Upload folder using huggingface_hub

added_tokens.json

@@ -0,0 +1,11 @@
+{
+  "</p>": 92552,
+  "<p>": 92551,
+  "<|action_end|>": 92547,
+  "<|action_start|>": 92546,
+  "<|im_end|>": 92545,
+  "<|im_start|>": 92544,
+  "<|interpreter|>": 92548,
+  "<|plugin|>": 92549,
+  "[SEG]": 92550
+}

config.json

@@ -0,0 +1,39 @@
+{
+  "_name_or_path": "MBZUAI/GeoPixel-7B-RES",
+  "architectures": [
+    "GeoPixelForCausalLM"
+  ],
+  "attn_implementation": "flash_attention_2",
+  "auto_map": {
+    "AutoConfig": "configuration_internlm_xcomposer2.InternLMXcomposer2Config",
+    "AutoModel": "modeling_internlm_xcomposer2.InternLMXComposer2ForCausalLM",
+    "AutoModelForCausalLM": "geopixel.GeoPixelForCausalLM"
+  },
+  "bias": false,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "max_length": 16384,
+  "max_position_embeddings": 24576,
+  "model_type": "internlm2",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 32,
+  "num_key_value_heads": 8,
+  "out_dim": 256,
+  "pad_token_id": 0,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": {
+    "factor": 2.0,
+    "type": "dynamic"
+  },
+  "rope_theta": 1000000,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "train_mask_decoder": true,
+  "transformers_version": "4.33.2",
+  "use_cache": false,
+  "vocab_size": 92553
+}

configuration_internlm_xcomposer2.py

@@ -0,0 +1,150 @@
+# coding=utf-8
+# Copyright (c) The InternLM team and The HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on transformers/src/transformers/models/llama/configuration_llama.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" InternLM2 model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+INTERNLM2_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+class InternLMXcomposer2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`InternLM2Model`]. It is used to instantiate
+    an InternLM2 model according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the InternLM2-7B.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 32000):
+            Vocabulary size of the InternLM2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`InternLM2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 11008):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
+            `num_attention_heads`.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings(`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        Example:
+
+    """
+    model_type = "internlm2"
+    _auto_class = "AutoConfig"
+
+    def __init__(  # pylint: disable=W0102
+        self,
+        vocab_size=103168,
+        hidden_size=4096,
+        intermediate_size=11008,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=None,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        tie_word_embeddings=False,
+        bias=True,
+        rope_theta=10000,
+        rope_scaling=None,
+        attn_implementation="flash_attention_2",
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.bias = bias
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+        self.num_key_value_heads = num_key_value_heads
+
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self._rope_scaling_validation()
+
+        self.attn_implementation = attn_implementation
+        if self.attn_implementation is None:
+            self.attn_implementation = "flash_attention_2"
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+    def _rope_scaling_validation(self):
+        """
+        Validate the `rope_scaling` configuration.
+        """
+        if self.rope_scaling is None:
+            return
+
+        if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2:
+            raise ValueError(
+                "`rope_scaling` must be a dictionary with with two fields, `type` and `factor`, "
+                f"got {self.rope_scaling}"
+            )
+        rope_scaling_type = self.rope_scaling.get("type", None)
+        rope_scaling_factor = self.rope_scaling.get("factor", None)
+        if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]:
+            raise ValueError(
+                f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}"
+            )
+        if rope_scaling_factor is None or not isinstance(rope_scaling_factor, float) or rope_scaling_factor < 1.0:
+            raise ValueError(f"`rope_scaling`'s factor field must be a float >= 1, got {rope_scaling_factor}")

generation_config.json

@@ -0,0 +1,9 @@
+{
+  "_from_model_config": true,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "max_length": 16384,
+  "pad_token_id": 2,
+  "transformers_version": "4.33.2",
+  "use_cache": false
+}

