repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
mixture-of-diffusers
mixture-of-diffusers-master/generate_grid_from_json.py
import argparse import datetime from diffusers import LMSDiscreteScheduler, DDIMScheduler import json from pathlib import Path import torch from mixdiff.tiling import StableDiffusionTilingPipeline def generate_grid(generation_arguments): model_id = "CompVis/stable-diffusion-v1-4" # Prepared scheduler if generation_arguments["scheduler"] == "ddim": scheduler = DDIMScheduler() elif generation_arguments["scheduler"] == "lms": scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000) else: raise ValueError(f"Unrecognized scheduler {generation_arguments['scheduler']}") pipe = StableDiffusionTilingPipeline.from_pretrained(model_id, scheduler=scheduler, use_auth_token=True).to("cuda:0") pipeargs = { "guidance_scale": generation_arguments["gc"], "num_inference_steps": generation_arguments["steps"], "seed": generation_arguments["seed"], "prompt": generation_arguments["prompt"], "tile_height": generation_arguments["tile_height"], "tile_width": generation_arguments["tile_width"], "tile_row_overlap": generation_arguments["tile_row_overlap"], "tile_col_overlap": generation_arguments["tile_col_overlap"], "guidance_scale_tiles": generation_arguments["gc_tiles"], "cpu_vae": generation_arguments["cpu_vae"] if "cpu_vae" in generation_arguments else False, } if "seed_tiles" in generation_arguments: pipeargs = {**pipeargs, "seed_tiles": generation_arguments["seed_tiles"]} if "seed_tiles_mode" in generation_arguments: pipeargs = {**pipeargs, "seed_tiles_mode": generation_arguments["seed_tiles_mode"]} if "seed_reroll_regions" in generation_arguments: pipeargs = {**pipeargs, "seed_reroll_regions": generation_arguments["seed_reroll_regions"]} image = pipe(**pipeargs)["sample"][0] outname = "output" outpath = "./outputs" Path(outpath).mkdir(parents=True, exist_ok=True) image.save(f"{outpath}/{outname}.png") if __name__ == "__main__": parser = argparse.ArgumentParser(description='Generate a stable diffusion grid using a JSON file with all configuration parameters.') parser.add_argument('config', type=str, help='Path to configuration file') args = parser.parse_args() with open(args.config, "r") as f: generation_arguments = json.load(f) generate_grid(generation_arguments)
2,454
49.102041
145
py
mixture-of-diffusers
mixture-of-diffusers-master/mixdiff/canvas.py
from copy import deepcopy from dataclasses import asdict, dataclass from enum import Enum import numpy as np from numpy import pi, exp, sqrt import re import torch from torchvision.transforms.functional import resize from tqdm.auto import tqdm from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer from typing import List, Optional, Tuple, Union from diffusers.models import AutoencoderKL, UNet2DConditionModel from diffusers.pipeline_utils import DiffusionPipeline from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker class MaskModes(Enum): """Modes in which the influence of diffuser is masked""" CONSTANT = "constant" GAUSSIAN = "gaussian" QUARTIC = "quartic" # See https://en.wikipedia.org/wiki/Kernel_(statistics) class RerollModes(Enum): """Modes in which the reroll regions operate""" RESET = "reset" # Completely reset the random noise in the region EPSILON = "epsilon" # Alter slightly the latents in the region @dataclass class CanvasRegion: """Class defining a rectangular region in the canvas""" row_init: int # Region starting row in pixel space (included) row_end: int # Region end row in pixel space (not included) col_init: int # Region starting column in pixel space (included) col_end: int # Region end column in pixel space (not included) region_seed: int = None # Seed for random operations in this region noise_eps: float = 0.0 # Deviation of a zero-mean gaussian noise to be applied over the latents in this region. Useful for slightly "rerolling" latents def __post_init__(self): # Initialize arguments if not specified if self.region_seed is None: self.region_seed = np.random.randint(9999999999) # Check coordinates are non-negative for coord in [self.row_init, self.row_end, self.col_init, self.col_end]: if coord < 0: raise ValueError(f"A CanvasRegion must be defined with non-negative indices, found ({self.row_init}, {self.row_end}, {self.col_init}, {self.col_end})") # Check coordinates are divisible by 8, else we end up with nasty rounding error when mapping to latent space for coord in [self.row_init, self.row_end, self.col_init, self.col_end]: if coord // 8 != coord / 8: raise ValueError(f"A CanvasRegion must be defined with locations divisible by 8, found ({self.row_init}-{self.row_end}, {self.col_init}-{self.col_end})") # Check noise eps is non-negative if self.noise_eps < 0: raise ValueError(f"A CanvasRegion must be defined noises eps non-negative, found {self.noise_eps}") # Compute coordinates for this region in latent space self.latent_row_init = self.row_init // 8 self.latent_row_end = self.row_end // 8 self.latent_col_init = self.col_init // 8 self.latent_col_end = self.col_end // 8 @property def width(self): return self.col_end - self.col_init @property def height(self): return self.row_end - self.row_init def get_region_generator(self, device="cpu"): """Creates a torch.Generator based on the random seed of this region""" # Initialize region generator return torch.Generator(device).manual_seed(self.region_seed) @property def __dict__(self): return asdict(self) @dataclass class DiffusionRegion(CanvasRegion): """Abstract class defining a region where some class of diffusion process is acting""" pass @dataclass class RerollRegion(CanvasRegion): """Class defining a rectangular canvas region in which initial latent noise will be rerolled""" reroll_mode: RerollModes = RerollModes.RESET.value @dataclass class Text2ImageRegion(DiffusionRegion): """Class defining a region where a text guided diffusion process is acting""" prompt: str = "" # Text prompt guiding the diffuser in this region guidance_scale: float = 7.5 # Guidance scale of the diffuser in this region. If None, randomize mask_type: MaskModes = MaskModes.GAUSSIAN.value # Kind of weight mask applied to this region mask_weight: float = 1.0 # Global weights multiplier of the mask tokenized_prompt = None # Tokenized prompt encoded_prompt = None # Encoded prompt def __post_init__(self): super().__post_init__() # Mask weight cannot be negative if self.mask_weight < 0: raise ValueError(f"A Text2ImageRegion must be defined with non-negative mask weight, found {self.mask_weight}") # Mask type must be an actual known mask if self.mask_type not in [e.value for e in MaskModes]: raise ValueError(f"A Text2ImageRegion was defined with mask {self.mask_type}, which is not an accepted mask ({[e.value for e in MaskModes]})") # Randomize arguments if given as None if self.guidance_scale is None: self.guidance_scale = np.random.randint(5, 30) # Clean prompt self.prompt = re.sub(' +', ' ', self.prompt).replace("\n", " ") def tokenize_prompt(self, tokenizer): """Tokenizes the prompt for this diffusion region using a given tokenizer""" self.tokenized_prompt = tokenizer(self.prompt, padding="max_length", max_length=tokenizer.model_max_length, truncation=True, return_tensors="pt") def encode_prompt(self, text_encoder, device): """Encodes the previously tokenized prompt for this diffusion region using a given encoder""" assert self.tokenized_prompt is not None, ValueError("Prompt in diffusion region must be tokenized before encoding") self.encoded_prompt = text_encoder(self.tokenized_prompt.input_ids.to(device))[0] @dataclass class Image2ImageRegion(DiffusionRegion): """Class defining a region where an image guided diffusion process is acting""" reference_image: torch.FloatTensor = None strength: float = 0.8 # Strength of the image def __post_init__(self): super().__post_init__() if self.reference_image is None: raise ValueError("Must provide a reference image when creating an Image2ImageRegion") if self.strength < 0 or self.strength > 1: raise ValueError(f'The value of strength should in [0.0, 1.0] but is {self.strength}') # Rescale image to region shape self.reference_image = resize(self.reference_image, size=[self.height, self.width]) def encode_reference_image(self, encoder, device, generator, cpu_vae=False): """Encodes the reference image for this Image2Image region into the latent space""" # Place encoder in CPU or not following the parameter cpu_vae if cpu_vae: # Note here we use mean instead of sample, to avoid moving also generator to CPU, which is troublesome self.reference_latents = encoder.cpu().encode(self.reference_image).latent_dist.mean.to(device) else: self.reference_latents = encoder.encode(self.reference_image.to(device)).latent_dist.sample(generator=generator) self.reference_latents = 0.18215 * self.reference_latents @property def __dict__(self): # This class requires special casting to dict because of the reference_image tensor. Otherwise it cannot be casted to JSON # Get all basic fields from parent class super_fields = {key: getattr(self, key) for key in DiffusionRegion.__dataclass_fields__.keys()} # Pack other fields return { **super_fields, "reference_image": self.reference_image.cpu().tolist(), "strength": self.strength } @dataclass class MaskWeightsBuilder: """Auxiliary class to compute a tensor of weights for a given diffusion region""" latent_space_dim: int # Size of the U-net latent space nbatch: int = 1 # Batch size in the U-net def compute_mask_weights(self, region: DiffusionRegion) -> torch.tensor: """Computes a tensor of weights for a given diffusion region""" MASK_BUILDERS = { MaskModes.CONSTANT.value: self._constant_weights, MaskModes.GAUSSIAN.value: self._gaussian_weights, MaskModes.QUARTIC.value: self._quartic_weights, } return MASK_BUILDERS[region.mask_type](region) def _constant_weights(self, region: DiffusionRegion) -> torch.tensor: """Computes a tensor of constant for a given diffusion region""" latent_width = region.latent_col_end - region.latent_col_init latent_height = region.latent_row_end - region.latent_row_init return torch.ones(self.nbatch, self.latent_space_dim, latent_height, latent_width) * region.mask_weight def _gaussian_weights(self, region: DiffusionRegion) -> torch.tensor: """Generates a gaussian mask of weights for tile contributions""" latent_width = region.latent_col_end - region.latent_col_init latent_height = region.latent_row_end - region.latent_row_init var = 0.01 midpoint = (latent_width - 1) / 2 # -1 because index goes from 0 to latent_width - 1 x_probs = [exp(-(x-midpoint)*(x-midpoint)/(latent_width*latent_width)/(2*var)) / sqrt(2*pi*var) for x in range(latent_width)] midpoint = (latent_height -1) / 2 y_probs = [exp(-(y-midpoint)*(y-midpoint)/(latent_height*latent_height)/(2*var)) / sqrt(2*pi*var) for y in range(latent_height)] weights = np.outer(y_probs, x_probs) * region.mask_weight return torch.tile(torch.tensor(weights), (self.nbatch, self.latent_space_dim, 1, 1)) def _quartic_weights(self, region: DiffusionRegion) -> torch.tensor: """Generates a quartic mask of weights for tile contributions The quartic kernel has bounded support over the diffusion region, and a smooth decay to the region limits. """ quartic_constant = 15. / 16. support = (np.array(range(region.latent_col_init, region.latent_col_end)) - region.latent_col_init) / (region.latent_col_end - region.latent_col_init - 1) * 1.99 - (1.99 / 2.) x_probs = quartic_constant * np.square(1 - np.square(support)) support = (np.array(range(region.latent_row_init, region.latent_row_end)) - region.latent_row_init) / (region.latent_row_end - region.latent_row_init - 1) * 1.99 - (1.99 / 2.) y_probs = quartic_constant * np.square(1 - np.square(support)) weights = np.outer(y_probs, x_probs) * region.mask_weight return torch.tile(torch.tensor(weights), (self.nbatch, self.latent_space_dim, 1, 1)) class StableDiffusionCanvasPipeline(DiffusionPipeline): """Stable Diffusion pipeline that mixes several diffusers in the same canvas""" def __init__( self, vae: AutoencoderKL, text_encoder: CLIPTextModel, tokenizer: CLIPTokenizer, unet: UNet2DConditionModel, scheduler: Union[DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler], safety_checker: StableDiffusionSafetyChecker, feature_extractor: CLIPFeatureExtractor, ): super().__init__() self.register_modules( vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, safety_checker=safety_checker, feature_extractor=feature_extractor, ) def decode_latents(self, latents, cpu_vae=False): """Decodes a given array of latents into pixel space""" # scale and decode the image latents with vae if cpu_vae: lat = deepcopy(latents).cpu() vae = deepcopy(self.vae).cpu() else: lat = latents vae = self.vae lat = 1 / 0.18215 * lat image = vae.decode(lat).sample image = (image / 2 + 0.5).clamp(0, 1) image = image.cpu().permute(0, 2, 3, 1).numpy() return self.numpy_to_pil(image) def get_latest_timestep_img2img(self, num_inference_steps, strength): """Finds the latest timesteps where an img2img strength does not impose latents anymore""" # get the original timestep using init_timestep offset = self.scheduler.config.get("steps_offset", 0) init_timestep = int(num_inference_steps * (1 - strength)) + offset init_timestep = min(init_timestep, num_inference_steps) t_start = min(max(num_inference_steps - init_timestep + offset, 0), num_inference_steps-1) latest_timestep = self.scheduler.timesteps[t_start] return latest_timestep @torch.no_grad() def __call__( self, canvas_height: int, canvas_width: int, regions: List[DiffusionRegion], num_inference_steps: Optional[int] = 50, seed: Optional[int] = 12345, reroll_regions: Optional[List[RerollRegion]] = None, cpu_vae: Optional[bool] = False, decode_steps: Optional[bool] = False ): if reroll_regions is None: reroll_regions = [] batch_size = 1 if decode_steps: steps_images = [] # Prepare scheduler self.scheduler.set_timesteps(num_inference_steps, device=self.device) # Split diffusion regions by their kind text2image_regions = [region for region in regions if isinstance(region, Text2ImageRegion)] image2image_regions = [region for region in regions if isinstance(region, Image2ImageRegion)] # Prepare text embeddings for region in text2image_regions: region.tokenize_prompt(self.tokenizer) region.encode_prompt(self.text_encoder, self.device) # Create original noisy latents using the timesteps latents_shape = (batch_size, self.unet.config.in_channels, canvas_height // 8, canvas_width // 8) generator = torch.Generator(self.device).manual_seed(seed) init_noise = torch.randn(latents_shape, generator=generator, device=self.device) # Reset latents in seed reroll regions, if requested for region in reroll_regions: if region.reroll_mode == RerollModes.RESET.value: region_shape = (latents_shape[0], latents_shape[1], region.latent_row_end - region.latent_row_init, region.latent_col_end - region.latent_col_init) init_noise[:, :, region.latent_row_init:region.latent_row_end, region.latent_col_init:region.latent_col_end] = torch.randn(region_shape, generator=region.get_region_generator(self.device), device=self.device) # Apply epsilon noise to regions: first diffusion regions, then reroll regions all_eps_rerolls = regions + [r for r in reroll_regions if r.reroll_mode == RerollModes.EPSILON.value] for region in all_eps_rerolls: if region.noise_eps > 0: region_noise = init_noise[:, :, region.latent_row_init:region.latent_row_end, region.latent_col_init:region.latent_col_end] eps_noise = torch.randn(region_noise.shape, generator=region.get_region_generator(self.device), device=self.device) * region.noise_eps init_noise[:, :, region.latent_row_init:region.latent_row_end, region.latent_col_init:region.latent_col_end] += eps_noise # scale the initial noise by the standard deviation required by the scheduler latents = init_noise * self.scheduler.init_noise_sigma # Get unconditional embeddings for classifier free guidance in text2image regions for region in text2image_regions: max_length = region.tokenized_prompt.input_ids.shape[-1] uncond_input = self.tokenizer( [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt" ) uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0] # For classifier free guidance, we need to do two forward passes. # Here we concatenate the unconditional and text embeddings into a single batch # to avoid doing two forward passes region.encoded_prompt = torch.cat([uncond_embeddings, region.encoded_prompt]) # Prepare image latents for region in image2image_regions: region.encode_reference_image(self.vae, device=self.device, generator=generator) # Prepare mask of weights for each region mask_builder = MaskWeightsBuilder(latent_space_dim=self.unet.config.in_channels, nbatch=batch_size) mask_weights = [mask_builder.compute_mask_weights(region).to(self.device) for region in text2image_regions] # Diffusion timesteps for i, t in tqdm(enumerate(self.scheduler.timesteps)): # Diffuse each region noise_preds_regions = [] # text2image regions for region in text2image_regions: region_latents = latents[:, :, region.latent_row_init:region.latent_row_end, region.latent_col_init:region.latent_col_end] # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([region_latents] * 2) # scale model input following scheduler rules latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) # predict the noise residual noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=region.encoded_prompt)["sample"] # perform guidance noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) noise_pred_region = noise_pred_uncond + region.guidance_scale * (noise_pred_text - noise_pred_uncond) noise_preds_regions.append(noise_pred_region) # Merge noise predictions for all tiles noise_pred = torch.zeros(latents.shape, device=self.device) contributors = torch.zeros(latents.shape, device=self.device) # Add each tile contribution to overall latents for region, noise_pred_region, mask_weights_region in zip(text2image_regions, noise_preds_regions, mask_weights): noise_pred[:, :, region.latent_row_init:region.latent_row_end, region.latent_col_init:region.latent_col_end] += noise_pred_region * mask_weights_region contributors[:, :, region.latent_row_init:region.latent_row_end, region.latent_col_init:region.latent_col_end] += mask_weights_region # Average overlapping areas with more than 1 contributor noise_pred /= contributors noise_pred = torch.nan_to_num(noise_pred) # Replace NaNs by zeros: NaN can appear if a position is not covered by any DiffusionRegion # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents).prev_sample # Image2Image regions: override latents generated by the scheduler for region in image2image_regions: influence_step = self.get_latest_timestep_img2img(num_inference_steps, region.strength) # Only override in the timesteps before the last influence step of the image (given by its strength) if t > influence_step: timestep = t.repeat(batch_size) region_init_noise = init_noise[:, :, region.latent_row_init:region.latent_row_end, region.latent_col_init:region.latent_col_end] region_latents = self.scheduler.add_noise(region.reference_latents, region_init_noise, timestep) latents[:, :, region.latent_row_init:region.latent_row_end, region.latent_col_init:region.latent_col_end] = region_latents if decode_steps: steps_images.append(self.decode_latents(latents, cpu_vae)) # scale and decode the image latents with vae image = self.decode_latents(latents, cpu_vae) output = {"sample": image} if decode_steps: output = {**output, "steps_images": steps_images} return output
20,181
49.836272
224
py
mixture-of-diffusers
mixture-of-diffusers-master/mixdiff/tiling.py
from enum import Enum import inspect from ligo.segments import segment from typing import List, Optional, Tuple, Union import torch from tqdm.auto import tqdm from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer from diffusers.models import AutoencoderKL, UNet2DConditionModel from diffusers.pipeline_utils import DiffusionPipeline from diffusers.schedulers import DDIMScheduler, PNDMScheduler from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker from diffusers.schedulers import LMSDiscreteScheduler from .extrasmixin import StableDiffusionExtrasMixin class StableDiffusionTilingPipeline(DiffusionPipeline, StableDiffusionExtrasMixin): def __init__( self, vae: AutoencoderKL, text_encoder: CLIPTextModel, tokenizer: CLIPTokenizer, unet: UNet2DConditionModel, scheduler: Union[DDIMScheduler, PNDMScheduler], safety_checker: StableDiffusionSafetyChecker, feature_extractor: CLIPFeatureExtractor, ): super().__init__() self.register_modules( vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet, scheduler=scheduler, safety_checker=safety_checker, feature_extractor=feature_extractor, ) class SeedTilesMode(Enum): """Modes in which the latents of a particular tile can be re-seeded""" FULL = "full" EXCLUSIVE = "exclusive" @torch.no_grad() def __call__( self, prompt: Union[str, List[List[str]]], num_inference_steps: Optional[int] = 50, guidance_scale: Optional[float] = 7.5, eta: Optional[float] = 0.0, seed: Optional[int] = None, tile_height: Optional[int] = 512, tile_width: Optional[int] = 512, tile_row_overlap: Optional[int] = 256, tile_col_overlap: Optional[int] = 256, guidance_scale_tiles: Optional[List[List[float]]] = None, seed_tiles: Optional[List[List[int]]] = None, seed_tiles_mode: Optional[Union[str, List[List[str]]]] = "full", seed_reroll_regions: Optional[List[Tuple[int, int, int, int, int]]] = None, cpu_vae: Optional[bool] = False, ): if not isinstance(prompt, list) or not all(isinstance(row, list) for row in prompt): raise ValueError(f"`prompt` has to be a list of lists but is {type(prompt)}") grid_rows = len(prompt) grid_cols = len(prompt[0]) if not all(len(row) == grid_cols for row in prompt): raise ValueError(f"All prompt rows must have the same number of prompt columns") if not isinstance(seed_tiles_mode, str) and (not isinstance(seed_tiles_mode, list) or not all(isinstance(row, list) for row in seed_tiles_mode)): raise ValueError(f"`seed_tiles_mode` has to be a string or list of lists but is {type(prompt)}") if isinstance(seed_tiles_mode, str): seed_tiles_mode = [[seed_tiles_mode for _ in range(len(row))] for row in prompt] if any(mode not in (modes := [mode.value for mode in self.SeedTilesMode]) for row in seed_tiles_mode for mode in row): raise ValueError(f"Seed tiles mode must be one of {modes}") if seed_reroll_regions is None: seed_reroll_regions = [] batch_size = 1 # create original noisy latents using the timesteps height = tile_height + (grid_rows - 1) * (tile_height - tile_row_overlap) width = tile_width + (grid_cols - 1) * (tile_width - tile_col_overlap) latents_shape = (batch_size, self.unet.config.in_channels, height // 8, width // 8) generator = torch.Generator("cuda").manual_seed(seed) latents = torch.randn(latents_shape, generator=generator, device=self.device) # overwrite latents for specific tiles if provided if seed_tiles is not None: for row in range(grid_rows): for col in range(grid_cols): if (seed_tile := seed_tiles[row][col]) is not None: mode = seed_tiles_mode[row][col] if mode == self.SeedTilesMode.FULL.value: row_init, row_end, col_init, col_end = _tile2latent_indices(row, col, tile_width, tile_height, tile_row_overlap, tile_col_overlap) else: row_init, row_end, col_init, col_end = _tile2latent_exclusive_indices(row, col, tile_width, tile_height, tile_row_overlap, tile_col_overlap, grid_rows, grid_cols) tile_generator = torch.Generator("cuda").manual_seed(seed_tile) tile_shape = (latents_shape[0], latents_shape[1], row_end - row_init, col_end - col_init) latents[:, :, row_init:row_end, col_init:col_end] = torch.randn(tile_shape, generator=tile_generator, device=self.device) # overwrite again for seed reroll regions for row_init, row_end, col_init, col_end, seed_reroll in seed_reroll_regions: row_init, row_end, col_init, col_end = _pixel2latent_indices(row_init, row_end, col_init, col_end) # to latent space coordinates reroll_generator = torch.Generator("cuda").manual_seed(seed_reroll) region_shape = (latents_shape[0], latents_shape[1], row_end - row_init, col_end - col_init) latents[:, :, row_init:row_end, col_init:col_end] = torch.randn(region_shape, generator=reroll_generator, device=self.device) # Prepare scheduler accepts_offset = "offset" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys()) extra_set_kwargs = {} if accepts_offset: extra_set_kwargs["offset"] = 1 self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs) # if we use LMSDiscreteScheduler, let's make sure latents are multiplied by sigmas if isinstance(self.scheduler, LMSDiscreteScheduler): latents = latents * self.scheduler.sigmas[0] # get prompts text embeddings text_input = [ [ self.tokenizer( col, padding="max_length", max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt", ) for col in row ] for row in prompt ] text_embeddings = [ [ self.text_encoder(col.input_ids.to(self.device))[0] for col in row ] for row in text_input ] # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` # corresponds to doing no classifier free guidance. do_classifier_free_guidance = guidance_scale > 1.0 # TODO: also active if any tile has guidance scale # get unconditional embeddings for classifier free guidance if do_classifier_free_guidance: for i in range(grid_rows): for j in range(grid_cols): max_length = text_input[i][j].input_ids.shape[-1] uncond_input = self.tokenizer( [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt" ) uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0] # For classifier free guidance, we need to do two forward passes. # Here we concatenate the unconditional and text embeddings into a single batch # to avoid doing two forward passes text_embeddings[i][j] = torch.cat([uncond_embeddings, text_embeddings[i][j]]) # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 # and should be between [0, 1] accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) extra_step_kwargs = {} if accepts_eta: extra_step_kwargs["eta"] = eta # Mask for tile weights strenght tile_weights = self._gaussian_weights(tile_width, tile_height, batch_size) # Diffusion timesteps for i, t in tqdm(enumerate(self.scheduler.timesteps)): # Diffuse each tile noise_preds = [] for row in range(grid_rows): noise_preds_row = [] for col in range(grid_cols): px_row_init, px_row_end, px_col_init, px_col_end = _tile2latent_indices(row, col, tile_width, tile_height, tile_row_overlap, tile_col_overlap) tile_latents = latents[:, :, px_row_init:px_row_end, px_col_init:px_col_end] # expand the latents if we are doing classifier free guidance latent_model_input = torch.cat([tile_latents] * 2) if do_classifier_free_guidance else tile_latents latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) # predict the noise residual noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embeddings[row][col])["sample"] # perform guidance if do_classifier_free_guidance: noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) guidance = guidance_scale if guidance_scale_tiles is None or guidance_scale_tiles[row][col] is None else guidance_scale_tiles[row][col] noise_pred_tile = noise_pred_uncond + guidance * (noise_pred_text - noise_pred_uncond) noise_preds_row.append(noise_pred_tile) noise_preds.append(noise_preds_row) # Stitch noise predictions for all tiles noise_pred = torch.zeros(latents.shape, device=self.device) contributors = torch.zeros(latents.shape, device=self.device) # Add each tile contribution to overall latents for row in range(grid_rows): for col in range(grid_cols): px_row_init, px_row_end, px_col_init, px_col_end = _tile2latent_indices(row, col, tile_width, tile_height, tile_row_overlap, tile_col_overlap) noise_pred[:, :, px_row_init:px_row_end, px_col_init:px_col_end] += noise_preds[row][col] * tile_weights contributors[:, :, px_row_init:px_row_end, px_col_init:px_col_end] += tile_weights # Average overlapping areas with more than 1 contributor noise_pred /= contributors # compute the previous noisy sample x_t -> x_t-1 latents = self.scheduler.step(noise_pred, t, latents).prev_sample # scale and decode the image latents with vae image = self.decode_latents(latents, cpu_vae) return {"sample": image} def _gaussian_weights(self, tile_width, tile_height, nbatches): """Generates a gaussian mask of weights for tile contributions""" from numpy import pi, exp, sqrt import numpy as np latent_width = tile_width // 8 latent_height = tile_height // 8 var = 0.01 midpoint = (latent_width - 1) / 2 # -1 because index goes from 0 to latent_width - 1 x_probs = [exp(-(x-midpoint)*(x-midpoint)/(latent_width*latent_width)/(2*var)) / sqrt(2*pi*var) for x in range(latent_width)] midpoint = latent_height / 2 y_probs = [exp(-(y-midpoint)*(y-midpoint)/(latent_height*latent_height)/(2*var)) / sqrt(2*pi*var) for y in range(latent_height)] weights = np.outer(y_probs, x_probs) return torch.tile(torch.tensor(weights, device=self.device), (nbatches, self.unet.config.in_channels, 1, 1)) def _tile2pixel_indices(tile_row, tile_col, tile_width, tile_height, tile_row_overlap, tile_col_overlap): """Given a tile row and column numbers returns the range of pixels affected by that tiles in the overall image Returns a tuple with: - Starting coordinates of rows in pixel space - Ending coordinates of rows in pixel space - Starting coordinates of columns in pixel space - Ending coordinates of columns in pixel space """ px_row_init = 0 if tile_row == 0 else tile_row * (tile_height - tile_row_overlap) px_row_end = px_row_init + tile_height px_col_init = 0 if tile_col == 0 else tile_col * (tile_width - tile_col_overlap) px_col_end = px_col_init + tile_width return px_row_init, px_row_end, px_col_init, px_col_end def _pixel2latent_indices(px_row_init, px_row_end, px_col_init, px_col_end): """Translates coordinates in pixel space to coordinates in latent space""" return px_row_init // 8, px_row_end // 8, px_col_init // 8, px_col_end // 8 def _tile2latent_indices(tile_row, tile_col, tile_width, tile_height, tile_row_overlap, tile_col_overlap): """Given a tile row and column numbers returns the range of latents affected by that tiles in the overall image Returns a tuple with: - Starting coordinates of rows in latent space - Ending coordinates of rows in latent space - Starting coordinates of columns in latent space - Ending coordinates of columns in latent space """ px_row_init, px_row_end, px_col_init, px_col_end = _tile2pixel_indices(tile_row, tile_col, tile_width, tile_height, tile_row_overlap, tile_col_overlap) return _pixel2latent_indices(px_row_init, px_row_end, px_col_init, px_col_end) def _tile2latent_exclusive_indices(tile_row, tile_col, tile_width, tile_height, tile_row_overlap, tile_col_overlap, rows, columns): """Given a tile row and column numbers returns the range of latents affected only by that tile in the overall image Returns a tuple with: - Starting coordinates of rows in latent space - Ending coordinates of rows in latent space - Starting coordinates of columns in latent space - Ending coordinates of columns in latent space """ row_init, row_end, col_init, col_end = _tile2latent_indices(tile_row, tile_col, tile_width, tile_height, tile_row_overlap, tile_col_overlap) row_segment = segment(row_init, row_end) col_segment = segment(col_init, col_end) # Iterate over the rest of tiles, clipping the region for the current tile for row in range(rows): for column in range(columns): if row != tile_row and column != tile_col: clip_row_init, clip_row_end, clip_col_init, clip_col_end = _tile2latent_indices(row, column, tile_width, tile_height, tile_row_overlap, tile_col_overlap) row_segment = row_segment - segment(clip_row_init, clip_row_end) col_segment = col_segment - segment(clip_col_init, clip_col_end) #return row_init, row_end, col_init, col_end return row_segment[0], row_segment[1], col_segment[0], col_segment[1]
15,330
52.232639
218
py
mixture-of-diffusers
mixture-of-diffusers-master/mixdiff/imgtools.py
import numpy as np import torch from PIL import Image, ImageFilter def preprocess_image(image): """Preprocess an input image Same as https://github.com/huggingface/diffusers/blob/1138d63b519e37f0ce04e027b9f4a3261d27c628/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_img2img.py#L44 """ w, h = image.size w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 image = image.resize((w, h), resample=Image.LANCZOS) image = np.array(image).astype(np.float32) / 255.0 image = image[None].transpose(0, 3, 1, 2) image = torch.from_numpy(image) return 2.0 * image - 1.0 def preprocess_mask(mask, smoothing=None): """Preprocess an inpainting mask""" mask = mask.convert("L") if smoothing is not None: smoothed = mask.filter(ImageFilter.GaussianBlur(smoothing)) mask = Image.composite(mask, smoothed, mask) # Original mask values kept as 1, out of mask get smoothed mask.save("outputs/smoothed_mask.png") # FIXME w, h = mask.size w, h = map(lambda x: x - x % 32, (w, h)) # resize to integer multiple of 32 mask = mask.resize((w // 8, h // 8), resample=Image.NEAREST) mask = np.array(mask).astype(np.float32) / 255.0 mask = np.tile(mask, (4, 1, 1)) mask = mask[None].transpose(0, 1, 2, 3) # what does this step do? mask = 1 - mask # repaint white, keep black mask = torch.from_numpy(mask) return mask
1,458
39.527778
180
py
ACME
ACME-master/test.py
import os import time import random import numpy as np import torch import torch.nn as nn import torch.nn.parallel import torch.optim import torch.utils.data import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models import torch.backends.cudnn as cudnn from data_loader import get_loader from args import get_parser from models import * from tqdm import tqdm import pdb import torch.nn.functional as F from triplet_loss import * import pickle from build_vocab import Vocabulary from torch.autograd import Variable from torch.nn.utils.rnn import pack_padded_sequence import torchvision.utils as vutils device = [0] with open(opts.vocab_path, 'rb') as f: vocab = pickle.load(f) image_model = ImageEmbedding() image_model = torch.nn.DataParallel(image_model, device_ids=device).cuda() image_model_pre = torch.load('acme/model_e045_v1.pkl') image_model.load_state_dict(image_model_pre) recipe_model = TextEmbedding() recipe_model = torch.nn.DataParallel(recipe_model, device_ids=device).cuda() recipe_model_pre = torch.load('acme/model_e045_v2.pkl') recipe_model.load_state_dict(recipe_model_pre) fc_sia = nn.Sequential( nn.Linear(opts.embDim, opts.embDim), nn.BatchNorm1d(opts.embDim), nn.Tanh(), ).cuda() fc_sia.load_state_dict(torch.load('acme/model_e045_v8.pkl')) np.random.seed(opts.seed) def main(): normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) val_transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224)]) val_loader = get_loader(opts.img_path, val_transform, vocab, opts.data_path, partition='test', batch_size=opts.batch_size, shuffle=False, num_workers=opts.workers, pin_memory=True) print('Validation loader prepared.') test(val_loader) def test(test_loader): image_model.eval() recipe_model.eval() for i, data in enumerate(tqdm(test_loader)): with torch.no_grad(): img_emd_modal = image_model(data[0][0].cuda()) recipe_emb_modal = recipe_model(data[0][1].cuda(), data[0][2].cuda(), data[0][3].cuda(), data[0][4].cuda()) img_emd_modal = norm(fc_sia(img_emd_modal)) recipe_emb_modal = norm(fc_sia(recipe_emb_modal)) if i==0: data0 = img_emd_modal.data.cpu().numpy() data1 = recipe_emb_modal.data.cpu().numpy() else: data0 = np.concatenate((data0,img_emd_modal.data.cpu().numpy()),axis=0) data1 = np.concatenate((data1,recipe_emb_modal.data.cpu().numpy()),axis=0) medR_i2t, recall_i2t = rank_i2t(opts, data0, data1) print('I2T Val medR {medR:.4f}\t' 'Recall {recall}'.format(medR=medR_i2t, recall=recall_i2t)) medR_t2i, recall_t2i = rank_t2i(opts, data0, data1) print('T2I Val medR {medR:.4f}\t' 'Recall {recall}'.format(medR=medR_t2i, recall=recall_t2i)) return recall_i2t, recall_t2i, medR_i2t, medR_t2i def rank_i2t(opts, img_embeds, rec_embeds): random.seed(opts.seed) im_vecs = img_embeds instr_vecs = rec_embeds # Ranker N = 10000 idxs = range(N) glob_rank = [] glob_recall = {1:0.0,5:0.0,10:0.0} for i in range(10): ids = random.sample(range(0,len(img_embeds)), N) im_sub = im_vecs[ids,:] instr_sub = instr_vecs[ids,:] med_rank = [] recall = {1:0.0,5:0.0,10:0.0} for ii in idxs: distance = {} for j in range(N): distance[j] = np.linalg.norm(im_sub[ii] - instr_sub[j]) distance_sorted = sorted(distance.items(), key=lambda x:x[1]) pos = np.where(np.array(distance_sorted) == distance[ii])[0][0] if (pos+1) == 1: recall[1]+=1 if (pos+1) <=5: recall[5]+=1 if (pos+1)<=10: recall[10]+=1 # store the position med_rank.append(pos+1) for i in recall.keys(): recall[i]=recall[i]/N med = np.median(med_rank) for i in recall.keys(): glob_recall[i]+=recall[i] glob_rank.append(med) for i in glob_recall.keys(): glob_recall[i] = glob_recall[i]/10 return np.average(glob_rank), glob_recall def rank_t2i(opts, img_embeds, rec_embeds): random.seed(opts.seed) im_vecs = img_embeds instr_vecs = rec_embeds # Ranker N = 10000 idxs = range(N) glob_rank = [] glob_recall = {1:0.0,5:0.0,10:0.0} for i in range(10): ids = random.sample(range(0,len(img_embeds)), N) im_sub = im_vecs[ids,:] instr_sub = instr_vecs[ids,:] med_rank = [] recall = {1:0.0,5:0.0,10:0.0} for ii in idxs: distance = {} for j in range(N): distance[j] = np.linalg.norm(instr_sub[ii] - im_sub[j]) distance_sorted = sorted(distance.items(), key=lambda x:x[1]) pos = np.where(np.array(distance_sorted) == distance[ii])[0][0] if (pos+1) == 1: recall[1]+=1 if (pos+1) <=5: recall[5]+=1 if (pos+1)<=10: recall[10]+=1 # store the position med_rank.append(pos+1) for i in recall.keys(): recall[i]=recall[i]/N med = np.median(med_rank) for i in recall.keys(): glob_recall[i]+=recall[i] glob_rank.append(med) for i in glob_recall.keys(): glob_recall[i] = glob_recall[i]/10 return np.average(glob_rank), glob_recall if __name__ == '__main__': main()
5,828
28.145
119
py
ACME
ACME-master/args.py
import argparse def get_parser(): parser = argparse.ArgumentParser(description='tri-joint parameters') # general parser.add_argument('--seed', default=1234, type=int) parser.add_argument('--device', default=[0], type=list) # data parser.add_argument('--img_path', default='../im2recipe-Pytorch/data') parser.add_argument('--data_path', default='data/food_data/') parser.add_argument('--workers', default=10, type=int) parser.add_argument('--vocab_path', type=str, default='data/new_word_dict.pkl', help='path for vocabulary wrapper') # model parser.add_argument('--batch_size', default=64, type=int) # im2recipe model parser.add_argument('--embDim', default=1024, type=int) parser.add_argument('--nRNNs', default=1, type=int) parser.add_argument('--srnnDim', default=1024, type=int) parser.add_argument('--irnnDim', default=300, type=int) parser.add_argument('--imfeatDim', default=2048, type=int) parser.add_argument('--stDim', default=1024, type=int) parser.add_argument('--ingrW2VDim', default=300, type=int) parser.add_argument('--maxSeqlen', default=20, type=int) parser.add_argument('--maxIngrs', default=20, type=int) parser.add_argument('--maxImgs', default=5, type=int) parser.add_argument('--numClasses', default=1048, type=int) #img-text parser.add_argument('--hidden_size', type=int , default=512, help='dimension of lstm hidden states') parser.add_argument('--num_layers', type=int , default=1, help='number of layers in lstm') #text-img parser.add_argument('--Z_DIM', type=int , default=100, help='noise dimension for image generation') parser.add_argument('--DF_DIM', type=int , default=64, help='D dimension') parser.add_argument('--GF_DIM', type=int , default=64, help='G dimension') parser.add_argument('--EMBEDDING_DIM', type=int , default=128, help='embedding dimension') parser.add_argument('--R_NUM', type=int , default=2, help='resudial unit number') parser.add_argument('--BRANCH_NUM', type=int , default=3, help='the number of the stages') parser.add_argument('--B_CONDITION', type=bool , default=True, help='if use condition loss') # training parser.add_argument('--lr', default=0.0001, type=float) parser.add_argument('--momentum', default=0.9, type=float) parser.add_argument('--weight_decay', default=0, type=float) parser.add_argument('--epochs', default=50, type=int) parser.add_argument('--start_epoch', default=0, type=int) parser.add_argument('--ingrW2V', default='data/vocab.bin',type=str) # dataset parser.add_argument('--maxlen', default=20, type=int) parser.add_argument('--vocab', default = 'vocab.txt', type=str) parser.add_argument('--dataset', default = '../data/recipe1M/', type=str) parser.add_argument('--sthdir', default = '../data/', type=str) return parser
2,906
47.45
119
py
ACME
ACME-master/data_loader.py
from __future__ import print_function import torch.utils.data as data from PIL import Image import os import sys import pickle import numpy as np import lmdb import torch import pdb import torchvision.transforms as transforms import nltk from build_vocab import Vocabulary from args import get_parser parser = get_parser() opts = parser.parse_args() def default_loader(path): try: im = Image.open(path).convert('RGB') return im except: return Image.new('RGB', (224, 224), 'white') class ImagerLoader(data.Dataset): def __init__(self, img_path, transform=None, vocab=None, loader=default_loader, data_path=None, partition=None): if data_path == None: raise Exception('No data path specified.') if partition is None: raise Exception('Unknown partition type %s.' % partition) else: self.partition = partition self.env = lmdb.open(os.path.join(img_path, partition + '_lmdb'), max_readers=1, readonly=True, lock=False, readahead=False, meminit=False) with open(os.path.join('data/food_data/' + partition + '_ids.pkl'), 'rb') as f: self.ids = pickle.load(f, encoding='latin1') with open(os.path.join('data/food_data/' + partition + '_split.pkl'), 'rb') as f: self.split = pickle.load(f, encoding='latin1') self.imgPath = img_path self.maxInst = 20 self.transform = transform self.loader = loader self.vocab = vocab def __getitem__(self, index): # for background with self.env.begin(write=False) as txn: serialized_sample = txn.get(self.ids[index].encode()) sample = pickle.loads(serialized_sample, encoding='latin1') imgs = sample['imgs'] food_id = self.ids[index] if self.partition != 'train': imgIdx = 0 else: imgIdx = np.random.choice(range(min(5, len(imgs)))) loader_path = [imgs[imgIdx]['id'][i] for i in range(4)] loader_path = os.path.join(*loader_path) path = os.path.join(self.imgPath, self.partition, loader_path, imgs[imgIdx]['id']) # instructions instrs = sample['intrs'] itr_ln = len(instrs) t_inst = np.zeros((self.maxInst, np.shape(instrs)[1]), dtype=np.float32) t_inst[:itr_ln][:] = instrs instrs = torch.FloatTensor(t_inst) # ingredients ingrs = sample['ingrs'].astype(int) ingrs = torch.LongTensor(ingrs) igr_ln = max(np.nonzero(sample['ingrs'])[0]) + 1 # image img = self.loader(path) img = self.transform(img) normalize = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])]) re_img = transforms.Resize(128)(img) img = normalize(img) ret = normalize(re_img) class_label = sample['classes'] - 1 vocab = self.vocab selec_ingrs = set(self.split[food_id]['ingredients']) ingr_cap = [] ingr_cap.append(vocab('<start>')) one_hot_vec = torch.zeros(4102) for i in list(selec_ingrs): one_hot_vec[vocab(str(i).lower())] = 1 ingr_cap = torch.Tensor(ingr_cap) # output ## also output the length of captions, which could be used in LSTM prediction return img, instrs, itr_ln, ingrs, igr_ln, \ ingr_cap, class_label, ret, one_hot_vec, food_id def __len__(self): return len(self.ids) def collate_fn(data): """Creates mini-batch tensors from the list of tuples (image, caption). We should build custom collate_fn rather than using default collate_fn, because merging caption (including padding) is not supported in default. Args: data: list of tuple (image, caption). - image: torch tensor of shape (3, 256, 256). - caption: torch tensor of shape (?); variable length. Returns: images: torch tensor of shape (batch_size, 3, 256, 256). targets: torch tensor of shape (batch_size, padded_length). lengths: list; valid length for each padded caption. """ # Sort a data list by caption length (descending order). data.sort(key=lambda x: len(x[5]), reverse=True) img, instrs, itr_ln, ingrs, igr_ln,\ ingr_cap, class_label, ret, one_hot_vec, food_id = zip(*data) # Merge images (from tuple of 3D tensor to 4D tensor). images = torch.stack(img, 0) instrs = torch.stack(instrs, 0) itr_ln = torch.LongTensor(list(itr_ln)) ingrs = torch.stack(ingrs, 0) igr_ln = torch.LongTensor(list(igr_ln)) class_label = torch.LongTensor(list(class_label)) ret = torch.stack(ret, 0) # Merge captions (from tuple of 1D tensor to 2D tensor). lengths = [len(cap) for cap in ingr_cap] targets = torch.zeros(len(ingr_cap), max(lengths)).long() for i, cap in enumerate(ingr_cap): end = lengths[i] targets[i, :end] = cap[:end] one_hot_vec = torch.stack(one_hot_vec, 0) return [images, instrs, itr_ln, ingrs, igr_ln, list(food_id)], \ [images, instrs, itr_ln, ingrs, igr_ln, targets, lengths, class_label, ret, one_hot_vec] def get_loader(img_path, transform, vocab, data_path, partition, batch_size, shuffle, num_workers, pin_memory): data_loader = torch.utils.data.DataLoader(ImagerLoader(img_path, transform, vocab, data_path=data_path, partition=partition), batch_size=batch_size, shuffle=shuffle, num_workers=num_workers, pin_memory=pin_memory, drop_last=True, collate_fn=collate_fn) return data_loader
6,119
35
115
py
ACME
ACME-master/triplet_loss.py
from __future__ import print_function import torch from torch import nn from torch.autograd import Variable class TripletLoss(object): """Modified from Tong Xiao's open-reid (https://github.com/Cysu/open-reid). Related Triplet Loss theory can be found in paper 'In Defense of the Triplet Loss for Person Re-Identification'.""" def __init__(self, device, margin=None): self.margin = margin self.device = device if margin is not None: self.ranking_loss = nn.MarginRankingLoss(margin=margin) else: self.ranking_loss = nn.SoftMarginLoss() def __call__(self, dist_ap, dist_an): with torch.cuda.device(self.device[0]): y = Variable(dist_an.data.new().resize_as_(dist_an.data).fill_(1)).cuda() if self.margin is not None: loss = self.ranking_loss(dist_an, dist_ap, y) else: loss = self.ranking_loss(dist_an - dist_ap, y) return loss def normalize(x, axis=-1): x = 1. * x / (torch.norm(x, 2, axis, keepdim=True).expand_as(x) + 1e-12) return x def euclidean_dist(x, y): """ Args: x: pytorch Variable, with shape [m, d] y: pytorch Variable, with shape [n, d] Returns: dist: pytorch Variable, with shape [m, n] """ m, n = x.size(0), y.size(0) xx = torch.pow(x, 2).sum(1, keepdim=True).expand(m, n) # import pdb; pdb.set_trace() yy = torch.pow(y, 2).sum(1, keepdim=True).expand(n, m).t() dist = xx + yy dist.addmm_(1, -2, x, y.t()) dist = dist.clamp(min=1e-12).sqrt() # for numerical stability return dist def hard_example_mining(dist_mat, labels, return_inds=False): assert len(dist_mat.size()) == 2 assert dist_mat.size(0) == dist_mat.size(1) N = dist_mat.size(0) # shape [N, N] is_pos = labels.expand(N, N).eq(labels.expand(N, N).t()) is_neg = labels.expand(N, N).ne(labels.expand(N, N).t()) # `dist_ap` means distance(anchor, positive) # both `dist_ap` and `relative_p_inds` with shape [N, 1] dist_ap, relative_p_inds = torch.max( dist_mat[is_pos].contiguous().view(N, -1), 1, keepdim=True) # `dist_an` means distance(anchor, negative) # both `dist_an` and `relative_n_inds` with shape [N, 1] dist_an, relative_n_inds = torch.min( dist_mat[is_neg].contiguous().view(N, -1), 1, keepdim=True) # shape [N] dist_ap = dist_ap.squeeze(1) dist_an = dist_an.squeeze(1) if return_inds: # shape [N, N] ind = (labels.new().resize_as_(labels) .copy_(torch.arange(0, N).long()) .unsqueeze( 0).expand(N, N)) # shape [N, 1] p_inds = torch.gather( ind[is_pos].contiguous().view(N, -1), 1, relative_p_inds.data) n_inds = torch.gather( ind[is_neg].contiguous().view(N, -1), 1, relative_n_inds.data) # shape [N] p_inds = p_inds.squeeze(1) n_inds = n_inds.squeeze(1) return dist_ap, dist_an, p_inds, n_inds return dist_ap, dist_an def global_loss(tri_loss, global_feat, labels, normalize_feature=False): if normalize_feature: global_feat = normalize(global_feat, axis=-1) # shape [N, N] dist_mat = euclidean_dist(global_feat, global_feat) dist_ap, dist_an = hard_example_mining( dist_mat, labels, return_inds=False) loss = tri_loss(dist_ap, dist_an) return loss, dist_ap, dist_an, dist_mat
3,225
30.940594
79
py
ACME
ACME-master/models.py
import torch import torch.nn as nn import torch.nn.parallel import torch.legacy as legacy import torch.optim import torch.utils.data import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models import torchwordemb from args import get_parser import pdb import torch.nn.functional as F from torch.autograd import Variable # ============================================================================= parser = get_parser() opts = parser.parse_args() # # ============================================================================= ############################################################## ## some codes modified from Han Zhang's Stack-GAN (https://github.com/hanzhanggit/StackGAN) ## and Salvador's im2recipe (https://github.com/torralba-lab/im2recipe-Pytorch) ############################################################## class GLU(nn.Module): def __init__(self): super(GLU, self).__init__() def forward(self, x): nc = x.size(1) assert nc % 2 == 0, 'channels dont divide 2!' nc = int(nc/2) return x[:, :nc] * torch.sigmoid(x[:, nc:]) def conv3x3(in_planes, out_planes): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=1, padding=1, bias=False) # ############## G networks ################################################ # Upsale the spatial size by a factor of 2 def upBlock(in_planes, out_planes): block = nn.Sequential( nn.Upsample(scale_factor=2, mode='nearest'), conv3x3(in_planes, out_planes * 2), nn.BatchNorm2d(out_planes * 2), GLU() ) return block # Keep the spatial size def Block3x3_relu(in_planes, out_planes): block = nn.Sequential( conv3x3(in_planes, out_planes * 2), nn.BatchNorm2d(out_planes * 2), GLU() ) return block class ResBlock(nn.Module): def __init__(self, channel_num): super(ResBlock, self).__init__() self.block = nn.Sequential( conv3x3(channel_num, channel_num * 2), nn.BatchNorm2d(channel_num * 2), GLU(), conv3x3(channel_num, channel_num), nn.BatchNorm2d(channel_num) ) def forward(self, x): residual = x out = self.block(x) out += residual return out class TableModule(nn.Module): def __init__(self): super(TableModule, self).__init__() def forward(self, x, dim): y = torch.cat(x, dim) return y def norm(input, p=2, dim=1, eps=1e-12): return input / input.norm(p,dim,keepdim=True).clamp(min=eps).expand_as(input) # Skip-thoughts LSTM class stRNN(nn.Module): def __init__(self): super(stRNN, self).__init__() self.lstm = nn.LSTM(input_size=opts.stDim, hidden_size=opts.srnnDim, bidirectional=False, batch_first=True) def forward(self, x, sq_lengths): # here we use a previous LSTM to get the representation of each instruction # sort sequence according to the length sorted_len, sorted_idx = sq_lengths.sort(0, descending=True) index_sorted_idx = sorted_idx\ .view(-1,1,1).expand_as(x) sorted_inputs = x.gather(0, index_sorted_idx.long()) # pack sequence packed_seq = torch.nn.utils.rnn.pack_padded_sequence( sorted_inputs, sorted_len.cpu().data.numpy(), batch_first=True) # pass it to the lstm out, hidden = self.lstm(packed_seq) # unsort the output _, original_idx = sorted_idx.sort(0, descending=False) unpacked, _ = torch.nn.utils.rnn.pad_packed_sequence(out, batch_first=True) unsorted_idx = original_idx.view(-1,1,1).expand_as(unpacked) # we get the last index of each sequence in the batch idx = (sq_lengths-1).view(-1,1).expand(unpacked.size(0), unpacked.size(2)).unsqueeze(1) # we sort and get the last element of each sequence output = unpacked.gather(0, unsorted_idx.long()).gather(1,idx.long()) output = output.view(output.size(0),output.size(1)*output.size(2)) return output class ingRNN(nn.Module): def __init__(self): super(ingRNN, self).__init__() self.irnn = nn.LSTM(input_size=opts.ingrW2VDim, hidden_size=opts.irnnDim, bidirectional=True, batch_first=True) _, vec = torchwordemb.load_word2vec_bin(opts.ingrW2V) self.embs = nn.Embedding(vec.size(0), opts.ingrW2VDim, padding_idx=0) # not sure about the padding idx self.embs.weight.data.copy_(vec) # self.embs = nn.Embedding(3122, opts.ingrW2VDim) # self.irnn = nn.LSTM(input_size=opts.ingrW2VDim, hidden_size=opts.irnnDim, bidirectional=True, batch_first=True) def forward(self, x, sq_lengths): # we get the w2v for each element of the ingredient sequence x = self.embs(x) # torch.Size([64, 20, 300]) # sort sequence according to the length sorted_len, sorted_idx = sq_lengths.sort(0, descending=True) index_sorted_idx = sorted_idx\ .view(-1,1,1).expand_as(x) sorted_inputs = x.gather(0, index_sorted_idx.long()) # pack sequence packed_seq = torch.nn.utils.rnn.pack_padded_sequence( sorted_inputs, sorted_len.cpu().data.numpy(), batch_first=True) # pass it to the rnn out, hidden = self.irnn(packed_seq) # unsort the output _, original_idx = sorted_idx.sort(0, descending=False) # LSTM # bi-directional unsorted_idx = original_idx.view(1,-1,1).expand_as(hidden[0]) # 2 directions x batch_size x num features, we transpose 1st and 2nd dimension output = hidden[0].gather(1,unsorted_idx).transpose(0,1).contiguous() # torch.Size([64, 2, 300]) output = output.view(output.size(0),output.size(1)*output.size(2)) return output class TextEmbedding(nn.Module): def __init__(self): super(TextEmbedding, self).__init__() self.recipe_embedding = nn.Sequential( nn.Linear(opts.irnnDim*2 + opts.srnnDim, opts.embDim), nn.Tanh(), ) self.semantic_branch = nn.Linear(opts.embDim, opts.numClasses) self.fc_recipe = nn.Sequential( nn.Linear(opts.embDim, opts.embDim), nn.BatchNorm1d(opts.embDim), nn.Tanh(), ) self.stRNN_ = stRNN() self.ingRNN_ = ingRNN() self.table = TableModule() def forward(self, y1, y2, z1, z2, ingrs_emb=None): if torch.is_tensor(ingrs_emb): recipe_emb = self.table([self.stRNN_(y1,y2), ingrs_emb],1) # joining on the last dim else: ingrs_emb = self.ingRNN_(z1,z2) recipe_emb = self.table([self.stRNN_(y1,y2), ingrs_emb],1) # joining on the last dim recipe_emb_domain = self.recipe_embedding(recipe_emb) output = recipe_emb_domain return output class ImageEmbedding(nn.Module): def __init__(self): super(ImageEmbedding, self).__init__() resnet = models.resnet50(pretrained=True) modules = list(resnet.children())[:-1] # we do not use the last fc layer. self.visionMLP = nn.Sequential(*modules) self.visual_embedding = nn.Sequential( nn.Linear(opts.imfeatDim, opts.embDim), nn.Tanh(), ) self.semantic_branch = nn.Linear(opts.embDim, opts.numClasses) self.fc_visual = nn.Sequential( nn.Linear(opts.embDim, opts.embDim), nn.BatchNorm1d(opts.embDim), nn.Tanh(), ) def forward(self, x): visual_emb = self.visionMLP(x) visual_emb = visual_emb.view(visual_emb.size(0), -1) # batch_size * 2048 visual_emb_domain = self.visual_embedding(visual_emb) output = visual_emb_domain return output class MultiLabelNet(nn.Module): def __init__(self): super(MultiLabelNet, self).__init__() self.model = nn.Sequential( nn.Linear(opts.embDim, 512), nn.LeakyReLU(0.2, inplace=True), nn.Linear(512, 256), nn.LeakyReLU(0.2, inplace=True), nn.Linear(256, 4102) ) self.ingrs_linear = nn.Sequential( nn.Linear(opts.embDim, 512), nn.LeakyReLU(0.2, inplace=True), nn.Linear(512, opts.numClasses) ) def forward(self, features): output = self.model(features) return [nn.Sigmoid()(output), self.ingrs_linear(features)] class DecoderRNN(nn.Module): def __init__(self, embed_size, hidden_size, vocab_size, num_layers, max_seq_length=100): """Set the hyper-parameters and build the layers.""" super(DecoderRNN, self).__init__() self.embed = nn.Embedding(vocab_size, embed_size) # self.lstm = nn.LSTM(embed_size, hidden_size, num_layers, batch_first=True) self.lstm = nn.LSTM(input_size=embed_size, hidden_size=opts.irnnDim, bidirectional=False, batch_first=True) self.linear = nn.Linear(opts.irnnDim, vocab_size) self.ingrs_linear = nn.Linear(opts.irnnDim, opts.irnnDim*2) self.max_seg_length = max_seq_length def forward(self, features, captions, lengths): """Decode image feature vectors and generates captions.""" embeddings = self.embed(captions) embeddings = torch.cat((features.unsqueeze(1), embeddings), 1) packed = torch.nn.utils.rnn.pack_padded_sequence(embeddings, lengths, batch_first=True) outs, hidden = self.lstm(packed) _, original_idx = lengths.sort(0, descending=False) unsorted_idx = original_idx.view(1,-1,1).expand_as(hidden[0]) output = hidden[0].gather(1,unsorted_idx).transpose(0,1).contiguous() # output = output.view(output.size(0),output.size(1)*output.size(2)) output_fea = self.ingrs_linear(output.view(output.size(0),output.size(1)*output.size(2))) outputs = self.linear(outs[0]) return [outputs, output_fea] def sample(self, features, states=None): """Generate captions for given image features using greedy search.""" sampled_ids = [] inputs = features.unsqueeze(1) for i in range(self.max_seg_length): outs, states = self.lstm(inputs, states) # outs: (batch_size, 1, hidden_size) outputs = self.linear(outs.squeeze(1)) # outputs: (batch_size, vocab_size) _, predicted = outputs.max(1) # predicted: (batch_size) sampled_ids.append(predicted) inputs = self.embed(predicted) # inputs: (batch_size, embed_size) inputs = inputs.unsqueeze(1) # inputs: (batch_size, 1, embed_size) sampled_ids = torch.stack(sampled_ids, 1) # sampled_ids: (batch_size, max_seq_length) return sampled_ids class text_emb_discriminator(nn.Module): def __init__(self): super(text_emb_discriminator, self).__init__() self.model = nn.Sequential( nn.Linear(opts.irnnDim * 2, 512), nn.LeakyReLU(0.2, inplace=True), nn.Linear(512, 256), nn.LeakyReLU(0.2, inplace=True), nn.Linear(256, opts.numClasses * 2) ) def forward(self, fea): output = self.model(fea) return output class cross_modal_discriminator(nn.Module): def __init__(self): super(cross_modal_discriminator, self).__init__() self.model = nn.Sequential( nn.Linear(opts.embDim, 512), nn.LeakyReLU(0.2, inplace=True), nn.Linear(512, 256), nn.LeakyReLU(0.2, inplace=True), nn.Linear(256, 1) ) def forward(self, fea): output = self.model(fea) return output class CA_NET(nn.Module): # some code is modified from vae examples # (https://github.com/pytorch/examples/blob/master/vae/main.py) def __init__(self): super(CA_NET, self).__init__() self.t_dim = opts.embDim self.ef_dim = opts.EMBEDDING_DIM self.fc = nn.Linear(self.t_dim, self.ef_dim * 4, bias=True) self.relu = GLU() def encode(self, text_embedding): x = self.relu(self.fc(text_embedding)) mu = x[:, :self.ef_dim] logvar = x[:, self.ef_dim:] return mu, logvar def reparametrize(self, mu, logvar): with torch.cuda.device(logvar.get_device()): std = logvar.mul(0.5).exp_() eps = torch.FloatTensor(std.size()).normal_().cuda() eps = Variable(eps) return eps.mul(std).add_(mu) def forward(self, text_embedding): mu, logvar = self.encode(text_embedding) c_code = self.reparametrize(mu, logvar) return c_code, mu, logvar class INIT_STAGE_G(nn.Module): def __init__(self, ngf): super(INIT_STAGE_G, self).__init__() self.gf_dim = ngf if opts.B_CONDITION: self.in_dim = opts.Z_DIM + opts.EMBEDDING_DIM else: self.in_dim = opts.Z_DIM self.define_module() def define_module(self): in_dim = self.in_dim ngf = self.gf_dim self.fc = nn.Sequential( nn.Linear(in_dim, ngf * 4 * 4 * 2, bias=False), nn.BatchNorm1d(ngf * 4 * 4 * 2), GLU()) self.upsample1 = upBlock(ngf, ngf // 2) self.upsample2 = upBlock(ngf // 2, ngf // 4) self.upsample3 = upBlock(ngf // 4, ngf // 8) self.upsample4 = upBlock(ngf // 8, ngf // 16) def forward(self, z_code, c_code=None): if opts.B_CONDITION and c_code is not None: in_code = torch.cat((c_code, z_code), 1) else: in_code = z_code # state size 16ngf x 4 x 4 out_code = self.fc(in_code) out_code = out_code.view(-1, self.gf_dim, 4, 4) # state size 8ngf x 8 x 8 out_code = self.upsample1(out_code) # state size 4ngf x 16 x 16 out_code = self.upsample2(out_code) # state size 2ngf x 32 x 32 out_code = self.upsample3(out_code) # state size ngf x 64 x 64 out_code = self.upsample4(out_code) return out_code class NEXT_STAGE_G(nn.Module): def __init__(self, ngf, num_residual=opts.R_NUM): super(NEXT_STAGE_G, self).__init__() self.gf_dim = ngf if opts.B_CONDITION: self.ef_dim = opts.EMBEDDING_DIM else: self.ef_dim = opts.Z_DIM self.num_residual = num_residual self.define_module() def _make_layer(self, block, channel_num): layers = [] for i in range(self.num_residual): layers.append(block(channel_num)) return nn.Sequential(*layers) def define_module(self): ngf = self.gf_dim efg = self.ef_dim self.jointConv = Block3x3_relu(ngf + efg, ngf) self.residual = self._make_layer(ResBlock, ngf) self.upsample = upBlock(ngf, ngf // 2) def forward(self, h_code, c_code): s_size = h_code.size(2) c_code = c_code.view(-1, self.ef_dim, 1, 1) c_code = c_code.repeat(1, 1, s_size, s_size) # state size (ngf+egf) x in_size x in_size h_c_code = torch.cat((c_code, h_code), 1) # state size ngf x in_size x in_size out_code = self.jointConv(h_c_code) out_code = self.residual(out_code) # state size ngf/2 x 2in_size x 2in_size out_code = self.upsample(out_code) return out_code class GET_IMAGE_G(nn.Module): def __init__(self, ngf): super(GET_IMAGE_G, self).__init__() self.gf_dim = ngf self.img = nn.Sequential( conv3x3(ngf, 3), nn.Tanh() ) def forward(self, h_code): out_img = self.img(h_code) return out_img class G_NET(nn.Module): def __init__(self): super(G_NET, self).__init__() self.gf_dim = opts.GF_DIM self.define_module() def define_module(self): if opts.B_CONDITION: self.ca_net = CA_NET() if opts.BRANCH_NUM > 0: self.h_net1 = INIT_STAGE_G(self.gf_dim * 16) self.img_net1 = GET_IMAGE_G(self.gf_dim) if opts.BRANCH_NUM > 1: self.h_net2 = NEXT_STAGE_G(self.gf_dim) self.img_net2 = GET_IMAGE_G(self.gf_dim // 2) if opts.BRANCH_NUM > 2: self.h_net3 = NEXT_STAGE_G(self.gf_dim // 2) self.img_net3 = GET_IMAGE_G(self.gf_dim // 4) if opts.BRANCH_NUM > 3: # Recommended structure (mainly limited by GPU memory), and not test yet self.h_net4 = NEXT_STAGE_G(self.gf_dim // 4, num_residual=1) self.img_net4 = GET_IMAGE_G(self.gf_dim // 8) if opts.BRANCH_NUM > 4: self.h_net4 = NEXT_STAGE_G(self.gf_dim // 8, num_residual=1) self.img_net4 = GET_IMAGE_G(self.gf_dim // 16) def forward(self, z_code, text_embedding=None): if opts.B_CONDITION and text_embedding is not None: c_code, mu, logvar = self.ca_net(text_embedding) else: c_code, mu, logvar = z_code, None, None fake_imgs = [] h_code1 = self.h_net1(z_code, c_code) h_code2 = self.h_net2(h_code1, c_code) fake_img2 = self.img_net2(h_code2) fake_imgs.append(fake_img2) return fake_imgs, mu, logvar # ############## D networks ################################################ def Block3x3_leakRelu(in_planes, out_planes): block = nn.Sequential( conv3x3(in_planes, out_planes), nn.BatchNorm2d(out_planes), nn.LeakyReLU(0.2, inplace=True) ) return block # Downsale the spatial size by a factor of 2 def downBlock(in_planes, out_planes): block = nn.Sequential( nn.Conv2d(in_planes, out_planes, 4, 2, 1, bias=False), nn.BatchNorm2d(out_planes), nn.LeakyReLU(0.2, inplace=True) ) return block # Downsale the spatial size by a factor of 16 def encode_image_by_16times(ndf): encode_img = nn.Sequential( # --> state size. ndf x in_size/2 x in_size/2 nn.Conv2d(3, ndf, 4, 2, 1, bias=False), nn.LeakyReLU(0.2, inplace=True), # --> state size 2ndf x x in_size/4 x in_size/4 nn.Conv2d(ndf, ndf * 2, 4, 2, 1, bias=False), nn.BatchNorm2d(ndf * 2), nn.LeakyReLU(0.2, inplace=True), # --> state size 4ndf x in_size/8 x in_size/8 nn.Conv2d(ndf * 2, ndf * 4, 4, 2, 1, bias=False), nn.BatchNorm2d(ndf * 4), nn.LeakyReLU(0.2, inplace=True), # --> state size 8ndf x in_size/16 x in_size/16 nn.Conv2d(ndf * 4, ndf * 8, 4, 2, 1, bias=False), nn.BatchNorm2d(ndf * 8), nn.LeakyReLU(0.2, inplace=True) ) return encode_img # For 128 x 128 images class D_NET128(nn.Module): def __init__(self): super(D_NET128, self).__init__() self.df_dim = opts.DF_DIM self.ef_dim = opts.EMBEDDING_DIM self.define_module() def define_module(self): ndf = self.df_dim efg = self.ef_dim self.img_code_s16 = encode_image_by_16times(ndf) self.img_code_s32 = downBlock(ndf * 8, ndf * 16) self.img_code_s32_1 = Block3x3_leakRelu(ndf * 16, ndf * 8) self.logits = nn.Sequential( nn.Conv2d(ndf * 8, 1, kernel_size=4, stride=4), nn.Sigmoid()) self.logits_class = nn.Sequential( nn.Conv2d(ndf * 8, 1, kernel_size=4, stride=4), nn.Sigmoid()) if opts.B_CONDITION: self.jointConv = Block3x3_leakRelu(ndf * 8 + efg, ndf * 8) self.uncond_branch = nn.Sequential( nn.Conv2d(ndf * 8, ndf * 4, kernel_size=4, stride=4), # nn.BatchNorm2d(ndf * 4), nn.LeakyReLU(0.2, inplace=True)) self.semantic_branch = nn.Linear(ndf * 4, opts.numClasses) def forward(self, x_var, c_code=None): x_code = self.img_code_s16(x_var) x_code = self.img_code_s32(x_code) x_code = self.img_code_s32_1(x_code) if opts.B_CONDITION and c_code is not None: c_code = c_code.view(-1, self.ef_dim, 1, 1) c_code = c_code.repeat(1, 1, 4, 4) # state size (ngf+egf) x 4 x 4 h_c_code = torch.cat((c_code, x_code), 1) # state size ngf x in_size x in_size h_c_code = self.jointConv(h_c_code) else: h_c_code = x_code output = self.logits(h_c_code) if opts.B_CONDITION: out_uncond = self.uncond_branch(x_code) out_uncond = out_uncond.view(out_uncond.size(0), -1) out_uncond = self.semantic_branch(out_uncond) return [output.view(-1), out_uncond] else: return [output.view(-1)]
21,061
34.698305
121
py
ACME
ACME-master/train.py
import os import time import random import numpy as np import torch import torch.nn as nn import torch.nn.parallel import torch.optim import torch.utils.data import torchvision.transforms as transforms import torchvision.datasets as datasets import torchvision.models as models import torch.backends.cudnn as cudnn from data_loader import get_loader from args import get_parser from models import * from torch.optim import lr_scheduler from tqdm import tqdm import pdb import torch.nn.functional as F from triplet_loss import * import pickle from build_vocab import Vocabulary from torch.autograd import Variable from torch.nn.utils.rnn import pack_padded_sequence import torchvision.utils as vutils # ============================================================================= parser = get_parser() opts = parser.parse_args() device = [0] with open(opts.vocab_path, 'rb') as f: vocab = pickle.load(f) # ============================================================================= ##load models image_model = torch.nn.DataParallel(ImageEmbedding().cuda(), device_ids=device) recipe_model = torch.nn.DataParallel(TextEmbedding().cuda(), device_ids=device) netG = torch.nn.DataParallel(G_NET().cuda(), device_ids=device) multi_label_net = torch.nn.DataParallel(MultiLabelNet().cuda(), device_ids=device) cm_discriminator = torch.nn.DataParallel(cross_modal_discriminator().cuda(), device_ids=device) text_discriminator = torch.nn.DataParallel(text_emb_discriminator().cuda(), device_ids=device) netsD = torch.nn.DataParallel(D_NET128().cuda(), device_ids=device) ## load loss functions triplet_loss = TripletLoss(device, margin=0.3) img2text_criterion = nn.MultiLabelMarginLoss().cuda() weights_class = torch.Tensor(opts.numClasses).fill_(1) weights_class[0] = 0 class_criterion = nn.CrossEntropyLoss(weight=weights_class).cuda() GAN_criterion = nn.BCELoss().cuda() nz = opts.Z_DIM noise = Variable(torch.FloatTensor(opts.batch_size, nz)).cuda() fixed_noise = Variable(torch.FloatTensor(opts.batch_size, nz).normal_(0, 1)).cuda() real_labels = Variable(torch.FloatTensor(opts.batch_size).fill_(1)).cuda() fake_labels = Variable(torch.FloatTensor(opts.batch_size).fill_(0)).cuda() fc_sia = nn.Sequential( nn.Linear(opts.embDim, opts.embDim), nn.BatchNorm1d(opts.embDim), nn.Tanh(), ).cuda() model_list = [image_model, recipe_model, netG, multi_label_net, cm_discriminator, text_discriminator, netsD, fc_sia] optimizer = torch.optim.Adam([ {'params': image_model.parameters()}, {'params': recipe_model.parameters()}, {'params': netG.parameters()}, {'params': multi_label_net.parameters()} ], lr=opts.lr, betas=(0.5, 0.999)) optimizers_imgD = torch.optim.Adam(netsD.parameters(), lr=opts.lr, betas=(0.5, 0.999)) optimizer_cmD = torch.optim.Adam(cm_discriminator.parameters(), lr=opts.lr, betas=(0.5, 0.999)) label = list(range(0, opts.batch_size)) label.extend(label) label = np.array(label) label = torch.tensor(label).cuda().long() method = 'acme' save_folder = method os.makedirs(save_folder, exist_ok=True) epoch_trace_f_dir = os.path.join(save_folder, "trace_" + method + ".csv") with open(epoch_trace_f_dir, "w") as f: f.write("epoch,lr,I2R,R@1,R@5,R@10,R2I,R@1,R@5,R@10\n") def main(): # data preparation, loaders normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_transform = transforms.Compose([ transforms.Resize(256), transforms.RandomCrop(224), transforms.RandomHorizontalFlip()]) val_transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224)]) cudnn.benchmark = True # preparing the training laoder train_loader = get_loader(opts.img_path, train_transform, vocab, opts.data_path, partition='train', batch_size=opts.batch_size, shuffle=True, num_workers=opts.workers, pin_memory=True) print('Training loader prepared.') # preparing validation loader val_loader = get_loader(opts.img_path, val_transform, vocab, opts.data_path, partition='test', batch_size=opts.batch_size, shuffle=False, num_workers=opts.workers, pin_memory=True) print('Validation loader prepared.') best_val_i2t = {1:0.0,5:0.0,10:0.0} best_val_t2i = {1:0.0,5:0.0,10:0.0} best_epoch_i2t = 0 best_epoch_t2i = 0 for epoch in range(0, opts.epochs): train(train_loader, epoch, val_loader) recall_i2t, recall_t2i, medR_i2t, medR_t2i = validate(val_loader) with open(epoch_trace_f_dir, "a") as f: lr = optimizer.param_groups[1]['lr'] f.write("{},{},{},{},{},{},{},{},{},{}\n".format\ (epoch,lr,medR_i2t,recall_i2t[1],recall_i2t[5],recall_i2t[10],\ medR_t2i,recall_t2i[1],recall_t2i[5],recall_t2i[10])) for keys in best_val_i2t: if recall_i2t[keys] > best_val_i2t[keys]: best_val_i2t = recall_i2t best_epoch = epoch+1 model_num = 1 for model_n in model_list: filename = save_folder + '/model_e%03d_v%d.pkl' % (epoch+1, model_num) torch.save(model_n.state_dict(), filename) model_num += 1 break print("best: ", best_epoch, best_val_i2t) print('params lr: %f' % optimizer.param_groups[1]['lr']) if epoch == 30: optimizer.param_groups[0]['lr'] = 0.00001 optimizer.param_groups[1]['lr'] = 0.00001 optimizer.param_groups[2]['lr'] = 0.00001 optimizer.param_groups[3]['lr'] = 0.00001 optimizers_imgD.param_groups[0]['lr'] = 0.00001 optimizer_cmD.param_groups[0]['lr'] = 0.00001 def train_Dnet(idx, real_imgs, fake_imgs, mu, label_class): netD = netsD real_imgs = real_imgs[idx] fake_imgs = fake_imgs[idx] real_logits = netD(real_imgs, mu.detach()) fake_logits = netD(fake_imgs.detach(), mu.detach()) lossD_real = GAN_criterion(real_logits[0], real_labels) lossD_fake = GAN_criterion(fake_logits[0], fake_labels) lossD = lossD_real + lossD_fake return lossD def KL_loss(mu, logvar): KLD_element = mu.pow(2).add_(logvar.exp()).mul_(-1).add_(1).add_(logvar) KLD = torch.mean(KLD_element).mul_(-0.5) return KLD def train_Gnet(idx, real_imgs, fake_imgs, mu, logvar, label_class): netD = netsD real_imgs = real_imgs[idx] fake_imgs = fake_imgs[idx] real_logits = netD(real_imgs, mu) fake_logits = netD(fake_imgs, mu) lossG_fake = GAN_criterion(fake_logits[0], real_labels) lossG_real_cond = class_criterion(real_logits[1], label_class) lossG_fake_cond = class_criterion(fake_logits[1], label_class) lossG_cond = lossG_real_cond + lossG_fake_cond lossG = lossG_fake + lossG_cond kl_loss = KL_loss(mu, logvar) * 2 lossG = kl_loss + lossG return lossG def compute_gradient_penalty(D, real_samples, fake_samples): """Calculates the gradient penalty loss for WGAN GP""" # Random weight term for interpolation between real and fake samples alpha = torch.cuda.FloatTensor(np.random.random((real_samples.size(0), 1))) # Get random interpolation between real and fake samples interpolates = (alpha * real_samples + ((1 - alpha) * fake_samples)).requires_grad_(True) d_interpolates = D(interpolates) fake = torch.autograd.Variable(torch.cuda.FloatTensor(real_samples.shape[0], 1).fill_(1.0), requires_grad=False) # Get gradient w.r.t. interpolates gradients = torch.autograd.grad( outputs=d_interpolates, # fack samples inputs=interpolates, # real samples grad_outputs=fake, create_graph=True, retain_graph=True, only_inputs=True, )[0] gradients = gradients.view(gradients.size(0), -1) gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean() return gradient_penalty def train(train_loader, epoch, val_loader): tri_losses = AverageMeter() img_losses = AverageMeter() text_losses = AverageMeter() cmG_losses = AverageMeter() image_model.train() recipe_model.train() for i, data in enumerate(tqdm(train_loader)): img_emd_modal = image_model(data[0][0].cuda()) recipe_emb_modal = recipe_model(data[0][1].cuda(), data[0][2].cuda(), data[0][3].cuda(), data[0][4].cuda()) ################################################################ # modal-level fusion ################################################################ real_validity = cm_discriminator(img_emd_modal.detach()) fake_validity = cm_discriminator(recipe_emb_modal.detach()) gradient_penalty = compute_gradient_penalty(cm_discriminator, img_emd_modal.detach(), recipe_emb_modal.detach()) loss_cmD = -torch.mean(real_validity) + torch.mean(fake_validity) + 10 * gradient_penalty optimizer_cmD.zero_grad() loss_cmD.backward() optimizer_cmD.step() g_fake_validity = cm_discriminator(recipe_emb_modal) loss_cmG = -torch.mean(g_fake_validity) ################################################################ # cross-modal retrieval ################################################################ img_id_fea = norm(fc_sia(img_emd_modal)) rec_id_fea = norm(fc_sia(recipe_emb_modal)) tri_loss = global_loss(triplet_loss, torch.cat((img_id_fea, rec_id_fea)), label)[0] ################################################################ # translation consistency label_class = data[1][7].cuda() real_imgs = [] real_imgs.append(data[1][8].cuda()) ingr_cap = data[1][5].cuda() lengths = torch.tensor(data[1][6]).cuda() targets = pack_padded_sequence(ingr_cap, lengths, batch_first=True)[0] one_hot_cap = data[1][9].cuda().long() ################################################################ # img2text ################################################################ recipe_out = multi_label_net(img_id_fea) loss_i2t = img2text_criterion(recipe_out[0], one_hot_cap) loss_t_class = class_criterion(recipe_out[1], label_class) loss_text = loss_i2t + loss_t_class ############################################################### # text2img ############################################################### noise.data.normal_(0, 1) fake_imgs, mu, logvar = netG(noise, rec_id_fea) lossD_total = 0 lossD = train_Dnet(0, real_imgs, fake_imgs, mu, label_class) optimizers_imgD.zero_grad() lossD.backward() optimizers_imgD.step() lossG = train_Gnet(0, real_imgs, fake_imgs, mu, logvar, label_class) loss_img = lossG if loss_text.item() < loss_img.item(): loss_img = (loss_text.item()/loss_img.item()) * loss_img else: loss_text = (loss_img.item()/loss_text.item()) * loss_text loss_g = loss_img + loss_text ############################################################### # back-propogate ############################################################### loss = tri_loss + 0.005 * loss_cmG + 0.002 * loss_g tri_losses.update(tri_loss.item(), data[0][0].size(0)) img_losses.update(loss_img.item(), data[0][0].size(0)) text_losses.update(loss_text.item(), data[0][0].size(0)) cmG_losses.update(loss_cmG.item(), data[0][0].size(0)) optimizer.zero_grad() loss.backward() optimizer.step() print(epoch) print('Epoch: {0} ' 'tri loss {tri_loss.val:.4f} ({tri_loss.avg:.4f}), ' 'cm loss {loss_cmG.val:.4f} ({loss_cmG.avg:.4f}), ' 'img loss {img_losses.val:.4f} ({img_losses.avg:.4f}), ' 'text loss {loss_text.val:.4f} ({loss_text.avg:.4f})' .format( epoch, tri_loss=tri_losses, loss_cmG=cmG_losses, img_losses=img_losses, loss_text=text_losses)) def validate(val_loader): # switch to evaluate mode image_model.eval() recipe_model.eval() end = time.time() for i, data in enumerate(tqdm(val_loader)): with torch.no_grad(): img_emd_modal = image_model(data[0][0].cuda()) recipe_emb_modal = recipe_model(data[0][1].cuda(), data[0][2].cuda(), data[0][3].cuda(), data[0][4].cuda()) img_emd_modal = norm(fc_sia(img_emd_modal)) recipe_emb_modal = norm(fc_sia(recipe_emb_modal)) if i==0: data0 = img_emd_modal.data.cpu().numpy() data1 = recipe_emb_modal.data.cpu().numpy() else: data0 = np.concatenate((data0,img_emd_modal.data.cpu().numpy()),axis=0) data1 = np.concatenate((data1,recipe_emb_modal.data.cpu().numpy()),axis=0) medR_i2t, recall_i2t = rank_i2t(opts, data0, data1) print('I2T Val medR {medR:.4f}\t' 'Recall {recall}'.format(medR=medR_i2t, recall=recall_i2t)) medR_t2i, recall_t2i = rank_t2i(opts, data0, data1) print('T2I Val medR {medR:.4f}\t' 'Recall {recall}'.format(medR=medR_t2i, recall=recall_t2i)) return recall_i2t, recall_t2i, medR_i2t, medR_t2i def rank_i2t(opts, img_embeds, rec_embeds): random.seed(opts.seed) im_vecs = img_embeds instr_vecs = rec_embeds # Ranker N = 1000 idxs = range(N) glob_rank = [] glob_recall = {1:0.0,5:0.0,10:0.0} for i in range(10): ids = random.sample(range(0,len(img_embeds)), N) im_sub = im_vecs[ids,:] instr_sub = instr_vecs[ids,:] med_rank = [] recall = {1:0.0,5:0.0,10:0.0} for ii in idxs: distance = {} for j in range(N): distance[j] = np.linalg.norm(im_sub[ii] - instr_sub[j]) distance_sorted = sorted(distance.items(), key=lambda x:x[1]) pos = np.where(np.array(distance_sorted) == distance[ii])[0][0] if (pos+1) == 1: recall[1]+=1 if (pos+1) <=5: recall[5]+=1 if (pos+1)<=10: recall[10]+=1 # store the position med_rank.append(pos+1) for i in recall.keys(): recall[i]=recall[i]/N med = np.median(med_rank) for i in recall.keys(): glob_recall[i]+=recall[i] glob_rank.append(med) for i in glob_recall.keys(): glob_recall[i] = glob_recall[i]/10 return np.average(glob_rank), glob_recall def rank_t2i(opts, img_embeds, rec_embeds): random.seed(opts.seed) im_vecs = img_embeds instr_vecs = rec_embeds # Ranker N = 1000 idxs = range(N) glob_rank = [] glob_recall = {1:0.0,5:0.0,10:0.0} for i in range(10): ids = random.sample(range(0,len(img_embeds)), N) im_sub = im_vecs[ids,:] instr_sub = instr_vecs[ids,:] med_rank = [] recall = {1:0.0,5:0.0,10:0.0} for ii in idxs: distance = {} for j in range(N): distance[j] = np.linalg.norm(instr_sub[ii] - im_sub[j]) distance_sorted = sorted(distance.items(), key=lambda x:x[1]) pos = np.where(np.array(distance_sorted) == distance[ii])[0][0] if (pos+1) == 1: recall[1]+=1 if (pos+1) <=5: recall[5]+=1 if (pos+1)<=10: recall[10]+=1 # store the position med_rank.append(pos+1) for i in recall.keys(): recall[i]=recall[i]/N med = np.median(med_rank) for i in recall.keys(): glob_recall[i]+=recall[i] glob_rank.append(med) for i in glob_recall.keys(): glob_recall[i] = glob_recall[i]/10 return np.average(glob_rank), glob_recall class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n self.avg = self.sum / self.count if __name__ == '__main__': main()
16,687
34.430998
120
py
ACME
ACME-master/build_vocab.py
import nltk import pickle import argparse from collections import Counter class Vocabulary(object): """Simple vocabulary wrapper.""" def __init__(self): self.word2idx = {} self.idx2word = {} self.idx = 0 def add_word(self, word): if not word in self.word2idx: self.word2idx[word] = self.idx self.idx2word[self.idx] = word self.idx += 1 def __call__(self, word): if not word in self.word2idx: return self.word2idx['<unk>'] return self.word2idx[word] def __len__(self): return len(self.word2idx) def build_vocab(path, threshold): """Build a simple vocabulary wrapper.""" dataset = ['train', 'val', 'test'] # dataset = ['train'] counter = Counter() for i in dataset: data_path = path + i + '_split.pkl' with open(data_path, "rb") as f: data = pickle.load(f) ids = data.keys() for j, id in enumerate(ids): ingredient = data[id]['ingredients'] for k in range(len(ingredient)): # tokens = nltk.tokenize.word_tokenize(ingredient[k].lower()) tokens = [ingredient[k].lower()] counter.update(tokens) if (j+1) % 1000 == 0: print("[{}/{}] Tokenized the ingredients.".format(j+1, len(ids))) # If the word frequency is less than 'threshold', then the word is discarded. words = [word for word, cnt in counter.items() if cnt >= threshold] # Create a vocab wrapper and add some special tokens. vocab = Vocabulary() vocab.add_word('<pad>') vocab.add_word('<start>') vocab.add_word('<end>') vocab.add_word('<unk>') vocab.add_word(',') for i, word in enumerate(words): vocab.add_word(word) import pdb; pdb.set_trace() return vocab def main(args): vocab = build_vocab(path=args.caption_path, threshold=args.threshold) vocab_path = args.vocab_path with open(vocab_path, 'wb') as f: pickle.dump(vocab, f) print("Total vocabulary size: {}".format(len(vocab))) print("Saved the vocabulary wrapper to '{}'".format(vocab_path)) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--caption_path', type=str, default='../im2recipe-Pytorch/data/recipe1M/', help='path for train annotation file') parser.add_argument('--vocab_path', type=str, default='./data/ingredient_vocab.pkl', help='path for saving vocabulary wrapper') parser.add_argument('--threshold', type=int, default=4, help='minimum word count threshold') args = parser.parse_args() main(args)
2,783
31.752941
89
py
czsl
czsl-main/test.py
