AttributeError: 'ASTModel' object has no attribute 'unfold'

[michaelschwob]

I'm relatively new to Python, so this may be a simple error to debug. However, I'm working through the example in the README to understand what the output of these models is like. My python version is 3.9.18 on macOS Sonoma 14.2.1 with M1 Max chip; I tried to get python 3.7.4, but that is not feasible with the M1 chip -- hopefully, 3.9.18 will suffice and is not the cause of this error.

Below is my MWE:

###
### Module Management
###

## needed to use ASTModel class
import sys
sys.path.insert(0, './')
from ast_models import *

###
### Example from Repo
###

# pretraining stage
# suppose you have an unlabled dataset with avg length of 1024 frames (i.e., 10.24s)
input_tdim = 1024
# create a 16*16 patch based AST model for pretraining.
# note, we don't use patch split overlap in pretraining, so fstride=fshape and tstride=tshape
ast_mdl = ASTModel(
             fshape=16, tshape=16, fstride=16, tstride=16,
             input_fdim=128, input_tdim=input_tdim, model_size='base',
             pretrain_stage=True)
# # alternatively, create a frame based AST model
# ast_mdl = ASTModel(
#              fshape=128, tshape=2, fstride=128, tstride=2,
#              input_fdim=128, input_tdim=input_tdim, model_size='base',
#              pretrain=True)

# do pretraining, see src/traintest_mask.py for our full pretraining code
# input in shape [batch_size, input_tdim, input_fdim]
test_input = torch.zeros([10, input_tdim, 128])
# mask 100 patches for both discriminative and generative loss
acc, nce_loss = ast_mdl(test_input, task='pretrain_mpc', mask_patch=100)
mse_loss = ast_mdl(test_input, task='pretrain_mpg', mask_patch=100)
loss = nce_loss + 10 * mse_loss
# do back propagate and update the model, etc

# after pretraining, save the pretrained model.
# the code is designed for Dataparallel model
ast_mdl = torch.nn.DataParallel(ast_mdl)
torch.save(ast_mdl.state_dict(), './test_mdl.pth')

# fine-tuning stage
# now you have a labeled dataset you want to finetune AST on
# suppose the avg length is 100 frames (1s) and there are 35 classes
# the fshape and tshape must be same in pretraining and finetuning
# but fstride and tstride can be different in pretraining and finetuning
# using smaller strides improves the performance but also increase the computational overhead
# set pretrain_stage as False since now is in the finetuning stage
# provide the path of the pretrained model you want to load
input_tdim = 100  # fine-tuning data length can be different with pretraining data length
ast_mdl = ASTModel(label_dim=35,
             fshape=16, tshape=16, fstride=10, tstride=10,
             input_fdim=128, input_tdim=input_tdim, model_size='base',
             pretrain_stage=False, load_pretrained_mdl_path='./test_mdl.pth')
# # alternatively, use a frame based AST model
# ast_mdl = ASTModel(label_dim=35,
#              fshape=128, tshape=2, fstride=128, tstride=1,
#              input_fdim=128, input_tdim=input_tdim, model_size='base',
#              pretrain_stage=False, load_pretrained_mdl_path='./test_mdl.pth')

# do finetuning, see src/traintest.py for our finetuning code
test_input = torch.zeros([10, input_tdim, 128])
prediction = ast_mdl(test_input, task='ft_avgtok')
# output should in shape [batch_size, label_dim]
print(prediction.shape)
# calculate the loss, do back propagate, etc

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
test_input = torch.zeros([1, input_tdim, 128]).to(device)
acc, nce = ast_mdl(test_input, task='pretrain_mpc', mask_patch=100)
# you can visualize the mask
pred, masked = ast_mdl(test_input, task='visualize_mask', mask_patch=100)
plt.imshow(masked[0,0])
plt.show()

Here, both calls to ast_mdl() result in the following error:

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/schwob/.pyenv/versions/3.9.18/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "/Users/schwob/.pyenv/versions/3.9.18/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1520, in _call_impl
    return forward_call(*args, **kwargs)
  File "/Users/schwob/Desktop/ssast/src/models/ast_models.py", line 448, in forward
    return self.mpc(x, mask_patch=mask_patch, cluster=cluster)
  File "/Users/schwob/Desktop/ssast/src/models/ast_models.py", line 293, in mpc
    input = self.unfold(x).transpose(1, 2)
  File "/Users/schwob/.pyenv/versions/3.9.18/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1688, in __getattr__
    raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
AttributeError: 'ASTModel' object has no attribute 'unfold'

The unfolding error only occurs once we define ast_mdl as a fine-tuning model.