geopixel.py

@@ -0,0 +1,418 @@
+from typing import List, Optional, Tuple, Union
+
+import os
+import torch
+import numpy as np
+import torch.nn as nn
+import matplotlib.pyplot as plt
+from PIL import Image
+import torch.nn.functional as F
+from transformers.modeling_outputs import CausalLMOutputWithPast
+from model.IXC.modeling_internlm_xcomposer2 import InternLMXComposer2ForCausalLM
+from model.IXC.modeling_internlm2 import InternLM2Model
+from model.sam2.build_sam import build_sam2_hf
+from model.sam2.utils.transforms import SAM2Transforms
+from transformers import TextStreamer
+try:
+    from transformers.generation.streamers import BaseStreamer
+except:  # noqa # pylint: disable=bare-except
+    BaseStreamer = None
+
+
+def dice_loss(
+    inputs: torch.Tensor,
+    targets: torch.Tensor,
+    num_masks: float,
+    scale=1000,  # 100000.0,
+    eps=1e-6,
+):
+    """
+    Compute the DICE loss, similar to generalized IOU for masks
+    Args:
+        inputs: A float tensor of arbitrary shape.
+                The predictions for each example.
+        targets: A float tensor with the same shape as inputs. Stores the binary
+                 classification label for each element in inputs
+                (0 for the negative class and 1 for the positive class).
+    """
+    inputs = inputs.sigmoid()
+    inputs = inputs.flatten(1, 2)
+    targets = targets.flatten(1, 2)
+    numerator = 2 * (inputs / scale * targets).sum(-1)
+    denominator = (inputs / scale).sum(-1) + (targets / scale).sum(-1)
+    loss = 1 - (numerator + eps) / (denominator + eps)
+    loss = loss.sum() / (num_masks + 1e-8)
+    return loss
+
+
+def sigmoid_ce_loss(
+    inputs: torch.Tensor,
+    targets: torch.Tensor,
+    num_masks: float,
+):
+    """
+    Args:
+        inputs: A float tensor of arbitrary shape.
+                The predictions for each example.
+        targets: A float tensor with the same shape as inputs. Stores the binary
+                 classification label for each element in inputs
+                (0 for the negative class and 1 for the positive class).
+    Returns:
+        Loss tensor
+    """
+    loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
+    loss = loss.flatten(1, 2).mean(1).sum() / (num_masks + 1e-8)
+    return loss
+
+
+class GeoPixelMetaModel:
+    def __init__(
+        self,
+        config,
+        **kwargs,
+    ):
+        super(GeoPixelMetaModel, self).__init__(config)
+        self.config = config
+        self.config.train_mask_decoder = getattr(self.config, "train_mask_decoder", kwargs.get("train_mask_decoder", False))
+        self.config.out_dim = getattr(self.config, "out_dim", kwargs.get("out_dim", 256))
+        self.vision_pretrained = kwargs.get("vision_pretrained", None)
+        self.initialize_geopixel_modules(self.config)
+
+    def initialize_geopixel_modules(self, config):
+        # grounding vision model
+        self.visual_model = build_sam2_hf(self.vision_pretrained)
+
+        self._transform = SAM2Transforms(
+                    resolution=self.visual_model.image_size,
+                    mask_threshold=0.0,
+                    max_hole_area=0.0,
+                    max_sprinkle_area=0.0,
+                )
+        # Spatial dim for backbone feature maps
+        self._bb_feat_sizes = [
+            (256, 256),
+            (128, 128),
+            (64, 64),
+        ]
+        
+        for param in self.visual_model.parameters():
+            param.requires_grad = False
+
+        if config.train_mask_decoder:
+            self.visual_model.sam_mask_decoder.train()
+            for param in self.visual_model.sam_mask_decoder.parameters():
+                param.requires_grad = True
+
+        # text projection layer
+        in_dim = config.hidden_size 
+        out_dim = config.out_dim    
+        text_projection_layers = [
+            nn.Linear(in_dim, in_dim),
+            nn.ReLU(inplace=True),
+            nn.Linear(in_dim, out_dim),
+            nn.Dropout(0.0),
+        ]
+        self.text_hidden_fcs = nn.ModuleList([nn.Sequential(*text_projection_layers)])
+        self.text_hidden_fcs.train()
+        for param in self.text_hidden_fcs.parameters():
+            param.requires_grad = True
+
+
+class GeoPixelModel(GeoPixelMetaModel, InternLM2Model):
+    def __init__(
+        self,
+        config,
+        **kwargs,
+    ):