# Torch imports import torch from torch.utils.tensorboard import SummaryWriter import torch.backends.cudnn as cudnn import numpy as np from flags import DATA_FOLDER cudnn.benchmark = True # Python imports import tqdm from tqdm import tqdm import os from os.path import join as ospj # Local imports from data import dataset as dset from models.common import Evaluator from utils.utils import load_args from utils.config_model import configure_model from flags import parser device = 'cuda' if torch.cuda.is_available() else 'cpu' def main(): # Get arguments and start logging args = parser.parse_args() logpath = args.logpath config = [os.path.join(logpath, _) for _ in os.listdir(logpath) if _.endswith('yml')][0] load_args(config, args) # Get dataset trainset = dset.CompositionDataset( root=os.path.join(DATA_FOLDER,args.data_dir), phase='train', split=args.splitname, model=args.image_extractor, update_features=args.update_features, train_only=args.train_only, subset=args.subset, open_world=args.open_world ) valset = dset.CompositionDataset( root=os.path.join(DATA_FOLDER,args.data_dir), phase='val', split=args.splitname, model=args.image_extractor, subset=args.subset, update_features=args.update_features, open_world=args.open_world ) valoader = torch.utils.data.DataLoader( valset, batch_size=args.test_batch_size, shuffle=False, num_workers=8) testset = dset.CompositionDataset( root=os.path.join(DATA_FOLDER,args.data_dir), phase='test', split=args.splitname, model =args.image_extractor, subset=args.subset, update_features = args.update_features, open_world=args.open_world ) testloader = torch.utils.data.DataLoader( testset, batch_size=args.test_batch_size, shuffle=False, num_workers=args.workers) # Get model and optimizer image_extractor, model, optimizer = configure_model(args, trainset) args.extractor = image_extractor args.load = ospj(logpath,'ckpt_best_auc.t7') checkpoint = torch.load(args.load) if image_extractor: try: image_extractor.load_state_dict(checkpoint['image_extractor']) image_extractor.eval() except: print('No Image extractor in checkpoint') model.load_state_dict(checkpoint['net']) model.eval() threshold = None if args.open_world and args.hard_masking: assert args.model == 'compcos', args.model + ' does not have hard masking.' if args.threshold is not None: threshold = args.threshold else: evaluator_val = Evaluator(valset, model) unseen_scores = model.compute_feasibility().to('cpu') seen_mask = model.seen_mask.to('cpu') min_feasibility = (unseen_scores+seen_mask*10.).min() max_feasibility = (unseen_scores-seen_mask*10.).max() thresholds = np.linspace(min_feasibility,max_feasibility, num=args.threshold_trials) best_auc = 0. best_th = -10 with torch.no_grad(): for th in thresholds: results = test(image_extractor,model,valoader,evaluator_val,args,threshold=th,print_results=False) auc = results['AUC'] if auc > best_auc: best_auc = auc best_th = th print('New best AUC',best_auc) print('Threshold',best_th) threshold = best_th evaluator = Evaluator(testset, model) with torch.no_grad(): test(image_extractor, model, testloader, evaluator, args, threshold) def test(image_extractor, model, testloader, evaluator, args, threshold=None, print_results=True): if image_extractor: image_extractor.eval() model.eval() accuracies, all_sub_gt, all_attr_gt, all_obj_gt, all_pair_gt, all_pred = [], [], [], [], [], [] for idx, data in tqdm(enumerate(testloader), total=len(testloader), desc='Testing'): data = [d.to(device) for d in data] if image_extractor: data[0] = image_extractor(data[0]) if threshold is None: _, predictions = model(data) else: _, predictions = model.val_forward_with_threshold(data,threshold) attr_truth, obj_truth, pair_truth = data[1], data[2], data[3] all_pred.append(predictions) all_attr_gt.append(attr_truth) all_obj_gt.append(obj_truth) all_pair_gt.append(pair_truth) if args.cpu_eval: all_attr_gt, all_obj_gt, all_pair_gt = torch.cat(all_attr_gt), torch.cat(all_obj_gt), torch.cat(all_pair_gt) else: all_attr_gt, all_obj_gt, all_pair_gt = torch.cat(all_attr_gt).to('cpu'), torch.cat(all_obj_gt).to( 'cpu'), torch.cat(all_pair_gt).to('cpu') all_pred_dict = {} # Gather values as dict of (attr, obj) as key and list of predictions as values if args.cpu_eval: for k in all_pred[0].keys(): all_pred_dict[k] = torch.cat( [all_pred[i][k].to('cpu') for i in range(len(all_pred))]) else: for k in all_pred[0].keys(): all_pred_dict[k] = torch.cat( [all_pred[i][k] for i in range(len(all_pred))]) # Calculate best unseen accuracy results = evaluator.score_model(all_pred_dict, all_obj_gt, bias=args.bias, topk=args.topk) stats = evaluator.evaluate_predictions(results, all_attr_gt, all_obj_gt, all_pair_gt, all_pred_dict, topk=args.topk) result = '' for key in stats: result = result + key + ' ' + str(round(stats[key], 4)) + '| ' result = result + args.name if print_results: print(f'Results') print(result) return results if __name__ == '__main__': main()
6,227
32.12766
120
py
czsl
czsl-main/train.py
# Torch imports import torch import torch.nn as nn import torch.nn.functional as F from torch.utils.tensorboard import SummaryWriter import torch.backends.cudnn as cudnn cudnn.benchmark = True # Python imports import tqdm from tqdm import tqdm import os from os.path import join as ospj import csv #Local imports from data import dataset as dset from models.common import Evaluator from utils.utils import save_args, load_args from utils.config_model import configure_model from flags import parser, DATA_FOLDER best_auc = 0 best_hm = 0 compose_switch = True device = 'cuda' if torch.cuda.is_available() else 'cpu' def main(): # Get arguments and start logging args = parser.parse_args() load_args(args.config, args) logpath = os.path.join(args.cv_dir, args.name) os.makedirs(logpath, exist_ok=True) save_args(args, logpath, args.config) writer = SummaryWriter(log_dir = logpath, flush_secs = 30) # Get dataset trainset = dset.CompositionDataset( root=os.path.join(DATA_FOLDER,args.data_dir), phase='train', split=args.splitname, model =args.image_extractor, num_negs=args.num_negs, pair_dropout=args.pair_dropout, update_features = args.update_features, train_only= args.train_only, open_world=args.open_world ) trainloader = torch.utils.data.DataLoader( trainset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers) testset = dset.CompositionDataset( root=os.path.join(DATA_FOLDER,args.data_dir), phase=args.test_set, split=args.splitname, model =args.image_extractor, subset=args.subset, update_features = args.update_features, open_world=args.open_world ) testloader = torch.utils.data.DataLoader( testset, batch_size=args.test_batch_size, shuffle=False, num_workers=args.workers) # Get model and optimizer image_extractor, model, optimizer = configure_model(args, trainset) args.extractor = image_extractor train = train_normal evaluator_val = Evaluator(testset, model) print(model) start_epoch = 0 # Load checkpoint if args.load is not None: checkpoint = torch.load(args.load) if image_extractor: try: image_extractor.load_state_dict(checkpoint['image_extractor']) if args.freeze_features: print('Freezing image extractor') image_extractor.eval() for param in image_extractor.parameters(): param.requires_grad = False except: print('No Image extractor in checkpoint') model.load_state_dict(checkpoint['net']) start_epoch = checkpoint['epoch'] print('Loaded model from ', args.load) for epoch in tqdm(range(start_epoch, args.max_epochs + 1), desc = 'Current epoch'): train(epoch, image_extractor, model, trainloader, optimizer, writer) if model.is_open and args.model=='compcos' and ((epoch+1)%args.update_feasibility_every)==0 : print('Updating feasibility scores') model.update_feasibility(epoch+1.) if epoch % args.eval_val_every == 0: with torch.no_grad(): # todo: might not be needed test(epoch, image_extractor, model, testloader, evaluator_val, writer, args, logpath) print('Best AUC achieved is ', best_auc) print('Best HM achieved is ', best_hm) def train_normal(epoch, image_extractor, model, trainloader, optimizer, writer): ''' Runs training for an epoch ''' if image_extractor: image_extractor.train() model.train() # Let's switch to training train_loss = 0.0 for idx, data in tqdm(enumerate(trainloader), total=len(trainloader), desc = 'Training'): data = [d.to(device) for d in data] if image_extractor: data[0] = image_extractor(data[0]) loss, _ = model(data) optimizer.zero_grad() loss.backward() optimizer.step() train_loss += loss.item() train_loss = train_loss/len(trainloader) writer.add_scalar('Loss/train_total', train_loss, epoch) print('Epoch: {}| Loss: {}'.format(epoch, round(train_loss, 2))) def test(epoch, image_extractor, model, testloader, evaluator, writer, args, logpath): ''' Runs testing for an epoch ''' global best_auc, best_hm def save_checkpoint(filename): state = { 'net': model.state_dict(), 'epoch': epoch, 'AUC': stats['AUC'] } if image_extractor: state['image_extractor'] = image_extractor.state_dict() torch.save(state, os.path.join(logpath, 'ckpt_{}.t7'.format(filename))) if image_extractor: image_extractor.eval() model.eval() accuracies, all_sub_gt, all_attr_gt, all_obj_gt, all_pair_gt, all_pred = [], [], [], [], [], [] for idx, data in tqdm(enumerate(testloader), total=len(testloader), desc='Testing'): data = [d.to(device) for d in data] if image_extractor: data[0] = image_extractor(data[0]) _, predictions = model(data) attr_truth, obj_truth, pair_truth = data[1], data[2], data[3] all_pred.append(predictions) all_attr_gt.append(attr_truth) all_obj_gt.append(obj_truth) all_pair_gt.append(pair_truth) if args.cpu_eval: all_attr_gt, all_obj_gt, all_pair_gt = torch.cat(all_attr_gt), torch.cat(all_obj_gt), torch.cat(all_pair_gt) else: all_attr_gt, all_obj_gt, all_pair_gt = torch.cat(all_attr_gt).to('cpu'), torch.cat(all_obj_gt).to( 'cpu'), torch.cat(all_pair_gt).to('cpu') all_pred_dict = {} # Gather values as dict of (attr, obj) as key and list of predictions as values if args.cpu_eval: for k in all_pred[0].keys(): all_pred_dict[k] = torch.cat( [all_pred[i][k].to('cpu') for i in range(len(all_pred))]) else: for k in all_pred[0].keys(): all_pred_dict[k] = torch.cat( [all_pred[i][k] for i in range(len(all_pred))]) # Calculate best unseen accuracy results = evaluator.score_model(all_pred_dict, all_obj_gt, bias=args.bias, topk=args.topk) stats = evaluator.evaluate_predictions(results, all_attr_gt, all_obj_gt, all_pair_gt, all_pred_dict, topk=args.topk) stats['a_epoch'] = epoch result = '' # write to Tensorboard for key in stats: writer.add_scalar(key, stats[key], epoch) result = result + key + ' ' + str(round(stats[key], 4)) + '| ' result = result + args.name print(f'Test Epoch: {epoch}') print(result) if epoch > 0 and epoch % args.save_every == 0: save_checkpoint(epoch) if stats['AUC'] > best_auc: best_auc = stats['AUC'] print('New best AUC ', best_auc) save_checkpoint('best_auc') if stats['best_hm'] > best_hm: best_hm = stats['best_hm'] print('New best HM ', best_hm) save_checkpoint('best_hm') # Logs with open(ospj(logpath, 'logs.csv'), 'a') as f: w = csv.DictWriter(f, stats.keys()) if epoch == 0: w.writeheader() w.writerow(stats) if __name__ == '__main__': try: main() except KeyboardInterrupt: print('Best AUC achieved is ', best_auc) print('Best HM achieved is ', best_hm)
7,551
31.13617
120
py
czsl
czsl-main/models/svm.py
import numpy as np import tqdm from data import dataset as dset import os from utils import utils import torch from torch.autograd import Variable import h5py from sklearn.svm import LinearSVC from sklearn.model_selection import GridSearchCV import torch.nn.functional as F from joblib import Parallel, delayed import glob import scipy.io from sklearn.calibration import CalibratedClassifierCV import argparse parser = argparse.ArgumentParser() parser.add_argument('--batch_size', type=int, default=512) parser.add_argument('--dataset', default='mitstates', help='mitstates|zappos') parser.add_argument('--data_dir', default='data/mit-states/') parser.add_argument('--generate', action ='store_true', default=False) parser.add_argument('--evalsvm', action ='store_true', default=False) parser.add_argument('--evaltf', action ='store_true', default=False) parser.add_argument('--completed', default='tensor-completion/completed/complete.mat') args = parser.parse_args() #----------------------------------------------------------------------------------------# search_params = {'C': np.logspace(-5,5,11)} def train_svm(Y_train, sample_weight=None): try: clf = GridSearchCV(LinearSVC(class_weight='balanced', fit_intercept=False), search_params, scoring='f1', cv=4) clf.fit(X_train, Y_train, sample_weight=sample_weight) except: clf = LinearSVC(C=0.1, class_weight='balanced', fit_intercept=False) if Y_train.sum()==len(Y_train) or Y_train.sum()==0: return None clf.fit(X_train, Y_train, sample_weight=sample_weight) return clf def generate_svms(): # train an SVM for every attribute, object and pair primitive Y = [(train_attrs==attr).astype(np.int) for attr in range(len(dataset.attrs))] attr_clfs = Parallel(n_jobs=32, verbose=16)(delayed(train_svm)(Y[attr]) for attr in range(len(dataset.attrs))) for attr, clf in enumerate(attr_clfs): print (attr, dataset.attrs[attr]) print ('params:', clf.best_params_) print ('-'*30) torch.save(clf.best_estimator_, '%s/svm/attr_%d'%(args.data_dir, attr)) Y = [(train_objs==obj).astype(np.int) for obj in range(len(dataset.objs))] obj_clfs = Parallel(n_jobs=32, verbose=16)(delayed(train_svm)(Y[obj]) for obj in range(len(dataset.objs))) for obj, clf in enumerate(obj_clfs): print (obj, dataset.objs[obj]) print ('params:', clf.best_params_) print ('-'*30) torch.save(clf.best_estimator_, '%s/svm/obj_%d'%(args.data_dir, obj)) Y, Y_attr, sample_weight = [], [], [] for idx, (attr, obj) in enumerate(dataset.train_pairs): Y_train = ((train_attrs==attr)*(train_objs==obj)).astype(np.int) # reweight instances to get a little more training data Y_train_attr = (train_attrs==attr).astype(np.int) instance_weights = 0.1*Y_train_attr instance_weights[Y_train.nonzero()[0]] = 1.0 Y.append(Y_train) Y_attr.append(Y_train_attr) sample_weight.append(instance_weights) pair_clfs = Parallel(n_jobs=32, verbose=16)(delayed(train_svm)(Y[pair]) for pair in range(len(dataset.train_pairs))) for idx, (attr, obj) in enumerate(dataset.train_pairs): clf = pair_clfs[idx] print (dataset.attrs[attr], dataset.objs[obj]) try: print ('params:', clf.best_params_) torch.save(clf.best_estimator_, '%s/svm/pair_%d_%d'%(args.data_dir, attr, obj)) except: print ('FAILED! #positive:', Y[idx].sum(), len(Y[idx])) return def make_svm_tensor(): subs, vals, size = [], [], (len(dataset.attrs), len(dataset.objs), X.shape[1]) fullsubs, fullvals = [], [] composite_clfs = glob.glob('%s/svm/pair*'%args.data_dir) print ('%d composite classifiers found'%(len(composite_clfs))) for clf in tqdm.tqdm(composite_clfs): _, attr, obj = os.path.basename(clf).split('_') attr, obj = int(attr), int(obj) clf = torch.load(clf) weight = clf.coef_.squeeze() for i in range(len(weight)): subs.append((attr, obj, i)) vals.append(weight[i]) for attr, obj in dataset.pairs: for i in range(X.shape[1]): fullsubs.append((attr, obj, i)) subs, vals = np.array(subs), np.array(vals).reshape(-1,1) fullsubs, fullvals = np.array(fullsubs), np.ones(len(fullsubs)).reshape(-1,1) savedat = {'subs':subs, 'vals':vals, 'size':size, 'fullsubs':fullsubs, 'fullvals':fullvals} scipy.io.savemat('tensor-completion/incomplete/%s.mat'%args.dataset, savedat) print (subs.shape, vals.shape, size) def evaluate_svms(): attr_clfs = [torch.load('%s/svm/attr_%d'%(args.data_dir, attr)) for attr in range(len(dataset.attrs))] obj_clfs = [torch.load('%s/svm/obj_%d'%(args.data_dir, obj)) for obj in range(len(dataset.objs))] # Calibrate all classifiers first Y = [(train_attrs==attr).astype(np.int) for attr in range(len(dataset.attrs))] for attr in tqdm.tqdm(range(len(dataset.attrs))): clf = attr_clfs[attr] calibrated = CalibratedClassifierCV(clf, method='sigmoid', cv='prefit') calibrated.fit(X_train, Y[attr]) attr_clfs[attr] = calibrated Y = [(train_objs==obj).astype(np.int) for obj in range(len(dataset.objs))] for obj in tqdm.tqdm(range(len(dataset.objs))): clf = obj_clfs[obj] calibrated = CalibratedClassifierCV(clf, method='sigmoid', cv='prefit') calibrated.fit(X_train, Y[obj]) obj_clfs[obj] = calibrated # Generate all the scores attr_scores, obj_scores = [], [] for attr in tqdm.tqdm(range(len(dataset.attrs))): clf = attr_clfs[attr] score = clf.predict_proba(X_test)[:,1] attr_scores.append(score) attr_scores = np.vstack(attr_scores) for obj in tqdm.tqdm(range(len(dataset.objs))): clf = obj_clfs[obj] score = clf.predict_proba(X_test)[:,1] obj_scores.append(score) obj_scores = np.vstack(obj_scores) attr_pred = torch.from_numpy(attr_scores).transpose(0,1) obj_pred = torch.from_numpy(obj_scores).transpose(0,1) x = [None, Variable(torch.from_numpy(test_attrs)).long(), Variable(torch.from_numpy(test_objs)).long(), Variable(torch.from_numpy(test_pairs)).long()] attr_pred, obj_pred, _ = utils.generate_prediction_tensors([attr_pred, obj_pred], dataset, x[2].data, source='classification') attr_match, obj_match, zsl_match, gzsl_match, fixobj_match = utils.performance_stats(attr_pred, obj_pred, x) print (attr_match.mean(), obj_match.mean(), zsl_match.mean(), gzsl_match.mean(), fixobj_match.mean()) def evaluate_tensorcompletion(): def parse_tensor(fl): tensor = scipy.io.loadmat(fl) nz_idx = zip(*(tensor['subs'])) composite_clfs = np.zeros((len(dataset.attrs), len(dataset.objs), X.shape[1])) composite_clfs[nz_idx[0], nz_idx[1], nz_idx[2]] = tensor['vals'].squeeze() return composite_clfs, nz_idx, tensor['vals'].squeeze() # see recon error tr_file = 'tensor-completion/incomplete/%s.mat'%args.dataset ts_file = args.completed tr_clfs, tr_nz_idx, tr_vals = parse_tensor(tr_file) ts_clfs, ts_nz_idx, ts_vals = parse_tensor(ts_file) print (tr_vals.min(), tr_vals.max(), tr_vals.mean()) print (ts_vals.min(), ts_vals.max(), ts_vals.mean()) print ('Completed Tensor: %s'%args.completed) # see train recon error err = 1.0*((tr_clfs[tr_nz_idx[0], tr_nz_idx[1], tr_nz_idx[2]]-ts_clfs[tr_nz_idx[0], tr_nz_idx[1], tr_nz_idx[2]])**2).sum()/(len(tr_vals)) print ('recon error:', err) # Create and scale classifiers for each pair clfs = {} test_pair_set = set(map(tuple, dataset.test_pairs.numpy().tolist())) for idx, (attr, obj) in tqdm.tqdm(enumerate(dataset.pairs), total=len(dataset.pairs)): clf = LinearSVC(fit_intercept=False) clf.fit(np.eye(2), [0,1]) if (attr, obj) in test_pair_set: X_ = X_test Y_ = (test_attrs==attr).astype(np.int)*(test_objs==obj).astype(np.int) clf.coef_ = ts_clfs[attr, obj][None,:] else: X_ = X_train Y_ = (train_attrs==attr).astype(np.int)*(train_objs==obj).astype(np.int) clf.coef_ = tr_clfs[attr, obj][None,:] calibrated = CalibratedClassifierCV(clf, method='sigmoid', cv='prefit') calibrated.fit(X_, Y_) clfs[(attr, obj)] = calibrated scores = {} for attr, obj in tqdm.tqdm(dataset.pairs): score = clfs[(attr, obj)].predict_proba(X_test)[:,1] scores[(attr, obj)] = torch.from_numpy(score).float().unsqueeze(1) x = [None, Variable(torch.from_numpy(test_attrs)).long(), Variable(torch.from_numpy(test_objs)).long(), Variable(torch.from_numpy(test_pairs)).long()] attr_pred, obj_pred, _ = utils.generate_prediction_tensors(scores, dataset, x[2].data, source='manifold') attr_match, obj_match, zsl_match, gzsl_match, fixobj_match = utils.performance_stats(attr_pred, obj_pred, x) print (attr_match.mean(), obj_match.mean(), zsl_match.mean(), gzsl_match.mean(), fixobj_match.mean()) #----------------------------------------------------------------------------------------# if args.dataset == 'mitstates': DSet = dset.MITStatesActivations elif args.dataset == 'zappos': DSet = dset.UTZapposActivations dataset = DSet(root=args.data_dir, phase='train') train_idx, train_attrs, train_objs, train_pairs = map(np.array, zip(*dataset.train_data)) test_idx, test_attrs, test_objs, test_pairs = map(np.array, zip(*dataset.test_data)) X = dataset.activations.numpy() X_train, X_test = X[train_idx,:], X[test_idx,:] print (len(dataset.attrs), len(dataset.objs), len(dataset.pairs)) print (X_train.shape, X_test.shape) if args.generate: generate_svms() make_svm_tensor() if args.evalsvm: evaluate_svms() if args.evaltf: evaluate_tensorcompletion()
9,922
40.345833
154
py
czsl
czsl-main/models/visual_product.py
import torch import torch.nn as nn import torch.nn.functional as F import torchvision.models as models import numpy as np from .common import MLP class VisualProductNN(nn.Module): def __init__(self, dset, args): super(VisualProductNN, self).__init__() self.attr_clf = MLP(dset.feat_dim, len(dset.attrs), 2, relu = False) self.obj_clf = MLP(dset.feat_dim, len(dset.objs), 2, relu = False) self.dset = dset def train_forward(self, x): img, attrs, objs = x[0],x[1], x[2] attr_pred = self.attr_clf(img) obj_pred = self.obj_clf(img) attr_loss = F.cross_entropy(attr_pred, attrs) obj_loss = F.cross_entropy(obj_pred, objs) loss = attr_loss + obj_loss return loss, None def val_forward(self, x): img = x[0] attr_pred = F.softmax(self.attr_clf(img), dim =1) obj_pred = F.softmax(self.obj_clf(img), dim = 1) scores = {} for itr, (attr, obj) in enumerate(self.dset.pairs): attr_id, obj_id = self.dset.attr2idx[attr], self.dset.obj2idx[obj] score = attr_pred[:,attr_id] * obj_pred[:, obj_id] scores[(attr, obj)] = score return None, scores def forward(self, x): if self.training: loss, pred = self.train_forward(x) else: with torch.no_grad(): loss, pred = self.val_forward(x) return loss, pred
1,457
29.375
78
py
czsl
czsl-main/models/symnet.py
import torch import torch.nn as nn import torch.nn.functional as F import numpy as np from .word_embedding import load_word_embeddings from .common import MLP device = 'cuda' if torch.cuda.is_available() else 'cpu' class Symnet(nn.Module): def __init__(self, dset, args): super(Symnet, self).__init__() self.dset = dset self.args = args self.num_attrs = len(dset.attrs) self.num_objs = len(dset.objs) self.image_embedder = MLP(dset.feat_dim, args.emb_dim, relu = False) self.ce_loss = nn.CrossEntropyLoss() self.mse_loss = nn.MSELoss() self.nll_loss = nn.NLLLoss() self.softmax = nn.Softmax(dim=1) # Attribute embedder and object classifier self.attr_embedder = nn.Embedding(len(dset.attrs), args.emb_dim) self.obj_classifier = MLP(args.emb_dim, len(dset.objs), num_layers = 2, relu = False, dropout = True, layers = [512]) # original paper uses a second fc classifier for final product self.attr_classifier = MLP(args.emb_dim, len(dset.attrs), num_layers = 2, relu = False, dropout = True, layers = [512]) # CoN and DecoN self.CoN_fc_attention = MLP(args.emb_dim, args.emb_dim, num_layers = 2, relu = False, dropout = True, layers = [512]) self.CoN_emb = MLP(args.emb_dim + args.emb_dim, args.emb_dim, num_layers = 2, relu = False, dropout = True, layers = [768]) self.DecoN_fc_attention = MLP(args.emb_dim, args.emb_dim, num_layers = 2, relu = False, dropout = True, layers = [512]) self.DecoN_emb = MLP(args.emb_dim + args.emb_dim, args.emb_dim, num_layers = 2, relu = False, dropout = True, layers = [768]) # if args.glove_init: pretrained_weight = load_word_embeddings(args.emb_init, dset.attrs) self.attr_embedder.weight.data.copy_(pretrained_weight) for param in self.attr_embedder.parameters(): param.requires_grad = False def CoN(self, img_embedding, attr_embedding): attention = torch.sigmoid(self.CoN_fc_attention(attr_embedding)) img_embedding = attention*img_embedding + img_embedding hidden = torch.cat([img_embedding, attr_embedding], dim = 1) output = self.CoN_emb(hidden) return output def DeCoN(self, img_embedding, attr_embedding): attention = torch.sigmoid(self.DecoN_fc_attention(attr_embedding)) img_embedding = attention*img_embedding + img_embedding hidden = torch.cat([img_embedding, attr_embedding], dim = 1) output = self.DecoN_emb(hidden) return output def distance_metric(self, a, b): return torch.norm(a-b, dim = -1) def RMD_prob(self, feat_plus, feat_minus, repeat_img_feat): """return attribute classification probability with our RMD""" # feat_plus, feat_minus: shape=(bz, #attr, dim_emb) # d_plus: distance between feature before&after CoN # d_minus: distance between feature before&after DecoN d_plus = self.distance_metric(feat_plus, repeat_img_feat).reshape(-1, self.num_attrs) d_minus = self.distance_metric(feat_minus, repeat_img_feat).reshape(-1, self.num_attrs) # not adding softmax because it is part of cross entropy loss return d_minus - d_plus def train_forward(self, x): pos_image_feat, pos_attr_id, pos_obj_id = x[0], x[1], x[2] neg_attr_id = x[4][:,0] batch_size = pos_image_feat.size(0) loss = [] pos_attr_emb = self.attr_embedder(pos_attr_id) neg_attr_emb = self.attr_embedder(neg_attr_id) pos_img = self.image_embedder(pos_image_feat) # rA = remove positive attribute A # aA = add positive attribute A # rB = remove negative attribute B # aB = add negative attribute B pos_rA = self.DeCoN(pos_img, pos_attr_emb) pos_aA = self.CoN(pos_img, pos_attr_emb) pos_rB = self.DeCoN(pos_img, neg_attr_emb) pos_aB = self.CoN(pos_img, neg_attr_emb) # get all attr embedding #attr, embedding attr_emb = torch.LongTensor(np.arange(self.num_attrs)).to(device) attr_emb = self.attr_embedder(attr_emb) tile_attr_emb = attr_emb.repeat(batch_size, 1) # (batch*attr, dim_emb) # Now we calculate all the losses if self.args.lambda_cls_attr > 0: # Original image score_pos_A = self.attr_classifier(pos_img) loss_cls_pos_a = self.ce_loss(score_pos_A, pos_attr_id) # After removing pos_attr score_pos_rA_A = self.attr_classifier(pos_rA) total = sum(score_pos_rA_A) # prob_pos_rA_A = 1 - self.softmax(score_pos_rA_A) # loss_cls_pos_rA_a = self.nll_loss(torch.log(prob_pos_rA_A), pos_attr_id) loss_cls_pos_rA_a = self.ce_loss(total - score_pos_rA_A, pos_attr_id) #should be maximum for the gt label # rmd time repeat_img_feat = torch.repeat_interleave(pos_img, self.num_attrs, 0) #(batch*attr, dim_rep) feat_plus = self.CoN(repeat_img_feat, tile_attr_emb) feat_minus = self.DeCoN(repeat_img_feat, tile_attr_emb) score_cls_rmd = self.RMD_prob(feat_plus, feat_minus, repeat_img_feat) loss_cls_rmd = self.ce_loss(score_cls_rmd, pos_attr_id) loss_cls_attr = self.args.lambda_cls_attr*sum([loss_cls_pos_a, loss_cls_pos_rA_a, loss_cls_rmd]) loss.append(loss_cls_attr) if self.args.lambda_cls_obj > 0: # Original image score_pos_O = self.obj_classifier(pos_img) loss_cls_pos_o = self.ce_loss(score_pos_O, pos_obj_id) # After removing pos attr score_pos_rA_O = self.obj_classifier(pos_rA) loss_cls_pos_rA_o = self.ce_loss(score_pos_rA_O, pos_obj_id) # After adding neg attr score_pos_aB_O = self.obj_classifier(pos_aB) loss_cls_pos_aB_o = self.ce_loss(score_pos_aB_O, pos_obj_id) loss_cls_obj = self.args.lambda_cls_obj * sum([loss_cls_pos_o, loss_cls_pos_rA_o, loss_cls_pos_aB_o]) loss.append(loss_cls_obj) if self.args.lambda_sym > 0: loss_sys_pos = self.mse_loss(pos_aA, pos_img) loss_sys_neg = self.mse_loss(pos_rB, pos_img) loss_sym = self.args.lambda_sym * (loss_sys_pos + loss_sys_neg) loss.append(loss_sym) ##### Axiom losses if self.args.lambda_axiom > 0: loss_clo = loss_inv = loss_com = 0 # closure pos_aA_rA = self.DeCoN(pos_aA, pos_attr_emb) pos_rB_aB = self.CoN(pos_rB, neg_attr_emb) loss_clo = self.mse_loss(pos_aA_rA, pos_rA) + self.mse_loss(pos_rB_aB, pos_aB) # invertibility pos_rA_aA = self.CoN(pos_rA, pos_attr_emb) pos_aB_rB = self.DeCoN(pos_aB, neg_attr_emb) loss_inv = self.mse_loss(pos_rA_aA, pos_img) + self.mse_loss(pos_aB_rB, pos_img) # commutative pos_aA_rB = self.DeCoN(pos_aA, neg_attr_emb) pos_rB_aA = self.DeCoN(pos_rB, pos_attr_emb) loss_com = self.mse_loss(pos_aA_rB, pos_rB_aA) loss_axiom = self.args.lambda_axiom * (loss_clo + loss_inv + loss_com) loss.append(loss_axiom) # triplet loss if self.args.lambda_trip > 0: pos_triplet = F.triplet_margin_loss(pos_img, pos_aA, pos_rA) neg_triplet = F.triplet_margin_loss(pos_img, pos_rB, pos_aB) loss_triplet = self.args.lambda_trip * (pos_triplet + neg_triplet) loss.append(loss_triplet) loss = sum(loss) return loss, None def val_forward(self, x): pos_image_feat, pos_attr_id, pos_obj_id = x[0], x[1], x[2] batch_size = pos_image_feat.shape[0] pos_img = self.image_embedder(pos_image_feat) repeat_img_feat = torch.repeat_interleave(pos_img, self.num_attrs, 0) #(batch*attr, dim_rep) # get all attr embedding #attr, embedding attr_emb = torch.LongTensor(np.arange(self.num_attrs)).to(device) attr_emb = self.attr_embedder(attr_emb) tile_attr_emb = attr_emb.repeat(batch_size, 1) # (batch*attr, dim_emb) feat_plus = self.CoN(repeat_img_feat, tile_attr_emb) feat_minus = self.DeCoN(repeat_img_feat, tile_attr_emb) score_cls_rmd = self.RMD_prob(feat_plus, feat_minus, repeat_img_feat) prob_A_rmd = F.softmax(score_cls_rmd, dim = 1) score_obj = self.obj_classifier(pos_img) prob_O = F.softmax(score_obj, dim = 1) scores = {} for itr, (attr, obj) in enumerate(self.dset.pairs): attr_id, obj_id = self.dset.attr2idx[attr], self.dset.obj2idx[obj] score = prob_A_rmd[:,attr_id] * prob_O[:, obj_id] scores[(attr, obj)] = score return None, scores def forward(self, x): if self.training: loss, pred = self.train_forward(x) else: with torch.no_grad(): loss, pred = self.val_forward(x) return loss, pred
9,135
41.691589
188
py
czsl
czsl-main/models/manifold_methods.py
import torch import torch.nn as nn import torch.nn.functional as F import numpy as np from .word_embedding import load_word_embeddings from .common import MLP, Reshape from flags import DATA_FOLDER device = 'cuda' if torch.cuda.is_available() else 'cpu' class ManifoldModel(nn.Module): def __init__(self, dset, args): super(ManifoldModel, self).__init__() self.args = args self.dset = dset def get_all_ids(relevant_pairs): # Precompute validation pairs attrs, objs = zip(*relevant_pairs) attrs = [dset.attr2idx[attr] for attr in attrs] objs = [dset.obj2idx[obj] for obj in objs] pairs = [a for a in range(len(relevant_pairs))] attrs = torch.LongTensor(attrs).to(device) objs = torch.LongTensor(objs).to(device) pairs = torch.LongTensor(pairs).to(device) return attrs, objs, pairs #Validation self.val_attrs, self.val_objs, self.val_pairs = get_all_ids(self.dset.pairs) # for indivual projections self.uniq_attrs, self.uniq_objs = torch.arange(len(self.dset.attrs)).long().to(device), \ torch.arange(len(self.dset.objs)).long().to(device) self.factor = 2 # Precompute training compositions if args.train_only: self.train_attrs, self.train_objs, self.train_pairs = get_all_ids(self.dset.train_pairs) else: self.train_attrs, self.train_objs = self.val_subs, self.val_attrs, self.val_objs if args.lambda_aux > 0 or args.lambda_cls_attr > 0 or args.lambda_cls_obj > 0: print('Initializing classifiers') self.obj_clf = nn.Linear(args.emb_dim, len(dset.objs)) self.attr_clf = nn.Linear(args.emb_dim, len(dset.attrs)) def train_forward_bce(self, x): img, attrs, objs = x[0], x[1], x[2] neg_attrs, neg_objs = x[4][:,0],x[5][:,0] #todo: do a less hacky version img_feat = self.image_embedder(img) # Sample 25% positive and 75% negative pairs labels = np.random.binomial(1, 0.25, attrs.shape[0]) labels = torch.from_numpy(labels).bool().to(device) sampled_attrs, sampled_objs = neg_attrs.clone(), neg_objs.clone() sampled_attrs[labels] = attrs[labels] sampled_objs[labels] = objs[labels] labels = labels.float() composed_clf = self.compose(attrs, objs) p = torch.sigmoid((img_feat*composed_clf).sum(1)) loss = F.binary_cross_entropy(p, labels) return loss, None def train_forward_triplet(self, x): img, attrs, objs = x[0], x[1], x[2] neg_attrs, neg_objs = x[4][:,0], x[5][:,0] #todo:do a less hacky version img_feats = self.image_embedder(img) positive = self.compose(attrs, objs) negative = self.compose(neg_attrs, neg_objs) loss = F.triplet_margin_loss(img_feats, positive, negative, margin = self.args.margin) # Auxiliary object/ attribute prediction loss both need to be correct if self.args.lambda_aux > 0: obj_pred = self.obj_clf(positive) attr_pred = self.attr_clf(positive) loss_aux = F.cross_entropy(attr_pred, attrs) + F.cross_entropy(obj_pred, objs) loss += self.args.lambda_aux * loss_aux return loss, None def val_forward_distance(self, x): img = x[0] batch_size = img.shape[0] img_feats = self.image_embedder(img) scores = {} pair_embeds = self.compose(self.val_attrs, self.val_objs) for itr, pair in enumerate(self.dset.pairs): pair_embed = pair_embeds[itr, None].expand(batch_size, pair_embeds.size(1)) score = self.compare_metric(img_feats, pair_embed) scores[pair] = score return None, scores def val_forward_distance_fast(self, x): img = x[0] batch_size = img.shape[0] img_feats = self.image_embedder(img) pair_embeds = self.compose(self.val_attrs, self.val_objs) # Evaluate all pairs batch_size, pairs, features = img_feats.shape[0], pair_embeds.shape[0], pair_embeds.shape[1] img_feats = img_feats[:,None,:].expand(-1, pairs, -1) pair_embeds = pair_embeds[None,:,:].expand(batch_size, -1, -1) diff = (img_feats - pair_embeds)**2 score = diff.sum(2) * -1 scores = {} for itr, pair in enumerate(self.dset.pairs): scores[pair] = score[:,self.dset.all_pair2idx[pair]] return None, scores def val_forward_direct(self, x): img = x[0] batch_size = img.shape[0] img_feats = self.image_embedder(img) pair_embeds = self.compose(self.val_attrs, self.val_objs).permute(1,0) # Evaluate all pairs score = torch.matmul(img_feats, pair_embeds) scores = {} for itr, pair in enumerate(self.dset.pairs): scores[pair] = score[:,self.dset.all_pair2idx[pair]] return None, scores def forward(self, x): if self.training: loss, pred = self.train_forward(x) else: with torch.no_grad(): loss, pred = self.val_forward(x) return loss, pred ######################################### class RedWine(ManifoldModel): def __init__(self, dset, args): super(RedWine, self).__init__(dset, args) self.image_embedder = lambda img: img self.compare_metric = lambda img_feats, pair_embed: torch.sigmoid((img_feats*pair_embed).sum(1)) self.train_forward = self.train_forward_bce self.val_forward = self.val_forward_distance_fast in_dim = dset.feat_dim if args.clf_init else args.emb_dim self.T = nn.Sequential( nn.Linear(2*in_dim, 3*in_dim), nn.LeakyReLU(0.1, True), nn.Linear(3*in_dim, 3*in_dim//2), nn.LeakyReLU(0.1, True), nn.Linear(3*in_dim//2, dset.feat_dim)) self.attr_embedder = nn.Embedding(len(dset.attrs), in_dim) self.obj_embedder = nn.Embedding(len(dset.objs), in_dim) # initialize the weights of the embedders with the svm weights if args.emb_init: pretrained_weight = load_word_embeddings(args.emb_init, dset.attrs) self.attr_embedder.weight.data.copy_(pretrained_weight) pretrained_weight = load_word_embeddings(args.emb_init, dset.objs) self.obj_embedder.weight.data.copy_(pretrained_weight) elif args.clf_init: for idx, attr in enumerate(dset.attrs): at_id = self.dset.attr2idx[attr] weight = torch.load('%s/svm/attr_%d'%(args.data_dir, at_id)).coef_.squeeze() self.attr_embedder.weight[idx].data.copy_(torch.from_numpy(weight)) for idx, obj in enumerate(dset.objs): obj_id = self.dset.obj2idx[obj] weight = torch.load('%s/svm/obj_%d'%(args.data_dir, obj_id)).coef_.squeeze() self.obj_embedder.weight[idx].data.copy_(torch.from_numpy(weight)) else: print ('init must be either glove or clf') return if args.static_inp: for param in self.attr_embedder.parameters(): param.requires_grad = False for param in self.obj_embedder.parameters(): param.requires_grad = False def compose(self, attrs, objs): attr_wt = self.attr_embedder(attrs) obj_wt = self.obj_embedder(objs) inp_wts = torch.cat([attr_wt, obj_wt], 1) # 2D composed_clf = self.T(inp_wts) return composed_clf class LabelEmbedPlus(ManifoldModel): def __init__(self, dset, args): super(LabelEmbedPlus, self).__init__(dset, args) if 'conv' in args.image_extractor: self.image_embedder = torch.nn.Sequential(torch.nn.Conv2d(dset.feat_dim,args.emb_dim,7), torch.nn.ReLU(True), Reshape(-1,args.emb_dim) ) else: self.image_embedder = MLP(dset.feat_dim, args.emb_dim) self.compare_metric = lambda img_feats, pair_embed: -F.pairwise_distance(img_feats, pair_embed) self.train_forward = self.train_forward_triplet self.val_forward = self.val_forward_distance_fast input_dim = dset.feat_dim if args.clf_init else args.emb_dim self.attr_embedder = nn.Embedding(len(dset.attrs), input_dim) self.obj_embedder = nn.Embedding(len(dset.objs), input_dim) self.T = MLP(2*input_dim, args.emb_dim, num_layers= args.nlayers) # init with word embeddings if args.emb_init: pretrained_weight = load_word_embeddings(args.emb_init, dset.attrs) self.attr_embedder.weight.data.copy_(pretrained_weight) pretrained_weight = load_word_embeddings(args.emb_init, dset.objs) self.obj_embedder.weight.data.copy_(pretrained_weight) # init with classifier weights elif args.clf_init: for idx, attr in enumerate(dset.attrs): at_id = dset.attrs.index(attr) weight = torch.load('%s/svm/attr_%d'%(args.data_dir, at_id)).coef_.squeeze() self.attr_embedder.weight[idx].data.copy_(torch.from_numpy(weight)) for idx, obj in enumerate(dset.objs): obj_id = dset.objs.index(obj) weight = torch.load('%s/svm/obj_%d'%(args.data_dir, obj_id)).coef_.squeeze() self.obj_emb.weight[idx].data.copy_(torch.from_numpy(weight)) # static inputs if args.static_inp: for param in self.attr_embedder.parameters(): param.requires_grad = False for param in self.obj_embedder.parameters(): param.requires_grad = False def compose(self, attrs, objs): inputs = [self.attr_embedder(attrs), self.obj_embedder(objs)] inputs = torch.cat(inputs, 1) output = self.T(inputs) return output class AttributeOperator(ManifoldModel): def __init__(self, dset, args): super(AttributeOperator, self).__init__(dset, args) self.image_embedder = MLP(dset.feat_dim, args.emb_dim) self.compare_metric = lambda img_feats, pair_embed: -F.pairwise_distance(img_feats, pair_embed) self.val_forward = self.val_forward_distance_fast self.attr_ops = nn.ParameterList([nn.Parameter(torch.eye(args.emb_dim)) for _ in range(len(self.dset.attrs))]) self.obj_embedder = nn.Embedding(len(dset.objs), args.emb_dim) if args.emb_init: pretrained_weight = load_word_embeddings('glove', dset.objs) self.obj_embedder.weight.data.copy_(pretrained_weight) self.inverse_cache = {} if args.lambda_ant>0 and args.dataset=='mitstates': antonym_list = open(DATA_FOLDER+'/data/antonyms.txt').read().strip().split('\n') antonym_list = [l.split() for l in antonym_list] antonym_list = [[self.dset.attrs.index(a1), self.dset.attrs.index(a2)] for a1, a2 in antonym_list] antonyms = {} antonyms.update({a1:a2 for a1, a2 in antonym_list}) antonyms.update({a2:a1 for a1, a2 in antonym_list}) self.antonyms, self.antonym_list = antonyms, antonym_list if args.static_inp: for param in self.obj_embedder.parameters(): param.requires_grad = False def apply_ops(self, ops, rep): out = torch.bmm(ops, rep.unsqueeze(2)).squeeze(2) out = F.relu(out) return out def compose(self, attrs, objs): obj_rep = self.obj_embedder(objs) attr_ops = torch.stack([self.attr_ops[attr.item()] for attr in attrs]) embedded_reps = self.apply_ops(attr_ops, obj_rep) return embedded_reps def apply_inverse(self, img_rep, attrs): inverse_ops = [] for i in range(img_rep.size(0)): attr = attrs[i] if attr not in self.inverse_cache: self.inverse_cache[attr] = self.attr_ops[attr].inverse() inverse_ops.append(self.inverse_cache[attr]) inverse_ops = torch.stack(inverse_ops) # (B,512,512) obj_rep = self.apply_ops(inverse_ops, img_rep) return obj_rep def train_forward(self, x): img, attrs, objs = x[0], x[1], x[2] neg_attrs, neg_objs, inv_attrs, comm_attrs = x[4][:,0], x[5][:,0], x[6], x[7] batch_size = img.size(0) loss = [] anchor = self.image_embedder(img) obj_emb = self.obj_embedder(objs) pos_ops = torch.stack([self.attr_ops[attr.item()] for attr in attrs]) positive = self.apply_ops(pos_ops, obj_emb) neg_obj_emb = self.obj_embedder(neg_objs) neg_ops = torch.stack([self.attr_ops[attr.item()] for attr in neg_attrs]) negative = self.apply_ops(neg_ops, neg_obj_emb) loss_triplet = F.triplet_margin_loss(anchor, positive, negative, margin=self.args.margin) loss.append(loss_triplet) # Auxiliary object/attribute loss if self.args.lambda_aux>0: obj_pred = self.obj_clf(positive) attr_pred = self.attr_clf(positive) loss_aux = F.cross_entropy(attr_pred, attrs) + F.cross_entropy(obj_pred, objs) loss.append(self.args.lambda_aux*loss_aux) # Inverse Consistency if self.args.lambda_inv>0: obj_rep = self.apply_inverse(anchor, attrs) new_ops = torch.stack([self.attr_ops[attr.item()] for attr in inv_attrs]) new_rep = self.apply_ops(new_ops, obj_rep) new_positive = self.apply_ops(new_ops, obj_emb) loss_inv = F.triplet_margin_loss(new_rep, new_positive, positive, margin=self.args.margin) loss.append(self.args.lambda_inv*loss_inv) # Commutative Operators if self.args.lambda_comm>0: B = torch.stack([self.attr_ops[attr.item()] for attr in comm_attrs]) BA = self.apply_ops(B, positive) AB = self.apply_ops(pos_ops, self.apply_ops(B, obj_emb)) loss_comm = ((AB-BA)**2).sum(1).mean() loss.append(self.args.lambda_comm*loss_comm) # Antonym Consistency if self.args.lambda_ant>0: select_idx = [i for i in range(batch_size) if attrs[i].item() in self.antonyms] if len(select_idx)>0: select_idx = torch.LongTensor(select_idx).cuda() attr_subset = attrs[select_idx] antonym_ops = torch.stack([self.attr_ops[self.antonyms[attr.item()]] for attr in attr_subset]) Ao = anchor[select_idx] if self.args.lambda_inv>0: o = obj_rep[select_idx] else: o = self.apply_inverse(Ao, attr_subset) BAo = self.apply_ops(antonym_ops, Ao) loss_cycle = ((BAo-o)**2).sum(1).mean() loss.append(self.args.lambda_ant*loss_cycle) loss = sum(loss) return loss, None def forward(self, x): loss, pred = super(AttributeOperator, self).forward(x) self.inverse_cache = {} return loss, pred
15,296
40.681199
118
py
czsl
czsl-main/models/common.py
import torch import torch.nn as nn import torch.nn.functional as F import numpy as np import copy from scipy.stats import hmean device = 'cuda' if torch.cuda.is_available() else 'cpu' class MLP(nn.Module): ''' Baseclass to create a simple MLP Inputs inp_dim: Int, Input dimension out-dim: Int, Output dimension num_layer: Number of hidden layers relu: Bool, Use non linear function at output bias: Bool, Use bias ''' def __init__(self, inp_dim, out_dim, num_layers = 1, relu = True, bias = True, dropout = False, norm = False, layers = []): super(MLP, self).__init__() mod = [] incoming = inp_dim for layer in range(num_layers - 1): if len(layers) == 0: outgoing = incoming else: outgoing = layers.pop(0) mod.append(nn.Linear(incoming, outgoing, bias = bias)) incoming = outgoing if norm: mod.append(nn.LayerNorm(outgoing)) # mod.append(nn.BatchNorm1d(outgoing)) mod.append(nn.ReLU(inplace = True)) # mod.append(nn.LeakyReLU(inplace=True, negative_slope=0.2)) if dropout: mod.append(nn.Dropout(p = 0.5)) mod.append(nn.Linear(incoming, out_dim, bias = bias)) if relu: mod.append(nn.ReLU(inplace = True)) # mod.append(nn.LeakyReLU(inplace=True, negative_slope=0.2)) self.mod = nn.Sequential(*mod) def forward(self, x): return self.mod(x) class Reshape(nn.Module): def __init__(self, *args): super(Reshape, self).__init__() self.shape = args def forward(self, x): return x.view(self.shape) def calculate_margines(domain_embedding, gt, margin_range = 5): ''' domain_embedding: pairs * feats gt: batch * feats ''' batch_size, pairs, features = gt.shape[0], domain_embedding.shape[0], domain_embedding.shape[1] gt_expanded = gt[:,None,:].expand(-1, pairs, -1) domain_embedding_expanded = domain_embedding[None, :, :].expand(batch_size, -1, -1) margin = (gt_expanded - domain_embedding_expanded)**2 margin = margin.sum(2) max_margin, _ = torch.max(margin, dim = 0) margin /= max_margin margin *= margin_range return margin def l2_all_batched(image_embedding, domain_embedding): ''' Image Embedding: Tensor of Batch_size * pairs * Feature_dim domain_embedding: Tensor of pairs * Feature_dim ''' pairs = image_embedding.shape[1] domain_embedding_extended = image_embedding[:,None,:].expand(-1,pairs,-1) l2_loss = (image_embedding - domain_embedding_extended) ** 2 l2_loss = l2_loss.sum(2) l2_loss = l2_loss.sum() / l2_loss.numel() return l2_loss def same_domain_triplet_loss(image_embedding, trip_images, gt, hard_k = None, margin = 2): ''' Image Embedding: Tensor of Batch_size * Feature_dim Triplet Images: Tensor of Batch_size * num_pairs * Feature_dim GT: Tensor of Batch_size ''' batch_size, pairs, features = trip_images.shape batch_iterator = torch.arange(batch_size).to(device) image_embedding_expanded = image_embedding[:,None,:].expand(-1, pairs, -1) diff = (image_embedding_expanded - trip_images)**2 diff = diff.sum(2) positive_anchor = diff[batch_iterator, gt][:,None] positive_anchor = positive_anchor.expand(-1, pairs) # Calculating triplet loss triplet_loss = positive_anchor - diff + margin # Setting positive anchor loss to 0 triplet_loss[batch_iterator, gt] = 0 # Removing easy triplets triplet_loss[triplet_loss < 0] = 0 # If only mining hard triplets if hard_k: triplet_loss, _ = triplet_loss.topk(hard_k) # Counting number of valid pairs num_positive_triplets = triplet_loss[triplet_loss > 1e-16].size(0) # Calculating the final loss triplet_loss = triplet_loss.sum() / (num_positive_triplets + 1e-16) return triplet_loss def cross_domain_triplet_loss(image_embedding, domain_embedding, gt, hard_k = None, margin = 2): ''' Image Embedding: Tensor of Batch_size * Feature_dim Domain Embedding: Tensor of Num_pairs * Feature_dim gt: Tensor of Batch_size with ground truth labels margin: Float of margin Returns: Triplet loss of all valid triplets ''' batch_size, pairs, features = image_embedding.shape[0], domain_embedding.shape[0], domain_embedding.shape[1] batch_iterator = torch.arange(batch_size).to(device) # Now dimensions will be Batch_size * Num_pairs * Feature_dim image_embedding = image_embedding[:,None,:].expand(-1, pairs, -1) domain_embedding = domain_embedding[None,:,:].expand(batch_size, -1, -1) # Calculating difference diff = (image_embedding - domain_embedding)**2 diff = diff.sum(2) # Getting the positive pair positive_anchor = diff[batch_iterator, gt][:,None] positive_anchor = positive_anchor.expand(-1, pairs) # Calculating triplet loss triplet_loss = positive_anchor - diff + margin # Setting positive anchor loss to 0 triplet_loss[batch_iterator, gt] = 0 # Removing easy triplets triplet_loss[triplet_loss < 0] = 0 # If only mining hard triplets if hard_k: triplet_loss, _ = triplet_loss.topk(hard_k) # Counting number of valid pairs num_positive_triplets = triplet_loss[triplet_loss > 1e-16].size(0) # Calculating the final loss triplet_loss = triplet_loss.sum() / (num_positive_triplets + 1e-16) return triplet_loss def same_domain_triplet_loss_old(image_embedding, positive_anchor, negative_anchor, margin = 2): ''' Image Embedding: Tensor of Batch_size * Feature_dim Positive anchor: Tensor of Batch_size * Feature_dim negative anchor: Tensor of Batch_size * negs *Feature_dim ''' batch_size, negs, features = negative_anchor.shape dist_pos = (image_embedding - positive_anchor)**2 dist_pos = dist_pos.sum(1) dist_pos = dist_pos[:, None].expand(-1, negs) image_embedding_expanded = image_embedding[:,None,:].expand(-1, negs, -1) dist_neg = (image_embedding_expanded - negative_anchor)**2 dist_neg = dist_neg.sum(2) triplet_loss = dist_pos - dist_neg + margin triplet_loss[triplet_loss < 0] = 0 num_positive_triplets = triplet_loss[triplet_loss > 1e-16].size(0) triplet_loss = triplet_loss.sum() / (num_positive_triplets + 1e-16) return triplet_loss def pairwise_distances(x, y=None): ''' Input: x is a Nxd matrix y is an optional Mxd matirx Output: dist is a NxM matrix where dist[i,j] is the square norm between x[i,:] and y[j,:] if y is not given then use 'y=x'. i.e. dist[i,j] = ||x[i,:]-y[j,:]||^2 ''' x_norm = (x**2).sum(1).view(-1, 1) if y is not None: y_t = torch.transpose(y, 0, 1) y_norm = (y**2).sum(1).view(1, -1) else: y_t = torch.transpose(x, 0, 1) y_norm = x_norm.view(1, -1) dist = x_norm + y_norm - 2.0 * torch.mm(x, y_t) # Ensure diagonal is zero if x=y # if y is None: # dist = dist - torch.diag(dist.diag) return torch.clamp(dist, 0.0, np.inf) class Evaluator: def __init__(self, dset, model): self.dset = dset # Convert text pairs to idx tensors: [('sliced', 'apple'), ('ripe', 'apple'), ...] --> torch.LongTensor([[0,1],[1,1], ...]) pairs = [(dset.attr2idx[attr], dset.obj2idx[obj]) for attr, obj in dset.pairs] self.train_pairs = [(dset.attr2idx[attr], dset.obj2idx[obj]) for attr, obj in dset.train_pairs] self.pairs = torch.LongTensor(pairs) # Mask over pairs that occur in closed world # Select set based on phase if dset.phase == 'train': print('Evaluating with train pairs') test_pair_set = set(dset.train_pairs) test_pair_gt = set(dset.train_pairs) elif dset.phase == 'val': print('Evaluating with validation pairs') test_pair_set = set(dset.val_pairs + dset.train_pairs) test_pair_gt = set(dset.val_pairs) else: print('Evaluating with test pairs') test_pair_set = set(dset.test_pairs + dset.train_pairs) test_pair_gt = set(dset.test_pairs) self.test_pair_dict = [(dset.attr2idx[attr], dset.obj2idx[obj]) for attr, obj in test_pair_gt] self.test_pair_dict = dict.fromkeys(self.test_pair_dict, 0) # dict values are pair val, score, total for attr, obj in test_pair_gt: pair_val = dset.pair2idx[(attr,obj)] key = (dset.attr2idx[attr], dset.obj2idx[obj]) self.test_pair_dict[key] = [pair_val, 0, 0] if dset.open_world: masks = [1 for _ in dset.pairs] else: masks = [1 if pair in test_pair_set else 0 for pair in dset.pairs] self.closed_mask = torch.BoolTensor(masks) # Mask of seen concepts seen_pair_set = set(dset.train_pairs) mask = [1 if pair in seen_pair_set else 0 for pair in dset.pairs] self.seen_mask = torch.BoolTensor(mask) # Object specific mask over which pairs occur in the object oracle setting oracle_obj_mask = [] for _obj in dset.objs: mask = [1 if _obj == obj else 0 for attr, obj in dset.pairs] oracle_obj_mask.append(torch.BoolTensor(mask)) self.oracle_obj_mask = torch.stack(oracle_obj_mask, 0) # Decide if the model under evaluation is a manifold model or not self.score_model = self.score_manifold_model # Generate mask for each settings, mask scores, and get prediction labels def generate_predictions(self, scores, obj_truth, bias = 0.0, topk = 5): # (Batch, #pairs) ''' Inputs scores: Output scores obj_truth: Ground truth object Returns results: dict of results in 3 settings ''' def get_pred_from_scores(_scores, topk): ''' Given list of scores, returns top 10 attr and obj predictions Check later ''' _, pair_pred = _scores.topk(topk, dim = 1) #sort returns indices of k largest values pair_pred = pair_pred.contiguous().view(-1) attr_pred, obj_pred = self.pairs[pair_pred][:, 0].view(-1, topk), \ self.pairs[pair_pred][:, 1].view(-1, topk) return (attr_pred, obj_pred) results = {} orig_scores = scores.clone() mask = self.seen_mask.repeat(scores.shape[0],1) # Repeat mask along pairs dimension scores[~mask] += bias # Add bias to test pairs # Unbiased setting # Open world setting --no mask, all pairs of the dataset results.update({'open': get_pred_from_scores(scores, topk)}) results.update({'unbiased_open': get_pred_from_scores(orig_scores, topk)}) # Closed world setting - set the score for all Non test pairs to -1e10, # this excludes the pairs from set not in evaluation mask = self.closed_mask.repeat(scores.shape[0], 1) closed_scores = scores.clone() closed_scores[~mask] = -1e10 closed_orig_scores = orig_scores.clone() closed_orig_scores[~mask] = -1e10 results.update({'closed': get_pred_from_scores(closed_scores, topk)}) results.update({'unbiased_closed': get_pred_from_scores(closed_orig_scores, topk)}) # Object_oracle setting - set the score to -1e10 for all pairs where the true object does Not participate, can also use the closed score mask = self.oracle_obj_mask[obj_truth] oracle_obj_scores = scores.clone() oracle_obj_scores[~mask] = -1e10 oracle_obj_scores_unbiased = orig_scores.clone() oracle_obj_scores_unbiased[~mask] = -1e10 results.update({'object_oracle': get_pred_from_scores(oracle_obj_scores, 1)}) results.update({'object_oracle_unbiased': get_pred_from_scores(oracle_obj_scores_unbiased, 1)}) return results def score_clf_model(self, scores, obj_truth, topk = 5): ''' Wrapper function to call generate_predictions for CLF models ''' attr_pred, obj_pred = scores # Go to CPU attr_pred, obj_pred, obj_truth = attr_pred.to('cpu'), obj_pred.to('cpu'), obj_truth.to('cpu') # Gather scores (P(a), P(o)) for all relevant (a,o) pairs # Multiply P(a) * P(o) to get P(pair) attr_subset = attr_pred.index_select(1, self.pairs[:,0]) # Return only attributes that are in our pairs obj_subset = obj_pred.index_select(1, self.pairs[:, 1]) scores = (attr_subset * obj_subset) # (Batch, #pairs) results = self.generate_predictions(scores, obj_truth) results['biased_scores'] = scores return results def score_manifold_model(self, scores, obj_truth, bias = 0.0, topk = 5): ''' Wrapper function to call generate_predictions for manifold models ''' # Go to CPU scores = {k: v.to('cpu') for k, v in scores.items()} obj_truth = obj_truth.to(device) # Gather scores for all relevant (a,o) pairs scores = torch.stack( [scores[(attr,obj)] for attr, obj in self.dset.pairs], 1 ) # (Batch, #pairs) orig_scores = scores.clone() results = self.generate_predictions(scores, obj_truth, bias, topk) results['scores'] = orig_scores return results def score_fast_model(self, scores, obj_truth, bias = 0.0, topk = 5): ''' Wrapper function to call generate_predictions for manifold models ''' results = {} mask = self.seen_mask.repeat(scores.shape[0],1) # Repeat mask along pairs dimension scores[~mask] += bias # Add bias to test pairs mask = self.closed_mask.repeat(scores.shape[0], 1) closed_scores = scores.clone() closed_scores[~mask] = -1e10 _, pair_pred = closed_scores.topk(topk, dim = 1) #sort returns indices of k largest values pair_pred = pair_pred.contiguous().view(-1) attr_pred, obj_pred = self.pairs[pair_pred][:, 0].view(-1, topk), \ self.pairs[pair_pred][:, 1].view(-1, topk) results.update({'closed': (attr_pred, obj_pred)}) return results def evaluate_predictions(self, predictions, attr_truth, obj_truth, pair_truth, allpred, topk = 1): # Go to CPU attr_truth, obj_truth, pair_truth = attr_truth.to('cpu'), obj_truth.to('cpu'), pair_truth.to('cpu') pairs = list( zip(list(attr_truth.numpy()), list(obj_truth.numpy()))) seen_ind, unseen_ind = [], [] for i in range(len(attr_truth)): if pairs[i] in self.train_pairs: seen_ind.append(i) else: unseen_ind.append(i) seen_ind, unseen_ind = torch.LongTensor(seen_ind), torch.LongTensor(unseen_ind) def _process(_scores): # Top k pair accuracy # Attribute, object and pair attr_match = (attr_truth.unsqueeze(1).repeat(1, topk) == _scores[0][:, :topk]) obj_match = (obj_truth.unsqueeze(1).repeat(1, topk) == _scores[1][:, :topk]) # Match of object pair match = (attr_match * obj_match).any(1).float() attr_match = attr_match.any(1).float() obj_match = obj_match.any(1).float() # Match of seen and unseen pairs seen_match = match[seen_ind] unseen_match = match[unseen_ind] ### Calculating class average accuracy # local_score_dict = copy.deepcopy(self.test_pair_dict) # for pair_gt, pair_pred in zip(pairs, match): # # print(pair_gt) # local_score_dict[pair_gt][2] += 1.0 #increase counter # if int(pair_pred) == 1: # local_score_dict[pair_gt][1] += 1.0 # # Now we have hits and totals for classes in evaluation set # seen_score, unseen_score = [], [] # for key, (idx, hits, total) in local_score_dict.items(): # score = hits/total # if bool(self.seen_mask[idx]) == True: # seen_score.append(score) # else: # unseen_score.append(score) seen_score, unseen_score = torch.ones(512,5), torch.ones(512,5) return attr_match, obj_match, match, seen_match, unseen_match, \ torch.Tensor(seen_score+unseen_score), torch.Tensor(seen_score), torch.Tensor(unseen_score) def _add_to_dict(_scores, type_name, stats): base = ['_attr_match', '_obj_match', '_match', '_seen_match', '_unseen_match', '_ca', '_seen_ca', '_unseen_ca'] for val, name in zip(_scores, base): stats[type_name + name] = val ##################### Match in places where corrent object obj_oracle_match = (attr_truth == predictions['object_oracle'][0][:, 0]).float() #object is already conditioned obj_oracle_match_unbiased = (attr_truth == predictions['object_oracle_unbiased'][0][:, 0]).float() stats = dict(obj_oracle_match = obj_oracle_match, obj_oracle_match_unbiased = obj_oracle_match_unbiased) #################### Closed world closed_scores = _process(predictions['closed']) unbiased_closed = _process(predictions['unbiased_closed']) _add_to_dict(closed_scores, 'closed', stats) _add_to_dict(unbiased_closed, 'closed_ub', stats) #################### Calculating AUC scores = predictions['scores'] # getting score for each ground truth class correct_scores = scores[torch.arange(scores.shape[0]), pair_truth][unseen_ind] # Getting top predicted score for these unseen classes max_seen_scores = predictions['scores'][unseen_ind][:, self.seen_mask].topk(topk, dim=1)[0][:, topk - 1] # Getting difference between these scores unseen_score_diff = max_seen_scores - correct_scores # Getting matched classes at max bias for diff unseen_matches = stats['closed_unseen_match'].bool() correct_unseen_score_diff = unseen_score_diff[unseen_matches] - 1e-4 # sorting these diffs correct_unseen_score_diff = torch.sort(correct_unseen_score_diff)[0] magic_binsize = 20 # getting step size for these bias values bias_skip = max(len(correct_unseen_score_diff) // magic_binsize, 1) # Getting list biaslist = correct_unseen_score_diff[::bias_skip] seen_match_max = float(stats['closed_seen_match'].mean()) unseen_match_max = float(stats['closed_unseen_match'].mean()) seen_accuracy, unseen_accuracy = [], [] # Go to CPU base_scores = {k: v.to('cpu') for k, v in allpred.items()} obj_truth = obj_truth.to('cpu') # Gather scores for all relevant (a,o) pairs base_scores = torch.stack( [allpred[(attr,obj)] for attr, obj in self.dset.pairs], 1 ) # (Batch, #pairs) for bias in biaslist: scores = base_scores.clone() results = self.score_fast_model(scores, obj_truth, bias = bias, topk = topk) results = results['closed'] # we only need biased results = _process(results) seen_match = float(results[3].mean()) unseen_match = float(results[4].mean()) seen_accuracy.append(seen_match) unseen_accuracy.append(unseen_match) seen_accuracy.append(seen_match_max) unseen_accuracy.append(unseen_match_max) seen_accuracy, unseen_accuracy = np.array(seen_accuracy), np.array(unseen_accuracy) area = np.trapz(seen_accuracy, unseen_accuracy) for key in stats: stats[key] = float(stats[key].mean()) harmonic_mean = hmean([seen_accuracy, unseen_accuracy], axis = 0) max_hm = np.max(harmonic_mean) idx = np.argmax(harmonic_mean) if idx == len(biaslist): bias_term = 1e3 else: bias_term = biaslist[idx] stats['biasterm'] = float(bias_term) stats['best_unseen'] = np.max(unseen_accuracy) stats['best_seen'] = np.max(seen_accuracy) stats['AUC'] = area stats['hm_unseen'] = unseen_accuracy[idx] stats['hm_seen'] = seen_accuracy[idx] stats['best_hm'] = max_hm return stats
20,689
39.174757
144
py
czsl
czsl-main/models/gcn.py
import numpy as np import scipy.sparse as sp import math import torch import torch.nn as nn import torch.nn.functional as F from torch.nn.init import xavier_uniform_ device = 'cuda' if torch.cuda.is_available() else 'cpu' def normt_spm(mx, method='in'): if method == 'in': mx = mx.transpose() rowsum = np.array(mx.sum(1)) r_inv = np.power(rowsum, -1).flatten() r_inv[np.isinf(r_inv)] = 0. r_mat_inv = sp.diags(r_inv) mx = r_mat_inv.dot(mx) return mx if method == 'sym': rowsum = np.array(mx.sum(1)) r_inv = np.power(rowsum, -0.5).flatten() r_inv[np.isinf(r_inv)] = 0. r_mat_inv = sp.diags(r_inv) mx = mx.dot(r_mat_inv).transpose().dot(r_mat_inv) return mx def spm_to_tensor(sparse_mx): sparse_mx = sparse_mx.tocoo().astype(np.float32) indices = torch.from_numpy(np.vstack( (sparse_mx.row, sparse_mx.col))).long() values = torch.from_numpy(sparse_mx.data) shape = torch.Size(sparse_mx.shape) return torch.sparse.FloatTensor(indices, values, shape) class GraphConv(nn.Module): def __init__(self, in_channels, out_channels, dropout=False, relu=True): super().__init__() if dropout: self.dropout = nn.Dropout(p=0.5) else: self.dropout = None self.layer = nn.Linear(in_channels, out_channels) if relu: # self.relu = nn.LeakyReLU(negative_slope=0.2) self.relu = nn.ReLU() else: self.relu = None def forward(self, inputs, adj): if self.dropout is not None: inputs = self.dropout(inputs) outputs = torch.mm(adj, torch.mm(inputs, self.layer.weight.T)) + self.layer.bias if self.relu is not None: outputs = self.relu(outputs) return outputs class GCN(nn.Module): def __init__(self, adj, in_channels, out_channels, hidden_layers): super().__init__() adj = normt_spm(adj, method='in') adj = spm_to_tensor(adj) self.adj = adj.to(device) self.train_adj = self.adj hl = hidden_layers.split(',') if hl[-1] == 'd': dropout_last = True hl = hl[:-1] else: dropout_last = False i = 0 layers = [] last_c = in_channels for c in hl: if c[0] == 'd': dropout = True c = c[1:] else: dropout = False c = int(c) i += 1 conv = GraphConv(last_c, c, dropout=dropout) self.add_module('conv{}'.format(i), conv) layers.append(conv) last_c = c conv = GraphConv(last_c, out_channels, relu=False, dropout=dropout_last) self.add_module('conv-last', conv) layers.append(conv) self.layers = layers def forward(self, x): if self.training: for conv in self.layers: x = conv(x, self.train_adj) else: for conv in self.layers: x = conv(x, self.adj) return F.normalize(x) ### GCNII class GraphConvolution(nn.Module): def __init__(self, in_features, out_features, dropout=False, relu=True, residual=False, variant=False): super(GraphConvolution, self).__init__() self.variant = variant if self.variant: self.in_features = 2*in_features else: self.in_features = in_features if dropout: self.dropout = nn.Dropout(p=0.5) else: self.dropout = None if relu: self.relu = nn.ReLU() else: self.relu = None self.out_features = out_features self.residual = residual self.layer = nn.Linear(self.in_features, self.out_features, bias = False) def reset_parameters(self): stdv = 1. / math.sqrt(self.out_features) self.weight.data.uniform_(-stdv, stdv) def forward(self, input, adj , h0 , lamda, alpha, l): if self.dropout is not None: input = self.dropout(input) theta = math.log(lamda/l+1) hi = torch.spmm(adj, input) if self.variant: support = torch.cat([hi,h0],1) r = (1-alpha)*hi+alpha*h0 else: support = (1-alpha)*hi+alpha*h0 r = support mm_term = torch.mm(support, self.layer.weight.T) output = theta*mm_term+(1-theta)*r if self.residual: output = output+input if self.relu is not None: output = self.relu(output) return output class GCNII(nn.Module): def __init__(self, adj, in_channels , out_channels, hidden_dim, hidden_layers, lamda, alpha, variant, dropout = True): super(GCNII, self).__init__() self.alpha = alpha self.lamda = lamda adj = normt_spm(adj, method='in') adj = spm_to_tensor(adj) self.adj = adj.to(device) i = 0 layers = nn.ModuleList() self.fc_dim = nn.Linear(in_channels, hidden_dim) self.relu = nn.ReLU() self.dropout = nn.Dropout() for i, c in enumerate(range(hidden_layers)): conv = GraphConvolution(hidden_dim, hidden_dim, variant=variant, dropout=dropout) layers.append(conv) self.layers = layers self.fc_out = nn.Linear(hidden_dim, out_channels) def forward(self, x): _layers = [] layer_inner = self.relu(self.fc_dim(self.dropout(x))) # layer_inner = x _layers.append(layer_inner) for i,con in enumerate(self.layers): layer_inner = con(layer_inner,self.adj,_layers[0],self.lamda,self.alpha,i+1) layer_inner = self.fc_out(self.dropout(layer_inner)) return layer_inner
5,869
27.634146
122
py
czsl
czsl-main/models/word_embedding.py
import torch import numpy as np from flags import DATA_FOLDER def load_word_embeddings(emb_type, vocab): if emb_type == 'glove': embeds = load_glove_embeddings(vocab) elif emb_type == 'fasttext': embeds = load_fasttext_embeddings(vocab) elif emb_type == 'word2vec': embeds = load_word2vec_embeddings(vocab) elif emb_type == 'ft+w2v': embeds1 = load_fasttext_embeddings(vocab) embeds2 = load_word2vec_embeddings(vocab) embeds = torch.cat([embeds1, embeds2], dim = 1) print('Combined embeddings are ',embeds.shape) elif emb_type == 'ft+gl': embeds1 = load_fasttext_embeddings(vocab) embeds2 = load_glove_embeddings(vocab) embeds = torch.cat([embeds1, embeds2], dim = 1) print('Combined embeddings are ',embeds.shape) elif emb_type == 'ft+ft': embeds1 = load_fasttext_embeddings(vocab) embeds = torch.cat([embeds1, embeds1], dim = 1) print('Combined embeddings are ',embeds.shape) elif emb_type == 'gl+w2v': embeds1 = load_glove_embeddings(vocab) embeds2 = load_word2vec_embeddings(vocab) embeds = torch.cat([embeds1, embeds2], dim = 1) print('Combined embeddings are ',embeds.shape) elif emb_type == 'ft+w2v+gl': embeds1 = load_fasttext_embeddings(vocab) embeds2 = load_word2vec_embeddings(vocab) embeds3 = load_glove_embeddings(vocab) embeds = torch.cat([embeds1, embeds2, embeds3], dim = 1) print('Combined embeddings are ',embeds.shape) else: raise ValueError('Invalid embedding') return embeds def load_fasttext_embeddings(vocab): custom_map = { 'Faux.Fur': 'fake fur', 'Faux.Leather': 'fake leather', 'Full.grain.leather': 'thick leather', 'Hair.Calf': 'hairy leather', 'Patent.Leather': 'shiny leather', 'Boots.Ankle': 'ankle boots', 'Boots.Knee.High': 'kneehigh boots', 'Boots.Mid-Calf': 'midcalf boots', 'Shoes.Boat.Shoes': 'boatshoes', 'Shoes.Clogs.and.Mules': 'clogs shoes', 'Shoes.Flats': 'flats shoes', 'Shoes.Heels': 'heels', 'Shoes.Loafers': 'loafers', 'Shoes.Oxfords': 'oxford shoes', 'Shoes.Sneakers.and.Athletic.Shoes': 'sneakers', 'traffic_light': 'traficlight', 'trash_can': 'trashcan', 'dry-erase_board' : 'dry_erase_board', 'black_and_white' : 'black_white', 'eiffel_tower' : 'tower' } vocab_lower = [v.lower() for v in vocab] vocab = [] for current in vocab_lower: if current in custom_map: vocab.append(custom_map[current]) else: vocab.append(current) import fasttext.util ft = fasttext.load_model(DATA_FOLDER+'/fast/cc.en.300.bin') embeds = [] for k in vocab: if '_' in k: ks = k.split('_') emb = np.stack([ft.get_word_vector(it) for it in ks]).mean(axis=0) else: emb = ft.get_word_vector(k) embeds.append(emb) embeds = torch.Tensor(np.stack(embeds)) print('Fasttext Embeddings loaded, total embeddings: {}'.format(embeds.size())) return embeds def load_word2vec_embeddings(vocab): # vocab = [v.lower() for v in vocab] from gensim import models model = models.KeyedVectors.load_word2vec_format( DATA_FOLDER+'/w2v/GoogleNews-vectors-negative300.bin', binary=True) custom_map = { 'Faux.Fur': 'fake_fur', 'Faux.Leather': 'fake_leather', 'Full.grain.leather': 'thick_leather', 'Hair.Calf': 'hair_leather', 'Patent.Leather': 'shiny_leather', 'Boots.Ankle': 'ankle_boots', 'Boots.Knee.High': 'knee_high_boots', 'Boots.Mid-Calf': 'midcalf_boots', 'Shoes.Boat.Shoes': 'boat_shoes', 'Shoes.Clogs.and.Mules': 'clogs_shoes', 'Shoes.Flats': 'flats_shoes', 'Shoes.Heels': 'heels', 'Shoes.Loafers': 'loafers', 'Shoes.Oxfords': 'oxford_shoes', 'Shoes.Sneakers.and.Athletic.Shoes': 'sneakers', 'traffic_light': 'traffic_light', 'trash_can': 'trashcan', 'dry-erase_board' : 'dry_erase_board', 'black_and_white' : 'black_white', 'eiffel_tower' : 'tower' } embeds = [] for k in vocab: if k in custom_map: k = custom_map[k] if '_' in k and k not in model: ks = k.split('_') emb = np.stack([model[it] for it in ks]).mean(axis=0) else: emb = model[k] embeds.append(emb) embeds = torch.Tensor(np.stack(embeds)) print('Word2Vec Embeddings loaded, total embeddings: {}'.format(embeds.size())) return embeds def load_glove_embeddings(vocab): ''' Inputs emb_file: Text file with word embedding pairs e.g. Glove, Processed in lower case. vocab: List of words Returns Embedding Matrix ''' vocab = [v.lower() for v in vocab] emb_file = DATA_FOLDER+'/glove/glove.6B.300d.txt' model = {} # populating a dictionary of word and embeddings for line in open(emb_file, 'r'): line = line.strip().split(' ') # Word-embedding wvec = torch.FloatTensor(list(map(float, line[1:]))) model[line[0]] = wvec # Adding some vectors for UT Zappos custom_map = { 'faux.fur': 'fake_fur', 'faux.leather': 'fake_leather', 'full.grain.leather': 'thick_leather', 'hair.calf': 'hair_leather', 'patent.leather': 'shiny_leather', 'boots.ankle': 'ankle_boots', 'boots.knee.high': 'knee_high_boots', 'boots.mid-calf': 'midcalf_boots', 'shoes.boat.shoes': 'boat_shoes', 'shoes.clogs.and.mules': 'clogs_shoes', 'shoes.flats': 'flats_shoes', 'shoes.heels': 'heels', 'shoes.loafers': 'loafers', 'shoes.oxfords': 'oxford_shoes', 'shoes.sneakers.and.athletic.shoes': 'sneakers', 'traffic_light': 'traffic_light', 'trash_can': 'trashcan', 'dry-erase_board' : 'dry_erase_board', 'black_and_white' : 'black_white', 'eiffel_tower' : 'tower', 'nubuck' : 'grainy_leather', } embeds = [] for k in vocab: if k in custom_map: k = custom_map[k] if '_' in k: ks = k.split('_') emb = torch.stack([model[it] for it in ks]).mean(dim=0) else: emb = model[k] embeds.append(emb) embeds = torch.stack(embeds) print('Glove Embeddings loaded, total embeddings: {}'.format(embeds.size())) return embeds
6,636
35.26776
90
py
czsl
czsl-main/models/modular_methods.py
# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # import torch import torch.nn as nn import torch.nn.functional as F import torchvision.models as tmodels import numpy as np from .word_embedding import load_word_embeddings import itertools import math import collections from torch.distributions.bernoulli import Bernoulli import pdb import sys class Flatten(nn.Module): def forward(self, input): return input.view(input.size(0), -1) class GatingSampler(nn.Module): """Docstring for GatingSampler. """ def __init__(self, gater, stoch_sample=True, temperature=1.0): """TODO: to be defined1. """ nn.Module.__init__(self) self.gater = gater self._stoch_sample = stoch_sample self._temperature = temperature def disable_stochastic_sampling(self): self._stoch_sample = False def enable_stochastic_sampling(self): self._stoch_sample = True def forward(self, tdesc=None, return_additional=False, gating_wt=None): if self.gater is None and return_additional: return None, None elif self.gater is None: return None if gating_wt is not None: return_wts = gating_wt gating_g = self.gater(tdesc, gating_wt=gating_wt) else: gating_g = self.gater(tdesc) return_wts = None if isinstance(gating_g, tuple): return_wts = gating_g[1] gating_g = gating_g[0] if not self._stoch_sample: sampled_g = gating_g else: raise (NotImplementedError) if return_additional: return sampled_g, return_wts return sampled_g class GatedModularNet(nn.Module): """ An interface for creating modular nets """ def __init__(self, module_list, start_modules=None, end_modules=None, single_head=False, chain=False): """TODO: to be defined1. :module_list: TODO :g: TODO """ nn.Module.__init__(self) self._module_list = nn.ModuleList( [nn.ModuleList(m) for m in module_list]) self.num_layers = len(self._module_list) if start_modules is not None: self._start_modules = nn.ModuleList(start_modules) else: self._start_modules = None if end_modules is not None: self._end_modules = nn.ModuleList(end_modules) self.num_layers += 1 else: self._end_modules = None self.sampled_g = None self.single_head = single_head self._chain = chain def forward(self, x, sampled_g=None, t=None, return_feat=False): """TODO: Docstring for forward. :x: Input data :g: Gating tensor (#Task x )#num_layer x #num_mods x #num_mods :t: task ID :returns: TODO """ if t is None: t_not_set = True t = torch.tensor([0] * x.shape[0], dtype=x.dtype).long() else: t_not_set = False t = t.squeeze() if self._start_modules is not None: prev_out = [mod(x) for mod in self._start_modules] else: prev_out = [x] if sampled_g is None: # NON-Gated Module network prev_out = sum(prev_out) / float(len(prev_out)) #prev_out = torch.mean(prev_out, 0) for li in range(len(self._module_list)): prev_out = sum([ mod(prev_out) for mod in self._module_list[li] ]) / float(len(self._module_list[li])) features = prev_out if self._end_modules is not None: if t_not_set or self.single_head: prev_out = self._end_modules[0](prev_out) else: prev_out = torch.cat([ self._end_modules[tid](prev_out[bi:bi + 1]) for bi, tid in enumerate(t) ], 0) if return_feat: return prev_out, features return prev_out else: # Forward prop with sampled Gs for li in range(len(self._module_list)): curr_out = [] for j in range(len(self._module_list[li])): gind = j if not self._chain else 0 # Dim: #Batch x C module_in_wt = sampled_g[li + 1][gind] # Module input weights rearranged to match inputs module_in_wt = module_in_wt.transpose(0, 1) add_dims = prev_out[0].dim() + 1 - module_in_wt.dim() module_in_wt = module_in_wt.view(*module_in_wt.shape, *([1] * add_dims)) module_in_wt = module_in_wt.expand( len(prev_out), *prev_out[0].shape) module_in = sum([ module_in_wt[i] * prev_out[i] for i in range(len(prev_out)) ]) mod = self._module_list[li][j] curr_out.append(mod(module_in)) prev_out = curr_out # Output modules (with sampled Gs) if self._end_modules is not None: li = self.num_layers - 1 if t_not_set or self.single_head: # Dim: #Batch x C module_in_wt = sampled_g[li + 1][0] # Module input weights rearranged to match inputs module_in_wt = module_in_wt.transpose(0, 1) add_dims = prev_out[0].dim() + 1 - module_in_wt.dim() module_in_wt = module_in_wt.view(*module_in_wt.shape, *([1] * add_dims)) module_in_wt = module_in_wt.expand( len(prev_out), *prev_out[0].shape) module_in = sum([ module_in_wt[i] * prev_out[i] for i in range(len(prev_out)) ]) features = module_in prev_out = self._end_modules[0](module_in) else: curr_out = [] for bi, tid in enumerate(t): # Dim: #Batch x C gind = tid if not self._chain else 0 module_in_wt = sampled_g[li + 1][gind] # Module input weights rearranged to match inputs module_in_wt = module_in_wt.transpose(0, 1) add_dims = prev_out[0].dim() + 1 - module_in_wt.dim() module_in_wt = module_in_wt.view( *module_in_wt.shape, *([1] * add_dims)) module_in_wt = module_in_wt.expand( len(prev_out), *prev_out[0].shape) module_in = sum([ module_in_wt[i] * prev_out[i] for i in range(len(prev_out)) ]) features = module_in mod = self._end_modules[tid] curr_out.append(mod(module_in[bi:bi + 1])) prev_out = curr_out prev_out = torch.cat(prev_out, 0) if return_feat: return prev_out, features return prev_out class CompositionalModel(nn.Module): def __init__(self, dset, args): super(CompositionalModel, self).__init__() self.args = args self.dset = dset # precompute validation pairs attrs, objs = zip(*self.dset.pairs) attrs = [dset.attr2idx[attr] for attr in attrs] objs = [dset.obj2idx[obj] for obj in objs] self.val_attrs = torch.LongTensor(attrs).cuda() self.val_objs = torch.LongTensor(objs).cuda() def train_forward_softmax(self, x): img, attrs, objs = x[0], x[1], x[2] neg_attrs, neg_objs = x[4], x[5] inv_attrs, comm_attrs = x[6], x[7] sampled_attrs = torch.cat((attrs.unsqueeze(1), neg_attrs), 1) sampled_objs = torch.cat((objs.unsqueeze(1), neg_objs), 1) img_ind = torch.arange(sampled_objs.shape[0]).unsqueeze(1).repeat( 1, sampled_attrs.shape[1]) flat_sampled_attrs = sampled_attrs.view(-1) flat_sampled_objs = sampled_objs.view(-1) flat_img_ind = img_ind.view(-1) labels = torch.zeros_like(sampled_attrs[:, 0]).long() self.composed_g = self.compose(flat_sampled_attrs, flat_sampled_objs) cls_scores, feat = self.comp_network( img[flat_img_ind], self.composed_g, return_feat=True) pair_scores = cls_scores[:, :1] pair_scores = pair_scores.view(*sampled_attrs.shape) loss = 0 loss_cls = F.cross_entropy(pair_scores, labels) loss += loss_cls loss_obj = torch.FloatTensor([0]) loss_attr = torch.FloatTensor([0]) loss_sparse = torch.FloatTensor([0]) loss_unif = torch.FloatTensor([0]) loss_aux = torch.FloatTensor([0]) acc = (pair_scores.argmax(1) == labels).sum().float() / float( len(labels)) all_losses = {} all_losses['total_loss'] = loss all_losses['main_loss'] = loss_cls all_losses['aux_loss'] = loss_aux all_losses['obj_loss'] = loss_obj all_losses['attr_loss'] = loss_attr all_losses['sparse_loss'] = loss_sparse all_losses['unif_loss'] = loss_unif return loss, all_losses, acc, (pair_scores, feat) def val_forward(self, x): img = x[0] batch_size = img.shape[0] pair_scores = torch.zeros(batch_size, len(self.val_attrs)) pair_feats = torch.zeros(batch_size, len(self.val_attrs), self.args.emb_dim) pair_bs = len(self.val_attrs) for pi in range(math.ceil(len(self.val_attrs) / pair_bs)): self.compose_g = self.compose( self.val_attrs[pi * pair_bs:(pi + 1) * pair_bs], self.val_objs[pi * pair_bs:(pi + 1) * pair_bs]) compose_g = self.compose_g expanded_im = img.unsqueeze(1).repeat( 1, compose_g[0][0].shape[0], *tuple([1] * (img.dim() - 1))).view(-1, *img.shape[1:]) expanded_compose_g = [[ g.unsqueeze(0).repeat(batch_size, *tuple([1] * g.dim())).view( -1, *g.shape[1:]) for g in layer_g ] for layer_g in compose_g] this_pair_scores, this_feat = self.comp_network( expanded_im, expanded_compose_g, return_feat=True) featnorm = torch.norm(this_feat, p=2, dim=-1) this_feat = this_feat.div( featnorm.unsqueeze(-1).expand_as(this_feat)) this_pair_scores = this_pair_scores[:, :1].view(batch_size, -1) this_feat = this_feat.view(batch_size, -1, self.args.emb_dim) pair_scores[:, pi * pair_bs:pi * pair_bs + this_pair_scores.shape[1]] = this_pair_scores[:, :] pair_feats[:, pi * pair_bs:pi * pair_bs + this_pair_scores.shape[1], :] = this_feat[:] scores = {} feats = {} for i, (attr, obj) in enumerate(self.dset.pairs): scores[(attr, obj)] = pair_scores[:, i] feats[(attr, obj)] = pair_feats[:, i] # return None, (scores, feats) return None, scores def forward(self, x, with_grad=False): if self.training: loss, loss_aux, acc, pred = self.train_forward(x) else: loss_aux = torch.Tensor([0]) loss = torch.Tensor([0]) if not with_grad: with torch.no_grad(): acc, pred = self.val_forward(x) else: acc, pred = self.val_forward(x) # return loss, loss_aux, acc, pred return loss, pred class GatedGeneralNN(CompositionalModel): """Docstring for GatedCompositionalModel. """ def __init__(self, dset, args, num_layers=2, num_modules_per_layer=3, stoch_sample=False, use_full_model=False, num_classes=[2], gater_type='general'): """TODO: to be defined1. :dset: TODO :args: TODO """ CompositionalModel.__init__(self, dset, args) self.train_forward = self.train_forward_softmax self.compose_type = 'nn' #todo: could be different gating_in_dim = 128 if args.emb_init: gating_in_dim = 300 elif args.clf_init: gating_in_dim = 512 if self.compose_type == 'nn': tdim = gating_in_dim * 2 inter_tdim = self.args.embed_rank # Change this to allow only obj, only attr gatings self.attr_embedder = nn.Embedding( len(dset.attrs) + 1, gating_in_dim, padding_idx=len(dset.attrs), ) self.obj_embedder = nn.Embedding( len(dset.objs) + 1, gating_in_dim, padding_idx=len(dset.objs), ) # initialize the weights of the embedders with the svm weights if args.emb_init: pretrained_weight = load_word_embeddings( args.emb_init, dset.attrs) self.attr_embedder.weight[:-1, :].data.copy_(pretrained_weight) pretrained_weight = load_word_embeddings( args.emb_init, dset.objs) self.obj_embedder.weight.data[:-1, :].copy_(pretrained_weight) elif args.clf_init: for idx, attr in enumerate(dset.attrs): at_id = self.dset.attr2idx[attr] weight = torch.load( '%s/svm/attr_%d' % (args.data_dir, at_id)).coef_.squeeze() self.attr_embedder.weight[idx].data.copy_( torch.from_numpy(weight)) for idx, obj in enumerate(dset.objs): obj_id = self.dset.obj2idx[obj] weight = torch.load( '%s/svm/obj_%d' % (args.data_dir, obj_id)).coef_.squeeze() self.obj_embedder.weight[idx].data.copy_( torch.from_numpy(weight)) else: n_attr = len(dset.attrs) gating_in_dim = 300 tdim = gating_in_dim * 2 + n_attr self.attr_embedder = nn.Embedding( n_attr, n_attr, ) self.attr_embedder.weight.data.copy_( torch.from_numpy(np.eye(n_attr))) self.obj_embedder = nn.Embedding( len(dset.objs) + 1, gating_in_dim, padding_idx=len(dset.objs), ) pretrained_weight = load_word_embeddings( '/home/ubuntu/workspace/czsl/data/glove/glove.6B.300d.txt', dset.objs) self.obj_embedder.weight.data[:-1, :].copy_(pretrained_weight) else: raise (NotImplementedError) self.comp_network, self.gating_network, self.nummods, _ = modular_general( num_layers=num_layers, num_modules_per_layer=num_modules_per_layer, feat_dim=dset.feat_dim, inter_dim=args.emb_dim, stoch_sample=stoch_sample, use_full_model=use_full_model, tdim=tdim, inter_tdim=inter_tdim, gater_type=gater_type, ) if args.static_inp: for param in self.attr_embedder.parameters(): param.requires_grad = False for param in self.obj_embedder.parameters(): param.requires_grad = False def compose(self, attrs, objs): obj_wt = self.obj_embedder(objs) if self.compose_type == 'nn': attr_wt = self.attr_embedder(attrs) inp_wts = torch.cat([attr_wt, obj_wt], 1) # 2D else: raise (NotImplementedError) composed_g, composed_g_wt = self.gating_network( inp_wts, return_additional=True) return composed_g class GeneralNormalizedNN(nn.Module): """Docstring for GatedCompositionalModel. """ def __init__(self, num_layers, num_modules_per_layer, in_dim, inter_dim): """TODO: to be defined1. """ nn.Module.__init__(self) self.start_modules = [nn.Sequential()] self.layer1 = [[ nn.Sequential( nn.Linear(in_dim, inter_dim), nn.BatchNorm1d(inter_dim), nn.ReLU()) for _ in range(num_modules_per_layer) ]] if num_layers > 1: self.layer2 = [[ nn.Sequential( nn.BatchNorm1d(inter_dim), nn.Linear(inter_dim, inter_dim), nn.BatchNorm1d(inter_dim), nn.ReLU()) for _m in range(num_modules_per_layer) ] for _l in range(num_layers - 1)] self.avgpool = nn.Sequential() self.fc = [ nn.Sequential(nn.BatchNorm1d(inter_dim), nn.Linear(inter_dim, 1)) ] class GeneralGatingNN(nn.Module): def __init__( self, num_mods, tdim, inter_tdim, randinit=False, ): """TODO: to be defined1. :num_mods: TODO :tdim: TODO """ nn.Module.__init__(self) self._num_mods = num_mods self._tdim = tdim self._inter_tdim = inter_tdim task_outdim = self._inter_tdim self.task_linear1 = nn.Linear(self._tdim, task_outdim, bias=False) self.task_bn1 = nn.BatchNorm1d(task_outdim) self.task_linear2 = nn.Linear(task_outdim, task_outdim, bias=False) self.task_bn2 = nn.BatchNorm1d(task_outdim) self.joint_linear1 = nn.Linear(task_outdim, task_outdim, bias=False) self.joint_bn1 = nn.BatchNorm1d(task_outdim) num_out = [[1]] + [[ self._num_mods[i - 1] for _ in range(self._num_mods[i]) ] for i in range(1, len(self._num_mods))] count = 0 out_ind = [] for i in range(len(num_out)): this_out_ind = [] for j in range(len(num_out[i])): this_out_ind.append([count, count + num_out[i][j]]) count += num_out[i][j] out_ind.append(this_out_ind) self.out_ind = out_ind self.out_count = count self.joint_linear2 = nn.Linear(task_outdim, count, bias=False) def apply_init(m): if isinstance(m, nn.Linear): nn.init.normal_(m.weight, 0, 0.1) if m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.BatchNorm1d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) for m in self.modules(): if isinstance(m, nn.ModuleList): for subm in m: if isinstance(subm, nn.ModuleList): for subsubm in subm: apply_init(subsubm) else: apply_init(subm) else: apply_init(m) if not randinit: self.joint_linear2.weight.data.zero_() def forward(self, tdesc=None): """TODO: Docstring for function. :arg1: TODO :returns: TODO """ if tdesc is None: return None x = tdesc task_embeds1 = F.relu(self.task_bn1(self.task_linear1(x))) joint_embed = task_embeds1 joint_embed = self.joint_linear2(joint_embed) joint_embed = [[ joint_embed[:, self.out_ind[i][j][0]:self.out_ind[i][j][1]] for j in range(len(self.out_ind[i])) ] for i in range(len(self.out_ind))] gating_wt = joint_embed prob_g = [[F.softmax(wt, -1) for wt in gating_wt[i]] for i in range(len(gating_wt))] return prob_g, gating_wt def modularize_network( model, stoch_sample=False, use_full_model=False, tdim=200, inter_tdim=200, gater_type='general', single_head=True, num_classes=[2], num_lookup_gating=10, ): # Copy start modules and end modules start_modules = model.start_modules end_modules = [ nn.Sequential(model.avgpool, Flatten(), fci) for fci in model.fc ] # Create module_list as list of lists [[layer1 modules], [layer2 modules], ...] module_list = [] li = 1 while True: if hasattr(model, 'layer{}'.format(li)): module_list.extend(getattr(model, 'layer{}'.format(li))) li += 1 else: break num_module_list = [len(start_modules)] + [len(layer) for layer in module_list] \ + [len(end_modules)] gated_model_func = GatedModularNet gated_net = gated_model_func( module_list, start_modules=start_modules, end_modules=end_modules, single_head=single_head) gater_func = GeneralGatingNN gater = gater_func( num_mods=num_module_list, tdim=tdim, inter_tdim=inter_tdim) fan_in = num_module_list if use_full_model: gater = None # Create Gating Sampler gating_sampler = GatingSampler(gater=gater, stoch_sample=stoch_sample) return gated_net, gating_sampler, num_module_list, fan_in def modular_general( num_layers, num_modules_per_layer, feat_dim, inter_dim, stoch_sample=False, use_full_model=False, num_classes=[2], single_head=True, gater_type='general', tdim=300, inter_tdim=300, num_lookup_gating=10, ): # First create a ResNext model model = GeneralNormalizedNN( num_layers, num_modules_per_layer, feat_dim, inter_dim, ) # Modularize the model and create gating funcs gated_net, gating_sampler, num_module_list, fan_in = modularize_network( model, stoch_sample, use_full_model=use_full_model, gater_type=gater_type, tdim=tdim, inter_tdim=inter_tdim, single_head=single_head, num_classes=num_classes, num_lookup_gating=num_lookup_gating, ) return gated_net, gating_sampler, num_module_list, fan_in
23,151
34.454824
90
py
czsl
czsl-main/models/compcos.py
