update

examples/clip/data/pairs.csv

@@ -1,3 +1,3 @@
 title	filepath
 a very typical bus station	0.jpg
-young confused girl standing in front of a wardrobe	1.jpg
+the jetty : different types of plants to establish a variety of ecosystems .	1.jpg

examples/clip/deepspeed.json

@@ -0,0 +1,48 @@
+{
+    "train_micro_batch_size_per_gpu": 64,
+    "gradient_accumulation_steps": 1,
+    "steps_per_print": 100,
+    "gradient_clipping": 1.0,
+    "zero_optimization": {
+      "stage": 2,
+      "contiguous_gradients": false,
+      "overlap_comm": true,
+      "reduce_scatter": true,
+      "reduce_bucket_size": 5e7,
+      "allgather_bucket_size": 5e7,
+      "cpu_offload": true
+    },
+    "scheduler": {
+      "type": "WarmupLR",
+      "params": {
+          "warmup_min_lr": 0,
+          "warmup_max_lr": 1e-5,
+          "warmup_num_steps": 2000
+      }
+   },
+    "zero_allow_untested_optimizer": true,
+    "fp16": {
+      "enabled": true,
+      "loss_scale": 0,
+      "loss_scale_window": 1000,
+      "hysteresis": 2,
+      "min_loss_scale": 1
+    },
+    "optimizer": {
+      "type": "Adam",
+      "params": {
+        "lr": 1e-5,
+        "weight_decay": 0.1,
+        "betas": [
+          0.9,
+          0.98
+        ],
+        "eps": 1e-6
+      }
+    },
+    "activation_checkpointing": {
+      "partition_activations": true,
+      "contiguous_memory_optimization": false
+    },
+    "wall_clock_breakdown": false
+  }

examples/clip/hostfile

@@ -0,0 +1 @@
+127.0.0.1 slots=2

examples/clip/inference_clip.py

@@ -1,38 +1,24 @@
 import torch
-import sys
-sys.path.append("/mnt/wchh/FlagAI-internal")
 from PIL import Image
-from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
 from flagai.auto_model.auto_loader import AutoLoader
+from flagai.data.dataset.mm.clip_dataset import clip_transform
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
-loader = AutoLoader(task_name="cl", #contrastive learning
-                    model_name="clip-base-p32-224",
-                    model_dir="/mnt/clip_models/")
+loader = AutoLoader(task_name="txt_img_matching", #contrastive learning
+                    model_name="clip-base-p32-224")
 
 model = loader.get_model()
 model.eval()
 model.to(device)
 tokenizer = loader.get_tokenizer()
-n_px = model.image_size
-print(n_px)
-def _convert_image_to_rgb(image):
-    return image.convert("RGB")
-image_transform = Compose([
-    Resize(n_px, interpolation=Image.BICUBIC),
-    CenterCrop(n_px),
-    _convert_image_to_rgb,
-    ToTensor(),
-    Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)),
-])
+transform = clip_transform(img_size=model.image_size)
 
-def test_inference():
+def inference():
     image = Image.open("./CLIP.png")
-    image = image_transform(image).unsqueeze(0).to(device)
+    image = transform(image).unsqueeze(0).to(device)
     text = tokenizer.tokenize(["a diagram", "a dog", "a cat"]).to(device)
 
-    import  numpy as np
     with torch.no_grad():
         image_features = model.encode_image(image)
         text_features = model.encode_text(text)
@@ -41,4 +27,4 @@ def test_inference():
     print(text_probs.cpu().numpy()[0].tolist())
 
 if __name__=="__main__":
-    test_inference()
+    inference()