Ultimately, my goal is to plot the probabilities of class assignment for the test input, so if there's a simpler way to do this, I am all ears.

Thank you for your time.

[michaelschwob]

Update: I'm still getting the same error using the same code with Python 3.7.4 on an Ubuntu laptop, so it doesn't seem that the python or module versions are the issue.

[YuanGongND]

did you change anything from my code?

[michaelschwob]

Aside from the "module management" section in the quoted code snippet, I have not.

[YuanGongND]

Can you run this? https://colab.research.google.com/github/YuanGongND/ast/blob/master/colab/AST_Inference_Demo.ipynb

[michaelschwob]

Yes, I can run the AST Colab script perfectly on my end.

[YuanGongND]

This looks like a package version issue, either with torch or timm. What is the version of your timm package? It has to be 0.4.5

[michaelschwob]

My timm package is 0.4.5. I think I'm seeing the cause for the error. Below is your ast_models.py file with a few lines I added (tagged with "# added !!"). It seems that when pretrain_stage == False, there are attributes that are not defined that are being called in the mpc function.

# -*- coding: utf-8 -*-
# @Time    : 7/16/21 3:12 AM
# @Author  : Yuan Gong
# @Affiliation  : Massachusetts Institute of Technology
# @Email   : [email protected]
# @File    : ast_models.py

# the unified ast models for all pretraining/fine-tuning tasks.

import torch.nn as nn
import torch
import sys
sys.path.append("/data/sls/scratch/yuangong/aed-trans/src/models/")
sys.path.append("/data/sls/scratch/yuangong/aed-trans/src/")
from timm.models.layers import trunc_normal_
import timm
import numpy as np
from timm.models.layers import to_2tuple
from random import randrange
from matplotlib import pyplot as plt
import random

# override the timm package to relax the input shape constraint.
class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)

    def forward(self, x):
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x

def get_sinusoid_encoding(n_position, d_hid):
    ''' Sinusoid position encoding table '''

    def get_position_angle_vec(position):
        return [position / np.power(10000, 2 * (hid_j // 2) / d_hid) for hid_j in range(d_hid)]

    sinusoid_table = np.array([get_position_angle_vec(pos_i) for pos_i in range(n_position)])
    sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])  # dim 2i
    sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])  # dim 2i+1

    return torch.FloatTensor(sinusoid_table).unsqueeze(0)

class ASTModel(nn.Module):
    def __init__(self, label_dim=527,
                 fshape=128, tshape=2, fstride=128, tstride=2,
                 input_fdim=128, input_tdim=1024, model_size='base',
                 pretrain_stage=True, load_pretrained_mdl_path=None):

        super(ASTModel, self).__init__()
        assert timm.__version__ == '0.4.5', 'Please use timm == 0.4.5, the code might not be compatible with newer versions.'

        # override timm input shape restriction
        timm.models.vision_transformer.PatchEmbed = PatchEmbed

        # pretrain the AST models
        if pretrain_stage == True:
            if load_pretrained_mdl_path != None:
                raise ValueError('Setting load_pretrained_mdl_path at pretraining stage is useless, pretraining is always from scratch, please change it to None.')
            if fstride != fshape or tstride != tshape:
                raise ValueError('fstride != fshape or tstride != tshape, they must be same at the pretraining stage, patch split overlapping is not supported.')