import torch import torch.nn as nn import torch.nn.functional as F from .word_embedding import load_word_embeddings from .common import MLP from itertools import product device = 'cuda' if torch.cuda.is_available() else 'cpu' def compute_cosine_similarity(names, weights, return_dict=True): pairing_names = list(product(names, names)) normed_weights = F.normalize(weights,dim=1) similarity = torch.mm(normed_weights, normed_weights.t()) if return_dict: dict_sim = {} for i,n in enumerate(names): for j,m in enumerate(names): dict_sim[(n,m)]=similarity[i,j].item() return dict_sim return pairing_names, similarity.to('cpu') class CompCos(nn.Module): def __init__(self, dset, args): super(CompCos, self).__init__() self.args = args self.dset = dset def get_all_ids(relevant_pairs): # Precompute validation pairs attrs, objs = zip(*relevant_pairs) attrs = [dset.attr2idx[attr] for attr in attrs] objs = [dset.obj2idx[obj] for obj in objs] pairs = [a for a in range(len(relevant_pairs))] attrs = torch.LongTensor(attrs).to(device) objs = torch.LongTensor(objs).to(device) pairs = torch.LongTensor(pairs).to(device) return attrs, objs, pairs # Validation self.val_attrs, self.val_objs, self.val_pairs = get_all_ids(self.dset.pairs) # for indivual projections self.uniq_attrs, self.uniq_objs = torch.arange(len(self.dset.attrs)).long().to(device), \ torch.arange(len(self.dset.objs)).long().to(device) self.factor = 2 self.scale = self.args.cosine_scale if dset.open_world: self.train_forward = self.train_forward_open self.known_pairs = dset.train_pairs seen_pair_set = set(self.known_pairs) mask = [1 if pair in seen_pair_set else 0 for pair in dset.pairs] self.seen_mask = torch.BoolTensor(mask).to(device) * 1. self.activated = False # Init feasibility-related variables self.attrs = dset.attrs self.objs = dset.objs self.possible_pairs = dset.pairs self.validation_pairs = dset.val_pairs self.feasibility_margin = (1-self.seen_mask).float() self.epoch_max_margin = self.args.epoch_max_margin self.cosine_margin_factor = -args.margin # Intantiate attribut-object relations, needed just to evaluate mined pairs self.obj_by_attrs_train = {k: [] for k in self.attrs} for (a, o) in self.known_pairs: self.obj_by_attrs_train[a].append(o) # Intantiate attribut-object relations, needed just to evaluate mined pairs self.attrs_by_obj_train = {k: [] for k in self.objs} for (a, o) in self.known_pairs: self.attrs_by_obj_train[o].append(a) else: self.train_forward = self.train_forward_closed # Precompute training compositions if args.train_only: self.train_attrs, self.train_objs, self.train_pairs = get_all_ids(self.dset.train_pairs) else: self.train_attrs, self.train_objs, self.train_pairs = self.val_attrs, self.val_objs, self.val_pairs try: self.args.fc_emb = self.args.fc_emb.split(',') except: self.args.fc_emb = [self.args.fc_emb] layers = [] for a in self.args.fc_emb: a = int(a) layers.append(a) self.image_embedder = MLP(dset.feat_dim, int(args.emb_dim), relu=args.relu, num_layers=args.nlayers, dropout=self.args.dropout, norm=self.args.norm, layers=layers) # Fixed self.composition = args.composition input_dim = args.emb_dim self.attr_embedder = nn.Embedding(len(dset.attrs), input_dim) self.obj_embedder = nn.Embedding(len(dset.objs), input_dim) # init with word embeddings if args.emb_init: pretrained_weight = load_word_embeddings(args.emb_init, dset.attrs) self.attr_embedder.weight.data.copy_(pretrained_weight) pretrained_weight = load_word_embeddings(args.emb_init, dset.objs) self.obj_embedder.weight.data.copy_(pretrained_weight) # static inputs if args.static_inp: for param in self.attr_embedder.parameters(): param.requires_grad = False for param in self.obj_embedder.parameters(): param.requires_grad = False # Composition MLP self.projection = nn.Linear(input_dim * 2, args.emb_dim) def freeze_representations(self): print('Freezing representations') for param in self.image_embedder.parameters(): param.requires_grad = False for param in self.attr_embedder.parameters(): param.requires_grad = False for param in self.obj_embedder.parameters(): param.requires_grad = False def compose(self, attrs, objs): attrs, objs = self.attr_embedder(attrs), self.obj_embedder(objs) inputs = torch.cat([attrs, objs], 1) output = self.projection(inputs) output = F.normalize(output, dim=1) return output def compute_feasibility(self): obj_embeddings = self.obj_embedder(torch.arange(len(self.objs)).long().to('cuda')) obj_embedding_sim = compute_cosine_similarity(self.objs, obj_embeddings, return_dict=True) attr_embeddings = self.attr_embedder(torch.arange(len(self.attrs)).long().to('cuda')) attr_embedding_sim = compute_cosine_similarity(self.attrs, attr_embeddings, return_dict=True) feasibility_scores = self.seen_mask.clone().float() for a in self.attrs: for o in self.objs: if (a, o) not in self.known_pairs: idx = self.dset.all_pair2idx[(a, o)] score_obj = self.get_pair_scores_objs(a, o, obj_embedding_sim) score_attr = self.get_pair_scores_attrs(a, o, attr_embedding_sim) score = (score_obj + score_attr) / 2 feasibility_scores[idx] = score self.feasibility_scores = feasibility_scores return feasibility_scores * (1 - self.seen_mask.float()) def get_pair_scores_objs(self, attr, obj, obj_embedding_sim): score = -1. for o in self.objs: if o!=obj and attr in self.attrs_by_obj_train[o]: temp_score = obj_embedding_sim[(obj,o)] if temp_score>score: score=temp_score return score def get_pair_scores_attrs(self, attr, obj, attr_embedding_sim): score = -1. for a in self.attrs: if a != attr and obj in self.obj_by_attrs_train[a]: temp_score = attr_embedding_sim[(attr, a)] if temp_score > score: score = temp_score return score def update_feasibility(self,epoch): self.activated = True feasibility_scores = self.compute_feasibility() self.feasibility_margin = min(1.,epoch/self.epoch_max_margin) * \ (self.cosine_margin_factor*feasibility_scores.float().to(device)) def val_forward(self, x): img = x[0] img_feats = self.image_embedder(img) img_feats_normed = F.normalize(img_feats, dim=1) pair_embeds = self.compose(self.val_attrs, self.val_objs).permute(1, 0) # Evaluate all pairs score = torch.matmul(img_feats_normed, pair_embeds) scores = {} for itr, pair in enumerate(self.dset.pairs): scores[pair] = score[:, self.dset.all_pair2idx[pair]] return None, scores def val_forward_with_threshold(self, x, th=0.): img = x[0] img_feats = self.image_embedder(img) img_feats_normed = F.normalize(img_feats, dim=1) pair_embeds = self.compose(self.val_attrs, self.val_objs).permute(1, 0) # Evaluate all pairs score = torch.matmul(img_feats_normed, pair_embeds) # Note: Pairs are already aligned here mask = (self.feasibility_scores>=th).float() score = score*mask + (1.-mask)*(-1.) scores = {} for itr, pair in enumerate(self.dset.pairs): scores[pair] = score[:, self.dset.all_pair2idx[pair]] return None, scores def train_forward_open(self, x): img, attrs, objs, pairs = x[0], x[1], x[2], x[3] img_feats = self.image_embedder(img) pair_embed = self.compose(self.train_attrs, self.train_objs).permute(1, 0) img_feats_normed = F.normalize(img_feats, dim=1) pair_pred = torch.matmul(img_feats_normed, pair_embed) if self.activated: pair_pred += (1 - self.seen_mask) * self.feasibility_margin loss_cos = F.cross_entropy(self.scale * pair_pred, pairs) else: pair_pred = pair_pred * self.seen_mask + (1 - self.seen_mask) * (-10) loss_cos = F.cross_entropy(self.scale * pair_pred, pairs) return loss_cos.mean(), None def train_forward_closed(self, x): img, attrs, objs, pairs = x[0], x[1], x[2], x[3] img_feats = self.image_embedder(img) pair_embed = self.compose(self.train_attrs, self.train_objs).permute(1, 0) img_feats_normed = F.normalize(img_feats, dim=1) pair_pred = torch.matmul(img_feats_normed, pair_embed) loss_cos = F.cross_entropy(self.scale * pair_pred, pairs) return loss_cos.mean(), None def forward(self, x): if self.training: loss, pred = self.train_forward(x) else: with torch.no_grad(): loss, pred = self.val_forward(x) return loss, pred
10,133
36.533333
111
py
czsl
czsl-main/models/graph_method.py
import torch import torch.nn as nn import torch.nn.functional as F import numpy as np from .common import MLP from .gcn import GCN, GCNII from .word_embedding import load_word_embeddings import scipy.sparse as sp def adj_to_edges(adj): # Adj sparse matrix to list of edges rows, cols = np.nonzero(adj) edges = list(zip(rows.tolist(), cols.tolist())) return edges def edges_to_adj(edges, n): # List of edges to Adj sparse matrix edges = np.array(edges) adj = sp.coo_matrix((np.ones(len(edges)), (edges[:, 0], edges[:, 1])), shape=(n, n), dtype='float32') return adj device = 'cuda' if torch.cuda.is_available() else 'cpu' class GraphFull(nn.Module): def __init__(self, dset, args): super(GraphFull, self).__init__() self.args = args self.dset = dset self.val_forward = self.val_forward_dotpr self.train_forward = self.train_forward_normal # Image Embedder self.num_attrs, self.num_objs, self.num_pairs = len(dset.attrs), len(dset.objs), len(dset.pairs) self.pairs = dset.pairs if self.args.train_only: train_idx = [] for current in dset.train_pairs: train_idx.append(dset.all_pair2idx[current]+self.num_attrs+self.num_objs) self.train_idx = torch.LongTensor(train_idx).to(device) self.args.fc_emb = self.args.fc_emb.split(',') layers = [] for a in self.args.fc_emb: a = int(a) layers.append(a) if args.nlayers: self.image_embedder = MLP(dset.feat_dim, args.emb_dim, num_layers= args.nlayers, dropout = self.args.dropout, norm = self.args.norm, layers = layers, relu = True) all_words = list(self.dset.attrs) + list(self.dset.objs) self.displacement = len(all_words) self.obj_to_idx = {word: idx for idx, word in enumerate(self.dset.objs)} self.attr_to_idx = {word: idx for idx, word in enumerate(self.dset.attrs)} if args.graph_init is not None: path = args.graph_init graph = torch.load(path) embeddings = graph['embeddings'].to(device) adj = graph['adj'] self.embeddings = embeddings else: embeddings = self.init_embeddings(all_words).to(device) adj = self.adj_from_pairs() self.embeddings = embeddings hidden_layers = self.args.gr_emb if args.gcn_type == 'gcn': self.gcn = GCN(adj, self.embeddings.shape[1], args.emb_dim, hidden_layers) else: self.gcn = GCNII(adj, self.embeddings.shape[1], args.emb_dim, args.hidden_dim, args.gcn_nlayers, lamda = 0.5, alpha = 0.1, variant = False) def init_embeddings(self, all_words): def get_compositional_embeddings(embeddings, pairs): # Getting compositional embeddings from base embeddings composition_embeds = [] for (attr, obj) in pairs: attr_embed = embeddings[self.attr_to_idx[attr]] obj_embed = embeddings[self.obj_to_idx[obj]+self.num_attrs] composed_embed = (attr_embed + obj_embed) / 2 composition_embeds.append(composed_embed) composition_embeds = torch.stack(composition_embeds) print('Compositional Embeddings are ', composition_embeds.shape) return composition_embeds # init with word embeddings embeddings = load_word_embeddings(self.args.emb_init, all_words) composition_embeds = get_compositional_embeddings(embeddings, self.pairs) full_embeddings = torch.cat([embeddings, composition_embeds], dim=0) return full_embeddings def update_dict(self, wdict, row,col,data): wdict['row'].append(row) wdict['col'].append(col) wdict['data'].append(data) def adj_from_pairs(self): def edges_from_pairs(pairs): weight_dict = {'data':[],'row':[],'col':[]} for i in range(self.displacement): self.update_dict(weight_dict,i,i,1.) for idx, (attr, obj) in enumerate(pairs): attr_idx, obj_idx = self.attr_to_idx[attr], self.obj_to_idx[obj] + self.num_attrs self.update_dict(weight_dict, attr_idx, obj_idx, 1.) self.update_dict(weight_dict, obj_idx, attr_idx, 1.) node_id = idx + self.displacement self.update_dict(weight_dict,node_id,node_id,1.) self.update_dict(weight_dict, node_id, attr_idx, 1.) self.update_dict(weight_dict, node_id, obj_idx, 1.) self.update_dict(weight_dict, attr_idx, node_id, 1.) self.update_dict(weight_dict, obj_idx, node_id, 1.) return weight_dict edges = edges_from_pairs(self.pairs) adj = sp.csr_matrix((edges['data'], (edges['row'], edges['col'])), shape=(len(self.pairs)+self.displacement, len(self.pairs)+self.displacement)) return adj def train_forward_normal(self, x): img, attrs, objs, pairs = x[0], x[1], x[2], x[3] if self.args.nlayers: img_feats = self.image_embedder(img) else: img_feats = (img) current_embeddings = self.gcn(self.embeddings) if self.args.train_only: pair_embed = current_embeddings[self.train_idx] else: pair_embed = current_embeddings[self.num_attrs+self.num_objs:self.num_attrs+self.num_objs+self.num_pairs,:] pair_embed = pair_embed.permute(1,0) pair_pred = torch.matmul(img_feats, pair_embed) loss = F.cross_entropy(pair_pred, pairs) return loss, None def val_forward_dotpr(self, x): img = x[0] if self.args.nlayers: img_feats = self.image_embedder(img) else: img_feats = (img) current_embedddings = self.gcn(self.embeddings) pair_embeds = current_embedddings[self.num_attrs+self.num_objs:self.num_attrs+self.num_objs+self.num_pairs,:].permute(1,0) score = torch.matmul(img_feats, pair_embeds) scores = {} for itr, pair in enumerate(self.dset.pairs): scores[pair] = score[:,self.dset.all_pair2idx[pair]] return None, scores def val_forward_distance_fast(self, x): img = x[0] img_feats = (self.image_embedder(img)) current_embeddings = self.gcn(self.embeddings) pair_embeds = current_embeddings[self.num_attrs+self.num_objs:,:] batch_size, pairs, features = img_feats.shape[0], pair_embeds.shape[0], pair_embeds.shape[1] img_feats = img_feats[:,None,:].expand(-1, pairs, -1) pair_embeds = pair_embeds[None,:,:].expand(batch_size, -1, -1) diff = (img_feats - pair_embeds)**2 score = diff.sum(2) * -1 scores = {} for itr, pair in enumerate(self.dset.pairs): scores[pair] = score[:,self.dset.all_pair2idx[pair]] return None, scores def forward(self, x): if self.training: loss, pred = self.train_forward(x) else: with torch.no_grad(): loss, pred = self.val_forward(x) return loss, pred
7,321
33.701422
151
py
czsl
czsl-main/models/image_extractor.py
import torch import torch.nn as nn import torch.nn.functional as F from torchvision import models from torchvision.models.resnet import ResNet, BasicBlock class ResNet18_conv(ResNet): def __init__(self): super(ResNet18_conv, self).__init__(BasicBlock, [2, 2, 2, 2]) def forward(self, x): # change forward here x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) return x def get_image_extractor(arch = 'resnet18', pretrained = True, feature_dim = None, checkpoint = ''): ''' Inputs arch: Base architecture pretrained: Bool, Imagenet weights feature_dim: Int, output feature dimension checkpoint: String, not implemented Returns Pytorch model ''' if arch == 'resnet18': model = models.resnet18(pretrained = pretrained) if feature_dim is None: model.fc = nn.Sequential() else: model.fc = nn.Linear(512, feature_dim) if arch == 'resnet18_conv': model = ResNet18_conv() model.load_state_dict(models.resnet18(pretrained=True).state_dict()) elif arch == 'resnet50': model = models.resnet50(pretrained = pretrained) if feature_dim is None: model.fc = nn.Sequential() else: model.fc = nn.Linear(2048, feature_dim) elif arch == 'resnet50_cutmix': model = models.resnet50(pretrained = pretrained) checkpoint = torch.load('/home/ubuntu/workspace/pretrained/resnet50_cutmix.tar') model.load_state_dict(checkpoint['state_dict'], strict=False) if feature_dim is None: model.fc = nn.Sequential() else: model.fc = nn.Linear(2048, feature_dim) elif arch == 'resnet152': model = models.resnet152(pretrained = pretrained) if feature_dim is None: model.fc = nn.Sequential() else: model.fc = nn.Linear(2048, feature_dim) elif arch == 'vgg16': model = models.vgg16(pretrained = pretrained) modules = list(model.classifier.children())[:-3] model.classifier=torch.nn.Sequential(*modules) if feature_dim is not None: model.classifier[3]=torch.nn.Linear(4096,feature_dim) return model
2,428
29.3625
99
py
czsl
czsl-main/utils/reorganize_utzap.py
# Copyright (c) Facebook, Inc. and its affiliates. # All rights reserved. # # This source code is licensed under the license found in the # LICENSE file in the root directory of this source tree. # """ Reorganize the UT-Zappos dataset to resemble the MIT-States dataset root/attr_obj/img1.jpg root/attr_obj/img2.jpg root/attr_obj/img3.jpg ... """ import os import torch import shutil import tqdm DATA_FOLDER= "ROOT_FOLDER" root = DATA_FOLDER+'/ut-zap50k/' os.makedirs(root+'/images',exist_ok=True) data = torch.load(root+'/metadata_compositional-split-natural.t7') for instance in tqdm.tqdm(data): image, attr, obj = instance['_image'], instance['attr'], instance['obj'] old_file = '%s/_images/%s'%(root, image) new_dir = '%s/images/%s_%s/'%(root, attr, obj) os.makedirs(new_dir, exist_ok=True) shutil.copy(old_file, new_dir)
835
25.125
73
py
czsl
czsl-main/utils/config_model.py
import torch import torch.optim as optim from models.image_extractor import get_image_extractor from models.visual_product import VisualProductNN from models.manifold_methods import RedWine, LabelEmbedPlus, AttributeOperator from models.modular_methods import GatedGeneralNN from models.graph_method import GraphFull from models.symnet import Symnet from models.compcos import CompCos device = 'cuda' if torch.cuda.is_available() else 'cpu' def configure_model(args, dataset): image_extractor = None is_open = False if args.model == 'visprodNN': model = VisualProductNN(dataset, args) elif args.model == 'redwine': model = RedWine(dataset, args) elif args.model == 'labelembed+': model = LabelEmbedPlus(dataset, args) elif args.model == 'attributeop': model = AttributeOperator(dataset, args) elif args.model == 'tmn': model = GatedGeneralNN(dataset, args, num_layers=args.nlayers, num_modules_per_layer=args.nmods) elif args.model == 'symnet': model = Symnet(dataset, args) elif args.model == 'graphfull': model = GraphFull(dataset, args) elif args.model == 'compcos': model = CompCos(dataset, args) if dataset.open_world and not args.train_only: is_open = True else: raise NotImplementedError model = model.to(device) if args.update_features: print('Learnable image_embeddings') image_extractor = get_image_extractor(arch = args.image_extractor, pretrained = True) image_extractor = image_extractor.to(device) # configuring optimizer if args.model=='redwine': optim_params = filter(lambda p: p.requires_grad, model.parameters()) elif args.model=='attributeop': attr_params = [param for name, param in model.named_parameters() if 'attr_op' in name and param.requires_grad] other_params = [param for name, param in model.named_parameters() if 'attr_op' not in name and param.requires_grad] optim_params = [{'params':attr_params, 'lr':0.1*args.lr}, {'params':other_params}] elif args.model=='tmn': gating_params = [ param for name, param in model.named_parameters() if 'gating_network' in name and param.requires_grad ] network_params = [ param for name, param in model.named_parameters() if 'gating_network' not in name and param.requires_grad ] optim_params = [ { 'params': network_params, }, { 'params': gating_params, 'lr': args.lrg }, ] else: model_params = [param for name, param in model.named_parameters() if param.requires_grad] optim_params = [{'params':model_params}] if args.update_features: ie_parameters = [param for name, param in image_extractor.named_parameters()] optim_params.append({'params': ie_parameters, 'lr': args.lrg}) optimizer = optim.Adam(optim_params, lr=args.lr, weight_decay=args.wd) model.is_open = is_open return image_extractor, model, optimizer
3,177
37.289157
123
py
czsl
czsl-main/utils/utils.py
import os from os.path import join as ospj import torch import random import copy import shutil import sys import yaml def chunks(l, n): """Yield successive n-sized chunks from l.""" for i in range(0, len(l), n): yield l[i:i + n] def get_norm_values(norm_family = 'imagenet'): ''' Inputs norm_family: String of norm_family Returns mean, std : tuple of 3 channel values ''' if norm_family == 'imagenet': mean, std = [0.485, 0.456, 0.406], [0.229, 0.224, 0.225] else: raise ValueError('Incorrect normalization family') return mean, std def save_args(args, log_path, argfile): shutil.copy('train.py', log_path) modelfiles = ospj(log_path, 'models') try: shutil.copy(argfile, log_path) except: print('Config exists') try: shutil.copytree('models/', modelfiles) except: print('Already exists') with open(ospj(log_path,'args_all.yaml'),'w') as f: yaml.dump(args, f, default_flow_style=False, allow_unicode=True) with open(ospj(log_path, 'args.txt'), 'w') as f: f.write('\n'.join(sys.argv[1:])) class UnNormalizer: ''' Unnormalize a given tensor using mean and std of a dataset family Inputs norm_family: String, dataset tensor: Torch tensor Outputs tensor: Unnormalized tensor ''' def __init__(self, norm_family = 'imagenet'): self.mean, self.std = get_norm_values(norm_family=norm_family) self.mean, self.std = torch.Tensor(self.mean).view(1, 3, 1, 1), torch.Tensor(self.std).view(1, 3, 1, 1) def __call__(self, tensor): return (tensor * self.std) + self.mean def load_args(filename, args): with open(filename, 'r') as stream: data_loaded = yaml.safe_load(stream) for key, group in data_loaded.items(): for key, val in group.items(): setattr(args, key, val)
1,943
28.454545
111
py
czsl
czsl-main/data/dataset.py
#external libs import numpy as np from tqdm import tqdm from PIL import Image import os import random from os.path import join as ospj from glob import glob #torch libs from torch.utils.data import Dataset import torch import torchvision.transforms as transforms #local libs from utils.utils import get_norm_values, chunks from models.image_extractor import get_image_extractor from itertools import product device = 'cuda' if torch.cuda.is_available() else 'cpu' class ImageLoader: def __init__(self, root): self.root_dir = root def __call__(self, img): img = Image.open(ospj(self.root_dir,img)).convert('RGB') #We don't want alpha return img def dataset_transform(phase, norm_family = 'imagenet'): ''' Inputs phase: String controlling which set of transforms to use norm_family: String controlling which normaliztion values to use Returns transform: A list of pytorch transforms ''' mean, std = get_norm_values(norm_family=norm_family) if phase == 'train': transform = transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean, std) ]) elif phase == 'val' or phase == 'test': transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean, std) ]) elif phase == 'all': transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize(mean, std) ]) else: raise ValueError('Invalid transform') return transform def filter_data(all_data, pairs_gt, topk = 5): ''' Helper function to clean data ''' valid_files = [] with open('/home/ubuntu/workspace/top'+str(topk)+'.txt') as f: for line in f: valid_files.append(line.strip()) data, pairs, attr, obj = [], [], [], [] for current in all_data: if current[0] in valid_files: data.append(current) pairs.append((current[1],current[2])) attr.append(current[1]) obj.append(current[2]) counter = 0 for current in pairs_gt: if current in pairs: counter+=1 print('Matches ', counter, ' out of ', len(pairs_gt)) print('Samples ', len(data), ' out of ', len(all_data)) return data, sorted(list(set(pairs))), sorted(list(set(attr))), sorted(list(set(obj))) # Dataset class now class CompositionDataset(Dataset): ''' Inputs root: String of base dir of dataset phase: String train, val, test split: String dataset split subset: Boolean if true uses a subset of train at each epoch num_negs: Int, numbers of negative pairs per batch pair_dropout: Percentage of pairs to leave in current epoch ''' def __init__( self, root, phase, split = 'compositional-split', model = 'resnet18', norm_family = 'imagenet', subset = False, num_negs = 1, pair_dropout = 0.0, update_features = False, return_images = False, train_only = False, open_world=False ): self.root = root self.phase = phase self.split = split self.num_negs = num_negs self.pair_dropout = pair_dropout self.norm_family = norm_family self.return_images = return_images self.update_features = update_features self.feat_dim = 512 if 'resnet18' in model else 2048 # todo, unify this with models self.open_world = open_world self.attrs, self.objs, self.pairs, self.train_pairs, \ self.val_pairs, self.test_pairs = self.parse_split() self.train_data, self.val_data, self.test_data = self.get_split_info() self.full_pairs = list(product(self.attrs,self.objs)) # Clean only was here self.obj2idx = {obj: idx for idx, obj in enumerate(self.objs)} self.attr2idx = {attr : idx for idx, attr in enumerate(self.attrs)} if self.open_world: self.pairs = self.full_pairs self.all_pair2idx = {pair: idx for idx, pair in enumerate(self.pairs)} if train_only and self.phase == 'train': print('Using only train pairs') self.pair2idx = {pair : idx for idx, pair in enumerate(self.train_pairs)} else: print('Using all pairs') self.pair2idx = {pair : idx for idx, pair in enumerate(self.pairs)} if self.phase == 'train': self.data = self.train_data elif self.phase == 'val': self.data = self.val_data elif self.phase == 'test': self.data = self.test_data elif self.phase == 'all': print('Using all data') self.data = self.train_data + self.val_data + self.test_data else: raise ValueError('Invalid training phase') self.all_data = self.train_data + self.val_data + self.test_data print('Dataset loaded') print('Train pairs: {}, Validation pairs: {}, Test Pairs: {}'.format( len(self.train_pairs), len(self.val_pairs), len(self.test_pairs))) print('Train images: {}, Validation images: {}, Test images: {}'.format( len(self.train_data), len(self.val_data), len(self.test_data))) if subset: ind = np.arange(len(self.data)) ind = ind[::len(ind) // 1000] self.data = [self.data[i] for i in ind] # Keeping a list of all pairs that occur with each object self.obj_affordance = {} self.train_obj_affordance = {} for _obj in self.objs: candidates = [attr for (_, attr, obj) in self.train_data+self.test_data if obj==_obj] self.obj_affordance[_obj] = list(set(candidates)) candidates = [attr for (_, attr, obj) in self.train_data if obj==_obj] self.train_obj_affordance[_obj] = list(set(candidates)) self.sample_indices = list(range(len(self.data))) self.sample_pairs = self.train_pairs # Load based on what to output self.transform = dataset_transform(self.phase, self.norm_family) self.loader = ImageLoader(ospj(self.root, 'images')) if not self.update_features: feat_file = ospj(root, model+'_featurers.t7') print(f'Using {model} and feature file {feat_file}') if not os.path.exists(feat_file): with torch.no_grad(): self.generate_features(feat_file, model) self.phase = phase activation_data = torch.load(feat_file) self.activations = dict( zip(activation_data['files'], activation_data['features'])) self.feat_dim = activation_data['features'].size(1) print('{} activations loaded'.format(len(self.activations))) def parse_split(self): ''' Helper function to read splits of object atrribute pair Returns all_attrs: List of all attributes all_objs: List of all objects all_pairs: List of all combination of attrs and objs tr_pairs: List of train pairs of attrs and objs vl_pairs: List of validation pairs of attrs and objs ts_pairs: List of test pairs of attrs and objs ''' def parse_pairs(pair_list): ''' Helper function to parse each phase to object attrribute vectors Inputs pair_list: path to textfile ''' with open(pair_list, 'r') as f: pairs = f.read().strip().split('\n') pairs = [line.split() for line in pairs] pairs = list(map(tuple, pairs)) attrs, objs = zip(*pairs) return attrs, objs, pairs tr_attrs, tr_objs, tr_pairs = parse_pairs( ospj(self.root, self.split, 'train_pairs.txt') ) vl_attrs, vl_objs, vl_pairs = parse_pairs( ospj(self.root, self.split, 'val_pairs.txt') ) ts_attrs, ts_objs, ts_pairs = parse_pairs( ospj(self.root, self.split, 'test_pairs.txt') ) #now we compose all objs, attrs and pairs all_attrs, all_objs = sorted( list(set(tr_attrs + vl_attrs + ts_attrs))), sorted( list(set(tr_objs + vl_objs + ts_objs))) all_pairs = sorted(list(set(tr_pairs + vl_pairs + ts_pairs))) return all_attrs, all_objs, all_pairs, tr_pairs, vl_pairs, ts_pairs def get_split_info(self): ''' Helper method to read image, attrs, objs samples Returns train_data, val_data, test_data: List of tuple of image, attrs, obj ''' data = torch.load(ospj(self.root, 'metadata_{}.t7'.format(self.split))) train_data, val_data, test_data = [], [], [] for instance in data: image, attr, obj, settype = instance['image'], instance['attr'], \ instance['obj'], instance['set'] curr_data = [image, attr, obj] if attr == 'NA' or (attr, obj) not in self.pairs or settype == 'NA': # Skip incomplete pairs, unknown pairs and unknown set continue if settype == 'train': train_data.append(curr_data) elif settype == 'val': val_data.append(curr_data) else: test_data.append(curr_data) return train_data, val_data, test_data def get_dict_data(self, data, pairs): data_dict = {} for current in pairs: data_dict[current] = [] for current in data: image, attr, obj = current data_dict[(attr, obj)].append(image) return data_dict def reset_dropout(self): ''' Helper function to sample new subset of data containing a subset of pairs of objs and attrs ''' self.sample_indices = list(range(len(self.data))) self.sample_pairs = self.train_pairs # Using sampling from random instead of 2 step numpy n_pairs = int((1 - self.pair_dropout) * len(self.train_pairs)) self.sample_pairs = random.sample(self.train_pairs, n_pairs) print('Sampled new subset') print('Using {} pairs out of {} pairs right now'.format( n_pairs, len(self.train_pairs))) self.sample_indices = [ i for i in range(len(self.data)) if (self.data[i][1], self.data[i][2]) in self.sample_pairs ] print('Using {} images out of {} images right now'.format( len(self.sample_indices), len(self.data))) def sample_negative(self, attr, obj): ''' Inputs attr: String of valid attribute obj: String of valid object Returns Tuple of a different attribute, object indexes ''' new_attr, new_obj = self.sample_pairs[np.random.choice( len(self.sample_pairs))] while new_attr == attr and new_obj == obj: new_attr, new_obj = self.sample_pairs[np.random.choice( len(self.sample_pairs))] return (self.attr2idx[new_attr], self.obj2idx[new_obj]) def sample_affordance(self, attr, obj): ''' Inputs attr: String of valid attribute obj: String of valid object Return Idx of a different attribute for the same object ''' new_attr = np.random.choice(self.obj_affordance[obj]) while new_attr == attr: new_attr = np.random.choice(self.obj_affordance[obj]) return self.attr2idx[new_attr] def sample_train_affordance(self, attr, obj): ''' Inputs attr: String of valid attribute obj: String of valid object Return Idx of a different attribute for the same object from the training pairs ''' new_attr = np.random.choice(self.train_obj_affordance[obj]) while new_attr == attr: new_attr = np.random.choice(self.train_obj_affordance[obj]) return self.attr2idx[new_attr] def generate_features(self, out_file, model): ''' Inputs out_file: Path to save features model: String of extraction model ''' # data = self.all_data data = ospj(self.root,'images') files_before = glob(ospj(data, '**', '*.jpg'), recursive=True) files_all = [] for current in files_before: parts = current.split('/') if "cgqa" in self.root: files_all.append(parts[-1]) else: files_all.append(os.path.join(parts[-2],parts[-1])) transform = dataset_transform('test', self.norm_family) feat_extractor = get_image_extractor(arch = model).eval() feat_extractor = feat_extractor.to(device) image_feats = [] image_files = [] for chunk in tqdm( chunks(files_all, 512), total=len(files_all) // 512, desc=f'Extracting features {model}'): files = chunk imgs = list(map(self.loader, files)) imgs = list(map(transform, imgs)) feats = feat_extractor(torch.stack(imgs, 0).to(device)) image_feats.append(feats.data.cpu()) image_files += files image_feats = torch.cat(image_feats, 0) print('features for %d images generated' % (len(image_files))) torch.save({'features': image_feats, 'files': image_files}, out_file) def __getitem__(self, index): ''' Call for getting samples ''' index = self.sample_indices[index] image, attr, obj = self.data[index] # Decide what to output if not self.update_features: img = self.activations[image] else: img = self.loader(image) img = self.transform(img) data = [img, self.attr2idx[attr], self.obj2idx[obj], self.pair2idx[(attr, obj)]] if self.phase == 'train': all_neg_attrs = [] all_neg_objs = [] for curr in range(self.num_negs): neg_attr, neg_obj = self.sample_negative(attr, obj) # negative for triplet lose all_neg_attrs.append(neg_attr) all_neg_objs.append(neg_obj) neg_attr, neg_obj = torch.LongTensor(all_neg_attrs), torch.LongTensor(all_neg_objs) #note here if len(self.train_obj_affordance[obj])>1: inv_attr = self.sample_train_affordance(attr, obj) # attribute for inverse regularizer else: inv_attr = (all_neg_attrs[0]) comm_attr = self.sample_affordance(inv_attr, obj) # attribute for commutative regularizer data += [neg_attr, neg_obj, inv_attr, comm_attr] # Return image paths if requested as the last element of the list if self.return_images and self.phase != 'train': data.append(image) return data def __len__(self): ''' Call for length ''' return len(self.sample_indices)
15,653
34.336343
106
py
MotifClass
MotifClass-master/text_classification/main.py
# The code structure is adapted from the WeSTClass implementation # https://github.com/yumeng5/WeSTClass import os os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = "0" import numpy as np np.random.seed(1234) from time import time from model import WSTC, f1 from keras.optimizers import SGD from gen import augment, pseudodocs from load_data import load_dataset from gensim.models import word2vec from gensim.models import KeyedVectors from sklearn import preprocessing def load_embedding(vocabulary_inv, num_class, dataset_name, embedding_name): model_name = f'{dataset_name}/embedding_{embedding_name}' if os.path.exists(model_name): # embedding_model = word2vec.Word2Vec.load(model_name) embedding_model = KeyedVectors.load_word2vec_format(model_name, binary = False, unicode_errors='ignore') print("Loading existing embedding vectors {}...".format(model_name)) else: print("Cannot find the embedding file!") embedding_weights = {key: embedding_model[word] if word in embedding_model else np.random.uniform(-0.25, 0.25, embedding_model.vector_size) for key, word in vocabulary_inv.items()} label_name = f'../{dataset_name}_data/labels.txt' if os.path.exists(label_name): centers = [None for _ in range(num_class)] with open(label_name) as fin: for idx, line in enumerate(fin): label = line.strip() centers[idx] = embedding_model[label] / np.linalg.norm(embedding_model[label]) return embedding_weights, centers def write_output(write_path, y_pred, perm): invperm = np.zeros(len(perm), dtype='int32') for i,v in enumerate(perm): invperm[v] = i y_pred = y_pred[invperm] with open(os.path.join(write_path, 'out.txt'), 'w') as f: for val in y_pred: f.write(str(val) + '\n') print("Classification results are written in {}".format(os.path.join(write_path, 'out.txt'))) return if __name__ == "__main__": import argparse parser = argparse.ArgumentParser(description='main', formatter_class=argparse.ArgumentDefaultsHelpFormatter) ### Basic settings ### # dataset selection: MAG-CS (default), Amazon parser.add_argument('--dataset', default='mag', choices=['mag', 'amazon']) # embedding files: joint embedding (default) parser.add_argument('--embedding', default='joint') # whether ground truth labels are available for evaluation: True (default), False parser.add_argument('--with_evaluation', default='True', choices=['True', 'False']) ### Training settings ### # mini-batch size for both pre-training and self-training: 256 (default) parser.add_argument('--batch_size', default=256, type=int) # training epochs: dataset-specific parser.add_argument('--pretrain_epochs', default=None, type=int) ### Hyperparameters settings ### # number of generated pseudo documents per class: dataset-specific parser.add_argument('--num_generated_docs', default=None, type=int) # keyword vocabulary size (gamma): 50 (default) parser.add_argument('--gamma', default=50, type=int) # vmf concentration parameter when synthesizing documents (kappa): dataset-specific parser.add_argument('--kappa', default=None, type=float) # number of copies each retrieved/generated pseudo-labeled document has in one epoch parser.add_argument('--ratio', default=10, type=int) ### Dummy arguments (please ignore) ### # weak supervision selection: labeled documents (default) parser.add_argument('--sup_source', default='docs', choices=['docs']) # maximum self-training iterations: 0 (default) parser.add_argument('--maxiter', default=0, type=int) # self-training update interval: None (default) parser.add_argument('--update_interval', default=None, type=int) # background word distribution weight (alpha): 0.0 (default) parser.add_argument('--alpha', default=0.0, type=float) # self-training stopping criterion (delta): None (default) parser.add_argument('--delta', default=0.1, type=float) # trained model directory: None (default) parser.add_argument('--trained_weights', default=None) args = parser.parse_args() print(args) alpha = args.alpha gamma = args.gamma delta = args.delta ratio = args.ratio word_embedding_dim = 100 update_interval = 50 self_lr = 1e-3 if args.dataset == 'mag': pretrain_epochs = 100 max_sequence_length = 200 num_generated_docs = 50 kappa = 200 beta = num_generated_docs * ratio elif args.dataset == 'amazon': pretrain_epochs = 100 max_sequence_length = 200 num_generated_docs = 100 kappa = 150 beta = num_generated_docs * ratio decay = 1e-6 if args.update_interval is not None: update_interval = args.update_interval if args.pretrain_epochs is not None: pretrain_epochs = args.pretrain_epochs if args.with_evaluation == 'True': with_evaluation = True else: with_evaluation = False if args.sup_source == 'docs': x, y, word_counts, vocabulary, vocabulary_inv_list, len_avg, len_std, word_sup_list, sup_idx, perm = \ load_dataset(args.dataset, model='cnn', sup_source=args.sup_source, with_evaluation=with_evaluation, truncate_len=max_sequence_length) np.random.seed(1234) vocabulary_inv = {key: value for key, value in enumerate(vocabulary_inv_list)} vocab_sz = len(vocabulary_inv) n_classes = len(word_sup_list) if x.shape[1] < max_sequence_length: max_sequence_length = x.shape[1] x = x[:, :max_sequence_length] sequence_length = max_sequence_length print("\n### Input preparation ###") embedding_weights, centers = load_embedding(vocabulary_inv, n_classes, args.dataset, args.embedding) embedding_mat = np.array([np.array(embedding_weights[word]) for word in vocabulary_inv]) wstc = WSTC(input_shape=x.shape, n_classes=n_classes, y=y, model='cnn', vocab_sz=vocab_sz, embedding_matrix=embedding_mat, word_embedding_dim=word_embedding_dim) if args.trained_weights is None: print("\n### Phase 1: vMF distribution fitting & pseudo document generation ###") word_sup_array = np.array([np.array([vocabulary[word] for word in word_class_list]) for word_class_list in word_sup_list]) total_counts = sum(word_counts[ele] for ele in word_counts) total_counts -= word_counts[vocabulary_inv_list[0]] background_array = np.zeros(vocab_sz) for i in range(1,vocab_sz): background_array[i] = word_counts[vocabulary_inv[i]]/total_counts seed_docs, seed_label = pseudodocs(word_sup_array, gamma, background_array, sequence_length, len_avg, len_std, beta, alpha, vocabulary_inv, embedding_mat, centers, kappa, 'cnn', './results/{}/{}/phase1/'.format(args.dataset, 'cnn')) if args.sup_source == 'docs': num_real_doc = len(sup_idx.flatten()) * ratio real_seed_docs, real_seed_label = augment(x, sup_idx, num_real_doc) seed_docs = np.concatenate((seed_docs, real_seed_docs), axis=0) seed_label = np.concatenate((seed_label, real_seed_label), axis=0) perm_seed = np.random.permutation(len(seed_label)) seed_docs = seed_docs[perm_seed] seed_label = seed_label[perm_seed] print('\n### Phase 2: pre-training with pseudo documents ###') wstc.pretrain(x=seed_docs, pretrain_labels=seed_label, sup_idx=sup_idx, optimizer=SGD(lr=0.1, momentum=0.9), epochs=pretrain_epochs, batch_size=args.batch_size, save_dir='./results/{}/{}/phase2'.format(args.dataset, 'cnn')) y_pred = wstc.predict(x) if y is not None: f1_macro, f1_micro = np.round(f1(y, y_pred), 5) print('F1 score after pre-training: f1_macro = {}, f1_micro = {}'.format(f1_macro, f1_micro)) else: print("\n### Directly loading trained weights ###") wstc.load_weights(args.trained_weights) y_pred = wstc.predict(x) if y is not None: f1_macro, f1_micro = np.round(f1(y, y_pred), 5) print('F1 score: f1_macro = {}, f1_micro = {}'.format(f1_macro, f1_micro)) print("\n### Generating outputs ###") write_output('./' + args.dataset, y_pred, perm)
7,853
36.759615
137
py
MotifClass
MotifClass-master/text_classification/model.py
import numpy as np np.random.seed(1234) import os from time import time import csv import keras.backend as K # K.set_session(K.tf.Session(config=K.tf.ConfigProto(intra_op_parallelism_threads=30, inter_op_parallelism_threads=30))) from keras.engine.topology import Layer from keras.layers import Dense, Input, Convolution1D, Embedding, GlobalMaxPooling1D, GRU, TimeDistributed from keras.layers.merge import Concatenate from keras.models import Model from keras import initializers, regularizers, constraints from keras.initializers import VarianceScaling, RandomUniform from sklearn.metrics import f1_score def f1(y_true, y_pred): y_true = y_true.astype(np.int64) assert y_pred.size == y_true.size f1_macro = f1_score(y_true, y_pred, average='macro') f1_micro = f1_score(y_true, y_pred, average='micro') return f1_macro, f1_micro def ConvolutionLayer(input_shape, n_classes, filter_sizes=[2, 3, 4, 5], num_filters=20, word_trainable=False, vocab_sz=None, embedding_matrix=None, word_embedding_dim=100, hidden_dim=20, act='relu', init='ones'): x = Input(shape=(input_shape,), name='input') z = Embedding(vocab_sz, word_embedding_dim, input_length=(input_shape,), name="embedding", weights=[embedding_matrix], trainable=word_trainable)(x) conv_blocks = [] for sz in filter_sizes: conv = Convolution1D(filters=num_filters, kernel_size=sz, padding="valid", activation=act, strides=1, kernel_initializer=init)(z) conv = GlobalMaxPooling1D()(conv) conv_blocks.append(conv) z = Concatenate()(conv_blocks) if len(conv_blocks) > 1 else conv_blocks[0] z = Dense(hidden_dim, activation="relu")(z) y = Dense(n_classes, activation="softmax")(z) return Model(inputs=x, outputs=y, name='classifier') def dot_product(x, kernel): if K.backend() == 'tensorflow': return K.squeeze(K.dot(x, K.expand_dims(kernel)), axis=-1) else: return K.dot(x, kernel) class AttentionWithContext(Layer): def __init__(self, W_regularizer=None, u_regularizer=None, b_regularizer=None, W_constraint=None, u_constraint=None, b_constraint=None, init='glorot_uniform', bias=True, **kwargs): self.supports_masking = True self.init = init self.W_regularizer = regularizers.get(W_regularizer) self.u_regularizer = regularizers.get(u_regularizer) self.b_regularizer = regularizers.get(b_regularizer) self.W_constraint = constraints.get(W_constraint) self.u_constraint = constraints.get(u_constraint) self.b_constraint = constraints.get(b_constraint) self.bias = bias super(AttentionWithContext, self).__init__(**kwargs) def build(self, input_shape): assert len(input_shape) == 3 self.W = self.add_weight(shape=(input_shape[-1], input_shape[-1],), initializer=self.init, name='{}_W'.format(self.name), regularizer=self.W_regularizer, constraint=self.W_constraint) if self.bias: self.b = self.add_weight(shape=(input_shape[-1],), initializer='zero', name='{}_b'.format(self.name), regularizer=self.b_regularizer, constraint=self.b_constraint) self.u = self.add_weight(shape=(input_shape[-1],), initializer=self.init, name='{}_u'.format(self.name), regularizer=self.u_regularizer, constraint=self.u_constraint) super(AttentionWithContext, self).build(input_shape) def compute_mask(self, input, input_mask=None): return None def call(self, x, mask=None): uit = dot_product(x, self.W) if self.bias: uit += self.b uit = K.tanh(uit) ait = dot_product(uit, self.u) a = K.exp(ait) if mask is not None: a *= K.cast(mask, K.floatx()) a /= K.cast(K.sum(a, axis=1, keepdims=True) + K.epsilon(), K.floatx()) a = K.expand_dims(a) weighted_input = x * a return K.sum(weighted_input, axis=1) def compute_output_shape(self, input_shape): return input_shape[0], input_shape[-1] def HierAttLayer(input_shape, n_classes, word_trainable=False, vocab_sz=None, embedding_matrix=None, word_embedding_dim=100, gru_dim=100, fc_dim=100): sentence_input = Input(shape=(input_shape[2],), dtype='int32') embedded_sequences = Embedding(vocab_sz, word_embedding_dim, input_length=input_shape[2], weights=[embedding_matrix], trainable=word_trainable)(sentence_input) l_lstm = GRU(gru_dim, return_sequences=True)(embedded_sequences) l_dense = TimeDistributed(Dense(fc_dim))(l_lstm) l_att = AttentionWithContext()(l_dense) sentEncoder = Model(sentence_input, l_att) x = Input(shape=(input_shape[1], input_shape[2]), dtype='int32') review_encoder = TimeDistributed(sentEncoder)(x) l_lstm_sent = GRU(gru_dim, return_sequences=True)(review_encoder) l_dense_sent = TimeDistributed(Dense(fc_dim))(l_lstm_sent) l_att_sent = AttentionWithContext()(l_dense_sent) y = Dense(n_classes, activation='softmax')(l_att_sent) return Model(inputs=x, outputs=y, name='classifier') class WSTC(object): def __init__(self, input_shape, n_classes=None, init=RandomUniform(minval=-0.01, maxval=0.01), y=None, model='cnn', vocab_sz=None, word_embedding_dim=100, embedding_matrix=None ): super(WSTC, self).__init__() self.input_shape = input_shape self.y = y self.n_classes = n_classes if model == 'cnn': self.classifier = ConvolutionLayer(self.input_shape[1], n_classes=n_classes, vocab_sz=vocab_sz, embedding_matrix=embedding_matrix, word_embedding_dim=word_embedding_dim, init=init) elif model == 'rnn': self.classifier = HierAttLayer(self.input_shape, n_classes=n_classes, vocab_sz=vocab_sz, embedding_matrix=embedding_matrix, word_embedding_dim=word_embedding_dim) self.model = self.classifier self.sup_list = {} def pretrain(self, x, pretrain_labels, sup_idx=None, optimizer='adam', loss='kld', epochs=200, batch_size=256, save_dir=None): self.classifier.compile(optimizer=optimizer, loss=loss) print("\nNeural model summary: ") self.model.summary() if sup_idx is not None: for i, seed_idx in enumerate(sup_idx): for idx in seed_idx: self.sup_list[idx] = i # begin pretraining t0 = time() print('\nPretraining...') self.classifier.fit(x, pretrain_labels, batch_size=batch_size, epochs=epochs) print('Pretraining time: {:.2f}s'.format(time() - t0)) if save_dir is not None: if not os.path.exists(save_dir): os.makedirs(save_dir) self.classifier.save_weights(save_dir + '/pretrained.h5') print('Pretrained model saved to {}/pretrained.h5'.format(save_dir)) self.pretrained = True def load_weights(self, weights): self.model.load_weights(weights) def predict(self, x): q = self.model.predict(x, verbose=0) return q.argmax(1) def target_distribution(self, q, power=2): weight = q**power / q.sum(axis=0) p = (weight.T / weight.sum(axis=1)).T for i in self.sup_list: p[i] = 0 p[i][self.sup_list[i]] = 1 return p def compile(self, optimizer='sgd', loss='kld'): self.model.compile(optimizer=optimizer, loss=loss) def fit(self, x, y=None, maxiter=5e4, batch_size=256, tol=0.1, power=2, update_interval=140, save_dir=None, save_suffix=''): print('Update interval: {}'.format(update_interval)) pred = self.classifier.predict(x) y_pred = np.argmax(pred, axis=1) y_pred_last = np.copy(y_pred) # logging file if not os.path.exists(save_dir): os.makedirs(save_dir) logfile = open(save_dir + '/self_training_log_{}.csv'.format(save_suffix), 'w') logwriter = csv.DictWriter(logfile, fieldnames=['iter', 'f1_macro', 'f1_micro']) logwriter.writeheader() index = 0 index_array = np.arange(x.shape[0]) for ite in range(int(maxiter)): if ite % update_interval == 0: q = self.model.predict(x, verbose=0) y_pred = q.argmax(axis=1) p = self.target_distribution(q, power) print('\nIter {}: '.format(ite)) if y is not None: f1_macro, f1_micro = np.round(f1(y, y_pred), 5) logdict = dict(iter=ite, f1_macro=f1_macro, f1_micro=f1_micro) logwriter.writerow(logdict) print('f1_macro = {}, f1_micro = {}'.format(f1_macro, f1_micro)) # check stop criterion delta_label = np.sum(y_pred != y_pred_last).astype(np.float) / y_pred.shape[0] y_pred_last = np.copy(y_pred) print('Fraction of documents with label changes: {} %'.format(np.round(delta_label*100, 3))) if ite > 0 and delta_label < tol/100: print('\nFraction: {} % < tol: {} %'.format(np.round(delta_label*100, 3), tol)) print('Reached tolerance threshold. Stopping training.') logfile.close() break # train on batch idx = index_array[index * batch_size: min((index+1) * batch_size, x.shape[0])] self.model.train_on_batch(x=x[idx], y=p[idx]) index = index + 1 if (index + 1) * batch_size <= x.shape[0] else 0 ite += 1 logfile.close() if save_dir is not None: self.model.save_weights(save_dir + '/final.h5') print("Final model saved to: {}/final.h5".format(save_dir)) return self.predict(x)
9,038
32.354244
124
py
lgv-geometric-transferability
lgv-geometric-transferability-main/attack_csgld_pgd_torch.py
""" Implementation of the attacks used in the article """ import numpy as np import pandas as pd import torch import argparse import time import os import sys import re from tqdm import tqdm import random from random import shuffle from utils.data import CIFAR10, CIFAR100, ImageNet, MNIST from utils.helpers import keyvalue, guess_model, guess_and_load_model, load_classifier, load_classifier_ensemble, list_models, \ compute_accuracy_from_nested_list_models, save_numpy, compute_norm, guess_method, USE_CUDA, DEVICE from utils.attacks import ExtendedProjectedGradientDescentPyTorch from art.attacks.evasion import CarliniLInfMethod from utils.models import LightNestedEnsemble from torchattacks import Square, CW, AutoAttack, OnePixel, PGD, PGDL2, APGD, FAB, DeepFool, MultiAttack from torch.utils.data import TensorDataset, DataLoader import torch.backends.cudnn as cudnn cudnn.benchmark = True cudnn.deterministic = True # parse args parser = argparse.ArgumentParser(description="Craft PGD adv ex with each update computed on a different samples from an ensemble of models trained with cSGLD") parser.add_argument("dirs_models", nargs='+', help="Path to directory containing all the models file of the ensemble model. Also support single path to a model file.") parser.add_argument('--attack', choices=['PGD', 'PGD_ta', 'APGD', 'FAB', 'Square', 'AutoAttack', 'CW', 'OnePixel', 'DeepFool'], default='PGD', help="Attack to craft adversarial examples. Only PGD supports momentum, .") parser.add_argument('--n-iter', type=int, default=None, help="Number of iterations to perform. If None (default), set to the number of samples.") parser.add_argument("--norm", choices=['1', '2', 'inf'], default='2', help="Type of L-norm to use. Default: 2") parser.add_argument("--max-norm", type=float, required=True, help="Max L-norm of the perturbation") parser.add_argument("--norm-step", type=float, required=True, help="Max norm at each step.") parser.add_argument('--n-ensemble', type=int, default=1, help="Number of samples to ensemble. Default: 1") parser.add_argument('--shuffle', action='store_true', help="Random order of models vs sequential order of the MCMC (default)") parser.add_argument('--n-random-init', type=int, default=0, help="Number of random restarts to perform. 0: no random init.") parser.add_argument('--grad-noise-std', type=float, default=None, help="Add Gaussian noise to gradients with the specified standard deviation.") parser.add_argument('--temperature', type=float, default=None, help="Temperature scaling the logits of the surrogate model. Deactivated if None (default).") parser.add_argument('--skip-first-n-models', type=int, default=0, help="Number of models samples to discard") parser.add_argument('--n-models-cycle', type=int, help="Number of models samples per cycle (only used for limit-n-samples-per-cycle or limit-n-cycles)") parser.add_argument('--limit-n-samples-per-cycle', type=int, default=None, help="Takes into account only the first n samples inside a cycle, droping off the last ones. Default: None (desactivated)") parser.add_argument('--method-samples-per-cycle', choices=['interval', 'true_interval', 'first', 'last'], default='interval', help="Method to select samples inside cycle. Use interval for cycle based surrogate, true_interval for non-cyclical surrogate.") parser.add_argument('--limit-n-cycles', type=int, default=None, help="Takes into account only the first n cycles, droping off the last ones. Default: None (desactivated)") # test time transferability improvements parser.add_argument('--ghost-attack', action='store_true', help="Load each model as a Ghost network (default: no model alteration)") parser.add_argument('--input-diversity', action='store_true', help="Add input diversity to each model (default: no model alteration)") parser.add_argument('--skip-gradient-method', action='store_true', help="Add Skip Gradient Method (SGM) backward hook to each surrogate model (default: no model alteration)") parser.add_argument('--translation-invariant', action='store_true', help="Apply translation invariance kernel to gradient (default: regular gradient)") parser.add_argument('--target-defense-randomization', action='store_true', help="The target model is loaded with defense randomization (default: regular target). Set to True with --translation-invariant.") parser.add_argument("--momentum", type=float, default=None, help="Apply momentum to gradients (default: regular gradient)") # target model parser.add_argument("--model-target-path", nargs='+', default=None, help="Path to the target models.") parser.add_argument("--csv-export", default=None, help="Path to CSV where to export data about target.") parser.add_argument("--csv-key-val", nargs='*', metavar="KEY=VALUE", action=keyvalue, help="Add the keys as columns with the corresponding values to the exported CSV.") parser.add_argument("--export-target-per-iter", type=int, default=None, help="Export target acc each N iterations in csv-export file. Default (None) 1 line for final data.") # others parser.add_argument("--n-examples", type=int, default=None, help="Craft adv ex on a subset of test examples. If None " "(default), perturbate all the test set. If " "model-target-path is set, extract the subset from " "the examples correctly predicted by it.") parser.add_argument("--data-path", default=None, help="Path of data. Only supported for ImageNet.") parser.add_argument("--validation", action='store_true', help="Craft adversarial examples from a validation set built from train set (of size: 2 x n_examples). Default: no validation set, examples from test set.") parser.add_argument("--seed", type=int, default=None, help="Set random seed") parser.add_argument("--batch-size", type=int, default=128, help="Batch size. Try a lower value if out of memory (especially for high values of --ensemble-inner).") parser.add_argument("--force-add", type=float, default=None, help="Add this scalar to the example. Use for compatibility with model trained on other range of pixels") parser.add_argument("--force-divide", type=float, default=None, help="Divide the example ex by this scalar. Use for compatibility with model trained on other range of pixels") parser.add_argument("--skip-accuracy-computation", action='store_true', help="Do not compute accuracies. To be used for full test set.") args = parser.parse_args() if args.norm == 'inf': args.norm = np.inf else: args.norm = int(args.norm) # check args if args.limit_n_samples_per_cycle or args.limit_n_cycles: if not args.n_models_cycle: raise ValueError("If a limit is set in the number of models to consider, you have to precise the number of samples per cycle.") if args.validation and not args.n_examples: raise ValueError('For validation set, please provide its size with n-examples arg') # set random seed if args.seed: torch.manual_seed(args.seed) torch.cuda.manual_seed_all(args.seed) np.random.seed(args.seed) random.seed(args.seed) # detect models if re.match('.+\\.pth?(\\.tar)?$', args.dirs_models[0]) and len(args.dirs_models) == 1: # link to single model if not os.path.isfile(args.dirs_models[0]): raise ValueError('Non-existing path surrogate file passed') paths_ensembles = [[args.dirs_models[0], ], ] else: paths_ensembles = [list_models(x) for x in args.dirs_models] print(f'Ensembles of models detected: {[len(x) for x in paths_ensembles]}') if args.skip_first_n_models: print(f'Discarding the first {args.skip_first_n_models} models') paths_ensembles = [x[args.skip_first_n_models:] for x in paths_ensembles] if any([len(x) == 0 for x in paths_ensembles]): raise ValueError('Empty model ensemble') if args.n_models_cycle: if any([len(x) % args.n_models_cycle != 0 for x in paths_ensembles]): print('Warning: Number of models is not a multiple of the number of models per cycle') if args.limit_n_cycles: if any([len(x) < args.limit_n_cycles * args.n_models_cycle for x in paths_ensembles]): raise ValueError(f'One of the ensemble is smaller than what expected ({ args.limit_n_cycles * args.n_models_cycle })') if args.limit_n_samples_per_cycle: if args.limit_n_samples_per_cycle > args.n_models_cycle: raise ValueError('Limit to nb samples > nb of samples per cycle.') # load test/val data validation_size = 2 * args.n_examples if args.validation else None # if validation data, set the size of val dataset 2xn_examples (to have enough correctly predicted examples) if 'CIFAR100' in args.dirs_models[0]: data = CIFAR100(batch_size=args.batch_size, validation=validation_size, seed=args.seed) elif 'CIFAR10' in args.dirs_models[0]: data = CIFAR10(batch_size=args.batch_size, validation=validation_size, seed=args.seed) elif 'ImageNet' in args.dirs_models[0]: data = ImageNet(batch_size=args.batch_size, path=args.data_path, validation=validation_size, seed=args.seed) elif 'MNIST' in args.dirs_models[0]: data = MNIST(batch_size=args.batch_size, validation=validation_size, seed=args.seed) else: raise NotImplementedError('Dataset not supported') model_target = None if args.model_target_path: # load target and select n_examples correctly predicted by it # target model is loaded with randomization defense for translation invariance models_target = [guess_and_load_model(path_model=x, data=data, defense_randomization=(args.translation_invariant or args.target_defense_randomization)) for x in args.model_target_path] X, y = data.correctly_predicted_to_numpy(models=models_target, train=False, validation=args.validation, N=args.n_examples, seed=args.seed) else: X, y = data.to_numpy(train=False, validation=args.validation, N=args.n_examples, seed=args.seed) if args.force_add: X += args.force_add data.min_pixel_value += args.force_add data.max_pixel_value += args.force_add if args.force_divide: X /= args.force_divide data.min_pixel_value /= args.force_divide data.max_pixel_value /= args.force_divide # limit cycles or samples per cycles if args.limit_n_cycles or args.limit_n_samples_per_cycle: paths_ensembles_lim = [] for i_ens, paths_models in enumerate(paths_ensembles): paths_ensembles_lim.append([]) for i, path_model in enumerate(paths_models): # stop if limit is set on the number of cycles to consider if args.limit_n_cycles: if i >= args.limit_n_cycles * args.n_models_cycle: break # only add current model for selected indexes if args.limit_n_samples_per_cycle: # select index (at regular interval, always including the last) max_index = args.n_models_cycle-1 if args.method_samples_per_cycle == 'interval': indexes_to_keep = [int(x.left) for x in pd.interval_range(start=0, end=max_index, periods=args.limit_n_samples_per_cycle-1)] + [max_index] elif args.method_samples_per_cycle == 'true_interval': indexes_to_keep = [int(x.left) for x in pd.interval_range(start=0, end=max_index+1, periods=args.limit_n_samples_per_cycle)] elif args.method_samples_per_cycle == 'last': indexes_to_keep = list(range(max_index - args.limit_n_samples_per_cycle+1, max_index+1)) elif args.method_samples_per_cycle == 'first': indexes_to_keep = list(range(0, args.limit_n_samples_per_cycle)) else: raise NotImplementedError('Method not supported.') if (i % args.n_models_cycle) not in indexes_to_keep: continue paths_ensembles_lim[i_ens].append(path_model) paths_ensembles = paths_ensembles_lim if any([len(x) != len(paths_ensembles[0]) for x in paths_ensembles]): raise NotImplementedError('All ensembles should have the same number of models.') print(f'Ensembles of models used: {[len(x) for x in paths_ensembles]}') # shuffle models if args.shuffle: for paths_models in paths_ensembles: shuffle(paths_models) # don't load unused models (if nb models > nb iters) if args.n_iter: max_nb_models_used = args.n_iter * args.n_ensemble for i, paths_models in enumerate(paths_ensembles): if len(paths_models) > max_nb_models_used: paths_ensembles[i] = paths_models[:max_nb_models_used] if len(args.dirs_models) > 1 and args.n_ensemble > 1: raise ValueError('Attacking multiple ensembles doesn\'t support n-ensemble arg.') # create nested list of models (ensemble > model) # [ens1: [m1, m2, m3, m4], ens2: [m5, m6, m7, m8]] ensemble_list = [] for i, path_model in enumerate(paths_ensembles[0]): # if we have multiple MCMC chains, we ensemble if len(paths_ensembles) > 1: ensemble_list.append([x[i] for x in paths_ensembles]) else: # if args.n_ensemble, we ensemble models from the same MCMC chain if len(ensemble_list) == 0: # avoid IndexError at first iteration ensemble_list.append([path_model, ]) elif len(ensemble_list[-1]) >= args.n_ensemble: ensemble_list.append([path_model, ]) else: ensemble_list[-1].append(path_model) # load each models and create ART classifier ensemble_classifiers = [] # list of ART classifiers. Each one has the logits fused list_ensemble_models = [] # nested list of torch models for i, ensemble_path in enumerate(ensemble_list): # only 1 model to attack if len(ensemble_path) == 1: model = guess_and_load_model(ensemble_path[0], data=data, load_as_ghost=args.ghost_attack, input_diversity=args.input_diversity, skip_gradient=args.skip_gradient_method, temperature=args.temperature) classifier = load_classifier(model, data=data) list_ensemble_models.append([model]) # if ensembling, store path_model to a list and build the ensembling model else: models_to_ensemble = [] for j, path_model in enumerate(ensemble_path): # load next model and continue only if ensemble is done models_to_ensemble.append(guess_and_load_model(path_model, data=data, load_as_ghost=args.ghost_attack, input_diversity=args.input_diversity, temperature=args.temperature, force_cpu=False)) classifier = load_classifier_ensemble(models_to_ensemble, data=data) list_ensemble_models.append(models_to_ensemble) ensemble_classifiers.append(classifier) del classifier # compute benign acc if not args.skip_accuracy_computation: acc_ens_prob, loss_ens_prob, predict_correct_ens = compute_accuracy_from_nested_list_models(list_ensemble=list_ensemble_models, X=X, y=y, data=data, export_predict=True) print(f"Accuracy on ensemble benign test examples: {acc_ens_prob*100:.3f}% (loss: {loss_ens_prob:.3f}).") # time code if USE_CUDA: torch.cuda.synchronize() start_time = time.perf_counter() if args.attack == 'PGD': attack = ExtendedProjectedGradientDescentPyTorch( estimators=ensemble_classifiers, targeted=False, norm=args.norm, eps=args.max_norm, eps_step=args.norm_step, max_iter=args.n_iter, num_random_init=args.n_random_init, batch_size=args.batch_size, translation_invariant=args.translation_invariant, momentum=args.momentum, grad_noise_std=args.grad_noise_std, models_target_dict={name: models_target[i] for i,name in enumerate(args.model_target_path)} if args.export_target_per_iter else None, freq_eval_target=args.export_target_per_iter, data=data ) X_adv = attack.generate(x=X, y=y) elif args.attack == 'CW' and args.norm == np.inf: ensemble_models = LightNestedEnsemble(list_models=list_ensemble_models, order=None) ensemble_classifier = load_classifier(ensemble_models, data=data) attack = CarliniLInfMethod( classifier=ensemble_classifier, targeted=False, eps=args.max_norm, max_iter=args.n_iter, batch_size=args.batch_size, learning_rate=0.01 ) X_adv = attack.generate(x=X, y=y) else: # attacks from torchattacks ensemble_models = LightNestedEnsemble(list_models=list_ensemble_models, order=None) # we take care of the order before norm_ta = f'L{args.norm}' if args.attack == 'PGD_ta' and norm_ta == 'Linf': attacks_list = [PGD(ensemble_models, eps=args.max_norm, alpha=args.norm_step, steps=args.n_iter, random_start=args.n_random_init > 0) for x in range(max(1, args.n_random_init))] attack = MultiAttack(attacks_list) elif args.attack == 'APGD': attack = APGD(ensemble_models, norm=norm_ta, eps=args.max_norm, steps=args.n_iter, n_restarts=args.n_random_init, loss='ce', seed=args.seed) elif args.attack == 'FAB': attack = FAB(ensemble_models, norm=norm_ta, eps=args.max_norm, steps=args.n_iter, n_restarts=args.n_random_init, seed=args.seed, n_classes=data.num_classes) elif args.attack == 'Square': attack = Square(ensemble_models, norm=norm_ta, eps=args.max_norm, n_queries=args.n_iter, n_restarts=1, loss='ce', seed=args.seed) elif args.attack == 'AutoAttack': attack = AutoAttack(ensemble_models, norm=norm_ta, eps=args.max_norm, n_classes=data.num_classes, seed=args.seed) elif args.attack == 'DeepFool': if args.norm != 2: raise ValueError('Only L2 norm supported for DeepFool attack') print('Warming: max-norm ignored for DeepFool attack!') attack = DeepFool(ensemble_models, steps=args.n_iter) elif args.attack == 'CW': if args.norm != 2: raise ValueError('Only L2 norm supported for CW attack') print('Warming: max-norm ignored for CW attack!') attacks_list = [CW(ensemble_models, c=c, steps=1000, lr=0.1, kappa=30) for c in [0.1, 1, 10, 100]] attack = MultiAttack(attacks_list) elif args.attack == 'OnePixel': print('Warming: norm ignored for OnePixel attack, max_norm used as nb pixels!') attack = OnePixel(ensemble_models, pixels=args.max_norm, steps=75, popsize=400, inf_batch=args.batch_size) else: raise NotImplementedError(f'Attack not implemented.') # implement batchsize X_dataset = TensorDataset(torch.tensor(X).to(DEVICE), torch.tensor(y).to(DEVICE)) X_loader = DataLoader(X_dataset, batch_size=args.batch_size, shuffle=False) X_adv = np.zeros((0,)+data.get_input_shape()[1:]) for X_batch, y_batch in tqdm(X_loader, desc='Batch'): X_adv_batch = attack(X_batch, y_batch).detach().cpu().numpy() X_adv = np.vstack((X_adv, X_adv_batch)) if X.shape != X_adv.shape: raise RuntimeError(f'X and X_adv do not have the same shape: {X.shape} ; {X_adv.shape}') if USE_CUDA: torch.cuda.synchronize() end_time = time.perf_counter() model_name_list = [guess_model(x) for x in args.dirs_models] # print stats if not args.skip_accuracy_computation: acc_ens_prob_adv, loss_ens_prob_adv = compute_accuracy_from_nested_list_models(list_ensemble=list_ensemble_models, X=X_adv, y=y, data=data) lpnorm = compute_norm(X_adv=X_adv, X=X, norm=args.norm) print( f"Surrogate stats after {args.n_iter} iters: Accuracy: {acc_ens_prob_adv * 100:.3f}%, Loss: {loss_ens_prob_adv:.3f} (from {loss_ens_prob:.3f}), " f"L{args.norm}-norm: mean {lpnorm.mean():.5f} (min {lpnorm.min():.5f} max {lpnorm.max():.5f}), Nb examples: {X_adv.shape[0]}, " f"Time: {(end_time - start_time) / 60:.3f} min") if args.csv_export: if not args.model_target_path: raise ValueError('Target model should be specified to export CSV.') for i, model_target in enumerate(models_target): acc_target_adv, loss_target_adv, acc_target_adv_ensok = compute_accuracy_from_nested_list_models([[model_target,],], X=X_adv, y=y, data=data, export_mask=predict_correct_ens) # transfer_rate_target = 1 - accuracy on adversarial examples predicted correctly both by the target and the surrogate transfer_rate_target = 1 - acc_target_adv_ensok nb_examples_transfer_rate = predict_correct_ens.sum().cpu().item() acc_target_original, loss_target_original = compute_accuracy_from_nested_list_models([[model_target,],], X=X, y=y, data=data) print(f'* On target: { args.model_target_path[i] }') print(f" Attack success rate: {(1-acc_target_adv) * 100:.3f} % (transfer rate: {transfer_rate_target * 100:.3f}% on {nb_examples_transfer_rate} examples)") print(f" Loss on target: {loss_target_adv:.3f} (vs. original {loss_target_original:.3f})") dict_metrics = args.csv_key_val.copy() if args.csv_key_val else dict() dict_metrics.update({ 'model_target': f"{'defense_randomization/' if (args.translation_invariant or args.target_defense_randomization) else ''}{args.model_target_path[i]}", 'arch_target': guess_model(args.model_target_path[i]), 'model_surrogate': args.dirs_models[0], 'surrogate_type': guess_method(args.dirs_models[0]), 'surrogate_archs': '_'.join(model_name_list), # 'surrogate_size_ensembles': len(paths_ensembles[0]), # nb models per arch 'surrogate_size_ensembles': args.n_models_cycle * args.limit_n_cycles if args.limit_n_cycles else len(paths_ensembles[0]), # nb models per arch 'norm_type': args.norm, 'norm_max': args.max_norm, 'norm_step': args.norm_step, 'n_iter': args.n_iter, 'n_ensemble': args.n_ensemble, 'n_random_init': args.n_random_init, 'momentum': args.momentum, 'shuffle': args.shuffle, 'ghost': args.ghost_attack, 'input_diversity': args.input_diversity, 'translation_invariant': args.translation_invariant, #'adv_fail_rate': acc_target_adv, # X contains only correctly predicted examples 'adv_success_rate': 1-acc_target_adv, 'transfer_rate': transfer_rate_target, 'loss_adv': loss_target_adv, 'loss_original': loss_target_original, 'adv_norm_mean': lpnorm.mean(), 'adv_norm_min': lpnorm.min(), 'adv_norm_max': lpnorm.max(), 'limit_samples_cycle': args.limit_n_samples_per_cycle, 'limit_cycles': args.limit_n_cycles, 'surrogate_acc_original_ex': acc_ens_prob, 'surrogate_acc_adv_ex': acc_ens_prob_adv, 'surrogate_loss_original_ex': loss_ens_prob, 'surrogate_loss_adv_ex': loss_ens_prob_adv, 'target_acc_original_ex': acc_target_original, 'acc_original_ex': acc_ens_prob, 'nb_adv': X_adv.shape[0], 'nb_adv_transf_rate': nb_examples_transfer_rate, # different nb of examples to compute the transfer rate 'dataset': 'val' if args.validation else 'test', 'time': end_time - start_time, 'transferability_techniques': f"{'MI_' if args.momentum else ''}{'ghost_' if args.ghost_attack else ''}{'DI_' if args.input_diversity else ''}{'TI_' if args.translation_invariant else ''}{'SGM_' if args.skip_gradient_method else ''}", 'grad_noise_std': args.grad_noise_std, 'temperature': args.temperature, 'attack': args.attack, 'args': ' '.join(sys.argv[1:]), }) df_metrics = pd.DataFrame([dict_metrics, ]) if args.export_target_per_iter: stats_target_dict = attack.get_target_accuracy_per_iter(args.model_target_path[i]) # duplicate the df line to the number of iterations df_metrics = pd.concat([df_metrics] * len(stats_target_dict['acc']), ignore_index=True) df_metrics['n_iter'] = stats_target_dict['iter'] df_metrics['adv_fail_rate'] = stats_target_dict['acc'] df_metrics['adv_success_rate'] = 1 - df_metrics['adv_fail_rate'] df_metrics['loss_adv'] = stats_target_dict['loss'] # create dir and append one line to csv os.makedirs(os.path.dirname(args.csv_export), exist_ok=True) df_metrics.to_csv(args.csv_export, mode='a', header=not os.path.exists(args.csv_export), index=False)
24,661
61.753181
254
py
lgv-geometric-transferability
lgv-geometric-transferability-main/compute_accuracy.py
import argparse import torch import torch.nn.functional as F import numpy as np import pandas as pd import os import sys import torchvision.datasets as datasets import torchvision.transforms as transforms from utils.helpers import list_models, guess_and_load_model, guess_method from utils.data import ImageNet def nll(outputs, labels): labels = labels.astype(int) idx = (np.arange(labels.size), labels) ps = outputs[idx] nll = -np.sum(np.log(ps)) return nll # parse args parser = argparse.ArgumentParser(description="Compute the accuracy of an ensemble of models") parser.add_argument("dir_models", help="Path to directory containing all the models file of the ensemble model") parser.add_argument("--data-path", default=None, help="Path of data. Only supported for ImageNet.") parser.add_argument("--batch-size", type=int, default=128, help="Batch size. Try a lower value if out of memory (especially for high values of --ensemble-inner).") parser.add_argument("--num-workers", type=int, default=10) parser.add_argument("--temperature", type=float, default=1, help="Apply temperature scaling.") parser.add_argument("--validation", type=int, default=None, help="Compute on a validation dataset of provided size (subset of the train test).") parser.add_argument("--seed", type=int, default=None, help="Random seed. Important to set with validation flag to have the same set.") parser.add_argument("--csv-export", default=None, help="Path of CSV to export.") args = parser.parse_args() if not args.seed and args.validation: raise ValueError('Provide random seed for validation set.') data = ImageNet(batch_size=args.batch_size, path=args.data_path) path_ensemble = list_models(args.dir_models) model_ensemble = [] for i, filename in enumerate(path_ensemble): model = guess_and_load_model(filename, data=data, load_as_ghost=False, input_diversity=False, skip_gradient=False) model.eval() model_ensemble.append(model) num_classes = 1000 traindir = os.path.join(args.data_path, 'train') valdir = os.path.join(args.data_path, 'val') val_transform = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), ]) if args.validation: # create a validation set from the train set train_dataset = datasets.ImageFolder(traindir, val_transform) # fix generator for reproducibility dataset, _ = torch.utils.data.random_split(train_dataset, lengths=[args.validation, len(train_dataset) - args.validation], generator=torch.Generator().manual_seed(args.seed)) else: # use the imagenet val set as test set dataset = datasets.ImageFolder(valdir, val_transform) loader = torch.utils.data.DataLoader( dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) predictions = np.zeros((len(loader.dataset), num_classes)) targets = np.zeros(len(loader.dataset)) k = 0 for input, target in loader: input = input.cuda(non_blocking=True) # torch.manual_seed(i) with torch.no_grad(): for model in model_ensemble: output = model(input) cur_preds = output.cpu().numpy() / args.temperature predictions[k:k+input.size()[0]] += F.softmax(output, dim=1).cpu().numpy() cur_targets = target.numpy() targets[k:(k+target.size(0))] = cur_targets k += input.size()[0] predictions = predictions / len(model_ensemble) test_acc = np.mean(np.argmax(predictions, axis=1) == targets) test_nll = nll(predictions, targets) / predictions.shape[0] test_ce = F.cross_entropy(predictions, targets) print("--- VAL ---" if args.validation else "--- TEST ---") print(f"Ensemble {args.dir_models} of {len(model_ensemble)} models") print(" Accuracy:", test_acc) print(" NLL:", test_nll) print(" CE:", test_ce) if args.csv_export: df_metrics = pd.DataFrame([{ 'dir_models': args.dir_models, 'temperature': args.temperature, 'dataset': 'val' if args.validation else 'test', 'model_type': guess_method(args.dir_models), 'accuracy': test_acc, 'nll': test_nll, 'nb_ex': len(loader.dataset), 'n_models': len(model_ensemble), 'args': ' '.join(sys.argv[1:]), },]) os.makedirs(os.path.dirname(args.csv_export), exist_ok=True) df_metrics.to_csv(args.csv_export, mode='a', header=not os.path.exists(args.csv_export), index=False)
4,542
37.82906
163
py
lgv-geometric-transferability
lgv-geometric-transferability-main/lgv/imagenet/analyse_weights_space.py
import pandas as pd import random import os import argparse from tqdm import tqdm import numpy as np import torch from torchvision import models as tmodels import torchvision.datasets as datasets import torchvision.transforms as transforms #from pyhessian import hessian from utils.data import ImageNet from utils.helpers import list_models, guess_and_load_model, guess_model from utils.pca_weights import PcaWeights, model2vector, models2tensor, vector2model # from utils.subspace_inference.posteriors import SWAG # from utils.subspace_inference.posteriors.proj_model import SubspaceModel # from utils.subspace_inference import utils, losses from utils.subspace_inference.utils import save_checkpoint def parse_args(): parser = argparse.ArgumentParser(description="Analyse weights space: PCA and shift subspace XPs") parser.add_argument("dir_models_pca", help="Path to directory containing all the models file of the ensemble model") parser.add_argument("path_original_model", help="Path to original model") parser.add_argument("--xp", choices=['PCA_projection', 'translate', 'hessian'], default='PCA_projection') parser.add_argument("--pca-rank", type=int, default=5) parser.add_argument("--export-dir", default=None, help="If set, export projected models on the first --export-ranks components") parser.add_argument("--export-ranks", nargs='*', type=int, default=[], help="If set, export projected models on the first --export-ranks components." "If muliple values are provided, export values recurvively on different subfolder. Must be > 0.") parser.add_argument("--update-bn", action='store_true', help="Update BN after projection to original space") parser.add_argument("--alpha-translate", type=float, default=1., help="Multiply deviations by this constant.") parser.add_argument("--dir-models-translate", default=None, help="Path to directory containing the models from " "another local maximum, to which we translate the " "dir_models_pca") parser.add_argument("--limit-n-export-models", type=int, default=None, help="Limit the number of exported model by randomly sampling them. Default: None, no limit.") # parser.add_argument("--path-original-model-translate", help="Path to original model of the other local maximum, to which we translate the dir_models_pca") parser.add_argument("--data-path", default=None, help="Path of data. Only supported for ImageNet.") parser.add_argument("--batch-size", type=int, default=128, help="Batch size. Try a lower value if out of memory (especially for high values of --ensemble-inner).") parser.add_argument("--num-workers", type=int, default=10) parser.add_argument("--seed", type=int, default=None, help="Random seed passed to PCA.") args = parser.parse_args() return args def main(args): np.random.seed(args.seed) data = ImageNet(batch_size=args.batch_size, path=args.data_path) traindir = os.path.join(args.data_path, 'train') valdir = os.path.join(args.data_path, 'val') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_dataset = datasets.ImageFolder( traindir, transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])) train_loader = torch.utils.data.DataLoader( train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) val_loader = torch.utils.data.DataLoader( datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])), batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) loaders = {'train': train_loader, 'test': val_loader} paths_pca = list_models(args.dir_models_pca) models_pca = [] for i, filename in enumerate(paths_pca): model = guess_and_load_model(filename, data=data, load_as_ghost=False, input_diversity=False, skip_gradient=False) models_pca.append(model) model_original = guess_and_load_model(args.path_original_model, data=data, load_as_ghost=False, input_diversity=False, skip_gradient=False) model_cfg = getattr(tmodels, guess_model(paths_pca[0])) # build a PCA of the target models # swag_model = SWAG(model_cfg, # subspace_type='pca', # subspace_kwargs={ # 'max_rank': len(models_pca), # collect all the models # 'pca_rank': args.pca_rank, # }, # num_classes=data.num_classes) # swag_model.cuda() # # emulate model collection # for model in models_pca: # swag_model.collect_model(model) # # print('Setting SWA solution (as subspace shift vector)') # swag_model.set_swa() # utils.bn_update(loaders["train"], swag_model, verbose=True, subset=0.1) # print(f" Metrics on test-set: {utils.eval(loaders['test'], swag_model, losses.cross_entropy)}") # print('Building subspace') # mean, variance, cov_factor = swag_model.get_space() # print(np.linalg.norm(cov_factor, axis=1)) # subspace = SubspaceModel(mean, cov_factor) if args.xp == 'hessian': raise NotImplementedError('XP not yet implemented') if args.xp == 'PCA_projection': # build a PCA of the target models pca = PcaWeights(model_cfg, rank=args.pca_rank, train_loader=loaders['train'], seed=args.seed) pca.fit(models_pca) # save the origin as SWA (0 rank) Z0 = np.zeros((1, args.pca_rank)) model_swa = pca.inverse_transform(Z0, update_bn=args.update_bn)[0] # analyse PCA print(f"Explained variance ratio: {pca.pca.explained_variance_ratio_}") total_cum_var_ratio = np.cumsum(pca.pca.explained_variance_ratio_) print(f"Total cumulated explained variance: {total_cum_var_ratio}") # print(f"Explained variance: {pca.pca.explained_variance_}") print(f"Singular values: {pca.pca.singular_values_}") df_metrics_pca = pd.DataFrame({ 'dir_models': args.dir_models_pca, 'dim': list(range(args.pca_rank)), 'expl_var': pca.pca.explained_variance_ratio_, 'totcum_expl_var': total_cum_var_ratio, 'singular_values': pca.pca.singular_values_, }) # project parameters into subspace, and into original space Z = pca.transform(models=models_pca) # original model Z_original = pca.transform(models=[model_original]) Z_all = np.concatenate((Z0, Z_original, Z), axis=0) labels_all = ['SWA'] + ['Original'] + ['Collected models'] * len(Z) # viz in comet: # experiment.log_embedding( # vectors=Z_all, # labels=labels_all, # title="Collected models (PCA)" # ) # extract models print('Projecting models...') model_original_proj = pca.inverse_transform(Z_original, update_bn=args.update_bn)[0] if args.export_dir: os.makedirs(args.export_dir, exist_ok=True) df_metrics_pca.to_csv(os.path.join(args.export_dir, 'metrics_pca.csv'), index=False) # save projected models export_dir = os.path.join(args.export_dir, 'dims_0') save_checkpoint(export_dir, name='model_swa', state_dict=model_swa.state_dict()) export_dir = os.path.join(args.export_dir, f'original_proj') save_checkpoint(export_dir, name='model_original_proj', state_dict=model_original_proj.state_dict()) for i in tqdm(args.export_ranks, desc="Export dims"): Z_proj = Z.copy() # i between 1 and pca_rank+1 Z_proj[:, i:] = 0 # project to the first i components models_pca_proj = pca.inverse_transform(Z_proj, update_bn=args.update_bn) export_dir = os.path.join(args.export_dir, f'dims_{i}') for j, model in enumerate(models_pca_proj): save_checkpoint(export_dir, name='models_pca_proj', sample=j, state_dict=model.state_dict()) if args.xp == 'translate': if not args.dir_models_translate: # or not args.path_original_model_translate raise ValueError('dir_models_translate should de provided') print('Translating to new local minimum...') w_original = model2vector(model_original) # compute first SWA W = models2tensor(models_pca) w_swa = torch.mean(W, 0) #w_swa = model2vector(model_swa) # can be check with: pca.transform(W=torch.reshape(w_swa, (1, w_swa.shape[0]))) # load other collected models paths_new_models = list_models(args.dir_models_translate) new_models = [] for i, filename in enumerate(paths_new_models): model = guess_and_load_model(filename, data=data, load_as_ghost=False, input_diversity=False, skip_gradient=False) new_models.append(model) W_new = models2tensor(new_models) w_new_swa = torch.mean(W_new, 0) # update BN and save new SWA # model_swa_new = vector2model(w_new_swa, model_cfg, update_bn=args.update_bn, train_loader=train_loader) # export_dir = os.path.join(args.export_dir, 'SWA_new') # save_checkpoint(export_dir, name='model_new_swa', state_dict=model_swa_new.state_dict()) # save another embedding with the new original model and the new collected models translated to the first SWA # Z_new_original_translated = pca.transform(W=torch.reshape(w_new_original - w_new_swa + w_swa, (1, w_new_swa.shape[0]))) # Z_new_translated = pca.transform(W=W_new - w_new_swa + w_swa) # Z_all = np.concatenate((Z_all, Z_new_original_translated, Z_new_translated), axis=0) # print(Z_new_translated.shape) # labels_all = labels_all + ['Original model 2 translated'] + ['Collected models 2'] * len(new_models) # experiment.log_embedding( # vectors=Z_all, # labels=labels_all, # title="Collected models (PCA) + translated new models" # ) # for translation_type in range(1, 4): # if translation_type == 1: # print(' ...Strategy 1: new_original_model + collected_models - original_model') # elif translation_type == 2: # print(' ...Strategy 2: new_original_model + collected_models - swa') # elif translation_type == 3: # print(' ...Strategy 3: new_swa + collected_models - swa') # else: # raise RuntimeError('Undefined translation_type') dot_prod_two_subspace_basis = np.zeros((len(models_pca), W_new.shape[0])) print('Translating: LGV-SWA_new + (LGV - LGV-SWA) ') index_to_export = None if args.limit_n_export_models: print(f'...Limiting Export to {args.limit_n_export_models} randomly picked models') random.seed(args.seed) index_to_export = random.sample(range(0, len(models_pca)), args.limit_n_export_models) for i, model_pca in enumerate(tqdm(models_pca, desc=f"Translation")): w_pca = model2vector(model_pca) # cosine similarity compute dot product of basis vectors of deviations for j in range(W_new.shape[0]): # dot product b/w unit vectors v1 = w_pca - w_swa v2 = W_new[j,:] - w_new_swa if not (len(v1.shape) == len(v2.shape) == 1): raise RuntimeError('Should compute cosine sim on vectors') dot_prod_two_subspace_basis[i,j] = (torch.dot(v1, v2) / (torch.linalg.norm(v1, ord=2) * torch.linalg.norm(v2, ord=2))).cpu().numpy() # angles_two_subspace_basis[i,j] = np.arccos(np.clip(torch.dot(v1_u, v2_u).cpu().numpy(), -1.0, 1.0)) # print(f'Dot product of {i} deviation LGV1 with all LGV2 deviations: {dot_prod_two_subspace_basis[i,:]}') # debug if (index_to_export is None) or (i in index_to_export): w_pca_trans = w_new_swa + args.alpha_translate * (w_pca - w_swa) model_pca_trans = vector2model(w=w_pca_trans, model_cfg=model_cfg, update_bn=args.update_bn, train_loader=train_loader) if args.export_dir: # export_dir = os.path.join(args.export_dir, f'translation_{translation_type}') export_dir = os.path.join(args.export_dir, f'translation_deviations_to_new_swa') save_checkpoint(export_dir, name='models_translated', sample=i, state_dict=model_pca_trans.state_dict()) print(f'\nCosine similarity of all basis vectors from the two LGV:') print(f' mean:{ dot_prod_two_subspace_basis.mean().item() }') print(f' mean abs:{ np.abs(dot_prod_two_subspace_basis).mean().item() }') print(f' min:{ dot_prod_two_subspace_basis.min().item() }') print(f' max:{ dot_prod_two_subspace_basis.max().item() }') if args.export_dir: pd.DataFrame(dot_prod_two_subspace_basis).to_csv(os.path.join(export_dir, 'dot_prod_two_subspace_basis.csv')) if __name__ == '__main__': args = parse_args() main(args)