+        super(GeoPixelModel, self).__init__(config, **kwargs)
+        self.config.use_cache = False
+
+
+class GeoPixelForCausalLM(InternLMXComposer2ForCausalLM):
+    def __init__(self,config,**kwargs,):
+        
+        self.ce_loss_weight = kwargs.pop("ce_loss_weight", None)
+        self.dice_loss_weight = kwargs.pop("dice_loss_weight", None)
+        self.bce_loss_weight = kwargs.pop("bce_loss_weight", None)
+        self.seg_token_idx = kwargs.pop("seg_token_idx")
+
+        super().__init__(config)
+        self.model = GeoPixelModel(config, **kwargs)
+        self.vocab_size = config.vocab_size
+        self.output = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.post_init()
+
+    def encode_g_img(self, image):
+        """
+        Calculates the image embeddings for the provided image
+        Arguments:
+          image (np.ndarray or str)
+        """
+        if image is None:
+            return None
+        if isinstance(image, str):
+            _, ext = os.path.splitext(image)
+            if ext.lower() in {'.jpg', '.jpeg', '.png', '.gif', '.bmp', '.webp','.tif'}:
+                image = Image.open(image)
+                w, h = image.size
+                _orig_hw = [(h, w)] 
+            else:
+                print ('Unknow input format', image)
+                return None
+        else:
+            assert isinstance(image, torch.Tensor)
+            _orig_hw = [image.shape[:2]]
+        image = self.model._transform(image)
+        image = image[None, ...].to(self.device)
+        assert ( len(image.shape) == 4 and image.shape[1] == 3), f"image must be of size 1x3xHxW, got {image.shape}"
+        features = self.get_visual_embs(image)   
+        return features,_orig_hw
+
+    def get_visual_embs(self, img_batch: torch.FloatTensor):
+        with torch.no_grad():
+            torch.cuda.empty_cache()
+            img_batch = img_batch.to(self.device)
+            batch_size = img_batch.shape[0]
+            assert (
+                len(img_batch.shape) == 4 and img_batch.shape[1] == 3
+            ), f"grounding_img_batch must be of size Bx3xHxW, got {img_batch.shape}"
+            backbone_out = self.model.visual_model.forward_image(img_batch)
+            _, vision_feats, _, _ = self.model.visual_model._prepare_backbone_features(backbone_out)
+            if self.model.visual_model.directly_add_no_mem_embed:
+                vision_feats[-1] = vision_feats[-1] + self.model.visual_model.no_mem_embed
+            feats = [
+                feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
+                for feat, feat_size in zip(vision_feats[::-1], self.model._bb_feat_sizes[::-1])
+            ][::-1]
+            features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        return features
+    
+    def forward(self, **kwargs):
+        return super().forward(**kwargs) if "past_key_values" in kwargs else self.model_forward(**kwargs)
+    
+    def model_forward(
+            self,
+            inference: bool = False,
+            **kwargs,
+    ):
+        samples = kwargs.get('samples', None)
+        if samples and samples['data_type'][0] == 'grounding': 
+            kwargs['output_hidden_states'] = True
+            kwargs['use_cache'] = False
+
+            torch.cuda.empty_cache()
+            outputs = super().forward(**kwargs)
+
+            if inference:
+                assert len(samples['text_input']) == 1 and len(samples['image'][0]) == 1 #single image and single query
+                output_hidden_states = [outputs.hidden_states]
+                outputs = None
+            else:
+                output_hidden_states = outputs.hidden_states
+
+            hidden_states = []
+            assert len(self.model.text_hidden_fcs) == 1
+            hidden_states.append(self.model.text_hidden_fcs[0](output_hidden_states[-1]))
+            last_hidden_state = torch.stack(hidden_states, dim=-1).sum(dim=-1)
+
+            seg_token_mask = outputs.seg_token_mask
+            pred_embeddings = [states[masks] for states, masks in zip(last_hidden_state, seg_token_mask)]
+            image_g_batch = torch.cat(samples['image_g'][0],dim = 0)
+            image_g_features = self.get_visual_embs(image_g_batch)
+            ori_hw = samples['ori_hw'][0]
+            all_pred_masks = []
+            for i in range(len(pred_embeddings)): #(bs,)