            # if AudioSet pretraining is not used (but ImageNet pretraining may still apply)
            if model_size == 'tiny':
                self.v = timm.create_model('vit_deit_tiny_distilled_patch16_224', pretrained=False)
                self.heads, self.depth = 3, 12
                self.cls_token_num = 2
            elif model_size == 'small':
                self.v = timm.create_model('vit_deit_small_distilled_patch16_224', pretrained=False)
                self.heads, self.depth = 6, 12
                self.cls_token_num = 2
            elif model_size == 'base':
                self.v = timm.create_model('vit_deit_base_distilled_patch16_384', pretrained=False)
                self.heads, self.depth = 12, 12
                self.cls_token_num = 2
            elif model_size == 'base_nokd':
                self.v = timm.create_model('vit_deit_base_patch16_384', pretrained=False)
                self.heads, self.depth = 12, 12
                self.cls_token_num = 1
            else:
                raise Exception('Model size must be one of tiny, small, base, base_nokd')

            self.original_num_patches = self.v.patch_embed.num_patches
            self.oringal_hw = int(self.original_num_patches ** 0.5)
            self.original_embedding_dim = self.v.pos_embed.shape[2]

            # SSL Pretraining Code
            self.softmax = nn.Softmax(dim=-1)
            self.lsoftmax = nn.LogSoftmax(dim=-1)
            self.fshape, self.tshape = fshape, tshape
            self.fstride, self.tstride = fstride, tstride
            self.input_fdim, self.input_tdim = input_fdim, input_tdim
            # this is a trick to make state_dict to track pretraining input_fdim and input_tdim and save them by using torch.save
            self.p_input_fdim, self.p_input_tdim = nn.Parameter(torch.tensor(input_fdim), requires_grad=False), nn.Parameter(torch.tensor(input_tdim), requires_grad=False)

            # masked patch classification (discriminative objective) layer
            # we use two layers for pretext task, but using a single layer has similar performance.
            # we map the output of transformer (768-dim for base models) to 256-dim patch input space, and then dot product with flattened patch input (also 256-dim) to calculate loss.
            # alternatively, you can map the output of transformer to 768-dim patch embedding space, and dot product with patch embedding. Performance-wise they are similar, but map to 256 space is more efficient.
            self.cpredlayer = nn.Sequential(nn.Linear(self.original_embedding_dim, self.original_embedding_dim), nn.ReLU(), nn.Linear(self.original_embedding_dim, 256))
            # masked patch reconstruction (generative objective) layer
            self.gpredlayer = nn.Sequential(nn.Linear(self.original_embedding_dim, self.original_embedding_dim), nn.ReLU(), nn.Linear(self.original_embedding_dim, 256))
            self.unfold = torch.nn.Unfold(kernel_size=(fshape, tshape), stride=(fstride, tstride))

            # we use learnable mask embedding (follow the BEIT paper), but using a fixed mask embedding (e.g., 0) leads to same performance.
            self.mask_embed = nn.Parameter(torch.zeros([1, 1, self.original_embedding_dim]))
            self.mask_embed = torch.nn.init.xavier_normal_(self.mask_embed)

            # get the intermediate shape
            self.p_f_dim, self.p_t_dim = self.get_shape(fstride, tstride, input_fdim, input_tdim, fshape, tshape)
            num_patches = self.p_f_dim * self.p_t_dim
            self.num_patches = num_patches
            self.v.patch_embed.num_patches = num_patches
            print('pretraining patch split stride: frequency={:d}, time={:d}'.format(fstride, tstride))
            print('pretraining patch shape: frequency={:d}, time={:d}'.format(fshape, tshape))
            print('pretraining patch array dimension: frequency={:d}, time={:d}'.format(self.p_f_dim, self.p_t_dim))
            print('pretraining number of patches={:d}'.format(num_patches))

            # the linear patch projection layer, use 1 channel for spectrogram rather than the original 3 channels for RGB images.
            new_proj = torch.nn.Conv2d(1, self.original_embedding_dim, kernel_size=(fshape, tshape), stride=(fstride, tstride))
            self.v.patch_embed.proj = new_proj

            # use trainable positional embedding
            new_pos_embed = nn.Parameter(torch.zeros(1, self.v.patch_embed.num_patches + self.cls_token_num, self.original_embedding_dim))
            self.v.pos_embed = new_pos_embed
            trunc_normal_(self.v.pos_embed, std=.02)