13,744
54.873984
179
py
lgv-geometric-transferability
lgv-geometric-transferability-main/lgv/imagenet/analyse_feature_space.py
""" Interpolate between adv ex from 2 surrogate in feature space """ import os import sys import torch import math import random import argparse import numpy as np import pandas as pd from math import sqrt from tqdm import tqdm from utils.n_sphere import convert_spherical, convert_rectangular from utils.data import CIFAR10, ImageNet from utils.helpers import guess_and_load_model, load_classifier, list_models, project_on_sphere, compute_accuracy_from_nested_list_models, compute_norm, flatten from utils.attacks import ExtendedProjectedGradientDescentPyTorch # parse args parser = argparse.ArgumentParser(description="Interpolation of adversarial examples from two surrogate") parser.add_argument("path_model_1", help="Path to model 1. Could be either a directory of models or a path to a model.") parser.add_argument("path_model_2", help="Path to model 2. Could be either a directory of models or a path to a model.") parser.add_argument("--xp", choices=['interpolation', 'disk'], default='interpolation', help="Type of L-norm to use. Default: 2") parser.add_argument("--path_target", nargs='+', help="Path to target directory") parser.add_argument("--norm", choices=['1', '2', 'inf'], default='2', help="Type of L-norm to use. Default: 2") parser.add_argument("--max-norm", type=float, required=True, help="Max L-norm of the perturbation") parser.add_argument("--csv-export", default=None, help="Path to CSV where to export data about target.") # all xp parser.add_argument("--n-examples", type=int, default=2000, help="Craft adv ex on a subset of test examples. If None " "(default), perturbate all the test set. If " "model-target-path is set, extract the subset from " "the examples correctly predicted by it.") # xp interpolation parser.add_argument("--n-interpolation", type=int, default=100, help="Number of adv ex to compute along the generated path (include the two original ones).") parser.add_argument('--interpolation-method', choices=['linear', 'proj_sphere', 'polar', 'linear_hyperspherical_coord'], default='linear_hyperspherical_coord', help="Interpolation method between the 2 adv examples") parser.add_argument("--alpha-range", type=float, default=1, help="Max alpha interpolation coef. alpha in [-alphamax, 1+alphamax]") # xp disk parser.add_argument("--n-points", type=int, default=500, help="Number of points to evaluate loss in the disk.") parser.add_argument("--grid", choices=['grid', 'sunflower'], default='grid', help="Type of grid to generate points in the disk.") # others parser.add_argument("--data-path", default=None, help="Path of data. Only supported for ImageNet.") parser.add_argument("--seed", type=int, default=None, help="Set random seed") parser.add_argument("--batch-size", type=int, default=128, help="Batch size. Try a lower value if out of memory (especially for high values of --ensemble-inner).") args = parser.parse_args() if args.norm == 'inf': args.norm = np.inf else: args.norm = int(args.norm) if args.seed: torch.manual_seed(args.seed) torch.cuda.manual_seed_all(args.seed) np.random.seed(args.seed) random.seed(args.seed) if 'CIFAR10' in args.path_model_1: data = CIFAR10(batch_size=args.batch_size) elif 'ImageNet' in args.path_model_1: data = ImageNet(batch_size=args.batch_size, path=args.data_path) else: raise ValueError('dataset not supported') path_model_list_1 = list_models(args.path_model_1) path_model_list_2 = list_models(args.path_model_2) list_model_1 = [guess_and_load_model(path, data=data) for path in path_model_list_1] list_model_2 = [guess_and_load_model(path, data=data) for path in path_model_list_2] list_classifier_1 = [load_classifier(model, data=data) for model in list_model_1] list_classifier_2 = [load_classifier(model, data=data) for model in list_model_2] models_target = [guess_and_load_model(path_model=x, data=data) for x in args.path_target] X, y = data.correctly_predicted_to_numpy(models=models_target, train=False, N=args.n_examples, seed=args.seed) attack1 = ExtendedProjectedGradientDescentPyTorch(estimators=list_classifier_1, targeted=False, norm=args.norm, eps=args.max_norm, eps_step=args.max_norm / 10., max_iter=50, num_random_init=0, batch_size=args.batch_size) attack2 = ExtendedProjectedGradientDescentPyTorch(estimators=list_classifier_2, targeted=False, norm=args.norm, eps=args.max_norm, eps_step=args.max_norm / 10., max_iter=50, num_random_init=0, batch_size=args.batch_size) X_adv1 = attack1.generate(x=X, y=y) X_adv2 = attack2.generate(x=X, y=y) delta1 = X_adv1 - X delta2 = X_adv2 - X print('Natural accuracy:') for i, model_target in enumerate(models_target): acc_target_original, loss_target_original = compute_accuracy_from_nested_list_models([[model_target, ], ], X=X, y=y, data=data) print(f' * {args.path_target[i]}: {acc_target_original*100}% ; loss: {loss_target_original}') if args.xp == "interpolation": alpha_list = np.linspace(-args.alpha_range, 1+args.alpha_range, args.n_interpolation, endpoint=True).tolist() metrics_list = [] if args.interpolation_method == 'linear_hyperspherical_coord': # get nb digits precision for n-sphere rounding precision_dtype = np.finfo(delta1.dtype).precision # perturbations in spherical coordinates delta1_sc = convert_spherical(flatten(delta1), digits=precision_dtype) delta2_sc = convert_spherical(flatten(delta2), digits=precision_dtype) for alpha in tqdm(alpha_list, desc="Interpolate adv ex"): if args.interpolation_method == 'linear': X_adv_interp = X + (1-alpha) * delta1 + alpha * delta2 elif args.interpolation_method == 'proj_sphere': X_adv_interp = X + (1-alpha) * delta1 + alpha * delta2 X_adv_interp = project_on_sphere(X=X, X_adv=X_adv_interp, data=data, size=args.max_norm, norm=args.norm) elif args.interpolation_method == 'polar': # https://www.inference.vc/high-dimensional-gaussian-distributions-are-soap-bubble/ # valid under some condition X_adv_interp = X + sqrt(1-alpha)*delta1 + sqrt(alpha)*delta2 elif args.interpolation_method == 'linear_hyperspherical_coord': if args.norm != 2: raise ValueError('linear_hyperspherical_coord interpolation only guarantee for 2-norm') # the most appropriate interpolation # https://en.wikipedia.org/wiki/N-sphere#Spherical_coordinates # interpolate in spherical coordinates delta_inter_sc = (1-alpha) * delta1_sc + alpha * delta2_sc # if we extrapolate, the perturbation may be outside the L2 ball. We project it into the L2 sphere # the first dim is the l2 norm (radius of the sphere) examples_to_proj = (delta_inter_sc[:, 0] > args.max_norm) delta_inter_sc[examples_to_proj, 0] = args.max_norm delta_inter = convert_rectangular(delta_inter_sc, digits=precision_dtype) X_adv_interp = X + delta_inter.reshape(X.shape) else: raise ValueError('Interpolation method not supported') # check norm perturbation lpnorm = compute_norm(X_adv=X_adv_interp, X=X, norm=args.norm) if (lpnorm > args.max_norm + 1e-6).any(): print(f'For alpha={alpha}, nb examples outside the Lp ball: {(lpnorm > args.max_norm + 1e-6).sum()}') acc_surrogate1, loss_surrogate1 = compute_accuracy_from_nested_list_models([list_model_1, ], X=X_adv_interp, y=y, data=data) metrics_list.append({'model': args.path_model_1, 'type_model': 'surrogate', 'alpha': alpha, 'adv_accuracy': acc_surrogate1, 'adv_loss': loss_surrogate1}) acc_surrogate2, loss_surrogate2 = compute_accuracy_from_nested_list_models([list_model_2, ], X=X_adv_interp, y=y, data=data) metrics_list.append({'model': args.path_model_2, 'type_model': 'surrogate', 'alpha': alpha, 'adv_accuracy': acc_surrogate2, 'adv_loss': loss_surrogate2}) for i, model_target in enumerate(models_target): acc_target, loss_target = compute_accuracy_from_nested_list_models([[model_target, ], ], X=X_adv_interp, y=y, data=data) metrics_list.append({'model': args.path_target[i], 'type_model': 'target', 'alpha': alpha, 'adv_accuracy': acc_target, 'adv_loss': loss_target}) df_metrics = pd.DataFrame(metrics_list) df_metrics['norm'] = args.norm df_metrics['max_norm'] = args.max_norm df_metrics['n_examples'] = args.n_examples df_metrics['n_interpolation'] = args.n_interpolation df_metrics['interpolation_method'] = args.interpolation_method df_metrics['seed'] = args.seed os.makedirs(os.path.dirname(args.csv_export), exist_ok=True) df_metrics.to_csv(args.csv_export, header=True, index=False) # for i, path_target in enumerate(args.path_target): # df_ = df_metrics.query(f'model == "{path_target}"') # experiment.log_curve(f"target_{path_target}", x=df_['alpha'], y=df_['adv_loss']) # # df_ = df_metrics.query(f'model == "{args.path_model_1}"') # experiment.log_curve(f"target_{args.path_model_1}", x=df_['alpha'], y=df_['adv_loss']) # # df_ = df_metrics.query(f'model == "{args.path_model_2}"') # experiment.log_curve(f"target_{args.path_model_2}", x=df_['alpha'], y=df_['adv_loss']) elif args.xp == 'disk': metrics_list = [] # define plane with 3 points: X, X_adv1, X_adv2 using Gram-Schmidt # quote from SWA paper "Suppose we have three weight vectors w1, w2, w3. # We set u = (w2−w1), v = (w3−w1)− <w3−w1, w2− w1>/||w2− w1||2 · (w2 − w1). # Then the normalized vectors u_ = u/||u||, v_ = v/||v|| form an orthonormal basis in the plane containing w1, w2, w3." u = flatten(delta1) # (X_adv1−X) u_ = u / np.linalg.norm(u, axis=1, keepdims=True) # normalised vector u__ = u / np.linalg.norm(u, axis=1, keepdims=True)**2 # tmp v = flatten(delta2) - flatten(np.diagonal(np.dot(flatten(delta2), u__.T))) * u # delta2 = Xadv2-X # check with np.isclose( flatten(X_adv2 - X)[0] - np.dot(flatten(X_adv2 - X)[0], u__[0]) * u[0] , v[0] ) v_ = v / np.linalg.norm(v, axis=1, keepdims=True) # normalised vector if args.grid == 'grid': nb_points_per_axis = math.ceil(math.sqrt(args.n_points)) x_values = np.linspace(-args.max_norm, args.max_norm, nb_points_per_axis, endpoint=True) x1, x2 = np.meshgrid(x_values, x_values) x1, x2 = x1.flatten(), x2.flatten() elif args.grid == 'sunflower': # then we generate N points in the disk of radius epsilon # 2D sunflower spiral algorithm https://stackoverflow.com/a/44164075/6253027 indices = np.arange(0, args.n_points, dtype=float) + 0.5 r = np.sqrt(indices / args.n_points) * args.max_norm # max_norm = epsilon theta = np.pi * (1 + 5 ** 0.5) * indices x1, x2 = r * np.cos(theta), r * np.sin(theta) else: raise ValueError("Wrong type of grid") # save position of u and v df_Xadv1 = pd.DataFrame({ 'model': args.path_model_1, 'type_model': 'surrogate_1', 'x1': np.diagonal(np.dot(flatten(delta1), u_.T)), 'x2': np.diagonal(np.dot(flatten(delta1), v_.T)), }) df_Xadv2 = pd.DataFrame({ 'model': args.path_model_2, 'type_model': 'surrogate_2', 'x1': np.diagonal(np.dot(flatten(delta2), u_.T)), 'x2': np.diagonal(np.dot(flatten(delta2), v_.T)), }) df_Xadv = df_Xadv1.append(df_Xadv2, ignore_index=True) print(f'Position of Xadv1: {df_Xadv1["x1"].mean()}, {df_Xadv1["x2"].mean()}') print(f'Position of Xadv2: {df_Xadv2["x1"].mean()}, {df_Xadv2["x2"].mean()}') os.makedirs(os.path.dirname(args.csv_export), exist_ok=True) path_csv_xadv = args.csv_export.replace('.csv', '__ref_xadv.csv') df_Xadv.to_csv(path_csv_xadv, header=True, index=False) # and generate the examples to evaluate # "A point P with coordinates(x, y) in the plane would then be given by P = w1+x·u_ + y·v_" # for each point in disk for i in tqdm(range(args.n_points), desc="Predicting points"): X_disk = X + (x1[i] * u_ + x2[i] * v_).reshape(X.shape) if args.grid != 'grid': lpnorm = compute_norm(X_adv=X_disk, X=X, norm=args.norm) if (lpnorm > args.max_norm + 1e-6).any(): print(f'At point #{i}, nb examples outside the Lp ball: {(lpnorm > args.max_norm + 1e-6).sum()}') acc_surrogate1, loss_surrogate1 = compute_accuracy_from_nested_list_models([list_model_1, ], X=X_disk, y=y, data=data) metrics_list.append( {'model': args.path_model_1, 'type_model': 'surrogate', 'x1': x1[i], 'x2': x2[i], 'adv_accuracy': acc_surrogate1, 'adv_loss': loss_surrogate1}) acc_surrogate2, loss_surrogate2 = compute_accuracy_from_nested_list_models([list_model_2, ], X=X_disk, y=y, data=data) metrics_list.append( {'model': args.path_model_2, 'type_model': 'surrogate', 'x1': x1[i], 'x2': x2[i], 'adv_accuracy': acc_surrogate2, 'adv_loss': loss_surrogate2}) for j, model_target in enumerate(models_target): acc_target, loss_target = compute_accuracy_from_nested_list_models([[model_target, ], ], X=X_disk, y=y, data=data) metrics_list.append( {'model': args.path_target[j], 'type_model': 'target', 'x1': x1[i], 'x2': x2[i], 'adv_accuracy': acc_target, 'adv_loss': loss_target}) df_metrics = pd.DataFrame(metrics_list) df_metrics['norm'] = args.norm df_metrics['max_norm'] = args.max_norm df_metrics['n_examples'] = args.n_examples df_metrics['n_points'] = args.n_points df_metrics['seed'] = args.seed df_metrics['xp'] = args.xp df_metrics.to_csv(args.csv_export, header=True, index=False) else: raise ValueError("Wrong XP id")
14,705
54.91635
215
py
lgv-geometric-transferability
lgv-geometric-transferability-main/lgv/imagenet/generate_parametric_path.py
import os import argparse from tqdm import tqdm import numpy as np import torch from torchvision import models as tmodels import torchvision.datasets as datasets import torchvision.transforms as transforms from utils.data import ImageNet from utils.helpers import guess_and_load_model, guess_model from utils.pca_weights import model2vector, models2tensor, vector2model from utils.subspace_inference.utils import save_checkpoint def parse_args(): parser = argparse.ArgumentParser(description="Export models along a path connecting two models") parser.add_argument("path_model_1", help="Path to first model") parser.add_argument("path_model_2", help="Path to 2nd model") parser.add_argument("--names-models", nargs=2, metavar=('name1', 'name2'), help="Names of the two models") parser.add_argument("--alphas", default=None, nargs='*', type=float, help="Overwrite default list of alphas. If specified, n-models is ignored.") parser.add_argument("--n-models", default=100, type=int, help="Number of models to export") parser.add_argument("--export-dir", default=None, help="Dir to export models and CSV") parser.add_argument("--update-bn", action='store_true', help="Update BN of produced models") parser.add_argument("--data-path", default=None, help="Path of data. Only supported for ImageNet.") parser.add_argument("--batch-size", type=int, default=128, help="Batch size. Try a lower value if out of memory (especially for high values of --ensemble-inner).") parser.add_argument("--num-workers", type=int, default=10) parser.add_argument("--seed", type=int, default=None, help="Random seed passed to PCA.") args = parser.parse_args() return args def main(args): np.random.seed(args.seed) data = ImageNet(batch_size=args.batch_size, path=args.data_path) traindir = os.path.join(args.data_path, 'train') valdir = os.path.join(args.data_path, 'val') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_dataset = datasets.ImageFolder( traindir, transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])) train_loader = torch.utils.data.DataLoader( train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) val_loader = torch.utils.data.DataLoader( datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])), batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) loaders = {'train': train_loader, 'test': val_loader} model1 = guess_and_load_model(args.path_model_1, data=data, load_as_ghost=False, input_diversity=False, skip_gradient=False) model2 = guess_and_load_model(args.path_model_2, data=data, load_as_ghost=False, input_diversity=False, skip_gradient=False) model_cfg = getattr(tmodels, guess_model(args.path_model_1)) theta1 = model2vector(model1) theta2 = model2vector(model2) # alpha_min, alpha_max = args.alpha_range alpha_min, alpha_max = -1, 2 # hardcoded to avoid corner cases if args.alphas: alpha_list = list(args.alphas) else: alpha_list = np.linspace(alpha_min, alpha_max, args.n_models, endpoint=True).tolist() if 0. not in alpha_list: alpha_list += [0.] if 1. not in alpha_list: alpha_list += [1.] print(f'Start exporting the following alpha values: {["{0:0.2f}".format(i) for i in alpha_list]}') for alpha in tqdm(alpha_list, desc="Export interpolated models"): theta_alpha = torch.lerp(theta1, theta2, alpha).detach().clone() model_alpha = vector2model(theta_alpha, model_cfg, update_bn=args.update_bn, train_loader=loaders['train']) export_dir = os.path.join(args.export_dir, f'alpha_{alpha}') save_checkpoint(export_dir, name='model_interpolated', sample=alpha, state_dict=model_alpha.state_dict()) if __name__ == '__main__': args = parse_args() main(args)
4,285
46.622222
167
py
lgv-geometric-transferability
lgv-geometric-transferability-main/lgv/imagenet/train_swag_imagenet.py
import argparse import os import random import sys import time import tabulate from collections import OrderedDict import torch import torch.nn.functional as F import torchvision.models import timm from utils.swag import data from utils.subspace_inference import utils, losses #from utils.swag.posteriors import SWAG from utils.subspace_inference.posteriors import SWAG parser = argparse.ArgumentParser(description="SGD/SWA training") parser.add_argument( "--dir", type=str, default=None, required=True, help="training directory (default: None)", ) parser.add_argument( "--data_path", type=str, default=None, required=True, metavar="PATH", help="path to datasets location (default: None)", ) parser.add_argument( "--batch_size", type=int, default=256, metavar="N", help="input batch size (default: 256)", ) parser.add_argument( "--num_workers", type=int, default=4, metavar="N", help="number of workers (default: 4)", ) parser.add_argument( "--model", type=str, default=None, required=True, metavar="MODEL", help="model name (default: None)", ) parser.add_argument( "--pretrained", action="store_true", help="pretrained model usage flag (default: off)", ) parser.add_argument( "--pretrained_ckpt", type=str, default=None, help="pretrained behavior from model checkpoint (default: off)", ) parser.add_argument( "--parallel", action="store_true", help="data parallel model switch (default: off)" ) parser.add_argument( "--resume", type=str, default=None, metavar="CKPT", help="checkpoint to resume training from (default: None). Should be trained on torchvision data normalization.", ) parser.add_argument( "--epochs", type=int, default=5, metavar="N", help="number of epochs to train (default: 5)", ) parser.add_argument( "--save_freq", type=int, default=1, metavar="N", help="save frequency (default: 1)" ) parser.add_argument( "--eval_freq", type=int, default=1, metavar="N", help="evaluation frequency (default: 1)", ) parser.add_argument( "--eval_freq_swa", type=int, default=1, metavar="N", help="evaluation frequency of SWA solution, need BN update (default: 1)", ) parser.add_argument( "--eval_start", action='store_true', help="evaluation of the initial model (default: deactivated)", ) parser.add_argument( "--lr_init", type=float, default=0.01, metavar="LR", help="initial learning rate (default: 0.01)", ) parser.add_argument( "--momentum", type=float, default=0.9, metavar="M", help="SGD momentum (default: 0.9)", ) parser.add_argument( "--wd", type=float, default=1e-4, help="weight decay (default: 1e-4)" ) parser.add_argument("--swa", action="store_true", help="swa usage flag (default: off)") parser.add_argument( "--swa_cpu", action="store_true", help="store swag on cpu (default: off)" ) parser.add_argument( "--swa_start", type=float, default=161, metavar="N", help="SWA start epoch number (default: 161)", ) parser.add_argument( "--swa_lr", type=float, default=0.02, metavar="LR", help="SWA LR (default: 0.02)" ) parser.add_argument( "--swa_freq", type=int, default=4, metavar="N", help="SWA model collection frequency/ num samples per epoch (default: 4)", ) parser.add_argument('--max_num_models', type=int, default=20, help='maximum number of SWAG models to save') parser.add_argument( "--swa_resume", type=str, default=None, metavar="CKPT", help="checkpoint to restor SWA from (default: None)", ) parser.add_argument( "--loss", type=str, default="CE", help="loss to use for training model (default: Cross-entropy)", ) parser.add_argument( "--seed", type=int, default=1, metavar="S", help="random seed (default: 1)" ) parser.add_argument("--no_schedule", action="store_true", help="store schedule") # parser.add_argument('--inference', choices=['low_rank_gaussian', 'projected_sgd'], default='low_rank_gaussian') parser.add_argument('--subspace', choices=['covariance', 'pca', 'freq_dir'], default='covariance') parser.add_argument('--no-save-swag', action='store_true', default="Don't save swag checkpoint") args = parser.parse_args() args.device = None if torch.cuda.is_available(): args.device = torch.device("cuda") else: args.device = torch.device("cpu") print("Preparing directory %s" % args.dir) os.makedirs(args.dir, exist_ok=True) with open(os.path.join(args.dir, "command.sh"), "w") as f: f.write(" ".join(sys.argv)) f.write("\n") # torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = True random.seed(args.seed) torch.manual_seed(args.seed) torch.cuda.manual_seed(args.seed) print("Loading ImageNet from %s" % (args.data_path)) loaders, num_classes = data.loaders(args.data_path, args.batch_size, args.num_workers) if 'timm_' in args.model: print(f"Loading { '' }model from timm: {args.model}") print(f"Loading { 'pretrained ' if args.pretrained else ''}model from timm: {args.model}") arch_ = args.model.replace('timm_', '') model = timm.create_model(arch_, pretrained=args.pretrained) config = timm.data.resolve_data_config({}, model=model) if config['mean'] != (0.485, 0.456, 0.406) or config['std'] != (0.229, 0.224, 0.225): raise NotImplementedError(f"This model requires non-default normalization values. got: {config['mean'], config['std']}") model_class = getattr(timm.models, arch_) else: print("Using torchvision model %s" % args.model) model_class = getattr(torchvision.models, args.model) print("Preparing model") model = model_class(pretrained=args.pretrained, num_classes=num_classes) model.to(args.device) if args.swa: print("SWAG training") swag_model = SWAG(model_class, subspace_type=args.subspace, subspace_kwargs={'max_rank': args.max_num_models}, num_classes=num_classes) args.swa_device = "cpu" if args.swa_cpu else args.device swag_model.to(args.swa_device) if args.pretrained: print(f"Starting from pretrained model") model.to(args.swa_device) swag_model.collect_model(model) model.to(args.device) if args.pretrained_ckpt: print(f"Starting from chkpt: {args.pretrained_ckpt}") checkpoint = torch.load(args.pretrained_ckpt) try: model.load_state_dict(checkpoint["state_dict"]) except RuntimeError: new_state_dict = OrderedDict() for k, v in checkpoint['state_dict'].items(): name = k[7:] # remove `module.` new_state_dict[name] = v model.load_state_dict(new_state_dict) # TODO: also load optimizer if available model.to(args.swa_device) swag_model.collect_model(model) model.to(args.device) else: print("SGD training") def schedule(epoch): if args.swa and epoch >= args.swa_start: return args.swa_lr else: return args.lr_init * (0.1 ** (epoch // 30)) # use a slightly modified loss function that allows input of model if args.loss == "CE": criterion = losses.cross_entropy # criterion = F.cross_entropy elif args.loss == "adv_CE": criterion = losses.adversarial_cross_entropy else: raise NotImplementedError('criterion not implemented') optimizer = torch.optim.SGD( model.parameters(), lr=args.lr_init, momentum=args.momentum, weight_decay=args.wd ) if args.parallel: print("Using Data Parallel model") model = torch.nn.parallel.DataParallel(model) start_epoch = 0 if args.resume is not None: print("Resume training from %s" % args.resume) checkpoint = torch.load(args.resume) start_epoch = checkpoint["epoch"] model.load_state_dict(checkpoint["state_dict"]) optimizer.load_state_dict(checkpoint["optimizer"]) if args.swa and args.swa_resume is not None: checkpoint = torch.load(args.swa_resume) swag_model.subspace.rank = torch.tensor(0) swag_model.load_state_dict(checkpoint["state_dict"]) columns = ["ep", "lr", "tr_loss", "tr_acc", "te_loss", "te_acc", "time", "mem_usage"] if args.swa: columns = columns[:-2] + ["swa_te_loss", "swa_te_acc"] + columns[-2:] swag_res = {"loss": None, "accuracy": None} # deactivate original model save # utils.save_checkpoint( # args.dir, # start_epoch, # state_dict=model.state_dict(), # optimizer=optimizer.state_dict(), # ) if args.eval_start: print("START CKPT EVAL TEST") init_res = utils.eval(loaders["test"], model, criterion, verbose=True) print(f"Loss: {init_res['loss']} ; Acc: {init_res['accuracy']}") print("START CKPT EVAL TRAIN") init_res = utils.eval(loaders["train"], model, criterion, verbose=True) print(f"Loss: {init_res['loss']} ; Acc: {init_res['accuracy']}") num_iterates = 0 for epoch in range(start_epoch, args.epochs): time_ep = time.time() if not args.no_schedule: lr = schedule(epoch) utils.adjust_learning_rate(optimizer, lr) else: lr = args.lr_init print("EPOCH %d. TRAIN" % (epoch + 1)) if args.swa and (epoch + 1) > args.swa_start: subset = 1.0 / args.swa_freq for i in range(args.swa_freq): print("PART %d/%d" % (i + 1, args.swa_freq)) train_res = utils.train_epoch( loaders["train"], model, criterion, optimizer, subset=subset, verbose=True, ) num_iterates += 1 utils.save_checkpoint( args.dir, num_iterates, name="iter", state_dict=model.state_dict() ) model.to(args.swa_device) swag_model.collect_model(model) model.to(args.device) else: train_res = utils.train_epoch( loaders["train"], model, criterion, optimizer, verbose=True ) if ( epoch == 0 or epoch % args.eval_freq == args.eval_freq - 1 or epoch == args.epochs - 1 ): print("EPOCH %d. EVAL" % (epoch + 1)) test_res = utils.eval(loaders["test"], model, criterion, verbose=True) else: test_res = {"loss": None, "accuracy": None} if args.swa and (epoch + 1) > args.swa_start: if ( epoch == args.swa_start or epoch % args.eval_freq_swa == args.eval_freq_swa - 1 or epoch == args.epochs - 1 ): swag_res = {"loss": None, "accuracy": None} swag_model.to(args.device) #swag_model.sample(0.0) swag_model.set_swa() print("EPOCH %d. SWA BN" % (epoch + 1)) utils.bn_update(loaders["train"], swag_model, verbose=True, subset=0.1) print("EPOCH %d. SWA EVAL" % (epoch + 1)) swag_res = utils.eval(loaders["test"], swag_model, criterion, verbose=True) swag_model.to(args.swa_device) else: swag_res = {"loss": None, "accuracy": None} if (epoch + 1) % args.save_freq == 0: if args.swa: if not args.no_save_swag: utils.save_checkpoint( args.dir, epoch + 1, name="swag", state_dict=swag_model.state_dict() ) else: utils.save_checkpoint( args.dir, epoch + 1, state_dict=model.state_dict(), optimizer=optimizer.state_dict(), ) time_ep = time.time() - time_ep memory_usage = torch.cuda.memory_allocated() / (1024.0 ** 3) values = [ epoch + 1, lr, train_res["loss"], train_res["accuracy"], test_res["loss"], test_res["accuracy"], time_ep, memory_usage, ] if args.swa: values = values[:-2] + [swag_res["loss"], swag_res["accuracy"]] + values[-2:] table = tabulate.tabulate([values], columns, tablefmt="simple", floatfmt="8.4f") table = table.split("\n") table = "\n".join([table[1]] + table) print(table) if args.epochs % args.save_freq != 0: if args.swa: if not args.no_save_swag: utils.save_checkpoint( args.dir, args.epochs, name="swag", state_dict=swag_model.state_dict() ) else: utils.save_checkpoint(args.dir, args.epochs, state_dict=model.state_dict())
12,515
29.378641
128
py
lgv-geometric-transferability
lgv-geometric-transferability-main/lgv/imagenet/generate_noisy_models.py
import pandas as pd import os from pathlib import Path import argparse import random from tqdm import tqdm import numpy as np import torch from torchvision import models as tmodels import torchvision.datasets as datasets import torchvision.transforms as transforms from utils.data import ImageNet from utils.helpers import guess_and_load_model, guess_model, list_models from utils.pca_weights import model2vector, models2tensor, vector2model from utils.subspace_inference.utils import save_checkpoint def parse_args(): parser = argparse.ArgumentParser(description="Export models with added randomness") parser.add_argument("path_model", help="Path to model") parser.add_argument("--xp", default='gaussian_noise', choices=['gaussian_noise', 'random_1D', 'random_ensemble_equivalent', 'gaussian_subspace'], help="Name of the experiment to run. gaussian_noise to add Gaussian noise to weights (iid noise" " per model). random_1D to export n-models along a single random direction in weights. " "space. random_ensemble_equivalent to export for all models of path-ref-ensemble a model " "at the same L2 distance from path_model but in a random direction") parser.add_argument("--std", default=1, type=float, help="Standard deviation of parameter noise. Used only for 'gaussian_noise' XP") parser.add_argument("--max-norm", default=1, type=float, help="Max 2-norm in weights space to sample equality spaced n-models along the single random direction. Used only for 'random_1D' XP") parser.add_argument("--path-ref-ensemble", default=None, help="Ensemble of reference from which we generate a similar one but with random directions. Used only for 'random_ensemble_equivalent' XP") parser.add_argument("--n-models", default=10, type=int, help="Number of models to export") parser.add_argument("--export-dir", default=None, help="Dir to export models and CSV") parser.add_argument("--update-bn", action='store_true', help="Update BN of produced models") parser.add_argument("--data-path", default=None, help="Path of data. Only supported for ImageNet.") parser.add_argument("--batch-size", type=int, default=128, help="Batch size.") parser.add_argument("--num-workers", type=int, default=10) parser.add_argument("--seed", type=int, default=None, help="Random seed passed to PCA.") args = parser.parse_args() return args def main(args): np.random.seed(args.seed) torch.manual_seed(args.seed) random.seed(args.seed) data = ImageNet(batch_size=args.batch_size, path=args.data_path) traindir = os.path.join(args.data_path, 'train') valdir = os.path.join(args.data_path, 'val') normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) train_dataset = datasets.ImageFolder( traindir, transforms.Compose([ transforms.RandomResizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ])) train_loader = torch.utils.data.DataLoader( train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) val_loader = torch.utils.data.DataLoader( datasets.ImageFolder(valdir, transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ])), batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers, pin_memory=True) loaders = {'train': train_loader, 'test': val_loader} model = guess_and_load_model(args.path_model, data=data, load_as_ghost=False, input_diversity=False, skip_gradient=False) if args.path_ref_ensemble: paths_ref_ensemble = list_models(args.path_ref_ensemble) models_ref_ensemble = [] for i, filename in enumerate(paths_ref_ensemble): model_tmp = guess_and_load_model(filename, data=data, load_as_ghost=False, input_diversity=False, skip_gradient=False) models_ref_ensemble.append(model_tmp) w_ref_ensemble_list = [model2vector(x) for x in models_ref_ensemble] model_cfg = getattr(tmodels, guess_model(args.path_model)) w = model2vector(model) if args.xp == 'gaussian_noise': for i in tqdm(range(args.n_models), desc="Export models"): w_random = w.detach().clone() + torch.randn(w.shape) * args.std model_noisy = vector2model(w_random, model_cfg, update_bn=args.update_bn, train_loader=loaders['train']) save_checkpoint(args.export_dir, name='model_noisy', sample=i, state_dict=model_noisy.state_dict()) elif args.xp == 'random_1D': rand_vect = torch.randn(w.shape) rand_vect = rand_vect / torch.linalg.norm(rand_vect, ord=2) norm_list = np.linspace(0, args.max_norm, num=args.n_models).tolist() for i, norm in enumerate(tqdm(norm_list, desc="Export models")): w_random = w.detach().clone() + rand_vect * norm model_noisy = vector2model(w_random, model_cfg, update_bn=args.update_bn, train_loader=loaders['train']) export_dir = os.path.join(args.export_dir, f'norm_{norm}') save_checkpoint(export_dir, name='model_noisy_1D', sample=i, state_dict=model_noisy.state_dict()) elif args.xp == 'random_ensemble_equivalent': # for each model, we compute the distance b/w w_ref and w, generate a random direction (uniform from unit sphere) # add it to w, update BN, and export for i, w_ref in enumerate(w_ref_ensemble_list): filename_ref = Path(paths_ref_ensemble[i]).stem ref_dist = torch.linalg.norm(w_ref - w, ord=2) rand_vect = torch.randn(w.shape) rand_vect = rand_vect / torch.linalg.norm(rand_vect, ord=2) w_random = w.detach().clone() + rand_vect * ref_dist print(f"Export model {filename_ref}; ||w_ref - w||: {ref_dist}; ||w_random - w||: {torch.linalg.norm(w_random - w, ord=2)} ") model_noisy = vector2model(w_random, model_cfg, update_bn=args.update_bn, train_loader=loaders['train']) save_checkpoint(args.export_dir, name=filename_ref, state_dict=model_noisy.state_dict()) elif args.xp == 'gaussian_subspace': # random gaussian in the LGV deviation subspace n_samples = len(models_ref_ensemble) n_weights = w.shape[0] W = models2tensor(models_ref_ensemble) D = (W-w).T # deviation matrix for i in tqdm(range(args.n_models), desc="Export models"): w_random = w.detach().clone() + 1 / (n_samples-1)**0.5 * (D @ torch.randn(n_samples)) * args.std model_noisy = vector2model(w_random, model_cfg, update_bn=args.update_bn, train_loader=loaders['train']) save_checkpoint(args.export_dir, name='model_gaussian_subspace', sample=i, state_dict=model_noisy.state_dict()) if __name__ == '__main__': args = parse_args() main(args)
7,194
56.103175
201
py
lgv-geometric-transferability
lgv-geometric-transferability-main/lgv/imagenet/hessian/compute_hessian.py
#* # @file Different utility functions # Copyright (c) Zhewei Yao, Amir Gholami # All rights reserved. # This file is part of PyHessian library. # # PyHessian is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # PyHessian is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with PyHessian. If not, see <http://www.gnu.org/licenses/>. #* from __future__ import print_function import argparse from collections import OrderedDict from utils_hessian import * from torchvision.models import resnet50 from pyhessian import hessian # Settings parser = argparse.ArgumentParser(description='PyTorch Example') parser.add_argument( '--mini-hessian-batch-size', type=int, default=200, help='input batch size for mini-hessian batch (default: 200)') parser.add_argument('--hessian-batch-size', type=int, default=200, help='input batch size for hessian (default: 200)') parser.add_argument('--seed', type=int, default=1, help='random seed (default: 1)') parser.add_argument('--cuda', action='store_false', help='do we use gpu or not') parser.add_argument('--resume', type=str, default='', help='get the checkpoint') parser.add_argument('--data-path', type=str, default='', help='get the checkpoint') args = parser.parse_args() # set random seed to reproduce the work torch.manual_seed(args.seed) if args.cuda: torch.cuda.manual_seed(args.seed) for arg in vars(args): print(arg, getattr(args, arg)) # get dataset train_loader, test_loader = getData(name='imagenet_without_dataaugmentation', train_bs=args.mini_hessian_batch_size, test_bs=1, path=args.data_path ) ############## # Get the hessian data ############## assert (args.hessian_batch_size % args.mini_hessian_batch_size == 0) #assert (50000 % args.hessian_batch_size == 0) # TODO batch_num = args.hessian_batch_size // args.mini_hessian_batch_size if batch_num == 1: for inputs, labels in train_loader: hessian_dataloader = (inputs, labels) break else: hessian_dataloader = [] for i, (inputs, labels) in enumerate(train_loader): hessian_dataloader.append((inputs, labels)) if i == batch_num - 1: break # get model model = resnet50() if args.cuda: print('loading to gpu') model = model.cuda() #model = torch.nn.DataParallel(model) criterion = nn.CrossEntropyLoss() # label loss ################### # Get model checkpoint, get saving folder ################### if args.resume == '': raise Exception("please choose the trained model") state_dict = torch.load(args.resume) if 'state_dict' in state_dict: state_dict = state_dict['state_dict'] try: model.load_state_dict(state_dict) except RuntimeError: new_state_dict = OrderedDict() for k, v in state_dict.items(): name = k[7:] # remove `module.` new_state_dict[name] = v # load params model.load_state_dict(new_state_dict) ###################################################### # Begin the computation ###################################################### # turn model to eval mode model.eval() if batch_num == 1: hessian_comp = hessian(model, criterion, data=hessian_dataloader, cuda=args.cuda) else: hessian_comp = hessian(model, criterion, dataloader=hessian_dataloader, cuda=args.cuda) print( '********** finish data loading and begin Hessian computation **********') top_eigenvalues, _ = hessian_comp.eigenvalues() print('\n***Top Eigenvalues: ', top_eigenvalues) trace = hessian_comp.trace() print('\n***Trace: ', np.mean(trace)) #density_eigen, density_weight = hessian_comp.density()
4,571
30.531034
78
py
lgv-geometric-transferability
lgv-geometric-transferability-main/lgv/imagenet/hessian/utils_hessian.py
#* # @file Different utility functions # Copyright (c) Zhewei Yao, Amir Gholami # All rights reserved. # This file is part of PyHessian library. # # PyHessian is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by # the Free Software Foundation, either version 3 of the License, or # (at your option) any later version. # # PyHessian is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with PyHessian. If not, see <http://www.gnu.org/licenses/>. #* import os import numpy as np import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torchvision import datasets, transforms from torch.autograd import Variable def getData(name='cifar10', train_bs=128, test_bs=1000, path=None): """ Get the dataloader """ if name == 'cifar10': transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) trainset = datasets.CIFAR10(root='../data', train=True, download=True, transform=transform_train) train_loader = torch.utils.data.DataLoader(trainset, batch_size=train_bs, shuffle=True) testset = datasets.CIFAR10(root='../data', train=False, download=False, transform=transform_test) test_loader = torch.utils.data.DataLoader(testset, batch_size=test_bs, shuffle=False) if name == 'cifar10_without_dataaugmentation': transform_train = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) trainset = datasets.CIFAR10(root='../data', train=True, download=True, transform=transform_train) train_loader = torch.utils.data.DataLoader(trainset, batch_size=train_bs, shuffle=True) testset = datasets.CIFAR10(root='../data', train=False, download=False, transform=transform_test) test_loader = torch.utils.data.DataLoader(testset, batch_size=test_bs, shuffle=False) if name == 'imagenet_without_dataaugmentation': transform_train = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), ]) transform_test = transforms.Compose([ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)), ]) traindir = os.path.join(path, 'train') trainset = datasets.ImageFolder(root=traindir, transform=transform_train) train_loader = torch.utils.data.DataLoader(trainset, batch_size=train_bs, shuffle=True) testdir = os.path.join(path, 'validation') testset = datasets.ImageFolder(root=testdir, transform=transform_test) test_loader = torch.utils.data.DataLoader(testset, batch_size=test_bs, shuffle=False) return train_loader, test_loader def test(model, test_loader, cuda=True): """ Get the test performance """ model.eval() correct = 0 total_num = 0 for data, target in test_loader: if cuda: data, target = data.cuda(), target.cuda() output = model(data) pred = output.data.max( 1, keepdim=True)[1] # get the index of the max log-probability correct += pred.eq(target.data.view_as(pred)).cpu().sum().item() total_num += len(data) print('testing_correct: ', correct / total_num, '\n') return correct / total_num
5,701
39.728571
75
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/modelsghostpreresnet.py
""" PreResNet model definition ported from https://github.com/bearpaw/pytorch-classification/blob/master/models/cifar/preresnet.py ----- Adapted to add skip connection erosion Do not use to train a model. Only for inference. Train on regular PreResNet """ import torch import torch.nn as nn import torchvision.transforms as transforms import math __all__ = ['PreResNet110', 'PreResNet164'] def conv3x3(in_planes, out_planes, stride=1): return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.bn1 = nn.BatchNorm2d(inplanes) self.relu = nn.ReLU(inplace=True) self.conv1 = conv3x3(inplanes, planes, stride) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = conv3x3(planes, planes) self.downsample = downsample self.stride = stride self.planes = planes def forward(self, x): residual = x out = self.bn1(x) out = self.relu(out) out = self.conv1(out) out = self.bn2(out) out = self.relu(out) out = self.conv2(out) if self.downsample is not None: residual = self.downsample(x) # added for Ghost # random scalar with optimal value reported in the original implementation of Ghost Paper # https://github.com/LiYingwei/ghost-network/blob/master/nets/resnet_v2_50.py#L129 random_range = 0.22 # uniform distribution [1-random_range, 1+random_range] if residual.size(1) != self.planes * self.expansion or len(residual.size()) != 4: raise ValueError('wrong dimension') perturb_var = torch.rand((1, residual.size(1), 1, 1), requires_grad=False, device=residual.device) * random_range * 2 + ( 1 - random_range) out += residual * perturb_var return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.bn1 = nn.BatchNorm2d(inplanes) self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn3 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride self.planes = planes def forward(self, x): residual = x out = self.bn1(x) out = self.relu(out) out = self.conv1(out) out = self.bn2(out) out = self.relu(out) out = self.conv2(out) out = self.bn3(out) out = self.relu(out) out = self.conv3(out) if self.downsample is not None: residual = self.downsample(x) # added for Ghost # random scalar with optimal value reported in the original implementation of Ghost Paper # https://github.com/LiYingwei/ghost-network/blob/master/nets/resnet_v2_50.py#L129 random_range = 0.22 # uniform distribution [1-random_range, 1+random_range] if residual.size(1) != self.planes * self.expansion or len(residual.size()) != 4: raise ValueError('wrong dimension') perturb_var = torch.rand((1,residual.size(1),1,1), requires_grad=False, device=residual.device) * random_range * 2 + (1-random_range) out += residual * perturb_var return out class PreResNet(nn.Module): def __init__(self, num_classes=10, depth=110): super(PreResNet, self).__init__() if depth >= 44: assert (depth - 2) % 9 == 0, 'depth should be 9n+2' n = (depth - 2) // 9 block = Bottleneck else: assert (depth - 2) % 6 == 0, 'depth should be 6n+2' n = (depth - 2) // 6 block = BasicBlock self.inplanes = 16 self.conv1 = nn.Conv2d(3, 16, kernel_size=3, padding=1, bias=False) self.layer1 = self._make_layer(block, 16, n) self.layer2 = self._make_layer(block, 32, n, stride=2) self.layer3 = self._make_layer(block, 64, n, stride=2) self.bn = nn.BatchNorm2d(64 * block.expansion) self.relu = nn.ReLU(inplace=True) self.avgpool = nn.AvgPool2d(8) self.fc = nn.Linear(64 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), ) layers = list() layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.layer1(x) # 32x32 x = self.layer2(x) # 16x16 x = self.layer3(x) # 8x8 x = self.bn(x) x = self.relu(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x class PreResNet110: base = PreResNet args = list() kwargs = {'depth': 110} transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), #transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), #transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) class PreResNet164: base = PreResNet args = list() kwargs = {'depth': 164} transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), #transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), #transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ])
6,999
32.653846
141
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/n_sphere.py
# N-sphere Convert to Spherical or Rectangular Coordination # improve n-sphere package with numerical stability and basic vectorization: https://pypi.org/project/n-sphere/ import numpy as np import math import torch SUPPORTED_TYPES = ['Tensor', 'ndarray', 'list'] def convert_spherical(input, digits=6, tol=1e-8): input_type = type(input).__name__ if input_type not in SUPPORTED_TYPES: raise ValueError("Unsupported type") result = [] if input_type == 'list': input = np.array(input) # over 2-dimension (current available 2-dimension) if (input.ndim == 1): r = 0 for i in range(0, len(input)): r += input[i] * input[i] r = math.sqrt(r) convert = [r] for i in range(0, len(input) - 2): convert.append(round(math.acos(input[i] / (r + tol)), digits)) r = math.sqrt(r * r - input[i] * input[i]) if input[-2] >= 0: convert.append(round(math.acos(input[-2] / (r + tol)), digits)) else: convert.append(round(2 * math.pi - math.acos(input[-2] / (r + tol)), digits)) result = convert if input_type == 'ndarray': result = np.array(result) elif input_type == 'Tensor': result = torch.stack(result) else: result = np.zeros(input.shape) ssq_cum = np.sum(input[:,-2:] ** 2, axis=1) result[:, -1] = np.arccos(input[:,-2] / np.sqrt(ssq_cum + tol)) mask = input[:, -1] < 0 result[mask, -1] = 2 * np.pi - result[mask, -1] for i in range(2, input.shape[1]): ssq_cum = ssq_cum + input[:, -i-1]**2 result[:, -i] = np.arccos(input[:, -i-1] / np.sqrt(ssq_cum + tol)) result[:,0] = np.sqrt(ssq_cum) if input_type == 'Tensor': result = torch.from_numpy(result) return result def convert_rectangular(input, digits=6): input_type = type(input).__name__ if input_type not in SUPPORTED_TYPES: raise ValueError("Unsupported type") if input_type == 'list': input = np.array(input) if input.ndim == 1: result = [] r = input[0] multi_sin = 1 convert = [] if input_type == 'Tensor': for i in range(1, len(input) - 1): convert.append(r * multi_sin * math.cos(input[i])) multi_sin *= math.sin(input[i]) convert.append(r * multi_sin * math.cos(input[-1])) convert.append(r * multi_sin * math.sin(input[-1])) convert = np.array(convert) convert = torch.from_numpy(convert) else: for i in range(1, len(input) - 1): convert.append(round(r * multi_sin * math.cos(input[i]), digits)) multi_sin *= math.sin(input[i]) convert.append(round(r * multi_sin * math.cos(input[-1]), digits)) convert.append(round(r * multi_sin * math.sin(input[-1]), digits)) if input_type != 'list': convert = np.array(convert) result = convert else: # over 2-dimension result = np.zeros(input.shape) r = input[:, 0] multi_sin = np.zeros(input.shape[0]) + 1 for i in range(1, input.shape[1] - 1): result[:, i - 1] = r * multi_sin * np.cos(input[:, i]) multi_sin *= np.sin(input[:, i]) result[:, i] = r * multi_sin * np.cos(input[:, i + 1]) result[:, i + 1] = r * multi_sin * np.sin(input[:, i + 1]) if input_type == 'Tensor': result = torch.from_numpy(result) return result
3,609
37.817204
111
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/pca_weights.py
import torch from sklearn.decomposition import PCA from utils.subspace_inference.utils import flatten, bn_update def model2vector(model): """ Transform a pytorch model into its weight Tensor :param model: pytorch model :return: tensor of size (n_weights,) """ w = flatten([param.detach().cpu() for param in model.parameters()]) return w def models2tensor(models): """ Transform a list of pytorch model into its weight Tensor :param models: list of pytorch model :return: tensor of size (n_models, n_weights) """ n_weights = sum(p.numel() for p in models[0].parameters()) W = torch.empty(0, n_weights, dtype=torch.float32) for model in models: w = model2vector(model) W = torch.cat((W, w.view(1, -1)), dim=0) return W def vector2model(w, model_cfg, update_bn=True, train_loader=None, **kwargs): """ Build a pytorch model from given weight vector :param w: tensor of size (1, n_weights) :param model_cfg: model class :param update_bn: Update or not the BN statistics :param train_loader: Data loader to update BN stats :param kwargs: args passed to model_cfg :return: pytorch model """ if update_bn and not train_loader: raise ValueError('train_loader must be provided with update_bn') w = w.detach().clone() new_model = model_cfg(**kwargs).cuda() offset = 0 for param in new_model.parameters(): param.data.copy_(w[offset:offset + param.numel()].view(param.size()).to('cuda')) offset += param.numel() if update_bn: bn_update(train_loader, new_model, verbose=False, subset=0.1) new_model.eval() return new_model class PcaWeights: def __init__(self, model_cfg, rank=20, train_loader=None, seed=None): """ PCA on pytorch models :param model_cfg: class of the models :param rank: number of PCA components :param train_loader: train loader to update BN of produced models """ self.model_cfg = model_cfg self.rank = rank self.train_loader = train_loader self.pca = PCA(n_components=self.rank, svd_solver='auto', random_state=seed) def fit(self, models): """ Fit PCA to models :param models: list of pytorch model """ num_models = len(models) if self.rank > 0.8*num_models: print('Randomized SVD might not be ideal for rank PCA > 80% number of models') W = models2tensor(models) if W.shape[0] != num_models: raise RuntimeError('Wrong dimension of W') W = W.numpy() self.pca.fit(W) def transform(self, models=None, W=None, components=None): """ Transform a model into the PCA subspace :param models: list of pytorch model :param W: pytorch tensor of the weights. Ignored if models is specified. :param components: list of int corresponding to the components to keep. `None` keep all `rank` components :return: """ if not models and not torch.is_tensor(W): raise ValueError('models or W should be defined') if models: W = models2tensor(models) W = W.numpy() Z = self.pca.transform(W) if components: Z = Z[:, components] return Z def inverse_transform(self, Z, update_bn=True, **kwargs): """ Inverse transform from latent space to model :param Z: Component vector :param update_bn: Update BN layers of models :param kwargs: args passed to model class :return: list of models """ W = self.pca.inverse_transform(Z) W = torch.from_numpy(W) new_models = [] for i in range(W.shape[0]): w = W[i, :] new_model = vector2model(w=w, model_cfg=self.model_cfg, update_bn=update_bn, train_loader=self.train_loader, **kwargs) new_models.append(new_model) return new_models
4,046
34.191304
120
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/optimizers.py
""" File adapted from https://github.com/JavierAntoran/Bayesian-Neural-Networks """ from torch.optim.optimizer import Optimizer, required import numpy as np import torch class SGLD(Optimizer): """ SGLD optimiser based on pytorch's SGD. Note that the weight decay is specified in terms of the gaussian prior sigma. """ def __init__(self, params, lr=1e-2, prior_sigma=np.inf): weight_decay = 1 / (prior_sigma ** 2) if not 0.0 <= lr: raise ValueError("Invalid learning rate: {}".format(lr)) if not 0.0 <= weight_decay: raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) defaults = dict(lr=lr, weight_decay=weight_decay) super(SGLD, self).__init__(params, defaults) @torch.no_grad() def step(self): """ Performs a single optimization step. """ loss = None for group in self.param_groups: weight_decay = group['weight_decay'] for p in group['params']: if p.grad is None: continue d_p = p.grad.data if weight_decay != 0: d_p = d_p.add(p, alpha=weight_decay) langevin_noise = p.data.new(p.data.size()).normal_(mean=0, std=1) / np.sqrt(group['lr']) p.data = p.data.add(0.5 * d_p + langevin_noise, alpha=-group['lr']) return loss class pSGLD(Optimizer): """ RMSprop preconditioned SGLD using pytorch rmsprop implementation. """ def __init__(self, params, lr=1e-2, prior_sigma=np.inf, alpha=0.99, eps=1e-8): weight_decay = 1 / (prior_sigma ** 2) if not 0.0 <= lr: raise ValueError("Invalid learning rate: {}".format(lr)) if not 0.0 <= eps: raise ValueError("Invalid epsilon value: {}".format(eps)) if not 0.0 <= weight_decay: raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) if not 0.0 <= alpha: raise ValueError("Invalid alpha value: {}".format(alpha)) defaults = dict(lr=lr, weight_decay=weight_decay, alpha=alpha, eps=eps) super(pSGLD, self).__init__(params, defaults) @torch.no_grad() def step(self): """ Performs a single optimization step. """ loss = None for group in self.param_groups: weight_decay = group['weight_decay'] for p in group['params']: if p.grad is None: continue d_p = p.grad.data state = self.state[p] if len(state) == 0: state['step'] = 0 state['square_avg'] = torch.zeros_like(p.data) square_avg = state['square_avg'] alpha = group['alpha'] state['step'] += 1 if weight_decay != 0: d_p = d_p.add(p, alpha=weight_decay) # sqavg x alpha + (1-alph) sqavg *(elemwise) sqavg square_avg.mul_(alpha).addcmul_(d_p, d_p, value=1 - alpha) avg = square_avg.sqrt().add_(group['eps']) langevin_noise = p.data.new(p.data.size()).normal_(mean=0, std=1) / np.sqrt(group['lr']) p.data = p.data.add(0.5 * d_p.div_(avg) + langevin_noise / torch.sqrt(avg), alpha=-group['lr']) return loss
3,430
30.190909
111
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/data.py
import os import logging import torch import torchvision import torchvision.datasets as datasets import numpy as np from torchvision import transforms from .helpers import list_models, guess_and_load_model, DEVICE def check_args(method): def inner(ref, **kwargs): if kwargs.get('validation', False) and not ref.valset: raise ValueError('Trying to call validation without creating it') return method(ref, **kwargs) return inner class DataBase: trainset = valset = testset = None transform_train = transform_test = None trainloader = None valloader = None testloader = None transform = None classes = None use_cuda = torch.cuda.is_available() device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") def __init__(self, batch_size=64, num_workers=0, validation=None, normalize=False, seed=None): """ Base class for data :param batch_size: Batch size for data loader. :param num_workers: Number of workers. :param validation: Split the training set into a training set and into a validation set. :param normalize: False, data in [0,1] (default) for adversarial crafting; True normalize data for training. :param seed: Random seed. """ if hasattr(self, 'min_pixel_value') + hasattr(self, 'max_pixel_value') < 2: self.min_pixel_value = 1e8 self.max_pixel_value = -1e8 for images, _ in self.testloader: min_pixel = torch.min(images) max_pixel = torch.max(images) if min_pixel < self.min_pixel_value: self.min_pixel_value = min_pixel if max_pixel > self.max_pixel_value: self.max_pixel_value = max_pixel self.classes = self.testset.classes self.num_classes = len(self.classes) self.batch_size = batch_size self.num_workers = num_workers self.validation = validation self.seed = seed if self.validation: if type(self.validation) != int: raise ValueError('validation should be a int of the size of the val set') if self.seed is None: logging.warning('It is recommended to provided random seed to reproducibility') generator = None else: generator = torch.Generator().manual_seed(self.seed) # create validation set of provided size self.trainset, self.valset = torch.utils.data.random_split(self.trainset, lengths=[len(self.trainset) - self.validation, self.validation], generator=generator) self.trainloader.data = self.trainset # update dataset loader self.valset.transform = self.transform_test # set test transform to val dataset self.valloader = torch.utils.data.DataLoader(self.valset, batch_size=batch_size, pin_memory=self.use_cuda, shuffle=False, num_workers=num_workers) def get_input_shape(self): return tuple(self.trainset.data.shape) @check_args def to_numpy(self, train=False, validation=False, N=None, seed=None): """ Return dataset as numpy array Becareful, data has to be able to be loaded into memory. :param train: bool, train set :param validation: bool, validation set. If train is False, test set. :param N: int, max number of examples to import :return: X, y: numpy arrays """ if train: set = self.trainset elif validation: set = self.valset else: set = self.testset if N is None: N = len(set) if seed: torch.manual_seed(seed) loader = torch.utils.data.DataLoader(set, batch_size=N, shuffle=(train or N < len(set))) load_tmp = next(iter(loader)) X = load_tmp[0].numpy() y = load_tmp[1].numpy() return X, y @check_args def correctly_predicted_to_numpy(self, model=None, models=None, train=False, validation=False, N=None, seed=None): """ Return the examples correcty predicted by model in the dataset as numpy arrays :param model: pytorch model :param models: list of pytorch models :param train: bool, train or test set :param validation: bool, validation set. If train is False, test set. :param N: int, max number of examples to import :param seed: int, fix random seed for reproducibility and select same examples :return: X, y: numpy arrays """ if not model and not models: raise ValueError('model or models should be defined') if model and models: raise ValueError('model and models cannot be both defined') if model: models = [model, ] if train: set = self.trainset elif validation: set = self.valset else: set = self.testset if N is None: N = len(set) if seed: torch.manual_seed(seed) loader = torch.utils.data.DataLoader(set, batch_size=self.batch_size, shuffle=(train or N < len(set))) X = np.zeros((N,) + self.get_input_shape()[1:], dtype='float32') y = np.zeros((N,), dtype='int') nb_images_saved = 0 for images, labels in loader: images_ = images.to(DEVICE) idx_ok = torch.zeros(len(images)) for model in models: outputs = model(images_) _, predicted = torch.max(outputs.data, 1) predicted = predicted.cpu() idx_ok += predicted == labels idx_ok = idx_ok == len(models) images_ok = images[idx_ok,] labels_ok = labels[idx_ok,] nb_images_to_append = min(idx_ok.sum().item(), N-nb_images_saved) X[nb_images_saved:(nb_images_saved+nb_images_to_append),] = images_ok[0:nb_images_to_append,].numpy() y[nb_images_saved:nb_images_saved+nb_images_to_append] = labels_ok[0:nb_images_to_append].numpy() nb_images_saved += nb_images_to_append if nb_images_saved >= N: break X = X[0:nb_images_saved,] y = y[0:nb_images_saved,] if not (X.shape[0] == y.shape[0] <= N): raise RuntimeError("Array shape unexpected") if X.shape[0] < N < len(set): logging.warning('Number of examples lower than requested') return X, y @check_args def compute_accuracy(self, model, train=False, validation=False): """ Compute the accuracy on the test or train data :param model: Pytorch NN :param train: compute on the train set :param train: compute on the validation set. If train is False, test set. :return: float """ if train: loader = self.trainloader elif validation: loader = self.valloader else: loader = self.testloader model.eval() correct = 0 total = 0 with torch.no_grad(): for batch in loader: inputs, labels = batch[0].to(self.device, non_blocking=True), batch[1].to(self.device, non_blocking=True) outputs = model(inputs) _, predicted = torch.max(outputs.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() return correct / total class MNIST(DataBase): def __init__(self, batch_size, num_workers=0, path='data', validation=None, normalize=False, seed=None): if normalize: transform = transforms.Compose( [transforms.ToTensor(), transforms.Normalize(mean=(0.1307,), std=(0.3081,))]) else: transform = transforms.Compose([transforms.ToTensor(),]) self.transform_train = self.transform_test = self.transform = transform self.trainset = torchvision.datasets.MNIST(root=path, train=True, download=True, transform=self.transform) self.trainloader = torch.utils.data.DataLoader(self.trainset, batch_size=batch_size, pin_memory=self.use_cuda, shuffle=True, num_workers=num_workers) self.testset = torchvision.datasets.MNIST(root=path, train=False, download=True, transform=self.transform) self.testloader = torch.utils.data.DataLoader(self.testset, batch_size=batch_size, pin_memory=self.use_cuda, shuffle=False, num_workers=num_workers) super().__init__(batch_size=batch_size, num_workers=num_workers, validation=validation, seed=seed) def get_input_shape(self): return (1, 1, 28, 28) class CIFAR10(DataBase): def __init__(self, batch_size, num_workers=0, path='data', validation=None, normalize=False, seed=None): if normalize: normalize_transform = transforms.Normalize(mean=(0.49139968, 0.48215841, 0.44653091), std=(0.24703223, 0.24348513, 0.26158784)) else: normalize_transform = transforms.Normalize(mean=(0., 0., 0.), std=(1., 1., 1.)) #normalize = transforms.Normalize(mean=(0.5, 0.5, 0.5), # std=(0.5, 0.5, 0.5)) self.transform_train = transforms.Compose( [transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize_transform]) self.transform_test = transforms.Compose( [transforms.ToTensor(), normalize_transform]) self.trainset = torchvision.datasets.CIFAR10(root=path, train=True, download=True, transform=self.transform_train) self.trainloader = torch.utils.data.DataLoader(self.trainset, batch_size=batch_size, pin_memory=self.use_cuda, shuffle=True, num_workers=num_workers) self.testset = torchvision.datasets.CIFAR10(root=path, train=False, download=True, transform=self.transform_test) self.testloader = torch.utils.data.DataLoader(self.testset, batch_size=batch_size, pin_memory=self.use_cuda, shuffle=False, num_workers=num_workers) super().__init__(batch_size=batch_size, num_workers=num_workers, validation=validation, seed=seed) def get_input_shape(self): return (1, 3, 32, 32) class CIFAR100(DataBase): def __init__(self, batch_size, num_workers=0, path='data', validation=None, normalize=False, seed=None): if normalize: normalize_transform = transforms.Normalize((0.50707516, 0.48654887, 0.44091784), (0.26733429, 0.25643846, 0.27615047)) else: normalize_transform = transforms.Normalize((0., 0., 0.), (1., 1., 1.)) self.transform_train = transforms.Compose( [transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize_transform]) self.transform_test = transforms.Compose( [transforms.ToTensor(), normalize_transform]) self.trainset = torchvision.datasets.CIFAR100(root=path, train=True, download=True, transform=self.transform_train) self.trainloader = torch.utils.data.DataLoader(self.trainset, batch_size=batch_size, pin_memory=self.use_cuda, shuffle=True, num_workers=num_workers) self.testset = torchvision.datasets.CIFAR100(root=path, train=False, download=True, transform=self.transform_test) self.testloader = torch.utils.data.DataLoader(self.testset, batch_size=batch_size, pin_memory=self.use_cuda, shuffle=False, num_workers=num_workers) super().__init__(batch_size=batch_size, num_workers=num_workers, validation=validation, seed=seed) class ImageNet(DataBase): def __init__(self, batch_size, path='/work/projects/bigdata_sets/ImageNet/ILSVRC2012/raw-data/', num_workers=0, validation=None, normalize=False, input_size=224, resize_size=256, seed=None): # default: input_size=224, resize_size=256 # inception: input_size=299, resize_size=342 # if DATAPATH env var is set, overright default value self.input_size = input_size self.resize_size = resize_size if os.environ.get('DATAPATH', False): path = os.environ.get('DATAPATH') if normalize: print('Loading ImageNet with normalization.') normalize_transform = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) else: normalize_transform = transforms.Normalize(mean=[0., 0., 0.], std=[1., 1., 1.]) traindir = os.path.join(path, 'train') self.transform_train = transforms.Compose([ transforms.RandomResizedCrop(input_size), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize_transform, ]) self.trainset = datasets.ImageFolder(traindir, self.transform_train) self.trainloader = torch.utils.data.DataLoader( self.trainset, batch_size=batch_size, shuffle=True, num_workers=num_workers, pin_memory=self.use_cuda) testdir = os.path.join(path, 'validation') self.transform_test = transforms.Compose([ transforms.Resize(resize_size), transforms.CenterCrop(input_size), transforms.ToTensor(), normalize_transform, ]) self.testset = datasets.ImageFolder(testdir, self.transform_test) self.testloader = torch.utils.data.DataLoader( self.testset, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=self.use_cuda) self.min_pixel_value = 0 self.max_pixel_value = 1 super().__init__(batch_size=batch_size, num_workers=num_workers, validation=validation, seed=seed) def get_input_shape(self): return (1, 3, self.input_size, self.input_size)
14,626
45.582803
130
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/utils_sgm.py
""" Code from the following paper: @inproceedings{wu2020skip, title={Skip connections matter: On the transferability of adversarial examples generated with resnets}, author={Wu, Dongxian and Wang, Yisen and Xia, Shu-Tao and Bailey, James and Ma, Xingjun}, booktitle={ICLR}, year={2020} } https://github.com/csdongxian/skip-connections-matter """ import numpy as np import torch import torch.nn as nn def backward_hook(gamma): # implement SGM through grad through ReLU def _backward_hook(module, grad_in, grad_out): if isinstance(module, nn.ReLU): return (gamma * grad_in[0],) return _backward_hook def backward_hook_norm(module, grad_in, grad_out): # normalize the gradient to avoid gradient explosion or vanish grad_in_norm = [] for i, grad_in_i in enumerate(grad_in): std = torch.std(grad_in_i) grad_in_norm.append(grad_in_i / std) return tuple(grad_in_norm) def register_hook_for_resnet(model, arch, gamma): # There is only 1 ReLU in Conv module of ResNet-18/34 # and 2 ReLU in Conv module ResNet-50/101/152 if arch in ['resnet50', 'resnet101', 'resnet152']: gamma = np.power(gamma, 0.5) backward_hook_sgm = backward_hook(gamma) for name, module in model.named_modules(): if 'relu' in name and not '0.relu' in name: # only to the last ReLU of each layer module.register_backward_hook(backward_hook_sgm) # e.g., 1.layer1.1, 1.layer4.2, ... # if len(name.split('.')) == 3: if len(name.split('.')) >= 2 and 'layer' in name.split('.')[-2]: module.register_backward_hook(backward_hook_norm) def register_hook_for_preresnet(model, arch, gamma): # There are 3 ReLU in Conv module of PreResNet110 if arch in ['PreResNet110']: gamma = np.power(gamma, 1/3) else: raise ValueError('Arch not supported') backward_hook_sgm = backward_hook(gamma) for name, module in model.named_modules(): if 'relu' in name and 'layer' in name and '.0.relu' not in name: # .0.relu to skip the relu of downsampling ('.0.' is important to keep layerX.10.relu) # do not apply to last relu 'relu' module.register_backward_hook(backward_hook_sgm) # e.g., 1.layer1.1, 1.layer4.2, ... # if len(name.split('.')) == 3: if len(name.split('.')) >= 2 and 'layer' in name.split('.')[-2]: module.register_backward_hook(backward_hook_norm) def register_hook_for_densenet(model, arch, gamma): # There are 2 ReLU in Conv module of DenseNet-121/169/201. gamma = np.power(gamma, 0.5) backward_hook_sgm = backward_hook(gamma) for name, module in model.named_modules(): if 'relu' in name and not 'transition' in name: module.register_backward_hook(backward_hook_sgm)
2,861
35.692308
107
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/layers.py
import torch from PIL import Image from torchvision.transforms import functional as F class RandomResizePad(torch.nn.Module): def __init__(self, min_resize): super().__init__() self.min_resize = min_resize def forward(self, img): size_original = img.size() if size_original[-1] != size_original[-2]: raise ValueError("Only squared images supported") random_size = int(torch.randint(low=self.min_resize, high=size_original[-1], size=(1,))) pad_margin = size_original[-1] - random_size img = F.resize(img, size=random_size, interpolation=Image.NEAREST) pad_top = int(torch.randint(low=0, high=pad_margin, size=(1,))) pad_bottom = pad_margin - pad_top pad_right = int(torch.randint(low=0, high=pad_margin, size=(1,))) pad_left = pad_margin - pad_right img = F.pad(img, padding=[pad_left, pad_top, pad_right, pad_bottom], fill=0) if img.size() != size_original: raise RuntimeError("Output size is not the same than input size.") return img
1,083
40.692308
96
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/modelsghost.py
# adapted from torchvision ResNet implementation https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py # Add skip connection erosion # do not use to train a model. Only for inference. Train on regular torchvision resnet import torch from torch import Tensor import torch.nn as nn try: from torchvision.models.utils import load_state_dict_from_url except ImportError: from torch.hub import load_state_dict_from_url from typing import Type, Any, Callable, Union, List, Optional __all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152', 'resnext50_32x4d', 'resnext101_32x8d', 'wide_resnet50_2', 'wide_resnet101_2'] model_urls = { 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', 'resnext50_32x4d': 'https://download.pytorch.org/models/resnext50_32x4d-7cdf4587.pth', 'resnext101_32x8d': 'https://download.pytorch.org/models/resnext101_32x8d-8ba56ff5.pth', 'wide_resnet50_2': 'https://download.pytorch.org/models/wide_resnet50_2-95faca4d.pth', 'wide_resnet101_2': 'https://download.pytorch.org/models/wide_resnet101_2-32ee1156.pth', } def conv3x3(in_planes: int, out_planes: int, stride: int = 1, groups: int = 1, dilation: int = 1) -> nn.Conv2d: """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=dilation, groups=groups, bias=False, dilation=dilation) def conv1x1(in_planes: int, out_planes: int, stride: int = 1) -> nn.Conv2d: """1x1 convolution""" return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) class BasicBlock(nn.Module): expansion: int = 1 def __init__( self, inplanes: int, planes: int, stride: int = 1, downsample: Optional[nn.Module] = None, groups: int = 1, base_width: int = 64, dilation: int = 1, norm_layer: Optional[Callable[..., nn.Module]] = None ) -> None: super(BasicBlock, self).__init__() if norm_layer is None: norm_layer = nn.BatchNorm2d if groups != 1 or base_width != 64: raise ValueError('BasicBlock only supports groups=1 and base_width=64') if dilation > 1: raise NotImplementedError("Dilation > 1 not supported in BasicBlock") # Both self.conv1 and self.downsample layers downsample the input when stride != 1 self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = norm_layer(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = norm_layer(planes) self.downsample = downsample self.stride = stride def forward(self, x: Tensor) -> Tensor: identity = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: identity = self.downsample(x) # added for Ghost # random scalar with optimal value reported in the original implementation of Ghost Paper # https://github.com/LiYingwei/ghost-network/blob/master/nets/resnet_v2_50.py#L129 random_range = 0.22 # uniform distribution [1-random_range, 1+random_range] if identity.size(1) != self.planes * self.expansion or len(identity.size()) != 4: raise ValueError('wrong dimension') perturb_var = torch.rand((1, identity.size(1), 1, 1), requires_grad=False, device=identity.device) * random_range * 2 + ( 1 - random_range) out += identity * perturb_var out = self.relu(out) return out class Bottleneck(nn.Module): # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2) # while original implementation places the stride at the first 1x1 convolution(self.conv1) # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385. # This variant is also known as ResNet V1.5 and improves accuracy according to # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch. expansion: int = 4 def __init__( self, inplanes: int, planes: int, stride: int = 1, downsample: Optional[nn.Module] = None, groups: int = 1, base_width: int = 64, dilation: int = 1, norm_layer: Optional[Callable[..., nn.Module]] = None ) -> None: super(Bottleneck, self).__init__() if norm_layer is None: norm_layer = nn.BatchNorm2d width = int(planes * (base_width / 64.)) * groups # Both self.conv2 and self.downsample layers downsample the input when stride != 1 self.conv1 = conv1x1(inplanes, width) self.bn1 = norm_layer(width) self.conv2 = conv3x3(width, width, stride, groups, dilation) self.bn2 = norm_layer(width) self.conv3 = conv1x1(width, planes * self.expansion) self.bn3 = norm_layer(planes * self.expansion) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride self.planes = planes def forward(self, x: Tensor) -> Tensor: identity = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: identity = self.downsample(x) # added for Ghost # random scalar with optimal value reported in the original implementation of Ghost Paper # https://github.com/LiYingwei/ghost-network/blob/master/nets/resnet_v2_50.py#L129 random_range = 0.22 # uniform distribution [1-random_range, 1+random_range] if identity.size(1) != self.planes * self.expansion or len(identity.size()) != 4: raise ValueError('wrong dimension') perturb_var = torch.rand((1,identity.size(1),1,1), requires_grad=False, device=identity.device) * random_range * 2 + (1-random_range) out += identity * perturb_var out = self.relu(out) return out class ResNet(nn.Module): def __init__( self, block: Type[Union[BasicBlock, Bottleneck]], layers: List[int], num_classes: int = 1000, zero_init_residual: bool = False, groups: int = 1, width_per_group: int = 64, replace_stride_with_dilation: Optional[List[bool]] = None, norm_layer: Optional[Callable[..., nn.Module]] = None ) -> None: super(ResNet, self).__init__() if norm_layer is None: norm_layer = nn.BatchNorm2d self._norm_layer = norm_layer self.inplanes = 64 self.dilation = 1 if replace_stride_with_dilation is None: # each element in the tuple indicates if we should replace # the 2x2 stride with a dilated convolution instead replace_stride_with_dilation = [False, False, False] if len(replace_stride_with_dilation) != 3: raise ValueError("replace_stride_with_dilation should be None " "or a 3-element tuple, got {}".format(replace_stride_with_dilation)) self.groups = groups self.base_width = width_per_group self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = norm_layer(self.inplanes) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dilate=replace_stride_with_dilation[0]) self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dilate=replace_stride_with_dilation[1]) self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dilate=replace_stride_with_dilation[2]) self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) self.fc = nn.Linear(512 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) # Zero-initialize the last BN in each residual branch, # so that the residual branch starts with zeros, and each residual block behaves like an identity. # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 if zero_init_residual: for m in self.modules(): if isinstance(m, Bottleneck): nn.init.constant_(m.bn3.weight, 0) # type: ignore[arg-type] elif isinstance(m, BasicBlock): nn.init.constant_(m.bn2.weight, 0) # type: ignore[arg-type] def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int, stride: int = 1, dilate: bool = False) -> nn.Sequential: norm_layer = self._norm_layer downsample = None previous_dilation = self.dilation if dilate: self.dilation *= stride stride = 1 if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( conv1x1(self.inplanes, planes * block.expansion, stride), norm_layer(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample, self.groups, self.base_width, previous_dilation, norm_layer)) self.inplanes = planes * block.expansion for _ in range(1, blocks): layers.append(block(self.inplanes, planes, groups=self.groups, base_width=self.base_width, dilation=self.dilation, norm_layer=norm_layer)) return nn.Sequential(*layers) def _forward_impl(self, x: Tensor) -> Tensor: # See note [TorchScript super()] x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.avgpool(x) x = torch.flatten(x, 1) x = self.fc(x) return x def forward(self, x: Tensor) -> Tensor: return self._forward_impl(x) def _resnet( arch: str, block: Type[Union[BasicBlock, Bottleneck]], layers: List[int], pretrained: bool, progress: bool, **kwargs: Any ) -> ResNet: model = ResNet(block, layers, **kwargs) if pretrained: state_dict = load_state_dict_from_url(model_urls[arch], progress=progress) model.load_state_dict(state_dict) return model def resnet18(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""ResNet-18 model from `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ return _resnet('resnet18', BasicBlock, [2, 2, 2, 2], pretrained, progress, **kwargs) def resnet34(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""ResNet-34 model from `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ return _resnet('resnet34', BasicBlock, [3, 4, 6, 3], pretrained, progress, **kwargs) def resnet50(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""ResNet-50 model from `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ return _resnet('resnet50', Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs) def resnet101(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""ResNet-101 model from `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ return _resnet('resnet101', Bottleneck, [3, 4, 23, 3], pretrained, progress, **kwargs) def resnet152(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""ResNet-152 model from `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ return _resnet('resnet152', Bottleneck, [3, 8, 36, 3], pretrained, progress, **kwargs) def resnext50_32x4d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""ResNeXt-50 32x4d model from `"Aggregated Residual Transformation for Deep Neural Networks" <https://arxiv.org/pdf/1611.05431.pdf>`_. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ kwargs['groups'] = 32 kwargs['width_per_group'] = 4 return _resnet('resnext50_32x4d', Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs) def resnext101_32x8d(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""ResNeXt-101 32x8d model from `"Aggregated Residual Transformation for Deep Neural Networks" <https://arxiv.org/pdf/1611.05431.pdf>`_. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ kwargs['groups'] = 32 kwargs['width_per_group'] = 8 return _resnet('resnext101_32x8d', Bottleneck, [3, 4, 23, 3], pretrained, progress, **kwargs) def wide_resnet50_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""Wide ResNet-50-2 model from `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_. The model is the same as ResNet except for the bottleneck number of channels which is twice larger in every block. The number of channels in outer 1x1 convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048 channels, and in Wide ResNet-50-2 has 2048-1024-2048. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ kwargs['width_per_group'] = 64 * 2 return _resnet('wide_resnet50_2', Bottleneck, [3, 4, 6, 3], pretrained, progress, **kwargs) def wide_resnet101_2(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet: r"""Wide ResNet-101-2 model from `"Wide Residual Networks" <https://arxiv.org/pdf/1605.07146.pdf>`_. The model is the same as ResNet except for the bottleneck number of channels which is twice larger in every block. The number of channels in outer 1x1 convolutions is the same, e.g. last block in ResNet-50 has 2048-512-2048 channels, and in Wide ResNet-50-2 has 2048-1024-2048. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet progress (bool): If True, displays a progress bar of the download to stderr """ kwargs['width_per_group'] = 64 * 2 return _resnet('wide_resnet101_2', Bottleneck, [3, 4, 23, 3], pretrained, progress, **kwargs)
17,173
40.383133
141
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/models.py
