiment)
+    # save_folder = os.path.join(prefix,
+    #                            "yelp_polarity_iter15000_as128_scalar1.0_cycle-auto_prenc-start_wsTrue_lg-averaged_attn_add_attn_beta1.0_reg-kld_attn_mode-none_ffn_option-parallel_ffn_enc_layer-8_dec_layer-12_zdim-768_optFalse_ftFalse_zrate-0.1_fb-1sd-42_3.25")
     print('Loading model weights...')
     state = torch.load(os.path.join(save_folder, 'model_best_val.pt'))  # , map_location='cpu' model_latest.pt
     if 'module' in list(state.keys())[0]:  # model_path is data parallel model with attr 'module'