        # use a pretrained models for finetuning
        elif pretrain_stage == False:
            self.unfold = torch.nn.Unfold(kernel_size=(fshape, tshape), stride=(fstride, tstride)) # added !!
            device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
            if load_pretrained_mdl_path == None:
                raise ValueError('Please set load_pretrained_mdl_path to load a pretrained models.')
            sd = torch.load(load_pretrained_mdl_path, map_location=device)
            # get the fshape and tshape, input_fdim and input_tdim in the pretraining stage
            try:
                p_fshape, p_tshape = sd['module.v.patch_embed.proj.weight'].shape[2], sd['module.v.patch_embed.proj.weight'].shape[3]
                p_input_fdim, p_input_tdim = sd['module.p_input_fdim'].item(), sd['module.p_input_tdim'].item()
            except:
                raise  ValueError('The model loaded is not from a torch.nn.Dataparallel object. Wrap it with torch.nn.Dataparallel and try again.')

            print('now load a SSL pretrained models from ' + load_pretrained_mdl_path)
            # during pretraining, fstride=fshape and tstride=tshape because no patch overlapping is used
            # here, input_fdim and input_tdim should be that used in pretraining, not that in the fine-tuning.
            # we need to know input_fdim and input_tdim to do positional embedding cut/interpolation.
            # generally it should be better to use same input_fdim during pretraining and finetuning, but input_tdim can be safely different
            audio_model = ASTModel(fstride=p_fshape, tstride=p_tshape, fshape=p_fshape, tshape=p_tshape,
                                   input_fdim=p_input_fdim, input_tdim=p_input_tdim, pretrain_stage=True, model_size=model_size)
            audio_model = torch.nn.DataParallel(audio_model)
            audio_model.load_state_dict(sd, strict=False)

            self.v = audio_model.module.v
            self.original_embedding_dim = self.v.pos_embed.shape[2]
            self.cls_token_num = audio_model.module.cls_token_num

            # mlp head for fine-tuning
            self.mlp_head = nn.Sequential(nn.LayerNorm(self.original_embedding_dim),
                                          nn.Linear(self.original_embedding_dim, label_dim))

            f_dim, t_dim = self.get_shape(fstride, tstride, input_fdim, input_tdim, fshape, tshape)
            # patch array dimension during pretraining
            p_f_dim, p_t_dim = audio_model.module.p_f_dim, audio_model.module.p_t_dim
            self.p_f_dim = p_f_dim # added !!
            self.p_t_dim = p_t_dim # added !!
            num_patches = f_dim * t_dim
            self.num_patches = num_patches # added !!
            p_num_patches = p_f_dim * p_t_dim
            self.v.patch_embed.num_patches = num_patches
            print('fine-tuning patch split stride: frequncey={:d}, time={:d}'.format(fstride, tstride))
            print('fine-tuning number of patches={:d}'.format(num_patches))

            # patch shape should be same for pretraining and fine-tuning
            if fshape != p_fshape or tshape != p_tshape:
                raise ValueError('The patch shape of pretraining and fine-tuning is not consistant, pretraining: f={:d}, t={:d}, finetuning: f={:d}, t={:d}'.format(p_fshape, p_tshape, fshape, tshape))

            # patch split stride generally should be different for pretraining and fine-tuning, as patch split overlapping is only used in finetuning
            # during pretraining, p_fshape = p_fstride and p_tshape = p_tstride
            if fstride != p_fshape or tstride != p_tshape:
                # initialize a new patch embedding layer with desired new stride.
                new_proj = torch.nn.Conv2d(1, self.original_embedding_dim, kernel_size=(fshape, tshape), stride=(fstride, tstride))
                # but the weights of patch embedding layer is still got from the pretrained models
                new_proj.weight = torch.nn.Parameter(torch.sum(self.v.patch_embed.proj.weight, dim=1).unsqueeze(1))
                new_proj.bias = self.v.patch_embed.proj.bias
                self.v.patch_embed.proj = new_proj

            new_pos_embed = self.v.pos_embed[:, self.cls_token_num:, :].detach().reshape(1, p_num_patches, self.original_embedding_dim).transpose(1, 2).reshape(1, self.original_embedding_dim, p_f_dim, p_t_dim)
            # cut or interpolate the positional embedding
            if t_dim < p_t_dim:
                new_pos_embed = new_pos_embed[:, :, :, int(p_t_dim/2) - int(t_dim / 2): int(p_t_dim/2) - int(t_dim / 2) + t_dim]
            else:
                new_pos_embed = torch.nn.functional.interpolate(new_pos_embed, size=(8, t_dim), mode='bilinear')
            if f_dim < p_f_dim:
                new_pos_embed = new_pos_embed[:, :, int(p_f_dim/2) - int(f_dim / 2): int(p_f_dim/2) - int(f_dim / 2) + t_dim, :]
            else:
                new_pos_embed = torch.nn.functional.interpolate(new_pos_embed, size=(f_dim, t_dim), mode='bilinear')