import torch from torch import nn import torch.nn.functional as F from random import randrange, shuffle class ModelWithTemperature(nn.Module): """ A thin decorator, which wraps a model with temperature scaling. Code adapted from https://github.com/gpleiss/temperature_scaling/blob/master/temperature_scaling.py model (nn.Module): A classification neural network NB: Output of the neural network should be the classification logits, NOT the softmax (or log softmax)! """ def __init__(self, model, temperature=1.): super(ModelWithTemperature, self).__init__() self.model = model self.temperature = torch.tensor(temperature) def forward(self, input): logits = self.model(input) return self.temperature_scale(logits) def temperature_scale(self, logits): """ Perform temperature scaling on logits """ # Expand temperature to match the size of logits temperature = self.temperature.expand(logits.size(0), logits.size(1)) return logits / temperature def to(self, *args, **kwargs): self.model = self.model.to(*args, **kwargs) self.temperature = self.temperature.to(*args, **kwargs) super(ModelWithTemperature, self).to(*args, **kwargs) class MnistFc(nn.Module): def __init__(self, num_classes=10, pretrained=False): if pretrained: raise NotImplementedError() super(MnistFc, self).__init__() self.fc1 = nn.Linear(784, 1200) self.fc2 = nn.Linear(1200, 1200) self.fc3 = nn.Linear(1200, num_classes) def forward(self, x): x = torch.flatten(x, 1) x = self.fc1(x) x = F.relu(x) x = self.fc2(x) x = F.relu(x) output = self.fc3(x) return output class MnistSmallFc(nn.Module): """ One hidden layer FC """ def __init__(self, num_classes=10, pretrained=False, hidden_size=512): if pretrained: raise NotImplementedError() super(MnistSmallFc, self).__init__() self.hidden_size = hidden_size self.fc1 = nn.Linear(784, hidden_size) self.fc2 = nn.Linear(hidden_size, num_classes) def forward(self, x): x = torch.flatten(x, 1) x = self.fc1(x) x = F.relu(x) output = self.fc2(x) return output class MnistCnn(nn.Module): def __init__(self, num_classes=10, pretrained=False): if pretrained: raise NotImplementedError() super(MnistCnn, self).__init__() self.conv1 = nn.Conv2d(1, 32, kernel_size=3) self.conv2 = nn.Conv2d(32, 64, kernel_size=5) self.fc1 = nn.Linear(1024, 200) self.fc2 = nn.Linear(200, 200) self.fc3 = nn.Linear(200, num_classes) def forward(self, x): x = self.conv1(x) x = F.relu(x) x = F.max_pool2d(x, 2) x = self.conv2(x) x = F.relu(x) x = F.max_pool2d(x, 2) x = torch.flatten(x, 1) x = self.fc1(x) x = F.relu(x) x = self.fc2(x) x = F.relu(x) output = self.fc3(x) return output class CifarLeNet(nn.Module): def __init__(self, num_classes=10, pretrained=False): super(CifarLeNet, self).__init__() self.conv_1 = nn.Conv2d(3, 6, 5) self.conv_2 = nn.Conv2d(6, 16, 5) self.fc_1 = nn.Linear(16 * 5 * 5, 120) self.fc_2 = nn.Linear(120, 84) self.fc_3 = nn.Linear(84, num_classes) def forward(self, x): out = F.relu(self.conv_1(x)) out = F.max_pool2d(out, 2) out = F.relu(self.conv_2(out)) out = F.max_pool2d(out, 2) out = out.view(out.size(0), -1) out = F.relu(self.fc_1(out)) out = F.relu(self.fc_2(out)) out = self.fc_3(out) return out class TorchEnsemble(nn.Module): def __init__(self, models, ensemble_logits=False): """ :param models: list of pytorch models to ensemble :param ensemble_logits: True if ensemble logits, False to ensemble probabilities :return probablities if ensemble_logits is False, logits if True """ super(TorchEnsemble, self).__init__() if len(models) < 1: raise ValueError('Empty list of models') self.models = nn.ModuleList(models) self.ensemble_logits = ensemble_logits self.softmax = nn.Softmax(dim=1) def forward(self, x): # clone to make sure x is not changed by inplace methods if not self.ensemble_logits: x_list = [self.softmax(model(x.clone())) for model in self.models] # probs else: x_list = [model(x.clone()) for model in self.models] # logits x = torch.stack(x_list) # concat on dim 0 x = torch.mean(x, dim=0, keepdim=False) #for model in self.models: # xi = model(x.clone()) # clone to make sure x is not changed by inplace methods # x1 = self.modelA(x.clone()) # clone to make sure x is not changed by inplace methods # x1 = x1.view(x1.size(0), -1) # x2 = self.modelB(x) # x2 = x2.view(x2.size(0), -1) # x = torch.cat((x1, x2), dim=1) # x = self.classifier(F.relu(x)) return x class LightEnsemble(nn.Module): def __init__(self, models, order=None): """ Perform a single forward pass to one of the models when call forward() :param models: list of pytorch models to ensemble :param random: str, 'random' draw a model with replacement, 'shuffle' draw a model w/o replacement, None cycle in provided order (Default). :return logits """ super(LightEnsemble, self).__init__() self.n_models = len(models) if self.n_models < 1: raise ValueError('Empty list of models') if order == 'shuffle': shuffle(models) elif order in [None, 'random']: pass else: raise ValueError('Not supported order') self.models = nn.ModuleList(models) self.order = order self.f_count = 0 def forward(self, x): if self.order == 'random': index = randrange(0, self.n_models) else: index = self.f_count % self.n_models x = self.models[index](x) self.f_count += 1 return x class LightNestedEnsemble(nn.Module): def __init__(self, list_models, order=None): """ Perform ensemble a single list of models when call forward() :param models: nested list of pytorch models to ensemble :param random: str, 'random' draw a model with replacement, 'shuffle' draw a model w/o replacement, None cycle in provided order (Default). :return logits """ super(LightNestedEnsemble, self).__init__() self.n_ensembles = len(list_models) if self.n_ensembles < 1: raise ValueError('Empty list of models') if order == 'shuffle': shuffle(list_models) elif order in [None, 'random']: pass else: raise ValueError('Not supported order') self.models = nn.ModuleList([TorchEnsemble(models=x, ensemble_logits=True) for x in list_models]) self.order = order self.f_count = 0 def forward(self, x): if self.order == 'random': index = randrange(0, self.n_ensembles) else: index = self.f_count % self.n_ensembles x = self.models[index](x) self.f_count += 1 return x
7,591
32.444934
147
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/helpers.py
import os import re import glob import argparse import torch import numpy as np from collections import OrderedDict try: from art.classifiers import PyTorchClassifier except ModuleNotFoundError: from art.estimators.classification import PyTorchClassifier from .models import TorchEnsemble, CifarLeNet, MnistCnn, MnistFc, MnistSmallFc, ModelWithTemperature from .layers import RandomResizePad from .utils_sgm import register_hook_for_resnet, register_hook_for_preresnet, register_hook_for_densenet from pytorch_ensembles import models as pemodels from utils import modelsghost as ghostmodels from utils import modelsghostpreresnet as ghostpreresnet from torchvision import models as tvmodels import timm from robustbench.utils import load_model as rb_load_model from torchvision import transforms from torch import nn MCMC_OPTIMIZERS = ['SGLD', 'pSGLD'] USE_CUDA = torch.cuda.is_available() DEVICE = torch.device("cuda:0" if USE_CUDA else "cpu") PEMODELS_NAMES = ['BayesPreResNet110', 'PreResNet110', 'PreResNet164', 'VGG16BN', 'VGG19BN', 'WideResNet28x10', 'WideResNet28x10do'] TVMODELS_NAMES = ['BayesResNet50', "resnet50", "resnet18", "resnet152", "resnext50_32x4d", "mnasnet1_0", "densenet121", "densenet201", "mobilenet_v2", "wide_resnet50_2", "vgg19", "inception_v3", "inception_v2", "googlenet"] TIMODELS_NAMES = ["efficientnet_b0", "adv_inception_v3", "inception_resnet_v2", "tf_inception_v3", "timm_resnet152", "timm_resnext50_32x4d", "timm_wide_resnet50_2", "vit_base_patch16_224"] # timm_vgg19 timm_densenet201 timm_inception_v3 are imported from torchvision. do not use RBMODELS_NAMES = ["Salman2020Do_50_2", "Salman2020Do_R50", "Engstrom2019Robustness", "Wong2020Fast", "Salman2020Do_R18"] ALL_MODELS_NAMES = ['MnistFc', 'MnistSmallFc', 'MnistCnn', 'LeNet'] + PEMODELS_NAMES + TVMODELS_NAMES + TIMODELS_NAMES + RBMODELS_NAMES class keyvalue(argparse.Action): """ Parse key-values command line argument Code from https://www.geeksforgeeks.org/python-key-value-pair-using-argparse/ """ # Constructor calling def __call__(self, parser, namespace, values, option_string=None): setattr(namespace, self.dest, dict()) for value in values: # split it into key and value key, value = value.split('=') # assign into dictionary getattr(namespace, self.dest)[key] = value def args2paths(args, index_model=None): """ Create a string corresponding to path to save current model :param args: train.py command line arguments namespace :param index_model: int for a specific element of ensemble models or for a model sample :return: set of pytorch model path and metrics CSV file path """ if args.optimizer in MCMC_OPTIMIZERS: filename = f'{index_model:04}.pth' filename_metrics = 'metrics.csv' relative_path = f'{args.dataset}/{args.architecture}/mcmc_samples/{args.optimizer}_bs{args.batch_size}_lr{args.lr}_lrd{"plateau" if args.lr_decay_on_plateau else args.lr_decay}_psig{args.prior_sigma:.1f}_s{args.samples}_si{args.sampling_interval}_bi{args.burnin}_seed{args.seed}' else: # DNN if index_model is None: # single DNN filename = f'model_{args.optimizer}_bs{args.batch_size}_lr{args.lr}_lrd{"plateau" if args.lr_decay_on_plateau else args.lr_decay}_psig{args.prior_sigma:.1f}_ep{args.epochs}_seed{args.seed}.pth' filename_metrics = filename.replace('model', 'metrics').replace('.pth', '.csv') relative_path = f'{args.dataset}/{args.architecture}/single_model/' else: # Ensemble of DNN filename = f'{index_model:04}.pth' filename_metrics = 'metrics.csv' relative_path = f'{args.dataset}/{args.architecture}/dnn_ensemble/{args.optimizer}_bs{args.batch_size}_lr{args.lr}_lrd{"plateau" if args.lr_decay_on_plateau else args.lr_decay}_psig{args.prior_sigma:.1f}_ep{args.epochs}_seed{args.seed}' os.makedirs(os.path.join(args.output, relative_path), exist_ok=True) return os.path.join(args.output, relative_path, filename), \ os.path.join(args.output, relative_path, filename_metrics) def list_models(models_dir): # pretrained model if re.match(r"^ImageNet/pretrained/(\w+)$", models_dir): return [models_dir, ] # path to single model if re.match('.+\\.pth?(\\.tar)?$', models_dir): if not os.path.isfile(models_dir): raise ValueError('Non-existing path surrogate file passed') return [models_dir, ] # directory of models path_models = glob.glob(f'{models_dir}/*.pt') path_models.extend(glob.glob(f'{models_dir}/*.pth')) path_models.extend(glob.glob(f'{models_dir}/*.pt.tar')) path_models.extend(glob.glob(f'{models_dir}/*.pth.tar')) path_models = sorted(path_models) return path_models def load_model(path_model, class_model, *args, **kwargs): model = class_model(*args, **kwargs) model.load_state_dict(torch.load(path_model)) model.eval() return model def load_list_models(models_dir, class_model, device=None, *args, **kwargs): path_models = list_models(models_dir) models = [] for path_model in path_models: model = load_model(path_model=path_model, class_model=class_model) if device: model.to(device) models.append(model) return models def guess_model(path_model): """ Return the name of the model """ candidates = [x for x in ALL_MODELS_NAMES if x in path_model] if len(candidates) == 1: return candidates[0] elif len(candidates) > 1: # pick the longest one return max(candidates, key=len) raise ValueError('Not able to guess model name') def guess_and_load_model(path_model, data, load_as_ghost=False, input_diversity=False, skip_gradient=False, defense_randomization=False, temperature=None, force_cpu=False): """ Load model from its path only (guessing the model class) :param path_model: str, path to the pt file to load :param data: data class :param load_as_ghost: load model as a Ghost Network. Only Skip Connection Erosion of resnet on ImageNet supported. :param input_diversity: add input diversity as first layer (p=0.5) :param skip_gradient: apply Skip Gradient Method with backward hook (gamma=0.5) :param defense_randomization: add input diversity as first layer (p=1) to be used as defense :param temperature: temperature scaling value. Deactivated if None (default). :param force_cpu: don't send model to GPU if True :return: pytorch instance of a model """ if load_as_ghost and not (('resnet' in path_model and 'ImageNet' in path_model) or ('PreResNet' in path_model and 'CIFAR10' in path_model)): raise ValueError('Ghost Networks only supports resnet on ImageNet, or PreResNet on CIFAR10.') if input_diversity and defense_randomization: raise ValueError('input_diversity and defense_randomization should not be set at the same time') if 'MNIST' in path_model: if load_as_ghost: raise NotImplementedError('Ghost MNIST models not supported') # model from utils/models.py if 'CNN' in path_model or 'MnistCnn' in path_model: model_name_list = ['MnistCnn'] model = MnistCnn(pretrained=False, num_classes=data.num_classes) elif 'FC' in path_model or 'MnistFc' in path_model: model_name_list = ['MnistFc'] model = MnistFc(pretrained=False, num_classes=data.num_classes) elif 'MnistSmallFc' in path_model: model_name_list = ['MnistSmallFc'] model = MnistSmallFc(pretrained=False, num_classes=data.num_classes) else: raise NotImplementedError('Model class unknown') model_loaded = torch.load(path_model, map_location=DEVICE) if 'model_state' in model_loaded: model_loaded = model_loaded['model_state'] try: model.load_state_dict(model_loaded) except RuntimeError: new_state_dict = OrderedDict() for k, v in model_loaded.items(): name = k[2:] # remove `1.` new_state_dict[name] = v # load params model.load_state_dict(new_state_dict) model = add_normalization_layer(model=model, mean=(0.1307,), std=(0.3081,)) arch = model_name_list[0] elif 'CIFAR' in path_model: # ghost models if load_as_ghost: if 'PreResNet' not in path_model: raise ValueError('Only PreResNet ghost models supported on CIFAR10') model_name_list = [x for x in PEMODELS_NAMES if x in path_model] if len(model_name_list) != 1: raise ValueError(f'Failed to extract model name: {model_name_list}') arch = getattr(ghostpreresnet, model_name_list[0]) model = arch.base(num_classes=data.num_classes, **arch.kwargs) # model from utils/models.py elif 'LeNet' in path_model: model_name_list = ['LeNet'] model = CifarLeNet(pretrained=False, num_classes=data.num_classes) # model from pytorch-ensembles elif 'BayesPreResNet110' in path_model: model_name_list = ['BayesPreResNet110'] arch = getattr(pemodels, 'BayesPreResNet110') model = arch.base(num_classes=data.num_classes, **arch.kwargs) elif len([x for x in PEMODELS_NAMES if x in path_model]) >= 1: # list model name in pemodels: [x for x in dir(pemodels) if x[0:2] != '__'] model_name_list = [x for x in PEMODELS_NAMES if x in path_model] if len(model_name_list) != 1: raise ValueError(f'Failed to extract model name: {model_name_list}') arch = getattr(pemodels, model_name_list[0]) model = arch.base(num_classes=data.num_classes, **arch.kwargs) else: raise NotImplementedError('Model class unknown') model_loaded = torch.load(path_model, map_location=DEVICE) if 'model_state' in model_loaded: model_loaded = model_loaded['model_state'] try: model.load_state_dict(model_loaded) except RuntimeError: new_state_dict = OrderedDict() for k, v in model_loaded.items(): name = k[2:] # remove `1.` new_state_dict[name] = v # load params model.load_state_dict(new_state_dict) # cSGLD trained on [0,1] data range: nothing to do # old dnn_ensemble trained on [-1, 1] data range if 'dnn_ensemble' in path_model or 'single_model' in path_model: model.to(DEVICE) model = add_normalization_layer(model=model, mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5)) arch = model_name_list[0] elif 'ImageNet' in path_model: if 'ImageNet/pretrained' in path_model: if load_as_ghost: raise ValueError('Ghost pretrained models not supported') a = re.match(r"^ImageNet/pretrained/(\w+)$", path_model) if a: arch = a.groups()[0] else: raise ValueError('Failed extracting name of pretrained model') if arch in TVMODELS_NAMES: model = tvmodels.__dict__[arch](pretrained=True) model = add_normalization_layer(model=model, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) elif arch in TIMODELS_NAMES: arch_ = arch.replace('timm_', '') model = timm.create_model(arch_, pretrained=True) config = timm.data.resolve_data_config({}, model=model) model = add_normalization_layer(model=model, mean=config['mean'], std=config['std']) elif arch in RBMODELS_NAMES: # https://github.com/RobustBench/robustbench#imagenet model = rb_load_model(model_name=arch, model_dir='~/.cache/robustbench/models', dataset='imagenet', threat_model="Linf") else: raise ValueError(f'Model {arch} not supported.') model.to(DEVICE) else: checkpoint_dict = torch.load(path_model, map_location=DEVICE) source_model = tvmodels if load_as_ghost: source_model = ghostmodels if 'cSGLD' in path_model or 'single_model' in path_model: arch = checkpoint_dict['arch'] if arch in TIMODELS_NAMES: if load_as_ghost: raise ValueError('Ghost timm models not supported') model = timm.create_model(arch, pretrained=False) else: model = source_model.__dict__[arch]() # try to load state_dir # some models were trained with dataparallel, some not. try: model.load_state_dict(checkpoint_dict['state_dict']) except RuntimeError: #model = torch.nn.parallel.DataParallel(model) new_state_dict = OrderedDict() for k, v in checkpoint_dict['state_dict'].items(): name = k[7:] # remove `module.` new_state_dict[name] = v # load params model.load_state_dict(new_state_dict) if 'withdatanorm' in path_model: # models trained with regular data norm model = add_normalization_layer(model=model, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) elif 'subspace_inference' in path_model: # all SI models are trained on normalized data a = re.match(r".*models(_target)?/ImageNet/(\w+)/.*", path_model) if a: arch = a.group(2) else: raise ValueError('Failed extracting archiecture from filename model') if arch in TIMODELS_NAMES: if load_as_ghost: raise ValueError('Ghost timm models not supported') model = timm.create_model(arch.replace('timm_', ''), pretrained=False) else: model = source_model.__dict__[arch]() state_dict = checkpoint_dict if 'state_dict' in checkpoint_dict: state_dict = checkpoint_dict['state_dict'] try: model.load_state_dict(state_dict) except RuntimeError: new_state_dict = OrderedDict() for k, v in state_dict.items(): name = k[7:] # remove `module.` new_state_dict[name] = v # load params model.load_state_dict(new_state_dict) model.to(DEVICE) # add normalization layer at first model = add_normalization_layer(model=model, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) elif ('deepens_imagenet' in path_model) or ('FGE' in path_model) or ('SSE' in path_model): # pretrained resnet50 from pytorch-ensemble arch = 'resnet50' model = source_model.__dict__[arch]() # models trained with dataparallel #model = torch.nn.parallel.DataParallel(model) new_state_dict = OrderedDict() for k, v in checkpoint_dict['state_dict'].items(): name = k[7:] # remove `module.` new_state_dict[name] = v model.load_state_dict(new_state_dict) model.to(DEVICE) # add normalization layer at first model = add_normalization_layer(model=model, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) elif 'SWAG' in path_model: # train by us from 1 deepens_imagenet checkpoint resnet50 arch = 'resnet50' model = source_model.__dict__[arch]() if 'state_dict' in checkpoint_dict: checkpoint_dict = checkpoint_dict['state_dict'] model.load_state_dict(checkpoint_dict) model.to(DEVICE) # add normalization layer at first model = add_normalization_layer(model=model, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) elif 'VI' in path_model: arch = 'BayesResNet50' model = pemodels.__dict__[arch]() new_state_dict = OrderedDict() model.load_state_dict(checkpoint_dict) model.to(DEVICE) # add normalization layer at first model = add_normalization_layer(model=model, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) else: raise ValueError('ImageNet model not recognized.') if 'inception' in path_model: # resize input to to 299 model = add_resize_layer(model=model, size=(299, 299)) else: raise ValueError('dataset not supported') model.eval() # to GPU if USE_CUDA and not force_cpu: model.to(DEVICE) if input_diversity: # add as first layer that randomly resize between [90%, 100%] and 0-pad to original size with probability 0.5 model = add_random_resize_layer(model=model, p=0.5, min_resize=round(data.get_input_shape()[2]*0.9)) model.to(DEVICE) model.eval() if skip_gradient: if arch in ['resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152']: register_hook_for_resnet(model, arch=arch, gamma=0.2) elif arch in ['PreResNet110']: gamma = float(os.getenv('ADV_TRANSFER_SGM_GAMMA_PRERESNET', '0.7').replace(',', '.')) register_hook_for_preresnet(model, arch=arch, gamma=gamma) elif arch in ['densenet121', 'densenet169', 'densenet201']: register_hook_for_densenet(model, arch=arch, gamma=0.5) else: raise ValueError(f'Arch { arch } not supported by Skip Gradient Method') if defense_randomization: # add as first layer that randomly resize between [90%, 100%] and 0-pad to original size with probability 1. # defense by https://arxiv.org/pdf/1711.01991.pdf model = add_random_resize_layer(model=model, p=1, min_resize=round(data.get_input_shape()[2]*0.9)) model.to(DEVICE) model.eval() if temperature: model = ModelWithTemperature(model, temperature=temperature) model.to(DEVICE) model.eval() return model def load_classifier(model, data): """ Load ART PyTorch classifier from pytorch model :param model: pytorch model instance :param data: data class :return: ART classifier """ # not used but mandatory for ART criterion = torch.nn.CrossEntropyLoss() optimizer = torch.optim.SGD(model.parameters(), lr=0.01) classifier = PyTorchClassifier( model=model, clip_values=(data.min_pixel_value, data.max_pixel_value), loss=criterion, optimizer=optimizer, input_shape=data.get_input_shape(), nb_classes=data.num_classes, device_type="gpu" if USE_CUDA else "cpu" ) #classifier.set_learning_phase(False) # automatically set by predict() calls return classifier def load_classifier_ensemble(models, **kwargs): """ Build an ART classifier of an ensemble of PyTorch models :param models: list of pytorch model instances :return: """ model = TorchEnsemble(models=models, ensemble_logits=True) if USE_CUDA: model.to(DEVICE) model.eval() return load_classifier(model, **kwargs) def predict_ensemble(models_dir, X, data): """ Compute prediction for each model inside models_dir :param models_dir: str path to pytorch's models :param X: pytorch tensor or numpy array :param data: data class instance :return: tuple of 2 numpy arrays of predicted labels with probs and logits ensembling """ torchdataset = torch.utils.data.TensorDataset(torch.Tensor(X)) loader = torch.utils.data.DataLoader( torchdataset, batch_size=data.batch_size, shuffle=False, num_workers=data.num_workers, pin_memory=USE_CUDA) #if not torch.is_tensor(X): # X = torch.Tensor(X) #if USE_CUDA: # X = X.to(DEVICE) path_models = list_models(models_dir) y_pred_ens_logit = torch.zeros((X.shape[0], data.num_classes)) y_pred_ens_prob = torch.zeros((X.shape[0], data.num_classes)) for path_model in path_models: model = guess_and_load_model(path_model=path_model, data=data) for i, (inputs,) in enumerate(loader): inputs = inputs.to(DEVICE, non_blocking=True) with torch.no_grad(): output = model(inputs) y_pred_ens_logit[i*data.batch_size:(i+1)*data.batch_size, :] += output.cpu() y_pred_ens_prob[i*data.batch_size:(i+1)*data.batch_size, :] += torch.nn.functional.softmax(output, dim=1).cpu() # clean del model, output if USE_CUDA: torch.cuda.empty_cache() y_pred_ens_logit /= len(path_models) y_pred_ens_logit = torch.nn.functional.softmax(y_pred_ens_logit, dim=1) y_pred_ens_prob /= len(path_models) label_pred_logit = np.argmax(y_pred_ens_logit.numpy(), axis=1) label_pred_prob = np.argmax(y_pred_ens_prob.numpy(), axis=1) return label_pred_prob, label_pred_logit def compute_accuracy_ensemble(models_dir, X, y, data): label_pred_prob, label_pred_logit = predict_ensemble(models_dir=models_dir, X=X, data=data) acc_prob = (label_pred_prob == y).mean() acc_logit = (label_pred_logit == y).mean() return acc_prob, acc_logit def compute_accuracy_multiple_ensemble(models_dirs, X, y, data): """ Compute the mean of the accuracies of several ensembles. """ acc_prob, acc_logit = 0., 0. for models_dir in models_dirs: acc_prob_tmp, acc_logit_tmp = compute_accuracy_ensemble(models_dir=models_dir, X=X, y=y, data=data) acc_prob += acc_prob_tmp acc_logit += acc_logit_tmp return acc_prob / len(models_dirs), acc_logit / len(models_dirs) def compute_accuracy_from_nested_list_models(list_ensemble, X, y, data, export_predict=False, export_mask=None, tol=1e-12): """ Compute prediction for each model inside list_ensemble :param list_ensemble: list of list of pytorch's models :param X: pytorch tensor or numpy array :param y: pytorch tensor or numpy array :param data: data class instance :param export_predict: bool, also export a boolean vector corresponding to correct predictions of examples :param export_mask: tensor, also export metrics computed on examples masked by provided tensor :param loss: str either 'CE' or 'NLL' :return: tuple of 2 scalar accuracy and CE loss """ #nb_models = np.sum([len(x) for x in list_ensemble]) # total nb of models if export_mask is not None: if export_mask.shape != (X.shape[0], ): raise ValueError('Wrong shape for export_mask') if not np.isin(export_mask.numpy(), [0, 1]).all(): raise ValueError('export_mask tensor should contain only 0,1 values') export_mask.bool().to(DEVICE) torchdataset = torch.utils.data.TensorDataset(torch.tensor(X, dtype=torch.float), torch.tensor(y)) loader = torch.utils.data.DataLoader( torchdataset, batch_size=data.batch_size, shuffle=False, num_workers=data.num_workers, pin_memory=USE_CUDA) loss = nn.NLLLoss(reduction='sum').to(DEVICE) correct = 0 total = 0 loss_sum = 0. predict_correct = torch.zeros((0,)) log_probs = [] with torch.no_grad(): for i, (inputs, labels) in enumerate(loader): inputs = inputs.to(DEVICE, non_blocking=True) labels = labels.to(DEVICE, non_blocking=True) y_pred_ens_batch = torch.zeros(inputs.size(0), data.num_classes).to(DEVICE, non_blocking=True) size_ensemble = 0 log_prob_models_batch = [] for list_model in list_ensemble: for model in list_model: output = model(inputs) # more numerically stable to save log_softmax then log sum exp log_prob_model_batch = torch.nn.functional.log_softmax(output, dim=1) log_prob_models_batch.append(log_prob_model_batch) # log sum exp of log probs, over models dims for each example and classes log_prob_ens_batch = torch.logsumexp(torch.dstack(log_prob_models_batch), dim=2) - np.log(len(log_prob_models_batch)) loss_sum += loss(log_prob_ens_batch, labels.long()).item() # NLLLoss takes log prob _, predicted = torch.max(log_prob_ens_batch.data, 1) correct += (predicted == labels).sum().item() predict_correct = torch.cat((predict_correct, (predicted == labels).cpu()), 0) total += labels.size(0) if export_predict: if predict_correct.shape[0] != X.shape[0]: raise RuntimeError('Unexpected shape of predict_correct vector') return correct / total, loss_sum / total, predict_correct.bool() if export_mask is not None: correct_masked = torch.masked_select(predict_correct, mask=export_mask).sum().cpu().item() total_masked = export_mask.sum().cpu().item() if total_masked == 0: total_masked = np.NaN return correct / total, loss_sum / total, correct_masked / total_masked return correct / total, loss_sum / total def save_numpy(array, path, filename): os.makedirs(path, exist_ok=True) np.save(os.path.join(path, filename), array) def flatten(X): return X.reshape((X.shape[0], -1)) def compute_norm(X_adv, X, norm=2): return np.linalg.norm(flatten(X_adv) - flatten(X), ord=norm, axis=1) def project_on_sphere(X, X_adv, data, size=4., norm=2): """ Project on sphere (not the ball!) of specified size :param X: np array :param X_adv: np array :param size: :param norm: Lp norm to use. Only 2 implemented :return: """ if norm != 2: raise NotImplementedError('Only L2 norm implemented') lpnorm = compute_norm(X_adv, X, norm=2) X_adv_proj = X + size / lpnorm.reshape((X.shape[0], 1, 1, 1)) * (X_adv - X) X_adv_proj = np.clip(X_adv_proj, data.min_pixel_value, data.max_pixel_value) return X_adv_proj def add_normalization_layer(model, mean, std): return torch.nn.Sequential( transforms.Normalize(mean=mean, std=std), model ) def add_resize_layer(model, size, **kwargs): return torch.nn.Sequential( transforms.Resize(size=size, **kwargs), model ) def add_random_resize_layer(model, p=0.5, min_resize=200): """ Add input diversity layer as first layer :param model: pytorch model :param p: probability to apply input diversity :param min_resize: minimum possible resize :return: """ return torch.nn.Sequential( transforms.RandomApply([ RandomResizePad(min_resize), ], p=p), model ) def guess_method(path_model): """ Return the name of the method to train the model """ METHODS_OTHER = ['SGLD', 'cSGLD', 'pSGLD', 'HMC', 'SVI', 'VI', 'SWAG', 'FGE', 'SSE', 'SI/pca/ESS', 'SI/random/ESS', 'SWA', 'collected_models', 'cSGD', 'ImageNet/pretrained'] METHODS_DNN = ['SGD', 'Adam', 'RMSprop', 'DNN', 'dnn_ensemble', 'single_model', 'deepens'] # always return 'dnn' candidates = [x for x in METHODS_OTHER if x in path_model] if len(candidates) > 1: # pick the longest one return max(candidates, key=len) elif len(candidates) == 1: return candidates[0] candidates = [x for x in METHODS_DNN if x in path_model] if len(candidates) >= 1: return 'dnn' raise ValueError('Not able to guess model training method')
28,359
45.79868
287
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/attacks.py
import torch import os import numpy as np import scipy.stats as st from art.attacks.evasion import FastGradientMethod, ProjectedGradientDescentPyTorch from art.classifiers import PyTorchClassifier from art.config import ART_NUMPY_DTYPE from art.utils import ( random_sphere, projection, ) from tqdm import tqdm from typing import Optional, Union, TYPE_CHECKING, Tuple, List from utils.helpers import DEVICE, compute_accuracy_from_nested_list_models class ExtendedFastGradientMethod(FastGradientMethod): attack_params = FastGradientMethod.attack_params + ['momentum', 'grad_momentum', 'translation_invariant'] def __init__(self, momentum=None, grad_momentum=None, translation_invariant=False, **kwargs): super().__init__(**kwargs) self.momentum = momentum self.grad_momentum = grad_momentum self.translation_invariant = translation_invariant if momentum and grad_momentum is None: raise ValueError("grad should be provided when using momentum") def _compute_perturbation(self, batch: np.ndarray, batch_labels: np.ndarray, mask: Optional[np.ndarray], batch_grad_momentum: Optional[np.ndarray]) -> \ Tuple[np.ndarray, np.ndarray]: # Pick a small scalar to avoid division by 0 tol = 10e-8 # Get gradient wrt loss; invert it if attack is targeted grad = self.estimator.loss_gradient(batch, batch_labels) * (1 - 2 * int(self.targeted)) if self.translation_invariant: # original implementation: #noise = tf.nn.depthwise_conv2d(noise, stack_kernel, strides=[1, 1, 1, 1], padding='SAME') # translation_invariant kernel #kernel = self.gkern(15, 3).astype(np.float32) #stack_kernel = np.stack([kernel, kernel, kernel]).swapaxes(2, 0) #stack_kernel = np.expand_dims(stack_kernel, 3) kernlen = 15 nb_channels = batch.shape[1] padding = int((kernlen - 1) / 2) # same padding with torch.no_grad(): kernel = torch.from_numpy(self.gkern(kernlen=kernlen).astype(np.float32)) stack_kernel = kernel.view(1, 1, kernlen, kernlen).repeat(nb_channels, 1, 1, 1).to(DEVICE) grad = torch.nn.functional.conv2d(torch.from_numpy(grad).to(DEVICE), stack_kernel, padding=padding, groups=nb_channels).cpu().numpy() if grad.shape != batch.shape: raise RuntimeError('Translation invariant gradient does not have the same dimension as input') # Apply norm bound def _apply_norm(grad, norm=self.norm, object_type=False): if norm in [np.inf, "inf"]: grad = np.sign(grad) elif norm == 1: if not object_type: ind = tuple(range(1, len(batch.shape))) else: ind = None grad = grad / (np.sum(np.abs(grad), axis=ind, keepdims=True) + tol) elif norm == 2: if not object_type: ind = tuple(range(1, len(batch.shape))) else: ind = None grad = grad / (np.sqrt(np.sum(np.square(grad), axis=ind, keepdims=True)) + tol) return grad # momentum if self.momentum: # scale L1-norm if batch.dtype == np.object: for i_sample in range(batch.shape[0]): grad[i_sample] = _apply_norm(grad[i_sample], norm=1, object_type=True) assert batch[i_sample].shape == grad[i_sample].shape else: grad = _apply_norm(grad, norm=1) # update moving avg #self.grad_momentum = self.grad_momentum * self.momentum + grad batch_grad_momentum = batch_grad_momentum * self.momentum + grad grad = batch_grad_momentum.copy() #self.grad_momentum = self.grad_momentum.detach().clone() # Apply mask if mask is not None: grad = np.where(mask == 0.0, 0.0, grad) # Apply norm bound if batch.dtype == np.object: for i_sample in range(batch.shape[0]): grad[i_sample] = _apply_norm(grad[i_sample], object_type=True) assert batch[i_sample].shape == grad[i_sample].shape else: grad = _apply_norm(grad) assert batch.shape == grad.shape return grad, batch_grad_momentum @staticmethod def gkern(kernlen=21, nsig=3): """Returns a 2D Gaussian kernel array. From the original implementation https://github.com/dongyp13/Translation-Invariant-Attacks/blob/master/attack_iter.py """ x = np.linspace(-nsig, nsig, kernlen) kern1d = st.norm.pdf(x) kernel_raw = np.outer(kern1d, kern1d) kernel = kernel_raw / kernel_raw.sum() return kernel def _compute( self, x: np.ndarray, x_init: np.ndarray, y: np.ndarray, mask: Optional[np.ndarray], eps: float, eps_step: float, project: bool, random_init: bool, ) -> np.ndarray: if random_init: n = x.shape[0] m = np.prod(x.shape[1:]).item() random_perturbation = random_sphere(n, m, eps, self.norm).reshape(x.shape).astype(ART_NUMPY_DTYPE) if mask is not None: random_perturbation = random_perturbation * (mask.astype(ART_NUMPY_DTYPE)) x_adv = x.astype(ART_NUMPY_DTYPE) + random_perturbation if self.estimator.clip_values is not None: clip_min, clip_max = self.estimator.clip_values x_adv = np.clip(x_adv, clip_min, clip_max) else: if x.dtype == np.object: x_adv = x.copy() else: x_adv = x.astype(ART_NUMPY_DTYPE) # Compute perturbation with implicit batching for batch_id in range(int(np.ceil(x.shape[0] / float(self.batch_size)))): batch_index_1, batch_index_2 = batch_id * self.batch_size, (batch_id + 1) * self.batch_size batch_index_2 = min(batch_index_2, x.shape[0]) batch = x_adv[batch_index_1:batch_index_2] batch_labels = y[batch_index_1:batch_index_2] batch_grad_momentum = self.grad_momentum[batch_index_1:batch_index_2] mask_batch = mask if mask is not None: # Here we need to make a distinction: if the masks are different for each input, we need to index # those for the current batch. Otherwise (i.e. mask is meant to be broadcasted), keep it as it is. if len(mask.shape) == len(x.shape): mask_batch = mask[batch_index_1:batch_index_2] # Get perturbation perturbation, batch_grad_momentum = self._compute_perturbation(batch, batch_labels, mask_batch, batch_grad_momentum) # Apply perturbation and clip x_adv[batch_index_1:batch_index_2] = self._apply_perturbation(batch, perturbation, eps_step) self.grad_momentum[batch_index_1:batch_index_2] = batch_grad_momentum if project: if x_adv.dtype == np.object: for i_sample in range(batch_index_1, batch_index_2): perturbation = projection(x_adv[i_sample] - x_init[i_sample], eps, self.norm) x_adv[i_sample] = x_init[i_sample] + perturbation else: perturbation = projection( x_adv[batch_index_1:batch_index_2] - x_init[batch_index_1:batch_index_2], eps, self.norm ) x_adv[batch_index_1:batch_index_2] = x_init[batch_index_1:batch_index_2] + perturbation return x_adv class ExtendedProjectedGradientDescentPyTorch(ProjectedGradientDescentPyTorch): """ Implement Light Version of PGD where only a subset of models is attack at each iterations. Also support some test-time techniques for transferability. :param models_target_dict: Dictionary of target models as values with the name as key. :param freq_acc_target: Compute accuracy on target each provided iterations. All iterations by default. :param data: Data class. For now only required with models_target_dict. """ attack_params = ProjectedGradientDescentPyTorch.attack_params + ['momentum', 'translation_invariant', 'grad_noise_std', 'models_target_dict', 'freq_eval_target', 'data'], def __init__( self, estimators: List[PyTorchClassifier], # pass a list of classifier momentum: float = None, translation_invariant: bool = False, grad_noise_std: float = None, models_target_dict=None, freq_eval_target=1, data=None, **kwargs ): self.estimators = estimators super().__init__(estimator=estimators[0], **kwargs) self.momentum = momentum self.translation_invariant = translation_invariant self.grad_noise_std = grad_noise_std # to report target accuracy at each freq_eval_target iteration self.freq_eval_target = freq_eval_target self.data = data if models_target_dict and data is None: raise ValueError('data param should be provided if models_target_dict is set.') self.models_target_dict = models_target_dict self.stats_per_iter = {k: {'iter': [], 'acc': [], 'loss': []} for k in models_target_dict} if models_target_dict else {} def generate(self, x: np.ndarray, y: Optional[np.ndarray] = None, **kwargs) -> np.ndarray: """ Generate adversarial samples and return them in an array. Modified to compute the successful adversarial on all the classifiers in the list. Not only one. :param x: An array with the original inputs. :param y: Target values (class labels) one-hot-encoded of shape `(nb_samples, nb_classes)` or indices of shape (nb_samples,). Only provide this parameter if you'd like to use true labels when crafting adversarial samples. Otherwise, model predictions are used as labels to avoid the "label leaking" effect (explained in this paper: https://arxiv.org/abs/1611.01236). Default is `None`. :param mask: An array with a mask broadcastable to input `x` defining where to apply adversarial perturbations. Shape needs to be broadcastable to the shape of x and can also be of the same shape as `x`. Any features for which the mask is zero will not be adversarially perturbed. :type mask: `np.ndarray` :return: An array holding the adversarial examples. """ import torch # lgtm [py/repeated-import] mask = self._get_mask(x, **kwargs) # Ensure eps is broadcastable self._check_compatibility_input_and_eps(x=x) # Check whether random eps is enabled self._random_eps() # Set up targets targets = self._set_targets(x, y) # Create dataset if mask is not None: # Here we need to make a distinction: if the masks are different for each input, we need to index # those for the current batch. Otherwise (i.e. mask is meant to be broadcasted), keep it as it is. if len(mask.shape) == len(x.shape): dataset = torch.utils.data.TensorDataset( torch.from_numpy(x.astype(ART_NUMPY_DTYPE)), torch.from_numpy(targets.astype(ART_NUMPY_DTYPE)), torch.from_numpy(mask.astype(ART_NUMPY_DTYPE)), ) else: dataset = torch.utils.data.TensorDataset( torch.from_numpy(x.astype(ART_NUMPY_DTYPE)), torch.from_numpy(targets.astype(ART_NUMPY_DTYPE)), torch.from_numpy(np.array([mask.astype(ART_NUMPY_DTYPE)] * x.shape[0])), ) else: dataset = torch.utils.data.TensorDataset( torch.from_numpy(x.astype(ART_NUMPY_DTYPE)), torch.from_numpy(targets.astype(ART_NUMPY_DTYPE)), ) data_loader = torch.utils.data.DataLoader( dataset=dataset, batch_size=self.batch_size, shuffle=False, drop_last=False ) # Start to compute adversarial examples adv_x = x.astype(ART_NUMPY_DTYPE) # Compute perturbation with batching for (batch_id, batch_all) in enumerate( tqdm(data_loader, desc="PGD - Batches", leave=False, disable=not self.verbose) ): if mask is not None: (batch, batch_labels, mask_batch) = batch_all[0], batch_all[1], batch_all[2] else: (batch, batch_labels, mask_batch) = batch_all[0], batch_all[1], None batch_index_1, batch_index_2 = batch_id * self.batch_size, (batch_id + 1) * self.batch_size # Compute batch_eps and batch_eps_step if isinstance(self.eps, np.ndarray): if len(self.eps.shape) == len(x.shape) and self.eps.shape[0] == x.shape[0]: batch_eps = self.eps[batch_index_1:batch_index_2] batch_eps_step = self.eps_step[batch_index_1:batch_index_2] else: batch_eps = self.eps batch_eps_step = self.eps_step else: batch_eps = self.eps batch_eps_step = self.eps_step for rand_init_num in range(max(1, self.num_random_init)): if rand_init_num == 0: # first iteration: use the adversarial examples as they are the only ones we have now adv_x[batch_index_1:batch_index_2] = self._generate_batch( x=batch, targets=batch_labels, mask=mask_batch, eps=batch_eps, eps_step=batch_eps_step ) else: adversarial_batch = self._generate_batch( x=batch, targets=batch_labels, mask=mask_batch, eps=batch_eps, eps_step=batch_eps_step ) # return the successful adversarial examples # modified: attack_success = self.compute_success_array( batch, batch_labels, adversarial_batch, self.targeted, batch_size=self.batch_size, ) adv_x[batch_index_1:batch_index_2][attack_success] = adversarial_batch[attack_success] # modified: # logger.info( # "Success rate of attack: %.2f%%", # 100 * self.compute_success(x, y, adv_x, self.targeted, batch_size=self.batch_size), # ) return adv_x def compute_success_array( self, x_clean: np.ndarray, labels: np.ndarray, x_adv: np.ndarray, targeted: bool = False, batch_size: int = 1, ) -> float: """ Compute the success rate of an attack based on clean samples, adversarial samples and targets or correct labels. Modified: use the list of estimators to compute the predictions :param x_clean: Original clean samples. :param labels: Correct labels of `x_clean` if the attack is untargeted, or target labels of the attack otherwise. :param x_adv: Adversarial samples to be evaluated. :param targeted: `True` if the attack is targeted. In that case, `labels` are treated as target classes instead of correct labels of the clean samples. :param batch_size: Batch size. :return: Percentage of successful adversarial samples. """ adv_results = np.zeros((x_adv.shape[0], self.estimators[0].nb_classes), dtype=np.float32) for classifier in self.estimators: adv_results += classifier.predict(x_adv, batch_size=batch_size) adv_preds = np.argmax(adv_results, axis=1) if targeted: attack_success = adv_preds == np.argmax(labels, axis=1) else: results = np.zeros((x_clean.shape[0], self.estimators[0].nb_classes), dtype=np.float32) for classifier in self.estimators: results += classifier.predict(x_clean, batch_size=batch_size) preds = np.argmax(results, axis=1) attack_success = adv_preds != preds return attack_success def _generate_batch( self, x: "torch.Tensor", targets: "torch.Tensor", mask: "torch.Tensor", eps: Union[int, float, np.ndarray], eps_step: Union[int, float, np.ndarray], ) -> np.ndarray: """ Generate a batch of adversarial samples and return them in an array. Each iteration is computed on a different estimator from estimators. :param x: An array with the original inputs. :param targets: Target values (class labels) one-hot-encoded of shape `(nb_samples, nb_classes)`. :param mask: An array with a mask to be applied to the adversarial perturbations. Shape needs to be broadcastable to the shape of x. Any features for which the mask is zero will not be adversarially perturbed. :param eps: Maximum perturbation that the attacker can introduce. :param eps_step: Attack step size (input variation) at each iteration. :return: Adversarial examples. """ inputs = x.to(self.estimator.device) targets = targets.to(self.estimator.device) y = torch.max(targets, dim=1).indices # class idx adv_x = inputs if mask is not None: mask = mask.to(self.estimator.device) # init grad_momentum at beginning of a batch self.grad_momentum = torch.zeros(x.size()).to(self.estimator.device) for i_max_iter in range(self.max_iter): # cycle between estimators self._estimator = self.estimators[i_max_iter % len(self.estimators)] adv_x = self._compute_torch( adv_x, inputs, targets, mask, eps, eps_step, self.num_random_init > 0 and i_max_iter == 0, ) # compute target accuracy at each freq_eval_target iteration (+ first/last iterations) if self.models_target_dict: if (i_max_iter % self.freq_eval_target == 0) or (i_max_iter == self.max_iter-1): for name_target, model_target in self.models_target_dict.items(): acc_target, loss_target = compute_accuracy_from_nested_list_models( [[model_target, ], ], X=adv_x.cpu(), y=y.cpu(), data=self.data) self.stats_per_iter[name_target]['iter'].append(i_max_iter) self.stats_per_iter[name_target]['acc'].append(acc_target) self.stats_per_iter[name_target]['loss'].append(loss_target) return adv_x.cpu().detach().numpy() def _compute_perturbation( self, x: "torch.Tensor", y: "torch.Tensor", mask: Optional["torch.Tensor"] ) -> "torch.Tensor": """ Compute perturbations. :param x: Current adversarial examples. :param y: Target values (class labels) one-hot-encoded of shape `(nb_samples, nb_classes)` or indices of shape (nb_samples,). Only provide this parameter if you'd like to use true labels when crafting adversarial samples. Otherwise, model predictions are used as labels to avoid the "label leaking" effect (explained in this paper: https://arxiv.org/abs/1611.01236). Default is `None`. :param mask: An array with a mask to be applied to the adversarial perturbations. Shape needs to be broadcastable to the shape of x. Any features for which the mask is zero will not be adversarially perturbed. :return: Perturbations. """ import torch # lgtm [py/repeated-import] # Pick a small scalar to avoid division by 0 tol = 10e-8 # Get gradient wrt loss; invert it if attack is targeted grad = self.estimator.loss_gradient(x=x, y=y) * (1 - 2 * int(self.targeted)) # added: if self.translation_invariant: # original implementation: #noise = tf.nn.depthwise_conv2d(noise, stack_kernel, strides=[1, 1, 1, 1], padding='SAME') # translation_invariant kernel #kernel = self.gkern(15, 3).astype(np.float32) #stack_kernel = np.stack([kernel, kernel, kernel]).swapaxes(2, 0) #stack_kernel = np.expand_dims(stack_kernel, 3) # kernlen = 15 kernlen = int(os.getenv('ADV_TRANSFER_TI_KERNEL_SIZE', '15')) nb_channels = x.size(1) padding = int((kernlen - 1) / 2) # same padding with torch.no_grad(): kernel = self.gkern(kernlen=kernlen).to(self.estimator.device) stack_kernel = kernel.view(1, 1, kernlen, kernlen).repeat(nb_channels, 1, 1, 1) grad = torch.nn.functional.conv2d(grad, stack_kernel, padding=padding, groups=nb_channels) if grad.shape != x.shape: raise RuntimeError('Translation invariant gradient does not have the same dimension as input') # added for momentum if self.momentum: # scale L1-norm ind = tuple(range(1, len(x.shape))) grad = grad / (torch.sum(grad.abs(), dim=ind, keepdims=True) + tol) # type: ignore # update moving avg with torch.no_grad(): self.grad_momentum = self.grad_momentum * self.momentum + grad grad = self.grad_momentum # add gaussian noise to gradients if self.grad_noise_std: grad += torch.randn(grad.shape).to(self.estimator.device) * self.grad_noise_std # Apply mask if mask is not None: grad = torch.where(mask == 0.0, torch.tensor(0.0), grad) # Apply norm bound if self.norm in ["inf", np.inf]: grad = grad.sign() elif self.norm == 1: ind = tuple(range(1, len(x.shape))) grad = grad / (torch.sum(grad.abs(), dim=ind, keepdims=True) + tol) # type: ignore elif self.norm == 2: ind = tuple(range(1, len(x.shape))) grad = grad / (torch.sqrt(torch.sum(grad * grad, axis=ind, keepdims=True)) + tol) # type: ignore assert x.shape == grad.shape return grad @staticmethod def gkern(kernlen=21, nsig=3): """Returns a 2D Gaussian kernel array. From the original implementation https://github.com/dongyp13/Translation-Invariant-Attacks/blob/master/attack_iter.py """ x = np.linspace(-nsig, nsig, kernlen) kern1d = st.norm.pdf(x) kernel_raw = np.outer(kern1d, kern1d) kernel = kernel_raw / kernel_raw.sum() # convert to tensor kernel = torch.from_numpy(kernel.astype(np.float32)) return kernel def get_target_accuracy_per_iter(self, name_target): if not self.models_target_dict: raise ValueError('models_target_dict must be specified') return self.stats_per_iter[name_target]
23,513
45.562376
174
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/swag/losses.py
import torch import torch.nn.functional as F def cross_entropy(model, input, target): # standard cross-entropy loss function output = model(input) loss = F.cross_entropy(output, target) return loss, output def adversarial_cross_entropy( model, input, target, lossfn=F.cross_entropy, epsilon=0.01 ): # loss function based on algorithm 1 of "simple and scalable uncertainty estimation using # deep ensembles," lakshminaraynan, pritzel, and blundell, nips 2017, # https://arxiv.org/pdf/1612.01474.pdf # note: the small difference bw this paper is that here the loss is only backpropped # through the adversarial loss rather than both due to memory constraints on preresnets # we can change back if we want to restrict ourselves to VGG-like networks (where it's fine). # scale epsilon by min and max (should be [0,1] for all experiments) # see algorithm 1 of paper scaled_epsilon = epsilon * (input.max() - input.min()) # force inputs to require gradient input.requires_grad = True # standard forwards pass output = model(input) loss = lossfn(output, target) # now compute gradients wrt input loss.backward(retain_graph=True) # now compute sign of gradients inputs_grad = torch.sign(input.grad) # perturb inputs and use clamped output inputs_perturbed = torch.clamp( input + scaled_epsilon * inputs_grad, 0.0, 1.0 ).detach() # inputs_perturbed.requires_grad = False input.grad.zero_() # model.zero_grad() outputs_perturbed = model(inputs_perturbed) # compute adversarial version of loss adv_loss = lossfn(outputs_perturbed, target) # return mean of loss for reasonable scalings return (loss + adv_loss) / 2.0, output def masked_loss(y_pred, y_true, void_class=11.0, weight=None, reduce=True): # masked version of crossentropy loss el = torch.ones_like(y_true) * void_class mask = torch.ne(y_true, el).long() y_true_tmp = y_true * mask loss = F.cross_entropy(y_pred, y_true_tmp, weight=weight, reduction="none") loss = mask.float() * loss if reduce: return loss.sum() / mask.sum() else: return loss, mask def seg_cross_entropy(model, input, target, weight=None): output = model(input) # use masked loss function loss = masked_loss(output, target, weight=weight) return {"loss": loss, "output": output} def seg_ale_cross_entropy(model, input, target, num_samples=50, weight=None): # requires two outputs for model(input) output = model(input) mean = output[:, 0, :, :, :] scale = output[:, 1, :, :, :].abs() output_distribution = torch.distributions.Normal(mean, scale) total_loss = 0 for _ in range(num_samples): sample = output_distribution.rsample() current_loss, mask = masked_loss(sample, target, weight=weight, reduce=False) total_loss = total_loss + current_loss.exp() mean_loss = total_loss / num_samples return {"loss": mean_loss.log().sum() / mask.sum(), "output": mean, "scale": scale}
3,094
28.47619
97
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/swag/utils.py
import itertools import torch import os import copy from datetime import datetime import math import numpy as np import tqdm import torch.nn.functional as F def flatten(lst): tmp = [i.contiguous().view(-1, 1) for i in lst] return torch.cat(tmp).view(-1) def unflatten_like(vector, likeTensorList): # Takes a flat torch.tensor and unflattens it to a list of torch.tensors # shaped like likeTensorList outList = [] i = 0 for tensor in likeTensorList: # n = module._parameters[name].numel() n = tensor.numel() outList.append(vector[:, i : i + n].view(tensor.shape)) i += n return outList def LogSumExp(x, dim=0): m, _ = torch.max(x, dim=dim, keepdim=True) return m + torch.log((x - m).exp().sum(dim=dim, keepdim=True)) def adjust_learning_rate(optimizer, lr): for param_group in optimizer.param_groups: param_group["lr"] = lr return lr def save_checkpoint(dir, epoch, name="checkpoint", **kwargs): state = {"epoch": epoch} state.update(kwargs) filepath = os.path.join(dir, "%s-%d.pt" % (name, epoch)) torch.save(state, filepath) def train_epoch( loader, model, criterion, optimizer, cuda=True, regression=False, verbose=False, subset=None, ): loss_sum = 0.0 correct = 0.0 verb_stage = 0 num_objects_current = 0 num_batches = len(loader) model.train() if subset is not None: num_batches = int(num_batches * subset) loader = itertools.islice(loader, num_batches) if verbose: loader = tqdm.tqdm(loader, total=num_batches) for i, (input, target) in enumerate(loader): if cuda: input = input.cuda(non_blocking=True) target = target.cuda(non_blocking=True) loss, output = criterion(model, input, target) optimizer.zero_grad() loss.backward() optimizer.step() loss_sum += loss.data.item() * input.size(0) if not regression: pred = output.data.argmax(1, keepdim=True) correct += pred.eq(target.data.view_as(pred)).sum().item() num_objects_current += input.size(0) if verbose and 10 * (i + 1) / num_batches >= verb_stage + 1: print( "Stage %d/10. Loss: %12.4f. Acc: %6.2f" % ( verb_stage + 1, loss_sum / num_objects_current, correct / num_objects_current * 100.0, ) ) verb_stage += 1 return { "loss": loss_sum / num_objects_current, "accuracy": None if regression else correct / num_objects_current * 100.0, } def eval(loader, model, criterion, cuda=True, regression=False, verbose=False): loss_sum = 0.0 correct = 0.0 num_objects_total = len(loader.dataset) model.eval() with torch.no_grad(): if verbose: loader = tqdm.tqdm(loader) for i, (input, target) in enumerate(loader): if cuda: input = input.cuda(non_blocking=True) target = target.cuda(non_blocking=True) loss, output = criterion(model, input, target) loss_sum += loss.item() * input.size(0) if not regression: pred = output.data.argmax(1, keepdim=True) correct += pred.eq(target.data.view_as(pred)).sum().item() return { "loss": loss_sum / num_objects_total, "accuracy": None if regression else correct / num_objects_total * 100.0, } def predict(loader, model, verbose=False): predictions = list() targets = list() model.eval() if verbose: loader = tqdm.tqdm(loader) offset = 0 with torch.no_grad(): for input, target in loader: input = input.cuda(non_blocking=True) output = model(input) batch_size = input.size(0) predictions.append(F.softmax(output, dim=1).cpu().numpy()) targets.append(target.numpy()) offset += batch_size return {"predictions": np.vstack(predictions), "targets": np.concatenate(targets)} def moving_average(net1, net2, alpha=1): for param1, param2 in zip(net1.parameters(), net2.parameters()): param1.data *= 1.0 - alpha param1.data += param2.data * alpha def _check_bn(module, flag): if issubclass(module.__class__, torch.nn.modules.batchnorm._BatchNorm): flag[0] = True def check_bn(model): flag = [False] model.apply(lambda module: _check_bn(module, flag)) return flag[0] def reset_bn(module): if issubclass(module.__class__, torch.nn.modules.batchnorm._BatchNorm): module.running_mean = torch.zeros_like(module.running_mean) module.running_var = torch.ones_like(module.running_var) def _get_momenta(module, momenta): if issubclass(module.__class__, torch.nn.modules.batchnorm._BatchNorm): momenta[module] = module.momentum def _set_momenta(module, momenta): if issubclass(module.__class__, torch.nn.modules.batchnorm._BatchNorm): module.momentum = momenta[module] def bn_update(loader, model, verbose=False, subset=None, **kwargs): """ BatchNorm buffers update (if any). Performs 1 epochs to estimate buffers average using train dataset. :param loader: train dataset loader for buffers average estimation. :param model: model being update :return: None """ if not check_bn(model): return model.train() momenta = {} model.apply(reset_bn) model.apply(lambda module: _get_momenta(module, momenta)) n = 0 num_batches = len(loader) with torch.no_grad(): if subset is not None: num_batches = int(num_batches * subset) loader = itertools.islice(loader, num_batches) if verbose: loader = tqdm.tqdm(loader, total=num_batches) for input, _ in loader: input = input.cuda(non_blocking=True) input_var = torch.autograd.Variable(input) b = input_var.data.size(0) momentum = b / (n + b) for module in momenta.keys(): module.momentum = momentum model(input_var, **kwargs) n += b model.apply(lambda module: _set_momenta(module, momenta)) def inv_softmax(x, eps=1e-10): return torch.log(x / (1.0 - x + eps)) def predictions(test_loader, model, seed=None, cuda=True, regression=False, **kwargs): # will assume that model is already in eval mode # model.eval() preds = [] targets = [] for input, target in test_loader: if seed is not None: torch.manual_seed(seed) if cuda: input = input.cuda(non_blocking=True) output = model(input, **kwargs) if regression: preds.append(output.cpu().data.numpy()) else: probs = F.softmax(output, dim=1) preds.append(probs.cpu().data.numpy()) targets.append(target.numpy()) return np.vstack(preds), np.concatenate(targets) def schedule(epoch, lr_init, epochs, swa, swa_start=None, swa_lr=None): t = (epoch) / (swa_start if swa else epochs) lr_ratio = swa_lr / lr_init if swa else 0.01 if t <= 0.5: factor = 1.0 elif t <= 0.9: factor = 1.0 - (1.0 - lr_ratio) * (t - 0.5) / 0.4 else: factor = lr_ratio return lr_init * factor
7,489
26.740741
86
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/swag/data.py
""" separate data loader for imagenet """ import os import torch import torchvision import torchvision.transforms as transforms from PIL import ImageFile ImageFile.LOAD_TRUNCATED_IMAGES = True def loaders(path, batch_size, num_workers, shuffle_train=True): train_dir = os.path.join(path, "train") # validation_dir = os.path.join(path, 'validation') validation_dir = os.path.join(path, "val") # transformations for pretrained models (https://github.com/pytorch/examples/blob/42e5b996718797e45c46a25c55b031e6768f8440/imagenet/main.py#L89-L101) normalize = transforms.Normalize( mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225] ) transform_train = transforms.Compose( [ transforms.RandomSizedCrop(224), transforms.RandomHorizontalFlip(), transforms.ToTensor(), normalize, ] ) transform_test = transforms.Compose( [ transforms.Resize(256), transforms.CenterCrop(224), transforms.ToTensor(), normalize, ] ) train_set = torchvision.datasets.ImageFolder(train_dir, transform=transform_train) test_set = torchvision.datasets.ImageFolder( validation_dir, transform=transform_test ) num_classes = 1000 return ( { "train": torch.utils.data.DataLoader( train_set, batch_size=batch_size, shuffle=shuffle_train, num_workers=num_workers, pin_memory=True, ), "test": torch.utils.data.DataLoader( test_set, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=True, ), }, num_classes, )
1,851
25.84058
153
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/swag/posteriors/swag.py
""" implementation of SWAG """ import torch import numpy as np import itertools from torch.distributions.normal import Normal import copy import gpytorch from gpytorch.lazy import RootLazyTensor, DiagLazyTensor, AddedDiagLazyTensor from gpytorch.distributions import MultivariateNormal from ..utils import flatten, unflatten_like def swag_parameters(module, params, no_cov_mat=True): for name in list(module._parameters.keys()): if module._parameters[name] is None: continue data = module._parameters[name].data module._parameters.pop(name) module.register_buffer("%s_mean" % name, data.new(data.size()).zero_()) module.register_buffer("%s_sq_mean" % name, data.new(data.size()).zero_()) if no_cov_mat is False: module.register_buffer( "%s_cov_mat_sqrt" % name, data.new_empty((0, data.numel())).zero_() ) params.append((module, name)) class SWAG(torch.nn.Module): def __init__( self, base, no_cov_mat=True, max_num_models=0, var_clamp=1e-30, *args, **kwargs ): super(SWAG, self).__init__() self.register_buffer("n_models", torch.zeros([1], dtype=torch.long)) self.params = list() self.no_cov_mat = no_cov_mat self.max_num_models = max_num_models self.var_clamp = var_clamp self.base = base(*args, **kwargs) self.base.apply( lambda module: swag_parameters( module=module, params=self.params, no_cov_mat=self.no_cov_mat ) ) def forward(self, *args, **kwargs): return self.base(*args, **kwargs) def sample(self, scale=1.0, cov=False, seed=None, block=False, fullrank=True): if seed is not None: torch.manual_seed(seed) if not block: self.sample_fullrank(scale, cov, fullrank) else: self.sample_blockwise(scale, cov, fullrank) def sample_blockwise(self, scale, cov, fullrank): for module, name in self.params: mean = module.__getattr__("%s_mean" % name) sq_mean = module.__getattr__("%s_sq_mean" % name) eps = torch.randn_like(mean) var = torch.clamp(sq_mean - mean ** 2, self.var_clamp) scaled_diag_sample = scale * torch.sqrt(var) * eps if cov is True: cov_mat_sqrt = module.__getattr__("%s_cov_mat_sqrt" % name) eps = cov_mat_sqrt.new_empty((cov_mat_sqrt.size(0), 1)).normal_() cov_sample = ( scale / ((self.max_num_models - 1) ** 0.5) ) * cov_mat_sqrt.t().matmul(eps).view_as(mean) if fullrank: w = mean + scaled_diag_sample + cov_sample else: w = mean + scaled_diag_sample else: w = mean + scaled_diag_sample module.__setattr__(name, w) def sample_fullrank(self, scale, cov, fullrank): scale_sqrt = scale ** 0.5 mean_list = [] sq_mean_list = [] if cov: cov_mat_sqrt_list = [] for (module, name) in self.params: mean = module.__getattr__("%s_mean" % name) sq_mean = module.__getattr__("%s_sq_mean" % name) if cov: cov_mat_sqrt = module.__getattr__("%s_cov_mat_sqrt" % name) cov_mat_sqrt_list.append(cov_mat_sqrt.cpu()) mean_list.append(mean.cpu()) sq_mean_list.append(sq_mean.cpu()) mean = flatten(mean_list) sq_mean = flatten(sq_mean_list) # draw diagonal variance sample var = torch.clamp(sq_mean - mean ** 2, self.var_clamp) var_sample = var.sqrt() * torch.randn_like(var, requires_grad=False) # if covariance draw low rank sample if cov: cov_mat_sqrt = torch.cat(cov_mat_sqrt_list, dim=1) cov_sample = cov_mat_sqrt.t().matmul( cov_mat_sqrt.new_empty( (cov_mat_sqrt.size(0),), requires_grad=False ).normal_() ) cov_sample /= (self.max_num_models - 1) ** 0.5 rand_sample = var_sample + cov_sample else: rand_sample = var_sample # update sample with mean and scale sample = mean + scale_sqrt * rand_sample sample = sample.unsqueeze(0) # unflatten new sample like the mean sample samples_list = unflatten_like(sample, mean_list) for (module, name), sample in zip(self.params, samples_list): module.__setattr__(name, sample.cuda()) def collect_model(self, base_model): for (module, name), base_param in zip(self.params, base_model.parameters()): mean = module.__getattr__("%s_mean" % name) sq_mean = module.__getattr__("%s_sq_mean" % name) # first moment mean = mean * self.n_models.item() / ( self.n_models.item() + 1.0 ) + base_param.data / (self.n_models.item() + 1.0) # second moment sq_mean = sq_mean * self.n_models.item() / ( self.n_models.item() + 1.0 ) + base_param.data ** 2 / (self.n_models.item() + 1.0) # square root of covariance matrix if self.no_cov_mat is False: cov_mat_sqrt = module.__getattr__("%s_cov_mat_sqrt" % name) # block covariance matrices, store deviation from current mean dev = (base_param.data - mean).view(-1, 1) cov_mat_sqrt = torch.cat((cov_mat_sqrt, dev.view(-1, 1).t()), dim=0) # remove first column if we have stored too many models if (self.n_models.item() + 1) > self.max_num_models: cov_mat_sqrt = cov_mat_sqrt[1:, :] module.__setattr__("%s_cov_mat_sqrt" % name, cov_mat_sqrt) module.__setattr__("%s_mean" % name, mean) module.__setattr__("%s_sq_mean" % name, sq_mean) self.n_models.add_(1) def load_state_dict(self, state_dict, strict=True): if not self.no_cov_mat: n_models = state_dict["n_models"].item() rank = min(n_models, self.max_num_models) for module, name in self.params: mean = module.__getattr__("%s_mean" % name) module.__setattr__( "%s_cov_mat_sqrt" % name, mean.new_empty((rank, mean.numel())).zero_(), ) super(SWAG, self).load_state_dict(state_dict, strict) def export_numpy_params(self, export_cov_mat=False): mean_list = [] sq_mean_list = [] cov_mat_list = [] for module, name in self.params: mean_list.append(module.__getattr__("%s_mean" % name).cpu().numpy().ravel()) sq_mean_list.append( module.__getattr__("%s_sq_mean" % name).cpu().numpy().ravel() ) if export_cov_mat: cov_mat_list.append( module.__getattr__("%s_cov_mat_sqrt" % name).cpu().numpy().ravel() ) mean = np.concatenate(mean_list) sq_mean = np.concatenate(sq_mean_list) var = sq_mean - np.square(mean) if export_cov_mat: return mean, var, cov_mat_list else: return mean, var def import_numpy_weights(self, w): k = 0 for module, name in self.params: mean = module.__getattr__("%s_mean" % name) s = np.prod(mean.shape) module.__setattr__(name, mean.new_tensor(w[k : k + s].reshape(mean.shape))) k += s def generate_mean_var_covar(self): mean_list = [] var_list = [] cov_mat_root_list = [] for module, name in self.params: mean = module.__getattr__("%s_mean" % name) sq_mean = module.__getattr__("%s_sq_mean" % name) cov_mat_sqrt = module.__getattr__("%s_cov_mat_sqrt" % name) mean_list.append(mean) var_list.append(sq_mean - mean ** 2.0) cov_mat_root_list.append(cov_mat_sqrt) return mean_list, var_list, cov_mat_root_list def compute_ll_for_block(self, vec, mean, var, cov_mat_root): vec = flatten(vec) mean = flatten(mean) var = flatten(var) cov_mat_lt = RootLazyTensor(cov_mat_root.t()) var_lt = DiagLazyTensor(var + 1e-6) covar_lt = AddedDiagLazyTensor(var_lt, cov_mat_lt) qdist = MultivariateNormal(mean, covar_lt) with gpytorch.settings.num_trace_samples( 1 ) and gpytorch.settings.max_cg_iterations(25): return qdist.log_prob(vec) def block_logdet(self, var, cov_mat_root): var = flatten(var) cov_mat_lt = RootLazyTensor(cov_mat_root.t()) var_lt = DiagLazyTensor(var + 1e-6) covar_lt = AddedDiagLazyTensor(var_lt, cov_mat_lt) return covar_lt.log_det() def block_logll(self, param_list, mean_list, var_list, cov_mat_root_list): full_logprob = 0 for i, (param, mean, var, cov_mat_root) in enumerate( zip(param_list, mean_list, var_list, cov_mat_root_list) ): # print('Block: ', i) block_ll = self.compute_ll_for_block(param, mean, var, cov_mat_root) full_logprob += block_ll return full_logprob def full_logll(self, param_list, mean_list, var_list, cov_mat_root_list): cov_mat_root = torch.cat(cov_mat_root_list, dim=1) mean_vector = flatten(mean_list) var_vector = flatten(var_list) param_vector = flatten(param_list) return self.compute_ll_for_block( param_vector, mean_vector, var_vector, cov_mat_root ) def compute_logdet(self, block=False): _, var_list, covar_mat_root_list = self.generate_mean_var_covar() if block: full_logdet = 0 for (var, cov_mat_root) in zip(var_list, covar_mat_root_list): block_logdet = self.block_logdet(var, cov_mat_root) full_logdet += block_logdet else: var_vector = flatten(var_list) cov_mat_root = torch.cat(covar_mat_root_list, dim=1) full_logdet = self.block_logdet(var_vector, cov_mat_root) return full_logdet def diag_logll(self, param_list, mean_list, var_list): logprob = 0.0 for param, mean, scale in zip(param_list, mean_list, var_list): logprob += Normal(mean, scale).log_prob(param).sum() return logprob def compute_logprob(self, vec=None, block=False, diag=False): mean_list, var_list, covar_mat_root_list = self.generate_mean_var_covar() if vec is None: param_list = [getattr(param, name) for param, name in self.params] else: param_list = unflatten_like(vec, mean_list) if diag: return self.diag_logll(param_list, mean_list, var_list) elif block is True: return self.block_logll( param_list, mean_list, var_list, covar_mat_root_list ) else: return self.full_logll(param_list, mean_list, var_list, covar_mat_root_list)
11,331
34.63522
88
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/losses.py
import torch import torch.nn.functional as F class GaussianLikelihood: """ Minus Gaussian likelihood for regression problems. Mean squared error (MSE) divided by `2 * noise_var`. """ def __init__(self, noise_var = 0.5): self.noise_var = noise_var self.mse = torch.nn.functional.mse_loss def __call__(self, model, input, target): output = model(input) if self.noise_var is not None: mse = self.mse(output, target) loss = mse / (2 * self.noise_var) return loss, output, {"mse": mse} else: mean = output[:,0].view_as(target) var = output[:,1].view_as(target) mse = self.mse(mean, target, reduction='none') mean_portion = mse / (2 * var) var_portion = 0.5 * torch.log(var) loss = mean_portion + var_portion return loss.mean(), output[:,0], {'mse': torch.mean((mean - target)**2.0)} def cross_entropy(model, input, target): # standard cross-entropy loss function output = model(input) loss = F.cross_entropy(output, target) return loss, output, {} def cross_entropy_output(output, target): # standard cross-entropy loss function loss = F.cross_entropy(output, target) return loss, {} def adversarial_cross_entropy(model, input, target, lossfn = F.cross_entropy, epsilon = 0.01): # loss function based on algorithm 1 of "simple and scalable uncertainty estimation using # deep ensembles," lakshminaraynan, pritzel, and blundell, nips 2017, # https://arxiv.org/pdf/1612.01474.pdf # note: the small difference bw this paper is that here the loss is only backpropped # through the adversarial loss rather than both due to memory constraints on preresnets # we can change back if we want to restrict ourselves to VGG-like networks (where it's fine). #scale epsilon by min and max (should be [0,1] for all experiments) #see algorithm 1 of paper scaled_epsilon = epsilon * (input.max() - input.min()) #force inputs to require gradient input.requires_grad = True #standard forwards pass output = model(input) loss = lossfn(output, target) #now compute gradients wrt input loss.backward(retain_graph = True) #now compute sign of gradients inputs_grad = torch.sign(input.grad) #perturb inputs and use clamped output inputs_perturbed = torch.clamp(input + scaled_epsilon * inputs_grad, 0.0, 1.0).detach() #inputs_perturbed.requires_grad = False input.grad.zero_() #model.zero_grad() outputs_perturbed = model(inputs_perturbed) #compute adversarial version of loss adv_loss = lossfn(outputs_perturbed, target) #return mean of loss for reasonable scalings return (loss + adv_loss)/2.0, output, {} def masked_loss(y_pred, y_true, void_class = 11., weight=None, reduce = True): # masked version of crossentropy loss el = torch.ones_like(y_true) * void_class mask = torch.ne(y_true, el).long() y_true_tmp = y_true * mask loss = F.cross_entropy(y_pred, y_true_tmp, weight=weight, reduction='none') loss = mask.float() * loss if reduce: return loss.sum()/mask.sum() else: return loss, mask def seg_cross_entropy(model, input, target, weight = None): output = model(input) # use masked loss function loss = masked_loss(output, target, weight=weight) return {'loss': loss, 'output': output} def seg_ale_cross_entropy(model, input, target, num_samples = 50, weight = None): #requires two outputs for model(input) output = model(input) mean = output[:, 0, :, :, :] scale = output[:, 1, :, :, :].abs() output_distribution = torch.distributions.Normal(mean, scale) total_loss = 0 for _ in range(num_samples): sample = output_distribution.rsample() current_loss, mask = masked_loss(sample, target, weight=weight, reduce=False) total_loss = total_loss + current_loss.exp() mean_loss = total_loss / num_samples return {'loss': mean_loss.log().sum() / mask.sum(), 'output': mean, 'scale': scale}
4,258
29.862319
97
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/utils.py
import itertools import torch import os import copy from datetime import datetime import math import numpy as np import tqdm from collections import defaultdict from time import gmtime, strftime import sys import torch.nn.functional as F def get_logging_print(fname): cur_time = strftime("%m-%d_%H:%M:%S", gmtime()) def print_func(*args): str_to_write = ' '.join(map(str, args)) filename = fname % cur_time if '%s' in fname else fname with open(filename, 'a') as f: f.write(str_to_write + '\n') f.flush() print(str_to_write) sys.stdout.flush() return print_func, fname % cur_time if '%s' in fname else fname def flatten(lst): tmp = [i.contiguous().view(-1,1) for i in lst] return torch.cat(tmp).view(-1) def unflatten_like(vector, likeTensorList): # Takes a flat torch.tensor and unflattens it to a list of torch.tensors # shaped like likeTensorList outList = [] i=0 for tensor in likeTensorList: #n = module._parameters[name].numel() n = tensor.numel() outList.append(vector[:,i:i+n].view(tensor.shape)) i+=n return outList def LogSumExp(x,dim=0): m,_ = torch.max(x,dim=dim,keepdim=True) return m + torch.log((x - m).exp().sum(dim=dim,keepdim=True)) def adjust_learning_rate(optimizer, lr): for param_group in optimizer.param_groups: param_group['lr'] = lr return lr def save_checkpoint(dir, epoch=None, name='checkpoint', sample=None, **kwargs): os.makedirs(dir, exist_ok=True) state = { 'epoch': epoch, 'sample': sample, } if epoch is not None: name = f'{name}-{epoch:05}.pt' elif sample is not None: name = f'{name}-{sample:05}.pt' else: name = f'{name}.pt' state.update(kwargs) filepath = os.path.join(dir, name) torch.save(state, filepath) def train_epoch(loader, model, criterion, optimizer, cuda=True, regression=False, verbose=False, subset=None): loss_sum = 0.0 stats_sum = defaultdict(float) correct = 0.0 verb_stage = 0 num_objects_current = 0 num_batches = len(loader) model.train() if subset is not None: num_batches = int(num_batches * subset) loader = itertools.islice(loader, num_batches) if verbose: loader = tqdm.tqdm(loader, total=num_batches) for i, (input, target) in enumerate(loader): if cuda: input = input.cuda(non_blocking=True) target = target.cuda(non_blocking=True) loss, output, stats = criterion(model, input, target) optimizer.zero_grad() loss.backward() optimizer.step() loss_sum += loss.data.item() * input.size(0) for key, value in stats.items(): stats_sum[key] += value * input.size(0) if not regression: pred = output.data.argmax(1, keepdim=True) correct += pred.eq(target.data.view_as(pred)).sum().item() num_objects_current += input.size(0) if verbose and 10 * (i + 1) / num_batches >= verb_stage + 1: print('Stage %d/10. Loss: %12.4f. Acc: %6.2f' % ( verb_stage + 1, loss_sum / num_objects_current, correct / num_objects_current * 100.0 )) verb_stage += 1 return { 'loss': loss_sum / num_objects_current, 'accuracy': None if regression else correct / num_objects_current * 100.0, 'stats': {key: value / num_objects_current for key, value in stats_sum.items()} } def eval(loader, model, criterion, cuda=True, regression=False, verbose=False): loss_sum = 0.0 correct = 0.0 stats_sum = defaultdict(float) num_objects_total = len(loader.dataset) model.eval() with torch.no_grad(): if verbose: loader = tqdm.tqdm(loader) for i, (input, target) in enumerate(loader): if cuda: input = input.cuda(non_blocking=True) target = target.cuda(non_blocking=True) loss, output, stats = criterion(model, input, target) loss_sum += loss.item() * input.size(0) for key, value in stats.items(): stats_sum[key] += value if not regression: pred = output.data.argmax(1, keepdim=True) correct += pred.eq(target.data.view_as(pred)).sum().item() return { 'loss': loss_sum / num_objects_total, 'accuracy': None if regression else correct / num_objects_total * 100.0, 'stats': {key: value / num_objects_total for key, value in stats_sum.items()} } def predict(loader, model, verbose=False): predictions = list() targets = list() model.eval() if verbose: loader = tqdm.tqdm(loader) offset = 0 with torch.no_grad(): for input, target in loader: input = input.cuda(non_blocking=True) output = model(input) batch_size = input.size(0) predictions.append(F.softmax(output, dim=1).cpu().numpy()) targets.append(target.numpy()) offset += batch_size return { 'predictions': np.vstack(predictions), 'targets': np.concatenate(targets) } def moving_average(net1, net2, alpha=1): for param1, param2 in zip(net1.parameters(), net2.parameters()): param1.data *= (1.0 - alpha) param1.data += param2.data * alpha def _check_bn(module, flag): if issubclass(module.__class__, torch.nn.modules.batchnorm._BatchNorm): flag[0] = True def check_bn(model): flag = [False] model.apply(lambda module: _check_bn(module, flag)) return flag[0] def reset_bn(module): if issubclass(module.__class__, torch.nn.modules.batchnorm._BatchNorm): module.running_mean = torch.zeros_like(module.running_mean) module.running_var = torch.ones_like(module.running_var) def _get_momenta(module, momenta): if issubclass(module.__class__, torch.nn.modules.batchnorm._BatchNorm): momenta[module] = module.momentum def _set_momenta(module, momenta): if issubclass(module.__class__, torch.nn.modules.batchnorm._BatchNorm): module.momentum = momenta[module] def bn_update(loader, model, verbose=False, subset=None, **kwargs): """ BatchNorm buffers update (if any). Performs 1 epochs to estimate buffers average using train dataset. :param loader: train dataset loader for buffers average estimation. :param model: model being update :return: None """ if not check_bn(model): return model.train() momenta = {} model.apply(reset_bn) model.apply(lambda module: _get_momenta(module, momenta)) n = 0 num_batches = len(loader) with torch.no_grad(): if subset is not None: num_batches = int(num_batches * subset) loader = itertools.islice(loader, num_batches) if verbose: loader = tqdm.tqdm(loader, total=num_batches) for input, _ in loader: input = input.cuda(non_blocking=True) input_var = torch.autograd.Variable(input) b = input_var.data.size(0) momentum = b / (n + b) for module in momenta.keys(): module.momentum = momentum model(input_var, **kwargs) n += b model.apply(lambda module: _set_momenta(module, momenta)) def inv_softmax(x, eps = 1e-10): return torch.log(x/(1.0 - x + eps)) def predictions(test_loader, model, seed=None, cuda=True, regression=False, **kwargs): #will assume that model is already in eval mode #model.eval() preds = [] targets = [] for input, target in test_loader: if seed is not None: torch.manual_seed(seed) if cuda: input = input.cuda(non_blocking=True) output = model(input, **kwargs) if regression: preds.append(output.cpu().data.numpy()) else: probs = F.softmax(output, dim=1) preds.append(probs.cpu().data.numpy()) targets.append(target.numpy()) return np.vstack(preds), np.concatenate(targets) def schedule(epoch, lr_init, epochs, swa, swa_start=None, swa_lr=None): t = (epoch) / (swa_start if swa else epochs) lr_ratio = swa_lr / lr_init if swa else 0.01 if t <= 0.5: factor = 1.0 elif t <= 0.9: factor = 1.0 - (1.0 - lr_ratio) * (t - 0.5) / 0.4 else: factor = lr_ratio return lr_init * factor def set_weights(model, vector, device=None): offset = 0 for param in model.parameters(): param.data.copy_(vector[offset:offset + param.numel()].view(param.size()).to(device)) offset += param.numel() def extract_parameters(model): params = [] for module in model.modules(): for name in list(module._parameters.keys()): if module._parameters[name] is None: continue param = module._parameters[name] params.append((module, name, param.size())) module._parameters.pop(name) return params def set_weights_old(params, w, device): offset = 0 for module, name, shape in params: size = np.prod(shape) value = w[offset:offset + size] setattr(module, name, value.view(shape).to(device)) offset += size
9,479
28.349845
110
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/data.py
import numpy as np import torch import torchvision import os c10_classes = np.array([ [0, 1, 2, 8, 9], [3, 4, 5, 6, 7] ], dtype=np.int32) def camvid_loaders(path, batch_size, num_workers, transform_train, transform_test, use_validation, val_size, shuffle_train=True, joint_transform=None, ft_joint_transform=None, ft_batch_size=1, **kwargs): #load training and finetuning datasets print(path) train_set = CamVid(root=path, split='train', joint_transform=joint_transform, transform=transform_train, **kwargs) ft_train_set = CamVid(root=path, split='train', joint_transform=ft_joint_transform, transform=transform_train, **kwargs) val_set = CamVid(root=path, split='val', joint_transform=None, transform=transform_test, **kwargs) test_set = CamVid(root=path, split='test', joint_transform=None, transform=transform_test, **kwargs) num_classes = 11 # hard coded labels ehre return {'train': torch.utils.data.DataLoader( train_set, batch_size=batch_size, shuffle=shuffle_train, num_workers=num_workers, pin_memory=True ), 'fine_tune': torch.utils.data.DataLoader( ft_train_set, batch_size=ft_batch_size, shuffle=shuffle_train, num_workers=num_workers, pin_memory=True ), 'val': torch.utils.data.DataLoader( val_set, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=True ), 'test': torch.utils.data.DataLoader( test_set, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=True )}, num_classes def svhn_loaders(path, batch_size, num_workers, transform_train, transform_test, use_validation, val_size, shuffle_train=True): train_set = torchvision.datasets.SVHN(root=path, split='train', download = True, transform = transform_train) if use_validation: test_set = torchvision.datasets.SVHN(root=path, split='train', download = True, transform = transform_test) train_set.data = train_set.data[:-val_size] train_set.labels = train_set.labels[:-val_size] test_set.data = test_set.data[-val_size:] test_set.labels = test_set.labels[-val_size:] else: print('You are going to run models on the test set. Are you sure?') test_set = torchvision.datasets.SVHN(root=path, split='test', download = True, transform = transform_test) num_classes = 10 return \ { 'train': torch.utils.data.DataLoader( train_set, batch_size=batch_size, shuffle=True and shuffle_train, num_workers=num_workers, pin_memory=True ), 'test': torch.utils.data.DataLoader( test_set, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=True ), }, \ num_classes def loaders(dataset, path, batch_size, num_workers, transform_train, transform_test, use_validation=True, val_size=5000, split_classes=None, shuffle_train=True, **kwargs): if dataset == 'CamVid': return camvid_loaders(path, batch_size=batch_size, num_workers=num_workers, transform_train=transform_train, transform_test=transform_test, use_validation=use_validation, val_size=val_size, **kwargs) path = os.path.join(path, dataset.lower()) ds = getattr(torchvision.datasets, dataset) if dataset == 'SVHN': return svhn_loaders(path, batch_size, num_workers, transform_train, transform_test, use_validation, val_size) else: ds = getattr(torchvision.datasets, dataset) if dataset == 'STL10': train_set = ds(root=path, split='train', download=True, transform=transform_train) num_classes = 10 cls_mapping = np.array([0, 2, 1, 3, 4, 5, 7, 6, 8, 9]) train_set.labels = cls_mapping[train_set.labels] elif dataset == 'ImageNet': train_set = ds(root=path, split='train', transform=transform_train) else: train_set = ds(root=path, train=True, download=True, transform=transform_train) num_classes = max(train_set.train_labels) + 1 if use_validation: print("Using train (" + str(len(train_set.train_data)-val_size) + ") + validation (" +str(val_size)+ ")") train_set.train_data = train_set.train_data[:-val_size] train_set.train_labels = train_set.train_labels[:-val_size] test_set = ds(root=path, train=True, download=True, transform=transform_test) test_set.train = False test_set.test_data = test_set.train_data[-val_size:] test_set.test_labels = test_set.train_labels[-val_size:] delattr(test_set, 'train_data') delattr(test_set, 'train_labels') else: print('You are going to run models on the test set. Are you sure?') if dataset == 'STL10': test_set = ds(root=path, split='test', download=True, transform=transform_test) test_set.labels = cls_mapping[test_set.labels] else: test_set = ds(root=path, train=False, download=True, transform=transform_test) if split_classes is not None: assert dataset == 'CIFAR10' assert split_classes in {0, 1} print('Using classes:', end='') print(c10_classes[split_classes]) train_mask = np.isin(train_set.train_labels, c10_classes[split_classes]) train_set.train_data = train_set.train_data[train_mask, :] train_set.train_labels = np.array(train_set.train_labels)[train_mask] train_set.train_labels = np.where(train_set.train_labels[:, None] == c10_classes[split_classes][None, :])[1].tolist() print('Train: %d/%d' % (train_set.train_data.shape[0], train_mask.size)) test_mask = np.isin(test_set.test_labels, c10_classes[split_classes]) test_set.test_data = test_set.test_data[test_mask, :] test_set.test_labels = np.array(test_set.test_labels)[test_mask] test_set.test_labels = np.where(test_set.test_labels[:, None] == c10_classes[split_classes][None, :])[1].tolist() print('Test: %d/%d' % (test_set.test_data.shape[0], test_mask.size)) return \ { 'train': torch.utils.data.DataLoader( train_set, batch_size=batch_size, shuffle=True and shuffle_train, num_workers=num_workers, pin_memory=True ), 'test': torch.utils.data.DataLoader( test_set, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=True ), }, \ num_classes def loaders_inc(dataset, path, num_chunks, batch_size, num_workers, transform_train, transform_test, use_validation=True, val_size=5000, shuffle_train=True, seed=1): assert dataset in {'MNIST', 'CIFAR10', 'CIFAR100'} path = os.path.join(path, dataset.lower()) ds = getattr(torchvision.datasets, dataset) train_set = ds(root=path, train=True, download=True, transform=transform_train) num_classes = int(max(train_set.train_labels)) + 1 num_samples = (train_set.train_data.shape[0] - val_size) if use_validation else train_set.train_data.shape[0] train_sets = list() offset = 0 random_state = np.random.RandomState(seed) order = random_state.permutation(train_set.train_data.shape[0]) for i in range(num_chunks, 0, -1): chunk_size = (num_samples + i - 1) // i tmp_set = ds(root=path, train=True, download=True, transform=transform_train) tmp_set.train_data = tmp_set.train_data[order[offset:offset + chunk_size]] tmp_set.train_labels = np.array(tmp_set.train_labels)[order[offset:offset + chunk_size]] train_sets.append(tmp_set) offset += chunk_size num_samples -= chunk_size print('Using train %d chunks: %s' % (num_chunks, str([tmp_set.train_data.shape[0] for tmp_set in train_sets]))) if use_validation: print('Using validation (%d)' % val_size) test_set = ds(root=path, train=True, download=True, transform=transform_test) test_set.train = False test_set.test_data = test_set.train_data[order[-val_size:]] test_set.test_labels = np.array(test_set.train_labels)[order[-val_size:]] delattr(test_set, 'train_data') delattr(test_set, 'train_labels') else: print('You are going to run models on the test set. Are you sure?') test_set = ds(root=path, train=False, download=True, transform=transform_test) return \ { 'train': [ torch.utils.data.DataLoader( tmp_set, batch_size=batch_size, shuffle=True and shuffle_train, num_workers=num_workers, pin_memory=True ) for tmp_set in train_sets ], 'test': torch.utils.data.DataLoader( test_set, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=True ), }, \ num_classes