+                if (pred_embeddings[i].numel()== 0):
+                    pred_masks.append([])
+                    continue
+                (sparse_embeddings, dense_embeddings,) = self.model.visual_model.sam_prompt_encoder(
+                    points=None,
+                    boxes=None,
+                    masks=None,
+                    text_embeds=pred_embeddings[i].unsqueeze(1),
+                )
+                batch_mode = (pred_embeddings[i].shape[0]>1)
+                high_res_features = [
+                    feat_level[i].unsqueeze(0)
+                    for feat_level in image_g_features["high_res_feats"]
+                ]
+                sparse_embeddings = sparse_embeddings.to(pred_embeddings[i].dtype)
+                image_g_embeds = image_g_features['image_embed'][i].unsqueeze(0).to(torch.bfloat16)
+                low_res_masks, _, _ , _ = self.model.visual_model.sam_mask_decoder(
+                    image_embeddings=image_g_embeds,
+                    image_pe=self.model.visual_model.sam_prompt_encoder.get_dense_pe(),
+                    sparse_prompt_embeddings=sparse_embeddings,
+                    dense_prompt_embeddings=dense_embeddings,
+                    repeat_image=batch_mode,
+                    multimask_output=False,
+                    high_res_features=high_res_features,
+                )
+                pred_masks = self.model._transform.postprocess_masks(
+                    low_res_masks,
+                    ori_hw[i],
+                )
+                all_pred_masks.append(pred_masks[:, 0])
+                
+
+            model_output = outputs
+            gt_masks =  samples['masks'][0]
+            pred_masks = all_pred_masks 
+
+            if inference:
+                return {
+                    "pred_masks": pred_masks,
+                    "gt_masks": gt_masks,
+                }
+
+            ce_loss = model_output.loss
+            ce_loss = ce_loss * self.ce_loss_weight
+            mask_bce_loss = 0
+            mask_dice_loss = 0
+            num_masks = 0
+
+            for batch_idx in range(len(pred_masks)): # for every image
+                cur_gt_masks = torch.stack(
+                    [
+                        torch.from_numpy(gt_mask).to(dtype=pred_masks[batch_idx].dtype, device=pred_masks[batch_idx].device)
+                        for gt_mask in gt_masks[batch_idx]
+                    ],
+                    dim=0
+                ) # expected (bs,H,W)
+                cur_pred_masks = pred_masks[batch_idx]
+                assert (
+                    cur_gt_masks.shape[0] == cur_pred_masks.shape[0]
+                ), "gt_masks.shape: {}, pred_masks.shape: {}".format(
+                    cur_gt_masks.shape, cur_pred_masks.shape
+                )
+                mask_bce_loss += (
+                    sigmoid_ce_loss(cur_pred_masks, cur_gt_masks, num_masks=cur_gt_masks.shape[0])
+                    * cur_gt_masks.shape[0]
+                )
+                mask_dice_loss += (
+                    dice_loss(cur_pred_masks, cur_gt_masks, num_masks=cur_gt_masks.shape[0])
+                    * cur_gt_masks.shape[0]
+                )
+                num_masks += cur_gt_masks.shape[0] 
+
+            mask_bce_loss = self.bce_loss_weight * mask_bce_loss / (num_masks + 1e-8)
+            mask_dice_loss = self.dice_loss_weight * mask_dice_loss / (num_masks + 1e-8)
+            mask_loss = mask_bce_loss + mask_dice_loss
+
+            loss = ce_loss + mask_loss
+            outputs = CausalLMOutputWithPast(
+                loss=loss,
+                logits=model_output.logits,
+                past_key_values=model_output.past_key_values,
+                hidden_states=output_hidden_states,
+                attentions=model_output.attentions,
+            )
+            outputs.ce_loss = ce_loss
+            outputs.mask_bce_loss = mask_bce_loss
+            outputs.mask_dice_loss = mask_dice_loss
+            outputs.mask_loss = mask_loss
+        else: 
+            outputs =  super().forward(**kwargs)
+        return outputs
+
+    def evaluate(
+        self,
+        tokenizer,
+        query: str,
+        images: List[Tuple[str, str]] = [],
+        hd_num: int = 9,
+        history: List[Tuple[str, str]] = [],
+        max_new_tokens: int = 1024,
+        stream: bool = False,
+        **kwargs,