            new_pos_embed = new_pos_embed.reshape(1, self.original_embedding_dim, num_patches).transpose(1, 2)
            self.v.pos_embed = nn.Parameter(torch.cat([self.v.pos_embed[:, :self.cls_token_num, :].detach(), new_pos_embed], dim=1))

    # get the shape of intermediate representation.
    def get_shape(self, fstride, tstride, input_fdim, input_tdim, fshape, tshape):
        test_input = torch.randn(1, 1, input_fdim, input_tdim)
        test_proj = nn.Conv2d(1, self.original_embedding_dim, kernel_size=(fshape, tshape), stride=(fstride, tstride))
        test_out = test_proj(test_input)
        f_dim = test_out.shape[2]
        t_dim = test_out.shape[3]
        return f_dim, t_dim

    # generate mask for 16*16 patch
    def gen_maskid_patch(self, sequence_len=512, mask_size=100, cluster=3):
        mask_id = []

        # randomize clutering factor in [3,6)
        cur_clus = randrange(cluster) + 3

        while len(list(set(mask_id))) <= mask_size:
            start_id = randrange(sequence_len)

            # this improves the efficiency, but might change the pretrained model
            # while start_id in mask_id:
            #     start_id = randrange(sequence_len)

            cur_mask = []
            for i in range(0, cur_clus):
                for j in range(0, cur_clus):
                    mask_cand = start_id + self.p_t_dim * i + j
                    if mask_cand > 0 and mask_cand < sequence_len:
                        cur_mask.append(mask_cand)
            mask_id = mask_id + cur_mask
        mask_id = list(set(mask_id))[:mask_size]
        return torch.tensor(mask_id)

    # using cluster for frame masking hurts the performance, so just use the naive random sampling
    def gen_maskid_frame(self, sequence_len=512, mask_size=100):
        mask_id = random.sample(range(0, sequence_len), mask_size)
        return torch.tensor(mask_id)

    def finetuningavgtok(self, x):
        B = x.shape[0]
        x = self.v.patch_embed(x)
        if self.cls_token_num == 2:
            cls_tokens = self.v.cls_token.expand(B, -1, -1)
            dist_token = self.v.dist_token.expand(B, -1, -1)
            x = torch.cat((cls_tokens, dist_token, x), dim=1)
        else:
            cls_tokens = self.v.cls_token.expand(B, -1, -1)
            x = torch.cat((cls_tokens, x), dim=1)
        x = x + self.v.pos_embed
        x = self.v.pos_drop(x)

        for blk_id, blk in enumerate(self.v.blocks):
            x = blk(x)
        x = self.v.norm(x)

        # average output of all tokens except cls token(s)
        x = torch.mean(x[:, self.cls_token_num:, :], dim=1)
        x = self.mlp_head(x)
        return x

    def finetuningcls(self, x):
        B = x.shape[0]
        x = self.v.patch_embed(x)
        if self.cls_token_num == 2:
            cls_tokens = self.v.cls_token.expand(B, -1, -1)
            dist_token = self.v.dist_token.expand(B, -1, -1)
            x = torch.cat((cls_tokens, dist_token, x), dim=1)
        else:
            cls_tokens = self.v.cls_token.expand(B, -1, -1)
            x = torch.cat((cls_tokens, x), dim=1)
        x = x + self.v.pos_embed
        x = self.v.pos_drop(x)

        for blk_id, blk in enumerate(self.v.blocks):
            x = blk(x)
        x = self.v.norm(x)

        # if models has two cls tokens (DEIT), average as the clip-level representation
        if self.cls_token_num == 2:
            x = (x[:, 0] + x[:, 1]) / 2
        else:
            x = x[:, 0]
        x = self.mlp_head(x)
        return x