9,867
39.77686
165
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/preresnet.py
""" PreResNet model definition ported from https://github.com/bearpaw/pytorch-classification/blob/master/models/cifar/preresnet.py """ import torch.nn as nn import torchvision.transforms as transforms import math __all__ = ['PreResNet110', 'PreResNet56', 'PreResNet8', 'PreResNet83', 'PreResNet164'] def conv3x3(in_planes, out_planes, stride=1): return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.bn1 = nn.BatchNorm2d(inplanes) self.relu = nn.ReLU(inplace=True) self.conv1 = conv3x3(inplanes, planes, stride) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = conv3x3(planes, planes) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.bn1(x) out = self.relu(out) out = self.conv1(out) out = self.bn2(out) out = self.relu(out) out = self.conv2(out) if self.downsample is not None: residual = self.downsample(x) out += residual return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.bn1 = nn.BatchNorm2d(inplanes) self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn3 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.bn1(x) out = self.relu(out) out = self.conv1(out) out = self.bn2(out) out = self.relu(out) out = self.conv2(out) out = self.bn3(out) out = self.relu(out) out = self.conv3(out) if self.downsample is not None: residual = self.downsample(x) out += residual return out class PreResNet(nn.Module): def __init__(self, num_classes=10, depth=110): super(PreResNet, self).__init__() if depth >= 44: assert (depth - 2) % 9 == 0, 'depth should be 9n+2' n = (depth - 2) // 9 block = Bottleneck else: assert (depth - 2) % 6 == 0, 'depth should be 6n+2' n = (depth - 2) // 6 block = BasicBlock self.inplanes = 16 self.conv1 = nn.Conv2d(3, 16, kernel_size=3, padding=1, bias=False) self.layer1 = self._make_layer(block, 16, n) self.layer2 = self._make_layer(block, 32, n, stride=2) self.layer3 = self._make_layer(block, 64, n, stride=2) self.bn = nn.BatchNorm2d(64 * block.expansion) self.relu = nn.ReLU(inplace=True) self.avgpool = nn.AvgPool2d(8) self.fc = nn.Linear(64 * block.expansion, num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), ) layers = list() layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.layer1(x) # 32x32 x = self.layer2(x) # 16x16 x = self.layer3(x) # 8x8 x = self.bn(x) x = self.relu(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.fc(x) return x class PreResNet164: base = PreResNet args = list() kwargs = {'depth': 164} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) class PreResNet110: base = PreResNet args = list() kwargs = {'depth': 110} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) class PreResNet83: base = PreResNet args = list() kwargs = {'depth': 83} transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) class PreResNet56: base = PreResNet args = list() kwargs = {'depth': 56} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) class PreResNet8: base = PreResNet args = list() kwargs = {'depth': 8} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ])
7,349
30.410256
103
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/regression_net.py
import math import torch import torch.nn as nn import torchvision.transforms as transforms try: import os os.sys.path.append("/home/izmailovpavel/Documents/Projects/curves/") import curves except: pass __all__ = [ 'RegNet', 'ToyRegNet', ] class MDropout(torch.nn.Module): def __init__(self, dim, p, inplace=False): super(MDropout, self).__init__() self.dim = dim self.p = p self.inplace = inplace self.register_buffer('mask', torch.ones(dim, dtype=torch.long)) def forward(self, input): if self.training: return torch.nn.functional.dropout(input, self.p, self.training, self.inplace) else: return input * self.mask.float().view(1, -1) * 1.0 / (1.0 - self.p) def sample(self): self.mask.bernoulli_(1.0 - self.p) def sample_masks(module): if isinstance(module, MDropout): module.sample() class SplitDim(nn.Module): def __init__(self, nonlin_col=1, nonlin_type=torch.nn.functional.softplus, correction = True): super(SplitDim, self).__init__() self.nonlinearity = nonlin_type self.col = nonlin_col if correction: self.var = torch.nn.Parameter(torch.zeros(1)) else: #equivalent to about 3e-7 when using softplus self.register_buffer('var', torch.ones(1, requires_grad = False)*-15.) self.correction = correction def forward(self, input): transformed_output = self.nonlinearity(input[:,self.col]) transformed_output = (transformed_output + self.nonlinearity(self.var)) stack_list = [input[:,:self.col], transformed_output.view(-1,1)] if self.col+1 < input.size(1): stack_list.append(input[:,(self.col+1):]) #print(self.nonlinearity(self.var).item(), transformed_output.mean().item()) output = torch.cat(stack_list,1) return output class RegNetBase(nn.Sequential): def __init__(self, dimensions, input_dim=1, output_dim=1, dropout=None, apply_var=True): super(RegNetBase, self).__init__() self.dimensions = [input_dim, *dimensions, output_dim] for i in range(len(self.dimensions) - 1): if dropout is not None and i > 0: self.add_module('dropout%d' % i, MDropout(self.dimensions[i], p=dropout)) self.add_module('linear%d' % i, torch.nn.Linear(self.dimensions[i], self.dimensions[i + 1])) if i < len(self.dimensions) - 2: self.add_module('relu%d' % i, torch.nn.ReLU()) if output_dim == 2: self.add_module('var_split', SplitDim(correction=apply_var)) def forward(self, x, output_features=False): if not output_features: return super().forward(x) else: print(self._modules.values()) print(list(self._modules.values())[:-2]) for module in list(self._modules.values())[:-3]: x = module(x) print(x.size()) return x class RegNetCurve(nn.Sequential): def __init__(self, dimensions, fix_points, input_dim=1, output_dim=1, dropout=None): super(RegNetCurve, self).__init__() self.dimensions = [input_dim, *dimensions, output_dim] for i in range(len(self.dimensions) - 1): if dropout is not None and i > 0: self.add_module('dropout%d' % i, MDropout(self.dimensions[i], p=dropout)) self.add_module('linear%d' % i, curves.Linear(self.dimensions[i], self.dimensions[i + 1], fix_points=fix_points)) if i < len(self.dimensions) - 2: self.add_module('tanh%d' % i, torch.nn.ReLU()) def forward(self, x, t): for module in self._modules.values(): if isinstance(module, curves.Linear): x = module(x, t) else: x = module(x) return x class RegNet: base = RegNetBase curve = RegNetCurve args = list() kwargs = {"dimensions": [1000, 1000, 500, 50, 2]} transform_train = transforms.Compose([ transforms.ToTensor(), ]) transform_test = transforms.Compose([ transforms.ToTensor(), ]) class ToyRegNet: base = RegNetBase curve = RegNetCurve args = list() kwargs = {"dimensions": [200, 50, 50, 50], "output_dim": 1, "input_dim": 2} transform_train = transforms.Compose([ transforms.ToTensor(), ]) transform_test = transforms.Compose([ transforms.ToTensor(), ])
4,672
31.006849
125
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/vgg.py
""" VGG model definition ported from https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py """ import math import torch.nn as nn import torchvision.transforms as transforms __all__ = ['VGG16', 'VGG16BN', 'VGG19', 'VGG19BN'] def make_layers(cfg, batch_norm=False): layers = list() in_channels = 3 for v in cfg: if v == 'M': layers += [nn.MaxPool2d(kernel_size=2, stride=2)] else: conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1) if batch_norm: layers += [conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)] else: layers += [conv2d, nn.ReLU(inplace=True)] in_channels = v return nn.Sequential(*layers) cfg = { 16: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'], 19: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'], } class VGG(nn.Module): def __init__(self, num_classes=10, depth=16, batch_norm=False): super(VGG, self).__init__() self.features = make_layers(cfg[depth], batch_norm) self.classifier = nn.Sequential( nn.Dropout(), nn.Linear(512, 512), nn.ReLU(True), nn.Dropout(), nn.Linear(512, 512), nn.ReLU(True), nn.Linear(512, num_classes), ) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) m.bias.data.zero_() def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.classifier(x) return x class Base: base = VGG args = list() kwargs = dict() transform_train = transforms.Compose([ transforms.RandomHorizontalFlip(), transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) #transforms.Normalize((0.4376821 , 0.4437697 , 0.47280442), (0.19803012, 0.20101562, 0.19703614)) ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) #transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) #transforms.Normalize((0.45242316, 0.45249584, 0.46897713), (0.21943445, 0.22656967, 0.22850613)) ]) class VGG16(Base): pass class VGG16BN(Base): kwargs = {'batch_norm': True} class VGG19(Base): kwargs = {'depth': 19} class VGG19BN(Base): kwargs = {'depth': 19, 'batch_norm': True}
2,841
27.707071
105
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/mlp.py
import torch.nn as nn import torchvision.transforms as transforms import torch __all__=['MLP', 'MLPBoston'] class MLPBase(nn.Module): def __init__(self, num_classes=0, in_dim=1, layers=2, hidden=7): super(MLPBase, self).__init__() out_layer_list = [hidden for i in range(layers)] if num_classes == 0: out_layer_list.append(1) #for regression else: out_layer_list.append(num_classes) in_layer_list = [hidden for i in range(layers)] in_layer_list.insert(0, in_dim) layers = [] for input, output in zip(in_layer_list, out_layer_list): layers.append(nn.Linear(input, output)) #add relu activations layers.append(nn.ReLU()) layers.pop() #remove final relu layer self.model = nn.Sequential(*layers) #self.log_noise = nn.Parameter(torch.log(torch.ones(1)*7)) print(self.model) def forward(self, x): return self.model(x) class MLP: base = MLPBase args = list() kwargs = {} transform_train = transforms.ToTensor() transform_test = transforms.ToTensor() class MLPBoston: base = MLPBase base.log_noise = nn.Parameter(torch.log(torch.ones(1)*7)) args = list() kwargs = {'in_dim': 13, 'layers': 1, 'hidden':50} transform_train = transforms.ToTensor() transform_test = transforms.ToTensor()
1,419
27.979592
68
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/layers.py
""" layer definitions for 100-layer tiramisu #from: https://github.com/bfortuner/pytorch_tiramisu """ import torch import torch.nn as nn class DenseLayer(nn.Sequential): def __init__(self, in_channels, growth_rate): super().__init__() self.add_module('norm', nn.BatchNorm2d(in_channels)) self.add_module('relu', nn.ReLU(True)) self.add_module('conv', nn.Conv2d(in_channels, growth_rate, kernel_size=3, stride=1, padding=1, bias=True)) self.add_module('drop', nn.Dropout(p=0.2)) def forward(self, x): return super().forward(x) class DenseBlock(nn.Module): def __init__(self, in_channels, growth_rate, n_layers, upsample=False): super().__init__() self.upsample = upsample self.layers = nn.ModuleList([DenseLayer( in_channels + i*growth_rate, growth_rate) for i in range(n_layers)]) def forward(self, x): if self.upsample: new_features = [] #we pass all previous activations into each dense layer normally #But we only store each dense layer's output in the new_features array for layer in self.layers: out = layer(x) x = torch.cat([x, out], 1) new_features.append(out) return torch.cat(new_features,1) else: for layer in self.layers: out = layer(x) x = torch.cat([x, out], 1) # 1 = channel axis return x class TransitionDown(nn.Sequential): def __init__(self, in_channels): super().__init__() self.add_module('norm', nn.BatchNorm2d(num_features=in_channels)) self.add_module('relu', nn.ReLU(inplace=True)) self.add_module('conv', nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0, bias=True)) self.add_module('drop', nn.Dropout2d(0.2)) self.add_module('maxpool', nn.MaxPool2d(2)) def forward(self, x): return super().forward(x) class TransitionUp(nn.Module): def __init__(self, in_channels, out_channels): super().__init__() self.convTrans = nn.ConvTranspose2d( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=2, padding=0, bias=True) def forward(self, x, skip): out = self.convTrans(x) out = center_crop(out, skip.size(2), skip.size(3)) out = torch.cat([out, skip], 1) return out class Bottleneck(nn.Sequential): def __init__(self, in_channels, growth_rate, n_layers): super().__init__() self.add_module('bottleneck', DenseBlock( in_channels, growth_rate, n_layers, upsample=True)) def forward(self, x): return super().forward(x) def center_crop(layer, max_height, max_width): _, _, h, w = layer.size() xy1 = (w - max_width) // 2 xy2 = (h - max_height) // 2 return layer[:, :, xy2:(xy2 + max_height), xy1:(xy1 + max_width)]
3,117
33.644444
82
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/wide_resnet.py
""" WideResNet model definition ported from https://github.com/meliketoy/wide-resnet.pytorch/blob/master/networks/wide_resnet.py """ import torchvision.transforms as transforms import torch.nn as nn import torch.nn.init as init import torch.nn.functional as F import math __all__ = ['WideResNet28x10'] def conv3x3(in_planes, out_planes, stride=1): return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=True) def conv_init(m): classname = m.__class__.__name__ if classname.find('Conv') != -1: init.xavier_uniform(m.weight, gain=math.sqrt(2)) init.constant(m.bias, 0) elif classname.find('BatchNorm') != -1: init.constant(m.weight, 1) init.constant(m.bias, 0) class WideBasic(nn.Module): def __init__(self, in_planes, planes, dropout_rate, stride=1): super(WideBasic, self).__init__() self.bn1 = nn.BatchNorm2d(in_planes) self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, bias=True) self.dropout = nn.Dropout(p=dropout_rate) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=True) self.shortcut = nn.Sequential() if stride != 1 or in_planes != planes: self.shortcut = nn.Sequential( nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride, bias=True), ) def forward(self, x): out = self.dropout(self.conv1(F.relu(self.bn1(x)))) out = self.conv2(F.relu(self.bn2(out))) out += self.shortcut(x) return out class WideResNet(nn.Module): def __init__(self, num_classes=10, depth=28, widen_factor=10, dropout_rate=0.): super(WideResNet, self).__init__() self.in_planes = 16 assert ((depth - 4) % 6 == 0), 'Wide-resnet depth should be 6n+4' n = (depth - 4) / 6 k = widen_factor nstages = [16, 16 * k, 32 * k, 64 * k] self.conv1 = conv3x3(3, nstages[0]) self.layer1 = self._wide_layer(WideBasic, nstages[1], n, dropout_rate, stride=1) self.layer2 = self._wide_layer(WideBasic, nstages[2], n, dropout_rate, stride=2) self.layer3 = self._wide_layer(WideBasic, nstages[3], n, dropout_rate, stride=2) self.bn1 = nn.BatchNorm2d(nstages[3], momentum=0.9) self.linear = nn.Linear(nstages[3], num_classes) def _wide_layer(self, block, planes, num_blocks, dropout_rate, stride): strides = [stride] + [1] * int(num_blocks - 1) layers = [] for stride in strides: layers.append(block(self.in_planes, planes, dropout_rate, stride)) self.in_planes = planes return nn.Sequential(*layers) def forward(self, x): out = self.conv1(x) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = F.relu(self.bn1(out)) out = F.avg_pool2d(out, 8) out = out.view(out.size(0), -1) out = self.linear(out) return out class WideResNet28x10: base = WideResNet args = list() kwargs = {'depth': 28, 'widen_factor': 10} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ])
3,660
32.587156
100
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/vgg_dropout.py
""" VGG model definition ported from https://github.com/pytorch/vision/blob/master/torchvision/models/vgg.py """ import math import torch.nn as nn import torchvision.transforms as transforms __all__ = ['VGG16Drop', 'VGG16BNDrop', 'VGG19Drop', 'VGG19BNDrop'] P = 0.05 def make_layers(cfg, batch_norm=False): layers = list() in_channels = 3 for v in cfg: if v == 'M': layers += [nn.MaxPool2d(kernel_size=2, stride=2)] else: conv2d = nn.Conv2d(in_channels, v, kernel_size=3, padding=1) if batch_norm: layers += [nn.Dropout(P), conv2d, nn.BatchNorm2d(v), nn.ReLU(inplace=True)] else: layers += [nn.Dropout(P), conv2d, nn.ReLU(inplace=True)] in_channels = v return nn.Sequential(*layers) cfg = { 16: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512, 'M', 512, 512, 512, 'M'], 19: [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 256, 'M', 512, 512, 512, 512, 'M', 512, 512, 512, 512, 'M'], } class VGGDrop(nn.Module): def __init__(self, num_classes=10, depth=16, batch_norm=False): super(VGGDrop, self).__init__() self.features = make_layers(cfg[depth], batch_norm) self.classifier = nn.Sequential( nn.Dropout(P), nn.Linear(512, 512), nn.ReLU(True), nn.Dropout(P), nn.Linear(512, 512), nn.ReLU(True), nn.Linear(512, num_classes), ) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) m.bias.data.zero_() def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.classifier(x) return x class Base: base = VGGDrop args = list() kwargs = dict() transform_train = transforms.Compose([ transforms.RandomHorizontalFlip(), transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) #transforms.Normalize((0.4376821 , 0.4437697 , 0.47280442), (0.19803012, 0.20101562, 0.19703614)) ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) #transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) #transforms.Normalize((0.45242316, 0.45249584, 0.46897713), (0.21943445, 0.22656967, 0.22850613)) ]) class VGG16Drop(Base): pass class VGG16BNDrop(Base): kwargs = {'batch_norm': True} class VGG19Drop(Base): kwargs = {'depth': 19} class VGG19BNDrop(Base): kwargs = {'depth': 19, 'batch_norm': True}
2,927
27.990099
105
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/wide_resnet_dropout.py
""" WideResNet model definition ported from https://github.com/meliketoy/wide-resnet.pytorch/blob/master/networks/wide_resnet.py """ import torchvision.transforms as transforms import torch.nn as nn import torch.nn.init as init import torch.nn.functional as F import math __all__ = ['WideResNet28x10Drop'] P = 0.05 def conv3x3(in_planes, out_planes, stride=1): return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=True) def conv_init(m): classname = m.__class__.__name__ if classname.find('Conv') != -1: init.xavier_uniform(m.weight, gain=math.sqrt(2)) init.constant(m.bias, 0) elif classname.find('BatchNorm') != -1: init.constant(m.weight, 1) init.constant(m.bias, 0) class WideBasic(nn.Module): def __init__(self, in_planes, planes, stride=1): super(WideBasic, self).__init__() self.bn1 = nn.BatchNorm2d(in_planes) self.dropout = nn.Dropout(p=P) self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, bias=True) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=True) self.shortcut = nn.Sequential() if stride != 1 or in_planes != planes: self.shortcut = nn.Sequential( nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride, bias=True), ) def forward(self, x): out = self.dropout(self.conv1(F.relu(self.bn1(x)))) out = self.conv2(F.relu(self.bn2(out))) out += self.shortcut(x) return out class WideResNetDrop(nn.Module): def __init__(self, num_classes=10, depth=28, widen_factor=10): super(WideResNetDrop, self).__init__() self.in_planes = 16 assert ((depth - 4) % 6 == 0), 'Wide-resnet depth should be 6n+4' n = (depth - 4) / 6 k = widen_factor nstages = [16, 16 * k, 32 * k, 64 * k] self.conv1 = conv3x3(3, nstages[0]) self.layer1 = self._wide_layer(WideBasic, nstages[1], n, stride=1) self.layer2 = self._wide_layer(WideBasic, nstages[2], n, stride=2) self.layer3 = self._wide_layer(WideBasic, nstages[3], n, stride=2) self.bn1 = nn.BatchNorm2d(nstages[3], momentum=0.9) self.linear = nn.Linear(nstages[3], num_classes) self.drop = nn.Dropout(P) def _wide_layer(self, block, planes, num_blocks, stride): strides = [stride] + [1] * int(num_blocks - 1) layers = [] for stride in strides: layers.append(block(self.in_planes, planes, stride)) self.in_planes = planes return nn.Sequential(*layers) def forward(self, x): out = self.conv1(x) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = F.relu(self.bn1(out)) out = F.avg_pool2d(out, 8) out = out.view(out.size(0), -1) out = self.drop(out) out = self.linear(out) return out class WideResNet28x10Drop: base = WideResNetDrop args = list() kwargs = {'depth': 28, 'widen_factor': 10} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ])
3,640
31.508929
100
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/models/preresnet_dropout.py
""" PreResNet model definition ported from https://github.com/bearpaw/pytorch-classification/blob/master/models/cifar/preresnet.py """ import torch.nn as nn import torchvision.transforms as transforms import math __all__ = ['PreResNet110Drop', 'PreResNet56Drop', 'PreResNet8Drop', 'PreResNet164Drop'] P = 0.01 def conv3x3(in_planes, out_planes, stride=1): return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.bn1 = nn.BatchNorm2d(inplanes) self.relu = nn.ReLU(inplace=True) self.drop = nn.Dropout(P) self.conv1 = conv3x3(inplanes, planes, stride) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = conv3x3(planes, planes) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.bn1(x) out = self.relu(out) out = self.drop(out) out = self.conv1(out) out = self.bn2(out) out = self.relu(out) out = self.drop(out) out = self.conv2(out) if self.downsample is not None: residual = self.downsample(x) out += residual return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.bn1 = nn.BatchNorm2d(inplanes) self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn3 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride self.drop = nn.Dropout(P) def forward(self, x): residual = x out = self.bn1(x) out = self.relu(out) out = self.drop(out) out = self.conv1(out) out = self.bn2(out) out = self.relu(out) out = self.drop(out) out = self.conv2(out) out = self.bn3(out) out = self.relu(out) out = self.drop(out) out = self.conv3(out) if self.downsample is not None: residual = self.downsample(x) out += residual return out class PreResNetDrop(nn.Module): def __init__(self, num_classes=10, depth=110): super(PreResNetDrop, self).__init__() assert (depth - 2) % 6 == 0, 'depth should be 6n+2' n = (depth - 2) // 6 block = Bottleneck if depth >= 44 else BasicBlock self.inplanes = 16 self.conv1 = nn.Conv2d(3, 16, kernel_size=3, padding=1, bias=False) self.layer1 = self._make_layer(block, 16, n) self.layer2 = self._make_layer(block, 32, n, stride=2) self.layer3 = self._make_layer(block, 64, n, stride=2) self.bn = nn.BatchNorm2d(64 * block.expansion) self.relu = nn.ReLU(inplace=True) self.avgpool = nn.AvgPool2d(8) self.fc = nn.Linear(64 * block.expansion, num_classes) self.drop = nn.Dropout(P) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Dropout(P), nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), ) layers = list() layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.layer1(x) # 32x32 x = self.layer2(x) # 16x16 x = self.layer3(x) # 8x8 x = self.bn(x) x = self.relu(x) x = self.avgpool(x) x = x.view(x.size(0), -1) x = self.drop(x) x = self.fc(x) return x class PreResNet164Drop: base = PreResNetDrop args = list() kwargs = {'depth': 164} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) class PreResNet110Drop: base = PreResNetDrop args = list() kwargs = {'depth': 110} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) class PreResNet56Drop: base = PreResNetDrop args = list() kwargs = {'depth': 56} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) class PreResNet8Drop: base = PreResNetDrop args = list() kwargs = {'depth': 8} transform_train = transforms.Compose([ transforms.Resize(32), transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) transform_test = transforms.Compose([ transforms.Resize(32), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ])
7,057
30.092511
103
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/ess.py
import torch import numpy as np from .elliptical_slice import elliptical_slice, slice_sample from .proj_model import ProjectedModel class EllipticalSliceSampling(torch.nn.Module): def __init__(self, base, subspace, var, loader, criterion, num_samples = 20, use_cuda = False, method='elliptical', *args, **kwargs): super(EllipticalSliceSampling, self).__init__() if method=='elliptical': self.slice_method = elliptical_slice if method=='slice': self.slice_method = slice_sample self.base_model = base(*args, **kwargs) if use_cuda: self.base_model.cuda() self.base_params = [] for name, param in self.base_model.named_parameters(): self.base_params.append([param, name, param.size()]) self.subspace = subspace self.var = var self.loader = loader self.criterion = criterion self.num_samples = num_samples self.use_cuda = use_cuda self.all_samples = None self.model = self.base_model def forward(self, *args, **kwargs): return self.model.forward(*args, **kwargs) def prior_sample(self, prior='identity', scale=1.0): if prior=='identity': cov_mat = np.eye(self.subspace.cov_factor.size(0)) elif prior=='schur': trans_cov_mat = self.subspace.cov_factor.matmul(self.subspace.cov_factor.subspace.t()).numpy() trans_cov_mat /= (self.swag_model.n_models.item() - 1) cov_mat = np.eye(self.subspace.cov_factor.size(0)) + trans_cov_mat else: raise NotImplementedError('Only schur and identity priors have been implemented') cov_mat *= scale sample = np.random.multivariate_normal(np.zeros(self.subspace.cov_factor.size(0)), cov_mat.astype(np.float64), 1)[0,:] return sample def log_pdf(self, params, temperature = 1., minibatch = False): params_tensor = torch.FloatTensor(params) params_tensor = params_tensor.view(-1) if self.use_cuda: params_tensor = params_tensor.cuda() with torch.no_grad(): proj_model = ProjectedModel(model=self.base_model, subspace = self.subspace, proj_params = params_tensor) loss = 0 num_datapoints = 0.0 for batch_num, (data, target) in enumerate(self.loader): if minibatch and batch_num > 0: break num_datapoints += data.size(0) if self.use_cuda: data, target = data.cuda(), target.cuda() batch_loss, _, _ = self.criterion(proj_model, data, target) loss += batch_loss loss = loss / (batch_num+1) * num_datapoints return -loss.cpu().numpy() / temperature def fit(self, use_cuda = True, prior='identity', scale=1.0, **kwargs): # initialize at prior mean = 0 current_sample = np.zeros(self.subspace.cov_factor.size(0)) all_samples = np.zeros((current_sample.size, self.num_samples)) logprobs = np.zeros(self.num_samples) for i in range(self.num_samples): prior_sample = self.prior_sample(prior=prior, scale=scale) current_sample, logprobs[i] = self.slice_method(initial_theta=current_sample, prior=prior_sample, lnpdf=self.log_pdf, **kwargs) # print(logprobs[i]) all_samples[:,i] = current_sample self.all_samples = all_samples return logprobs def sample(self, ind=None, *args, **kwargs): if ind is None: ind = np.random.randint(self.num_samples) rsample = torch.FloatTensor(self.all_samples[:,int(ind)]) #sample = self.subspace(torch.FloatTensor(rsample)).view(-1) if self.use_cuda: device = torch.device('cuda') else: device = torch.device('cpu') rsample = rsample.to(device) self.model = ProjectedModel(model=self.base_model, subspace=self.subspace, proj_params=rsample) return rsample
4,179
35.347826
126
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/inferences.py
""" inferences class w/in the subspace currently only fitting the Gaussian associated is implemented """ import abc import torch import numpy as np from torch.distributions import LowRankMultivariateNormal from .elliptical_slice import elliptical_slice from ..utils import unflatten_like, flatten, train_epoch from .proj_model import ProjectedModel from .vi_model import VIModel, ELBO class Inference(torch.nn.Module, metaclass=abc.ABCMeta): subclasses = {} @classmethod def register_subclass(cls, inference_type): def decorator(subclass): cls.subclasses[inference_type] = subclass return subclass return decorator @classmethod def create(cls, inference_type, **kwargs): if inference_type not in cls.subclasses: raise ValueError('Bad inference type {}'.format(inference_type)) return cls.subclasses[inference_type](**kwargs) @abc.abstractmethod def __init__(self, *args, **kwargs): super(Inference, self).__init__() @abc.abstractmethod def fit(self, mean, variance, cov_factor, *args, **kwargs): pass @abc.abstractmethod def sample(self, *args, **kwargs): pass @Inference.register_subclass('low_rank_gaussian') class LRGaussian(Inference): def __init__(self, base, base_args, base_kwargs, var_clamp=1e-6): super(LRGaussian, self).__init__() self.var_clamp = var_clamp self.dist = None def fit(self, mean, variance, cov_factor): # ensure variance >= var_clamp variance = torch.clamp(variance, self.var_clamp) # form a low rank (+ diagonal Gaussian) distribution when fitting self.dist = LowRankMultivariateNormal(loc=mean, cov_diag=variance, cov_factor=cov_factor.t()) def sample(self, scale=0.5, seed=None): if seed is not None: torch.manual_seed(seed) # x = \mu + L'z unscaled_sample = self.dist.rsample() # x' = \sqrt(scale) * (x - \mu) + \mu scaled_sample = (scale ** 0.5) * (unscaled_sample - self.dist.loc) + self.dist.loc return scaled_sample def log_prob(self, sample): return self.dist.log_prob(sample) @Inference.register_subclass('projected_sgd') class ProjSGD(Inference): def __init__(self, model, loader, criterion, epochs = 10, **kwargs): super(ProjSGD, self).__init__() self.kwargs = kwargs self.optimizer = None self.epochs = epochs self.mean, self.var, self.subspace = None, None, None self.optimizer = None self.proj_params = None self.loader, self.criterion = loader, criterion self.model = model def fit(self, mean, variance, subspace, use_cuda = True, **kwargs): if use_cuda and torch.cuda.is_available(): self.mean = mean.cuda() self.subspace = subspace.cuda() else: self.mean = mean self.subspace = subspace if self.proj_params is None: proj_params = torch.zeros(self.subspace.size(0), 1, dtype = self.subspace.dtype, device = self.subspace.device, requires_grad = True) print(proj_params.device) self.proj_model = ProjectedModel(model=self.model, mean=self.mean.unsqueeze(1), projection=self.subspace, proj_params=proj_params) # define optimizer self.optimizer = torch.optim.SGD([proj_params], **self.kwargs) else: proj_params = self.proj_params.clone() # now train projected parameters loss_vec = [] for _ in range(self.epochs): loss = train_epoch(loader=self.loader, optimizer=self.optimizer, model=self.proj_model, criterion=self.criterion, **kwargs) loss_vec.append( loss ) self.proj_params = proj_params return loss_vec def sample(self, *args, **kwargs): print(self.mean.size(), self.subspace.size(), self.proj_params.size()) map_sample = self.mean + self.subspace.t().matmul(self.proj_params.squeeze(1)) return map_sample.view(1,-1) @Inference.register_subclass('vi') class VI(Inference): def __init__(self, base, base_args, base_kwargs, rank, init_inv_softplus_simga=-6.0, prior_log_sigma=0.0): super(VI, self).__init__() self.vi_model = VIModel( base=base, base_args=base_args, base_kwargs=base_kwargs, rank=rank, init_inv_softplus_simga=init_inv_softplus_simga, prior_log_sigma=prior_log_sigma ) def fit(self, mean, variance, cov_factor, loader, criterion, epochs=100): print('Fitting VI') self.vi_model.set_subspace(mean, cov_factor) elbo = ELBO(criterion, len(loader.dataset)) optimizer = torch.optim.Adam([param for param in self.vi_model.parameters() if param.requires_grad]) for _ in range(epochs): train_res = train_epoch(loader, self.vi_model, elbo, optimizer) print(train_res) def sample(self): return self.vi_model.sample()
5,185
31.21118
145
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/vinf_model.py
import math import torch from ..utils import set_weights class VINFModel(torch.nn.Module): def __init__(self, base, subspace, flow, prior_log_sigma=1.0, *args, **kwargs): super(VINFModel, self).__init__() self.base_model = base(*args, **kwargs) self.flow = flow self.subspace = subspace self.rank = self.subspace.rank self.prior_log_sigma = prior_log_sigma def forward(self, input, t=None, *args, **kwargs): if t is None: t = self.flow.sample() w = self.subspace(t.squeeze()) set_weights(self.base_model, w, self.flow.device) return self.base_model(input, *args, **kwargs) def compute_kl_mc(self, t=None): if t is None: t = self.flow.sample() prior_logprob = - torch.norm(t.squeeze())**2 / (2 * math.exp(self.prior_log_sigma * 2)) return self.flow.log_prob(t) - prior_logprob def compute_entropy_mc(self, t=None): if t is None: t = self.flow.sample() return self.flow.log_prob(t) class ELBO_NF(object): def __init__(self, criterion, num_samples, temperature=1.): self.criterion = criterion self.num_samples = num_samples self.temperature = temperature def __call__(self, model, input, target): # likelihood term t = model.flow.sample() output = model(input, t=t) nll, _ = self.criterion(output, target) # kl term kl = model.compute_kl_mc(t) loss = nll + kl * self.temperature / self.num_samples # -elbo return loss, output, {"nll": nll.item(), "kl": kl.item()} class BenchmarkVINFModel(VINFModel): # same as a VINFModel, except with a fit method # for ease of benchmarking def __init__(self, loader, criterion, optimizer, epochs, base, subspace, flow, prior_log_sigma=3.0, lr=0.1, temperature=1., num_samples=45000, *args, **kwargs): super(BenchmarkVINFModel, self).__init__(base, subspace, flow, prior_log_sigma=prior_log_sigma) self.loader = loader self.criterion = criterion self.optimizer = torch.optim.Adam([param for param in self.parameters()], lr=lr) self.elbo = ELBO_NF(self.criterion, num_samples, temperature) def fit(self, *args, **kwargs): for epoch in range(self.epochs): train_res = train_epoch(self.loader, self, self.elbo, self.optimizer) values = ['%d/%d' % (epoch + 1, self.epochs), train_res['accuracy'], train_res['loss'], train_res['stats']['kl'], train_res['stats']['nll']] print(values)
2,662
33.584416
103
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/swag.py
import torch from ..utils import flatten, set_weights from .subspaces import Subspace class SWAG(torch.nn.Module): def __init__(self, base, subspace_type, subspace_kwargs=None, var_clamp=1e-6, *args, **kwargs): super(SWAG, self).__init__() self.base_model = base(*args, **kwargs) self.num_parameters = sum(param.numel() for param in self.base_model.parameters()) self.register_buffer('mean', torch.zeros(self.num_parameters)) self.register_buffer('sq_mean', torch.zeros(self.num_parameters)) self.register_buffer('n_models', torch.zeros(1, dtype=torch.long)) # Initialize subspace if subspace_kwargs is None: subspace_kwargs = dict() self.subspace = Subspace.create(subspace_type, num_parameters=self.num_parameters, **subspace_kwargs) self.var_clamp = var_clamp self.cov_factor = None self.model_device = 'cpu' # dont put subspace on cuda? def cuda(self, device=None): self.model_device = 'cuda' self.base_model.cuda(device=device) def to(self, *args, **kwargs): self.base_model.to(*args, **kwargs) device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs) self.model_device = device.type self.subspace.to(device=torch.device('cpu'), dtype=dtype, non_blocking=non_blocking) def forward(self, *args, **kwargs): return self.base_model(*args, **kwargs) def collect_model(self, base_model, *args, **kwargs): # need to refit the space after collecting a new model self.cov_factor = None w = flatten([param.detach().cpu() for param in base_model.parameters()]) # first moment self.mean.mul_(self.n_models.item() / (self.n_models.item() + 1.0)) self.mean.add_(w / (self.n_models.item() + 1.0)) # second moment self.sq_mean.mul_(self.n_models.item() / (self.n_models.item() + 1.0)) self.sq_mean.add_(w ** 2 / (self.n_models.item() + 1.0)) dev_vector = w - self.mean self.subspace.collect_vector(dev_vector, *args, **kwargs) self.n_models.add_(1) def _get_mean_and_variance(self): variance = torch.clamp(self.sq_mean - self.mean ** 2, self.var_clamp) return self.mean, variance def fit(self): if self.cov_factor is not None: return self.cov_factor = self.subspace.get_space() def set_swa(self): set_weights(self.base_model, self.mean, self.model_device) def sample(self, scale=0.5, diag_noise=True, cov_factor=True): self.fit() mean, variance = self._get_mean_and_variance() z = torch.zeros_like(variance) if cov_factor: eps_low_rank = torch.randn(self.cov_factor.size()[0]) z = self.cov_factor.t() @ eps_low_rank if diag_noise: z += variance.sqrt() * torch.randn_like(variance) z *= scale ** 0.5 sample = mean + z # apply to parameters set_weights(self.base_model, sample, self.model_device) return sample def get_space(self, export_cov_factor=True): mean, variance = self._get_mean_and_variance() if not export_cov_factor: return mean.clone(), variance.clone() else: self.fit() return mean.clone(), variance.clone(), self.cov_factor.clone()
3,487
34.232323
96
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/subspaces.py
""" subspace classes CovarianceSpace: covariance subspace PCASpace: PCA subspace FreqDirSpace: Frequent Directions Space """ import abc import torch import numpy as np from sklearn.decomposition import TruncatedSVD from sklearn.decomposition._pca import _assess_dimension from sklearn.utils.extmath import randomized_svd class Subspace(torch.nn.Module, metaclass=abc.ABCMeta): subclasses = {} @classmethod def register_subclass(cls, subspace_type): def decorator(subclass): cls.subclasses[subspace_type] = subclass return subclass return decorator @classmethod def create(cls, subspace_type, **kwargs): if subspace_type not in cls.subclasses: raise ValueError('Bad subspaces type {}'.format(subspace_type)) return cls.subclasses[subspace_type](**kwargs) def __init__(self): super(Subspace, self).__init__() @abc.abstractmethod def collect_vector(self, vector): pass @abc.abstractmethod def get_space(self): pass @Subspace.register_subclass('random') class RandomSpace(Subspace): def __init__(self, num_parameters, rank=20, method='dense'): assert method in ['dense', 'fastfood'] super(RandomSpace, self).__init__() self.num_parameters = num_parameters self.rank = rank self.method = method if method == 'dense': self.subspace = torch.randn(rank, num_parameters) if method == 'fastfood': raise NotImplementedError("FastFood transform hasn't been implemented yet") # random subspace is independent of data def collect_vector(self, vector): pass def get_space(self): return self.subspace @Subspace.register_subclass('covariance') class CovarianceSpace(Subspace): def __init__(self, num_parameters, max_rank=20): super(CovarianceSpace, self).__init__() self.num_parameters = num_parameters self.register_buffer('rank', torch.zeros(1, dtype=torch.long)) self.register_buffer('cov_mat_sqrt', torch.empty(0, self.num_parameters, dtype=torch.float32)) self.max_rank = max_rank def collect_vector(self, vector): if self.rank.item() + 1 > self.max_rank: self.cov_mat_sqrt = self.cov_mat_sqrt[1:, :] self.cov_mat_sqrt = torch.cat((self.cov_mat_sqrt, vector.view(1, -1)), dim=0) self.rank = torch.min(self.rank + 1, torch.as_tensor(self.max_rank)).view(-1) def get_space(self): return self.cov_mat_sqrt.clone() / (self.cov_mat_sqrt.size(0) - 1) ** 0.5 def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): rank = state_dict[prefix + 'rank'].item() self.cov_mat_sqrt = self.cov_mat_sqrt.new_empty((rank, self.cov_mat_sqrt.size()[1])) super(CovarianceSpace, self)._load_from_state_dict(state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) @Subspace.register_subclass('pca') class PCASpace(CovarianceSpace): def __init__(self, num_parameters, pca_rank=20, max_rank=20): super(PCASpace, self).__init__(num_parameters, max_rank=max_rank) # better phrasing for this condition? assert(pca_rank == 'mle' or isinstance(pca_rank, int)) if pca_rank != 'mle': assert 1 <= pca_rank <= max_rank self.pca_rank = pca_rank def get_space(self): cov_mat_sqrt_np = self.cov_mat_sqrt.clone().numpy() # perform PCA on DD' cov_mat_sqrt_np /= (max(1, self.rank.item() - 1))**0.5 if self.pca_rank == 'mle': pca_rank = self.rank.item() else: pca_rank = self.pca_rank pca_rank = max(1, min(pca_rank, self.rank.item())) # pca_decomp = TruncatedSVD(n_components=pca_rank) # pca_decomp.fit(cov_mat_sqrt_np) _, s, Vt = randomized_svd(cov_mat_sqrt_np, n_components=pca_rank, n_iter=5) # perform post-selection fitting if self.pca_rank == 'mle': eigs = s ** 2.0 ll = np.zeros(len(eigs)) correction = np.zeros(len(eigs)) # compute minka's PCA marginal log likelihood and the correction term for rank in range(len(eigs)): # secondary correction term based on the rank of the matrix + degrees of freedom m = cov_mat_sqrt_np.shape[1] * rank - rank * (rank + 1) / 2. correction[rank] = 0.5 * m * np.log(cov_mat_sqrt_np.shape[0]) ll[rank] = _assess_dimension(spectrum=eigs, rank=rank, # n_features=min(cov_mat_sqrt_np.shape), n_samples=max(cov_mat_sqrt_np.shape)) self.ll = ll self.corrected_ll = ll - correction self.pca_rank = np.nanargmax(self.corrected_ll) print('PCA Rank is: ', self.pca_rank) return torch.FloatTensor(s[:self.pca_rank, None] * Vt[:self.pca_rank, :]) else: return torch.FloatTensor(s[:, None] * Vt) @Subspace.register_subclass('freq_dir') class FreqDirSpace(CovarianceSpace): """ Frequent directions is an online PCA technique, cf. "Frequent directions: Simple and deterministic matrix sketching." """ def __init__(self, num_parameters, max_rank=20): super(FreqDirSpace, self).__init__(num_parameters, max_rank=max_rank) self.register_buffer('num_models', torch.zeros(1, dtype=torch.long)) self.delta = 0.0 self.normalized = False def collect_vector(self, vector): if self.rank >= 2 * self.max_rank: sketch = self.cov_mat_sqrt.numpy() [_, s, Vt] = np.linalg.svd(sketch, full_matrices=False) if s.size >= self.max_rank: current_delta = s[self.max_rank - 1] ** 2 self.delta += current_delta s = np.sqrt(s[:self.max_rank - 1] ** 2 - current_delta) self.cov_mat_sqrt = torch.from_numpy(s[:, None] * Vt[:s.size, :]) self.cov_mat_sqrt = torch.cat((self.cov_mat_sqrt, vector.view(1, -1)), dim=0) self.rank = torch.as_tensor(self.cov_mat_sqrt.size(0)) self.num_models.add_(1) self.normalized = False def get_space(self): if not self.normalized: sketch = self.cov_mat_sqrt.numpy() [_, s, Vt] = np.linalg.svd(sketch, full_matrices=False) self.cov_mat_sqrt = torch.from_numpy(s[:, None] * Vt) self.normalized = True curr_rank = min(self.rank.item(), self.max_rank) return self.cov_mat_sqrt[:curr_rank].clone() / max(1, self.num_models.item() - 1) ** 0.5
7,016
35.357513
121
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/realnvp.py
import math import numpy as np import torch from torch import nn from torch import distributions class RealNVP(nn.Module): def __init__(self, nets, nett, masks, prior, device=None): super().__init__() self.prior = prior self.mask = nn.Parameter(masks, requires_grad=False) self.t = torch.nn.ModuleList([nett() for _ in range(len(masks))]) self.s = torch.nn.ModuleList([nets() for _ in range(len(masks))]) self.to(device) self.device = device def g(self, z): x = z for i in reversed(range(len(self.mask))): mx = self.mask[i] * x tmx = self.t[i](mx) smx = self.s[i](mx) x = mx + (1 - self.mask[i]) * ((x - tmx) * torch.exp(-smx)) return x def f(self, x): z = x log_det_J = 0 for i in range(len(self.mask)): mz = self.mask[i] * z smz = self.s[i](mz) tmz = self.t[i](mz) z = mz + (1 - self.mask[i]) * (z * torch.exp(smz) + tmz) if x.dim() == 2: log_det_J += (smz * (1-self.mask[i])).sum(1) else: log_det_J += (smz * (1-self.mask[i])).sum(1, 2, 3) return z, log_det_J def log_prob(self, x): z, log_det_J = self.f(x) return self.prior.log_prob(z) + log_det_J def sample(self, bs=1): z = self.prior.sample(torch.Size([bs])) x = self.g(z) return x def construct_flow(D, coupling_layers_num=2, inner_dim=128, inner_layers=2, prior=None, device=None): def inner_seq(n, inner_dim): res = [] for _ in range(n): res.append(nn.Linear(inner_dim, inner_dim)) res.append(nn.ReLU()) return res class Nets(nn.Module): """ net for parametrizing scaling function in coupling layer """ def __init__(self, D, inner_dim, inner_layers): super().__init__() self.seq_part = nn.Sequential(nn.Linear(D, inner_dim), nn.ReLU(), *inner_seq(inner_layers, inner_dim), nn.Linear(inner_dim, D), nn.Tanh()) self.scale = nn.Parameter(torch.ones(D)) def forward(self, x): x = self.seq_part.forward(x) x = self.scale * x return x # a function that take no arguments and return a pytorch model, dim(X) -> dim(X) nets = lambda: Nets(D, inner_dim, inner_layers) nett = lambda: nn.Sequential(nn.Linear(D, inner_dim), nn.ReLU(), *inner_seq(inner_layers, inner_dim), nn.Linear(inner_dim, D)) if prior is None: prior = distributions.MultivariateNormal(torch.zeros(D).to(device), torch.eye(D).to(device)) d = D // 2 masks = torch.zeros(coupling_layers_num, D) for i in range(masks.size(0)): if i % 2: masks[i, :d] = 1. else: masks[i, d:] = 1. masks.to(device) return RealNVP(nets, nett, masks, prior, device=device)
3,244
32.112245
101
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/proj_model.py
import torch from ..utils import unflatten_like class SubspaceModel(torch.nn.Module): def __init__(self, mean, cov_factor): super(SubspaceModel, self).__init__() self.rank = cov_factor.size(0) self.register_buffer('mean', mean) self.register_buffer('cov_factor', cov_factor) def forward(self, t): return self.mean + self.cov_factor.t() @ t class ProjectedModel(torch.nn.Module): def __init__(self, proj_params, model, projection=None, mean=None, subspace=None): super(ProjectedModel, self).__init__() self.model = model if subspace is None: self.subspace = SubspaceModel(mean, projection) else: self.subspace = subspace if mean is None and subspace is None: raise NotImplementedError('Must enter either subspace or mean') self.proj_params = proj_params def update_params(self, vec, model): vec_list = unflatten_like(likeTensorList=list(model.parameters()), vector=vec.view(1,-1)) for param, v in zip(model.parameters(), vec_list): param.detach_() param.mul_(0.0).add_(v) def forward(self, *args, **kwargs): y = self.subspace(self.proj_params) self.update_params(y, self.model) return self.model(*args, **kwargs)
1,329
32.25
97
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/pyro.py
import numpy as np import torch import pyro import pyro.distributions as dist from pyro.infer.mcmc import NUTS, MCMC from pyro.nn import AutoRegressiveNN from ..utils import extract_parameters from ..utils import set_weights_old as set_weights class PyroModel(torch.nn.Module): def __init__(self, base, subspace, prior_log_sigma, likelihood_given_outputs, batch_size = 100, *args, **kwargs): super(PyroModel, self).__init__() self.base_model = base(*args, **kwargs) self.base_params = extract_parameters(self.base_model) #self.rank = cov_factor.size()[0] self.prior_log_sigma = prior_log_sigma self.likelihood = likelihood_given_outputs self.batch_size = batch_size self.subspace = subspace self.rank = self.subspace.cov_factor.size(0) def model(self, x, y): self.t = pyro.sample("t", dist.Normal(torch.zeros(self.rank), torch.ones(self.rank) * np.exp(self.prior_log_sigma)).to_event(1)) self.t = self.t.to(x.device) bs = self.batch_size num_batches = x.shape[0] // bs if x.shape[0] % bs: num_batches += 1 for i in pyro.plate("batches", num_batches): x_ = x[i * bs: (i+1)*bs] y_ = y[i * bs: (i+1)*bs] with pyro.plate("data" + str(i), x_.shape[0]): w = self.subspace(self.t) set_weights(self.base_params, w, self.t.device) z = self.base_model(x_) pyro.sample("y" + str(i), self.likelihood(z).to_event(1), obs=y_) def model_subsample(self, x, y): subsample_size = self.batch_size self.t = pyro.sample("t", dist.Normal(torch.zeros(self.rank), torch.ones(self.rank) * np.exp(self.prior_log_sigma)).to_event(1)) self.t = self.t.to(x.device) with pyro.plate("data", x.shape[0], subsample_size=subsample_size) as ind: #w = self.mean.to(self.t.device) + self.cov_factor.to(self.t.device).t() @ self.t w = self.subspace(self.t) set_weights(self.base_params, w, self.t.device) z = self.base_model(x[ind]) pyro.sample("y", self.likelihood(z).to_event(1), obs=y[ind]) def forward(self, *args, **kwargs): w = self.subspace(self.t) set_weights(self.base_params, w, self.t.device) return self.base_model(*args, **kwargs) class BenchmarkPyroModel(PyroModel): def __init__(self, base, subspace, prior_log_sigma, likelihood_given_outputs, batch_size = 100, kernel=NUTS, num_samples=30, kernel_kwargs={}, *args, **kwargs): super(BenchmarkPyroModel, self).__init__(base, subspace, prior_log_sigma, likelihood_given_outputs, batch_size=batch_size, *args, **kwargs) self.kernel = kernel(self.model, **kernel_kwargs) self.num_samples = num_samples #self.loader = loader self.all_samples = None self.mcmc_run = None def fit(self, inputs, targets, *args, **kwargs): self.mcmc_run = MCMC(self.kernel, num_samples=self.num_samples, warmup_steps=100).run(inputs, targets) self.all_samples = torch.cat(list(self.mcmc_run.marginal(sites="t").support(flatten=True).values()), dim=-1) def sample(self, ind=None, scale=1.0): if ind is None: ind = np.random.randint(self.num_samples) self.eval() self.t.set_(self.all_samples[int(ind), :]) class GaussianGuide(torch.nn.Module): def __init__(self, rank, init_inv_softplus_sigma=-3.0, eps=1e-6, with_mu=True): super(GaussianGuide, self).__init__() self.rank = rank self.eps = eps self.with_mu = with_mu self.init_inv_softplus_sigma = init_inv_softplus_sigma def model(self, *args, **kargs): if self.with_mu: self.mu = pyro.param("mu", torch.zeros(self.rank)) else: self.mu = torch.zeros(self.rank) self.inv_softplus_sigma = pyro.param("inv_softplus_sigma", torch.ones(self.rank) * (self.init_inv_softplus_sigma)) sigma = torch.nn.functional.softplus(self.inv_softplus_sigma) + self.eps self.t = pyro.sample("t", dist.Normal(self.mu, sigma).to_event(1)) return self.t class IAFGuide(torch.nn.Module): def __init__(self, rank, n_hid=[100]): super(IAFGuide, self).__init__() self.rank = rank self.n_hid = n_hid @property def sigma(self): sigma = torch.nn.functional.softplus(self.inv_softplus_sigma) return sigma def model(self, *args, **kargs): self.inv_softplus_sigma = pyro.param("inv_softplus_sigma", torch.ones(self.rank)) sigma = self.sigma#torch.nn.functional.softplus(self.inv_softplus_sigma) #base_dist = dist.Normal(torch.zeros(self.rank), torch.ones(self.rank)) # Pavel: introducing `sigma` in the IAF distribution makes training more # stable in tems of the scale of the distribution we are trying to learn base_dist = dist.Normal(torch.zeros(self.rank), sigma) ann = AutoRegressiveNN(self.rank, self.n_hid, skip_connections=True) iaf = dist.InverseAutoregressiveFlow(ann) iaf_module = pyro.module("my_iaf", iaf) iaf_dist = dist.TransformedDistribution(base_dist, [iaf]) self.t = pyro.sample("t", iaf_dist.to_event(1)) return self.t class TemperedCategorical(dist.Categorical): def __init__(self, temperature=1., *args, **kwargs): super(TemperedCategorical, self).__init__(*args, **kwargs) self.temperature = temperature def log_prob(self, value): ans = super(TemperedCategorical, self).log_prob(value) return ans / self.temperature def expand(self, batch_shape): # Blindly copied from pyro batch_shape = torch.Size(batch_shape) validate_args = self.__dict__.get('validate_args') if 'probs' in self.__dict__: probs = self.probs.expand(batch_shape + self.probs.shape[-1:]) return TemperedCategorical(temperature=self.temperature, probs=probs, validate_args=validate_args) else: logits = self.logits.expand(batch_shape + self.logits.shape[-1:]) return TemperedCategorical(temperature=self.temperature, logits=logits, validate_args=validate_args)
6,618
37.707602
116
py
lgv-geometric-transferability
lgv-geometric-transferability-main/utils/subspace_inference/posteriors/vi_model.py
import math import torch from ..utils import extract_parameters, train_epoch from ..utils import set_weights_old as set_weights class VIModel(torch.nn.Module): def __init__(self, base, subspace, init_inv_softplus_sigma=-3.0, prior_log_sigma=3.0, eps=1e-6, with_mu=True, *args, **kwargs): super(VIModel, self).__init__() self.base_model = base(*args, **kwargs) self.base_params = extract_parameters(self.base_model) self.subspace = subspace self.rank = self.subspace.rank self.prior_log_sigma = prior_log_sigma self.eps = eps self.with_mu = with_mu if with_mu: self.mu = torch.nn.Parameter(torch.zeros(self.rank)) self.inv_softplus_sigma = torch.nn.Parameter(torch.empty(self.rank).fill_(init_inv_softplus_sigma)) def forward(self, *args, **kwargs): device = self.inv_softplus_sigma.device sigma = torch.nn.functional.softplus(self.inv_softplus_sigma) + self.eps if self.with_mu: z = self.mu + torch.randn(self.rank, device=device) * sigma else: z = torch.randn(self.rank, device=device) * sigma w = self.subspace(z) set_weights(self.base_params, w, device) #set_weights(self.base_model, w, device) return self.base_model(*args, **kwargs) def sample(self, scale=1.): device = self.inv_softplus_sigma.device sigma = torch.nn.functional.softplus(self.inv_softplus_sigma.detach()) + self.eps if self.with_mu: z = self.mu + torch.randn(self.rank, device=device) * sigma * scale else: z = torch.randn(self.rank, device=device) * sigma * scale w = self.subspace(z) return w #def sample_z(self): # sigma = torch.nn.functional.softplus(self.inv_softplus_sigma.detach().cpu()) + self.eps # z = torch.randn(self.rank) * sigma # if self.with_mu: # z += self.mu.detach().cpu() # return z def compute_kl(self): sigma = torch.nn.functional.softplus(self.inv_softplus_sigma) + self.eps kl = torch.sum(self.prior_log_sigma - torch.log(sigma) + 0.5 * (sigma ** 2) / (math.exp(self.prior_log_sigma * 2))) if self.with_mu: kl += 0.5 * torch.sum(self.mu ** 2) / math.exp(self.prior_log_sigma * 2) return kl def compute_entropy(self): sigma = torch.nn.functional.softplus(self.inv_softplus_sigma) + self.eps return torch.sum(torch.log(sigma)) class ELBO(object): def __init__(self, criterion, num_samples, temperature=1.): self.criterion = criterion self.num_samples = num_samples self.temperature = temperature #print("In ELBO, temperature:", temperature) #print("In ELBO, num_samples:", num_samples) def __call__(self, model, input, target): nll, output, _ = self.criterion(model, input, target) kl = model.compute_kl() / self.num_samples kl *= self.temperature loss = nll + kl #loss = nll return loss, output, {'nll': nll.item(), 'kl': kl.item()} class BenchmarkVIModel(VIModel): # same as a VI model, except with a fit method # for ease of benchmarking def __init__(self, loader, criterion, epochs, base, subspace, init_inv_softplus_sigma=-3.0, prior_log_sigma=3.0, eps=1e-6, with_mu=True, lr=0.01, num_samples=45000, temperature=1.0, use_cuda=True, *args, **kwargs): super(BenchmarkVIModel, self).__init__(base, subspace, init_inv_softplus_sigma=-3.0, prior_log_sigma=prior_log_sigma, eps=eps, with_mu=with_mu, *args, **kwargs) self.use_cuda = use_cuda self.loader = loader self.criterion = criterion #print("Num Samples ELBO:", num_samples) self.optimizer = torch.optim.Adam([param for param in self.parameters()], lr=lr) self.elbo = ELBO(self.criterion, num_samples, temperature=temperature) self.epochs = epochs def fit(self, *args, **kwargs): for epoch in range(self.epochs): train_res = train_epoch(self.loader, self, self.elbo, self.optimizer, regression=True, cuda = self.use_cuda) values = ['%d/%d' % (epoch + 1, self.epochs), train_res['accuracy'], train_res['loss'], train_res['stats']['kl'], train_res['stats']['nll']] #print(values) #with torch.no_grad(): # print("sigma:", torch.nn.functional.softplus(self.inv_softplus_sigma.cpu())) # if self.with_mu: # print("mu:", self.mu.cpu())
4,676
38.635593
139
py
ActiveVisionManipulation
ActiveVisionManipulation-master/HER/envs/fakercnn_pusher.py
from HER.envs import bb_pusher import numpy as np from HER.rcnn import renderer import random from keras import backend as K from HER.rcnn import load_rcnn import tensorflow as tf from HER.envs.pusher import _tuple from ipdb import set_trace as st class BaxterEnv(bb_pusher.BaxterEnv): def __init__(self, *args, aux_rwd = False, test = False, fake_scheme = 'default', **kwargs): self.aux_rwd = aux_rwd self.fake_scheme = fake_scheme self.test = test super().__init__(*args, **kwargs) if test: self.initialize_rcnn() def reset_model(self): #we can't use last_mask or last_box BECAUSE it's misleading once the camera starts moving self.last_box_vis = None # if hasattr(self, 'last_mask'): # del self.last_mask #in case we fail at detection # if self.test: # self.last_box = None return super().reset_model() def calc_reward(self, state, return_success = False): rwd, succ = super().calc_reward(state, True) if self.aux_rwd and not self.test: rwd += state.aux if return_success: return rwd, succ else: return rwd def _get_obs(self): obs = super()._get_obs() if self.aux_rwd and not self.test: obs.aux = self.last_success return obs #we need to dig up some old code here... def initialize_rcnn(self): #need to mess around with a different grpah/session because the rcnn #and the HER code runs on separate graphs/sessions self.rcnn_graph = tf.Graph() self.rcnn_session = tf.Session(graph = self.rcnn_graph) K.set_session(self.rcnn_session) with self.rcnn_session.as_default(): with self.rcnn_graph.as_default(): self.rcnn = load_rcnn.load_rcnn() def _get_obj_bbox_from_rcnn(self): self.renderer.render_rgb() #throwaway rgb = self.renderer.render_rgb() ##### K.set_session(self.rcnn_session) with self.rcnn_session.as_default(): with self.rcnn_graph.as_default(): rcnn_out = self.rcnn.detect([rgb], verbose = 0)[0] rois = rcnn_out['rois'] #from ipdb import set_trace as st #st() DEBUG_RCNN = True if len(rois) == 0 or rcnn_out['scores'][0] < 0.98: if DEBUG_RCNN: print("rcnn detected nothing") box = None else: #rois are sorted by score, so we take the most confident detection y1, x1, y2, x2 = rois[0] box = (x1, x2, y1, y2) if DEBUG_RCNN: print("rcnn detection at", box) # self.last_box = box self.last_box_vis = box # return self.last_box return box def _get_obj_bbox(self): #if self.test and self.num_step > 1: if self.test: return self._get_obj_bbox_from_rcnn() self.renderer.render_rgb() #throwaway mask = self.renderer.render_box() modal_mask = self.renderer.render_box(override_amodal = False) unocc = np.sum(modal_mask) / (np.sum(mask) + 1E-9) #complete unocc = 100% success. 75% occluded means 50% detection chance if self.num_step == 1: #let's say we alwyas see it on ep 0 chance_success = 1.0 else: #actually this seems a bit too strict if self.fake_scheme == 'default': chance_success = np.sqrt(unocc) elif self.fake_scheme == 'piecewise': if unocc >= 0.2: chance_success = unocc*0.1+0.9 #so 0.92 at 0.2 else: chance_success = unocc*0.92*5.0 else: raise Exception('unknown fake_scheme') if not hasattr(self, 'bar'): self.bar = lambda: None self.bar.count = 0 self.bar.succ = 0 self.bar.unocc = 0.0 self.bar.count += 1 self.bar.unocc += unocc self.last_success = 0 if random.random() < chance_success: #if we succeeded #self.last_mask = mask self.last_success = 0.25 self.bar.succ += 1 else: #if we failed return None #elif not hasattr(self, 'last_mask'): #if we failed but don't have last mask #self.last_mask = np.zeros((self.img_params.full_imgH, self.img_params.full_imgW)) if self.debug: print('detect rate', self.bar.succ/self.bar.count) print('unocc', self.bar.unocc/self.bar.count) #return self._get_bbox_from_mask(self.last_mask, unocc) return self._get_bbox_from_mask(mask, unocc)
4,861
32.531034
97
py
ActiveVisionManipulation
ActiveVisionManipulation-master/HER/rcnn/Mask_RCNN/parallel_model.py
""" Mask R-CNN Multi-GPU Support for Keras. Copyright (c) 2017 Matterport, Inc. Licensed under the MIT License (see LICENSE for details) Written by Waleed Abdulla Ideas and a small code snippets from these sources: https://github.com/fchollet/keras/issues/2436 https://medium.com/@kuza55/transparent-multi-gpu-training-on-tensorflow-with-keras-8b0016fd9012 https://github.com/avolkov1/keras_experiments/blob/master/keras_exp/multigpu/ https://github.com/fchollet/keras/blob/master/keras/utils/training_utils.py """ import tensorflow as tf import keras.backend as K import keras.layers as KL import keras.models as KM class ParallelModel(KM.Model): """Subclasses the standard Keras Model and adds multi-GPU support. It works by creating a copy of the model on each GPU. Then it slices the inputs and sends a slice to each copy of the model, and then merges the outputs together and applies the loss on the combined outputs. """ def __init__(self, keras_model, gpu_count): """Class constructor. keras_model: The Keras model to parallelize gpu_count: Number of GPUs. Must be > 1 """ self.inner_model = keras_model self.gpu_count = gpu_count merged_outputs = self.make_parallel() super(ParallelModel, self).__init__(inputs=self.inner_model.inputs, outputs=merged_outputs) def __getattribute__(self, attrname): """Redirect loading and saving methods to the inner model. That's where the weights are stored.""" if 'load' in attrname or 'save' in attrname: return getattr(self.inner_model, attrname) return super(ParallelModel, self).__getattribute__(attrname) def summary(self, *args, **kwargs): """Override summary() to display summaries of both, the wrapper and inner models.""" super(ParallelModel, self).summary(*args, **kwargs) self.inner_model.summary(*args, **kwargs) def make_parallel(self): """Creates a new wrapper model that consists of multiple replicas of the original model placed on different GPUs. """ # Slice inputs. Slice inputs on the CPU to avoid sending a copy # of the full inputs to all GPUs. Saves on bandwidth and memory. input_slices = {name: tf.split(x, self.gpu_count) for name, x in zip(self.inner_model.input_names, self.inner_model.inputs)} output_names = self.inner_model.output_names outputs_all = [] for i in range(len(self.inner_model.outputs)): outputs_all.append([]) # Run the model call() on each GPU to place the ops there for i in range(self.gpu_count): with tf.device('/gpu:%d' % i): with tf.name_scope('tower_%d' % i): # Run a slice of inputs through this replica zipped_inputs = zip(self.inner_model.input_names, self.inner_model.inputs) inputs = [ KL.Lambda(lambda s: input_slices[name][i], output_shape=lambda s: (None,) + s[1:])(tensor) for name, tensor in zipped_inputs] # Create the model replica and get the outputs outputs = self.inner_model(inputs) if not isinstance(outputs, list): outputs = [outputs] # Save the outputs for merging back together later for l, o in enumerate(outputs): outputs_all[l].append(o) # Merge outputs on CPU with tf.device('/cpu:0'): merged = [] for outputs, name in zip(outputs_all, output_names): # If outputs are numbers without dimensions, add a batch dim. def add_dim(tensor): """Add a dimension to tensors that don't have any.""" if K.int_shape(tensor) == (): return KL.Lambda(lambda t: K.reshape(t, [1, 1]))(tensor) return tensor outputs = list(map(add_dim, outputs)) # Concatenate merged.append(KL.Concatenate(axis=0, name=name)(outputs)) return merged if __name__ == "__main__": # Testing code below. It creates a simple model to train on MNIST and # tries to run it on 2 GPUs. It saves the graph so it can be viewed # in TensorBoard. Run it as: # # python3 parallel_model.py import os import numpy as np import keras.optimizers from keras.datasets import mnist from keras.preprocessing.image import ImageDataGenerator GPU_COUNT = 2 # Root directory of the project ROOT_DIR = os.getcwd() # Directory to save logs and trained model MODEL_DIR = os.path.join(ROOT_DIR, "logs/parallel") def build_model(x_train, num_classes): # Reset default graph. Keras leaves old ops in the graph, # which are ignored for execution but clutter graph # visualization in TensorBoard. tf.reset_default_graph() inputs = KL.Input(shape=x_train.shape[1:], name="input_image") x = KL.Conv2D(32, (3, 3), activation='relu', padding="same", name="conv1")(inputs) x = KL.Conv2D(64, (3, 3), activation='relu', padding="same", name="conv2")(x) x = KL.MaxPooling2D(pool_size=(2, 2), name="pool1")(x) x = KL.Flatten(name="flat1")(x) x = KL.Dense(128, activation='relu', name="dense1")(x) x = KL.Dense(num_classes, activation='softmax', name="dense2")(x) return KM.Model(inputs, x, "digit_classifier_model") # Load MNIST Data (x_train, y_train), (x_test, y_test) = mnist.load_data() x_train = np.expand_dims(x_train, -1).astype('float32') / 255 x_test = np.expand_dims(x_test, -1).astype('float32') / 255 print('x_train shape:', x_train.shape) print('x_test shape:', x_test.shape) # Build data generator and model datagen = ImageDataGenerator() model = build_model(x_train, 10) # Add multi-GPU support. model = ParallelModel(model, GPU_COUNT) optimizer = keras.optimizers.SGD(lr=0.01, momentum=0.9, clipnorm=5.0) model.compile(loss='sparse_categorical_crossentropy', optimizer=optimizer, metrics=['accuracy']) model.summary() # Train model.fit_generator( datagen.flow(x_train, y_train, batch_size=64), steps_per_epoch=50, epochs=10, verbose=1, validation_data=(x_test, y_test), callbacks=[keras.callbacks.TensorBoard(log_dir=MODEL_DIR, write_graph=True)] )
6,863
38.448276
95
py
ActiveVisionManipulation
ActiveVisionManipulation-master/HER/rcnn/Mask_RCNN/model.py
""" Mask R-CNN The main Mask R-CNN model implemenetation. Copyright (c) 2017 Matterport, Inc. Licensed under the MIT License (see LICENSE for details) Written by Waleed Abdulla """ import os import sys import glob import random import math import datetime import itertools import json import re import logging from collections import OrderedDict import numpy as np import scipy.misc import tensorflow as tf import keras import keras.backend as K import keras.layers as KL import keras.initializers as KI import keras.engine as KE import keras.models as KM import utils # Requires TensorFlow 1.3+ and Keras 2.0.8+. from distutils.version import LooseVersion assert LooseVersion(tf.__version__) >= LooseVersion("1.3") assert LooseVersion(keras.__version__) >= LooseVersion('2.0.8') ############################################################ # Utility Functions ############################################################ def log(text, array=None): """Prints a text message. And, optionally, if a Numpy array is provided it prints it's shape, min, and max values. """ if array is not None: text = text.ljust(25) text += ("shape: {:20} min: {:10.5f} max: {:10.5f}".format( str(array.shape), array.min() if array.size else "", array.max() if array.size else "")) print(text) class BatchNorm(KL.BatchNormalization): """Batch Normalization class. Subclasses the Keras BN class and hardcodes training=False so the BN layer doesn't update during training. Batch normalization has a negative effect on training if batches are small so we disable it here. """ def call(self, inputs, training=None): return super(self.__class__, self).call(inputs, training=False) ############################################################ # Resnet Graph ############################################################ # Code adopted from: # https://github.com/fchollet/deep-learning-models/blob/master/resnet50.py def identity_block(input_tensor, kernel_size, filters, stage, block, use_bias=True): """The identity_block is the block that has no conv layer at shortcut # Arguments input_tensor: input tensor kernel_size: defualt 3, the kernel size of middle conv layer at main path filters: list of integers, the nb_filters of 3 conv layer at main path stage: integer, current stage label, used for generating layer names block: 'a','b'..., current block label, used for generating layer names """ nb_filter1, nb_filter2, nb_filter3 = filters conv_name_base = 'res' + str(stage) + block + '_branch' bn_name_base = 'bn' + str(stage) + block + '_branch' x = KL.Conv2D(nb_filter1, (1, 1), name=conv_name_base + '2a', use_bias=use_bias)(input_tensor) x = BatchNorm(axis=3, name=bn_name_base + '2a')(x) x = KL.Activation('relu')(x) x = KL.Conv2D(nb_filter2, (kernel_size, kernel_size), padding='same', name=conv_name_base + '2b', use_bias=use_bias)(x) x = BatchNorm(axis=3, name=bn_name_base + '2b')(x) x = KL.Activation('relu')(x) x = KL.Conv2D(nb_filter3, (1, 1), name=conv_name_base + '2c', use_bias=use_bias)(x) x = BatchNorm(axis=3, name=bn_name_base + '2c')(x) x = KL.Add()([x, input_tensor]) x = KL.Activation('relu', name='res' + str(stage) + block + '_out')(x) return x def conv_block(input_tensor, kernel_size, filters, stage, block, strides=(2, 2), use_bias=True): """conv_block is the block that has a conv layer at shortcut # Arguments input_tensor: input tensor kernel_size: defualt 3, the kernel size of middle conv layer at main path filters: list of integers, the nb_filters of 3 conv layer at main path stage: integer, current stage label, used for generating layer names block: 'a','b'..., current block label, used for generating layer names Note that from stage 3, the first conv layer at main path is with subsample=(2,2) And the shortcut should have subsample=(2,2) as well """ nb_filter1, nb_filter2, nb_filter3 = filters conv_name_base = 'res' + str(stage) + block + '_branch' bn_name_base = 'bn' + str(stage) + block + '_branch' x = KL.Conv2D(nb_filter1, (1, 1), strides=strides, name=conv_name_base + '2a', use_bias=use_bias)(input_tensor) x = BatchNorm(axis=3, name=bn_name_base + '2a')(x) x = KL.Activation('relu')(x) x = KL.Conv2D(nb_filter2, (kernel_size, kernel_size), padding='same', name=conv_name_base + '2b', use_bias=use_bias)(x) x = BatchNorm(axis=3, name=bn_name_base + '2b')(x) x = KL.Activation('relu')(x) x = KL.Conv2D(nb_filter3, (1, 1), name=conv_name_base + '2c', use_bias=use_bias)(x) x = BatchNorm(axis=3, name=bn_name_base + '2c')(x) shortcut = KL.Conv2D(nb_filter3, (1, 1), strides=strides, name=conv_name_base + '1', use_bias=use_bias)(input_tensor) shortcut = BatchNorm(axis=3, name=bn_name_base + '1')(shortcut) x = KL.Add()([x, shortcut]) x = KL.Activation('relu', name='res' + str(stage) + block + '_out')(x) return x def resnet_graph(input_image, architecture, stage5=False): assert architecture in ["resnet50", "resnet101"] # Stage 1 x = KL.ZeroPadding2D((3, 3))(input_image) x = KL.Conv2D(64, (7, 7), strides=(2, 2), name='conv1', use_bias=True)(x) x = BatchNorm(axis=3, name='bn_conv1')(x) x = KL.Activation('relu')(x) C1 = x = KL.MaxPooling2D((3, 3), strides=(2, 2), padding="same")(x) # Stage 2 x = conv_block(x, 3, [64, 64, 256], stage=2, block='a', strides=(1, 1)) x = identity_block(x, 3, [64, 64, 256], stage=2, block='b') C2 = x = identity_block(x, 3, [64, 64, 256], stage=2, block='c') # Stage 3 x = conv_block(x, 3, [128, 128, 512], stage=3, block='a') x = identity_block(x, 3, [128, 128, 512], stage=3, block='b') x = identity_block(x, 3, [128, 128, 512], stage=3, block='c') C3 = x = identity_block(x, 3, [128, 128, 512], stage=3, block='d') # Stage 4 x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a') block_count = {"resnet50": 5, "resnet101": 22}[architecture] for i in range(block_count): x = identity_block(x, 3, [256, 256, 1024], stage=4, block=chr(98 + i)) C4 = x # Stage 5 if stage5: x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a') x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b') C5 = x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c') else: C5 = None return [C1, C2, C3, C4, C5] ############################################################ # Proposal Layer ############################################################ def apply_box_deltas_graph(boxes, deltas): """Applies the given deltas to the given boxes. boxes: [N, 4] where each row is y1, x1, y2, x2 deltas: [N, 4] where each row is [dy, dx, log(dh), log(dw)] """ # Convert to y, x, h, w height = boxes[:, 2] - boxes[:, 0] width = boxes[:, 3] - boxes[:, 1] center_y = boxes[:, 0] + 0.5 * height center_x = boxes[:, 1] + 0.5 * width # Apply deltas center_y += deltas[:, 0] * height center_x += deltas[:, 1] * width height *= tf.exp(deltas[:, 2]) width *= tf.exp(deltas[:, 3]) # Convert back to y1, x1, y2, x2 y1 = center_y - 0.5 * height x1 = center_x - 0.5 * width y2 = y1 + height x2 = x1 + width result = tf.stack([y1, x1, y2, x2], axis=1, name="apply_box_deltas_out") return result def clip_boxes_graph(boxes, window): """ boxes: [N, 4] each row is y1, x1, y2, x2 window: [4] in the form y1, x1, y2, x2 """ # Split corners wy1, wx1, wy2, wx2 = tf.split(window, 4) y1, x1, y2, x2 = tf.split(boxes, 4, axis=1) # Clip y1 = tf.maximum(tf.minimum(y1, wy2), wy1) x1 = tf.maximum(tf.minimum(x1, wx2), wx1) y2 = tf.maximum(tf.minimum(y2, wy2), wy1) x2 = tf.maximum(tf.minimum(x2, wx2), wx1) clipped = tf.concat([y1, x1, y2, x2], axis=1, name="clipped_boxes") clipped.set_shape((clipped.shape[0], 4)) return clipped class ProposalLayer(KE.Layer): """Receives anchor scores and selects a subset to pass as proposals to the second stage. Filtering is done based on anchor scores and non-max suppression to remove overlaps. It also applies bounding box refinement deltas to anchors. Inputs: rpn_probs: [batch, anchors, (bg prob, fg prob)] rpn_bbox: [batch, anchors, (dy, dx, log(dh), log(dw))] Returns: Proposals in normalized coordinates [batch, rois, (y1, x1, y2, x2)] """ def __init__(self, proposal_count, nms_threshold, anchors, config=None, **kwargs): """ anchors: [N, (y1, x1, y2, x2)] anchors defined in image coordinates """ super(ProposalLayer, self).__init__(**kwargs) self.config = config self.proposal_count = proposal_count self.nms_threshold = nms_threshold self.anchors = anchors.astype(np.float32) def call(self, inputs): # Box Scores. Use the foreground class confidence. [Batch, num_rois, 1] scores = inputs[0][:, :, 1] # Box deltas [batch, num_rois, 4] deltas = inputs[1] deltas = deltas * np.reshape(self.config.RPN_BBOX_STD_DEV, [1, 1, 4]) # Base anchors anchors = self.anchors # Improve performance by trimming to top anchors by score # and doing the rest on the smaller subset. pre_nms_limit = min(6000, self.anchors.shape[0]) ix = tf.nn.top_k(scores, pre_nms_limit, sorted=True, name="top_anchors").indices scores = utils.batch_slice([scores, ix], lambda x, y: tf.gather(x, y), self.config.IMAGES_PER_GPU) deltas = utils.batch_slice([deltas, ix], lambda x, y: tf.gather(x, y), self.config.IMAGES_PER_GPU) anchors = utils.batch_slice(ix, lambda x: tf.gather(anchors, x), self.config.IMAGES_PER_GPU, names=["pre_nms_anchors"]) # Apply deltas to anchors to get refined anchors. # [batch, N, (y1, x1, y2, x2)] boxes = utils.batch_slice([anchors, deltas], lambda x, y: apply_box_deltas_graph(x, y), self.config.IMAGES_PER_GPU, names=["refined_anchors"]) # Clip to image boundaries. [batch, N, (y1, x1, y2, x2)] height, width = self.config.IMAGE_SHAPE[:2] window = np.array([0, 0, height, width]).astype(np.float32) boxes = utils.batch_slice(boxes, lambda x: clip_boxes_graph(x, window), self.config.IMAGES_PER_GPU, names=["refined_anchors_clipped"]) # Filter out small boxes # According to Xinlei Chen's paper, this reduces detection accuracy # for small objects, so we're skipping it. # Normalize dimensions to range of 0 to 1. normalized_boxes = boxes / np.array([[height, width, height, width]]) # Non-max suppression def nms(normalized_boxes, scores): indices = tf.image.non_max_suppression( normalized_boxes, scores, self.proposal_count, self.nms_threshold, name="rpn_non_max_suppression") proposals = tf.gather(normalized_boxes, indices) # Pad if needed padding = tf.maximum(self.proposal_count - tf.shape(proposals)[0], 0) proposals = tf.pad(proposals, [(0, padding), (0, 0)]) return proposals proposals = utils.batch_slice([normalized_boxes, scores], nms, self.config.IMAGES_PER_GPU) return proposals def compute_output_shape(self, input_shape): return (None, self.proposal_count, 4) ############################################################ # ROIAlign Layer ############################################################ def log2_graph(x): """Implementatin of Log2. TF doesn't have a native implemenation.""" return tf.log(x) / tf.log(2.0) class PyramidROIAlign(KE.Layer): """Implements ROI Pooling on multiple levels of the feature pyramid. Params: - pool_shape: [height, width] of the output pooled regions. Usually [7, 7] - image_shape: [height, width, channels]. Shape of input image in pixels Inputs: - boxes: [batch, num_boxes, (y1, x1, y2, x2)] in normalized coordinates. Possibly padded with zeros if not enough boxes to fill the array. - Feature maps: List of feature maps from different levels of the pyramid. Each is [batch, height, width, channels] Output: Pooled regions in the shape: [batch, num_boxes, height, width, channels]. The width and height are those specific in the pool_shape in the layer constructor. """ def __init__(self, pool_shape, image_shape, **kwargs): super(PyramidROIAlign, self).