File size: 5,446 Bytes
aa8012e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
""" OpenAI pretrained model functions

Adapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.
"""

import os
import warnings
from typing import List, Optional, Union

import torch

from .model import build_model_from_openai_state_dict, convert_weights_to_lp, get_cast_dtype
from .pretrained import get_pretrained_url, list_pretrained_models_by_tag, download_pretrained_from_url

__all__ = ["list_openai_models", "load_openai_model"]


def list_openai_models() -> List[str]:
    """Returns the names of available CLIP models"""
    return list_pretrained_models_by_tag('openai')


def load_openai_model(
        name: str,
        precision: Optional[str] = None,
        device: Optional[Union[str, torch.device]] = None,
        jit: bool = True,
        cache_dir: Optional[str] = None,
):
    """Load a CLIP model

    Parameters
    ----------
    name : str
        A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict
    precision: str
        Model precision, if None defaults to 'fp32' if device == 'cpu' else 'fp16'.
    device : Union[str, torch.device]
        The device to put the loaded model
    jit : bool
        Whether to load the optimized JIT model (default) or more hackable non-JIT model.
    cache_dir : Optional[str]
        The directory to cache the downloaded model weights

    Returns
    -------
    model : torch.nn.Module
        The CLIP model
    preprocess : Callable[[PIL.Image], torch.Tensor]
        A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input
    """
    if device is None:
        device = "cuda" if torch.cuda.is_available() else "cpu"
    if precision is None:
        precision = 'fp32' if device == 'cpu' else 'fp16'

    if get_pretrained_url(name, 'openai'):
        model_path = download_pretrained_from_url(get_pretrained_url(name, 'openai'), cache_dir=cache_dir)
    elif os.path.isfile(name):
        model_path = name
    else:
        raise RuntimeError(f"Model {name} not found; available models = {list_openai_models()}")

    try:
        # loading JIT archive
        model = torch.jit.load(model_path, map_location=device if jit else "cpu").eval()
        state_dict = None
    except RuntimeError:
        # loading saved state dict
        if jit:
            warnings.warn(f"File {model_path} is not a JIT archive. Loading as a state dict instead")
            jit = False
        state_dict = torch.load(model_path, map_location="cpu")

    if not jit:
        # Build a non-jit model from the OpenAI jitted model state dict
        cast_dtype = get_cast_dtype(precision)
        try:
            model = build_model_from_openai_state_dict(state_dict or model.state_dict(), cast_dtype=cast_dtype)
        except KeyError:
            sd = {k[7:]: v for k, v in state_dict["state_dict"].items()}
            model = build_model_from_openai_state_dict(sd, cast_dtype=cast_dtype)

        # model from OpenAI state dict is in manually cast fp16 mode, must be converted for AMP/fp32/bf16 use
        model = model.to(device)
        if precision.startswith('amp') or precision == 'fp32':
            model.float()
        elif precision == 'bf16':
            convert_weights_to_lp(model, dtype=torch.bfloat16)

        return model

    # patch the device names
    device_holder = torch.jit.trace(lambda: torch.ones([]).to(torch.device(device)), example_inputs=[])
    device_node = [n for n in device_holder.graph.findAllNodes("prim::Constant") if "Device" in repr(n)][-1]

    def patch_device(module):
        try:
            graphs = [module.graph] if hasattr(module, "graph") else []
        except RuntimeError:
            graphs = []

        if hasattr(module, "forward1"):
            graphs.append(module.forward1.graph)

        for graph in graphs:
            for node in graph.findAllNodes("prim::Constant"):
                if "value" in node.attributeNames() and str(node["value"]).startswith("cuda"):
                    node.copyAttributes(device_node)

    model.apply(patch_device)
    patch_device(model.encode_image)
    patch_device(model.encode_text)

    # patch dtype to float32 (typically for CPU)
    if precision == 'fp32':
        float_holder = torch.jit.trace(lambda: torch.ones([]).float(), example_inputs=[])
        float_input = list(float_holder.graph.findNode("aten::to").inputs())[1]
        float_node = float_input.node()

        def patch_float(module):
            try:
                graphs = [module.graph] if hasattr(module, "graph") else []
            except RuntimeError:
                graphs = []

            if hasattr(module, "forward1"):
                graphs.append(module.forward1.graph)

            for graph in graphs:
                for node in graph.findAllNodes("aten::to"):
                    inputs = list(node.inputs())
                    for i in [1, 2]:  # dtype can be the second or third argument to aten::to()
                        if inputs[i].node()["value"] == 5:
                            inputs[i].node().copyAttributes(float_node)

        model.apply(patch_float)
        patch_float(model.encode_image)
        patch_float(model.encode_text)
        model.float()

    # ensure image_size attr available at consistent location for both jit and non-jit
    model.visual.image_size = model.input_resolution.item()
    return model