+    ):
+        with torch.no_grad():
+            inputs, im_mask, _ = self.interleav_wrap_chat(query, images, history=history, hd_num=hd_num)        
+            inputs = {
+                k: v.to(self.device)
+                for k, v in inputs.items() if torch.is_tensor(v)
+            }
+            eos_token_id = [
+                tokenizer.eos_token_id,
+                #tokenizer.convert_tokens_to_ids(['[UNUSED_TOKEN_145]'])[0]
+            ]
+            all_pred_masks = []
+            
+            if stream:
+                streamer = TextStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
+            else: 
+                streamer = None
+
+            outputs = self.generate(
+                **inputs,
+                max_new_tokens=max_new_tokens,
+                im_mask=im_mask,
+                input_ids = None,
+                streamer= streamer,
+                num_beams=1,
+                do_sample=False,
+                temperature=1.0,
+                top_p= 1.0,
+                top_k = 0,
+                eos_token_id=eos_token_id,
+                repetition_penalty=1.0,
+                infer_mode = 'base',
+                output_hidden_states=True,
+                return_dict_in_generate=True,
+                **kwargs,
+            )
+            output_ids = outputs['sequences']
+            response = tokenizer.decode(output_ids[0].cpu().tolist(), skip_special_tokens=True)
+            response = response.replace("[UNUSED_TOKEN_145]","")
+            history = history + [(query, response)]
+            if len(images)==1 and isinstance(images[0], str): 
+                output_hidden_states = outputs.hidden_states[-1] 
+                seg_token_mask = output_ids[:, 1:-1] == self.seg_token_idx
+                inputs_embeds_len = inputs['inputs_embeds'].size(1)
+                seg_token_mask = torch.cat(
+                    [
+                        torch.zeros((seg_token_mask.shape[0], inputs_embeds_len)).bool().cuda(),
+                        seg_token_mask,
+                    ],
+                    dim=1,
+                )
+                hidden_states = []
+                assert len(self.model.text_hidden_fcs) == 1
+                hidden_states.append(self.model.text_hidden_fcs[0](output_hidden_states))
+                last_hidden_state = torch.stack(hidden_states, dim=-1).sum(dim=-1)
+                pred_embeddings = [states[masks] for states, masks in zip(last_hidden_state, seg_token_mask)]
+                image_g_features, ori_hw = self.encode_g_img(images[0]) 
+
+                for i in range(len(pred_embeddings)):
+                    if (pred_embeddings[i].numel()== 0):
+                        all_pred_masks.append([])
+                        continue
+                    (sparse_embeddings,dense_embeddings,) = self.model.visual_model.sam_prompt_encoder(
+                        points=None,
+                        boxes=None,
+                        masks=None,
+                        text_embeds=pred_embeddings[i].unsqueeze(1),
+                    )
+                    batch_mode = (pred_embeddings[i].shape[0]>1)
+                    high_res_features = [
+                        feat_level[i].unsqueeze(0)
+                        for feat_level in image_g_features["high_res_feats"]
+                    ]
+                    sparse_embeddings = sparse_embeddings.to(pred_embeddings[i].dtype)
+                    image_g_embeds = image_g_features['image_embed'][i].unsqueeze(0).to(torch.bfloat16)
+
+                    low_res_masks, _, _ , _  = self.model.visual_model.sam_mask_decoder(
+                        image_embeddings=image_g_embeds,
+                        image_pe=self.model.visual_model.sam_prompt_encoder.get_dense_pe(),
+                        sparse_prompt_embeddings=sparse_embeddings,
+                        dense_prompt_embeddings=dense_embeddings,
+                        repeat_image=batch_mode,
+                        multimask_output=False,
+                        high_res_features=high_res_features,
+                    )
+                    pred_masks = self.model._transform.postprocess_masks(
+                        low_res_masks,
+                        ori_hw[i],
+                    )
+                    all_pred_masks.append(pred_masks[:, 0])
+
+        return response, all_pred_masks