__init__(**kwargs) self.pool_shape = tuple(pool_shape) self.image_shape = tuple(image_shape) def call(self, inputs): # Crop boxes [batch, num_boxes, (y1, x1, y2, x2)] in normalized coords boxes = inputs[0] # Feature Maps. List of feature maps from different level of the # feature pyramid. Each is [batch, height, width, channels] feature_maps = inputs[1:] # Assign each ROI to a level in the pyramid based on the ROI area. y1, x1, y2, x2 = tf.split(boxes, 4, axis=2) h = y2 - y1 w = x2 - x1 # Equation 1 in the Feature Pyramid Networks paper. Account for # the fact that our coordinates are normalized here. # e.g. a 224x224 ROI (in pixels) maps to P4 image_area = tf.cast( self.image_shape[0] * self.image_shape[1], tf.float32) roi_level = log2_graph(tf.sqrt(h * w) / (224.0 / tf.sqrt(image_area))) roi_level = tf.minimum(5, tf.maximum( 2, 4 + tf.cast(tf.round(roi_level), tf.int32))) roi_level = tf.squeeze(roi_level, 2) # Loop through levels and apply ROI pooling to each. P2 to P5. pooled = [] box_to_level = [] for i, level in enumerate(range(2, 6)): ix = tf.where(tf.equal(roi_level, level)) level_boxes = tf.gather_nd(boxes, ix) # Box indicies for crop_and_resize. box_indices = tf.cast(ix[:, 0], tf.int32) # Keep track of which box is mapped to which level box_to_level.append(ix) # Stop gradient propogation to ROI proposals level_boxes = tf.stop_gradient(level_boxes) box_indices = tf.stop_gradient(box_indices) # Crop and Resize # From Mask R-CNN paper: "We sample four regular locations, so # that we can evaluate either max or average pooling. In fact, # interpolating only a single value at each bin center (without # pooling) is nearly as effective." # # Here we use the simplified approach of a single value per bin, # which is how it's done in tf.crop_and_resize() # Result: [batch * num_boxes, pool_height, pool_width, channels] pooled.append(tf.image.crop_and_resize( feature_maps[i], level_boxes, box_indices, self.pool_shape, method="bilinear")) # Pack pooled features into one tensor pooled = tf.concat(pooled, axis=0) # Pack box_to_level mapping into one array and add another # column representing the order of pooled boxes box_to_level = tf.concat(box_to_level, axis=0) box_range = tf.expand_dims(tf.range(tf.shape(box_to_level)[0]), 1) box_to_level = tf.concat([tf.cast(box_to_level, tf.int32), box_range], axis=1) # Rearrange pooled features to match the order of the original boxes # Sort box_to_level by batch then box index # TF doesn't have a way to sort by two columns, so merge them and sort. sorting_tensor = box_to_level[:, 0] * 100000 + box_to_level[:, 1] ix = tf.nn.top_k(sorting_tensor, k=tf.shape( box_to_level)[0]).indices[::-1] ix = tf.gather(box_to_level[:, 2], ix) pooled = tf.gather(pooled, ix) # Re-add the batch dimension pooled = tf.expand_dims(pooled, 0) return pooled def compute_output_shape(self, input_shape): return input_shape[0][:2] + self.pool_shape + (input_shape[1][-1], ) ############################################################ # Detection Target Layer ############################################################ def overlaps_graph(boxes1, boxes2): """Computes IoU overlaps between two sets of boxes. boxes1, boxes2: [N, (y1, x1, y2, x2)]. """ # 1. Tile boxes2 and repeate boxes1. This allows us to compare # every boxes1 against every boxes2 without loops. # TF doesn't have an equivalent to np.repeate() so simulate it # using tf.tile() and tf.reshape. b1 = tf.reshape(tf.tile(tf.expand_dims(boxes1, 1), [1, 1, tf.shape(boxes2)[0]]), [-1, 4]) b2 = tf.tile(boxes2, [tf.shape(boxes1)[0], 1]) # 2. Compute intersections b1_y1, b1_x1, b1_y2, b1_x2 = tf.split(b1, 4, axis=1) b2_y1, b2_x1, b2_y2, b2_x2 = tf.split(b2, 4, axis=1) y1 = tf.maximum(b1_y1, b2_y1) x1 = tf.maximum(b1_x1, b2_x1) y2 = tf.minimum(b1_y2, b2_y2) x2 = tf.minimum(b1_x2, b2_x2) intersection = tf.maximum(x2 - x1, 0) * tf.maximum(y2 - y1, 0) # 3. Compute unions b1_area = (b1_y2 - b1_y1) * (b1_x2 - b1_x1) b2_area = (b2_y2 - b2_y1) * (b2_x2 - b2_x1) union = b1_area + b2_area - intersection # 4. Compute IoU and reshape to [boxes1, boxes2] iou = intersection / union overlaps = tf.reshape(iou, [tf.shape(boxes1)[0], tf.shape(boxes2)[0]]) return overlaps def detection_targets_graph(proposals, gt_class_ids, gt_boxes, gt_masks, config): """Generates detection targets for one image. Subsamples proposals and generates target class IDs, bounding box deltas, and masks for each. Inputs: proposals: [N, (y1, x1, y2, x2)] in normalized coordinates. Might be zero padded if there are not enough proposals. gt_class_ids: [MAX_GT_INSTANCES] int class IDs gt_boxes: [MAX_GT_INSTANCES, (y1, x1, y2, x2)] in normalized coordinates. gt_masks: [height, width, MAX_GT_INSTANCES] of boolean type. Returns: Target ROIs and corresponding class IDs, bounding box shifts, and masks. rois: [TRAIN_ROIS_PER_IMAGE, (y1, x1, y2, x2)] in normalized coordinates class_ids: [TRAIN_ROIS_PER_IMAGE]. Integer class IDs. Zero padded. deltas: [TRAIN_ROIS_PER_IMAGE, NUM_CLASSES, (dy, dx, log(dh), log(dw))] Class-specific bbox refinements. masks: [TRAIN_ROIS_PER_IMAGE, height, width). Masks cropped to bbox boundaries and resized to neural network output size. Note: Returned arrays might be zero padded if not enough target ROIs. """ # Assertions asserts = [ tf.Assert(tf.greater(tf.shape(proposals)[0], 0), [proposals], name="roi_assertion"), ] with tf.control_dependencies(asserts): proposals = tf.identity(proposals) # Remove zero padding proposals, _ = trim_zeros_graph(proposals, name="trim_proposals") gt_boxes, non_zeros = trim_zeros_graph(gt_boxes, name="trim_gt_boxes") gt_class_ids = tf.boolean_mask(gt_class_ids, non_zeros, name="trim_gt_class_ids") gt_masks = tf.gather(gt_masks, tf.where(non_zeros)[:, 0], axis=2, name="trim_gt_masks") # Handle COCO crowds # A crowd box in COCO is a bounding box around several instances. Exclude # them from training. A crowd box is given a negative class ID. crowd_ix = tf.where(gt_class_ids < 0)[:, 0] non_crowd_ix = tf.where(gt_class_ids > 0)[:, 0] crowd_boxes = tf.gather(gt_boxes, crowd_ix) crowd_masks = tf.gather(gt_masks, crowd_ix, axis=2) gt_class_ids = tf.gather(gt_class_ids, non_crowd_ix) gt_boxes = tf.gather(gt_boxes, non_crowd_ix) gt_masks = tf.gather(gt_masks, non_crowd_ix, axis=2) # Compute overlaps matrix [proposals, gt_boxes] overlaps = overlaps_graph(proposals, gt_boxes) # Compute overlaps with crowd boxes [anchors, crowds] crowd_overlaps = overlaps_graph(proposals, crowd_boxes) crowd_iou_max = tf.reduce_max(crowd_overlaps, axis=1) no_crowd_bool = (crowd_iou_max < 0.001) # Determine postive and negative ROIs roi_iou_max = tf.reduce_max(overlaps, axis=1) # 1. Positive ROIs are those with >= 0.5 IoU with a GT box positive_roi_bool = (roi_iou_max >= 0.5) positive_indices = tf.where(positive_roi_bool)[:, 0] # 2. Negative ROIs are those with < 0.5 with every GT box. Skip crowds. negative_indices = tf.where(tf.logical_and(roi_iou_max < 0.5, no_crowd_bool))[:, 0] # Subsample ROIs. Aim for 33% positive # Positive ROIs positive_count = int(config.TRAIN_ROIS_PER_IMAGE * config.ROI_POSITIVE_RATIO) positive_indices = tf.random_shuffle(positive_indices)[:positive_count] positive_count = tf.shape(positive_indices)[0] # Negative ROIs. Add enough to maintain positive:negative ratio. r = 1.0 / config.ROI_POSITIVE_RATIO negative_count = tf.cast(r * tf.cast(positive_count, tf.float32), tf.int32) - positive_count negative_indices = tf.random_shuffle(negative_indices)[:negative_count] # Gather selected ROIs positive_rois = tf.gather(proposals, positive_indices) negative_rois = tf.gather(proposals, negative_indices) # Assign positive ROIs to GT boxes. positive_overlaps = tf.gather(overlaps, positive_indices) roi_gt_box_assignment = tf.argmax(positive_overlaps, axis=1) roi_gt_boxes = tf.gather(gt_boxes, roi_gt_box_assignment) roi_gt_class_ids = tf.gather(gt_class_ids, roi_gt_box_assignment) # Compute bbox refinement for positive ROIs deltas = utils.box_refinement_graph(positive_rois, roi_gt_boxes) deltas /= config.BBOX_STD_DEV # Assign positive ROIs to GT masks # Permute masks to [N, height, width, 1] transposed_masks = tf.expand_dims(tf.transpose(gt_masks, [2, 0, 1]), -1) # Pick the right mask for each ROI roi_masks = tf.gather(transposed_masks, roi_gt_box_assignment) # Compute mask targets boxes = positive_rois if config.USE_MINI_MASK: # Transform ROI corrdinates from normalized image space # to normalized mini-mask space. y1, x1, y2, x2 = tf.split(positive_rois, 4, axis=1) gt_y1, gt_x1, gt_y2, gt_x2 = tf.split(roi_gt_boxes, 4, axis=1) gt_h = gt_y2 - gt_y1 gt_w = gt_x2 - gt_x1 y1 = (y1 - gt_y1) / gt_h x1 = (x1 - gt_x1) / gt_w y2 = (y2 - gt_y1) / gt_h x2 = (x2 - gt_x1) / gt_w boxes = tf.concat([y1, x1, y2, x2], 1) box_ids = tf.range(0, tf.shape(roi_masks)[0]) masks = tf.image.crop_and_resize(tf.cast(roi_masks, tf.float32), boxes, box_ids, config.MASK_SHAPE) # Remove the extra dimension from masks. masks = tf.squeeze(masks, axis=3) # Threshold mask pixels at 0.5 to have GT masks be 0 or 1 to use with # binary cross entropy loss. masks = tf.round(masks) # Append negative ROIs and pad bbox deltas and masks that # are not used for negative ROIs with zeros. rois = tf.concat([positive_rois, negative_rois], axis=0) N = tf.shape(negative_rois)[0] P = tf.maximum(config.TRAIN_ROIS_PER_IMAGE - tf.shape(rois)[0], 0) rois = tf.pad(rois, [(0, P), (0, 0)]) roi_gt_boxes = tf.pad(roi_gt_boxes, [(0, N + P), (0, 0)]) roi_gt_class_ids = tf.pad(roi_gt_class_ids, [(0, N + P)]) deltas = tf.pad(deltas, [(0, N + P), (0, 0)]) masks = tf.pad(masks, [[0, N + P], (0, 0), (0, 0)]) return rois, roi_gt_class_ids, deltas, masks class DetectionTargetLayer(KE.Layer): """Subsamples proposals and generates target box refinement, class_ids, and masks for each. Inputs: proposals: [batch, N, (y1, x1, y2, x2)] in normalized coordinates. Might be zero padded if there are not enough proposals. gt_class_ids: [batch, MAX_GT_INSTANCES] Integer class IDs. gt_boxes: [batch, MAX_GT_INSTANCES, (y1, x1, y2, x2)] in normalized coordinates. gt_masks: [batch, height, width, MAX_GT_INSTANCES] of boolean type Returns: Target ROIs and corresponding class IDs, bounding box shifts, and masks. rois: [batch, TRAIN_ROIS_PER_IMAGE, (y1, x1, y2, x2)] in normalized coordinates target_class_ids: [batch, TRAIN_ROIS_PER_IMAGE]. Integer class IDs. target_deltas: [batch, TRAIN_ROIS_PER_IMAGE, NUM_CLASSES, (dy, dx, log(dh), log(dw), class_id)] Class-specific bbox refinements. target_mask: [batch, TRAIN_ROIS_PER_IMAGE, height, width) Masks cropped to bbox boundaries and resized to neural network output size. Note: Returned arrays might be zero padded if not enough target ROIs. """ def __init__(self, config, **kwargs): super(DetectionTargetLayer, self).__init__(**kwargs) self.config = config def call(self, inputs): proposals = inputs[0] gt_class_ids = inputs[1] gt_boxes = inputs[2] gt_masks = inputs[3] # Slice the batch and run a graph for each slice # TODO: Rename target_bbox to target_deltas for clarity names = ["rois", "target_class_ids", "target_bbox", "target_mask"] outputs = utils.batch_slice( [proposals, gt_class_ids, gt_boxes, gt_masks], lambda w, x, y, z: detection_targets_graph( w, x, y, z, self.config), self.config.IMAGES_PER_GPU, names=names) return outputs def compute_output_shape(self, input_shape): return [ (None, self.config.TRAIN_ROIS_PER_IMAGE, 4), # rois (None, 1), # class_ids (None, self.config.TRAIN_ROIS_PER_IMAGE, 4), # deltas (None, self.config.TRAIN_ROIS_PER_IMAGE, self.config.MASK_SHAPE[0], self.config.MASK_SHAPE[1]) # masks ] def compute_mask(self, inputs, mask=None): return [None, None, None, None] ############################################################ # Detection Layer ############################################################ def clip_to_window(window, boxes): """ window: (y1, x1, y2, x2). The window in the image we want to clip to. boxes: [N, (y1, x1, y2, x2)] """ boxes[:, 0] = np.maximum(np.minimum(boxes[:, 0], window[2]), window[0]) boxes[:, 1] = np.maximum(np.minimum(boxes[:, 1], window[3]), window[1]) boxes[:, 2] = np.maximum(np.minimum(boxes[:, 2], window[2]), window[0]) boxes[:, 3] = np.maximum(np.minimum(boxes[:, 3], window[3]), window[1]) return boxes def refine_detections_graph(rois, probs, deltas, window, config): """Refine classified proposals and filter overlaps and return final detections. Inputs: rois: [N, (y1, x1, y2, x2)] in normalized coordinates probs: [N, num_classes]. Class probabilities. deltas: [N, num_classes, (dy, dx, log(dh), log(dw))]. Class-specific bounding box deltas. window: (y1, x1, y2, x2) in image coordinates. The part of the image that contains the image excluding the padding. Returns detections shaped: [N, (y1, x1, y2, x2, class_id, score)] where coordinates are in image domain. """ # Class IDs per ROI class_ids = tf.argmax(probs, axis=1, output_type=tf.int32) # Class probability of the top class of each ROI indices = tf.stack([tf.range(probs.shape[0]), class_ids], axis=1) class_scores = tf.gather_nd(probs, indices) # Class-specific bounding box deltas deltas_specific = tf.gather_nd(deltas, indices) # Apply bounding box deltas # Shape: [boxes, (y1, x1, y2, x2)] in normalized coordinates refined_rois = apply_box_deltas_graph( rois, deltas_specific * config.BBOX_STD_DEV) # Convert coordiates to image domain # TODO: better to keep them normalized until later height, width = config.IMAGE_SHAPE[:2] refined_rois *= tf.constant([height, width, height, width], dtype=tf.float32) # Clip boxes to image window refined_rois = clip_boxes_graph(refined_rois, window) # Round and cast to int since we're deadling with pixels now refined_rois = tf.to_int32(tf.rint(refined_rois)) # TODO: Filter out boxes with zero area # Filter out background boxes keep = tf.where(class_ids > 0)[:, 0] # Filter out low confidence boxes if config.DETECTION_MIN_CONFIDENCE: conf_keep = tf.where(class_scores >= config.DETECTION_MIN_CONFIDENCE)[:, 0] keep = tf.sets.set_intersection(tf.expand_dims(keep, 0), tf.expand_dims(conf_keep, 0)) keep = tf.sparse_tensor_to_dense(keep)[0] # Apply per-class NMS # 1. Prepare variables pre_nms_class_ids = tf.gather(class_ids, keep) pre_nms_scores = tf.gather(class_scores, keep) pre_nms_rois = tf.gather(refined_rois, keep) unique_pre_nms_class_ids = tf.unique(pre_nms_class_ids)[0] def nms_keep_map(class_id): """Apply Non-Maximum Suppression on ROIs of the given class.""" # Indices of ROIs of the given class ixs = tf.where(tf.equal(pre_nms_class_ids, class_id))[:, 0] # Apply NMS class_keep = tf.image.non_max_suppression( tf.to_float(tf.gather(pre_nms_rois, ixs)), tf.gather(pre_nms_scores, ixs), max_output_size=config.DETECTION_MAX_INSTANCES, iou_threshold=config.DETECTION_NMS_THRESHOLD) # Map indicies class_keep = tf.gather(keep, tf.gather(ixs, class_keep)) # Pad with -1 so returned tensors have the same shape gap = config.DETECTION_MAX_INSTANCES - tf.shape(class_keep)[0] class_keep = tf.pad(class_keep, [(0, gap)], mode='CONSTANT', constant_values=-1) # Set shape so map_fn() can infer result shape class_keep.set_shape([config.DETECTION_MAX_INSTANCES]) return class_keep # 2. Map over class IDs nms_keep = tf.map_fn(nms_keep_map, unique_pre_nms_class_ids, dtype=tf.int64) # 3. Merge results into one list, and remove -1 padding nms_keep = tf.reshape(nms_keep, [-1]) nms_keep = tf.gather(nms_keep, tf.where(nms_keep > -1)[:, 0]) # 4. Compute intersection between keep and nms_keep keep = tf.sets.set_intersection(tf.expand_dims(keep, 0), tf.expand_dims(nms_keep, 0)) keep = tf.sparse_tensor_to_dense(keep)[0] # Keep top detections roi_count = config.DETECTION_MAX_INSTANCES class_scores_keep = tf.gather(class_scores, keep) num_keep = tf.minimum(tf.shape(class_scores_keep)[0], roi_count) top_ids = tf.nn.top_k(class_scores_keep, k=num_keep, sorted=True)[1] keep = tf.gather(keep, top_ids) # Arrange output as [N, (y1, x1, y2, x2, class_id, score)] # Coordinates are in image domain. detections = tf.concat([ tf.to_float(tf.gather(refined_rois, keep)), tf.to_float(tf.gather(class_ids, keep))[..., tf.newaxis], tf.gather(class_scores, keep)[..., tf.newaxis] ], axis=1) # Pad with zeros if detections < DETECTION_MAX_INSTANCES gap = config.DETECTION_MAX_INSTANCES - tf.shape(detections)[0] detections = tf.pad(detections, [(0, gap), (0, 0)], "CONSTANT") return detections class DetectionLayer(KE.Layer): """Takes classified proposal boxes and their bounding box deltas and returns the final detection boxes. Returns: [batch, num_detections, (y1, x1, y2, x2, class_id, class_score)] where coordinates are in image domain """ def __init__(self, config=None, **kwargs): super(DetectionLayer, self).__init__(**kwargs) self.config = config def call(self, inputs): rois = inputs[0] mrcnn_class = inputs[1] mrcnn_bbox = inputs[2] image_meta = inputs[3] # Run detection refinement graph on each item in the batch _, _, window, _ = parse_image_meta_graph(image_meta) detections_batch = utils.batch_slice( [rois, mrcnn_class, mrcnn_bbox, window], lambda x, y, w, z: refine_detections_graph(x, y, w, z, self.config), self.config.IMAGES_PER_GPU) # Reshape output # [batch, num_detections, (y1, x1, y2, x2, class_score)] in pixels return tf.reshape( detections_batch, [self.config.BATCH_SIZE, self.config.DETECTION_MAX_INSTANCES, 6]) def compute_output_shape(self, input_shape): return (None, self.config.DETECTION_MAX_INSTANCES, 6) # Region Proposal Network (RPN) def rpn_graph(feature_map, anchors_per_location, anchor_stride): """Builds the computation graph of Region Proposal Network. feature_map: backbone features [batch, height, width, depth] anchors_per_location: number of anchors per pixel in the feature map anchor_stride: Controls the density of anchors. Typically 1 (anchors for every pixel in the feature map), or 2 (every other pixel). Returns: rpn_logits: [batch, H, W, 2] Anchor classifier logits (before softmax) rpn_probs: [batch, H, W, 2] Anchor classifier probabilities. rpn_bbox: [batch, H, W, (dy, dx, log(dh), log(dw))] Deltas to be applied to anchors. """ # TODO: check if stride of 2 causes alignment issues if the featuremap # is not even. # Shared convolutional base of the RPN shared = KL.Conv2D(512, (3, 3), padding='same', activation='relu', strides=anchor_stride, name='rpn_conv_shared')(feature_map) # Anchor Score. [batch, height, width, anchors per location * 2]. x = KL.Conv2D(2 * anchors_per_location, (1, 1), padding='valid', activation='linear', name='rpn_class_raw')(shared) # Reshape to [batch, anchors, 2] rpn_class_logits = KL.Lambda( lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 2]))(x) # Softmax on last dimension of BG/FG. rpn_probs = KL.Activation( "softmax", name="rpn_class_xxx")(rpn_class_logits) # Bounding box refinement. [batch, H, W, anchors per location, depth] # where depth is [x, y, log(w), log(h)] x = KL.Conv2D(anchors_per_location * 4, (1, 1), padding="valid", activation='linear', name='rpn_bbox_pred')(shared) # Reshape to [batch, anchors, 4] rpn_bbox = KL.Lambda(lambda t: tf.reshape(t, [tf.shape(t)[0], -1, 4]))(x) return [rpn_class_logits, rpn_probs, rpn_bbox] def build_rpn_model(anchor_stride, anchors_per_location, depth): """Builds a Keras model of the Region Proposal Network. It wraps the RPN graph so it can be used multiple times with shared weights. anchors_per_location: number of anchors per pixel in the feature map anchor_stride: Controls the density of anchors. Typically 1 (anchors for every pixel in the feature map), or 2 (every other pixel). depth: Depth of the backbone feature map. Returns a Keras Model object. The model outputs, when called, are: rpn_logits: [batch, H, W, 2] Anchor classifier logits (before softmax) rpn_probs: [batch, W, W, 2] Anchor classifier probabilities. rpn_bbox: [batch, H, W, (dy, dx, log(dh), log(dw))] Deltas to be applied to anchors. """ input_feature_map = KL.Input(shape=[None, None, depth], name="input_rpn_feature_map") outputs = rpn_graph(input_feature_map, anchors_per_location, anchor_stride) return KM.Model([input_feature_map], outputs, name="rpn_model") ############################################################ # Feature Pyramid Network Heads ############################################################ def fpn_classifier_graph(rois, feature_maps, image_shape, pool_size, num_classes): """Builds the computation graph of the feature pyramid network classifier and regressor heads. rois: [batch, num_rois, (y1, x1, y2, x2)] Proposal boxes in normalized coordinates. feature_maps: List of feature maps from diffent layers of the pyramid, [P2, P3, P4, P5]. Each has a different resolution. image_shape: [height, width, depth] pool_size: The width of the square feature map generated from ROI Pooling. num_classes: number of classes, which determines the depth of the results Returns: logits: [N, NUM_CLASSES] classifier logits (before softmax) probs: [N, NUM_CLASSES] classifier probabilities bbox_deltas: [N, (dy, dx, log(dh), log(dw))] Deltas to apply to proposal boxes """ # ROI Pooling # Shape: [batch, num_boxes, pool_height, pool_width, channels] x = PyramidROIAlign([pool_size, pool_size], image_shape, name="roi_align_classifier")([rois] + feature_maps) # Two 1024 FC layers (implemented with Conv2D for consistency) x = KL.TimeDistributed(KL.Conv2D(1024, (pool_size, pool_size), padding="valid"), name="mrcnn_class_conv1")(x) x = KL.TimeDistributed(BatchNorm(axis=3), name='mrcnn_class_bn1')(x) x = KL.Activation('relu')(x) x = KL.TimeDistributed(KL.Conv2D(1024, (1, 1)), name="mrcnn_class_conv2")(x) x = KL.TimeDistributed(BatchNorm(axis=3), name='mrcnn_class_bn2')(x) x = KL.Activation('relu')(x) shared = KL.Lambda(lambda x: K.squeeze(K.squeeze(x, 3), 2), name="pool_squeeze")(x) # Classifier head mrcnn_class_logits = KL.TimeDistributed(KL.Dense(num_classes), name='mrcnn_class_logits')(shared) mrcnn_probs = KL.TimeDistributed(KL.Activation("softmax"), name="mrcnn_class")(mrcnn_class_logits) # BBox head # [batch, boxes, num_classes * (dy, dx, log(dh), log(dw))] x = KL.TimeDistributed(KL.Dense(num_classes * 4, activation='linear'), name='mrcnn_bbox_fc')(shared) # Reshape to [batch, boxes, num_classes, (dy, dx, log(dh), log(dw))] s = K.int_shape(x) mrcnn_bbox = KL.Reshape((s[1], num_classes, 4), name="mrcnn_bbox")(x) return mrcnn_class_logits, mrcnn_probs, mrcnn_bbox def build_fpn_mask_graph(rois, feature_maps, image_shape, pool_size, num_classes): """Builds the computation graph of the mask head of Feature Pyramid Network. rois: [batch, num_rois, (y1, x1, y2, x2)] Proposal boxes in normalized coordinates. feature_maps: List of feature maps from diffent layers of the pyramid, [P2, P3, P4, P5]. Each has a different resolution. image_shape: [height, width, depth] pool_size: The width of the square feature map generated from ROI Pooling. num_classes: number of classes, which determines the depth of the results Returns: Masks [batch, roi_count, height, width, num_classes] """ # ROI Pooling # Shape: [batch, boxes, pool_height, pool_width, channels] x = PyramidROIAlign([pool_size, pool_size], image_shape, name="roi_align_mask")([rois] + feature_maps) # Conv layers x = KL.TimeDistributed(KL.Conv2D(256, (3, 3), padding="same"), name="mrcnn_mask_conv1")(x) x = KL.TimeDistributed(BatchNorm(axis=3), name='mrcnn_mask_bn1')(x) x = KL.Activation('relu')(x) x = KL.TimeDistributed(KL.Conv2D(256, (3, 3), padding="same"), name="mrcnn_mask_conv2")(x) x = KL.TimeDistributed(BatchNorm(axis=3), name='mrcnn_mask_bn2')(x) x = KL.Activation('relu')(x) x = KL.TimeDistributed(KL.Conv2D(256, (3, 3), padding="same"), name="mrcnn_mask_conv3")(x) x = KL.TimeDistributed(BatchNorm(axis=3), name='mrcnn_mask_bn3')(x) x = KL.Activation('relu')(x) x = KL.TimeDistributed(KL.Conv2D(256, (3, 3), padding="same"), name="mrcnn_mask_conv4")(x) x = KL.TimeDistributed(BatchNorm(axis=3), name='mrcnn_mask_bn4')(x) x = KL.Activation('relu')(x) x = KL.TimeDistributed(KL.Conv2DTranspose(256, (2, 2), strides=2, activation="relu"), name="mrcnn_mask_deconv")(x) x = KL.TimeDistributed(KL.Conv2D(num_classes, (1, 1), strides=1, activation="sigmoid"), name="mrcnn_mask")(x) return x ############################################################ # Loss Functions ############################################################ def smooth_l1_loss(y_true, y_pred): """Implements Smooth-L1 loss. y_true and y_pred are typicallly: [N, 4], but could be any shape. """ diff = K.abs(y_true - y_pred) less_than_one = K.cast(K.less(diff, 1.0), "float32") loss = (less_than_one * 0.5 * diff**2) + (1 - less_than_one) * (diff - 0.5) return loss def rpn_class_loss_graph(rpn_match, rpn_class_logits): """RPN anchor classifier loss. rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive, -1=negative, 0=neutral anchor. rpn_class_logits: [batch, anchors, 2]. RPN classifier logits for FG/BG. """ # Squeeze last dim to simplify rpn_match = tf.squeeze(rpn_match, -1) # Get anchor classes. Convert the -1/+1 match to 0/1 values. anchor_class = K.cast(K.equal(rpn_match, 1), tf.int32) # Positive and Negative anchors contribute to the loss, # but neutral anchors (match value = 0) don't. indices = tf.where(K.not_equal(rpn_match, 0)) # Pick rows that contribute to the loss and filter out the rest. rpn_class_logits = tf.gather_nd(rpn_class_logits, indices) anchor_class = tf.gather_nd(anchor_class, indices) # Crossentropy loss loss = K.sparse_categorical_crossentropy(target=anchor_class, output=rpn_class_logits, from_logits=True) loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0)) return loss def rpn_bbox_loss_graph(config, target_bbox, rpn_match, rpn_bbox): """Return the RPN bounding box loss graph. config: the model config object. target_bbox: [batch, max positive anchors, (dy, dx, log(dh), log(dw))]. Uses 0 padding to fill in unsed bbox deltas. rpn_match: [batch, anchors, 1]. Anchor match type. 1=positive, -1=negative, 0=neutral anchor. rpn_bbox: [batch, anchors, (dy, dx, log(dh), log(dw))] """ # Positive anchors contribute to the loss, but negative and # neutral anchors (match value of 0 or -1) don't. rpn_match = K.squeeze(rpn_match, -1) indices = tf.where(K.equal(rpn_match, 1)) # Pick bbox deltas that contribute to the loss rpn_bbox = tf.gather_nd(rpn_bbox, indices) # Trim target bounding box deltas to the same length as rpn_bbox. batch_counts = K.sum(K.cast(K.equal(rpn_match, 1), tf.int32), axis=1) target_bbox = batch_pack_graph(target_bbox, batch_counts, config.IMAGES_PER_GPU) # TODO: use smooth_l1_loss() rather than reimplementing here # to reduce code duplication diff = K.abs(target_bbox - rpn_bbox) less_than_one = K.cast(K.less(diff, 1.0), "float32") loss = (less_than_one * 0.5 * diff**2) + (1 - less_than_one) * (diff - 0.5) loss = K.switch(tf.size(loss) > 0, K.mean(loss), tf.constant(0.0)) return loss def mrcnn_class_loss_graph(target_class_ids, pred_class_logits, active_class_ids): """Loss for the classifier head of Mask RCNN. target_class_ids: [batch, num_rois]. Integer class IDs. Uses zero padding to fill in the array. pred_class_logits: [batch, num_rois, num_classes] active_class_ids: [batch, num_classes]. Has a value of 1 for classes that are in the dataset of the image, and 0 for classes that are not in the dataset. """ target_class_ids = tf.cast(target_class_ids, 'int64') # Find predictions of classes that are not in the dataset. pred_class_ids = tf.argmax(pred_class_logits, axis=2) # TODO: Update this line to work with batch > 1. Right now it assumes all # images in a batch have the same active_class_ids pred_active = tf.gather(active_class_ids[0], pred_class_ids) # Loss loss = tf.nn.sparse_softmax_cross_entropy_with_logits( labels=target_class_ids, logits=pred_class_logits) # Erase losses of predictions of classes that are not in the active # classes of the image. loss = loss * pred_active # Computer loss mean. Use only predictions that contribute # to the loss to get a correct mean. loss = tf.reduce_sum(loss) / tf.reduce_sum(pred_active) return loss def mrcnn_bbox_loss_graph(target_bbox, target_class_ids, pred_bbox): """Loss for Mask R-CNN bounding box refinement. target_bbox: [batch, num_rois, (dy, dx, log(dh), log(dw))] target_class_ids: [batch, num_rois]. Integer class IDs. pred_bbox: [batch, num_rois, num_classes, (dy, dx, log(dh), log(dw))] """ # Reshape to merge batch and roi dimensions for simplicity. target_class_ids = K.reshape(target_class_ids, (-1,)) target_bbox = K.reshape(target_bbox, (-1, 4)) pred_bbox = K.reshape(pred_bbox, (-1, K.int_shape(pred_bbox)[2], 4)) # Only positive ROIs contribute to the loss. And only # the right class_id of each ROI. Get their indicies. positive_roi_ix = tf.where(target_class_ids > 0)[:, 0] positive_roi_class_ids = tf.cast( tf.gather(target_class_ids, positive_roi_ix), tf.int64) indices = tf.stack([positive_roi_ix, positive_roi_class_ids], axis=1) # Gather the deltas (predicted and true) that contribute to loss target_bbox = tf.gather(target_bbox, positive_roi_ix) pred_bbox = tf.gather_nd(pred_bbox, indices) # Smooth-L1 Loss loss = K.switch(tf.size(target_bbox) > 0, smooth_l1_loss(y_true=target_bbox, y_pred=pred_bbox), tf.constant(0.0)) loss = K.mean(loss) loss = K.reshape(loss, [1, 1]) return loss def mrcnn_mask_loss_graph(target_masks, target_class_ids, pred_masks): """Mask binary cross-entropy loss for the masks head. target_masks: [batch, num_rois, height, width]. A float32 tensor of values 0 or 1. Uses zero padding to fill array. target_class_ids: [batch, num_rois]. Integer class IDs. Zero padded. pred_masks: [batch, proposals, height, width, num_classes] float32 tensor with values from 0 to 1. """ # Reshape for simplicity. Merge first two dimensions into one. target_class_ids = K.reshape(target_class_ids, (-1,)) mask_shape = tf.shape(target_masks) target_masks = K.reshape(target_masks, (-1, mask_shape[2], mask_shape[3])) pred_shape = tf.shape(pred_masks) pred_masks = K.reshape(pred_masks, (-1, pred_shape[2], pred_shape[3], pred_shape[4])) # Permute predicted masks to [N, num_classes, height, width] pred_masks = tf.transpose(pred_masks, [0, 3, 1, 2]) # Only positive ROIs contribute to the loss. And only # the class specific mask of each ROI. positive_ix = tf.where(target_class_ids > 0)[:, 0] positive_class_ids = tf.cast( tf.gather(target_class_ids, positive_ix), tf.int64) indices = tf.stack([positive_ix, positive_class_ids], axis=1) # Gather the masks (predicted and true) that contribute to loss y_true = tf.gather(target_masks, positive_ix) y_pred = tf.gather_nd(pred_masks, indices) # Compute binary cross entropy. If no positive ROIs, then return 0. # shape: [batch, roi, num_classes] loss = K.switch(tf.size(y_true) > 0, K.binary_crossentropy(target=y_true, output=y_pred), tf.constant(0.0)) loss = K.mean(loss) loss = K.reshape(loss, [1, 1]) return loss ############################################################ # Data Generator ############################################################ def load_image_gt(dataset, config, image_id, augment=False, use_mini_mask=False): """Load and return ground truth data for an image (image, mask, bounding boxes). augment: If true, apply random image augmentation. Currently, only horizontal flipping is offered. use_mini_mask: If False, returns full-size masks that are the same height and width as the original image. These can be big, for example 1024x1024x100 (for 100 instances). Mini masks are smaller, typically, 224x224 and are generated by extracting the bounding box of the object and resizing it to MINI_MASK_SHAPE. Returns: image: [height, width, 3] shape: the original shape of the image before resizing and cropping. class_ids: [instance_count] Integer class IDs bbox: [instance_count, (y1, x1, y2, x2)] mask: [height, width, instance_count]. The height and width are those of the image unless use_mini_mask is True, in which case they are defined in MINI_MASK_SHAPE. """ # Load image and mask image = dataset.load_image(image_id) mask, class_ids = dataset.load_mask(image_id) shape = image.shape image, window, scale, padding = utils.resize_image( image, min_dim=config.IMAGE_MIN_DIM, max_dim=config.IMAGE_MAX_DIM, padding=config.IMAGE_PADDING) mask = utils.resize_mask(mask, scale, padding) # Random horizontal flips. if augment: if random.randint(0, 1): image = np.fliplr(image) mask = np.fliplr(mask) # Bounding boxes. Note that some boxes might be all zeros # if the corresponding mask got cropped out. # bbox: [num_instances, (y1, x1, y2, x2)] bbox = utils.extract_bboxes(mask) # Active classes # Different datasets have different classes, so track the # classes supported in the dataset of this image. active_class_ids = np.zeros([dataset.num_classes], dtype=np.int32) source_class_ids = dataset.source_class_ids[dataset.image_info[image_id]["source"]] active_class_ids[source_class_ids] = 1 # Resize masks to smaller size to reduce memory usage if use_mini_mask: mask = utils.minimize_mask(bbox, mask, config.MINI_MASK_SHAPE) # Image meta data image_meta = compose_image_meta(image_id, shape, window, active_class_ids) return image, image_meta, class_ids, bbox, mask def build_detection_targets(rpn_rois, gt_class_ids, gt_boxes, gt_masks, config): """Generate targets for training Stage 2 classifier and mask heads. This is not used in normal training. It's useful for debugging or to train the Mask RCNN heads without using the RPN head. Inputs: rpn_rois: [N, (y1, x1, y2, x2)] proposal boxes. gt_class_ids: [instance count] Integer class IDs gt_boxes: [instance count, (y1, x1, y2, x2)] gt_masks: [height, width, instance count] Grund truth masks. Can be full size or mini-masks. Returns: rois: [TRAIN_ROIS_PER_IMAGE, (y1, x1, y2, x2)] class_ids: [TRAIN_ROIS_PER_IMAGE]. Integer class IDs. bboxes: [TRAIN_ROIS_PER_IMAGE, NUM_CLASSES, (y, x, log(h), log(w))]. Class-specific bbox refinements. masks: [TRAIN_ROIS_PER_IMAGE, height, width, NUM_CLASSES). Class specific masks cropped to bbox boundaries and resized to neural network output size. """ assert rpn_rois.shape[0] > 0 assert gt_class_ids.dtype == np.int32, "Expected int but got {}".format( gt_class_ids.dtype) assert gt_boxes.dtype == np.int32, "Expected int but got {}".format( gt_boxes.dtype) assert gt_masks.dtype == np.bool_, "Expected bool but got {}".format( gt_masks.dtype) # It's common to add GT Boxes to ROIs but we don't do that here because # according to XinLei Chen's paper, it doesn't help. # Trim empty padding in gt_boxes and gt_masks parts instance_ids = np.where(gt_class_ids > 0)[0] assert instance_ids.shape[0] > 0, "Image must contain instances." gt_class_ids = gt_class_ids[instance_ids] gt_boxes = gt_boxes[instance_ids] gt_masks = gt_masks[:, :, instance_ids] # Compute areas of ROIs and ground truth boxes. rpn_roi_area = (rpn_rois[:, 2] - rpn_rois[:, 0]) * \ (rpn_rois[:, 3] - rpn_rois[:, 1]) gt_box_area = (gt_boxes[:, 2] - gt_boxes[:, 0]) * \ (gt_boxes[:, 3] - gt_boxes[:, 1]) # Compute overlaps [rpn_rois, gt_boxes] overlaps = np.zeros((rpn_rois.shape[0], gt_boxes.shape[0])) for i in range(overlaps.shape[1]): gt = gt_boxes[i] overlaps[:, i] = utils.compute_iou( gt, rpn_rois, gt_box_area[i], rpn_roi_area) # Assign ROIs to GT boxes rpn_roi_iou_argmax = np.argmax(overlaps, axis=1) rpn_roi_iou_max = overlaps[np.arange( overlaps.shape[0]), rpn_roi_iou_argmax] # GT box assigned to each ROI rpn_roi_gt_boxes = gt_boxes[rpn_roi_iou_argmax] rpn_roi_gt_class_ids = gt_class_ids[rpn_roi_iou_argmax] # Positive ROIs are those with >= 0.5 IoU with a GT box. fg_ids = np.where(rpn_roi_iou_max > 0.5)[0] # Negative ROIs are those with max IoU 0.1-0.5 (hard example mining) # TODO: To hard example mine or not to hard example mine, that's the question # bg_ids = np.where((rpn_roi_iou_max >= 0.1) & (rpn_roi_iou_max < 0.5))[0] bg_ids = np.where(rpn_roi_iou_max < 0.5)[0] # Subsample ROIs. Aim for 33% foreground. # FG fg_roi_count = int(config.TRAIN_ROIS_PER_IMAGE * config.ROI_POSITIVE_RATIO) if fg_ids.shape[0] > fg_roi_count: keep_fg_ids = np.random.choice(fg_ids, fg_roi_count, replace=False) else: keep_fg_ids = fg_ids # BG remaining = config.TRAIN_ROIS_PER_IMAGE - keep_fg_ids.shape[0] if bg_ids.shape[0] > remaining: keep_bg_ids = np.random.choice(bg_ids, remaining, replace=False) else: keep_bg_ids = bg_ids # Combine indicies of ROIs to keep keep = np.concatenate([keep_fg_ids, keep_bg_ids]) # Need more? remaining = config.TRAIN_ROIS_PER_IMAGE - keep.shape[0] if remaining > 0: # Looks like we don't have enough samples to maintain the desired # balance. Reduce requirements and fill in the rest. This is # likely different from the Mask RCNN paper. # There is a small chance we have neither fg nor bg samples. if keep.shape[0] == 0: # Pick bg regions with easier IoU threshold bg_ids = np.where(rpn_roi_iou_max < 0.5)[0] assert bg_ids.shape[0] >= remaining keep_bg_ids = np.random.choice(bg_ids, remaining, replace=False) assert keep_bg_ids.shape[0] == remaining keep = np.concatenate([keep, keep_bg_ids]) else: # Fill the rest with repeated bg rois. keep_extra_ids = np.random.choice( keep_bg_ids, remaining, replace=True) keep = np.concatenate([keep, keep_extra_ids]) assert keep.shape[0] == config.TRAIN_ROIS_PER_IMAGE, \ "keep doesn't match ROI batch size {}, {}".format( keep.shape[0], config.TRAIN_ROIS_PER_IMAGE) # Reset the gt boxes assigned to BG ROIs. rpn_roi_gt_boxes[keep_bg_ids, :] = 0 rpn_roi_gt_class_ids[keep_bg_ids] = 0 # For each kept ROI, assign a class_id, and for FG ROIs also add bbox refinement. rois = rpn_rois[keep] roi_gt_boxes = rpn_roi_gt_boxes[keep] roi_gt_class_ids = rpn_roi_gt_class_ids[keep] roi_gt_assignment = rpn_roi_iou_argmax[keep] # Class-aware bbox deltas. [y, x, log(h), log(w)] bboxes = np.zeros((config.TRAIN_ROIS_PER_IMAGE, config.NUM_CLASSES, 4), dtype=np.float32) pos_ids = np.where(roi_gt_class_ids > 0)[0] bboxes[pos_ids, roi_gt_class_ids[pos_ids]] = utils.box_refinement( rois[pos_ids], roi_gt_boxes[pos_ids, :4]) # Normalize bbox refinements bboxes /= config.BBOX_STD_DEV # Generate class-specific target masks. masks = np.zeros((config.TRAIN_ROIS_PER_IMAGE, config.MASK_SHAPE[0], config.MASK_SHAPE[1], config.NUM_CLASSES), dtype=np.float32) for i in pos_ids: class_id = roi_gt_class_ids[i] assert class_id > 0, "class id must be greater than 0" gt_id = roi_gt_assignment[i] class_mask = gt_masks[:, :, gt_id] if config.USE_MINI_MASK: # Create a mask placeholder, the size of the image placeholder = np.zeros(config.IMAGE_SHAPE[:2], dtype=bool) # GT box gt_y1, gt_x1, gt_y2, gt_x2 = gt_boxes[gt_id] gt_w = gt_x2 - gt_x1 gt_h = gt_y2 - gt_y1 # Resize mini mask to size of GT box placeholder[gt_y1:gt_y2, gt_x1:gt_x2] = \ np.round(scipy.misc.imresize(class_mask.astype(float), (gt_h, gt_w), interp='nearest') / 255.0).astype(bool) # Place the mini batch in the placeholder class_mask = placeholder # Pick part of the mask and resize it y1, x1, y2, x2 = rois[i].astype(np.int32) m = class_mask[y1:y2, x1:x2] mask = scipy.misc.imresize( m.astype(float), config.MASK_SHAPE, interp='nearest') / 255.0 masks[i, :, :, class_id] = mask return rois, roi_gt_class_ids, bboxes, masks def build_rpn_targets(image_shape, anchors, gt_class_ids, gt_boxes, config): """Given the anchors and GT boxes, compute overlaps and identify positive anchors and deltas to refine them to match their corresponding GT boxes. anchors: [num_anchors, (y1, x1, y2, x2)] gt_class_ids: [num_gt_boxes] Integer class IDs. gt_boxes: [num_gt_boxes, (y1, x1, y2, x2)] Returns: rpn_match: [N] (int32) matches between anchors and GT boxes. 1 = positive anchor, -1 = negative anchor, 0 = neutral rpn_bbox: [N, (dy, dx, log(dh), log(dw))] Anchor bbox deltas. """ # RPN Match: 1 = positive anchor, -1 = negative anchor, 0 = neutral rpn_match = np.zeros([anchors.shape[0]], dtype=np.int32) # RPN bounding boxes: [max anchors per image, (dy, dx, log(dh), log(dw))] rpn_bbox = np.zeros((config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4)) # Handle COCO crowds # A crowd box in COCO is a bounding box around several instances. Exclude # them from training. A crowd box is given a negative class ID. crowd_ix = np.where(gt_class_ids < 0)[0] if crowd_ix.shape[0] > 0: # Filter out crowds from ground truth class IDs and boxes non_crowd_ix = np.where(gt_class_ids > 0)[0] crowd_boxes = gt_boxes[crowd_ix] gt_class_ids = gt_class_ids[non_crowd_ix] gt_boxes = gt_boxes[non_crowd_ix] # Compute overlaps with crowd boxes [anchors, crowds] crowd_overlaps = utils.compute_overlaps(anchors, crowd_boxes) crowd_iou_max = np.amax(crowd_overlaps, axis=1) no_crowd_bool = (crowd_iou_max < 0.001) else: # All anchors don't intersect a crowd no_crowd_bool = np.ones([anchors.shape[0]], dtype=bool) # Compute overlaps [num_anchors, num_gt_boxes] overlaps = utils.compute_overlaps(anchors, gt_boxes) # Match anchors to GT Boxes # If an anchor overlaps a GT box with IoU >= 0.7 then it's positive. # If an anchor overlaps a GT box with IoU < 0.3 then it's negative. # Neutral anchors are those that don't match the conditions above, # and they don't influence the loss function. # However, don't keep any GT box unmatched (rare, but happens). Instead, # match it to the closest anchor (even if its max IoU is < 0.3). # # 1. Set negative anchors first. They get overwritten below if a GT box is # matched to them. Skip boxes in crowd areas. anchor_iou_argmax = np.argmax(overlaps, axis=1) anchor_iou_max = overlaps[np.arange(overlaps.shape[0]), anchor_iou_argmax] rpn_match[(anchor_iou_max < 0.3) & (no_crowd_bool)] = -1 # 2. Set an anchor for each GT box (regardless of IoU value). # TODO: If multiple anchors have the same IoU match all of them gt_iou_argmax = np.argmax(overlaps, axis=0) rpn_match[gt_iou_argmax] = 1 # 3. Set anchors with high overlap as positive. rpn_match[anchor_iou_max >= 0.7] = 1 # Subsample to balance positive and negative anchors # Don't let positives be more than half the anchors ids = np.where(rpn_match == 1)[0] extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE // 2) if extra > 0: # Reset the extra ones to neutral ids = np.random.choice(ids, extra, replace=False) rpn_match[ids] = 0 # Same for negative proposals ids = np.where(rpn_match == -1)[0] extra = len(ids) - (config.RPN_TRAIN_ANCHORS_PER_IMAGE - np.sum(rpn_match == 1)) if extra > 0: # Rest the extra ones to neutral ids = np.random.choice(ids, extra, replace=False) rpn_match[ids] = 0 # For positive anchors, compute shift and scale needed to transform them # to match the corresponding GT boxes. ids = np.where(rpn_match == 1)[0] ix = 0 # index into rpn_bbox # TODO: use box_refinement() rather than duplicating the code here for i, a in zip(ids, anchors[ids]): # Closest gt box (it might have IoU < 0.7) gt = gt_boxes[anchor_iou_argmax[i]] # Convert coordinates to center plus width/height. # GT Box gt_h = gt[2] - gt[0] gt_w = gt[3] - gt[1] gt_center_y = gt[0] + 0.5 * gt_h gt_center_x = gt[1] + 0.5 * gt_w # Anchor a_h = a[2] - a[0] a_w = a[3] - a[1] a_center_y = a[0] + 0.5 * a_h a_center_x = a[1] + 0.5 * a_w # Compute the bbox refinement that the RPN should predict. rpn_bbox[ix] = [ (gt_center_y - a_center_y) / a_h, (gt_center_x - a_center_x) / a_w, np.log(gt_h / a_h), np.log(gt_w / a_w), ] # Normalize rpn_bbox[ix] /= config.RPN_BBOX_STD_DEV ix += 1 return rpn_match, rpn_bbox def generate_random_rois(image_shape, count, gt_class_ids, gt_boxes): """Generates ROI proposals similar to what a region proposal network would generate. image_shape: [Height, Width, Depth] count: Number of ROIs to generate gt_class_ids: [N] Integer ground truth class IDs gt_boxes: [N, (y1, x1, y2, x2)] Ground truth boxes in pixels. Returns: [count, (y1, x1, y2, x2)] ROI boxes in pixels. """ # placeholder rois = np.zeros((count, 4), dtype=np.int32) # Generate random ROIs around GT boxes (90% of count) rois_per_box = int(0.9 * count / gt_boxes.shape[0]) for i in range(gt_boxes.shape[0]): gt_y1, gt_x1, gt_y2, gt_x2 = gt_boxes[i] h = gt_y2 - gt_y1 w = gt_x2 - gt_x1 # random boundaries r_y1 = max(gt_y1 - h, 0) r_y2 = min(gt_y2 + h, image_shape[0]) r_x1 = max(gt_x1 - w, 0) r_x2 = min(gt_x2 + w, image_shape[1]) # To avoid generating boxes with zero area, we generate double what # we need and filter out the extra. If we get fewer valid boxes # than we need, we loop and try again. while True: y1y2 = np.random.randint(r_y1, r_y2, (rois_per_box * 2, 2)) x1x2 = np.random.randint(r_x1, r_x2, (rois_per_box * 2, 2)) # Filter out zero area boxes threshold = 1 y1y2 = y1y2[np.abs(y1y2[:, 0] - y1y2[:, 1]) >= threshold][:rois_per_box] x1x2 = x1x2[np.abs(x1x2[:, 0] - x1x2[:, 1]) >= threshold][:rois_per_box] if y1y2.shape[0] == rois_per_box and x1x2.shape[0] == rois_per_box: break # Sort on axis 1 to ensure x1 <= x2 and y1 <= y2 and then reshape # into x1, y1, x2, y2 order x1, x2 = np.split(np.sort(x1x2, axis=1), 2, axis=1) y1, y2 = np.split(np.sort(y1y2, axis=1), 2, axis=1) box_rois = np.hstack([y1, x1, y2, x2]) rois[rois_per_box * i:rois_per_box * (i + 1)] = box_rois # Generate random ROIs anywhere in the image (10% of count) remaining_count = count - (rois_per_box * gt_boxes.shape[0]) # To avoid generating boxes with zero area, we generate double what # we need and filter out the extra. If we get fewer valid boxes # than we need, we loop and try again. while True: y1y2 = np.random.randint(0, image_shape[0], (remaining_count * 2, 2)) x1x2 = np.random.randint(0, image_shape[1], (remaining_count * 2, 2)) # Filter out zero area boxes threshold = 1 y1y2 = y1y2[np.abs(y1y2[:, 0] - y1y2[:, 1]) >= threshold][:remaining_count] x1x2 = x1x2[np.abs(x1x2[:, 0] - x1x2[:, 1]) >= threshold][:remaining_count] if y1y2.shape[0] == remaining_count and x1x2.shape[0] == remaining_count: break # Sort on axis 1 to ensure x1 <= x2 and y1 <= y2 and then reshape # into x1, y1, x2, y2 order x1, x2 = np.split(np.sort(x1x2, axis=1), 2, axis=1) y1, y2 = np.split(np.sort(y1y2, axis=1), 2, axis=1) global_rois = np.hstack([y1, x1, y2, x2]) rois[-remaining_count:] = global_rois return rois def data_generator(dataset, config, shuffle=True, augment=True, random_rois=0, batch_size=1, detection_targets=False): """A generator that returns images and corresponding target class ids, bounding box deltas, and masks. dataset: The Dataset object to pick data from config: The model config object shuffle: If True, shuffles the samples before every epoch augment: If True, applies image augmentation to images (currently only horizontal flips are supported) random_rois: If > 0 then generate proposals to be used to train the network classifier and mask heads. Useful if training the Mask RCNN part without the RPN. batch_size: How many images to return in each call detection_targets: If True, generate detection targets (class IDs, bbox deltas, and masks). Typically for debugging or visualizations because in trainig detection targets are generated by DetectionTargetLayer. Returns a Python generator. Upon calling next() on it, the generator returns two lists, inputs and outputs. The containtes of the lists differs depending on the received arguments: inputs list: - images: [batch, H, W, C] - image_meta: [batch, size of image meta] - rpn_match: [batch, N] Integer (1=positive anchor, -1=negative, 0=neutral) - rpn_bbox: [batch, N, (dy, dx, log(dh), log(dw))] Anchor bbox deltas. - gt_class_ids: [batch, MAX_GT_INSTANCES] Integer class IDs - gt_boxes: [batch, MAX_GT_INSTANCES, (y1, x1, y2, x2)] - gt_masks: [batch, height, width, MAX_GT_INSTANCES]. The height and width are those of the image unless use_mini_mask is True, in which case they are defined in MINI_MASK_SHAPE. outputs list: Usually empty in regular training. But if detection_targets is True then the outputs list contains target class_ids, bbox deltas, and masks. """ b = 0 # batch item index image_index = -1 image_ids = np.copy(dataset.image_ids) error_count = 0 # Anchors # [anchor_count, (y1, x1, y2, x2)] anchors = utils.generate_pyramid_anchors(config.RPN_ANCHOR_SCALES, config.RPN_ANCHOR_RATIOS, config.BACKBONE_SHAPES, config.BACKBONE_STRIDES, config.RPN_ANCHOR_STRIDE) # Keras requires a generator to run indefinately. while True: try: # Increment index to pick next image. Shuffle if at the start of an epoch. image_index = (image_index + 1) % len(image_ids) if shuffle and image_index == 0: np.random.shuffle(image_ids) # Get GT bounding boxes and masks for image. image_id = image_ids[image_index] image, image_meta, gt_class_ids, gt_boxes, gt_masks = \ load_image_gt(dataset, config, image_id, augment=augment, use_mini_mask=config.USE_MINI_MASK) # Skip images that have no instances. This can happen in cases # where we train on a subset of classes and the image doesn't # have any of the classes we care about. if not np.any(gt_class_ids > 0): continue # RPN Targets rpn_match, rpn_bbox = build_rpn_targets(image.shape, anchors, gt_class_ids, gt_boxes, config) # Mask R-CNN Targets if random_rois: rpn_rois = generate_random_rois( image.shape, random_rois, gt_class_ids, gt_boxes) if detection_targets: rois, mrcnn_class_ids, mrcnn_bbox, mrcnn_mask =\ build_detection_targets( rpn_rois, gt_class_ids, gt_boxes, gt_masks, config) # Init batch arrays if b == 0: batch_image_meta = np.zeros( (batch_size,) + image_meta.shape, dtype=image_meta.dtype) batch_rpn_match = np.zeros( [batch_size, anchors.shape[0], 1], dtype=rpn_match.dtype) batch_rpn_bbox = np.zeros( [batch_size, config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4], dtype=rpn_bbox.dtype) batch_images = np.zeros( (batch_size,) + image.shape, dtype=np.float32) batch_gt_class_ids = np.zeros( (batch_size, config.MAX_GT_INSTANCES), dtype=np.int32) batch_gt_boxes = np.zeros( (batch_size, config.MAX_GT_INSTANCES, 4), dtype=np.int32) if config.USE_MINI_MASK: batch_gt_masks = np.zeros((batch_size, config.MINI_MASK_SHAPE[0], config.MINI_MASK_SHAPE[1], config.MAX_GT_INSTANCES)) else: batch_gt_masks = np.zeros( (batch_size, image.shape[0], image.shape[1], config.MAX_GT_INSTANCES)) if random_rois: batch_rpn_rois = np.zeros( (batch_size, rpn_rois.shape[0], 4), dtype=rpn_rois.dtype) if detection_targets: batch_rois = np.zeros( (batch_size,) + rois.shape, dtype=rois.dtype) batch_mrcnn_class_ids = np.zeros( (batch_size,) + mrcnn_class_ids.shape, dtype=mrcnn_class_ids.dtype) batch_mrcnn_bbox = np.zeros( (batch_size,) + mrcnn_bbox.shape, dtype=mrcnn_bbox.dtype) batch_mrcnn_mask = np.zeros( (batch_size,) + mrcnn_mask.shape, dtype=mrcnn_mask.dtype) # If more instances than fits in the array, sub-sample from them. if gt_boxes.shape[0] > config.MAX_GT_INSTANCES: ids = np.random.choice( np.arange(gt_boxes.shape[0]), config.MAX_GT_INSTANCES, replace=False) gt_class_ids = gt_class_ids[ids] gt_boxes = gt_boxes[ids] gt_masks = gt_masks[:, :, ids] # Add to batch batch_image_meta[b] = image_meta batch_rpn_match[b] = rpn_match[:, np.newaxis] batch_rpn_bbox[b] = rpn_bbox batch_images[b] = mold_image(image.astype(np.float32), config) batch_gt_class_ids[b, :gt_class_ids.shape[0]] = gt_class_ids batch_gt_boxes[b, :gt_boxes.shape[0]] = gt_boxes batch_gt_masks[b, :, :, :gt_masks.shape[-1]] = gt_masks if random_rois: batch_rpn_rois[b] = rpn_rois if detection_targets: batch_rois[b] = rois batch_mrcnn_class_ids[b] = mrcnn_class_ids batch_mrcnn_bbox[b] = mrcnn_bbox batch_mrcnn_mask[b] = mrcnn_mask b += 1 # Batch full? if b >= batch_size: inputs = [batch_images, batch_image_meta, batch_rpn_match, batch_rpn_bbox, batch_gt_class_ids, batch_gt_boxes, batch_gt_masks] outputs = [] if random_rois: inputs.extend([batch_rpn_rois]) if detection_targets: inputs.extend([batch_rois]) # Keras requires that output and targets have the same number of dimensions batch_mrcnn_class_ids = np.expand_dims( batch_mrcnn_class_ids, -1) outputs.extend( [batch_mrcnn_class_ids, batch_mrcnn_bbox, batch_mrcnn_mask]) yield inputs, outputs # start a new batch b = 0 except (GeneratorExit, KeyboardInterrupt): raise except: # Log it and skip the image logging.exception("Error processing image {}".format( dataset.image_info[image_id])) error_count += 1 if error_count > 5: raise ############################################################ # MaskRCNN Class ############################################################ class MaskRCNN(): """Encapsulates the Mask RCNN model functionality. The actual Keras model is in the keras_model property. """ def __init__(self, mode, config, model_dir): """ mode: Either "training" or "inference" config: A Sub-class of the Config class model_dir: Directory to save training logs and trained weights """ assert mode in ['training', 'inference'] self.mode = mode self.config = config self.model_dir = model_dir self.set_log_dir() self.keras_model = self.build(mode=mode, config=config) def build(self, mode, config): """Build Mask R-CNN architecture. input_shape: The shape of the input image. mode: Either "training" or "inference". The inputs and outputs of the model differ accordingly. """ assert mode in ['training', 'inference'] # Image size must be dividable by 2 multiple times h, w = config.IMAGE_SHAPE[:2] if h / 2**6 != int(h / 2**6) or w / 2**6 != int(w / 2**6): raise Exception("Image size must be dividable by 2 at least 6 times " "to avoid fractions when downscaling and upscaling." "For example, use 256, 320, 384, 448, 512, ... etc. ") # Inputs input_image = KL.Input( shape=config.IMAGE_SHAPE.tolist(), name="input_image") input_image_meta = KL.Input(shape=[None], name="input_image_meta") if mode == "training": # RPN GT input_rpn_match = KL.Input( shape=[None, 1], name="input_rpn_match", dtype=tf.int32) input_rpn_bbox = KL.Input( shape=[None, 4], name="input_rpn_bbox", dtype=tf.float32) # Detection GT (class IDs, bounding boxes, and masks) # 1. GT Class IDs (zero padded) input_gt_class_ids = KL.Input( shape=[None], name="input_gt_class_ids", dtype=tf.int32) # 2. GT Boxes in pixels (zero padded) # [batch, MAX_GT_INSTANCES, (y1, x1, y2, x2)] in image coordinates input_gt_boxes = KL.Input( shape=[None, 4], name="input_gt_boxes", dtype=tf.float32) # Normalize coordinates h, w = K.shape(input_image)[1], K.shape(input_image)[2] image_scale = K.cast(K.stack([h, w, h, w], axis=0), tf.float32) gt_boxes = KL.Lambda(lambda x: x / image_scale)(input_gt_boxes) # 3. GT Masks (zero padded) # [batch, height, width, MAX_GT_INSTANCES] if config.USE_MINI_MASK: input_gt_masks = KL.Input( shape=[config.MINI_MASK_SHAPE[0], config.MINI_MASK_SHAPE[1], None], name="input_gt_masks", dtype=bool) else: input_gt_masks = KL.Input( shape=[config.IMAGE_SHAPE[0], config.IMAGE_SHAPE[1], None], name="input_gt_masks", dtype=bool) # Build the shared convolutional layers. # Bottom-up Layers # Returns a list of the last layers of each stage, 5 in total. # Don't create the thead (stage 5), so we pick the 4th item in the list. _, C2, C3, C4, C5 = resnet_graph(input_image, "resnet101", stage5=True) # Top-down Layers # TODO: add assert to varify feature map sizes match what's in config P5 = KL.Conv2D(256, (1, 1), name='fpn_c5p5')(C5) P4 = KL.Add(name="fpn_p4add")([ KL.UpSampling2D(size=(2, 2), name="fpn_p5upsampled")(P5), KL.Conv2D(256, (1, 1), name='fpn_c4p4')(C4)]) P3 = KL.Add(name="fpn_p3add")([ KL.UpSampling2D(size=(2, 2), name="fpn_p4upsampled")(P4), KL.Conv2D(256, (1, 1), name='fpn_c3p3')(C3)]) P2 = KL.Add(name="fpn_p2add")([ KL.UpSampling2D(size=(2, 2), name="fpn_p3upsampled")(P3), KL.Conv2D(256, (1, 1), name='fpn_c2p2')(C2)]) # Attach 3x3 conv to all P layers to get the final feature maps. P2 = KL.Conv2D(256, (3, 3), padding="SAME", name="fpn_p2")(P2) P3 = KL.Conv2D(256, (3, 3), padding="SAME", name="fpn_p3")(P3) P4 = KL.Conv2D(256, (3, 3), padding="SAME", name="fpn_p4")(P4) P5 = KL.Conv2D(256, (3, 3), padding="SAME", name="fpn_p5")(P5) # P6 is used for the 5th anchor scale in RPN. Generated by # subsampling from P5 with stride of 2. P6 = KL.MaxPooling2D(pool_size=(1, 1), strides=2, name="fpn_p6")(P5) # Note that P6 is used in RPN, but not in the classifier heads. rpn_feature_maps = [P2, P3, P4, P5, P6] mrcnn_feature_maps = [P2, P3, P4, P5] # Generate Anchors self.anchors = utils.generate_pyramid_anchors(config.RPN_ANCHOR_SCALES, config.RPN_ANCHOR_RATIOS, config.BACKBONE_SHAPES, config.BACKBONE_STRIDES, config.RPN_ANCHOR_STRIDE) # RPN Model rpn = build_rpn_model(config.RPN_ANCHOR_STRIDE, len(config.RPN_ANCHOR_RATIOS), 256) # Loop through pyramid layers layer_outputs = [] # list of lists for p in rpn_feature_maps: layer_outputs.append(rpn([p])) # Concatenate layer outputs # Convert from list of lists of level outputs to list of lists # of outputs across levels. # e.g. [[a1, b1, c1], [a2, b2, c2]] => [[a1, a2], [b1, b2], [c1, c2]] output_names = ["rpn_class_logits", "rpn_class", "rpn_bbox"] outputs = list(zip(*layer_outputs)) outputs = [KL.Concatenate(axis=1, name=n)(list(o)) for o, n in zip(outputs, output_names)] rpn_class_logits, rpn_class, rpn_bbox = outputs # Generate proposals # Proposals are [batch, N, (y1, x1, y2, x2)] in normalized coordinates # and zero padded. proposal_count = config.POST_NMS_ROIS_TRAINING if mode == "training"\ else config.POST_NMS_ROIS_INFERENCE rpn_rois = ProposalLayer(proposal_count=proposal_count, nms_threshold=config.RPN_NMS_THRESHOLD, name="ROI", anchors=self.anchors, config=config)([rpn_class, rpn_bbox]) if mode == "training": # Class ID mask to mark class IDs supported by the dataset the image # came from. _, _, _, active_class_ids = KL.Lambda(lambda x: parse_image_meta_graph(x), mask=[None, None, None, None])(input_image_meta) if not config.USE_RPN_ROIS: # Ignore predicted ROIs and use ROIs provided as an input. input_rois = KL.Input(shape=[config.POST_NMS_ROIS_TRAINING, 4], name="input_roi", dtype=np.int32) # Normalize coordinates to 0-1 range. target_rois = KL.Lambda(lambda x: K.cast( x, tf.float32) / image_scale[:4])(input_rois) else: target_rois = rpn_rois # Generate detection targets # Subsamples proposals and generates target outputs for training # Note that proposal class IDs, gt_boxes, and gt_masks are zero # padded. Equally, returned rois and targets are zero padded. rois, target_class_ids, target_bbox, target_mask =\ DetectionTargetLayer(config, name="proposal_targets")([ target_rois, input_gt_class_ids, gt_boxes, input_gt_masks]) # Network Heads # TODO: verify that this handles zero padded ROIs mrcnn_class_logits, mrcnn_class, mrcnn_bbox =\ fpn_classifier_graph(rois, mrcnn_feature_maps, config.IMAGE_SHAPE, config.POOL_SIZE, config.NUM_CLASSES) mrcnn_mask = build_fpn_mask_graph(rois, mrcnn_feature_maps, config.IMAGE_SHAPE, config.MASK_POOL_SIZE, config.NUM_CLASSES) # TODO: clean up (use tf.identify if necessary) output_rois = KL.Lambda(lambda x: x * 1, name="output_rois")(rois) # Losses rpn_class_loss = KL.Lambda(lambda x: rpn_class_loss_graph(*x), name="rpn_class_loss")( [input_rpn_match, rpn_class_logits]) rpn_bbox_loss = KL.Lambda(lambda x: rpn_bbox_loss_graph(config, *x), name="rpn_bbox_loss")( [input_rpn_bbox, input_rpn_match, rpn_bbox]) class_loss = KL.Lambda(lambda x: mrcnn_class_loss_graph(*x), name="mrcnn_class_loss")( [target_class_ids, mrcnn_class_logits, active_class_ids]) bbox_loss = KL.Lambda(lambda x: mrcnn_bbox_loss_graph(*x), name="mrcnn_bbox_loss")( [target_bbox, target_class_ids, mrcnn_bbox]) mask_loss = KL.Lambda(lambda x: mrcnn_mask_loss_graph(*x), name="mrcnn_mask_loss")( [target_mask, target_class_ids, mrcnn_mask]) # Model inputs = [input_image, input_image_meta, input_rpn_match, input_rpn_bbox, input_gt_class_ids, input_gt_boxes, input_gt_masks] if not config.USE_RPN_ROIS: inputs.append(input_rois) outputs = [rpn_class_logits, rpn_class, rpn_bbox, mrcnn_class_logits, mrcnn_class, mrcnn_bbox, mrcnn_mask, rpn_rois, output_rois, rpn_class_loss, rpn_bbox_loss, class_loss, bbox_loss, mask_loss] model = KM.Model(inputs, outputs, name='mask_rcnn') else: # Network Heads # Proposal classifier and BBox regressor heads mrcnn_class_logits, mrcnn_class, mrcnn_bbox =\ fpn_classifier_graph(rpn_rois, mrcnn_feature_maps, config.IMAGE_SHAPE, config.POOL_SIZE, config.NUM_CLASSES) # Detections # output is [batch, num_detections, (y1, x1, y2, x2, class_id, score)] in image coordinates detections = DetectionLayer(config, name="mrcnn_detection")( [rpn_rois, mrcnn_class, mrcnn_bbox, input_image_meta]) # Convert boxes to normalized coordinates # TODO: let DetectionLayer return normalized coordinates to avoid # unnecessary conversions h, w = config.IMAGE_SHAPE[:2] detection_boxes = KL.Lambda( lambda x: x[..., :4] / np.array([h, w, h, w]))(detections) # Create masks for detections mrcnn_mask = build_fpn_mask_graph(detection_boxes, mrcnn_feature_maps, config.IMAGE_SHAPE, config.MASK_POOL_SIZE, config.NUM_CLASSES) model = KM.Model([input_image, input_image_meta], [detections, mrcnn_class, mrcnn_bbox, mrcnn_mask, rpn_rois, rpn_class, rpn_bbox], name='mask_rcnn') # Add multi-GPU support. if config.GPU_COUNT > 1: from parallel_model import ParallelModel model = ParallelModel(model, config.GPU_COUNT) return model def find_last(self): """Finds the last checkpoint file of the last trained model in the model directory. Returns: log_dir: The directory where events and weights are saved checkpoint_path: the path to the last checkpoint file """ # Get directory names. Each directory corresponds to a model dir_names = next(os.walk(self.model_dir))[1] key = self.config.NAME.lower() dir_names = filter(lambda f: f.startswith(key), dir_names) dir_names = sorted(dir_names) if not dir_names: return None, None # Pick last directory dir_name = os.path.join(self.model_dir, dir_names[-1]) # Find the last checkpoint checkpoints = next(os.walk(dir_name))[2] checkpoints = filter(lambda f: f.startswith("mask_rcnn"), checkpoints) checkpoints = sorted(checkpoints) if not checkpoints: return dir_name, None checkpoint = os.path.join(dir_name, checkpoints[-1]) return dir_name, checkpoint def load_weights(self, filepath, by_name=False, exclude=None): """Modified version of the correspoding Keras function with the addition of multi-GPU support and the ability to exclude some layers from loading. exlude: list of layer names to excluce """ import h5py from keras.engine import topology if exclude: by_name = True if h5py is None: raise ImportError('`load_weights` requires h5py.') f = h5py.File(filepath, mode='r') if 'layer_names' not in f.attrs and 'model_weights' in f: f = f['model_weights'] # In multi-GPU training, we wrap the model. Get layers # of the inner model because they have the weights. keras_model = self.keras_model layers = keras_model.inner_model.layers if hasattr(keras_model, "inner_model")\ else keras_model.layers # Exclude some layers if exclude: layers = filter(lambda l: l.name not in exclude, layers) if by_name: topology.load_weights_from_hdf5_group_by_name(f, layers) else: topology.load_weights_from_hdf5_group(f, layers) if hasattr(f, 'close'): f.close() # Update the log directory self.set_log_dir(filepath) def get_imagenet_weights(self): """Downloads ImageNet trained weights from Keras. Returns path to weights file. """ from keras.utils.data_utils import get_file TF_WEIGHTS_PATH_NO_TOP = 'https://github.com/fchollet/deep-learning-models/'\ 'releases/download/v0.2/'\ 'resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5' weights_path = get_file('resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5', TF_WEIGHTS_PATH_NO_TOP, cache_subdir='models', md5_hash='a268eb855778b3df3c7506639542a6af') return weights_path def compile(self, learning_rate, momentum): """Gets the model ready for training. Adds losses, regularization, and metrics. Then calls the Keras compile() function. """ # Optimizer object optimizer = keras.optimizers.SGD(lr=learning_rate, momentum=momentum, clipnorm=5.0) # Add Losses # First, clear previously set losses to avoid duplication self.keras_model._losses = [] self.keras_model._per_input_losses = {} loss_names = ["rpn_class_loss", "rpn_bbox_loss", "mrcnn_class_loss", "mrcnn_bbox_loss", "mrcnn_mask_loss"] for name in loss_names: layer = self.keras_model.get_layer(name) if layer.output in self.keras_model.losses: continue self.keras_model.add_loss( tf.reduce_mean(layer.output, keep_dims=True)) # Add L2 Regularization # Skip gamma and beta weights of batch normalization layers. reg_losses = [keras.regularizers.l2(self.config.WEIGHT_DECAY)(w) / tf.cast(tf.size(w), tf.float32) for w in self.keras_model.trainable_weights if 'gamma' not in w.name and 'beta' not in w.name] self.keras_model.add_loss(tf.add_n(reg_losses)) # Compile self.keras_model.compile(optimizer=optimizer, loss=[ None] * len(self.keras_model.outputs)) # Add metrics for losses for name in loss_names: if name in self.keras_model.metrics_names: continue layer = self.keras_model.get_layer(name) self.keras_model.metrics_names.append(name) self.keras_model.metrics_tensors.append(tf.reduce_mean( layer.output, keep_dims=True)) def set_trainable(self, layer_regex, keras_model=None, indent=0, verbose=1): """Sets model layers as trainable if their names match the given regular expression. """ # Print message on the first call (but not on recursive calls) if verbose > 0 and keras_model is None: log("Selecting layers to train") keras_model = keras_model or self.keras_model # In multi-GPU training, we wrap the model. Get layers # of the inner model because they have the weights. layers = keras_model.inner_model.layers if hasattr(keras_model, "inner_model")\ else keras_model.layers for layer in layers: # Is the layer a model? if layer.__class__.__name__ == 'Model': print("In model: ", layer.name) self.set_trainable( layer_regex, keras_model=layer, indent=indent + 4) continue if not layer.weights: continue # Is it trainable? trainable = bool(re.fullmatch(layer_regex, layer.name)) # Update layer. If layer is a container, update inner layer. if layer.__class__.__name__ == 'TimeDistributed': layer.layer.trainable = trainable else: layer.trainable = trainable # Print trainble layer names if trainable and verbose > 0: log("{}{:20} ({})".format(" " * indent, layer.name, layer.__class__.__name__)) def set_log_dir(self, model_path=None): """Sets the model log directory and epoch counter. model_path: If None, or a format different from what this code uses then set a new log directory and start epochs from 0. Otherwise, extract the log directory and the epoch counter from the file name. """ # Set date and epoch counter as if starting a new model self.epoch = 0 now = datetime.datetime.now() # If we have a model path with date and epochs use them if model_path: # Continue from we left of. Get epoch and date from the file name # A sample model path might look like: # /path/to/logs/coco20171029T2315/mask_rcnn_coco_0001.h5 regex = r".*/\w+(\d{4})(\d{2})(\d{2})T(\d{2})(\d{2})/mask\_rcnn\_\w+(\d{4})\.h5" m = re.match(regex, model_path) if m: now = datetime.datetime(int(m.group(1)), int(m.group(2)), int(m.group(3)), int(m.group(4)), int(m.group(5))) self.epoch = int(m.group(6)) + 1 # Directory for training logs self.log_dir = os.path.join(self.model_dir, "{}{:%Y%m%dT%H%M}".format( self.config.NAME.lower(), now)) # Path to save after each epoch. Include placeholders that get filled by Keras. self.checkpoint_path = os.path.join(self.log_dir, "mask_rcnn_{}_*epoch*.h5".format( self.config.NAME.lower())) self.checkpoint_path = self.checkpoint_path.replace( "*epoch*", "{epoch:04d}") def train(self, train_dataset, val_dataset, learning_rate, epochs, layers): """Train the model. train_dataset, val_dataset: Training and validation Dataset objects. learning_rate: The learning rate to train with epochs: Number of training epochs. Note that previous training epochs are considered to be done alreay, so this actually determines the epochs to train in total rather than in this particaular call. layers: Allows selecting wich layers to train. It can be: - A regular expression to match layer names to train - One of these predefined values: heaads: The RPN, classifier and mask heads of the network all: All the layers 3+: Train Resnet stage 3 and up 4+: Train Resnet stage 4 and up 5+: Train Resnet stage 5 and up """ assert self.mode == "training", "Create model in training mode." # Pre-defined layer regular expressions layer_regex = { # all layers but the backbone "heads": r"(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)", # From a specific Resnet stage and up "3+": r"(res3.*)|(bn3.*)|(res4.*)|(bn4.*)|(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)", "4+": r"(res4.*)|(bn4.*)|(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)", "5+": r"(res5.*)|(bn5.*)|(mrcnn\_.*)|(rpn\_.*)|(fpn\_.*)", # All layers "all": ".*", } if layers in layer_regex.keys(): layers = layer_regex[layers] # Data generators train_generator = data_generator(train_dataset, self.config, shuffle=True, batch_size=self.config.BATCH_SIZE) val_generator = data_generator(val_dataset, self.config, shuffle=True, batch_size=self.config.BATCH_SIZE, augment=False) # Callbacks callbacks = [ keras.callbacks.TensorBoard(log_dir=self.log_dir, histogram_freq=0, write_graph=True, write_images=False), keras.callbacks.ModelCheckpoint(self.checkpoint_path, verbose=0, save_weights_only=True), ] # Train log("\nStarting at epoch {}. LR={}\n".format(self.epoch, learning_rate)) log("Checkpoint Path: {}".format(self.checkpoint_path)) self.set_trainable(layers) self.compile(learning_rate, self.config.LEARNING_MOMENTUM) # Work-around for Windows: Keras fails on Windows when using # multiprocessing workers. See discussion here: # https://github.com/matterport/Mask_RCNN/issues/13#issuecomment-353124009 if os.name is 'nt': workers = 0 else: workers = max(self.config.BATCH_SIZE // 2, 2) self.keras_model.fit_generator( train_generator, initial_epoch=self.epoch, epochs=epochs, steps_per_epoch=self.config.STEPS_PER_EPOCH, callbacks=callbacks, validation_data=next(val_generator), validation_steps=self.config.VALIDATION_STEPS, max_queue_size=100, workers=workers, use_multiprocessing=True, ) self.epoch = max(self.epoch, epochs) def mold_inputs(self, images): """Takes a list of images and modifies them to the format expected as an input to the neural network. images: List of image matricies [height,width,depth]. Images can have different sizes. Returns 3 Numpy matricies: molded_images: [N, h, w, 3]. Images resized and normalized. image_metas: [N, length of meta data]. Details about each image. windows: [N, (y1, x1, y2, x2)]. The portion of the image that has the original image (padding excluded). """ molded_images = [] image_metas = [] windows = [] for image in images: # Resize image to fit the model expected size # TODO: move resizing to mold_image() molded_image, window, scale, padding = utils.resize_image( image, min_dim=self.config.IMAGE_MIN_DIM, max_dim=self.config.IMAGE_MAX_DIM, padding=self.config.IMAGE_PADDING) molded_image = mold_image(molded_image, self.config) # Build image_meta image_meta = compose_image_meta( 0, image.shape, window, np.zeros([self.config.NUM_CLASSES], dtype=np.int32)) # Append molded_images.append(molded_image) windows.append(window) image_metas.append(image_meta) # Pack into arrays molded_images = np.stack(molded_images) image_metas = np.stack(image_metas) windows = np.stack(windows) return molded_images, image_metas, windows def unmold_detections(self, detections, mrcnn_mask, image_shape, window): """Reformats the detections of one image from the format of the neural network output to a format suitable for use in the rest of the application. detections: [N, (y1, x1, y2, x2, class_id, score)] mrcnn_mask: [N, height, width, num_classes] image_shape: [height, width, depth] Original size of the image before resizing window: [y1, x1, y2, x2] Box in the image where the real image is excluding the padding. Returns: boxes: [N, (y1, x1, y2, x2)] Bounding boxes in pixels class_ids: [N] Integer class IDs for each bounding box scores: [N] Float probability scores of the class_id masks: [height, width, num_instances] Instance masks """ # How many detections do we have? # Detections array is padded with zeros. Find the first class_id == 0. zero_ix = np.where(detections[:, 4] == 0)[0] N = zero_ix[0] if zero_ix.shape[0] > 0 else detections.shape[0] # Extract boxes, class_ids, scores, and class-specific masks boxes = detections[:N, :4] class_ids = detections[:N, 4].astype(np.int32) scores = detections[:N, 5] masks = mrcnn_mask[np.arange(N), :, :, class_ids] # Compute scale and shift to translate coordinates to image domain. h_scale = image_shape[0] / (window[2] - window[0]) w_scale = image_shape[1] / (window[3] - window[1]) scale = min(h_scale, w_scale) shift = window[:2] # y, x scales = np.array([scale, scale, scale, scale]) shifts = np.array([shift[0], shift[1], shift[0], shift[1]]) # Translate bounding boxes to image domain boxes = np.multiply(boxes - shifts, scales).astype(np.int32) # Filter out detections with zero area. Often only happens in early # stages of training when the network weights are still a bit random. exclude_ix = np.where( (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1]) <= 0)[0] if exclude_ix.shape[0] > 0: boxes = np.delete(boxes, exclude_ix, axis=0) class_ids = np.delete(class_ids, exclude_ix, axis=0) scores = np.delete(scores, exclude_ix, axis=0) masks = np.delete(masks, exclude_ix, axis=0) N = class_ids.shape[0] # Resize masks to original image size and set boundary threshold. full_masks = [] for i in range(N): # Convert neural network mask to full size mask full_mask = utils.unmold_mask(masks[i], boxes[i], image_shape) full_masks.append(full_mask) full_masks = np.stack(full_masks, axis=-1)\ if full_masks else np.empty((0,) + masks.shape[1:3]) return boxes, class_ids, scores, full_masks def detect(self, images, verbose=0): """Runs the detection pipeline. images: List of images, potentially of different sizes. Returns a list of dicts, one dict per image. The dict contains: rois: [N, (y1, x1, y2, x2)] detection bounding boxes class_ids: [N] int class IDs scores: [N] float probability scores for the class IDs masks: [H, W, N] instance binary masks """ assert self.mode == "inference", "Create model in inference mode." assert len( images) == self.config.BATCH_SIZE, "len(images) must be equal to BATCH_SIZE" if verbose: log("Processing {} images".format(len(images))) for image in images: log("image", image) # Mold inputs to format expected by the neural network molded_images, image_metas, windows = self.mold_inputs(images) if verbose: log("molded_images", molded_images) log("image_metas", image_metas) # Run object detection detections, mrcnn_class, mrcnn_bbox, mrcnn_mask, \ rois, rpn_class, rpn_bbox =\ self.keras_model.predict([molded_images, image_metas], verbose=0) # Process detections results = [] for i, image in enumerate(images): final_rois, final_class_ids, final_scores, final_masks =\ self.unmold_detections(detections[i], mrcnn_mask[i], image.shape, windows[i]) results.append({ "rois": final_rois, "class_ids": final_class_ids, "scores": final_scores, "masks": final_masks, }) return results def ancestor(self, tensor, name, checked=None): """Finds the ancestor of a TF tensor in the computation graph. tensor: TensorFlow symbolic tensor. name: Name of ancestor tensor to find checked: For internal use. A list of tensors that were already searched to avoid loops in traversing the graph. """ checked = checked if checked is not None else [] # Put a limit on how deep we go to avoid very long loops if len(checked) > 500: return None # Convert name to a regex and allow matching a number prefix # because Keras adds them automatically if isinstance(name, str): name = re.compile(name.replace("/", r"(\_\d+)*/")) parents = tensor.op.inputs for p in parents: if p in checked: continue if bool(re.fullmatch(name, p.name)): return p checked.append(p) a = self.ancestor(p, name, checked) if a is not None: return a return None def find_trainable_layer(self, layer): """If a layer is encapsulated by another layer, this function digs through the encapsulation and returns the layer that holds the weights. """ if layer.__class__.__name__ == 'TimeDistributed': return self.find_trainable_layer(layer.layer) return layer def get_trainable_layers(self): """Returns a list of layers that have weights.""" layers = [] # Loop through all layers for l in self.keras_model.layers: # If layer is a wrapper, find inner trainable layer l = self.find_trainable_layer(l) # Include layer if it has weights if l.get_weights(): layers.append(l) return layers def run_graph(self, images, outputs): """Runs a sub-set of the computation graph that computes the given outputs. outputs: List of tuples (name, tensor) to compute. The tensors are symbolic TensorFlow tensors and the names are for easy tracking. Returns an ordered dict of results. Keys are the names received in the input and values are Numpy arrays. """ model = self.keras_model # Organize desired outputs into an ordered dict outputs = OrderedDict(outputs) for o in outputs.values(): assert o is not None # Build a Keras function to run parts of the computation graph inputs = model.inputs if model.uses_learning_phase and not isinstance(K.learning_phase(), int): inputs += [K.learning_phase()] kf = K.function(model.inputs, list(outputs.values())) # Run inference molded_images, image_metas, windows = self.mold_inputs(images) # TODO: support training mode? # if TEST_MODE == "training": # model_in = [molded_images, image_metas, # target_rpn_match, target_rpn_bbox, # gt_boxes, gt_masks] # if not config.USE_RPN_ROIS: # model_in.append(target_rois) # if model.uses_learning_phase and not isinstance(K.learning_phase(), int): # model_in.append(1.) # outputs_np = kf(model_in) # else: model_in = [molded_images, image_metas] if model.uses_learning_phase and not isinstance(K.learning_phase(), int): model_in.append(0.) outputs_np = kf(model_in) # Pack the generated Numpy arrays into a a dict and log the results. outputs_np = OrderedDict([(k, v) for k, v in zip(outputs.keys(), outputs_np)]) for k, v in outputs_np.items(): log(k, v) return outputs_np ############################################################ # Data Formatting ############################################################ def compose_image_meta(image_id, image_shape, window, active_class_ids): """Takes attributes of an image and puts them in one 1D array. image_id: An int ID of the image. Useful for debugging. image_shape: [height, width, channels] window: (y1, x1, y2, x2) in pixels. The area of the image where the real image is (excluding the padding) active_class_ids: List of class_ids available in the dataset from which the image came. Useful if training on images from multiple datasets where not all classes are present in all datasets. """ meta = np.array( [image_id] + # size=1 list(image_shape) + # size=3 list(window) + # size=4 (y1, x1, y2, x2) in image cooredinates list(active_class_ids) # size=num_classes ) return meta def parse_image_meta_graph(meta): """Parses a tensor that contains image attributes to its components. See compose_image_meta() for more details. meta: [batch, meta length] where meta length depends on NUM_CLASSES """ image_id = meta[:, 0] image_shape = meta[:, 1:4] window = meta[:, 4:8] # (y1, x1, y2, x2) window of image in in pixels active_class_ids = meta[:, 8:] return [image_id, image_shape, window, active_class_ids] def mold_image(images, config): """Takes RGB images with 0-255 values and subtraces the mean pixel and converts it to float. Expects image colors in RGB order. """ return images.astype(np.float32) - config.MEAN_PIXEL def unmold_image(normalized_images, config): """Takes a image normalized with mold() and returns the original.""" return (normalized_images + config.MEAN_PIXEL).astype(np.uint8) ############################################################ # Miscellenous Graph Functions ############################################################ def trim_zeros_graph(boxes, name=None): """Often boxes are represented with matricies of shape [N, 4] and are padded with zeros. This removes zero boxes. boxes: [N, 4] matrix of boxes. non_zeros: [N] a 1D boolean mask identifying the rows to keep """ non_zeros = tf.cast(tf.reduce_sum(tf.abs(boxes), axis=1), tf.bool) boxes = tf.boolean_mask(boxes, non_zeros, name=name) return boxes, non_zeros def batch_pack_graph(x, counts, num_rows): """Picks different number of values from each row in x depending on the values in counts. """ outputs = [] for i in range(num_rows): outputs.append(x[i, :counts[i]]) return tf.concat(outputs, axis=0)
111,322
42.553599
115
py
SimPer
SimPer-main/src/simper.py
""" Minimal SimPer implementation & example training loops. """ import tensorflow as tf from networks import Featurizer, Classifier @tf.function def _max_cross_corr(feats_1, feats_2): # feats_1: 1 x T(# time stamp) # feats_2: M(# aug) x T(# time stamp) feats_2 = tf.cast(feats_2, feats_1.dtype) feats_1 = feats_1 - tf.math.reduce_mean(feats_1, axis=-1, keepdims=True) feats_2 = feats_2 - tf.math.reduce_mean(feats_2, axis=-1, keepdims=True) min_N = min(feats_1.shape[-1], feats_2.shape[-1]) padded_N = max(feats_1.shape[-1], feats_2.shape[-1]) * 2 feats_1_pad = tf.pad(feats_1, tf.constant([[0, 0], [0, padded_N - feats_1.shape[-1]]])) feats_2_pad = tf.pad(feats_2, tf.constant([[0, 0], [0, padded_N - feats_2.shape[-1]]])) feats_1_fft = tf.signal.rfft(feats_1_pad) feats_2_fft = tf.signal.rfft(feats_2_pad) X = feats_1_fft * tf.math.conj(feats_2_fft) power_norm = tf.cast( tf.math.reduce_std(feats_1, axis=-1, keepdims=True) * tf.math.reduce_std(feats_2, axis=-1, keepdims=True), X.dtype) power_norm = tf.where( tf.equal(power_norm, 0), tf.ones_like(power_norm), power_norm) X = X / power_norm cc = tf.signal.irfft(X) / (min_N - 1) max_cc = tf.math.reduce_max(cc, axis=-1) return max_cc @tf.function def batched_max_cross_corr(x, y): """ x: M(# aug) x T(# time stamp) y: M(# aug) x T(# time stamp) """ # Calculate distance for a single row of x. per_x_dist = lambda i: _max_cross_corr(x[i:(i + 1), :], y) # Compute and stack distances for all rows of x. dist = tf.map_fn(fn=per_x_dist, elems=tf.range(tf.shape(x)[0], dtype=tf.int64), fn_output_signature=x.dtype) return dist @tf.function def normed_psd(x, fps, zero_pad=0, high_pass=0.25, low_pass=15): """ x: M(# aug) x T(# time stamp) """ x = x - tf.math.reduce_mean(x, axis=-1, keepdims=True) if zero_pad > 0: L = x.shape[-1] x = tf.pad(x, tf.constant([[int(zero_pad / 2 * L), int(zero_pad / 2 * L)]])) x = tf.abs(tf.signal.rfft(x)) ** 2 Fn = fps / 2 freqs = tf.linspace(0., Fn, x.shape[-1]) use_freqs = tf.math.logical_and(freqs >= high_pass, freqs <= low_pass) use_freqs = tf.repeat(tf.expand_dims(use_freqs, 0), x.shape[0], axis=0) x = tf.reshape(x[use_freqs], (x.shape[0], -1)) # Normalize PSD denom = tf.math.reduce_euclidean_norm(x, axis=-1, keepdims=True) denom = tf.where(tf.equal(denom, 0), tf.ones_like(denom), denom) x = x / denom return x @tf.function def batched_normed_psd(x, y): """ x: M(# aug) x T(# time stamp) y: M(# aug) x T(# time stamp) """ return tf.matmul(normed_psd(x), normed_psd(y), transpose_b=True) def label_distance(labels_1, labels_2, dist_fn='l1', label_temperature=0.1): # labels: bsz x M(#augs) # output: bsz x M(#augs) x M(#augs) if dist_fn == 'l1': dist_mat = - tf.math.abs(labels_1[:, :, None] - labels_2[:, None, :]) elif dist_fn == 'l2': dist_mat = - tf.math.abs(labels_1[:, :, None] - labels_2[:, None, :]) ** 2 elif dist_fn == 'sqrt': dist_mat = - tf.math.abs(labels_1[:, :, None] - labels_2[:, None, :]) ** 0.5 else: raise NotImplementedError(f"`{dist_fn}` not implemented.") prob_mat = tf.nn.softmax(dist_mat / label_temperature, axis=-1) return prob_mat class SimPer(tf.keras.Model): def __init__(self, hparams): super(SimPer, self).__init__() self.hparams = hparams self.featurizer = Featurizer(self.hparams["n_frames"]) self.regressor = Classifier(self.featurizer.n_outputs, 1, False) self.network = tf.keras.Sequential([self.featurizer, self.regressor]) self.optimizer = tf.keras.optimizers.Adam(lr=self.hparams["lr"]) def update(self, minibatches): all_x, all_speed = minibatches # all_x: [bsz, 2*M, SSL_FRAMES, H, W, C] batch_size, num_augments = all_x.shape[0], all_x.shape[1] all_x = tf.reshape(all_x, [batch_size * num_augments] + all_x.shape[2:].as_list()) # [bsz, 2*M] -> [bsz, M, M] all_labels = label_distance(all_speed[:, :num_augments // 2], all_speed[:, num_augments // 2:], self.hparams["label_dist_fn"], self.hparams["label_temperature"]) criterion = tf.keras.losses.CategoricalCrossentropy(from_logits=True) with tf.GradientTape() as tape: all_z = self.featurizer(all_x) all_z = tf.reshape(all_z, [batch_size, num_augments, -1]) loss = 0 for feats, labels in zip(all_z, all_labels): feat_dist = globals()[self.hparams["feat_dist_fn"]]( feats[:num_augments // 2], feats[num_augments // 2:]) loss += criterion(y_pred=feat_dist, y_true=labels) loss /= batch_size gradients = tape.gradient(loss, self.featurizer.trainable_variables) self.optimizer.apply_gradients(zip(gradients, self.featurizer.trainable_variables)) return loss def predict(self, x, training: bool): return self.featurizer(x, training=training)
5,269
35.344828
91
py
SimPer
SimPer-main/src/networks.py
""" Example network architectures: - Featurizer (for representation learning) - Classifier (for downstream tasks) """ import tensorflow as tf from tensorflow.keras.layers import (Conv2D, Conv3D, Dense, Flatten, BatchNormalization, TimeDistributed, MaxPool2D, GlobalAveragePooling2D) class Featurizer(tf.keras.Model): def __init__(self, n_outputs): super(Featurizer, self).__init__() self.conv0 = Conv3D(64, (5, 3, 3), padding='same') self.conv1 = Conv3D(128, (5, 3, 3), padding='same') self.conv2 = Conv3D(128, (5, 3, 3), padding='same') self.conv3 = Conv3D(1, (1, 1, 1)) self.bn0 = BatchNormalization() self.bn1 = BatchNormalization() self.bn2 = BatchNormalization() self.bn3 = BatchNormalization() self.pool0 = TimeDistributed(MaxPool2D((2, 2))) self.pool1 = TimeDistributed(MaxPool2D((2, 2))) self.pool2 = TimeDistributed(MaxPool2D((2, 2))) self.pool3 = TimeDistributed(GlobalAveragePooling2D()) self.flatten = Flatten() self.n_outputs = n_outputs def call(self, x): x = self.conv0(x) x = self.bn0(x) x = tf.nn.relu(x) x = self.pool0(x) x = self.conv1(x) x = self.bn1(x) x = tf.nn.relu(x) x = self.pool1(x) # x = self.conv2(x) # x = self.bn2(x) # x = tf.nn.relu(x) # x = self.pool2(x) x = self.conv3(x) x = self.bn3(x) x = tf.nn.relu(x) x = self.pool3(x) x = self.flatten(x) return x class MLP(tf.keras.Model): def __init__(self, n_outputs): super(MLP, self).__init__() self.inputs = Dense(n_outputs) self.hidden = Dense(n_outputs) self.outputs = Dense(n_outputs) def call(self, x): x = self.inputs(x) x = tf.nn.relu(x) # x = self.hidden(x) # x = tf.nn.relu(x) x = self.outputs(x) return x def Classifier(in_features, out_features, nonlinear=False): if nonlinear: return tf.keras.Sequential( [Dense(in_features // 2, activation=tf.nn.relu), Dense(in_features // 4, activation=tf.nn.relu), Dense(out_features)]) else: return Dense(out_features)
2,324
26.678571
88
py
sampling_cf
sampling_cf-main/main.py