examples/clip/train_clip.py

@@ -0,0 +1,36 @@
+import torch
+from flagai.data.dataset.mm.clip_dataset import CsvDataset, clip_transform, collate_fn
+from flagai.trainer import Trainer
+from flagai.auto_model.auto_loader import AutoLoader
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+# cd examples/clip
+data_path = "./data/pairs.csv"
+img_dir = "./data/img"
+
+trainer = Trainer(env_type="pytorch",
+                  epochs=5,
+                  pytorch_device=device,
+                  batch_size=64,
+                  lr=1e-4,
+                  log_interval=10,
+                  )
+
+loader = AutoLoader(task_name="txt_img_matching",#contrastive learning
+                    model_name="clip-base-p32-224",
+                    )
+model = loader.get_model()
+tokenizer = loader.get_tokenizer()
+
+transform = clip_transform(img_size=model.image_size)
+train_dataset = CsvDataset(data_path,
+                            img_dir,
+                            transform,
+                            tokenizer)
+
+optimizer = torch.optim.Adam(model.parameters(), lr=1e-5)
+trainer.train(model,
+              optimizer=optimizer,
+              train_dataset=train_dataset,
+              collate_fn=collate_fn)
+

examples/clip/train_clip_deepspeed.py

@@ -0,0 +1,48 @@
+import torch
+from flagai.data.dataset.mm.clip_dataset import CsvDataset, clip_transform, collate_fn
+from flagai.trainer import Trainer
+from flagai.auto_model.auto_loader import AutoLoader
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+# cd examples/clip
+data_path = "./data/pairs.csv"#"/mnt/datasets/multimodal/ConceptualCaptions/Train_GCC-training_output.csv"
+img_dir = "./data/img"#"/mnt/datasets/multimodal/ConceptualCaptions"
+
+trainer = Trainer(
+    env_type="deepspeed",
+    experiment_name="clip",
+    batch_size=64,
+    num_gpus=2,
+    fp16=True,
+    gradient_accumulation_steps=1,
+    lr=1e-4,
+    weight_decay=1e-5,
+    epochs=5,
+    log_interval=1,
+    load_dir=None,
+    pytorch_device=device,
+    save_dir="clip_deepspeed",
+    save_interval=1000,
+    num_checkpoints=1,
+    hostfile="./deepspeed/hostfile",
+    training_script=__file__,
+    deepspeed_config="./deepspeed/deepspeed.json"
+)
+loader = AutoLoader(task_name="txt_img_matching",#contrastive learning
+                    model_name="clip-base-p32-224",
+                    )
+model = loader.get_model()
+tokenizer = loader.get_tokenizer()
+
+transform = clip_transform(img_size=model.image_size)
+train_dataset = CsvDataset(data_path,
+                            img_dir,
+                            transform,
+                            tokenizer)
+
+optimizer = torch.optim.Adam(model.parameters(), lr=1e-5)
+trainer.train(model,
+              optimizer=optimizer,
+              train_dataset=train_dataset,
+              collate_fn=collate_fn)
+

examples/glm_poetry_generation/generate.py

@@ -2,19 +2,17 @@
 #
 # Licensed under the Apache License, Version 2.0 (the "License")
 import torch
-import sys
-sys.path.append("/mnt/wchh/FlagAI-internal")
 from flagai.auto_model.auto_loader import AutoLoader
 from flagai.model.predictor.predictor import Predictor
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
 # Note "./checkpoints_poetry/{}/mp_rank_00_model_states.pt", {} is a directory in the checkpoints_poetry.
-model_save_path = "/mnt/finetune_models/glm_poetry/mp_rank_00_model_states.pt"#"./checkpoints_poetry/1/mp_rank_00_model_states.pt"
+model_save_path = "./checkpoints_poetry/1/mp_rank_00_model_states.pt"
 
 auto_loader = AutoLoader("seq2seq",
                          model_name="GLM-large-ch",
-                         model_dir="/data/chkpt/")
+                         model_dir="/data/state_dict/")
 model = auto_loader.get_model()
 tokenizer = auto_loader.get_tokenizer()
 