    # masked patch pretraining with discriminative objective
    def mpc(self, x, mask_patch, cluster, show_mask=False):
        input = self.unfold(x).transpose(1, 2)
        B = x.shape[0]
        # x in shape (batch_size, sequence_len, embedding dim)
        x = self.v.patch_embed(x)

        # encode the patch
        # size 12(batch_size) * 100(#mask_patch) * 768(hidden_dim), prepare to save the true values of masked samples
        encode_samples = torch.empty((B, mask_patch, 256), device=x.device, requires_grad=False).float()
        # size 12(batch_size) * 100(#mask_patch), index of masked patches
        mask_index = torch.empty((B, mask_patch), device=x.device, requires_grad=False).long()
        # size 12(batch_size) * 512(sequence_len) * 768(hidden_dim)
        mask_dense = torch.ones([x.shape[0], x.shape[1], x.shape[2]], device=x.device)

        # for each audio clip in the batch
        for i in range(B):
            # randomly generate #mask_patch mask indexes without duplicate
            if cluster == True:
                # use this if you are masking e.g. 16*16 patches
                mask_index[i] = self.gen_maskid_patch(self.num_patches, mask_patch)
            else:
                # use this if you are masking frame, i.e., 128*2 patches
                mask_index[i] = self.gen_maskid_frame(self.num_patches, mask_patch)
            # copy the masked embeddings, note gradients are stopped in this path
            encode_samples[i] = input[i, mask_index[i], :].clone().detach()
            # mask the encode samples with 0
            mask_dense[i, mask_index[i], :] = 0

        # follow BEIT paper, mask with learnable masking embedding, but no performance diff observed compared with masking with 0s.
        mask_tokens = self.mask_embed.expand(B, x.shape[1], -1)

        # mask the patch
        x = x * mask_dense + (1-mask_dense) * mask_tokens

        # pass through the Transformer layers
        cls_tokens = self.v.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
        dist_token = self.v.dist_token.expand(B, -1, -1)
        x = torch.cat((cls_tokens, dist_token, x), dim=1)
        x = x + self.v.pos_embed
        x = self.v.pos_drop(x)
        for blk in self.v.blocks:
            x = blk(x)
        x = self.v.norm(x)

        # prediction of the masked patch
        pred = torch.empty((B, mask_patch, 256), device=x.device).float()  # e.g. size 12*100*768
        for i in range(B):
            #  +2 for indexes because skipping the cls and dis token
            # we map the output of transformer (768-dim for base models) to 256-dim patch input space, and then dot product with flattened patch input (also 256-dim) to calculate loss.
            # alternatively, you can map the output of transformer to 768-dim patch embedding space, and dot product with patch embedding. Performance-wise they are similar, but map to 256 space is more efficient.
            pred[i] = self.cpredlayer(x[i, mask_index[i] + self.cls_token_num, :])

        # calculate the NCE loss
        nce = torch.tensor(0.0).to(x.device)
        correct = torch.tensor(0.0).to(x.device)
        for i in np.arange(0, B):
            # negative samples are from the same batch
            # 8/12/2022: has a difference with equation (1) in the ssast paper but (likely) performance-wise similar, see https://github.com/YuanGongND/ssast/issues/13
            total = torch.mm(encode_samples[i], torch.transpose(pred[i], 0, 1))  # e.g. size 100*100
            correct += torch.sum(torch.eq(torch.argmax(self.softmax(total), dim=0), torch.arange(0, mask_patch, device=x.device)))  # correct is a tensor
            nce += torch.sum(torch.diag(self.lsoftmax(total)))  # nce is a tensor
        acc = 1. * correct / (B * mask_patch)
        nce = nce / (-1. * B * mask_patch)

        # visualize the masked area, for probing test only, set show_mask = False for any training/inference.
        if show_mask == False:
            return acc, nce
        else:
            if B > 1:
                raise Exception('Currently only support single spectrogram probing test.')

            self.mask_correct = torch.nn.Parameter(torch.arange(0, mask_patch), requires_grad=False)

            pred = input.clone()  # [B, 512, 256]
            masked = input.clone()

            for i in range(B):
                result = [float(t) * 99 for t in torch.eq(torch.argmax(self.softmax(total), dim=0), self.mask_correct)]
                pred[i, mask_index[i], :] = torch.tensor(result).reshape(mask_patch, 1).expand(mask_patch, 256)
                masked[i, mask_index[i], :] = 99.0