pytorch_model-00001-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6eedea61e25c9184b4e2989a3c6d79982a6f4ee263041401cfe443319d04863
+size 9968330657

pytorch_model-00002-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0df18331981448f806bbea1932341d233c2dc9f71a888d16586001efd5c189a4
+size 9999750322

pytorch_model-00003-of-00003.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4cb744be7e9cc162d6947b63c95cad2370d844b6e5aa76a58a66f61f1bb2d54a
+size 2709063690

special_tokens_map.json

@@ -0,0 +1,32 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|action_start|>",
+    "<|action_end|>",
+    "<|interpreter|>",
+    "<|plugin|>"
+  ],
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": "<unk>",
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenization_internlm2.py

@@ -0,0 +1,236 @@
+# coding=utf-8
+# Copyright (c) The InternLM team and The HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on transformers/src/transformers/models/llama/tokenization_llama.py
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Tokenization classes for InternLM."""
+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+from transformers.tokenization_utils import PreTrainedTokenizer
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "./tokenizer.model"}
+
+PRETRAINED_VOCAB_FILES_MAP = {}
+
+
+# Modified from transformers.model.llama.tokenization_llama.LlamaTokenizer
+class InternLM2Tokenizer(PreTrainedTokenizer):
+    """
+    Construct a InternLM2 tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    model_input_names = ["input_ids", "attention_mask"]
+    _auto_class = "AutoTokenizer"
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        pad_token="</s>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        decode_with_prefix_space=False,
+        clean_up_tokenization_spaces=False,
+        **kwargs,
+    ):
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        self.vocab_file = vocab_file
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self.decode_with_prefix_space = decode_with_prefix_space
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+        self._no_prefix_space_tokens = None
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    @property
+    def no_prefix_space_tokens(self):
+        if self._no_prefix_space_tokens is None:
+            vocab = self.convert_ids_to_tokens(list(range(self.vocab_size)))
+            self._no_prefix_space_tokens = {i for i, tok in enumerate(vocab) if not tok.startswith("▁")}
+        return self._no_prefix_space_tokens
+
+    @property
+    def vocab_size(self):
+        """Returns vocab size"""
+        return self.sp_model.get_piece_size()
+
+    @property
+    def bos_token_id(self) -> Optional[int]:
+        return self.sp_model.bos_id()
+
+    @property
+    def eos_token_id(self) -> Optional[int]:
+        return self.sp_model.eos_id()
+
+    def get_vocab(self):
+        """Returns vocab as a dict"""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text):
+        """Returns a tokenized string."""
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def _maybe_add_prefix_space(self, tokens, decoded):
+        if tokens and tokens[0] not in self.no_prefix_space_tokens:
+            return " " + decoded
+        else:
+            return decoded
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for token in tokens:
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        out_string = self.clean_up_tokenization(out_string)
+        out_string = self._maybe_add_prefix_space(tokens=tokens, decoded=out_string)
+        return out_string[1:]
+
+    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        if self.add_bos_token:
+            bos_token_ids = [self.bos_token_id]
+        else:
+            bos_token_ids = []
+
+        output = bos_token_ids + token_ids_0
+
+        if token_ids_1 is not None:
+            output = output + token_ids_1
+
+        if self.add_eos_token:
+            output = output + [self.eos_token_id]
+
+        return output
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1)) + [1]
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. T5 does not make
+        use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of zeros.
+        """
+        eos = [self.eos_token_id]
+
+        if token_ids_1 is None:
+            return len(token_ids_0 + eos) * [0]
+        return len(token_ids_0 + eos + token_ids_1 + eos) * [0]

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f868398fc4e05ee1e8aeba95ddf18ddcc45b8bce55d5093bead5bbf80429b48b
+size 1477754

tokenizer_config.json

@@ -0,0 +1,99 @@
+{
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "92538": {
+      "content": "<|plugin|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "92539": {
+      "content": "<|interpreter|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "92540": {
+      "content": "<|action_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "92541": {
+      "content": "<|action_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "92542": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "92543": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|action_start|>",
+    "<|action_end|>",
+    "<|interpreter|>",
+    "<|plugin|>"
+  ],
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_internlm2.InternLM2Tokenizer",
+      null
+    ]
+  },
+  "bos_token": "<s>",
+  "chat_template": "{{ bos_token }}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "</s>",
+  "padding_side": "right",
+  "tokenizer_class": "InternLM2Tokenizer",
+  "unk_token": "<unk>"
+}