@@ -29,7 +27,7 @@
 for text in test_data:
     print('===============================================\n')
     print(text, ":")
-    for i in range(4):  #generate several times
+    for i in range(1):  #generate several times
         print("--------------sample %d :-------------------" % (i))
         print('-----------random sample: --------------')
         print(
@@ -39,9 +37,9 @@
                                                     top_p=.1,
                                                     repetition_penalty=4.0,
                                                     temperature=1.2))
-    print('-----------beam search: --------------')
-    print(
-        predictor.predict_generate_beamsearch(text,
-                                                out_max_length=66,
-                                                beam_size=10))
-    print()
+        print('-----------beam search: --------------')
+        print(
+            predictor.predict_generate_beamsearch(text,
+                                                  out_max_length=66,
+                                                  beam_size=10))
+        print()

flagai/auto_model/auto_loader.py

@@ -54,7 +54,7 @@ def __getattr__(self, name):
     "opt_seq2seq": ("flagai.model.opt_model","OPTModel"),
     "opt_lm": ("flagai.model.opt_model","OPTModel"),
     "vit_classification": ("flagai.model.vision.vit", "VisionTransformer"),
-    "clip_cl": ("flagai.model.mm.clip_model", "CLIP"),
+    "clip_txt_img_matching": ("flagai.model.mm.clip_model", "CLIP"),
 }
 
 # 4 columns : 1-package name,  2-class name, 3-model brief name, 4-model type
@@ -86,7 +86,10 @@ def __getattr__(self, name):
     "vit-large-p32-224":["flagai.model.vision.vit", "VisionTransformer", "vit", "vision"],
     "vit-large-p32-384":["flagai.model.vision.vit", "VisionTransformer", "vit", "vision"],
 
-    "clip-base-p32-224":["flagai.model.mm.clip_model", "CLIP", "clip", "mm"]
+    "clip-base-p32-224":["flagai.model.mm.clip_model", "CLIP", "clip", "mm"],
+    "clip-base-p16-224":["flagai.model.mm.clip_model", "CLIP", "clip", "mm"],
+    "clip-large-p14-224":["flagai.model.mm.clip_model", "CLIP", "clip", "mm"],
+    "clip-large-p14-336":["flagai.model.mm.clip_model", "CLIP", "clip", "mm"]
 }
 
 # 2 columns : 1-package name,  2-class name
@@ -118,7 +121,10 @@ def __getattr__(self, name):
     "opt-30b-en": ["flagai.data.tokenizer.opt.opt_en_tokenizer","OPTTokenizer"],
     "opt-66b-en": ["flagai.data.tokenizer.opt.opt_en_tokenizer","OPTTokenizer"],
 
-    "clip-base-p32-224":["flagai.data.tokenizer.clip.tokenizer", "ClipTokenizer"]
+    "clip-base-p32-224":["flagai.data.tokenizer.clip.tokenizer", "ClipTokenizer"],
+    "clip-base-p16-224":["flagai.data.tokenizer.clip.tokenizer", "ClipTokenizer"],
+    "clip-large-p14-224":["flagai.data.tokenizer.clip.tokenizer", "ClipTokenizer"],
+    "clip-large-p14-336":["flagai.data.tokenizer.clip.tokenizer", "ClipTokenizer"]
 
 }
 
@@ -198,7 +204,11 @@ def __init__(self,
                                  device=device,
                                  **kwargs)
 
-        model_id = _get_model_id(model_name)
+        try:
+            model_id = _get_model_id(model_name)
+        except:
+            print("Model hub is not reachable!")
+            model_id = -1
 
         print("*"*20, task_name, model_id, model_name)
         if model_type == "nlp":
@@ -242,7 +252,6 @@ def __init__(self,
             tokenizer_class = getattr(LazyImport(tokenizer_class[0]),
                                       tokenizer_class[1])
             if brief_model_name == "clip":
-                model_id = _get_model_id(model_name)
                 vocab_file = os.path.join(download_path, 'bpe_simple_vocab_16e6.txt.gz')
                 if not os.path.exists(vocab_file):
                     vocab_file = _get_vocab_path(download_path, "bpe_simple_vocab_16e6.txt.gz", model_id)