            # print(total)
            # print(self.softmax(total))
            # print(torch.argmax(self.softmax(total), dim=0))
            # print(self.mask_correct)
            # print(torch.eq(torch.argmax(self.softmax(total), dim=0), self.mask_correct))
            # print([float(t)*99 for t in torch.eq(torch.argmax(self.softmax(total), dim=0), self.mask_correct)])

            fold = torch.nn.Fold(output_size=([self.input_fdim, self.input_tdim]), kernel_size=(self.fshape, self.tshape), stride=(self.fstride, self.tstride))
            pred = fold(pred.transpose(1, 2))
            masked = fold(masked.transpose(1, 2))

            return pred, masked

    # # masked patch pretraining with generative objective
    def mpg(self, input, mask_patch, cluster):
        B = input.shape[0]
        x = self.v.patch_embed(input)
        input = self.unfold(input).transpose(1, 2)

        # size 12(batch_size) * 100(#mask_patch), index of masked patches
        mask_index = torch.empty((B, mask_patch), device=x.device, requires_grad=False).long()
        # size 12(batch_size) * 512(sequence_len) * 768(hidden_dim)
        mask_dense = torch.ones([x.shape[0], x.shape[1], x.shape[2]], device=x.device)
        for i in range(B):
            # randomly generate #mask_patch mask indexes without duplicate
            if cluster == True:
                # use this if you are masking e.g. 16*16 patches
                mask_index[i] = self.gen_maskid_patch(self.num_patches, mask_patch)
            else:
                # use this if you are masking frame, i.e., 128*2 patches
                mask_index[i] = self.gen_maskid_frame(self.num_patches, mask_patch)
            mask_dense[i, mask_index[i], :] = 0

        mask_tokens = self.mask_embed.expand(B, x.shape[1], -1)

        # follow BEIT paper, mask with learnable masking embedding, but no performance diff observed compared with masking with 0s.
        x = x * mask_dense + (1-mask_dense) * mask_tokens

        # go through the Transformer layers
        cls_tokens = self.v.cls_token.expand(B, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
        dist_token = self.v.dist_token.expand(B, -1, -1)
        x = torch.cat((cls_tokens, dist_token, x), dim=1)
        x = x + self.v.pos_embed
        x = self.v.pos_drop(x)
        for blk in self.v.blocks:
            x = blk(x)
        x = self.v.norm(x)

        pred = torch.empty((B, mask_patch, self.fshape * self.tshape), device=x.device).float()  # e.g. size 12*100*256
        target = torch.empty((B, mask_patch, self.fshape * self.tshape), device=x.device).float() # e.g. size 12*100*256

        for i in range(B):
            #  +2 for indexes because cls and dis token
            pred[i] = self.gpredlayer(x[i, mask_index[i] + self.cls_token_num, :])
            target[i] = input[i, mask_index[i], :]

        # calculate the MSE loss
        mse = torch.mean((pred - target) ** 2)

        return mse

    def forward(self, x, task, cluster=True, mask_patch=400):
        # expect input x = (batch_size, time_frame_num, frequency_bins), e.g., (12, 1024, 128)
        x = x.unsqueeze(1)
        x = x.transpose(2, 3)

        # finetuning (ft), use the mean of all token (patch) output as clip-level representation.
        # this is default for SSAST fine-tuning as during pretraining, supervision signal is given to each token, not the [cls] token
        if task == 'ft_avgtok':
            return self.finetuningavgtok(x)
        # alternatively, use the [cls] token output as clip-level representation.
        elif task == 'ft_cls':
            return self.finetuningcls(x)
        # pretraining, masked patch classification (discriminative objective)
        elif task == 'pretrain_mpc':
            return self.mpc(x, mask_patch=mask_patch, cluster=cluster)
        # pretraining, masked patch reconstruction (generative objective)
        elif task == 'pretrain_mpg':
            return self.mpg(x, mask_patch=mask_patch, cluster=cluster)
        elif task == 'visualize_mask':
            return self.mpc(x, mask_patch=mask_patch, cluster=cluster, show_mask=True)
        else:
            raise Exception('Task unrecognized.')