import os import time import importlib import datetime as dt from tqdm import tqdm from utils import file_write, log_end_epoch, INF, valid_hyper_params from data_path_constants import get_log_file_path, get_model_file_path # NOTE: No global-level torch imports as the GPU-ID is set through code def train(model, criterion, optimizer, reader, hyper_params, forgetting_events, track_events): import torch model.train() # Initializing metrics since we will calculate MSE on the train set on the fly metrics = {} # Initializations at = 0 # Train for one epoch, batch-by-batch loop = tqdm(reader) for data, y in loop: # Empty the gradients model.zero_grad() optimizer.zero_grad() # Forward pass output = model(data) # Compute per-interaction loss loss = criterion(output, y, return_mean = False) criterion.anneal(1.0 / float(len(reader) * hyper_params['epochs'])) # loop.set_description("Loss: {}".format(round(float(loss), 4))) # Track forgetting events if track_events: with torch.no_grad(): if hyper_params['task'] == 'explicit': forgetting_events[at : at+data[0].shape[0]] += loss.data else: pos_output, neg_output = output pos_output = pos_output.repeat(1, neg_output.shape[1]) num_incorrect = torch.sum((neg_output > pos_output).float(), -1) forgetting_events[at : at+data[0].shape[0]] += num_incorrect.data at += data[0].shape[0] # Backward pass loss = torch.mean(loss) loss.backward() optimizer.step() return metrics, forgetting_events def train_complete(hyper_params, train_reader, val_reader, model, model_class, track_events): import torch from loss import CustomLoss from eval import evaluate from torch_utils import is_cuda_available criterion = CustomLoss(hyper_params) optimizer = torch.optim.Adam( model.parameters(), lr=hyper_params['lr'], betas=(0.9, 0.98), weight_decay=hyper_params['weight_decay'] ) file_write(hyper_params['log_file'], str(model)) file_write(hyper_params['log_file'], "\nModel Built!\nStarting Training...\n") try: best_MSE = float(INF) best_AUC = -float(INF) best_HR = -float(INF) decreasing_streak = 0 forgetting_events = None if track_events: forgetting_events = torch.zeros(train_reader.num_interactions).float() if is_cuda_available: forgetting_events = forgetting_events.cuda() for epoch in range(1, hyper_params['epochs'] + 1): epoch_start_time = time.time() # Training for one epoch metrics, local_forgetted_count = train( model, criterion, optimizer, train_reader, hyper_params, forgetting_events, track_events ) # Calulating the metrics on the validation set if (epoch % hyper_params['validate_every'] == 0) or (epoch == 1): metrics = evaluate(model, criterion, val_reader, hyper_params, train_reader.item_propensity) metrics['dataset'] = hyper_params['dataset'] decreasing_streak += 1 # Save best model on validation set if hyper_params['task'] == 'explicit' and metrics['MSE'] < best_MSE: print("Saving model...") torch.save(model.state_dict(), hyper_params['model_path']) decreasing_streak, best_MSE = 0, metrics['MSE'] elif hyper_params['task'] != 'explicit' and metrics['AUC'] > best_AUC: print("Saving model...") torch.save(model.state_dict(), hyper_params['model_path']) decreasing_streak, best_AUC = 0, metrics['AUC'] elif hyper_params['task'] != 'explicit' and metrics['HR@10'] > best_HR: print("Saving model...") torch.save(model.state_dict(), hyper_params['model_path']) decreasing_streak, best_HR = 0, metrics['HR@10'] log_end_epoch(hyper_params, metrics, epoch, time.time() - epoch_start_time, metrics_on = '(VAL)') # Check if need to early-stop if 'early_stop' in hyper_params and decreasing_streak >= hyper_params['early_stop']: file_write(hyper_params['log_file'], "Early stopping..") break except KeyboardInterrupt: print('Exiting from training early') # Load best model and return it for evaluation on test-set if os.path.exists(hyper_params['model_path']): model = model_class(hyper_params) if is_cuda_available: model = model.cuda() model.load_state_dict(torch.load(hyper_params['model_path'])) model.eval() if track_events: forgetting_events = forgetting_events.cpu().numpy() / float(hyper_params['epochs']) return model, forgetting_events def train_neumf(hyper_params, train_reader, val_reader, track_events): from pytorch_models.NeuMF import GMF, MLP, NeuMF from torch_utils import is_cuda_available, xavier_init initial_path = hyper_params['model_path'] # Pre-Training the GMF Model hyper_params['model_path'] = initial_path[:-3] + "_gmf.pt" gmf_model = GMF(hyper_params) if is_cuda_available: gmf_model = gmf_model.cuda() xavier_init(gmf_model) gmf_model, _ = train_complete(hyper_params, train_reader, val_reader, gmf_model, GMF, track_events) # Pre-Training the MLP Model hyper_params['model_path'] = initial_path[:-3] + "_mlp.pt" mlp_model = MLP(hyper_params) if is_cuda_available: mlp_model = mlp_model.cuda() xavier_init(mlp_model) mlp_model, _ = train_complete(hyper_params, train_reader, val_reader, mlp_model, MLP, track_events) # Training the final NeuMF Model hyper_params['model_path'] = initial_path model = NeuMF(hyper_params) if is_cuda_available: model = model.cuda() model.init(gmf_model, mlp_model) model, forgetting_events = train_complete(hyper_params, train_reader, val_reader, model, NeuMF, track_events) # Remove GMF and MLP models mlp_path = initial_path[:-3] + "_mlp.pt" gmf_path = initial_path[:-3] + "_gmf.pt" os.remove(mlp_path) ; os.remove(gmf_path) return model, forgetting_events def main_pytorch(hyper_params, track_events = False, eval_full = True): from load_data import load_data from eval import evaluate from torch_utils import is_cuda_available, xavier_init, get_model_class from loss import CustomLoss if not valid_hyper_params(hyper_params): print("Invalid task combination specified, exiting.") return # Load the data readers train_reader, test_reader, val_reader, hyper_params = load_data(hyper_params, track_events = track_events) file_write(hyper_params['log_file'], "\n\nSimulation run on: " + str(dt.datetime.now()) + "\n\n") file_write(hyper_params['log_file'], "Data reading complete!") file_write(hyper_params['log_file'], "Number of train batches: {:4d}".format(len(train_reader))) file_write(hyper_params['log_file'], "Number of validation batches: {:4d}".format(len(val_reader))) file_write(hyper_params['log_file'], "Number of test batches: {:4d}".format(len(test_reader))) # Initialize & train the model start_time = time.time() if hyper_params['model_type'] == 'NeuMF': model, forgetting_events = train_neumf(hyper_params, train_reader, val_reader, track_events) else: model = get_model_class(hyper_params)(hyper_params) if is_cuda_available: model = model.cuda() xavier_init(model) model, forgetting_events = train_complete( hyper_params, train_reader, val_reader, model, get_model_class(hyper_params), track_events ) metrics = {} if eval_full: # Calculating MSE on test-set criterion = CustomLoss(hyper_params) metrics = evaluate(model, criterion, test_reader, hyper_params, train_reader.item_propensity, test = True) log_end_epoch(hyper_params, metrics, 'final', time.time() - start_time, metrics_on = '(TEST)') # We have no space left for storing the models os.remove(hyper_params['model_path']) del model, train_reader, test_reader, val_reader return metrics, forgetting_events def main_pop_rec(hyper_params): from load_data import load_data from eval import evaluate from loss import CustomLoss from pytorch_models.pop_rec import PopRec # Load the data readers train_reader, test_reader, val_reader, hyper_params = load_data(hyper_params) file_write(hyper_params['log_file'], "\n\nSimulation run on: " + str(dt.datetime.now()) + "\n\n") file_write(hyper_params['log_file'], "Data reading complete!") file_write(hyper_params['log_file'], "Number of test batches: {:4d}\n\n".format(len(test_reader))) # Make the model start_time = time.time() model = PopRec(hyper_params, train_reader.get_item_count_map()) # Calculating MSE on test-set criterion = CustomLoss(hyper_params) metrics = evaluate(model, criterion, test_reader, hyper_params, train_reader.item_propensity, test = True) log_end_epoch(hyper_params, metrics, 'final', time.time() - start_time, metrics_on = '(TEST)') del model, train_reader, test_reader, val_reader return metrics, None def main(hyper_params, gpu_id = None): if not valid_hyper_params(hyper_params): print("Invalid task combination specified, exiting.") return # Setting GPU ID for running entire code ## Very Very Imp. if gpu_id is not None: os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id) # torch.cuda.set_device(int(gpu_id)) # torch.cuda.empty_cache() # Set dataset specific hyper-params hyper_params.update( importlib.import_module('data_hyperparams.{}'.format(hyper_params['dataset'])).hyper_params ) # Learning rate is "highly" dataset AND model specific if 'lr' not in hyper_params: if hyper_params['model_type'] == 'SASRec': hyper_params['lr'] = 0.006 elif hyper_params['model_type'] == 'SVAE': hyper_params['lr'] = 0.02 elif hyper_params['model_type'] == 'MVAE': hyper_params['lr'] = 0.01 else: hyper_params['lr'] = 0.008 hyper_params['log_file'] = get_log_file_path(hyper_params) hyper_params['model_path'] = get_model_file_path(hyper_params) if hyper_params['model_type'] == 'pop_rec': main_pop_rec(hyper_params) else: main_pytorch(hyper_params) # torch.cuda.empty_cache() if __name__ == '__main__': from hyper_params import hyper_params main(hyper_params)
10,944
39.238971
114
py
sampling_cf
sampling_cf-main/data_genie.py
from data_genie.data_genie_config import * from data_genie.data_genie_trainers import * from data_genie.data_genie_data import OracleData from data_genie.data_genie_model import PointwiseDataGenie, PairwiseDataGenie # NOTE: Please edit the config in `data_genie/data_genie_config.py` before \ # running this trainer script print("Datasets:", ", ".join(datasets)) for embedding_type in [ 'handcrafted', 'unsupervised_gcn_dim_8_layers_1', 'unsupervised_gcn_dim_8_layers_3', 'unsupervised_gcn_dim_16_layers_1', 'unsupervised_gcn_dim_16_layers_3', 'unsupervised_gcn_dim_32_layers_1', ]: # If you want to try out different combinations of handcrafted features options = [ # [0, 1], # [0, 1, 3], # [0, 1, 4], # [0, 1, 5], [0, 1, 3, 4, 5], # [0, 1, 2, 3, 4, 5, 6], ] if embedding_type == 'handcrafted' else [ None ] # Create model for feats_to_keep in options: # Load data print("\n\n{} Using {} embedding {}\n\n".format("="*30, embedding_type, "="*30)) pointwise_data = OracleData(datasets, feats_to_keep, embedding_type, bsz = 128, pointwise = True) pairwise_data = OracleData(datasets, feats_to_keep, embedding_type, bsz = 128, pointwise = False) ########### Linear Regression print("\n\n{} Linear Regression {}\n\n".format("="*30, "="*30)) for C in [ 1e-4, 1e-2, 1, 1e2, 1e4 ]: train_linear_regression(pointwise_data, embedding_type, feats_to_keep, C = float(C)) ########### Logistic Regression print("\n\n{} Logistic Regression {}\n\n".format("="*30, "="*30)) for C in [ 1e-2, 1, 1e2, 1e4 ]: train_logistic_regression(pairwise_data, embedding_type, feats_to_keep, C = float(C)) ########### XGBoost print("\n\n{} XGBoost Regression {}\n\n".format("="*30, "="*30)) for max_depth in [ 2, 4, 6, 8, 10 ]: train_xgboost_regression(pointwise_data, embedding_type, feats_to_keep, max_depth = max_depth) print("\n\n{} XGBoost Classification {}\n\n".format("="*30, "="*30)) for max_depth in [ 2, 4, 6, 8, 10 ]: train_xgboost_bce(pairwise_data, embedding_type, feats_to_keep, max_depth = max_depth) ########## Neural network print("\n\n{} Data-Genie {}\n\n".format("="*30, "="*30)) for Analyzer in [ PointwiseDataGenie, PairwiseDataGenie ]: print("\nPointwise:" if Analyzer == PointwiseDataGenie else "\nPairwise:") # NOTE: These hyper-parameters were estimated using a basic grid-search LR = 0.001 WD = float(1e-6) DIM = 64 DROPOUT = 0.2 GRAPH_DIM = 64 GCN_LAYERS = 2 EPOCHS = 200 if embedding_type != 'handcrafted' else 20 train_pytorch( pointwise_data if Analyzer == PointwiseDataGenie else pairwise_data, Analyzer, feats_to_keep, LR, WD, DIM, DROPOUT, embedding_type, GRAPH_DIM, GCN_LAYERS, EPOCHS = EPOCHS, VALIDATE_EVERY = 1 )
2,771
34.088608
124
py
sampling_cf
sampling_cf-main/loss.py
import torch import torch.nn.functional as F from torch_utils import is_cuda_available class CustomLoss(torch.nn.Module): def __init__(self, hyper_params): super(CustomLoss, self).__init__() self.forward = { 'explicit': self.mse, 'implicit': self.bpr, 'sequential': self.bpr, }[hyper_params['task']] if hyper_params['model_type'] == "MVAE": self.forward = self.vae_loss if hyper_params['model_type'] == "SVAE": self.forward = self.svae_loss if hyper_params['model_type'] == "SASRec": self.forward = self.bce_sasrec self.torch_bce = torch.nn.BCEWithLogitsLoss() self.anneal_val = 0.0 self.hyper_params = hyper_params def mse(self, output, y, return_mean = True): mse = torch.pow(output - y, 2) if return_mean: return torch.mean(mse) return mse def bce_sasrec(self, output, pos, return_mean = True): pos_logits, neg_logits = output pos_labels, neg_labels = torch.ones(pos_logits.shape), torch.zeros(neg_logits.shape) if is_cuda_available: pos_labels, neg_labels = pos_labels.cuda(), neg_labels.cuda() indices = pos != self.hyper_params['total_items'] loss = self.torch_bce(pos_logits[indices], pos_labels[indices]) loss += self.torch_bce(neg_logits[indices], neg_labels[indices]) return loss def bpr(self, output, y, return_mean = True): pos_output, neg_output = output pos_output = pos_output.repeat(1, neg_output.shape[1]).view(-1) neg_output = neg_output.view(-1) loss = -F.logsigmoid(pos_output - neg_output) if return_mean: return torch.mean(loss) return loss def anneal(self, step_size): self.anneal_val += step_size self.anneal_val = max(self.anneal_val, 0.2) def vae_loss(self, output, y_true_s, return_mean = True): decoder_output, mu_q, logvar_q = output # Calculate KL Divergence loss kld = torch.mean(torch.sum(0.5 * (-logvar_q + torch.exp(logvar_q) + mu_q**2 - 1), -1)) # Calculate Likelihood decoder_output = F.log_softmax(decoder_output, -1) likelihood = torch.sum(-1.0 * y_true_s * decoder_output, -1) final = (self.anneal_val * kld) + (likelihood) if return_mean: return torch.mean(final) return final def svae_loss(self, output, y, return_mean = True): decoder_output, mu_q, logvar_q = output dec_shape = decoder_output.shape # [batch_size x seq_len x total_items] # Calculate KL Divergence loss kld = torch.mean(torch.sum(0.5 * (-logvar_q + torch.exp(logvar_q) + mu_q**2 - 1), -1)) # Don't compute loss on padded items y_true_s, y_indices = y keep_indices = y_indices != self.hyper_params['total_items'] y_true_s = y_true_s[keep_indices] decoder_output = decoder_output[keep_indices] # Calculate Likelihood decoder_output = F.log_softmax(decoder_output, -1) likelihood = torch.sum(-1.0 * y_true_s * decoder_output) likelihood = likelihood / float(dec_shape[0] * self.hyper_params['num_next']) final = (self.anneal_val * kld) + (likelihood) if return_mean: return torch.mean(final) return final
3,395
36.318681
94
py
sampling_cf
sampling_cf-main/torch_utils.py
import torch is_cuda_available = torch.cuda.is_available() if is_cuda_available: print("Using CUDA...\n") LongTensor = torch.cuda.LongTensor FloatTensor = torch.cuda.FloatTensor BoolTensor = torch.cuda.BoolTensor else: LongTensor = torch.LongTensor FloatTensor = torch.FloatTensor BoolTensor = torch.BoolTensor def get_model_class(hyper_params): from pytorch_models import MF, MVAE, SASRec, SVAE return { "bias_only": MF.MF, "MF_dot": MF.MF, "MF": MF.MF, "MVAE": MVAE.MVAE, "SVAE": SVAE.SVAE, "SASRec": SASRec.SASRec, }[hyper_params['model_type']] def xavier_init(model): for _, param in model.named_parameters(): try: torch.nn.init.xavier_uniform_(param.data) except: pass # just ignore those failed init layers
827
25.709677
59
py
sampling_cf
sampling_cf-main/eval.py
import torch import numpy as np from numba import jit, float32, float64, int64 from utils import INF def evaluate(model, criterion, reader, hyper_params, item_propensity, topk = [ 10, 100 ], test = False): metrics = {} # Do a negative sampled item-space evaluation (only on the validation set) # if the dataset is too big partial_eval = (not test) and hyper_params['partial_eval'] partial_eval = partial_eval and (hyper_params['model_type'] not in [ 'MVAE', 'SVAE', 'pop_rec' ]) if partial_eval: metrics['eval'] = 'partial' if hyper_params['task'] == 'explicit': metrics['MSE'] = 0.0 else: preds, y_binary = [], [] for kind in [ 'HR', 'NDCG', 'PSP' ]: for k in topk: metrics['{}@{}'.format(kind, k)] = 0.0 model.eval() with torch.no_grad(): for data, y in reader: output = model(data, eval = True) if hyper_params['model_type'] in [ 'MVAE', 'SVAE' ]: output, _, _ = output if hyper_params['model_type'] == 'SVAE': output = output[:, -1, :] if hyper_params['task'] == 'explicit': metrics['MSE'] += torch.sum(criterion(output, y, return_mean = False).data) else: function = evaluate_batch_partial if partial_eval else evaluate_batch metrics, temp_preds, temp_y = function(data, output, y, item_propensity, topk, metrics) preds += temp_preds y_binary += temp_y if hyper_params['task'] == 'explicit': metrics['MSE'] = round(float(metrics['MSE']) / reader.num_interactions, 4) else: # NOTE: sklearn's `roc_auc_score` is suuuuper slow metrics['AUC'] = round(fast_auc(np.array(y_binary), np.array(preds)), 4) for kind in [ 'HR', 'NDCG', 'PSP' ]: for k in topk: metrics['{}@{}'.format(kind, k)] = round( float(100.0 * metrics['{}@{}'.format(kind, k)]) / reader.num_interactions, 4 ) return metrics def evaluate_batch(data, output_batch, y, item_propensity, topk, metrics): # Y train_positive, test_positive_set = y # Data _, _, auc_negatives = data # AUC Stuff temp_preds, temp_y = [], [] logits_cpu = output_batch.cpu().numpy() for b in range(len(output_batch)): # Validation set could have 0 positive interactions if len(test_positive_set[b]) == 0: continue temp_preds += np.take(logits_cpu[b], np.array(list(test_positive_set[b]))).tolist() temp_y += [ 1.0 for _ in range(len(test_positive_set[b])) ] temp_preds += np.take(logits_cpu[b], auc_negatives[b]).tolist() temp_y += [ 0.0 for _ in range(len(auc_negatives[b])) ] # Marking train-set consumed items as negative INF for b in range(len(output_batch)): output_batch[b][ train_positive[b] ] = -INF _, indices = torch.topk(output_batch, min(item_propensity.shape[0], max(topk)), sorted = True) indices = indices.cpu().numpy().tolist() for k in topk: for b in range(len(output_batch)): num_pos = float(len(test_positive_set[b])) # Validation set could have 0 positive interactions after sampling if num_pos == 0: continue metrics['HR@{}'.format(k)] += float(len(set(indices[b][:k]) & test_positive_set[b])) / float(min(num_pos, k)) test_positive_sorted_psp = sorted([ item_propensity[x] for x in test_positive_set[b] ])[::-1] dcg, idcg, psp, max_psp = 0.0, 0.0, 0.0, 0.0 for at, pred in enumerate(indices[b][:k]): if pred in test_positive_set[b]: dcg += 1.0 / np.log2(at + 2) psp += float(item_propensity[pred]) / float(min(num_pos, k)) if at < num_pos: idcg += 1.0 / np.log2(at + 2) max_psp += test_positive_sorted_psp[at] metrics['NDCG@{}'.format(k)] += dcg / idcg metrics['PSP@{}'.format(k)] += psp / max_psp return metrics, temp_preds, temp_y def evaluate_batch_partial(data, output, y, item_propensity, topk, metrics): _, test_pos_items, _ = data test_pos_items = test_pos_items.cpu().numpy() pos_score, neg_score = output pos_score, neg_score = pos_score.cpu().numpy(), neg_score.cpu().numpy() temp_preds, temp_y, hr, ndcg, psp = evaluate_batch_partial_jit( pos_score, neg_score, test_pos_items, np.array(item_propensity), np.array(topk) ) for at_k, k in enumerate(topk): metrics['HR@{}'.format(k)] += hr[at_k] metrics['NDCG@{}'.format(k)] += ndcg[at_k] metrics['PSP@{}'.format(k)] += psp[at_k] return metrics, temp_preds.tolist(), temp_y.tolist() @jit('Tuple((float32[:], float32[:], float32[:], float32[:], float32[:]))(float32[:,:], float32[:,:], int64[:,:], float64[:], int64[:])') def evaluate_batch_partial_jit(pos_score, neg_score, test_pos_items, item_propensity, topk): temp_preds = np.zeros( ((pos_score.shape[0] * pos_score.shape[1]) + (neg_score.shape[0] * neg_score.shape[1])), dtype = np.float32 ) temp_y = np.zeros(temp_preds.shape, dtype = np.float32) at_preds = 0 hr_arr = np.zeros((len(topk)), dtype = np.float32) ndcg_arr = np.zeros((len(topk)), dtype = np.float32) psp_arr = np.zeros((len(topk)), dtype = np.float32) for b in range(len(pos_score)): pos, neg = pos_score[b, :], neg_score[b, :] # pos will be padded, un-pad it last_index = len(pos) - 1 while last_index > 0 and pos[last_index] == pos[last_index - 1]: last_index -= 1 pos = pos[:last_index + 1] # Add to AUC temp_preds[at_preds:at_preds+len(pos)] = pos temp_y[at_preds:at_preds+len(pos)] = 1 at_preds += len(pos) temp_preds[at_preds:at_preds+len(neg)] = neg temp_y[at_preds:at_preds+len(neg)] = 0 at_preds += len(neg) # get rank of all elements in pos temp_ranks = np.argsort(- np.concatenate((pos, neg))) # To maintain order pos_ranks = np.zeros(len(pos)) for at, r in enumerate(temp_ranks): if r < len(pos): pos_ranks[r] = at + 1 test_positive_sorted_psp = sorted([ item_propensity[x] for x in test_pos_items[b] ])[::-1] for at_k, k in enumerate(topk): num_pos = float(len(pos)) hr_arr[at_k] += np.sum(pos_ranks <= k) / float(min(num_pos, k)) dcg, idcg, psp, max_psp = 0.0, 0.0, 0.0, 0.0 for at, rank in enumerate(pos_ranks): if rank <= k: dcg += 1.0 / np.log2(rank + 1) # 1-based indexing psp += item_propensity[test_pos_items[b][at]] / float(min(num_pos, k)) idcg += 1.0 / np.log2(at + 2) max_psp += test_positive_sorted_psp[at] ndcg_arr[at_k] += dcg / idcg psp_arr[at_k] += psp / max_psp return temp_preds[:at_preds], temp_y[:at_preds], hr_arr, ndcg_arr, psp_arr @jit(float64(float64[:], float64[:])) def fast_auc(y_true, y_prob): y_true = y_true[np.argsort(y_prob)] nfalse, auc = 0, 0 for i in range(len(y_true)): nfalse += (1 - y_true[i]) auc += y_true[i] * nfalse return auc / (nfalse * (len(y_true) - nfalse))
7,384
38.704301
137
py
sampling_cf
sampling_cf-main/svp_handler.py
import numpy as np from collections import defaultdict from main import main_pytorch from data_path_constants import get_svp_log_file_path, get_svp_model_file_path class SVPHandler: def __init__(self, model_type, loss_type, hyper_params): hyper_params['model_type'] = model_type hyper_params['task'] = loss_type hyper_params['num_train_negs'] = 1 hyper_params['num_test_negs'] = 100 hyper_params['latent_size'] = 10 hyper_params['dropout'] = 0.3 hyper_params['weight_decay'] = float(1e-6) hyper_params['lr'] = 0.006 hyper_params['epochs'] = 50 hyper_params['validate_every'] = 5000 hyper_params['batch_size'] = 1024 self.hyper_params = hyper_params self.hyper_params['log_file'] = self.log_file self.hyper_params['model_path'] = self.model_file self.train_model() def train_model(self): _, self.forgetted_count = main_pytorch(self.hyper_params, track_events = True, eval_full = False) def forgetting_events(self, percent, data, index): # Keep those points which have the maximum forgetted count # => Remove those points which have the minimum forgetted count index_map = [] for at, i in enumerate(index): if i == 0: index_map.append(at) split_point = int(float(len(self.forgetted_count)) * (float(percent) / 100.0)) order = np.argsort(self.forgetted_count) order = list(map(lambda x: index_map[x], order)) remove_indices = order[:split_point] # If greedy for i in remove_indices: index[i] = -1 # Remove return index def forgetting_events_user(self, percent, data, index): # Keep those users which have the maximum forgetted count # Remove those users which have the minimum forgetted count index_map, user_map, hist, at, total = [], [], {}, 0, 0 for u in range(len(data)): for i, r, t in data[u]: if index[at] == 0: index_map.append(at) user_map.append(u) if u not in hist: hist[u] = 0 hist[u] += 1 total += 1 at += 1 user_forgetted_count = defaultdict(list) for train_at, cnt in enumerate(self.forgetted_count): user_forgetted_count[user_map[train_at]].append(cnt) user_forgetted_count = sorted(list(dict(user_forgetted_count).items()), key = lambda x: np.mean(x[1])) interactions_to_remove, removed, users_to_remove = total * (float(percent) / 100.0), 0, set() for u, _ in user_forgetted_count: if removed >= interactions_to_remove: break users_to_remove.add(u) removed += hist[u] for train_at in range(len(user_map)): if user_map[train_at] in users_to_remove: index[index_map[train_at]] = -1 return index def compute_freq(self, data, index, freq_type): freq, at = defaultdict(int), 0 for u in range(len(data)): for i, r, t in data[u]: if index[at] == 0: to_count = [ u, i ][freq_type] freq[to_count] += 1 at += 1 valid_users = list(freq.keys()) return list(map(lambda x: freq[x], valid_users)), dict(zip(valid_users, list(range(len(freq))))) def compute_prop(self, freq_vector, num_instances, A = 0.55, B = 1.5): C = (np.log(num_instances)-1)*np.power(B+1, A) wts = 1.0 + C*np.power(np.array(freq_vector)+B, -A) return np.ravel(wts) def forgetting_events_propensity(self, percent, data, index, pooling_method = 'max'): # Keep those points which have the maximum forgetted count # Remove those points which have the minimum forgetted count num_interactions = len(self.forgetted_count) user_freq, user_map = self.compute_freq(data, index, 0) user_propensity_vector = self.compute_prop(user_freq, num_interactions) item_freq, item_map = self.compute_freq(data, index, 1) item_propensity_vector = self.compute_prop(item_freq, num_interactions) interaction_propensity, at = [], 0 freq, at = defaultdict(int), 0 def pool(prop_u, prop_i): if pooling_method == 'sum': return prop_u + prop_i elif pooling_method == 'max': return max(prop_u, prop_i) for u in range(len(data)): for i, r, t in data[u]: if index[at] == 0: interaction_propensity.append( pool(user_propensity_vector[user_map[u]], item_propensity_vector[item_map[i]]) ) at += 1 assert len(interaction_propensity) == num_interactions # interaction_propensity actually estimates the `inverse` propensity, hence multiply updated_count = np.array(self.forgetted_count) * np.array(interaction_propensity) index_map = [] for at, i in enumerate(index): if i == 0: index_map.append(at) split_point = int(float(len(updated_count)) * (float(percent) / 100.0)) order = np.argsort(updated_count) order = list(map(lambda x: index_map[x], order)) remove_indices = order[:split_point] # If greedy for i in remove_indices: index[i] = -1 # Remove return index def forgetting_events_user_propensity(self, percent, data, index): # Keep those users which have the maximum forgetted count # Keep those users which have the maximum propensity --> minimum frequency # Remove those users which have the minimum forgetted count num_interactions = len(self.forgetted_count) user_freq, user_index_map = self.compute_freq(data, index, 0) user_propensity_vector = self.compute_prop(user_freq, num_interactions) index_map, user_map, hist, at, total = [], [], {}, 0, 0 for u in range(len(data)): for i, r, t in data[u]: if index[at] == 0: index_map.append(at) user_map.append(u) if u not in hist: hist[u] = 0 hist[u] += 1 total += 1 at += 1 user_forgetted_count = defaultdict(list) for train_at, cnt in enumerate(self.forgetted_count): u = user_map[train_at] user_forgetted_count[u].append(cnt * user_propensity_vector[user_index_map[u]]) user_forgetted_count = sorted(list(dict(user_forgetted_count).items()), key = lambda x: np.mean(x[1])) interactions_to_remove, removed, users_to_remove = total * (float(percent) / 100.0), 0, set() for u, _ in user_forgetted_count: if removed >= interactions_to_remove: break users_to_remove.add(u) removed += hist[u] for train_at in range(len(user_map)): if user_map[train_at] in users_to_remove: index[index_map[train_at]] = -1 return index @property def model_file(self): return get_svp_model_file_path(self.hyper_params) @property def log_file(self): return get_svp_log_file_path(self.hyper_params)
7,281
40.375
110
py
sampling_cf
sampling_cf-main/data_genie/data_genie_loss.py
import torch import torch.nn as nn import torch.nn.functional as F class PointwiseLoss(nn.Module): def __init__(self): super(PointwiseLoss, self).__init__() def forward(self, output, y, return_mean = True): loss = torch.pow(output - y, 2) if return_mean: return torch.mean(loss) return loss class PairwiseLoss(nn.Module): def __init__(self): super(PairwiseLoss, self).__init__() def forward(self, pos_output, neg_output, return_mean = True): loss = -F.logsigmoid(pos_output - neg_output) if return_mean: return torch.mean(loss) return loss
560
27.05
63
py
sampling_cf
sampling_cf-main/data_genie/data_genie_trainers.py
import time import torch import numpy as np from tqdm import tqdm from collections import defaultdict from sklearn.feature_selection import RFE from sklearn.metrics import roc_auc_score from xgboost import XGBClassifier, XGBRegressor from torch.utils.tensorboard import SummaryWriter from sklearn.linear_model import Ridge, LogisticRegression from data_genie.data_genie_config import * from data_genie.data_genie_model import PointwiseDataGenie from data_genie.data_genie_utils import LOGS_BASE, TENSORBOARD_BASE, MODELS_BASE from torch_utils import xavier_init from utils import log_end_epoch, INF def get_metrics(output, y, pointwise): output = np.array(output).reshape([ len(output) // len(all_samplers), len(all_samplers) ]) y = np.array(y).reshape([ len(y) // len(all_samplers), len(all_samplers) ]) metrics = defaultdict(list) for i in range(len(output)): pred, true = np.array(output[i]), np.array(y[i]) keep_indices = [] for j in range(len(true)): if true[j] >= -1 and true[j] <= 1: keep_indices.append(j) keep_indices = np.array(keep_indices) if len(keep_indices) == 0: continue pred, true = pred[keep_indices], true[keep_indices] if pointwise: metrics['MSE'] += ((pred - true) ** 2).tolist() if len(keep_indices) < 5: continue metrics['P@1'].append(float(true[np.argmax(pred)] == np.max(true))) # Since there can be many with same kendall's tau metrics['Rand P@1'].append(float(true[np.random.randint(0, len(true))] == np.max(true))) return { k: round(100.0 * np.mean(metrics[k]), 2) if k != 'MSE' else round(np.mean(metrics[k]), 4) for k in metrics } def validate(model, data, which_data, pointwise): model.model.eval() output, y = [], [] with torch.no_grad(): for complete, subset, task, metric, tau in data.test_iter(which_data): output_batch = model.model(complete, subset, task, metric) if pointwise: output_batch = torch.tanh(output_batch) output += output_batch.cpu().numpy().tolist() y += tau return get_metrics(output, y, pointwise) def train_one_epoch(model, data): model.model.train() metric, total = 0.0, 0.0 for data_batch in data.train_iter(): # Empty the gradients model.model.zero_grad() model.optimizer.zero_grad() # Forward pass loss, temp_metric, temp_total = model.step(data_batch) metric += temp_metric ; total += temp_total # Backward pass loss.backward() model.optimizer.step() return model.get_metric_dict(metric, total) def train_pytorch(data, Analyzer, feats_to_keep, lr, wd, dim, dropout, embedding_type, graph_dim, gcn_layers, EPOCHS, VALIDATE_EVERY): common_path = "{}_{}_dim_{}_drop_{}_lr_{}_wd_{}".format( "pointwise" if Analyzer == PointwiseDataGenie else "pairwise", embedding_type, dim, dropout, lr, wd ) if feats_to_keep is not None: common_path += "_feats_" + "-".join(map(str, feats_to_keep)) if embedding_type != "handcrafted": common_path += "_graph_dim_{}_gcn_layers_{}".format(graph_dim, gcn_layers) writer = SummaryWriter(log_dir = TENSORBOARD_BASE + common_path) model = Analyzer({ 'lr': lr, 'weight_decay': float(wd), 'dim': dim, 'dropout': dropout, 'feats': 5 + (len(feats_to_keep) * NUM_SAMPLES) if feats_to_keep is not None else None, 'log_file': LOGS_BASE + '{}.txt'.format(common_path), 'model_file': MODELS_BASE + '{}.pt'.format(common_path), 'embedding_type': embedding_type, 'graph_dim': graph_dim, 'gcn_layers': gcn_layers }, writer, xavier_init) # Train model start_time, best_metric = time.time(), -INF avg_p1 = 0.0 for epoch in tqdm(range(1, EPOCHS + 1)): epoch_start_time = time.time() train_metrics = train_one_epoch(model, data) for m in train_metrics: writer.add_scalar('Train/' + m, train_metrics[m], epoch) log_end_epoch(model.hyper_params, train_metrics, epoch, time.time() - epoch_start_time, metrics_on = '(TRAIN)', dont_print = True) if epoch % VALIDATE_EVERY == 0: val_metrics = validate(model, data, data.val, pointwise = Analyzer == PointwiseDataGenie) for m in val_metrics: writer.add_scalar('Validation/' + m, val_metrics[m], epoch) log_end_epoch(model.hyper_params, val_metrics, epoch, time.time() - epoch_start_time, metrics_on = '(VAL)', dont_print = True) test_metrics = validate(model, data, data.test, pointwise = Analyzer == PointwiseDataGenie) for m in test_metrics: writer.add_scalar('Test/' + m, test_metrics[m], epoch) log_end_epoch(model.hyper_params, test_metrics, epoch, time.time() - epoch_start_time, metrics_on = '(TEST)', dont_print = True) avg_p1 += test_metrics['Rand P@1'] if test_metrics["P@1"] > best_metric: model.save() ; best_metric = test_metrics["P@1"] model.load() test_metrics = validate(model, data, data.test, pointwise = Analyzer == PointwiseDataGenie) test_metrics['Rand P@1'] = round(avg_p1 / float(EPOCHS), 2) for m in test_metrics: writer.add_scalar('Test/' + m, test_metrics[m], EPOCHS + 1) log_end_epoch(model.hyper_params, test_metrics, "final", time.time() - start_time, metrics_on = '(TEST)') def train_linear_regression(data, embedding_type, feats_to_keep, C): start_time = time.time() log_file = LOGS_BASE + "linear_regression_{}_C_{}".format(embedding_type, C) if feats_to_keep is not None: log_file += "_feats_" + "-".join(map(str, feats_to_keep)) log_file += '.txt' x, y = data.sklearn_regression_data ########## Initially # model = Ridge(alpha = C).fit(x, y) ########## Backward selection model = RFE(Ridge(alpha = C, normalize = True), n_features_to_select = 10, step = 1).fit(x, y) train_mse = round(np.mean((y - model.predict(x)) ** 2), 4) train_var = round(np.var(y), 4) output, y = [], [] for complete, subset, task, metric, tau in data.test_iter(data.test): output += model.predict( data.sklearn_regression_feature(complete, subset, task.unsqueeze(-1), metric.unsqueeze(-1)) ).tolist() y += tau test_metrics = get_metrics(output, y, pointwise = True) test_metrics['Train MSE'] = train_mse test_metrics['Train Var'] = train_var log_end_epoch({ 'log_file': log_file }, test_metrics, "final", time.time() - start_time, metrics_on = '(TEST)') def train_logistic_regression(data, embedding_type, feats_to_keep, C): start_time = time.time() log_file = LOGS_BASE + "logistic_regression_{}_C_{}".format(embedding_type, C) if feats_to_keep is not None: log_file += "_feats_" + "-".join(map(str, feats_to_keep)) log_file += '.txt' x, y = data.sklearn_bce_data model = LogisticRegression(C = C, max_iter = 3000).fit(x, y) train_auc = round(roc_auc_score(y, model.predict_proba(x)[:, 1]), 4) output, y = [], [] for complete, subset, task, metric, tau in data.test_iter(data.test): output += model.predict_proba( data.sklearn_bce_feature(complete, subset, subset, task.unsqueeze(-1), metric.unsqueeze(-1)) )[:, 1].tolist() y += tau test_metrics = get_metrics(output, y, pointwise = False) test_metrics['Train AUC'] = train_auc log_end_epoch({ 'log_file': log_file }, test_metrics, "final", time.time() - start_time, metrics_on = '(TEST)') def train_xgboost_regression(data, embedding_type, feats_to_keep, max_depth): start_time = time.time() common_path = "xgboost_regression_{}_depth_{}".format(embedding_type, max_depth) if feats_to_keep is not None: common_path += "_feats_" + "-".join(map(str, feats_to_keep)) x, y = data.sklearn_regression_data model = XGBRegressor(max_depth = max_depth).fit(x, y) train_mse = round(np.mean((y - model.predict(x)) ** 2), 4) train_var = round(np.var(y), 4) output, y = [], [] for complete, subset, task, metric, tau in data.test_iter(data.test): output += model.predict( data.sklearn_regression_feature(complete, subset, task.unsqueeze(-1), metric.unsqueeze(-1)) ).tolist() y += tau test_metrics = get_metrics(output, y, pointwise = True) test_metrics['Train MSE'] = train_mse test_metrics['Train Var'] = train_var log_end_epoch({ 'log_file': LOGS_BASE + "{}.txt".format(common_path) }, test_metrics, "final", time.time() - start_time, metrics_on = '(TEST)') def train_xgboost_bce(data, embedding_type, feats_to_keep, max_depth): start_time = time.time() common_path = "xgboost_bce_{}_depth_{}".format(embedding_type, max_depth) if feats_to_keep is not None: common_path += "_feats_" + "-".join(map(str, feats_to_keep)) x, y = data.sklearn_bce_data model = XGBClassifier(max_depth = max_depth, use_label_encoder=False, eval_metric = "logloss").fit(x, y) train_auc = round(roc_auc_score(y, model.predict(x)), 4) output, y = [], [] for complete, subset, task, metric, tau in data.test_iter(data.test): output += model.predict( data.sklearn_bce_feature(complete, subset, subset, task.unsqueeze(-1), metric.unsqueeze(-1)) ).tolist() y += tau test_metrics = get_metrics(output, y, pointwise = False) test_metrics['Train AUC'] = train_auc log_end_epoch({ 'log_file': LOGS_BASE + "{}.txt".format(common_path) }, test_metrics, "final", time.time() - start_time, metrics_on = '(TEST)')
8,933
38.883929
144
py
sampling_cf
sampling_cf-main/data_genie/get_embeddings.py
import gc import os import dgl import snap import torch import numpy as np from tqdm import tqdm import networkx as nx from collections import defaultdict from data_genie.data_genie_config import * from data_genie.data_genie_utils import save_numpy, load_numpy from data_genie.data_genie_utils import EMBEDDINGS_PATH_GCN, EMBEDDINGS_PATH_HANDCRAFTED, INFOGRAPH_MODEL_PATH from load_data import DataHolder from data_path_constants import get_data_path, get_index_path from data_genie.InfoGraph.infograph_model import InfoGraph from data_genie.InfoGraph.infograph_dataset import SyntheticDataset from data_genie.InfoGraph.train_infograph import train_infograph, argument def get_embeddings(dataset, embedding_type): return { 'unsupervise': get_embeddings_gcn, 'handcrafted': get_embeddings_handcrafted, }[embedding_type[:11]](dataset, embedding_type) def get_embeddings_gcn(dataset, model_file): # Extract which InfoGraph model we are looking for from the specified `model_file` dim, layers = None, None splitted = model_file.split("_") for i, word in enumerate(splitted): if word == "dim": dim = int(splitted[i+1]) if word == "layers": layers = int(splitted[i+1]) PATH = EMBEDDINGS_PATH_GCN(dataset, dim, layers) if not os.path.exists(PATH + ".npy"): prep_gcn_embeddings(dim, layers) return load_numpy(PATH) def prep_gcn_embeddings(dim, layers): # Step-1: Check if the unsupervised GCN Model has been trained? if not os.path.exists(INFOGRAPH_MODEL_PATH(dim, layers)): print("Specified InfoGraph configuration not trained yet, training now..") infograph_args = argument() infograph_args.n_layers = layers infograph_args.hid_dim = dim train_infograph(infograph_args) print("Loading best InfoGraph model..") model = InfoGraph(dim, layers) model.load_state_dict(torch.load(INFOGRAPH_MODEL_PATH(dim, layers))) model.eval() print("Loading data..") # Keep the dimension of node features fixed to a reasonable value. # If you want to change this, please also change at `InfoGraph/infograph_model.py` line 134 dataset = SyntheticDataset(feature_dimension = 32) graphs, _ = map(list, zip(*dataset)) num_graphs = len(graphs) # Embeddings print("Getting GCN embeddings..") print(len(graphs)) BSZ = 32 # Batch-size for predicting dataset embeddings emb = np.zeros([ len(graphs), (dim * layers) + 5 ]) for b in tqdm(range(0, len(graphs), BSZ)): wholegraph = dgl.batch(graphs[b:b+BSZ]) wholegraph.ndata['attr'] = wholegraph.ndata['attr'].to(torch.float32) emb[b:b+BSZ] = np.hstack([ model.get_embedding(wholegraph).cpu().numpy(), np.array(dataset.basic_data_stats_features[b:b+BSZ], dtype = np.float32) ]) del graphs, dataset, model gc.collect() # NOTE: Since we'll anyways be training on ALL datasets, we'll prep embeddings for ALL datasets at once at = 0 for d in datasets: final = np.zeros([ total_embeddings, (dim * layers) + 5 ]) for task, metrics in scenarios: final[get_embedding_id(task, 'complete_data', 0)] = emb[at] ; at += 1 for sampling_percent in percent_rns_options: for sampling in all_samplers: final[get_embedding_id(task, sampling, sampling_percent)] = emb[at] ; at += 1 save_numpy(EMBEDDINGS_PATH_GCN(d, dim, layers), final) assert at == num_graphs def get_embeddings_handcrafted(dataset, model_file): PATH = EMBEDDINGS_PATH_HANDCRAFTED(dataset) if not os.path.exists(PATH + ".npy"): prep_handcrafted_embeddings(dataset, PATH) return load_numpy(PATH) def prep_handcrafted_embeddings(dataset, save_path): final = np.zeros([ total_embeddings, NUM_FEAUTRES ]) print("Getting handcrafted embeddings..") loop = tqdm(total = len(all_samplers) * len(scenarios) * len(percent_rns_options)) for task, metrics in scenarios: final[get_embedding_id(task, 'complete_data', 0)] = get_single_feature(dataset, task, 'complete_data', 0) for sampling_percent in percent_rns_options: for sampling in sampling_kinds: final[get_embedding_id(task, sampling, sampling_percent)] = get_single_feature(dataset, task, sampling, sampling_percent) loop.update(1) for svp_method in svp_methods: for sampling in sampling_svp: name = "svp_{}_{}".format(svp_method, sampling) final[get_embedding_id(task, name, sampling_percent)] = get_single_feature( dataset, task, "svp_{}".format(svp_method), sampling_percent, svp = sampling ) loop.update(1) loop.close() save_numpy(save_path, final) def get_single_feature(dataset, task, sampling, sampling_percent, svp = None): hyper_params = { 'dataset': dataset, 'task': task, 'sampling': sampling, 'sampling_percent': sampling_percent, 'sampling_svp': svp } data = DataHolder(get_data_path(hyper_params), get_index_path(hyper_params)) MIL = float(1e6) return list([ float(data.num_users) / MIL, float(data.num_items) / MIL ] + \ [ float(data.num_train_interactions) / MIL ] + \ [ float(data.num_val_interactions) / MIL ] + \ [ float(data.num_test_interactions) / MIL ] + \ degree_distribution(data) + \ graph_characteristics(data)) def degree_distribution(data): user_degree, item_degree = defaultdict(int), defaultdict(int) for at, (u, i, r) in enumerate(data.data): if data.index[at] != -1: user_degree[u] += 1 item_degree[i] += 1 user_degree = sorted([ user_degree[u] for u in user_degree ]) item_degree = sorted([ item_degree[i] for i in item_degree ]) return sample_distribution(user_degree) + sample_distribution(item_degree) def graph_characteristics(data): g, snap_g = create_nx_graph(data) return \ sample_distribution([ g.degree(n) for n in g.nodes() ]) + \ sample_distribution([ len(c) for c in nx.connected_components(g) ]) + \ sample_distribution(hop_plot(snap_g)) + \ sample_distribution(snap_g.GetEigVals(100), reverse = True) + \ sample_distribution(clustering_coefficient(snap_g)) def hop_plot(g): g.PlotHops("temp", "Random Graph hop plot", False, 128) f = open("hop.temp.tab", 'r') ; lines = f.readlines() ; f.close() for f in [ "hop.temp.png", "hop.temp.tab", "hop.temp.plt" ]: os.remove(f) return list(map(lambda x: float(x.strip().split("\t")[1]), lines[4:])) def clustering_coefficient(g): Cf, CfVec = g.GetClustCf(True, -1) return [ pair.GetVal2() for pair in CfVec ] def sample_distribution(distribution, reverse = False): distribution = sorted(distribution, reverse = reverse) ; n = len(distribution) to_pick = np.round(np.linspace(0, len(distribution) - 1, NUM_SAMPLES)).astype(int) return [ distribution[i] for i in to_pick ] def create_nx_graph(data): g = nx.Graph() snap_g = snap.TUNGraph.New() # Add nodes & edges user_map, item_map, node_num = {}, {}, 0 for at, (u, i, r) in enumerate(data.data): if data.index[at] != -1: if u not in user_map: user_map[u] = node_num g.add_node(node_num) snap_g.AddNode(node_num) node_num += 1 if i not in item_map: item_map[i] = node_num g.add_node(node_num) snap_g.AddNode(node_num) node_num += 1 g.add_edge(user_map[u], item_map[i]) snap_g.AddEdge(user_map[u], item_map[i]) assert node_num == g.number_of_nodes() return g, snap_g
7,131
33.960784
125
py
sampling_cf
sampling_cf-main/data_genie/data_genie_data.py
import torch import numpy as np from torch_utils import LongTensor, FloatTensor, is_cuda_available from data_genie.data_genie_config import * from data_genie.get_data import get_data_pointwise, get_data_pairwise from data_genie.get_embeddings import get_embeddings from data_genie.InfoGraph.infograph_dataset import SyntheticDataset # NOTE: Helper functions def sel(d, i): return list(map(lambda x: x[i], d)) def sel2(d, i): return list(map(lambda x: list(map(lambda y: y[i], x)), d)) # Main class for manipulating and iterating over training data for DataGenie class OracleData: def __init__(self, datasets, feats_to_keep, embedding_type, bsz, pointwise): self.feats_to_keep, self.bsz, self.embedding_type = feats_to_keep, bsz, embedding_type get_final_data = get_data_pointwise if pointwise else get_data_pairwise # Get all embeddings self.embeddings = {} for dataset in datasets: self.embeddings[dataset] = get_embeddings(dataset, embedding_type) self.embeddings = self.normalize_embeddings(self.embeddings) # Create data self.dataset_indices, train_data, val_data, test_data = [], None, None, None for dataset in datasets: self.dataset_indices.append(0 if train_data is None else len(train_data)) this_dataset = self.join(get_final_data(dataset), self.embeddings[dataset], pointwise) if train_data is None: train_data, val_data, test_data = this_dataset else: for at, container in enumerate([ train_data, val_data, test_data ]): for i in range(5): container[i] += this_dataset[at][i] print("\n{} data:".format("Pointwise" if pointwise else "Pairwise")) print("# of training points:", len(train_data[0])) print("# of validation points:", len(val_data[0]) // len(all_samplers)) print("# of testing points:", len(test_data[0]) // len(all_samplers)) # NOTE: Each entry inside `train_data`, `val_data`, `test_data` is a 5-tuple: # 0. Full-data embedding ID # 1. Subset embedding ID # 2. Task ID # 3. Metric ID # 4. Y: Kendall's Tau between the ranked list of recommendation algorithms \ # trained on full vs. the data-subset (what we want to predict w/ Data-Genie) # Move data to GPU self.val = [ FloatTensor(val_data[0]), FloatTensor(val_data[1]), LongTensor(val_data[2]), LongTensor(val_data[3]), val_data[4] ] self.test = [ FloatTensor(test_data[0]), FloatTensor(test_data[1]), LongTensor(test_data[2]), LongTensor(test_data[3]), test_data[4] ] if pointwise: self.train = [ FloatTensor(train_data[0]), FloatTensor(train_data[1]), LongTensor(train_data[2]), LongTensor(train_data[3]), FloatTensor(train_data[4]) ] else: self.train = [ FloatTensor(train_data[0]), FloatTensor(train_data[1]), FloatTensor(train_data[2]), LongTensor(train_data[3]), LongTensor(train_data[4]) ] def normalize_embeddings(self, embeddings): combined_embeddings = np.concatenate(list(embeddings.values()), axis=0) # GCN-embeddings if self.feats_to_keep is None: pass # Handcrafted embeddings else: mean, std = self.get_mean_std(combined_embeddings) for dataset in embeddings: embeddings[dataset] = self.select_and_norm(embeddings[dataset], mean, std) return embeddings def join(self, all_data, embeddings, pointwise): train, val, test = all_data def join(data, indices): for i in indices: data[i] = list(map(lambda x: embeddings[x], data[i])) return data test, val = map(lambda x: join(x, [ 0, 1 ]), [ test, val ]) if pointwise: train = join(train, [ 0, 1 ]) else: train = join(train, [ 0, 1, 2 ]) return [ train, val, test ] def shuffle_train(self): num_train_interactions = len(self.train[0]) rand_indices = np.arange(num_train_interactions) ; np.random.shuffle(rand_indices) rand_indices_tensor = LongTensor(rand_indices) for i in range(len(self.train)): self.train[i] = self.train[i][rand_indices_tensor] return self.train # Convert task/metric index to one-hot vector def one_hot(self, index, total): if index.shape[0] > 1: ret = torch.zeros(index.shape[0], total) if is_cuda_available: ret = ret.cuda() ret.scatter_(1, index, 1.0) return ret ret = torch.zeros(total) if is_cuda_available: ret = ret.cuda() ret.scatter_(0, index, 1.0) return ret def sklearn_regression_feature(self, complete, subset, task, metric): return torch.cat([ complete, subset, self.one_hot(task, 3), self.one_hot(metric, 4) ], axis = -1).cpu().numpy() def sklearn_bce_feature(self, complete, pos, neg, task, metric): return torch.cat([ complete, pos, neg, self.one_hot(task, 3), self.one_hot(metric, 4) ], axis = -1).cpu().numpy() @property def sklearn_regression_data(self): shuffled_train = self.shuffle_train() x, y = [], [] for b in range(shuffled_train[0].shape[0]): x.append(self.sklearn_regression_feature( shuffled_train[0][b], shuffled_train[1][b], shuffled_train[2][b].unsqueeze(-1), shuffled_train[3][b].unsqueeze(-1) )) y.append(float(shuffled_train[4][b])) return np.asarray(x), np.asarray(y) @property def sklearn_bce_data(self): shuffled_train = self.shuffle_train() x, y = [], [] for b in range(shuffled_train[0].shape[0]): pos_or_neg = np.random.uniform() if pos_or_neg > 0.5: x.append(self.sklearn_bce_feature( shuffled_train[0][b], shuffled_train[1][b], shuffled_train[2][b], shuffled_train[3][b].unsqueeze(-1), shuffled_train[4][b].unsqueeze(-1) )) y.append(1) else: x.append(self.sklearn_bce_feature( shuffled_train[0][b], shuffled_train[2][b], shuffled_train[1][b], shuffled_train[3][b].unsqueeze(-1), shuffled_train[4][b].unsqueeze(-1) )) y.append(0) return np.asarray(x), np.asarray(y) def train_iter(self): num_train_interactions = len(self.train[0]) shuffled_train = self.shuffle_train() for b in range(0, num_train_interactions, self.bsz): l, r = b, b + self.bsz yield shuffled_train[0][l:r], shuffled_train[1][l:r], shuffled_train[2][l:r], \ shuffled_train[3][l:r], shuffled_train[4][l:r] def test_iter(self, data): for b in range(0, len(data[0]), self.bsz): l, r = b, b + self.bsz yield data[0][l:r], data[1][l:r], data[2][l:r], \ data[3][l:r], data[4][l:r] def select_and_norm(self, data, mean, std): data = np.asarray(data) assert data.shape[-1] == NUM_FEAUTRES indices = [] for f in self.feats_to_keep: indices += list(range(NUM_SAMPLES * f, NUM_SAMPLES * (f+1))) mean = mean[np.array(indices)] ; std = std[np.array(indices)] indices = list(range(5)) + list(map(lambda x: x+5, indices)) indices = np.array(indices) if len(data.shape) == 2: data = data[:, indices] data[:, 5:] -= mean data[:, 5:] /= std else: data = data[:, :, indices] data[:, :, 5:] -= mean data[:, :, 5:] /= std return data def get_mean_std(self, combined_data): assert combined_data.shape[1] == NUM_FEAUTRES temp_data = combined_data[:, 5:NUM_FEAUTRES] std = np.array(list(map(lambda x: max(x, float(1e-6)), np.std(temp_data, axis = 0)))) return np.mean(temp_data, axis = 0), std
7,077
33.526829
136
py
sampling_cf
sampling_cf-main/data_genie/data_genie_model.py
import dgl import torch import torch.nn as nn from torch_utils import is_cuda_available from data_genie.data_genie_loss import PointwiseLoss, PairwiseLoss # NOTE: Below two are the training classes for data-genie: pointwise/pairwise class PointwiseDataGenie: def __init__(self, hyper_params, writer, xavier_init): self.hyper_params = hyper_params self.writer = writer self.model = DataGenie(hyper_params) if is_cuda_available: self.model.cuda() xavier_init(self.model) ; print(self.model) self.criterion = PointwiseLoss() self.optimizer = torch.optim.Adam( self.model.parameters(), lr=hyper_params['lr'], betas=(0.9, 0.98), weight_decay=hyper_params['weight_decay'] ) def step(self, data_batch): complete, subset, task, metric, y = data_batch # Forward pass output = torch.tanh(self.model(complete, subset, task, metric)) # Since in range [ -1, 1 ] loss = self.criterion(output, y, return_mean = True) return torch.mean(loss), float(torch.sum(loss)), float(output.shape[0]) def get_metric_dict(self, mse, total): return { 'MSE': round(mse / total, 4) } def save(self): torch.save(self.model.state_dict(), self.hyper_params['model_file']) def load(self): self.model.load_state_dict(torch.load(self.hyper_params['model_file'])) class PairwiseDataGenie: def __init__(self, hyper_params, writer, xavier_init): self.hyper_params = hyper_params self.writer = writer self.model = DataGenie(hyper_params) if is_cuda_available: self.model.cuda() xavier_init(self.model) ; print(self.model) self.criterion = PairwiseLoss() self.optimizer = torch.optim.Adam( self.model.parameters(), lr=hyper_params['lr'], betas=(0.9, 0.98), weight_decay=hyper_params['weight_decay'] ) def step(self, data_batch): complete, pos, neg, task, metric = data_batch # Forward pass pos_output = self.model(complete, pos, task, metric) neg_output = self.model(complete, neg, task, metric) loss = self.criterion(pos_output, neg_output, return_mean = True) # Metric auc = float(torch.sum(pos_output > neg_output)) return loss, auc, float(pos_output.shape[0]) def get_metric_dict(self, auc, total): return { 'AUC': round(auc / total, 4) } def save(self): torch.save(self.model.state_dict(), self.hyper_params['model_file']) def load(self): self.model.load_state_dict(torch.load(self.hyper_params['model_file'])) # NOTE: Below is the actual data-genie pytorch model class DataGenie(nn.Module): def __init__(self, hyper_params): super(DataGenie, self).__init__() self.feats = hyper_params['feats'] if self.feats is None: dim, layers = None, None splitted = hyper_params['embedding_type'].split("_") for i, word in enumerate(splitted): if word == "dim": dim = int(splitted[i+1]) if word == "layers": layers = int(splitted[i+1]) self.feats = (dim * layers) + 5 self.task_embedding = nn.Embedding(3, 3) self.metric_embedding = nn.Embedding(4, 4) self.final = nn.Sequential( nn.Linear((2 * self.feats) + 3 + 4, hyper_params['dim']), nn.Dropout(hyper_params['dropout']), nn.ReLU(), nn.Linear(hyper_params['dim'], 1), ) def forward(self, complete_feats, subset_feats, task, metric): return self.final(torch.cat([ complete_feats, subset_feats, self.task_embedding(task), self.metric_embedding(metric) ], axis = -1))[:, 0]
3,352
31.240385
92
py
sampling_cf
sampling_cf-main/data_genie/InfoGraph/infograph_dataset.py
from dgl import save_graphs, load_graphs from dgl.data import DGLDataset from tqdm import tqdm import numpy as np import networkx as nx import torch import dgl import os from load_data import DataHolder from data_path_constants import get_data_path, get_index_path from data_genie.data_genie_config import * from data_genie.data_genie_utils import save_numpy, load_numpy from data_genie.data_genie_utils import INFOGRAPH_CACHED_GRAPHS, INFOGRAPH_CACHED_DATA_STATS ''' This is a wrapper class which contains ALL the CF datasets (and their sampled subsets) that we want to train Data-Genie on ''' class SyntheticDataset(DGLDataset): def __init__(self, feature_dimension): self.feature_dimension = feature_dimension super().__init__(name='synthetic') def process(self): self.graphs = [] self.basic_data_stats_features = [] for dataset in datasets: single_dataset = SingleDataset(self.feature_dimension, dataset) self.graphs += single_dataset.graphs self.basic_data_stats_features += single_dataset.basic_data_stats_features self.orig_total = len(self.graphs) # Including None assert len(self.basic_data_stats_features) == len(self.graphs) # Remove very small subsets self.graphs = list(filter(lambda x: x is not None, self.graphs)) self.basic_data_stats_features = list(filter(lambda x: x is not None, self.basic_data_stats_features)) assert len(self.basic_data_stats_features) == len(self.graphs) def __getitem__(self, i): # NOTE: Return 0 as the graph-label # We don't care about the label since this is an unsupervised task return self.graphs[i], 0 def __len__(self): return len(self.graphs) class SingleDataset(DGLDataset): def __init__(self, feature_dimension, dataset): self.feature_dimension = feature_dimension self.dataset = dataset super().__init__(name='synthetic') # Randomly init all node features (no implicit representation of user/item nodes) def get_random_node_features(self, graph): graph.ndata['attr'] = torch.empty(graph.num_nodes(), self.feature_dimension) torch.nn.init.xavier_uniform_(graph.ndata['attr'], gain=torch.nn.init.calculate_gain('relu')) return graph # Make the user-item interaction graph of the given CF dataset def add_graph(self, hyper_params): data = DataHolder(get_data_path(hyper_params), get_index_path(hyper_params)) g = nx.Graph() # Add nodes & edges user_map, item_map, node_num = {}, {}, 0 for at, (u, i, r) in enumerate(data.data): if data.index[at] != -1: if u not in user_map: user_map[u] = node_num g.add_node(node_num) node_num += 1 if i not in item_map: item_map[i] = node_num g.add_node(node_num) node_num += 1 g.add_edge(user_map[u], item_map[i]) assert node_num == g.number_of_nodes() # If nodes are too less return None if node_num < 50: self.graphs.append(None) self.basic_data_stats_features.append(None) return None self.graphs.append(self.get_random_node_features(dgl.from_networkx(g))) MIL = float(1e6) self.basic_data_stats_features.append([ float(data.num_users) / MIL, float(data.num_items) / MIL, float(data.num_train_interactions) / MIL, float(data.num_val_interactions) / MIL, float(data.num_test_interactions) / MIL ]) def process(self): self.graphs = [] self.basic_data_stats_features = [] total_samplers = (len(svp_methods) * len(sampling_svp)) + len(sampling_kinds) loop = tqdm(total = len(scenarios) * (1 + (len(percent_rns_options) * total_samplers))) for task, metrics in scenarios: # Full dataset self.add_graph({ 'dataset': self.dataset, 'task': task, 'sampling': 'complete_data', }) ; loop.update(1) # Sub-sampled dataset for sampling_percent in percent_rns_options: for sampling in sampling_kinds: self.add_graph({ 'dataset': self.dataset, 'task': task, 'sampling': sampling, 'sampling_percent': sampling_percent, }) ; loop.update(1) for svp_method in svp_methods: for sampling in sampling_svp: self.add_graph({ 'dataset': self.dataset, 'task': task, 'sampling': "svp_{}".format(svp_method), 'sampling_svp': sampling, 'sampling_percent': sampling_percent, }) ; loop.update(1) loop.close() assert len(self.graphs) == len(self.basic_data_stats_features) def save(self): save_graphs(self.graph_path, self.graphs) save_numpy( self.basic_data_stats_features_path, np.array(self.basic_data_stats_features) ) def load(self): self.graphs, _ = load_graphs(self.graph_path) self.basic_data_stats_features = load_numpy(self.basic_data_stats_features_path).tolist() def has_cache(self): return os.path.exists(self.graph_path) @property def graph_path(self): return INFOGRAPH_CACHED_GRAPHS(self.dataset) @property def basic_data_stats_features_path(self): return INFOGRAPH_CACHED_DATA_STATS(self.dataset)
4,944
30.698718
104
py
sampling_cf
sampling_cf-main/data_genie/InfoGraph/infograph_model.py
''' Credit https://github.com/hengruizhang98/InfoGraph & https://github.com/fanyun-sun/InfoGraph ''' import torch as th import torch.nn as nn import torch.nn.functional as F from torch.nn import Sequential, ModuleList, Linear, ReLU, BatchNorm1d from dgl.nn import GINConv from dgl.nn.pytorch.glob import SumPooling, AvgPooling, SortPooling from data_genie.InfoGraph.infograph_utils import local_global_loss_ ''' Feedforward neural network''' class FeedforwardNetwork(nn.Module): ''' 3-layer feed-forward neural networks with jumping connections Parameters ----------- in_dim: int, Input feature size. hid_dim: int, Hidden feature size. Functions ----------- forward(feat): feat: Tensor, [N * D], input features ''' def __init__(self, in_dim, hid_dim): super(FeedforwardNetwork, self).__init__() self.block = Sequential( Linear(in_dim, hid_dim), ReLU(), Linear(hid_dim, hid_dim), ReLU(), Linear(hid_dim, hid_dim), ReLU() ) self.jump_con = Linear(in_dim, hid_dim) def forward(self, feat): block_out = self.block(feat) jump_out = self.jump_con(feat) out = block_out + jump_out return out ''' Unsupervised Setting ''' class GINEncoder(nn.Module): ''' Encoder based on dgl.nn.GINConv & dgl.nn.SortPooling Parameters ----------- in_dim: int, Input feature size. hid_dim: int, Hidden feature size. n_layer: int, number of GIN layers. Functions ----------- forward(graph, feat): graph: dgl.Graph, feat: Tensor, [N * D], node features ''' def __init__(self, in_dim, hid_dim, n_layer): super(GINEncoder, self).__init__() self.n_layer = n_layer self.convs = ModuleList() self.bns = ModuleList() for i in range(n_layer): if i == 0: n_in = in_dim else: n_in = hid_dim n_out = hid_dim block = Sequential( Linear(n_in, n_out), ReLU(), Linear(hid_dim, hid_dim) ) conv = GINConv(block, 'sum') bn = BatchNorm1d(hid_dim) self.convs.append(conv) self.bns.append(bn) # Pooling # self.pool = SumPooling() # self.pool = AvgPooling() self.pool = SortPooling(1) def forward(self, graph, feat): xs = [] x = feat for i in range(self.n_layer): x = F.relu(self.convs[i](graph, x)) x = self.bns[i](x) xs.append(x) local_emb = th.cat(xs, 1) # patch-level embedding global_emb = self.pool(graph, local_emb) # graph-level embedding return global_emb, local_emb class InfoGraph(nn.Module): r""" InfoGraph model for unsupervised setting Parameters ----------- in_dim: int Input feature size. hid_dim: int Hidden feature size. Functions ----------- forward(graph): graph: dgl.Graph """ def __init__(self, hid_dim, n_layer): super(InfoGraph, self).__init__() self.in_dim = 32 # Keep it fixed to a reasonable value for our experiments self.hid_dim = hid_dim self.n_layer = n_layer embedding_dim = hid_dim * n_layer self.encoder = GINEncoder(self.in_dim, hid_dim, n_layer) self.local_d = FeedforwardNetwork(embedding_dim, embedding_dim) # local discriminator (node-level) self.global_d = FeedforwardNetwork(embedding_dim, embedding_dim) # global discriminator (graph-level) # get_embedding function for evaluation the learned embeddings def get_embedding(self, graph): with th.no_grad(): feat = graph.ndata['attr'] global_emb, _ = self.encoder(graph, feat) return global_emb def forward(self, graph): feat = graph.ndata['attr'] graph_id = graph.ndata['graph_id'] global_emb, local_emb = self.encoder(graph, feat) global_h = self.global_d(global_emb) # global hidden representation local_h = self.local_d(local_emb) # local hidden representation measure = 'JSD' loss = local_global_loss_(local_h, global_h, graph_id, measure) return loss
4,413
25.590361
110
py
sampling_cf
sampling_cf-main/data_genie/InfoGraph/train_infograph.py
import dgl import time import argparse import torch as th from dgl.dataloading import GraphDataLoader from tqdm import tqdm from data_genie.data_genie_utils import INFOGRAPH_MODEL_PATH from data_genie.InfoGraph.infograph_model import InfoGraph from data_genie.InfoGraph.infograph_dataset import SyntheticDataset def argument(): parser = argparse.ArgumentParser(description='InfoGraph') # training params parser.add_argument('--gpu', type=int, default=-1, help='GPU index, default:-1, using CPU.') parser.add_argument('--epochs', type=int, default=20, help='Training epochs.') parser.add_argument('--batch_size', type=int, default=128, help='Training batch size.') parser.add_argument('--lr', type=float, default=0.01, help='Learning rate.') # model params parser.add_argument('--n_layers', type=int, default=2, help='Number of graph convolution layers before each pooling') parser.add_argument('--hid_dim', type=int, default=32, help='Hidden layer dimensionalities') args = parser.parse_args() # check cuda if args.gpu != -1 and th.cuda.is_available(): args.device = 'cuda:{}'.format(args.gpu) else: args.device = 'cpu' return args def collate(samples): ''' collate function for building graph dataloader''' graphs, labels = map(list, zip(*samples)) # generate batched graphs batched_graph = dgl.batch(graphs) n_nodes = batched_graph.num_nodes() # generate graph_id for each node within the batch graph_id = th.zeros(n_nodes).long() N = 0 id = 0 for graph in graphs: N_next = N + graph.num_nodes() graph_id[N:N_next] = id N = N_next id += 1 batched_graph.ndata['graph_id'] = graph_id return batched_graph, len(graphs) def train_infograph(args): # Step 1: Prepare graph data ===================================== # dataset = SyntheticDataset(feature_dimension = 32) print("Total # of graphs:", len(dataset.graphs), "/", dataset.orig_total, "\n") # creata dataloader for batch training dataloader = GraphDataLoader( dataset, batch_size=args.batch_size, collate_fn=collate, drop_last=False, shuffle=True ) # Step 2: Create model =================================================================== # print("DIMENSION:", args.hid_dim, "; GCN Layers:", args.n_layers) model = InfoGraph(args.hid_dim, args.n_layers) model = model.to(args.device) # Step 3: Create training components ===================================================== # optimizer = th.optim.Adam(model.parameters(), lr=args.lr) print('===== Before training ======') # Step 4: training epoches =============================================================== # best_loss = float(1e10) for epoch in range(1, args.epochs): loss_all = 0 model.train() start_time = time.time() for graph, n_graph in tqdm(dataloader): graph = graph.to(args.device) optimizer.zero_grad() loss = model(graph) loss.backward() optimizer.step() loss_all += loss.item() * n_graph mean_loss = loss_all / len(dataloader) print('Epoch {}, Loss {:.4f}, Time {:.4f}'.format(epoch, mean_loss, time.time() - start_time)) if mean_loss < best_loss: print("Saving...") th.save(model.state_dict(), INFOGRAPH_MODEL_PATH(args.hid_dim, args.n_layers)) best_loss = mean_loss if __name__ == '__main__': args = argument() ; print(args) train_infograph(args)
3,615
33.438095
121
py
sampling_cf
sampling_cf-main/data_genie/InfoGraph/infograph_utils.py
''' Credit: https://github.com/fanyun-sun/InfoGraph ''' import torch import torch as th import torch.nn.functional as F import math def local_global_loss_(l_enc, g_enc, graph_id, measure): num_graphs = g_enc.shape[0] num_nodes = l_enc.shape[0] device = g_enc.device pos_mask = th.zeros((num_nodes, num_graphs)).to(device) neg_mask = th.ones((num_nodes, num_graphs)).to(device) for nodeidx, graphidx in enumerate(graph_id): pos_mask[nodeidx][graphidx] = 1. neg_mask[nodeidx][graphidx] = 0. res = th.mm(l_enc, g_enc.t()) E_pos = get_positive_expectation(res * pos_mask, measure, average=False).sum() E_pos = E_pos / num_nodes E_neg = get_negative_expectation(res * neg_mask, measure, average=False).sum() E_neg = E_neg / (num_nodes * (num_graphs - 1)) return E_neg - E_pos def log_sum_exp(x, axis=None): """Log sum exp function Args: x: Input. axis: Axis over which to perform sum. Returns: torch.Tensor: log sum exp """ x_max = torch.max(x, axis)[0] y = torch.log((torch.exp(x - x_max)).sum(axis)) + x_max return y def raise_measure_error(measure): supported_measures = ['GAN', 'JSD', 'X2', 'KL', 'RKL', 'DV', 'H2', 'W1'] raise NotImplementedError( 'Measure `{}` not supported. Supported: {}'.format(measure, supported_measures)) def get_positive_expectation(p_samples, measure, average=True): """Computes the positive part of a divergence / difference. Args: p_samples: Positive samples. measure: Measure to compute for. average: Average the result over samples. Returns: torch.Tensor """ log_2 = math.log(2.) if measure == 'GAN': Ep = - F.softplus(-p_samples) elif measure == 'JSD': Ep = log_2 - F.softplus(- p_samples) elif measure == 'X2': Ep = p_samples ** 2 elif measure == 'KL': Ep = p_samples + 1. elif measure == 'RKL': Ep = -torch.exp(-p_samples) elif measure == 'DV': Ep = p_samples elif measure == 'H2': Ep = 1. - torch.exp(-p_samples) elif measure == 'W1': Ep = p_samples else: raise_measure_error(measure) if average: return Ep.mean() else: return Ep def get_negative_expectation(q_samples, measure, average=True): """Computes the negative part of a divergence / difference. Args: q_samples: Negative samples. measure: Measure to compute for. average: Average the result over samples. Returns: torch.Tensor """ log_2 = math.log(2.) if measure == 'GAN': Eq = F.softplus(-q_samples) + q_samples elif measure == 'JSD': Eq = F.softplus(-q_samples) + q_samples - log_2 elif measure == 'X2': Eq = -0.5 * ((torch.sqrt(q_samples ** 2) + 1.) ** 2) elif measure == 'KL': Eq = torch.exp(q_samples) elif measure == 'RKL': Eq = q_samples - 1. elif measure == 'DV': Eq = log_sum_exp(q_samples, 0) - math.log(q_samples.size(0)) elif measure == 'H2': Eq = torch.exp(q_samples) - 1. elif measure == 'W1': Eq = q_samples else: raise_measure_error(measure) if average: return Eq.mean() else: return Eq
3,376
26.680328
82
py
sampling_cf
sampling_cf-main/pytorch_models/SASRec.py
import torch import numpy as np import torch.nn as nn from torch_utils import LongTensor, BoolTensor, is_cuda_available class PointWiseFeedForward(nn.Module): def __init__(self, hidden_units, dropout_rate): super(PointWiseFeedForward, self).__init__() self.conv1 = nn.Conv1d(hidden_units, hidden_units, kernel_size=1) self.dropout1 = nn.Dropout(p=dropout_rate) self.relu = nn.ReLU() self.conv2 = nn.Conv1d(hidden_units, hidden_units, kernel_size=1) self.dropout2 = nn.Dropout(p=dropout_rate) def forward(self, inputs): outputs = self.dropout2(self.conv2(self.relu(self.dropout1(self.conv1(inputs.transpose(-1, -2)))))) outputs = outputs.transpose(-1, -2) # as Conv1D requires (N, C, Length) outputs += inputs return outputs class SASRec(torch.nn.Module): def __init__(self, hyper_params): super(SASRec, self).__init__() self.hyper_params = hyper_params self.item_num = hyper_params['total_items'] self.item_emb = torch.nn.Embedding(self.item_num+1, hyper_params['latent_size'], padding_idx=self.item_num) self.pos_emb = torch.nn.Embedding(hyper_params['max_seq_len'], hyper_params['latent_size']) # TO IMPROVE self.emb_dropout = torch.nn.Dropout(p=hyper_params['dropout']) self.attention_layernorms = torch.nn.ModuleList() # to be Q for self-attention self.attention_layers = torch.nn.ModuleList() self.forward_layernorms = torch.nn.ModuleList() self.forward_layers = torch.nn.ModuleList() self.last_layernorm = torch.nn.LayerNorm(hyper_params['latent_size'], eps=1e-8) for _ in range(hyper_params['num_blocks']): new_attn_layernorm = torch.nn.LayerNorm(hyper_params['latent_size'], eps=1e-8) self.attention_layernorms.append(new_attn_layernorm) new_attn_layer = torch.nn.MultiheadAttention( hyper_params['latent_size'], hyper_params['num_heads'], hyper_params['dropout'] ) self.attention_layers.append(new_attn_layer) new_fwd_layernorm = torch.nn.LayerNorm(hyper_params['latent_size'], eps=1e-8) self.forward_layernorms.append(new_fwd_layernorm) new_fwd_layer = PointWiseFeedForward(hyper_params['latent_size'], hyper_params['dropout']) self.forward_layers.append(new_fwd_layer) def log2feats(self, log_seqs): seqs = self.item_emb(log_seqs) seqs *= self.item_emb.embedding_dim ** 0.5 positions = np.tile(np.array(range(log_seqs.shape[1])), [log_seqs.shape[0], 1]) seqs += self.pos_emb(LongTensor(positions)) seqs = self.emb_dropout(seqs) timeline_mask = BoolTensor(log_seqs == self.item_num) seqs *= ~timeline_mask.unsqueeze(-1) # broadcast in last dim tl = seqs.shape[1] # time dim len for enforce causality temp = torch.ones((tl, tl), dtype=torch.bool) if is_cuda_available: temp = temp.cuda() attention_mask = ~torch.tril(temp) for i in range(len(self.attention_layers)): seqs = torch.transpose(seqs, 0, 1) Q = self.attention_layernorms[i](seqs) mha_outputs, _ = self.attention_layers[i](Q, seqs, seqs, attn_mask=attention_mask) # key_padding_mask=timeline_mask # need_weights=False) this arg do not work? seqs = Q + mha_outputs seqs = torch.transpose(seqs, 0, 1) seqs = self.forward_layernorms[i](seqs) seqs = self.forward_layers[i](seqs) seqs *= ~timeline_mask.unsqueeze(-1) log_feats = self.last_layernorm(seqs) # (U, T, C) -> (U, -1, C) return log_feats def get_score(self, log_feats, items): embs = self.item_emb(items) return (log_feats * embs).sum(dim=-1) def forward(self, data, eval = False): log_seqs, pos_seqs, neg_seqs = data # Embed sequence log_feats = self.log2feats(log_seqs) # user_ids hasn't been used yet if eval: log_feats = log_feats[:, -1, :] # Rank all items if pos_seqs is None: return torch.matmul( log_feats, self.item_emb.weight.transpose(0, 1) )[:, :-1] # Sampled evaluation orig_shape = neg_seqs.shape return self.get_score(log_feats.unsqueeze(1), pos_seqs), self.get_score( log_feats.unsqueeze(1).repeat(1, orig_shape[1], 1), neg_seqs ).view(orig_shape) # Dot product orig_shape = neg_seqs.shape return self.get_score(log_feats, pos_seqs).unsqueeze(-1).repeat(1, 1, orig_shape[2]), \ self.get_score( log_feats.unsqueeze(2).repeat(1, 1, orig_shape[2], 1).view(orig_shape[0], orig_shape[1] * orig_shape[2], -1), neg_seqs.view(orig_shape[0], orig_shape[1] * orig_shape[2]) ).view(orig_shape)
5,147
40.853659
122
py
sampling_cf
sampling_cf-main/pytorch_models/NeuMF.py
import torch import torch.nn as nn from pytorch_models.MF import BaseMF class GMF(BaseMF): def __init__(self, hyper_params): super(GMF, self).__init__(hyper_params) self.final = nn.Linear(hyper_params['latent_size'], 1) self.dropout = nn.Dropout(hyper_params['dropout']) def get_score(self, user_id, item_id): # For the FM user_bias = self.user_bias.gather(0, user_id.view(-1)).view(user_id.shape) item_bias = self.item_bias.gather(0, item_id.view(-1)).view(item_id.shape) # Embed Latent space user = self.dropout(self.user_embedding(user_id.view(-1))) item = self.dropout(self.item_embedding(item_id.view(-1))) joint = user * item rating = self.final(joint)[:, 0].view(user_id.shape) # [bsz] return user_bias + item_bias + self.global_bias + rating class MLP(BaseMF): def __init__(self, hyper_params): super(MLP, self).__init__(hyper_params) self.project = nn.Sequential( nn.Dropout(hyper_params['dropout']), nn.Linear(2 * hyper_params['latent_size'], hyper_params['latent_size']), nn.ReLU(), nn.Linear(hyper_params['latent_size'], hyper_params['latent_size']) ) self.final = nn.Linear(hyper_params['latent_size'], 1) self.dropout = nn.Dropout(hyper_params['dropout']) def get_score(self, user_id, item_id): # For the FM user_bias = self.user_bias.gather(0, user_id.view(-1)).view(user_id.shape) item_bias = self.item_bias.gather(0, item_id.view(-1)).view(item_id.shape) # Embed Latent space user = self.dropout(self.user_embedding(user_id.view(-1))) item = self.dropout(self.item_embedding(item_id.view(-1))) joint = torch.cat([ user, item ], dim = -1) joint = self.project(joint) rating = self.final(joint)[:, 0].view(user_id.shape) return user_bias + item_bias + self.global_bias + rating class NeuMF(BaseMF): def __init__(self, hyper_params): super(NeuMF, self).__init__(hyper_params, keep_gamma = False) self.gmf_user_embedding = nn.Embedding(hyper_params['total_users'], hyper_params['latent_size']) self.gmf_item_embedding = nn.Embedding(hyper_params['total_items'], hyper_params['latent_size']) self.mlp_user_embedding = nn.Embedding(hyper_params['total_users'], hyper_params['latent_size']) self.mlp_item_embedding = nn.Embedding(hyper_params['total_items'], hyper_params['latent_size']) self.project = nn.Sequential( nn.Dropout(hyper_params['dropout']), nn.Linear(2 * hyper_params['latent_size'], hyper_params['latent_size']), nn.ReLU(), nn.Linear(hyper_params['latent_size'], hyper_params['latent_size']) ) self.final = nn.Linear(2 * hyper_params['latent_size'], 1) self.dropout = nn.Dropout(hyper_params['dropout']) def init(self, gmf_model, mlp_model): with torch.no_grad(): self.gmf_user_embedding.weight.data = gmf_model.user_embedding.weight.data self.gmf_item_embedding.weight.data = gmf_model.item_embedding.weight.data self.mlp_user_embedding.weight.data = mlp_model.user_embedding.weight.data self.mlp_item_embedding.weight.data = mlp_model.item_embedding.weight.data for i in range(len(self.project)): try: self.project[i].weight.data = mlp_model.project[i].weight.data self.project[i].bias.data = mlp_model.project[i].bias.data except: pass self.final.weight.data = torch.cat([ gmf_model.final.weight.data, mlp_model.final.weight.data ], dim = -1) self.final.bias.data = 0.5 * (gmf_model.final.bias.data + mlp_model.final.bias.data) self.user_bias.data = 0.5 * (gmf_model.user_bias.data + mlp_model.user_bias.data) self.item_bias.data = 0.5 * (gmf_model.item_bias.data + mlp_model.item_bias.data) def get_score(self, user_id, item_id): # For the FM user_bias = self.user_bias.gather(0, user_id.view(-1)).view(user_id.shape) item_bias = self.item_bias.gather(0, item_id.view(-1)).view(item_id.shape) # GMF Part user = self.dropout(self.gmf_user_embedding(user_id.view(-1))) # [bsz x 32] item = self.dropout(self.gmf_item_embedding(item_id.view(-1))) # [bsz x 32] gmf_joint = user * item # MLP Part user = self.dropout(self.mlp_user_embedding(user_id.view(-1))) # [bsz x 32] item = self.dropout(self.mlp_item_embedding(item_id.view(-1))) # [bsz x 32] mlp_joint = torch.cat([ user, item ], dim = -1) mlp_joint = self.project(mlp_joint) # NeuMF final = torch.cat([ gmf_joint, mlp_joint ], dim = -1) rating = self.final(final)[:, 0].view(user_id.shape) # [bsz] return user_bias + item_bias + self.global_bias + rating
5,009
43.732143
118
py
sampling_cf
sampling_cf-main/pytorch_models/SVAE.py
import torch import numpy as np import torch.nn as nn from torch.autograd import Variable from torch_utils import is_cuda_available class Encoder(nn.Module): def __init__(self, hyper_params): super(Encoder, self).__init__() self.linear1 = nn.Linear( hyper_params['latent_size'], hyper_params['latent_size'] ) nn.init.xavier_normal_(self.linear1.weight) self.activation = nn.Tanh() def forward(self, x): x = self.linear1(x) x = self.activation(x) return x class Decoder(nn.Module): def __init__(self, hyper_params): super(Decoder, self).__init__() self.linear1 = nn.Linear(hyper_params['latent_size'], hyper_params['latent_size']) self.linear2 = nn.Linear(hyper_params['latent_size'], hyper_params['total_items']) nn.init.xavier_normal_(self.linear1.weight) nn.init.xavier_normal_(self.linear2.weight) self.activation = nn.Tanh() def forward(self, x): x = self.linear1(x) x = self.activation(x) x = self.linear2(x) return x class SVAE(nn.Module): def __init__(self, hyper_params): super(SVAE, self).__init__() self.hyper_params = hyper_params self.encoder = Encoder(hyper_params) self.decoder = Decoder(hyper_params) self.item_embed = nn.Embedding( hyper_params['total_items'] + 1, hyper_params['latent_size'], padding_idx=hyper_params['total_items'] ) self.gru = nn.GRU( hyper_params['latent_size'], hyper_params['latent_size'], batch_first = True, num_layers = 1 ) self.linear1 = nn.Linear(hyper_params['latent_size'], 2 * hyper_params['latent_size']) nn.init.xavier_normal_(self.linear1.weight) self.tanh = nn.Tanh() def sample_latent(self, h_enc): temp_out = self.linear1(h_enc) mu = temp_out[:, :self.hyper_params['latent_size']] log_sigma = temp_out[:, self.hyper_params['latent_size']:] sigma = torch.exp(log_sigma) std_z = torch.from_numpy(np.random.normal(0, 1, size=sigma.size())).float() if is_cuda_available: std_z = std_z.cuda() return mu + sigma * Variable(std_z, requires_grad=False), mu, log_sigma # Reparameterization trick def forward(self, data, eval = False): x, _, _ = data in_shape = x.shape # [bsz x seq_len] x = self.item_embed(x) # [bsz x seq_len x latent_size] rnn_out, _ = self.gru(x) # [bsz x seq_len x latent_size] rnn_out = rnn_out.contiguous().view(in_shape[0] * in_shape[1], -1) # [bsz * seq_len x latent_size] enc_out = self.encoder(rnn_out) # [bsz * seq_len x latent_size] sampled_z, z_mean, z_log_sigma = self.sample_latent(enc_out) # [bsz * seq_len x latent_size] dec_out = self.decoder(sampled_z) # [bsz * seq_len x total_items] dec_out = dec_out.view(in_shape[0], in_shape[1], -1) # [bsz x seq_len x total_items] return dec_out, z_mean, z_log_sigma
3,388
38.406977
107
py
sampling_cf
sampling_cf-main/pytorch_models/MVAE.py
import torch import numpy as np import torch.nn as nn from torch.autograd import Variable from torch_utils import is_cuda_available class Encoder(nn.Module): def __init__(self, hyper_params): super(Encoder, self).__init__() self.linear1 = nn.Linear( hyper_params['total_items'], hyper_params['latent_size'] ) nn.init.xavier_normal_(self.linear1.weight) self.activation = nn.Tanh() def forward(self, x): x = self.linear1(x) x = self.activation(x) return x class Decoder(nn.Module): def __init__(self, hyper_params): super(Decoder, self).__init__() self.linear1 = nn.Linear(hyper_params['latent_size'], hyper_params['latent_size']) self.linear2 = nn.Linear(hyper_params['latent_size'], hyper_params['total_items']) nn.init.xavier_normal_(self.linear1.weight) nn.init.xavier_normal_(self.linear2.weight) self.activation = nn.Tanh() def forward(self, x): x = self.linear1(x) x = self.activation(x) x = self.linear2(x) return x class MVAE(nn.Module): def __init__(self, hyper_params): super(MVAE, self).__init__() self.hyper_params = hyper_params self.encoder = Encoder(hyper_params) self.decoder = Decoder(hyper_params) self.linear1 = nn.Linear(hyper_params['latent_size'], 2 * hyper_params['latent_size']) nn.init.xavier_normal_(self.linear1.weight) self.tanh = nn.Tanh() def sample_latent(self, h_enc): temp_out = self.linear1(h_enc) mu = temp_out[:, :self.hyper_params['latent_size']] log_sigma = temp_out[:, self.hyper_params['latent_size']:] sigma = torch.exp(log_sigma) std_z = torch.from_numpy(np.random.normal(0, 1, size=sigma.size())).float() if is_cuda_available: std_z = std_z.cuda() return mu + sigma * Variable(std_z, requires_grad=False), mu, log_sigma # Reparameterization trick def forward(self, data, eval = False): x, _, _ = data enc_out = self.encoder(x) sampled_z, z_mean, z_log_sigma = self.sample_latent(enc_out) dec_out = self.decoder(sampled_z) return dec_out, z_mean, z_log_sigma
2,331
32.797101
107
py
sampling_cf
sampling_cf-main/pytorch_models/MF.py
import torch import torch.nn as nn import torch.nn.functional as F from torch_utils import LongTensor, FloatTensor class BaseMF(nn.Module): def __init__(self, hyper_params, keep_gamma = True): super(BaseMF, self).__init__() self.hyper_params = hyper_params # Declaring alpha, beta, gamma self.global_bias = nn.Parameter(FloatTensor([ 4.0 if hyper_params['task'] == 'explicit' else 0.5 ])) self.user_bias = nn.Parameter(FloatTensor([ 0.0 for _ in range(hyper_params['total_users']) ])) self.item_bias = nn.Parameter(FloatTensor([ 0.0 for _ in range(hyper_params['total_items']) ])) if keep_gamma: self.user_embedding = nn.Embedding(hyper_params['total_users'], hyper_params['latent_size']) self.item_embedding = nn.Embedding(hyper_params['total_items'], hyper_params['latent_size']) # For faster evaluation self.all_items_vector = LongTensor( list(range(hyper_params['total_items'])) ) def get_score(self, data): pass # Virtual function, implement in all sub-classes def forward(self, data, eval = False): user_id, pos_item_id, neg_items = data # Evaluation -- Rank all items if pos_item_id is None: ret = [] for b in range(user_id.shape[0]): ret.append(self.get_score( user_id[b].unsqueeze(-1).repeat(1, self.hyper_params['total_items']).view(-1), self.all_items_vector.view(-1) ).view(1, -1)) return torch.cat(ret) # Explicit feedback if neg_items is None: return self.get_score(user_id, pos_item_id.squeeze(-1)) # Implicit feedback return self.get_score( user_id.unsqueeze(-1).repeat(1, pos_item_id.shape[1]).view(-1), pos_item_id.view(-1) ).view(pos_item_id.shape), self.get_score( user_id.unsqueeze(-1).repeat(1, neg_items.shape[1]).view(-1), neg_items.view(-1) ).view(neg_items.shape) class MF(BaseMF): def __init__(self, hyper_params): keep_gamma = hyper_params['model_type'] != 'bias_only' super(MF, self).__init__(hyper_params, keep_gamma = keep_gamma) if keep_gamma: self.dropout = nn.Dropout(hyper_params['dropout']) if hyper_params['model_type'] == 'MF': latent_size = hyper_params['latent_size'] self.projection = nn.Sequential( nn.Dropout(hyper_params['dropout']), nn.Linear(2 * latent_size, latent_size), nn.ReLU(), nn.Linear(latent_size, latent_size) ) for m in self.projection: if isinstance(m, nn.Linear): nn.init.xavier_uniform_(m.weight) self.final = nn.Linear(2 * latent_size, 1) self.sigmoid = nn.Sigmoid() self.relu = nn.ReLU() def get_score(self, user_id, item_id): # For the FM user_bias = self.user_bias.gather(0, user_id.view(-1)).view(user_id.shape) item_bias = self.item_bias.gather(0, item_id.view(-1)).view(item_id.shape) if self.hyper_params['model_type'] == 'bias_only': return user_bias + item_bias + self.global_bias # Embed Latent space user = self.dropout(self.user_embedding(user_id.view(-1))) # [bsz x 32] item = self.dropout(self.item_embedding(item_id.view(-1))) # [bsz x 32] # Dot product if self.hyper_params['model_type'] == 'MF_dot': rating = torch.sum(user * item, dim = -1).view(user_id.shape) return user_bias + item_bias + self.global_bias + rating mf_vector = user * item cat = torch.cat([ user, item ], dim = -1) mlp_vector = self.projection(cat) # Concatenate and get single score cat = torch.cat([ mlp_vector, mf_vector ], dim = -1) rating = self.final(cat)[:, 0].view(user_id.shape) # [bsz] return user_bias + item_bias + self.global_bias + rating
4,075
38.960784
108
py