if __name__ == '__main__':
    # this is an example of how to use the SSAST model

    # pretraining stage
    # suppose you have an unlabled dataset with avg length of 1024 frames (i.e., 10.24s)
    input_tdim = 1024
    # create a 16*16 patch based AST model for pretraining.
    # note, we don't use patch split overlap in pretraining, so fstride=fshape and tstride=tshape
    ast_mdl = ASTModel(
                 fshape=16, tshape=16, fstride=16, tstride=16,
                 input_fdim=128, input_tdim=input_tdim, model_size='base',
                 pretrain_stage=True)
    # # alternatively, create a frame based AST model
    # ast_mdl = ASTModel(
    #              fshape=128, tshape=2, fstride=128, tstride=2,
    #              input_fdim=128, input_tdim=input_tdim, model_size='base',
    #              pretrain=True)

    # do pretraining, see src/traintest_mask.py for our full pretraining code
    # input in shape [batch_size, input_tdim, input_fdim]
    test_input = torch.zeros([10, input_tdim, 128])
    # mask 100 patches for both discriminative and generative loss
    acc, nce_loss = ast_mdl(test_input, task='pretrain_mpc', mask_patch=100)
    mse_loss = ast_mdl(test_input, task='pretrain_mpg', mask_patch=100)
    loss = nce_loss + 10 * mse_loss
    # do back propagate and update the model, etc

    # after pretraining, save the pretrained model.
    # the code is designed for Dataparallel model
    ast_mdl = torch.nn.DataParallel(ast_mdl)
    torch.save(ast_mdl.state_dict(), './test_mdl.pth')

    # fine-tuning stage
    # now you have a labeled dataset you want to finetune AST on
    # suppose the avg length is 100 frames (1s) and there are 35 classes
    # the fshape and tshape must be same in pretraining and finetuning
    # but fstride and tstride can be different in pretraining and finetuning
    # using smaller strides improves the performance but also increase the computational overhead
    # set pretrain_stage as False since now is in the finetuning stage
    # provide the path of the pretrained model you want to load
    input_tdim = 100  # fine-tuning data length can be different with pretraining data length
    ast_mdl = ASTModel(label_dim=35,
                 fshape=16, tshape=16, fstride=10, tstride=10,
                 input_fdim=128, input_tdim=input_tdim, model_size='base',
                 pretrain_stage=False, load_pretrained_mdl_path='./test_mdl.pth')
    # # alternatively, use a frame based AST model
    # ast_mdl = ASTModel(label_dim=35,
    #              fshape=128, tshape=2, fstride=128, tstride=1,
    #              input_fdim=128, input_tdim=input_tdim, model_size='base',
    #              pretrain_stage=False, load_pretrained_mdl_path='./test_mdl.pth')

    # do finetuning, see src/traintest.py for our finetuning code
    test_input = torch.zeros([10, input_tdim, 128])
    prediction = ast_mdl(test_input, task='ft_avgtok')
    # output should in shape [batch_size, label_dim]
    print(prediction.shape)
    # calculate the loss, do back propagate, etc

    # # (optional) do some probe test
    # test_input = torch.zeros([1, input_tdim, 128]).to(device)
    # acc, nce = ast_mdl(test_input, task='pretrain_mpc', mask_patch=100)
    # # you can visualize the mask
    # pred, masked = ast_mdl(test_input, task='visualize_mask', mask_patch=100)
    # plt.imshow(masked[0,0])
    # plt.show()

You may notice that I'm getting the errors when running the code you commented out at the end of this snippet. Perhaps its code I should not be using. My main objectives are (i) to be able to fine-tune the model with the pretrain_mpc and visualize_mask tasks and (ii) to understand the output of the ast_mdl() calls so that I can get probabilities for an audio recording belonging to each label. Is it as simple as using torch.softmax(prediction,dims=1)?

I have been using CNNs to classify audio recordings, and I am really intrigued with the ability to have self-supervised audio classification. I greatly appreciate your work and hope that I can accomplish these two objectives so that I can use your SSAST method for my project.

[michaelschwob]

Just to follow-up on this bug, it only came up because I tried to run the optional code you commented out at the end of the ast_models.py file. Perhaps the easiest solution would be to remove that comment.